Brookings Panel on Economic Activity

March 2021, 2014

Debt and Incomplete Financial Markets:

A Case for Nominal GDP Targeting
Kevin D. Sheedy, London School of Economics
Final conference draft

Abstract
Financial markets are incomplete, thus for many households borrowing is possible only by accepting a
financial contract that specifies a fixed repayment. However, the future income that will repay this debt
is uncertain, so risk can be inefficiently distributed. This paper argues that a monetary policy of nominal
GDP targeting can improve the functioning of incomplete financial markets when incomplete contracts
are written in terms of money. By insulating households nominal incomes from aggregate real shocks,
this policy effectively completes financial markets by stabilizing the ratio of debt to income. The paper
argues the objective of replicating complete financial markets should receive substantial weight even in
an environment with other frictions that have been used to justify a policy of strict inflation targeting.

Debt and Incomplete Financial Markets:

A Case for Nominal GDP Targeting
Kevin D. Sheedy
London School of Economics
First draft: 7th February 2012
This version: 10th March 2014

Abstract
Financial markets are incomplete, thus for many households borrowing is possible only by
accepting a financial contract that specifies a fixed repayment. However, the future income that
will repay this debt is uncertain, so risk can be inefficiently distributed. This paper argues that a
monetary policy of nominal GDP targeting can improve the functioning of incomplete financial
markets when incomplete contracts are written in terms of money. By insulating households
nominal incomes from aggregate real shocks, this policy effectively completes financial markets
by stabilizing the ratio of debt to income. The paper argues the objective of replicating
complete financial markets should receive substantial weight even in an environment with
other frictions that have been used to justify a policy of strict inflation targeting.
JEL classifications: E21; E31; E44; E52.
Keywords: incomplete markets; heterogeneous agents; risk sharing; nominal GDP targeting.
I

thank Carlos Carvalho, Wouter den Haan, Monique Ebell, Cosmin Ilut, Albert Marcet, Matthias
Paustian, David Romer, and George Selgin for helpful comments. The paper has also benefited from the

comments of seminar participants at Banque de France, U. Cambridge, CERGE-EI, Ecole

Polytechnique,
U. Lausanne, U. Maryland, National Bank of Serbia, New York Fed, U. Oxford, PUCRio, Sao Paulo School
of Economics, U. Southampton, U. St. Andrews, U. Warwick, the Anglo-French-Italian Macroeconomics
workshop, Birmingham Econometrics and Macroeconomics conference, Centre for Economic Performance
annual conference, Econometric Society North American summer meeting, EEA annual congress, ESSET,
ESSIM, Joint French Macro workshop, LACEA, LBS-CEPR conference Developments in Macroeconomics
and Finance, London Macroeconomics workshop, Midwest Macro Meeting, and NBER Summer Institute
in Monetary Economics.
LSE, CEP, CEPR, and CfM. Address: Department of Economics, London School of Economics and
Political Science, Houghton Street, London, WC2A 2AE, UK. Tel: +44 207 107 5022, Fax: +44 207 955
6592, Email: k.d.sheedy@lse.ac.uk, Website: http://personal.lse.ac.uk/sheedy.

Introduction

At the heart of any argument for a monetary policy strategy lies a view of what are the most
important frictions or market failures that monetary policy should seek to mitigate. The canonical
justification for inflation targeting as optimal monetary policy rests on the argument that pricing
frictions in goods markets are of particular concern (see, for example, Woodford, 2003). With
infrequent price adjustment owing to menu costs or other nominal rigidities, high or volatile inflation
leads to relative price distortions that impair the efficient operation of markets, and which directly
consumes time and resources in the process of setting prices. Inflation targeting is the appropriate
policy response to such frictions because it is able to move the economy closer to, or even replicate,
what the equilibrium would be if prices were flexible. In other words, inflation targeting is able to
undo or partially circumvent the frictions created by nominal price stickiness.1
This paper argues that nominal price stickiness may not be the most serious friction that monetary policy has to contend with. While the use of money as a unit of account in setting infrequently
adjusted goods prices is well documented, moneys role as a unit of account in writing financial
contracts is equally pervasive. Moreover, just as price stickiness means that nominal prices fail to
be fully state contingent, financial contracts are typically not contingent on all possible future states
of the world, for example, debt contracts that specify fixed nominal repayments. Financial contracts
might not be fully contingent for a variety of reasons, but one explanation could be that transaction
costs make it prohibitively expensive to write and enforce complicated and lengthy contracts. Many
agents, such as households, would find it difficult to issue liabilities with state-contingent repayments resembling equity or derivatives, and must instead rely on simple debt contracts if they are
to borrow. Thus, in a similar way to how menu costs can make prices sticky, transaction costs can
make financial markets incomplete.
This paper studies the implications for optimal monetary policy of such financial-market incompleteness in the form of non-contingent nominal debt contracts.2 The argument can be understood
in terms of which monetary policy strategy is able to undo or mitigate the adverse consequences of
financial-market incompleteness, just as inflation targeting can be understood as a means of circumventing the problem of nominal price stickiness. For both non-contingent nominal financial contracts
and nominal price stickiness, it is moneys role as a unit of account that is crucial, and in both cases,
optimal monetary policy is essentially the choice of a particular nominal anchor that makes money
best perform its unit-of-account function. But in spite of this formal similarity, the optimal nominal
1

In addition to the theoretical case, the more practical merits of implementing inflation targeting are discussed in
Bernanke, Laubach, Mishkin and Posen (1999).
2
It is increasingly argued that monetary policy must take account of financial-market frictions such as collateral
constraints or spreads between internal and external finance. These are different from the financial frictions emphasized in this paper. Starting from Bernanke, Gertler and Gilchrist (1999), there is now a substantial body of work
that integrates credit frictions of the kind found in Bernanke and Gertler (1989) or Kiyotaki and Moore (1997) into
monetary DSGE models. Recent work in this area includes Christiano, Motto and Rostagno (2010). These frictions
can magnify the effects of both shocks and monetary policy actions and make these effects more persistent. But
the existence of a quantitatively important credit channel does not in and of itself imply that optimal monetary
policy is necessarily so different from inflation targeting unless new types of financial shocks are introduced (Faia and
Monacelli, 2007, Carlstrom, Fuerst and Paustian, 2010, De Fiore and Tristani, 2012).

anchor turns out to be very different when the friction is financial-market incompleteness rather
than sticky prices.
One problem of non-contingent debt contracts for risk-averse households is that when borrowing
for long periods, there will be considerable uncertainty about the future income from which fixed
debt repayments must be made. The issue is not only idiosyncratic uncertainty households do not
know the future course the economy will take, which will affect their labour income. Will there be a
productivity slowdown, a deep and long-lasting recession, or even a lost decade of poor economic
performance to come? Or will unforeseen technological developments or terms-of-trade movements
boost future incomes, and good economic management successfully steer the economy on a path
of steady growth? Borrowers do not know what aggregate shocks are to come, but must fix their
contractual repayments prior to this information being revealed.
The simplicity of non-contingent debt contracts can be seen as coming at the price of bundling
together two fundamentally different transfers: a transfer of consumption from the future to the
present for borrowers, but also a transfer of aggregate risk to borrowers. The future consumption
of borrowers is paid for from the difference between their uncertain future incomes and their fixed
debt repayments. The more debt they have, the more their future income is effectively leveraged,
leading to greater consumption risk. The flip-side of borrowers leverage is that savers are able to
hold a risk-free asset, reducing their consumption risk.
To see the sense in which this bundling together of borrowing and a transfer of risk is inefficient,
consider what would happen in complete financial markets. Individuals would buy or sell statecontingent bonds (Arrow-Debreu securities) that make payoffs conditional on particular states of
the world (or equivalently, write loan contracts with different repayments across all states of the
world). Risk-averse borrowers would want to sell relatively few bonds paying off in future states of
the world where GDP and thus incomes are low, and sell relatively more in good states of the world.
As a result, contingent bonds paying off in bad states would be relatively expensive and those paying
off in good states relatively cheap. These price differences would entice savers to shift away from
non-contingent bonds and take on more risk in their portfolios. Given that the economy has no riskfree technology for transferring goods over time, and as aggregate risk cannot be diversified away,
the efficient outcome is for risk-averse households to share aggregate risk, and complete markets
allow this to be unbundled from decisions about how much to borrow or save.
The efficient financial contract between risk-averse borrowers and savers in an economy subject
to aggregate income risk (abstracting from idiosyncratic risk) turns out to have a close resemblance
to an equity share in GDP. In other words, borrowers repayments should fall during recessions
and rise during booms. This means the ratio of debt liabilities to GDP should be more stable than
it would be in a world of incomplete financial markets where debt liabilities are fixed in value while
GDP fluctuates.
With incomplete financial markets, monetary policy has a role to play in mitigating inefficiencies
because private debt contracts are typically denominated in terms of money. Hence, the real degree
of state-contingency in financial contracts is endogenous to monetary policy. If incomplete markets
were the only source of inefficiency in the economy then the optimal monetary policy would aim
2

to make nominally non-contingent debt contracts mimic through variation in the real value of the
monetary unit of account the efficient financial contract that would be chosen with complete financial
markets.
Given that the efficient financial contract between borrowers and savers resembles an equity
share in GDP, it follows that a goal of monetary policy should be to stabilize the ratio of debt
liabilities to GDP. With non-contingent nominal debt liabilities, this can be achieved by having a
non-contingent level of nominal income, in other words, a monetary policy that targets nominal
GDP. Nominal income thus replaces nominal goods prices as the optimal nominal anchor. The
intuition is that while the central bank cannot eliminate uncertainty about future real GDP, it can
in principle make the level of future nominal GDP (and hence the nominal income of an average
household) perfectly predictable. Removing uncertainty about future nominal income thus alleviates
the problem of nominal debt repayments being non-contingent.
A policy of nominal GDP targeting is generally in conflict with inflation targeting because any
fluctuations in real GDP would lead to fluctuations in inflation of the same size and in the opposite
direction. Recessions would feature higher inflation and booms would feature lower inflation, or
even deflation. These inflation fluctuations can be helpful because they induce variation in the real
value of the monetary unit of account, making it and the non-contingent debt contracts expressed
in terms of it behave more like equity. This promotes efficient risk sharing. A policy of strict
inflation targeting would fix the real value of the monetary unit of account, converting nominally
non-contingent debt into real non-contingent debt, which would imply an uneven and generally
inefficient distribution of risk.
The inflation fluctuations that occur with nominal GDP targeting would entail relative-price
distortions if goods prices were sticky, so the benefit of efficient risk sharing is most likely not
achieved without some cost.3 It is ultimately a quantitative question whether the inefficiency caused
by incomplete financial markets is more important than the inefficiency caused by relative-price
distortions, and thus whether nominal GDP targeting is preferable to inflation targeting.
This paper presents a model that allows optimal monetary policy to be studied analytically in
an incomplete-markets economy with heterogeneous households and aggregate risk, which can be
straightforwardly calibrated for quantitative analysis. The model contains two types of households,
relatively impatient households who will choose to become borrowers, and relatively patient households who will choose to become savers. Although households differ in their time preferences, they
are all risk averse, and are all exposed to the same labour income risk. Real GDP is uncertain
because of aggregate productivity shocks, but there are no idiosyncratic shocks. The economy is
3

The model features both a trade-off between efficiency in goods markets and efficiency in financial markets, and a
trade-off between the volatility of inflation and the volatility of financial-market variables. Such trade-offs are implicit
in recent debates, though there is no widely accepted argument for why stabilizing prices in goods markets causes
financial markets to malfunction. White (2009b) and Christiano, Ilut, Motto and Rostagno (2010) argue that stable
inflation is no guarantee of financial stability, and may even create conditions for financial instability. Contrary
to these arguments, the conventional view that monetary policy should not react to asset prices is advocated in
Bernanke and Gertler (2001). Woodford (2011) makes the point that flexible inflation targeting can be adapted to
accommodate financial stability concerns, and that it would be unwise to discard inflation targetings role in providing
a clear nominal anchor.

assumed to have no investment or storage technology, and is closed to international trade. There
are no government bonds and no fiat money, and no taxes or fiscal transfers. In this world, patient
households defer consumption by lending to impatient households, who can thus bring forwards
consumption by borrowing. It is assumed the only financial contract available is a non-contingent
nominal bond. The basic model contains no other frictions, initially assuming prices and wages are
fully flexible.
The concept of a natural debt-to-GDP ratio provides a useful benchmark for monetary policy
analysis. This is defined as the ratio of (state-contingent) gross debt liabilities to GDP that would
prevail were financial markets complete. This object is independent of monetary policy. The actual
debt-to-GDP ratio in an economy with incomplete markets would coincide with the natural debtto-GDP ratio if forecasts of future GDP were always correct ex post, but will in general fluctuate
around it when the economy is hit by shocks. The natural debt-to-GDP ratio is thus analogous to
concepts such as the natural rate of unemployment and the natural rate of interest.
If all movements in real GDP growth rates are unpredictable then the natural debt-to-GDP ratio
turns out to be constant (or if utility functions are logarithmic, the ratio is constant irrespective
of the statistical properties of GDP growth). Even when the natural debt-to-GDP ratio is not
completely constant, plausible calibrations suggest it would have a low volatility relative to real
GDP itself.
Since the equilibrium of an economy with complete financial markets would be Pareto efficient in
the absence of other frictions, the natural debt-to-GDP ratio also has desirable welfare properties.
A goal of monetary policy in an incomplete-markets economy is therefore to close the debt gap,
defined as the difference between the actual and natural debt-to-GDP ratios. It is shown that doing
so effectively completes the market in the sense that the equilibrium with incomplete markets would
then coincide with the hypothetical complete-markets equilibrium. Monetary policy can affect the
actual debt-to-GDP ratio and thus the debt gap because that ratio is equal to nominal debt liabilities
(which are non-contingent with incomplete markets) divided by nominal GDP, where the latter is
under the control of monetary policy.
When the natural debt-to-GDP ratio is constant, or when the maturity of debt contracts is
sufficiently long, closing the debt gap can be achieved by adopting a fixed target for the growth rate
(or level) of nominal GDP. With this logic, the central bank uses nominal GDP as an intermediate
target that achieves its ultimate goal of closing the debt gap. This turns out to be preferable to
targeting the debt-to-GDP ratio directly because a monetary policy that targets only a real financial
variable would leave the economy without a nominal anchor. Nominal GDP targeting uniquely pins
down the nominal value of real incomes and thus provides the economy with a well-defined nominal
anchor.
It is important to note that in an incomplete-markets economy hit by shocks, whatever action
a central bank takes or fails to take will have distributional consequences. Ex post, there will be
winners and losers from both the shocks themselves and the policy responses. Creditors lose out
when inflation is unexpectedly high, while debtors suffer when inflation is unexpectedly low. It might
then be thought surprising that inflation fluctuations would ever be desirable. However, the inflation
4

fluctuations implied by a nominal GDP target are not arbitrary fluctuations they are perfectly
correlated with the real GDP fluctuations that are the ultimate source of uncertainty in the economy,
and which themselves have distributional consequences when households are heterogeneous. For
households to share risk, it must be possible to make transfers ex post that act as insurance from
an ex-ante perspective. The result of the paper is that ex-ante efficient insurance requires inflation
fluctuations that are negatively correlated with real GDP (a countercyclical price level) to generate
the appropriate ex-post transfers between debtors and creditors.
It might be objected that there are infinitely many state-contingent consumption allocations
which would equally well satisfy the criterion of ex-ante efficiency. However, only one of these
the hypothetical complete-markets equilibrium associated with the natural debt-to-GDP ratio
could ever be implemented through monetary policy. Thus for a policymaker solely interested in
promoting efficiency, there is a unique optimal policy that does not require any explicit distributional
preferences to be introduced.
The model also makes predictions about how different monetary policies will affect the volatility
of financial-market variables such as the debt-to-GDP ratio. It is shown that policies implying an
inefficient distribution of risk, for example, inflation targeting, are associated with near-random
walk fluctuations in the debt-to-GDP ratio. On the other hand, with complete financial markets,
the persistence of fluctuations in the debt-to-GDP ratio would be bounded by the persistence of
shocks to real GDP growth. When a monetary policy is adopted that allows the economy to mimic
complete financial markets, the actual debt-to-GDP ratio inherits these less persistent dynamics.
In a model with both nominal price rigidities and incomplete financial markets, these findings
allow the tension between relative-price distortions and efficient risk sharing to be seen in more
familiar terms as a trade-off between price stability and financial stability. Determining which of
these objectives is the more important in practice can be done by studying a quantitative version
of the model. Nominal price rigidity is introduced using the standard Calvo model of staggered
price adjustment. With both incomplete financial markets and sticky prices, optimal monetary
policy is a convex combination of the optimal monetary policies that are appropriate for each of the
two frictions in isolation. After calibrating all the parameters of the model, the conclusion is that
replicating complete financial markets should receive around 90% of the weight.
This paper is related to a number of areas of the literature on monetary policy and financial
markets. First, there is the empirical work of Bach and Stephenson (1974), Cukierman, Lennan
and Papadia (1985), and more recently, Doepke and Schneider (2006), who document the effects of
inflation in redistributing wealth between debtors and creditors. The novelty here is in studying the
implications for optimal monetary policy in an environment where inflation fluctuations with such
distributional effects may actually be desirable precisely because of the incompleteness of financial
markets.
The basic idea of this paper (though not the modelling or quantitative analysis) has many
precedents in the literature. Selgin (1997) describes the ex-ante efficiency advantages of falling
prices in good times and rising prices in bad times when financial contracts are non-contingent,

though there is no formal modelling of the argument.4 A survey reviewing the long history of
this idea in monetary economics is given in Selgin (1995). In recent work, Koenig (2013) advances
the risk-sharing argument for nominal GDP targeting in the context of a two-period model.5 An
earlier theoretical paper is Pescatori (2007), who studies optimal monetary policy in an economy
with rich and poor households, in the sense of there being an exogenously specified distribution
of assets among otherwise identical households. In that environment, both inflation and interest
rate fluctuations have redistributional effects on rich and poor households, and the central bank
optimally chooses the mix between them (there is a need to change interest rates because prices
are sticky, with deviations from the natural rate of interest leading to undesirable fluctuations in
output). A related paper is Lee (2010), who develops a model where heterogeneous households
choose less than complete consumption insurance because of the presence of convex transaction
costs in accessing financial markets. Inflation fluctuations expose households to idiosyncratic labourincome risk because households work in specific sectors of the economy, and sectoral relative prices
are distorted by inflation when prices are sticky. This leads optimal monetary policy to put more
weight on stabilizing inflation. Differently from those two papers, the argument here is that inflation
fluctuations can actually play a positive role in completing otherwise incomplete financial markets.6
The idea that inflation fluctuations may have a positive role to play when financial markets are
incomplete is now long-established in the literature on government debt (and has also been recently
applied by Allen, Carletti and Gale (2011) in the context of the real value of the liquidity available
to the banking system). Bohn (1988) developed the theory that non-contingent nominal government
debt can be desirable because when combined with a suitable monetary policy, inflation can change
the real value of the debt in response to fiscal shocks that would otherwise require fluctuations in
distortionary tax rates.7
Quantitative analysis of optimal monetary policy of this kind was developed in Chari, Christiano
and Kehoe (1991) and expanded further in Chari and Kehoe (1999). One finding was that inflation
needs to be extremely volatile to complete financial markets. As a result, Schmitt-Grohe and
Uribe (2004) and Siu (2004) argued that once some nominal price rigidity is considered so that
4

Persson and Svensson (1989) is an early example of a model in the context of an international portfolio
allocation problem where it is important how monetary policy affects the risk characteristics of nominal debt.
5
Hoelle and Peiris (2013) study the efficiency properties of nominal GDP targeting in a large open economy
with flexible prices, and explore the question of implementability through the central banks balance sheet. OsorioRodrguez (2013) examines whether there remains a role for monetary policy in completing financial markets when
nominal debt is denominated in terms of foreign currency.
6
In other related work on incomplete markets and monetary policy, Akyol (2004) analyses optimal monetary policy
in an incomplete-markets economy where households hold fiat money for self insurance against idiosyncratic shocks.
Kryvtsov, Shukayev and Ueberfeldt (2011) study an overlapping generations model with fiat money where monetary
policy can improve upon the suboptimal level of saving by varying the expected inflation rate and thus the returns
to holding money.
7
There is also a literature that emphasizes the impact of monetary policy on the financial positions of firms or
entrepreneurs in an economy with incomplete financial markets. De Fiore, Teles and Tristani (2011) study a flexibleprice economy where there is a costly state verification problem for entrepreneurs who issue short-term nominal bonds.
Andres, Arce and Thomas (2010) consider entrepreneurs facing a binding collateral constraint who issue short-term
nominal bonds with an endogenously determined interest rate spread. Vlieghe (2010) also has entrepreneurs facing
a collateral constraint, and even though they issue real bonds, monetary policy still has real effects on the wealth
distribution because prices are sticky, so incomes are endogenous. In these papers, the wealth distribution matters
because of its effects on the ability of entrepreneurs to finance their operations.

inflation fluctuations have a cost, the optimal policy becomes very close to strict inflation targeting.
This paper shares the focus of that literature on using inflation fluctuations to complete markets,
but comes to a different conclusion regarding the magnitude of the required inflation fluctuations
and whether the costs of those fluctuations outweigh the benefits. First, the benefits of completing
markets in this paper are linked to the degree of risk aversion and the degree of heterogeneity among
households, which are in general unrelated to the benefits of avoiding fluctuations in distortionary
tax rates, and which prove to be large in the calibrated model. Second, the earlier results assumed
government debt with a very short maturity. With longer maturity debt (household debt in this
paper), the costs of the inflation fluctuations needed to complete the market are much reduced.8
This paper is also related to the literature on household debt. Iacoviello (2005) examines the
consequences of household borrowing constraints in a DSGE model, while Guerrieri and Lorenzoni
(2011) and Eggertsson and Krugman (2012) study how a tightening of borrowing constraints for
indebted households can push the economy into a liquidity trap. Differently from those papers, the
focus here is on the implications of household debt for optimal monetary policy. Furthermore, the
finding here that the presence of household debt substantially changes optimal monetary policy does
not depend on there being borrowing constraints, or even the feedback effects from debt to aggregate
output stressed in those papers. C
urdia and Woodford (2009) also study optimal monetary policy in
an economy with household borrowing and saving, but the focus there is on spreads between interest
rates for borrowers and savers, while their model assumes an insurance facility that rules out the
risk-sharing considerations studied here. Finally, the paper is related to the literature on nominal
GDP targeting (Meade, 1978, Bean, 1983, Hall and Mankiw, 1994, and more recently, Sumner, 2012)
but proposes a different argument in favour of that policy.
The plan of the paper is as follows. Section 2 sets out a basic model and examines what monetary
policies can support risk sharing when financial markets are incomplete. Section 3 introduces a
DSGE model that includes both incomplete financial markets and sticky prices, and hence a trade-off
between mitigating the incompleteness of financial markets and avoiding relative-price distortions.
Optimal monetary policy subject to this trade-off is studied in section 4. Section 5 shows how
the full model can be calibrated and presents a quantitative analysis of optimal monetary policy.
Finally, section 6 draws some conclusions.

A model of a pure credit economy

The analysis begins with a simplified model that studies household borrowing and saving in a finitehorizon endowment economy with incomplete financial markets. A full dynamic stochastic general
equilibrium model with incomplete markets together with labour supply, production, and sticky
prices is presented in section 3.
8

This point is made by Lustig, Sleet and Yeltekin (2008) in the context of government debt.

2.1

Assumptions

The economy contains two groups of households, borrowers (b) and savers (s), each making up
50% of a measure-one population. Household types are indexed by {b, s}. There are three time
periods t {0, 1, 2}. All households have preferences represented by the utility function:
#
"
1
1
1
C,1
C
C,0
,2
+
+ 2
,
[2.1]
U = E
1
1
1
where C,t is per-person consumption by households of type at time t. All households have the
same subjective discount factor and the same coefficient of relative risk aversion .
Real GDP Yt is an exogenous endowment. The level of GDP in period 0 is non-stochastic,
but subsequent real GDP growth rates gt = (Yt Yt1 )/Yt1 are uncertain. Household types are
distinguished by the shares they receive of this endowment at different dates The ratio of borrowers
per-person incomes to per-person real GDP is denoted by the parameter t , and hence household
incomes at time t are:
Yb,t = t Yt ,

and Ys,t = (2 t )Yt .

[2.2]

The income shares t are known with certainty in period 0. Given that both household types have
the same time preferences, the households labelled as borrowers will indeed to choose to borrow
from the savers in equilibrium when the sequence {0 , 1 , 2 } is increasing. In other words,
borrowers are those households with initially low incomes that anticipate higher incomes in the
future, while savers have initially high incomes that they anticipate will fall in the future. In what
follows, the analysis is simplified by assuming the particular monotonic sequence of income shares
below:


0 = 1 2 E (1 + g1 )1 (1 + g2 )1 , 1 = 1, and 2 = 2,
[2.3]
where it is assumed that the subjective discount factor is sufficiently low relative to expected real
GDP growth so that 0 0.
Financial markets are incomplete in the sense that only nominal bonds can be issued or held
by households. It is assumed that borrowers cannot issue liabilities with state-contingent nominal
payoffs, and all non-contingent bonds are denominated in terms of money. There is assumed to be
a single type of bond traded at each date. Each bond issued in period 0 is a promise to repay 1 unit
of money in period 1, and units of money in period 2. The parameter determines the duration,
or average maturity, of the bonds ( = 0 is one-period debt, while larger values of represent
long-term debt contracts). Each bond issued in period 1 is simply a promise to repay one unit of
money in period 2. Note that in period 1, outstanding bonds from period 0 are equivalent to
newly issued bonds (old bonds are therefore counted in terms of new-bond equivalents from period
1 onwards). Households can take positive or negative positions in bonds (save or borrow) with no
limit on borrowing except being able to repay in all states of the world. There is no default, and so
all bonds are risk free in nominal terms.
Households begin with no initial assets or debts, and must leave no debts at the end of period
8

2. The net bond position per person of type- households at the end of period t is denoted by B,t ,
the nominal bond price is Qt at date t, and the price of goods in terms of money is Pt . The flow
budget identities are:
C,0 +

Q0 B,0
= Y,0 ,
P0

C,1 +

Q1 B,1
(1 + Q1 )B,0
= Y,1 +
,
P1
P1

and C,2 = Y,2 +

B,1
, [2.4]
P2

the term (1 + Q1 ) representing the sum of the period-1 coupon payment from period-0 bonds and
the market value of the period-2 repayment, the latter equivalent in value to newly issued bonds.
Money in this economy is simply a unit of account used in writing financial contracts. Monetary
policy is assumed to determine the inflation rate t = (Pt Pt1 )/Pt1 at each date.

2.2

Equilibrium

Maximizing the utility function [2.1] subject to the budget identities [2.4] implies Euler equations
that must hold for both household types {b, s} at dates t {0, 1}:



C,t
= Et (1 + rt+1 )C,t+1
.
[2.5]
In equilibrium, the goods and bond markets must clear at all dates:
Bb,t Bs,t
+
= 0;
[2.6]
2
2
Cb,t Cs,t
+
= Yt .
[2.7]
2
2
In what follows, let c,t C,t /Yt denote the ratios of consumption to income, and define the
variables dt , lt , and rt as follows:
dt

1 (1 + Qt )Bb,t1
,
2
Pt

lt

1 Qt Bb,t
,
2 Pt

1 + rt

(1 + Qt )Pt1
.
Qt1 Pt

[2.8]

As will be confirmed in equilibrium, Bb,t 0 and Bs,t 0, so dt can be interpreted as the gross
debt-to-GDP ratio (the beginning-of-period value of debt liabilities per person relative to GDP),
and lt as the end-of-period value of all bonds issued per person relative to GDP, referred to as the
loans-to-GDP ratio. The variable rt is the ex-post real return on holding bonds between periods
t 1 and t. Note that this is not the same as the interest rate on those bonds, which refers to the
ex-ante expected real return, t = Et rt+1 . Finally, it is convenient to express the equations in terms
of the yield-to-maturity, denoted by jt , rather than the bond price Qt . Given the coupon payments
on the bonds issued in period 0 and 1, the price-yield relationships are:
Q0 =

+
,
1 + j0 (1 + j0 )2

and Q1 =

1
.
1 + j1

[2.9]

With the definitions [2.8], the equations of the model are:



1 + rt
dt =
lt1 ;
1 + gt
cb,t = t 2(dt lt ), cs,t = (2 t ) + 2(dt lt ),



c,t+1 ,
c
,t = Et (1 + rt+1 )(1 + gt+1 )
9

[2.10a]
with d0 = 0,

l2 = 0;

[2.10b]
[2.10c]

where [2.10a] follows directly from [2.8], [2.10b] is derived from the budget identities [2.4] and market
clearing condition [2.6], and [2.10c] from the Euler equations [2.5]. Finally, [2.8] and the definition
of the yield-to-maturity in [2.9] imply:
!



1 + 1+j
1 + j1
1 + j0
1
, and 1 + r2 =
.
[2.11]
1 + r1 =

1 + 1
1 + 1+j0
1 + 2
It is assumed that the parameter restriction in [2.3] always holds in what follows.
In the case where there is no uncertainty about the path of real GDP (g1 = g1 and g2 = g2 , where
g1 and g2 are non-stochastic), and where there are no unexpected changes in inflation (1 =
1 and
2 =
2 ), the system of equations [2.10a][2.10c] and [2.11] has the following solution:
cb,t = cs,t = 1,

(1 + g1 )
,

1
and d2 = ;
2

1 + r1 =

and 1 + r2 =

(1 + g2 )
;

[2.12a]

[2.12b]
d1 = (1 + g2 )1 ,
2
2
l0 = (1 + g1 )1 (1 + g2 )1 , and l1 = (1 + g2 )1 .
[2.12c]
2
2
The equilibrium interest rates (equal here to the ex-post real returns r1 and r2 ) are identical to
what would prevail if there were a representative household. The income shares in [2.3] imply that
in the absence of shocks, borrowers and savers would have the same levels of consumption. Given
the levels of income and consumption, the implied final debt-to-GDP ratio is 50%, and given [2.3],
the debt-to-GDP and loans-to-GDP ratios at earlier dates are discounted values of the final debtto-GDP ratio (adjusted for any real GDP growth). This steady state is independent of monetary
policy. The values of
1 and
2 together with the real interest rates r1 and r2 determine the nominal

bond yields j0 and j1 .

2.3

The complete financial markets benchmark

Consider an hypothetical economy that has complete financial markets but is otherwise identical
to the economy described above. Households now have access to a complete set of state-contingent
bonds (traded sequentially, period-by-period), denominated in real terms without loss of generality.

Let F,t
denote the net portfolio of contingent bonds per person held between periods t and t + 1 by
households of type (the asterisk signifies complete financial markets). The prices of these securities
in real terms relative to the conditional probabilities of the states at time t are denoted by Kt+1 , so

Et [Kt+1 F,t+1
] is the date-t cost of the date-t + 1 payoff F,t+1
.
In this version of the model, the flow budget identities [2.4] are replaced by:

,
C,t
+ Et [Kt+1 F,t+1
] = Y,t + F,t

[2.13]

together with initial and terminal conditions F,0

= 0 and F,3
= 0. The Euler equations for
maximizing utility [2.1] subject to [2.13] are:
!
 

Cb,t+1
Cs,t+1

= Kt+1 =
,
[2.14]

Cb,t
Cs,t

10

/2 = 0 replaces
/2+Fs,t
which must hold in all states of the world. The market-clearing condition Fb,t
equation [2.6].
To relate the economy with complete markets to its incomplete-markets equivalent, consider the
following definitions of variables dt , lt , and rt that will be seen to be the equivalents of the debt-toGDP ratio dt , the loans-to-GDP ratio lt , and the ex-post real return rt in the incomplete-markets
economy (as given in [2.8]):

dt

1 Fb,t
,
2 Yt

lt

]
1 Et [Kt+1 Fb,t+1
,
2
Yt

and 1 + rt

Fb,t
.

Et1 [Kt Fb,t

]

[2.15]

Debt in an economy with complete financial markets refers to the total gross value of the contingent
bonds repayable in the realized state of the world. Loans refers to the value of the whole portfolio
of contingent bonds issued by borrowers, and the (gross) ex-post real return is the state-contingent
value of the bonds repayable relative to the value of all the bonds previously issued.
The definitions [2.15] directly imply that equation [2.10a] must hold in terms of c,t , dt , lt ,
and rt . The budget identities [2.13] and the contingent bond-market clearing conditions imply

)Kt+1 ] = 1 follows
that [2.10b] holds in terms of the variables defined in [2.15]. Since Et [(1 + rt+1
from the definition in [2.15], the first-order conditions [2.10c] imply that [2.10c] must hold in terms
of c,t and rt . Hence, the block of equations [2.10a][2.10c] applies to both the incomplete- and
complete-markets economies.
The first-order condition [2.14] with complete markets has stronger implications than equation
[2.10c], though. It also requires
cb,t+1
cs,t+1
=
,
cb,t
cs,t

[2.16]

to hold in all states of the world. This equation says that consumption growth rates must always be
equalized between borrowers and lenders, in other words, households use complete financial markets
to share risk. This is not generally an implication of the equilibrium conditions [2.10a][2.10c] and
[2.11] with incomplete financial markets. Furthermore, since a complete-markets economy has no
restriction on the types of assets that households can buy and sell, equation [2.11] that determines the
ex-post real return on a portfolio of nominal bonds is now irrelevant to determining the equilibrium
of the complete-markets economy. The ex-post real return is now determined implicitly by the
portfolio of contingent securities that ensures the risk-sharing condition [2.16] holds.
The complete-markets equilibrium can be obtained analytically by solving the system of equations [2.10a][2.10c] and [2.16] (again under the parameter restriction in [2.3]). The equilibrium
consumption-GDP ratios are c,t = 1, so there is full risk sharing between borrowers and savers,
meaning that all households consumption levels perfectly co-move in response to shocks (the consumption levels are equal owing to the parameter restriction [2.3]). Complete financial markets
therefore allocate consumption efficiently across states of the world, as well as over time. The equilibrium values of other variables have similar expressions to those found in the non-stochastic case
[2.12] except that certain outcomes are replaced by expected values conditional on earlier informa-

11

tion:

1

d1 = E1 (1 + g2 )1 , and d2 = ;
[2.17a]
2
2



2 
[2.17b]
l0 = E (1 + g1 )1 (1 + g2 )1 , and l1 = E1 (1 + g2 )1 .
2
2
The final debt-to-GDP ratio d2 is non-stochastic, and earlier debt-to-GDP ratios depend only on
conditional expectations of future real GDP growth rates. Since the realization of shocks in period 1
can change these conditional expectations, d1 and l1 are stochastic in general. The expressions for l0 ,
l1 , and d1 can be interpreted as the present discounted values in periods 0 or 1 of a payoff proportional
to the final debt-to-GDP ratio d2 , evaluated using prices of contingent securities (which are equal
to households common stochastic discount factor according to [2.14]). Intuitively, the completemarkets portfolio is an equity share in future real GDP. This supports risk sharing by allowing the
repayments of borrowers to move exactly in line with their incomes.
Finally, observe that the complete-markets equilibrium has a particularly simple form in two
special cases. If the utility function is logarithmic ( = 1) or real GDP follows a random walk (gt
is i.i.d.) then dt and lt are all non-stochastic:



1
2

, d2 = , l0 =
, and l1 = , where = E (1 + gt )1 .
[2.18]
2
2
2
2
Note that the complete-markets equilibrium is entirely independent of monetary policy in all cases.
d1 =

2.4

Replicating complete financial markets

The block of equations [2.10a][2.10c] is common to the equilibrium conditions irrespective of

whether financial markets are complete or not. The only difference between the equilibrium conditions is that the incomplete-markets economy includes [2.11] instead of [2.16] in the completemarkets economy. Since equation [2.11] includes the inflation rates 1 and 2 , the ability of monetary
policy to engineer a suitable state-contingent path for inflation means that the ex-post real returns
in [2.11] can be chosen to generate the same consumption allocation as implied by the risk-sharing
condition [2.16].9 This is in turn equivalent to ensuring the actual debt-to-GDP ratio dt mimics its
hypothetical equilibrium value dt in the economy with complete financial markets.
The monetary policy that replicates complete financial markets in this way turns out to be
a nominal GDP target. Nominal GDP is denoted by Nt = Pt Yt , and its growth rate by nt =
(Nt Nt1 )/Nt1 . Since it is assumed that monetary policy can determine a state-contingent path
for inflation t , and as real GDP growth gt is exogenous, monetary policy can equally well be
specified as a sequence of nominal GDP growth rates nt . Equation [2.11] for the ex-post real returns
on nominal bonds can be written in terms of nominal GDP growth as follows:
!



1 + 1+j
1 + j0
1 + r2
1 + j1
1 + r1
1
=
, and
=
.
[2.19]

1 + g1
1 + n1
1 + 1+j0
1 + g2
1 + n2
9
That monetary policy is able exactly to replicate complete financial markets is owing to the model having a
representative borrower and a representative saver. With heterogeneity within these groups as well as between
them, exact replication will generally not be possible.

12

Consider first the replication argument in either of the special cases of log utility ( = 1) or
real GDP following a random walk (gt is an i.i.d. stochastic process). In these special cases, the
complete-markets debt-to-GDP ratios [2.18] that monetary policy is aiming to replicate are nonstochastic. Suppose monetary policy sets a non-stochastic path for nominal GDP, that is, n1 = n
1
and n2 = n
2 for some constants n
1 and n
2 . If the replication is successful, c,t = 1, and so both
households Euler equations [2.10c] are satisfied when 1 = Et [(1 + rt+1 )(1 + gt+1 ) ]. With n2 = n
2
and [2.19] this requires for t = 1:
1 + j1
1
1
=
= ,
1
1+n
2
E1 [(1 + g2 ) ]

using the definition of from [2.18]. It then follows from [2.19] that (1 + r2 )/(1 + g2 ) = 1/.
From the complete-markets solution [2.18] together with [2.10a], (1 + r2 )/(1 + g2 ) = 1/, so this
monetary policy ensures the ex-post real return r2 on nominal bonds is identical to that on the
1 and the solution
complete-markets portfolio r2 for all realizations of shocks. Similarly, with n1 = n
/(1 + j1 ) = /(1 + n
2 ) from above, the Euler equation at t = 0 requires:
1 + j0
1
1
1
=
,

1+n
1 1 + 1+j0
1 + 1+n2
and hence (1 + r1 )/(1 + g1 ) = 1/, which coincides with (1 + r1 )/(1 + g1 ) = 1/. Therefore,
r1 = r1 , and r2 = r2 , under this monetary policy, which establishes that dt = dt and c,t = c,t . A
nominal GDP target with any non-stochastic rates of nominal GDP growth succeeds in replicating
the complete-markets equilibrium and supporting risk-sharing among borrowers and savers. The
intuition is that if the numerator of the debt-to-GDP ratio (expressed in monetary units) is fixed
because nominal debt liabilities are not state contingent, the ratio can be stabilized by targeting the
denominator (expressed in monetary units), that is, ensuring that nominal incomes are predictable.
There is one other special case in which a monetary policy that makes nominal GDP growth
perfectly predictable manages to replicate complete financial markets, even when the completemarkets debt-to-GDP ratio d1 is stochastic. This is the case where borrowers do not need to issue
any new debt and do not need to refinance any existing debt after the initial time period. In the
model, this corresponds to the limiting case of pure long-term bonds where ( is the ratio
of the period-2 and period-1 coupon payments on a bond issued in period 0). If monetary policy
ensures that (1+n1 )(1+n2 ) = (1+
n)2 for some non-stochastic n
then the equilibrium of the economy
will coincide with the hypothetical complete-markets equilibrium. Unlike the earlier special cases,
here it is not necessary that both the period 1 and 2 nominal GDP growth rates are non-stochastic,
only that the cumulated growth rate over both time periods is perfectly predictable. Intuitively,
with long-term debt, monetary policy needs only to ensure that nominal incomes are predictable
when debt is actually repaid.
When is finite, some existing debt must be repaid by borrowers in period 1, requiring them
to issue some new bonds if they are to continue to borrow until period 2. This exposes them to
risk coming from uncertainty about the interest rate that will prevail in period 1. A monetary
policy that aims to replicate complete financial markets must then address refinancing risk as well

13

as income risk. In general, this requires a target for nominal GDP growth that changes when shocks
occur, though as will be seen, there will still be a long-run target for nominal GDP that is invariant
to shocks. If the maturity parameter is positive and finite then the class of monetary policies
that replicate complete financial markets are given by the following nominal GDP growth rates in
periods 1 and 2:
1 + n1 = (1 + n
1)

E [(1 + g2 )1 ]
,
E1 [(1 + g2 )1 ]

and 1 + n2 = (1 + n
2)

E1 [(1 + g2 )1 ]
,
E [(1 + g2 )1 ]

[2.20]

where n
1 and n
2 are any non-stochastic growth rates (these would be the actual nominal GDP
growth rates in the absence of shocks). Note that any such monetary policy has the implication
that
(1 + n1 )(1 + n2 ) = (1 + n
1 )(1 + n
2 ),

[2.21]

so the long-run target for nominal GDP must be non-stochastic. In the special cases of = 1 or
gt being i.i.d. that were analysed earlier, the requirements on nominal GDP growth rates in [2.20]
reduce simply to n1 = n
1 and n2 = n
2 . Intuitively, the problem of refinancing risk is absent in these
special cases, albeit for a different reason in each. When real GDP growth rates are independent
over time, there is no news that changes expected future real GDP growth, and thus no reason for
the equilibrium real interest rate to vary over time. On the other hand, with log utility, the real
interest rate changes by the same amount and in the same direction as any revision to expectations
of future growth. Higher real interest rates are then exactly offset by expectations of an improvement
in future incomes relative to current incomes. This leaves monetary policy needing only to provide
insurance against fluctuations in current incomes, for which a predictable nominal GDP growth rate
suffices.
As discussed above, in the special case of pure long-term debt ( ), [2.21] is needed, but
[2.20] need not hold. In the special case of pure short-term debt ( = 0), it can be shown that only
the condition on n1 in [2.20] is needed, together with the restriction that n2 takes on a value that
is perfectly predictable in period 1.

2.5

Consequences of optimal monetary policy for inflation

Optimal monetary policies that replicate complete financial markets have been characterized as
nominal GDP targets. By definition, stabilizing nominal GDP when there are fluctuations in real
GDP entails fluctuations in inflation. That inflation fluctuates is not in itself the desirable feature
of these policies, but rather that inflation displays a negative correlation with real GDP growth.
In other words, there is an optimal degree of countercyclicality of the price level. In cases where a
constant nominal GDP growth rate is optimal, inflation should move in the opposite direction to
real GDP growth by the same amount.
In general, when the required nominal GDP growth rate is not independent of the realization of

14

shocks (see [2.20]), the implied inflation rates are:

1 + 1 =

(1 + n
1 ) E [(1 + g2 )1 ]
,
(1 + g1 ) E1 [(1 + g2 )1 ]

and 1 + 2 =

(1 + n
1 ) E1 [(1 + g2 )1 ]
.
(1 + g2 ) E [(1 + g2 )1 ]

[2.22]

Now suppose that real GDP does not follow a random walk (gt is not i.i.d.), and utility is not
logarithmic ( 6= 1). Take the realistic case where > 1 (high risk aversion, low intertemporal
substitution), and where there is some mean reversion in the level of real GDP after a shock has
occurred (so the growth rates g1 and g2 are negatively correlated). Under these assumptions, the
term E1 [(1 + g2 )1 ] in [2.22] moves in the same direction as (1 + g1 ) following a shock to real GDP
growth in period 1. This means the required change in inflation is even larger than what would be
implied by non-stochastic target for the nominal GDP growth rate. Intuitively, any unexpected fall
in g1 leads to an increase in the real interest rate because g2 is expected to rise, and the increase in
the real interest rate is larger. The complete-markets debt-to-GDP ratio d1 from [2.17a] thus falls
because the interest rate effect dominates the growth effect. To decrease the debt-to-GDP ratio in
an economy with nominal bonds, it is necessary for inflation to rise by more than the fall in.
In period 2, equation [2.22] shows that both predictable and unpredictable changes in real GDP
growth g2 require movements in period 2 inflation in the opposite direction to real GDP growth.
Therefore, even in the case where optimal monetary policy is not a perfectly predictable nominal
GDP growth rate, it is unlikely that a constant inflation rate would ever be optimal. In other words,
the required nominal GDP growth rate would not ever be likely to move by more than real GDP
growth in the same direction. To reinforce this conclusion, note that [2.21] implies the cumulated
inflation rate over periods 1 and 2 must be
(1 + 1 )(1 + 2 ) =

(1 + n
1 )(1 + n
2)
.
(1 + g1 )(1 + g2 )

This means that the price level must be countercyclical on average over periods 1 and 2, and the
only case in which the long-run price level should be predictable is unrealistic one in which the
long-run level of real GDP is perfectly predictable.
Writing a target for monetary policy in terms of inflation and real GDP growth might be reminiscent of so-called flexible inflation targeting, and it might be tempting to interpret a nominal GDP
target as simply a relabelling of monetary policies of that kind. But aside from the rationale being
very different from that invoked to justify flexible inflation targeting, there is a fundamental difference between the two policies. Flexible inflation targeting can be formalized as a target criterion
in both inflation and the output gap, whereas here, the target criterion includes the actual growth
rate of output, not the output gap. Monetary policy is effective in completing financial markets
precisely because the level of output that appears in the target is not adjusted for any unexpected
changes in potential output, even though these have consequences for inflation.

2.6

Consequences of suboptimal monetary policy for financial markets

Following a monetary policy that replicates complete financial markets has consequences for fluctuations in inflation. Similarly, following a monetary policy that stabilizes inflation has observable
15

implications for fluctuations in financial markets in addition to failing to replicate the risk sharing
of complete financial markets. This is because the behaviour of variables such as the debt-to-GDP
ratio is very different in a complete-markets economy than in an economy with incomplete markets
and a sub-optimal monetary policy.
In the hypothetical complete-markets economy, the equilibrium debt-to-GDP ratio has some
striking properties. Intuitively, the stock of outstanding debt relative to GDP ought to behave
like a state variable because debt liabilities are predetermined. However, with access to complete
financial markets, households optimizing behaviour leads them to choose a portfolio of contingent
securities such that the implied debt-to-GDP ratio is actually a purely forward-looking variable: all
the expressions for dt and lt in [2.17a] depend only on expectations of future real GDP growth.
The intuition for this surprising result is that given the budget identities of households, full risk
sharing requires the financial wealth of savers and the financial liabilities of borrowers to move in
line with the value of future labour income. The debt-to-GDP ratio must therefore behave like an
asset price rather than a state variable: it must reflect a forecast of the economys future prospects
rather than a record of past choices and shocks. It follows that in the complete-markets economy,
past realizations of shocks would have no correlation with the current debt-to-GDP ratio except to
the extent that these have predictive power for the economys future fundamentals.
When monetary policy is used to replicate complete financial markets in an incomplete-markets
economy, fluctuations of inflation are used to affect the real value of nominal debt so as to mimic
the behaviour of the complete-markets debt-to-GDP ratio. But if monetary policy does not replicate complete financial markets then the debt-to-GDP behaves in line with what simple intuition
suggests: it is intrinsically serially correlated, so past shocks have a persistent effect on the subsequent evolution of the level of debt. With incomplete financial markets, optimizing behaviour by
households leads to consumption smoothing over time, but not generally across different states of
the world. Thus, following a shock, households financial wealth may diverge from the present discounted value of future labour income. Moreover, when this divergence occurs, because households
spread out the adjustment of consumption over time, the consequences of the shock for financial
wealth and consumption are long lasting.10
To illustrate these claims, assume for simplicity that all uncertainty about real GDP in period
2 has been resolved by period 1. Suppose that monetary policy in period 1 (together with the
monetary policy expected in period 2) has failed to replicate complete financial markets, so the
10

The stark difference compared to complete financial markets is analogous to some well-known results from the
literature on optimal fiscal policy under different assumptions about the completeness of the financial markets a
government has access to. That literature considers an environment where the government aims to find the least
distortionary means of financing a stochastic sequence of government spending, given that available fiscal instruments
entail distortions which are convex in tax rates. With incomplete financial markets (in the sense that the government
can issue only non-contingent bonds), Barro (1979) finds the government should aim to smooth tax rates, which
implies the stock of government debt follows a random walk. On the other hand, Lucas and Stokey (1983) assume
the government can issue a full set of contingent bonds. In that case, the government now smooths taxes across
states-of-the-world as well as time, and this means the value of outstanding government liabilities is now a purely
forward-looking variable depending on expectations of future fiscal fundamentals. These findings for the behaviour
of government debt mimic the findings for household debt here, with consumption smoothing (whether across time
or also across states of the world) playing the role of tax smoothing in the analysis of optimal fiscal policy.

16

actual debt-to-GDP ratio d1 differs from its complete-markets equivalent d1 . In period 2 there
are no new shocks to react to, and it is assumed that there are no unexpected shifts in monetary
policy unrelated to the economys fundamentals. It follows from [2.10a] and [2.10b] that there is no
uncertainty about c,2 as of period 1, and so the Euler equations [2.10c] imply cb,2 /cb,1 = cs,2 /cs,1 .
Given [2.10b], this means cb,2 = cb,1 and cs,2 = cs,1 . Solving [2.10a][2.10b] for c,1 , c,2 , and d2 in
terms of the state variable d1 , and comparing to the complete-markets equilibrium in [2.17a]:
d2 d2 =
cb,1

cb,1

d1 d1
,
1 + (1 + g2 )1
= cb,2

cb,2

and

2(d1 d1 )
=
,
1 + (1 + g2 )1

cs,1

cs,1

= cs,2

cs,2

2(d1 d1 )
=
.
1 + (1 + g2 )1

This analysis shows that the equilibrium debt-to-GDP ratio is positively related to the previous
periods debt-to-GDP ratio when monetary policy fails to replicate complete financial markets. The
intuition is that consumption smoothing implies that any deviation from the complete-markets
consumption allocation persists over time. While the positive serial correlation of the debt-to-GDP
seems an obvious property, it is important to remember that the debt-to-GDP ratio in a completemarkets economy does not have this feature: it is a purely forward-looking variable.

2.7

Consequences of inflation indexation of bonds

Given that savers holding nominal bonds are exposed to the risk of inflation fluctuations, it might be
thought desirable that bonds be indexed to inflation. To analyse this question, consider an otherwise
identical economy with incomplete markets but where all bonds have coupons indexed to the price
level Pt . To simplify the analysis, suppose all bonds have a maturity of one time period ( = 0).
Writing down flow budget constraints analogous to [2.4] and following all the subsequent steps
leads to a set of equilibrium conditions including [2.10a][2.10c], which now hold in terms of c,t , dt ,
lt , and rt (the superscript signifies the values of these variables with inflation indexation). The set
of equilibrium conditions differs only from the incomplete-markets economy with nominal bonds in
that equation [2.11] for the ex-post real return is replaced by
1 + r1 = 1 + y0 ,

and 1 + r2 = 1 + y1 ,

[2.23]

where yt is the real yield-to-maturity on the indexed bonds, which is both the ex-ante interest rate

and the ex-post real return on these bonds (t = yt = rt+1

).
It is clear that the equilibrium for all real variables is now independent of monetary policy. A
comparison of [2.11] and [2.23] shows that the equilibrium coincides with what the equilibrium of
the nominal-bonds economy would be if monetary was strict inflation targeting (with non-stochastic
inflation rates 1 =
1 and 2 =
2 ).
This result shows that moving to an economy with indexed bonds is generally worse if monetary
policy aims to replicate complete financial markets. Now, whatever the central bank tries to do
has no real effects, so there is no monetary policy intervention that can complete financial markets.
All that happens is indexation locks in the generally sub-optimal outcome that would prevail with

17

strict inflation targeting. This is in spite of the fact that savers are now protected from inflation
fluctuations. The intuition for these findings is best understood by considering an economy in which
both nominal and indexed bonds are available.

2.8

Portfolio choice and the inflation risk premium

The analysis so far has assumed only a single type of bond is available. While the general case of
many different assets is beyond the scope of this paper, it is helpful to consider the consequences of
there being both nominal and indexed bonds that can be bought or issued by households. Assume
for simplicity that all bonds have a maturity of one time period ( = 0).
Writing down a flow budget constraint for this case analogous to [2.4] and following the subsequent steps leads to equilibrium conditions [2.10a][2.10b] as before, which now hold in terms of c,t ,
dt , lt , and rt (the superscript signifies the case of both nominal and indexed bonds). The ex-post
real return rt is now the return on the whole portfolio of nominal and indexed bonds:
rt = (1 st1 )rt + st1 rt ,

[2.24]

where st denotes the portfolio share in indexed bonds (borrowers and savers must have the same
portfolio shares in equilibrium), and the ex-post real returns rt and rt on nominal and indexed
bonds individually are given by equations [2.11] and [2.23] as before. In this version of the model,

the Euler equations [2.10c] must hold for both rt+1 and rt+1
for both borrowers and savers. The

inflation risk premium $t is defined as $t = Et [(1 + rt+1 )/(1 + rt+1

)] 1.
First consider the implications of strict inflation targeting (1 =
1 and 2 =
2 ). Using the Euler
equations [2.10c] and comparing equations [2.11] and [2.23], it follows immediately that rt = rt , and
hence rt = rt = rt for any portfolio share st . The equilibrium of the economy is thus the same
as that where strict inflation targeting is followed in an economy with only nominal bonds (any
portfolio share st is an equilibrium). The inflation risk premium is zero ($t = 0).
Now consider the nominal GDP target [2.20] that replicates complete markets in the economy
with only nominal bonds. The policy still achieves this goal in the economy with both nominal
and indexed bonds, with the equilibrium portfolio share in indexed bonds being zero (st = 0). The
average inflation risk premia are given by:
E$1 =

E [(1 + g1 )(1 + g2 )1 ] E [(1 + g1 ) ]

1,
E [(1 + g1 )1 (1 + g2 )1 ]

E$2 =

E [(1 + g2 )] E [(1 + g2 ) ]
1.
E [(1 + g2 )1 ]

The expression for E$2 is always positive, and E$1 is generally positive (unambiguously so if = 1
or gt is i.i.d.).
These results show that for both strict inflation targeting and a nominal GDP target that
replicates complete financial markets, the equilibrium outcomes with both types of bonds are the
same as those of an economy with only nominal bonds. For strict inflation targeting, this is simply
because the absence of inflation fluctuations makes both types of bond equivalent, thus any portfolio
share can be an equilibrium. Perhaps more surprisingly, the ability to hold indexed bonds does not
change the equilibrium when monetary policy pursues a policy that replicates complete financial
18

markets, which does entail fluctuations in inflation. The equilibrium portfolio share of indexed
bonds is zero in this case, so savers do not attempt to protect themselves from inflation risk. The
reason is the existence of an inflation risk premium, which means that savers earn higher average
returns by holding nominal bonds, which compensate them for the risk they bear.
This perhaps shifts the question to why borrowers continue to issue nominal bonds when lower
real interest rates are available on indexed bonds. However, for borrowers, it is the indexed bond
that is riskier and the nominal bond that is safer because the former obliges the borrower to make
the same real repayments irrespective of real income. When monetary policy replicates complete
financial markets, the countercyclical price level makes nominal bonds behave like equity, so coupons
have a lower real value when real incomes are low, providing insurance to borrowers, for which they
are willing to pay a higher average real interest rate. The higher average real interest rate on nominal
bonds can thus equally well be seen as an insurance premium for borrowers as a risk premium
for savers, and this inflation risk premium is actually a desirable feature of monetary policy. The
inflation fluctuations with nominal GDP targeting are not simply generating risk for savers; inflation
is a hedge for borrowers against the underlying real risk in the economy.11

2.9

Goals and intermediate targets

The argument for a nominal GDP target in this paper is that it replicates the debt-to-GDP ratio
that would be found if the economy had complete financial markets. Thus, the debt-to-GDP ratio
(and finally the implied consumption allocation) is the ultimate objective of policy. Nominal GDP
is simply an intermediate target which helps achieve that goal. It might then be thought preferable
to target the debt-to-GDP ratio directly if this is what monetary policy is actually seeking to
influence. However, as [2.20] shows, there are many possible nominal GDP growth rates consistent
with replicating complete financial markets. So an obvious pitfall of targeting the debt-to-GDP
ratio directly is that it would fail to provide the economy with a nominal anchor. Once a specific
nominal GDP target consistent with [2.20] is chosen, this policy both replicates complete markets
and provides the economy with a nominal anchor.

2.10

Efficiency, fairness, and the distributional effects of policy

The earlier analysis took it for granted that replicating complete financial markets ought to be
an objective of monetary policy. The implicit justification for this is the efficiency properties of
the complete-markets equilibrium, which is Pareto efficient in the absence of any other distortions.
Thus, in spite of policy having distributional effects, the optimal policy can be viewed as supporting
11

It would also be possible to replicate complete financial markets in the two-bond economy with a lower variance
of inflation as long as monetary policy deviates from strict inflation targeting, ensuring there is some correlation
between inflation and real GDP growth. The reduction of the covariance of inflation and real GDP growth would
require all households to hold a positive or negative position in both nominal and indexed bonds, with the size of the
gross positions increasing as the covariance between inflation and real GDP growth shrinks. However, since the gross
positions are larger, monetary policy errors (not generating exactly the covariance between inflation and real growth
that households were expecting) lead to larger deviations from complete financial markets than when households only
need to buy or sell one type of bond. These issues are left for future research.

19

(ex-ante) Pareto efficiency in much the same way that analyses of optimal monetary policy have
pointed to other inefficiencies that policy might correct. However, this justification is incomplete in
one important respect: while the complete-markets equilibrium is Pareto efficient, there are infinitely
many other consumption allocations that would equally well satisfy the criterion of Pareto efficiency.
What might support singling out one particular Pareto-efficient allocation as the target for policy
when the choice among this set necessarily entails taking a stance on distributional questions?
Consider a policymaker who evaluates welfare using a weighted sum of the utilities of borrowers
and savers:
s
b
Ub +
Us ,
[2.25]
W =
2
2
where is the weight assigned to each household of type {b, s}, which reflect the policymakers
distributional preferences. Suppose the policymaker were a social planner who has access to a full
set of state-contingent transfers between households. In this case, the optimal monetary would be
to maximize the welfare function [2.25] subject only to the economys resource constraint, which is
the goods-market clearing condition [2.7]. The first-order conditions for this maximization problem
are:

C,t
= t ,

[2.26]

where t denotes Lagrangian multiplier on resource constraint at time t. This first-order condition
implies the risk-sharing condition [2.16] must hold for any Pareto weights . The resulting consumption allocation would be one of many first-best (Pareto efficient) allocations. Conversely, any
Pareto efficient allocation is a solution of the planners problem for some Pareto weights. The set of
first-best consumption allocations is characterized by those that satisfy the resource constraint and
the risk-sharing condition. The equilibrium of the economy with complete financial markets is one
of these first-best allocations.
Unlike the hypothetical social planner, a policymaker setting monetary policy can only influence
the allocation of consumption through monetary policys impact on inflation and interest rates.
Given the distributional preferences of the central bank as represented by , optimal monetary
policy would maximize welfare [2.25] subject to all the equilibrium conditions [2.10a][2.10c] and
[2.11] of the economy with incomplete financial markets. This maximization problem is subject
to additional constraints beyond the resource constraint, so the resulting second-best consumption
allocation is generally not Pareto efficient because there is a trade-off between efficiency and the
distributional preferences of the policymaker.
However, the nominal GDP targets described earlier are known to replicate the complete-markets
equilibrium, so monetary policy can achieve at least one first-best allocation. The complete-markets
equilibrium is also the only first-best allocation implementable through monetary policy. The intuition for this result is that the risk-sharing condition is necessary for any allocation to be first
best. It is also the only equation that distinguishes the equilibrium conditions with complete markets
from the equilibrium conditions with incomplete markets, with the latter being the implementability
constraints on monetary policy.
20

Therefore, an implicit assumption that justifies the focus on replicating the complete-markets
equilibrium through nominal GDP targeting is that the policymaker has a lexicographic preference
for efficiency over any explicit distributional concerns: any first-best allocation is always favoured
over a non-first-best allocation. In determining optimal monetary policy, the central bank simply adopts whatever Pareto weights are associated with the only implementable ex-ante efficient
allocation.

2.11

Discussion

The importance of the arguments for nominal GDP targeting in this paper obviously depends on
the plausibility of the incomplete-markets assumption in the context of household borrowing and
saving. It seems reasonable to suppose that households will not find it easy to borrow by issuing
Arrow-Debreu state-contingent bonds, but might there be other ways of reaching the same goal?
Issuance of state-contingent bonds is equivalent to households agreeing loan contracts with financial
intermediaries that specify a complete menu of state-contingent repayments. But such contracts
would be much more time consuming to write, harder to understand, and more complicated to enforce than conventional non-contingent loan contracts, as well as making monitoring and assessment
of default risk a more elaborate exercise.12 Moreover, unlike firms, households cannot issue securities
such as equity that feature state-contingent payments but do not require a complete description of
the schedule of payments in advance.13
Another possibility is that even if households are restricted to non-contingent borrowing, they can
hedge their exposure to future income risk by purchasing an asset with returns that are negatively
correlated with GDP. But there are several pitfalls to this. First, it may not be clear which asset
reliably has a negative correlation with GDP (even if GDP securities of the type proposed by
Shiller (1993) were available, borrowers would need a short position in these). Second, the required
gross positions for hedging may be very large. Third, a household already intending to borrow will
need to borrow even more to buy the asset for hedging purposes, and the amount of borrowing may
be limited by an initial down-payment constraint and subsequent margin calls. In practice, a typical
borrower does not have a significant portfolio of assets except for a house, and housing returns most
likely lack the negative correlation with GDP required for hedging the relevant risks.
In spite of these difficulties, it might be argued the case for the incomplete markets assumption
is overstated because the possibilities of renegotiation, default, and bankruptcy introduce some
12

For examples of theoretical work on endogenizing the incompleteness of markets through limited enforcement of
contracts or asymmetric information, see Kehoe and Levine (1993) and Cole and Kocherlakota (2001).
13
Consider an individual owner of a business that generates a stream of risky profits. If the firms only external
finance is non-contingent debt then the individual bears all the risk (except in the case of default). If the individual
wanted to share risk with other investors then one possibility would be to replace the non-contingent debt with
state-contingent bonds where the payoffs on these bonds are positively related to the firms profits. However, what
is commonly observed is not issuance of state-contingent bonds but equity financing. Issuing equity also allows for
risk sharing, but unlike state-contingent bonds does not need to spell out a schedule of payments in all states of the
world. There is no right to any specific payment in any specific state at any specific time, only the right of being
residual claimant. The lack of specific claims is balanced by control rights over the firm. However, there is no obvious
way to be residual claimant on or have control rights over a household.

21

contingency into apparently non-contingent debt contracts. However, default and bankruptcy allow
for only a crude form of contingency in extreme circumstances, and these options are not without
their costs. Renegotiation is also not costless, and evidence from consumer mortgages in both the
recent U.S. housing bust and the Great Depression suggests that the extent of renegotiation may be
inefficiently low (White, 2009a, Piskorski, Seru and Vig, 2010, Ghent, 2011). Furthermore, even expost efficient renegotiation of a contract with no contingencies written in ex ante need not actually
provide for efficient sharing of risk from an ex-ante perspective.
It is also possible to assess the completeness of markets indirectly through tests of the efficient
risk-sharing condition, which is equivalent to perfect correlation between consumption growth rates
of different households. These tests are the subject of a large literature (Cochrane, 1991, Nelson,
1994, Attanasio and Davis, 1996, Hayashi, Altonji and Kotlikoff, 1996), which has generally rejected
the hypothesis of full risk sharing.
Finally, even if financial markets are incomplete, the assumption that contracts are written in
terms of specifically nominal non-contingent payments is important for the analysis. The evidence
presented in Doepke and Schneider (2006) indicates that household balance sheets contain significant
quantities of nominal liabilities and assets (for assets, it is important to account for indirect exposure
via households ownership of firms and financial intermediaries). Furthermore, as pointed out by
Shiller (1997), indexation of private debt contracts is extremely rare. This suggests the models
assumptions are not unrealistic.
The workings of nominal GDP targeting can also be seen from its implications for inflation and
the real value of nominal liabilities. Indeed, nominal GDP targeting can be equivalently described
as a policy of inducing a perfect negative correlation between the price level and real GDP, and
ensuring these variables have the same volatility. When real GDP falls, inflation increases, which
reduces the real value of fixed nominal liabilities in proportion to the fall in real income, and vice
versa when real GDP rises. Thus the extent to which financial markets with non-contingent nominal
assets are sufficiently complete to allow for efficient risk sharing is endogenous to the monetary policy
regime: monetary policy can make the real value of fixed nominal repayments contingent on the
realization of shocks. A strict policy of inflation targeting would be inefficient because it converts
non-contingent nominal liabilities into non-contingent real liabilities. This points to an inherent
tension between price stability and the efficient operation of financial markets.14
That optimal monetary policy in a non-representative-agent model should feature inflation fluctuations is perhaps surprising given the long tradition of regarding inflation-induced unpredictability
in the real values of contractual payments as one of the most important of all inflations costs. As
discussed in Clarida, Gal and Gertler (1999), there is a widely held view that the difficulties this
induces in long-term financial planning ought to be regarded as the most significant cost of inflation, above the relative price distortions, menu costs, and deviations from the Friedman rule that
have been stressed in representative-agent models. The view that unanticipated inflation leads to
14

In a more general setting where the incompleteness of financial markets is endogenized, inflation fluctuations
induced by nominal GDP targeting may play a role in minimizing the costs of contract renegotiation or default when
the economy is hit by an aggregate shock.

22

inefficient or inequitable redistributions between debtors and creditors clearly presupposes a world
of incomplete markets, otherwise inflation would not have these effects. How then to reconcile this
argument with the result that the incompleteness of financial markets suggests nominal GDP targeting is desirable because it supports efficient risk sharing? (again, were markets complete, monetary
policy would be irrelevant to risk sharing because all opportunities would already be exploited)
While nominal GDP targeting does imply unpredictable inflation fluctuations, the resulting real
transfers between debtors and creditors are not an arbitrary redistribution they are perfectly correlated with the relevant fundamental shock: unpredictable movements in aggregate real incomes.
Since future consumption uncertainty is affected by income risk as well as risk from fluctuations in
the real value of nominal contracts, it is not necessarily the case that long-term financial planning
is compromised by inflation fluctuations that have known correlations with the economys fundamentals. An efficient distribution of risk requires just such fluctuations because the provision of
insurance is impossible without the possibility of ex-post transfers that cannot be predicted ex ante.
Unpredictable movements in inflation orthogonal to the economys fundamentals (such as would
occur in the presence of monetary-policy shocks) are inefficient from a risk-sharing perspective, but
there is no contradiction with nominal GDP targeting because such movements would only occur if
policy failed to stabilize nominal GDP.
It might be objected that if debtors and creditors really wanted such contingent transfers then
they would write them into the contracts they agree, and it would be wrong for the central bank to try
to second-guess their intentions. But the absence of such contingencies from observed contracts may
simply reflect market incompleteness rather than what would be rationally chosen in a frictionless
world. Reconciling the non-contingent nature of financial contracts with complete markets is not
impossible, but it would require both substantial differences in risk tolerance across households
and a high correlation of risk tolerance with whether a household is a saver or a borrower. With
assumptions on preferences that make borrowers risk neutral or savers extremely risk averse, it
would not be efficient to share risk, even if no frictions prevented households writing contracts that
implement it.
There are a number of problems with this alternative interpretation of the observed prevalence of
non-contingent contracts. First, there is no compelling evidence to suggest that borrowers really are
risk neutral or savers are extremely risk averse relative to borrowers. Second, while there is evidence
suggesting considerable heterogeneity in individuals risk tolerance (Barsky, Juster, Kimball and
Shapiro, 1997, Cohen and Einav, 2007), most of this heterogeneity is not explained by observable
characteristics such as age and net worth (even though many characteristics such as these have
some correlation with risk tolerance). The dispersion in risk tolerance among individuals with
similar observed characteristics also suggests there should be a wide range of types of financial
contract with different degrees of contingency. Risk neutral borrowers would agree non-contingent
contracts with risk-averse savers, but contingent contracts would be offered to risk-averse borrowers.
Another problem with the complete markets but different risk preferences interpretation relates
to the behaviour of the price level over time. While nominal GDP has never been an explicit target
of monetary policy, nominal GDP targetings implication of a countercyclical price level has been
23

largely true in the U.S. during the post-war period (Cooley and Ohanian, 1991), albeit with a
correlation coefficient much smaller than one in absolute value, and a lower volatility relative to
real GDP. Whether by accident or design, U.S. monetary policy has had to a partial extent the
features of nominal GDP targeting, resulting in the real values of fixed nominal payments positively
co-moving with real GDP (but by less) on average. In a world of complete markets with extreme
differences in risk tolerance between savers and borrowers, efficient contracts would undo the real
contingency of payments brought about by the countercyclicality of the price level, for example,
through indexation clauses. But as discussed in Shiller (1997), private nominal debt contracts have
survived in this environment without any noticeable shift towards indexation. Furthermore, both
the volatility of inflation and correlation of the price level with real GDP have changed significantly
over time (the high volatility 1970s versus the Great Moderation, and the countercyclicality of
the post-war price level versus its procyclicality during the inter-war period). The basic form of
non-contingent nominal contracts has remained constant in spite of this change.15
Finally, while the policy recommendation of this paper goes against the long tradition of citing
the avoidance of redistribution between debtors and creditors as an argument for price stability,
it is worth noting that there is a similarly ancient tradition in monetary economics (which can be
traced back at least to Bailey, 1837) of arguing that money prices should co-move inversely with
productivity to promote fairness between debtors and creditors. The idea is that if money prices
fall when productivity rises, those savers who receive fixed nominal incomes are able to share in
the gains, while the rise in prices at a time of falling productivity helps to ameliorate the burden
of repayment for borrowers. This is equivalent to stabilizing the money value of incomes, in other
words, nominal GDP targeting. The intellectual history of this idea (the productivity norm) is
thoroughly surveyed in Selgin (1995). Like the older literature, this paper places distributional
questions at the heart of monetary policy analysis, but studies policy through the lens of mitigating
inefficiencies in incomplete financial markets, rather than with looser notions of fairness.

3
3.1

Incomplete financial markets in a monetary DSGE model

Households

The economy contains a measure-one population of households. Time is discrete and households
are infinitely lived. There are equal numbers of two types of households, referred to as borrowers
and savers and indexed by type {b, s}. A representative household of type has preferences
given by the following utility function

(
) 1
1+ 1

`1

X
Y
C,t+` H,t+`
,

[3.1]
U,t =
Et
,t+
1
1

1
+

=1
`=0
15

It could be argued that part of the reluctance to adopt indexation is a desire to avoid eliminating the risk-sharing
offered by nominal contracts when the price level is countercyclical.

24

where C,t is per-household consumption of a composite good by type- households at time t, and
H,t is hours of labour supplied. The two types are distinguished by their subjective discount factors,
with ,t being the discount factor of type- households between time t and t + 1. Both types have
a constant coefficient of relative risk aversion given by , and a constant elasticity of intertemporal
substitution given by 1 . The household-specific Frisch elasticity of labour supply is .
Each household of type receives real income Y,t at time t, to be specified below. The discount
factor ,t of type- households is assumed to be the following:


C,t
,t =
, where (c) = c(1) ,
[3.2]
Y,t
and where the parameters b , s , and are such that 0 < b < s < and 0 < < 1. It is
assumed individual households of type take ,t as given, that is, they do not internalize the effect
of their own consumption on the discount factor.
There are two differences compared to a representative-household model with a standard timeseparable utility function. First, there is heterogeneity in discount factors because borrowers
are more impatient than savers (b < s ), all else equal. This is the key assumption that will
give rise to borrowing and saving in equilibrium by the households that have been referred to as
borrowers and savers. Second, discount factors display the marginal increasing impatience ( < 1)
property of Uzawa (1968), in that the discount factor is lower when consumption is higher (relative
to income), all else equal. This assumption is invoked for technical reasons because it ensures the
wealth distribution will be stationary around a well-defined non-stochastic steady state.16 That
households take discount factors as given is assumed for simplicity and is analogous to models of
external habits (see for example, Abel, 1990).17
The composite good C,t is a CES aggregate of a measure-one continuum of differentiated goods
indexed by [0, 1]. The aggregator is the same for both types of households and features a
constant elasticity of substitution between goods. Households allocate spending C,t () between
goods to minimize the nominal expenditure Pt C,t required to obtain C,t units of the consumption
aggregator:

 1
Z
Z
1
,
[3.3]
Pt C,t = min
Pt ()C,t ()d s.t. C,t =
C,t () d
{C,t ()}

[0,1]

[0,1]

where Pt () is the nominal price of good .

Households of type face a real wage w,t for their labour. All households own equal (nontradable) shareholdings in a measure-one continuum of firms, with firm paying real dividend Jt ().
All households are assumed to face a common lump-sum tax Tt in real terms. Real disposable
16

None of the qualitative results depends on being significantly below one ( = 1 is the standard case of fixed
discount factors), and moreover, the quantitative importance of the results does not vanish when is arbitrarily close
to one. The assumption is analogous to those employed in small open-economy models to ensure a stationary net
foreign asset position (see Schmitt-Grohe and Uribe, 2003). An alternative is to work with an overlapping generations
model where the utility function is entirely standard, which automatically has a stationary wealth distribution because
households have finite lives. This avenue was explored in an earlier working paper (Sheedy, 2013).
17
The assumption can be relaxed at the cost of more complicated algebra, but there is little impact on the results
when is close to one. The changes to the results are described in the appendix.

25

income for households of type is thus

Z
Jt ()d Tt .
Y,t = w,t H,t +

[3.4]

[0,1]

3.2

Incomplete financial markets

The only liability that can be issued by households is a non-contingent nominal bond. Households
can take positive or negative positions in this bond (save or borrow), and there is no limit on
borrowing other than being able to repay in all states of the world given non-negativity constraints
on consumption. With this restriction, no default will occur, and thus bonds are risk-free in nominal
terms.18
The nominal bond has the following structure. One newly issued bond at time t makes a stream
of coupon payments in subsequent time periods, paying 1 unit of money (a normalization) at time
t + 1, then units at t + 2, 2 at t + 3, and so on (0 < ). The geometric structure of
the coupon payments means that a bond issued at time t ` is after its time-t coupon payment
equivalent to a quantity ` of new date-t bonds. If Qt denotes the price in terms of money at time
t of one new bond then the absence of arbitrage opportunities requires that bonds issued at date
t ` have price ` Qt at time t. It therefore suffices to track the overall quantity of bonds in terms
of new-bond equivalents, rather than the quantities of each vintage separately.19
The flow budget identity at time t of households of type is:
C,t +

Qt B,t
(1 + Qt )B,t1
= Y,t +
,
Pt
Pt

[3.5]

where B,t denotes the outstanding quantity of bonds (in terms of new-bond equivalents) held (or
issued, if negative) by type- households at the end of period t. The term 1 + Qt refers to the
coupon payment plus the resale value of bonds acquired or issued in the past.

3.3

Firms

Firm [0, 1] is the monopoly producer of differentiated good . Goods are produced using an
aggregator of labour inputs. Production of good is denoted by Yt (), and firm s labour usage by
Ht (), and wt denotes the wage cost per unit of Ht (). The firm pays out all real profits at time t as
dividends Jt ():


1
Pt ()
P
t ()
Jt () =
Yt () wt Ht (), where Yt () = At Ht () 1+ , and Yt () =
Ct . [3.6]
Pt
Pt
The first equation following the definition of profits is the production function, with At denoting
the common exogenous productivity level, and where the parameter determines the extent of
diminishing returns to labour ( 0). The final equation in [3.6] is the demand function that arises
from the household expenditure minimization problem [3.3].
18

The model abstracts from the choice of default when repayment is feasible.
Woodford (2001) uses this modelling device to study long-term government debt. See Garriga, Kydland and

Sustek
(2013) for a richer model of mortgage contracts.
19

26

The labour input Ht () is an aggregator of labour supplied by the two types of households. Firms
are assumed to receive a proportional wage-bill subsidy at rate 1 . Firms choose labour inputs
H,t () to minimize the post-subsidy cost wt Ht () of obtaining a unit of the aggregate labour input
Ht ():
1

wt Ht () = min (1 1 ) (wb,t Hb,t + ws,t Hs,t ) s.t. Ht () = 2Hb,t () 2 Hs,t () 2 .

H,t ()

[3.7]

The labour aggregator has a Cobb-Douglas functional form, implying a unit elasticity of substitution
between different labour types.20

3.4

Sticky prices

Price adjustment is assumed to be staggered according to the Calvo (1983) pricing model. In each
time period, there is a probability that firm must continue to use its previous nominal price
Pt1 (). If at time t a firm does receive an opportunity to change price, it sets a reset price denoted
by Pt . The reset price is set to maximize the current and expected future stream of profits. Future
profits conditional on continuing to charge Pt are multiplied by the probability ` that the reset
price will actually remain in use ` periods ahead, and are discounted using the real interest rate t :

(
)
!1
!(1+)

`
X

wt+` Ct+`
Pt

Pt
Ct+` .

[3.8]
Et Q`1
max
1+
P
P
A
Pt
t+`
t+`
(1
+

)
t+
t
=0
`=0

3.5

Money and monetary policy

The economy is cash-less in that money is not required for transactions, but money is used as a
unit of account in writing financial contracts and in pricing goods. Monetary policy is assumed to
be able to determine a path for the price level Pt .

3.6

Fiscal policy

The only role of fiscal policy here is to provide the wage-bill subsidy to firms by collecting equal
amounts of a lump-sum tax from all households. It is assumed the fiscal budget is in balance, so
taxes Tt are set at the level required to fund the current subsidy:21
Z
1
Tt =
(wb,t Hb,t () + ws,t Hs,t ()) d.
[3.9]
[0,1]
20

The use of an aggregator of different labour types is the standard approach to studying the welfare costs of sticky
wages (Erceg, Henderson and Levin, 2000). Here, a unit elasticity of substitution is the most analytically convenient
assumption, though a priori it is not clear whether the labour of different household types is more substitutable or
more complementary than this. In the model, being a borrower is simply a matter of being more impatient than
savers, but empirically, the average borrower is likely to differ from the average saver in respects such as age that
might mean their labour is not perfectly substitutable.
21
The wage-bill subsidy is a standard assumption which ensures the economys steady state is not distorted (Woodford, 2003). A balanced-budget rule is assumed to avoid any interactions between fiscal policy and financial markets.

27

3.7

Market clearing

Market clearing in goods, labour, and bond markets requires:

1
1
Cb,t () + Cs,t () = Yt (), for all [0, 1];
2
Z2
1
H,t ()d = H,t , for all {b, s};
2
[0,1]
1
1
Bb,t + Bs,t = 0.
2
2

3.8

[3.10a]
[3.10b]
[3.10c]

Equilibrium

The derivation of the equilibrium conditions is presented in the appendix. The analysis of incomplete
financial markets follows the method used for the simple model in section 2, while other aspects of
the model are standard features of New Keynesian models with sticky prices.
The debt-to-GDP ratio dt , the loans-to-GDP ratio lt , and the ex-post real return rt are defined
as follows:
1 Qt Bb,t
(1 + Qt )Pt1
1 (1 + Qt )Bb,t1
, Lt
, and 1 + rt
,
[3.11]
Dt
2
Pt
2 Pt
Qt1 Pt
and the ratios of consumption to GDP are c,t C,t /Yt . The yield-to-maturity jt is defined by:
Qt =

X
`=1

`1
,
(1 + jt )`

implying jt =

1
1 + .
Qt

28

[3.12]

With these definitions in hand, the full set of equilibrium conditions is collected below:
t = Et rt+1 ;


1 + rt
dt =
lt1 ;
1 + gt

[3.13a]
[3.13b]

cb,t = 1 2(dt lt ),
"

and cs,t = 1 + 2(dt lt );


 #
c,t+1

(1 + rt+1 )(1 + gt+1 )

;
c,t

[3.13c]

1 = ,t Et

[3.13d]

(1)

,t = c,t
;
)
( `1
Y
lim
,t+ (1 + gt+1+ )1 c
,t+` lt+` = 0;

[3.13e]
[3.13f]

=0



1 + jt
1 + jt1
1 + rt =
;
1 + t
1 + jt
#
"( `

 )

Y
(1
+
g
)
c
t+
,t+
lim ` Et
,t+1
(1 + jt+` )1 = 0;
`
(1
+

)
c
t+
,t+1
=1
1 + gt =

xt = (1 )(1 + )

[3.13i]
++ 1+

t Yt

Q`

=1 (1 + gt+ )
Et Q`1
=0 (1 + t+ )
`=0

s
1+s

cb,t cs,t

1++ 1+

At

t = (1 + t )(1+) 1+
t1 + (1 )
(

b
1+b

[3.13h]

Yt
;
Yt1
1+

[3.13g]

1 (1 + t )
1

1(1+t )1
1

Q`

=1 (1

b
s
1+b + 1+s

 (1+)
1

+ t+

1
1+

 1+
1
)1

[3.13j]

[3.13k]

Q`

=1 (1

x
1 t+`

+ t+

)(1+)

= 0.

[3.13l]

The variable xt is the level of real marginal cost for a firm whose good sells at a price equal to the
general price level, and t is the effect of relative-price distortions on aggregate productivity.
In a steady state where exogenous productivity At is growing at a constant rate, there is a
steady-state rate of real GDP growth g. The steady-state consumption-income ratios are given by
cb = 1 ,

cs = 1 + ,

where

1 (b /s )(1)
,
1 + (b /s )(1)

[3.14]

where 0 < < 1. The term depends on the relative patience b /s of the two household types
and the utility-function parameters and . The steady-state discount factors and real interest
rate are:

 (1)
1
1

1 + g
(1)
(1)

b = s =
b
+ s
2
, = r =
1, where (1 + g)1 ,

[3.15]
and it is assumed g is low enough to ensure that 0 < < 1. In the steady state, the discount factors
29

of the two types are aligned at , which is effectively an average of the patience parameters b and
s . The steady-state debt-to-GDP ratio can be written in terms of and as follows:
d =

,
2(1 )

and l =

.
2(1 )

[3.16]

The model can be parameterized directly with and rather than the two patience parameters b
and s (leaving , , and g to be chosen separately). The term plays the usual role of the discount
factor in a representative-household economy given its relationship with the real interest rate (with
an adjustment for steady-state real GDP growth). The term quantifies the extent of heterogeneity
between borrower and saver households, which is related to the amount of borrowing and saving
that occurs in equilibrium, and hence to the debt-to-GDP ratio in [3.16]. Given equation [3.14],
can be interpreted as the debt service ratio because it is the net fraction of income transferred by
borrowers to savers.22 As will be seen, is a sufficient statistic for the extent of heterogeneity in
the economy, with 0 being the limiting case of a representative-household economy (b s ).
Finally, rather than specify the bond coupon parameter directly, the steady-state fraction of debt
that is not refinanced each period is taken to be a parameter. This fraction is = /((1+
)(1+ g)).

Optimal monetary policy

This section studies the features of optimal monetary policy in the full model of section 3. The first
step is to consider what the equilibrium of the economy would be in the absence of two key frictions:
incomplete financial markets and sticky prices.

4.1

A first-best benchmark

Consider an economy with complete financial markets, but which is otherwise identical to that
described in section 3 (in particular, this economy may have sticky prices). The equilibrium
consumption-GDP ratios and discount factors in this hypothetical economy are
cb,t = 1 ,

cs,t = 1 + ,

and b,t
= s,t
= ,

[4.1]

which are identical to the steady-state values of these variables from [3.14] and [3.15]. As expected,
this equilibrium features full risk sharing between borrowers and savers. The associated level of the
debt-to-GDP ratio is
"
#
`
X Y

[4.2]
dt = Et
` (1 + gt+ )1 .
2
=1
`=0
The complete-markets debt-to-GDP ratio dt is referred to as the natural debt-to-GDP ratio. This
terminology is motivated by analogy with such concepts as the natural rate of interest, the natural
rate of unemployment, and the natural level of output. In common with these other concepts, it
represents an economic outcome that would be achieved in the absence of a particular friction (here,
22

Strictly speaking, when real GDP growth is different from zero, is the debt service ratio net of new borrowing.

30

the friction is the inability to issue state-contingent bonds, in contrast to the frictions of imperfect
information or nominal rigidities that affect the supply of output or labour in many models).
A generalization of this concept is the equilibrium of an economy where complete financial
markets are open only from some date t0 onwards, taking as given the initial wealth distribution at
t0 1. The consumption ratios in this case are denoted by c,t|t0 , and the associated debt-to-GDP
ratio by dt|t0 . The values of c,t|t0 depend only on the predetermined wealth distribution at t0 1.
Now consider an economy with both complete financial markets (open from t0 onwards) and
fully flexible prices (the case where = 0). The equilibrium of this economy represents one firstbest allocation starting from date t0 . Given the complete-markets consumption ratios c,t|t0 (which
depend only on the wealth distribution at t0 1), the equilibrium level of output Yt|t 0 can be obtained,
which is referred to as the natural level of output:
! 1 1+
 b
 s
++
s
1+
b +
1++
1
1+
1+
1+
1+s

s
b
.
[4.3]
Yt|t 0 =
c b,t|tb0 c s,t|t
A
t
0
1+
It is instructive to consider the special cases of fully flexible prices or complete financial markets.
First suppose financial markets are incomplete but prices are fully flexible. Monetary policy can
be used to replicate complete financial markets as in the model of section 2, which implies that
households make the same labour supply decisions as they would in a complete-markets economy.
Since prices are fully flexible, this does not lead to any relative price distortions, and the aggregate
level of output is efficient.
The second special case is that of complete financial markets or a representative household
( = 0), but where prices are sticky. Because financial markets are complete, full risk sharing occurs
without any policy intervention. In this case, strict inflation targeting (with a zero inflation target,
t = 0) avoids any relative-price distortions and ensures the level of aggregate output is what would
prevail were prices were fully flexible, which is efficient.
Thus, with flexible prices, there is no trade-off between a policy that supports risk sharing and
achieving the optimal level of aggregate output. With complete financial markets, there is no tradeoff between avoiding relative-price distortions and achieving the optimal level of aggregate output.
However, with both incomplete financial markets and sticky prices, all three objectives of risk
sharing, avoiding relative-price distortions, and closing the output gap are in conflict. These tradeoffs are studied by deriving a loss function to approximate the welfare function, and approximations
of the equilibrium conditions that represent the constraints on monetary policy.

4.2

Policy tradeoffs

An exact analytical solution for optimal monetary policy is not available, so this section resorts to
finding the log-linear approximation (the first-order perturbation around the non-stochastic steady
state), which can be found analytically. The notational convention is that variables in a sans
serif font denote log deviations of the equivalent variables in roman letters from their steady-state
values (log deviations of interest rates, inflation rates, and growth rates are log deviations of the

31

corresponding gross rates; for variables that have no steady state, the sans serif letter simply denotes
the logarithm of that variable). It is also convenient to define the debt gap dt|t0 dt /dt|t0 , the
deviation of the actual debt-to-GDP ratio from its value with complete financial markets, and the
output gap Yt|t0 Yt /Yt|t 0 .
A parameter restriction on the Frisch elasticities of borrowers and savers is imposed that implies
the wealth distribution has no effect on aggregate labour supply (up to a first-order terms):
b = (1 )/(1 + ),

and s = (1 + )/(1 ),

[4.4]

where is the steady-state debt service ratio defined in [3.14] and is the effective aggregate Frisch
elasticity of labour supply (it is assumed that < 1/). In practice, this assumption does not
make much difference to the results, but it does simplify the analysis and make the model easier to
compare to a standard New Keynesian model. One implication is that the growth rate of output
with flexible prices gt does not depend on the initial wealth distribution, so the t0 subscript can
t . The case where the distribution of wealth does
be dropped here, and also from the output gap Y
affect aggregate labour supply is considered in the appendix.
= c
1 that support the first-best equilibrium with
Using the Pareto weights
,t0 |t0 /Y t0
|t0
flexible prices and complete financial markets (from t0 onwards), a second-order approximation
(around a steady state with zero inflation) of the welfare function is given below in terms of the loss
function Lt0 :




2 (1 )2

1 X tt0
d2t|t0
Et0
1+
Lt0 =
2
2
2 t=t
(1 )(1 )
(1 + )(1 + )
0
 

(1 + )
1 + 2
2
+
Yt . [4.5]
t + + +
(1 )(1 )

The units of the loss function are percentage equivalents of initial output. The loss function includes the squared debt-to-GDP gap dt|t0 , which is a sufficient statistic for the welfare loss due to
deviations from complete markets. The coefficient is increasing in risk aversion and the degree
of heterogeneity between borrower and saver households as measured by . Intuitively, greater risk
aversion increases the importance of the risk sharing found in complete financial markets, while
greater heterogeneity between households leads to larger financial positions of borrowers and savers,
so a given percentage change in financial wealth has a larger impact on consumption.
The second term in the loss function [4.5] is the squared inflation rate, which is a sufficient
statistic for the welfare loss due to relative-price distortions. This is a well-known property of Calvo
pricing, and the coefficient in the loss function is the same as in standard models (see Woodford,
2003). The coefficient of inflation is increasing in the price elasticity of demand because a higher
price sensitivity implies that a given amount of price dispersion now leads to greater dispersion of
the quantities produced of different goods for which preferences and production technologies are
identical. The coefficient is increasing in price stickiness because longer price spells imply that a
given amount of inflation leads to greater relative-price distortions. The coefficient is also increasing
in the parameter (the output elasticity of marginal cost), which will be seen to determine the

32

degree of real rigidity in the economy. Both nominal and real rigidity increase the welfare costs of
inflation.
The third term in the loss function is the squared output gap. This represents the losses from
output and employment deviating from their efficient levels. The coefficient is also the same as
found in a standard New Keynesian model.
Since the steady state of the economy is not distorted (there are no linear terms in the loss
function [4.5]), a first-order accurate approximation of optimal monetary policy can be obtained
by minimizing the loss function subject to first-order accurate approximations of the economys
equilibrium conditions. The relevant constraints are given below:
t,
(t Et t+1 ) = Y

where

(1 + )
,
(1 )(1 )

and + +

1+
;

Et dt+1|t0 = dt|t0 ;

[4.6a]
[4.6b]

jt1 jt
t|t + Y
t1|t t (1 )(1 ) (t Et1 t ) = r ;
dt|t0 + dt1|t0 Y
t
0
0
1

[4.6c]
lim ()` Et jt+` = 0,

[4.6d]

where rt is a variable that depends only on the exogenous growth rate gt of the natural level of
output:
rt = gt + (1 )

` (Et gt+`
Et1 gt+`
).

[4.7]

`=0

The first constraint [4.6a] is the standard New Keynesian Phillips curve linking inflation, expectations of future inflation, and the output gap. The coefficients are the same as in the standard
New Keynesian model. To understand the constraints [4.6b][4.6d] specific to the incompleteness of
financial markets, first consider the special case where the Frisch elasticity of labour supply is zero,
t is always zero.
in which case the output gap Y

4.3

Incomplete financial markets versus relative-price distortions

t = 0
With a zero Frisch elasticity of labour supply ( = 0), the Phillips curve [4.6a] reduces to Y
(since ). In this case, the constraint in [4.6c] simplifies to:
jt1 jt
dt|t0 + dt1|t0 t = rt .
1

[4.8]

The constraint [4.6b] determines the predictable component of the debt-to-GDP gap dt|t0 , and [4.8]
links inflation t and nominal bond yields jt to the unexpected component of the debt-to-GDP gap.
The final constraint [4.6d] is simply a transversality condition on the bond price or yield. Optimal
t)
monetary policy minimizes the loss function [4.5] (ignoring here the term in the output gap Y
subject to [4.6b], [4.8], and [4.6d].

33

The optimal monetary policy features fluctuations in the debt-to-GDP gap:

dt|t0 = dt1|t0 (1 )t ,

with t =

` 1 (1 ` ) (Et gt+` Et1 gt+` ) ,

[4.9a]

`=0

1

(1 + )(1 2 )(1 )2 (1 2 )(1 2 )
.
and = 1 +
2 (1 )(1 )(1 )2

[4.9b]

Optimal policy also features inflation fluctuations, with optimal inflation persistence determined by
the maturity of debt contracts (the parameter ):

(1 2 ) if t t + 1
t1
t
0
.
[4.10]
t =
(1 2 )
if t = t
t

Fluctuations in the growth rate gt of flexible-price output (due to shocks to TFP) lead to changes
in the ex-post real return on the complete-markets portfolio. To replicate complete financial markets,
the central bank needs to vary inflation and nominal bond yields so as to mimic this real return.
Overall, what matters are unexpected changes in the discounted sum of current and future growth
rates, adjusted for any mitigating (or aggravating) changes in real interest rates. This is the shock
t given in [4.9a]. The term (1 ` ) is the adjustment for changes in real interest rates caused by
revised expectations of the economys future growth prospects. The parameter is the elasticity
of the real interest rate with respect to expected real GDP growth. Since changes in (ex-ante) real
interest rates only matter to the extent that debt is refinanced, for growth expectations ` periods
ahead, the interest-rate effect is proportional to the fraction 1 ` of existing debt that will be
refinanced by then.
With sticky prices, replicating complete financial markets through variation in inflation is now
costly, so the central bank tolerates some deviation from complete markets.23 To the extent that in
[4.9b] is less than one, a shock t leads to fluctuations in the debt-to-GDP gap dt . These fluctuations
are persistent because of the constraint [4.6b]: the serial correlation of the debt-to-GDP gap is .
Once a non-zero debt gap arises at time t, there is no predictable future policy action that can undo
its future consequences.
If were equal to 1, equation [4.9a] shows that the debt-to-GDP gap would be completely
stabilized, and if were 0, equation [4.10] shows that inflation would be completely stabilized.
Since 0 < < 1, optimal monetary policy can be interpreted as a convex combination of strict
inflation targeting and a policy that replicates complete financial markets. As the responses of dt|t
0

and t to the shock t are linearly related to , the terms and 1 can be interpreted respectively
as the weights on completing financial markets and avoiding relative-price distortions. Comparing
equations [4.5] and [4.9b], it can be seen that is positively related to the ratio of the coefficients
of d2t|t0 and 2t in the loss function divided by (1 2 ) and (1 2 ).
Greater risk aversion () or more heterogeneity and hence more borrowing () increase the
23
If the nominal rigidity were sticky wages rather than sticky prices, the central bank would care about nominal
wage inflation rather than price inflation (Erceg, Henderson and Levin, 2000). In that case, the tension with the goal
of replicating complete financial markets is reduced because targeting nominal income growth is less likely to be in
conflict with stabilizing nominal wage inflation than stabilizing price inflation.

34

coefficient of d2t|t0 and thus ; a larger price elasticity of demand (), stickier prices (), or more real
rigidities () increase the coefficient of 2t and thus reduce . The optimal trade-off is also affected
by the constraints in [4.6b] and [4.8], which explain the presence of the terms (12 ) and (12 )
in the formula for . A greater value of increases the persistence of the debt-to-GDP gap, which
makes fluctuations in dt|t0 more costly than suggested by the loss function coefficient alone. The
parameter affects the link between bond yields and the debt-to-GDP gap. It is seen that an
increase in leads to a higher value of , the intuition for which is related to the optimal behaviour
of inflation. Finally, note that while the optimal policy responses depend on the stochastic process
for real GDP growth, the optimal weight does not.
Equation [4.10] shows that optimal monetary policy features inflation persistence with serial
correlation given by the debt maturity parameter = /(1 + g). The steady-state fractions of
existing and newly issued debt are and 1 respectively, so the result is that inflation should
return to its average value at the same rate at which debt is refinanced. At the extremes, one-period
debt ( = 0 and = 0) corresponds to serially uncorrelated inflation, while perpetuities ( = 1, for
which 1) correspond to near random-walk persistence of inflation.
To understand this, note that with one-period debt, the current bond yield jt disappears from
the constraint [4.8], thus the only way that policy can affect dt|t0 is through an unexpected change
in current inflation. With longer maturity debt, the range of policy options increases. Changes
in current bond yields jt are also relevant in addition to current inflation, and the bond yield is
affected by expectations of future inflation. The three constraints [4.6b], [4.8], and [4.6d] imply
P
`
`
dt|t0 = dt1|t0 t
`=0 () (Et t+` Et1 t+` ), where indicates the fraction of existing debt
that will not have been refinanced after ` time periods. This shows it is now possible to use inflation
that is spread out over time to influence the debt-to-GDP gap dt|t0 , not only inflation surprises.
Furthermore, this inflation smoothing is optimal because the welfare costs of inflation are
convex (inflation appears in the loss function [4.5] as 2t ), so the costs of a given cumulated amount
of inflation are smaller when spread out over a number of quarters or years than when all the
inflation occurs in just one quarter. This is analogous to the tax smoothing argument of Barro
(1979). Interestingly, the argument shows that high degrees of inflation persistence need not be
interpreted as a failure of policy. Differently from the tax smoothing argument, it is generally not
optimal for inflation to display random walk or near-random walk persistence unless debt contracts
are close to perpetuities. As the maturity parameter is reduced and thus falls significantly
below one, expectations of inflation far in the future have a smaller effect on bond yields than
inflation in the near future. The further in the future inflation is expected to occur, the less effective
it is at influencing real returns and thus the debt-to-GDP ratio.
Even if optimal monetary policy places a substantial weight on the problem of incomplete
markets compared to relative-price distortions, in what sense does monetary policy still resemble
a nominal GDP target? Optimal inflation smoothing Et t+1 = t (from [4.10]) pins down the
predictable component of the inflation rate, but this in itself is not a complete description of policy because it leaves unspecified the unexpected component of inflation. Together with inflation
smoothing, specifying how much inflation reacts to a shock on impact, or equivalently, how much
35

cumulated inflation will follow a shock, completely characterizes the path of all nominal variables.
It turns out that optimal monetary policy retains the essential feature of nominal GDP targeting
in generating a negative comovement between prices and output. However, because of the desire to
smooth inflation, the central bank should not aim to stabilize nominal GDP (or a weighted measure
of nominal GDP) exactly on a quarter-by-quarter basis. Instead, optimal policy can be formulated
as a long-run target for weighted nominal GDP. When real GDP is non-stationary because TFP
follows a non-stationary process, optimal monetary policy features cointegration between the price
level and output. This cointegrating relationship can be interpreted a long-run target for weighted
nominal GDP because there is some linear combination of prices and real output that is eventually
returned to a constant or a deterministic trend following a shock to output.

4.4

The ex-ante real interest rate and the output gap

In the general case where labour supply is elastic ( > 0), monetary policy can also affect the
output gap and the ex-ante real interest rate. This is important not only because of a concern for
stabilizing the output gap, but also because the output gap affects the debt-to-GDP gap through
the constraint [4.6c]. The main reason for this is the connection between the output gap and the
t+1 Y
t + , where
ex-ante real interest rate t . The real interest rate is given by t = Et Y
t|t0

t = Et [Yt+1 Yt ] is the real interest rate at the natural level of output. This is the source of the
t|t in [4.6c].
coefficient of Y
0
The problem of minimizing the loss function [4.5] subject to the constraints [4.6a][4.6d] (with
commitment at some distant past date t0 ) has a solution with the following properties:
P
`
the debt gap dt is an AR(1) process with an innovation proportional to t =
`=0 (1 (1

` ))(Et gt+`
Et1 gt+`
) and an autoregressive root ; inflation t is an ARMA(2,1) process with an
innovation proportional to t and autoregressive roots equal to and , where the latter is:
2

=
1++

1++

2

[4.11]

which satisfies 0 < < 1. With strict inflation targeting, the debt gap dt is an AR(1) process with
an innovation proportional to t and an autoregressive root . The solution for the debt gap is
therefore a multiple of the debt gap under the optimal policy, and the optimal policy weight on
incomplete financial markets is defined so that the debt gap under the optimal policy is equal to
the debt gap under strict inflation targeting multiplied by 1 .
To understand the new aspects of the optimal monetary policy problem with elastic labour
supply, note that policy can now influence three variables which have implications for the debt gap
and thus the extent to which complete financial markets are replicated: inflation, real GDP, and the
ex-ante real interest rate. Inflation affects the ex-post real return on nominal bonds and thus the
value of existing debt (as before). Real GDP (and hence the output gap) affects the denominator
of the debt-to-GDP ratio. The ex-ante real interest rate affects the ongoing costs of servicing debt
relative to the stream of current and future labour income (formally, the ex-ante real interest rate

36

influences the debt gap by changing the level of the debt-to-GDP ratio consistent with risk sharing).
By combining the four constraints [4.6a][4.6d], the evolution of the debt gap depends on the
shock t and three terms related respectively to inflation, the output gap, and the ex-ante real
interest rate.24 The effect of inflation on debt is related to the unexpected change in the term
Et [t + ()t+1 + ()2 t+2 + ] as before, where ` represents the stock of debt issued in the
past that has not been refinanced ` periods after the shock.
t has the effect of directly boosting real GDP growth at time
An increase in the output gap Y
t and thus reducing the debt-to-GDP ratio, but the impact on the debt gap is more subtle. Since
monetary policy has only a temporary influence on real GDP, extra growth now reduces overall
growth in the future by exactly the same amount. Given the link between growth expectations
and the debt-to-GDP ratio consistent with risk sharing (the log linearization of equation [4.2] is
P
`
dt = (1 )
`=1 Et gt+` ), the effect of the output gap on the debt gap actually depends on
t + Et [(Y
t+1 Y
t ) + 2 (Y
t+2 Y
t+1 ) + ], not just Y
t . With the New Keynesian Phillips curve
Y
[4.6a], it is seen that this term is equal to (1)(/)t , the reciprocal of the long-run Phillips curve
slope multiplied by current inflation. Since it is reasonable to set the discount factor close to one
(in which case the long-run Phillips curve is close to vertical), this term is negligible for all practical
purposes (it is not exactly zero because future growth is discounted relative to current growth, so
by bringing growth forwards, there is still a small positive effect). Monetary policy therefore cannot
have a sustainable impact on the burden of debt simply through its temporary effect on the level of
real GDP.
It does not follow however that expansionary or contractionary monetary policy has no effect on
the debt burden beyond its implications for ex-post real returns through inflation. There remains
the option of changing ex-ante real interest rates. Intuitively, expansionary monetary policy that
reduces the ex-ante real interest rate is effectively a transfer from savers to borrowers, what might
be labelled financial repression. While monetary policy cannot permanently affect real interest
rates, there is no reason why cutting the real interest rates now means real interest rates in the
future must be higher than they would otherwise have been (unlike real GDP growth, as discussed
above).
Changing ex-ante real interest rates thus provides monetary policy with an alternative to influencing the debt gap through the effect of inflation on ex-post real returns. In contrast to the
latter, which is effective only while debt contracts are not refinanced, the former is effective only
when refinancing does take place. For debt refinanced ` periods after a shock at time t, the impact
of monetary policy on the date-t debt burden is determined by the discounted sum of real interest
24

t is:
The precise expression for the evolution of the debt-to-GDP gap d
t = d
t1
d

t + (Et Et1 )

()` t+`

`=0

`
+ (1 ) (Et Et1 )t + (1 ) (Et Et1 )
() (t+` t+1+` ) .

`=0

37

gaps Et [`+1 t+` + `+2 t+`+1 + `+3 t+`+2 + ] from t + ` onwards. Given the New Keynesian
t+1 Y
t ] linking the real interest rate and output
Phillips curve [4.6a] and the equation t = Et [Y
gaps, these terms reduce to (/)`+1 Et [t+`+1 t+` ], meaning that the slope of the inflation
path over time is an indicator of financial repression through ex-ante real interest rates. With a
(steady-state) fraction (1 )` of debt issued prior to date t being refinanced at t + `, the overall
effect of changes in ex-ante real interest rates on the debt burden is given by the unexpected change
P
`
in (/)(1 )
`=0 () Et [t+`+1 t+` ]. Thus, the more the inflation trajectory is smoothed,
the smaller is the effect of monetary policy on the ex-ante real interest rate.
As the maturity of debt increases, successful financial repression requires an inflation trajectory
with a non-zero slope further in the future. Given the Phillips curve, the required inflation path
entails output gap fluctuations over a longer horizon, increasing the losses from following such
a policy. Financial repression is therefore not well suited to stabilizing the debt gap when debt
contracts have a long maturity, in which case a policy of influencing ex-post real returns through
inflation smoothly spread out over time is effective at a much lower cost in terms of the implied
inflation and output gap fluctuations. But for short-maturity debt where only immediate inflation
surprises can affect ex-post real returns, financial repression provides an additional tool for stabilizing
the debt gap, with the losses from following this policy being small when the short-run Phillips curve
is relatively flat.

Quantitative analysis of optimal monetary policy

This section presents a quantitative analysis of the nature of optimal monetary policy taking into
account all the features of the full model from section 3.

5.1

Calibration

Let T denote the length in years of one discrete time period in the model. The numerical results
presented here assume a quarterly frequency (T = 1/4), but the choice of frequency does not
significantly affect the results. The parameters of the model are , , , , , , , , and . As
far as possible, these parameters are set to match features of U.S. data.25 The baseline calibration
targets and the implied parameter values are given in Table 1 and justified below.
Consider first the parameters and (the choice of these parameters is equivalent to specifying
the patience parameters b and s ). These are calibrated to match evidence on the average price
and quantity of household debt. The price of debt is the average annual continuously compounded
real interest rate r paid by households for loans. As seen in [3.15], the steady-state growth-adjusted
real interest rate is related to . Let g denote the average annual real growth rate of the economy.
Given the length of the discrete time period in the model, 1 + = er T and 1 + g = eg T . Hence, using
25

The source for all data referred to below is Federal Reserve Economic Data (http://research.stlouisfed.
org/fred2).

38

Table 1: Baseline calibration

Calibration targets
Real GDP growth
Real interest rate
Debt-to-GDP ratio
Coefficient of relative risk aversion
Marginal propensity to consume
Frisch elasticity of labour supply
Average duration of debt
Price elasticity of demand
Marginal cost elasticity w.r.t. output
Average duration of price stickiness

Implied parameter values

g
r
D

1.7%
5%
Discount factor
130% Debt service ratio

6%

Discount factor elasticity

Tm

Debt maturity parameter

Tp

8/12

Calvo pricing parameter

0.992
8.6%
5
0.993
2
0.967
3
0.5
0.625

Notes: The parameters are derived from the calibration targets using equations [5.1][5.5]. The calibration
targets are specified in annual time units; the parameter values assume a quarterly model (T = 1/4).
Sources: See discussion in section 5.1.

[3.15], can be set to:

= e(r g )T .

[5.1]

From 1972 through to 2011 there was an average annual nominal interest rate of 8.8% on 30-year
mortgages, 10% on 4-year auto loans, and 13.7% on two-year personal loans, while the average annual
change in the personal consumption expenditure (PCE) price index over the same time period was
3.8%. The average credit-card interest rate between 1995 and 2011 was 14%. For comparison, 30year Treasury bonds had an average yield of 7.7% over the periods 19772001 and 20062011. The
implied real interest rates are 4.2% on Treasury bonds, 5% on mortgages, 6.2% on auto loans, 9.9%
on personal loans, and 12% on credit cards.26 The baseline real interest rate is set to the 5% rate on
mortgages as these constitute the bulk of household debt.27 The sensitivity analysis considers values
of r from 4% up to 7%. Over the period 19722011 used to calibrate the interest rate, the average
annual growth rate of real GDP per capita was 1.7%. Together with the baseline real interest rate
of 5%, this implies that 0.992 using [5.1]. Since many models used for monetary policy analysis
are typically calibrated assuming zero real trend growth, for comparison the sensitivity analysis also
considers values of g between 0% and 2%.
The relevant quantity variable for debt is the ratio of gross household debt to annual household
income, denoted by D.28 This corresponds to what is defined as the loans-to-GDP ratio l in the
26

Average PCE inflation over the periods 19772001 and 20062011 was 3.5%, and 2% over the period 19952011.
Mortgage debt was around 77% of all household debt on average during 20062010. The baseline real interest
rate is close to the conventional calibration used in many real business cycle models (King and Rebelo, 1999). The
mortgage rate implies a spread of 0.8% between the interest rates on loans to households and Treasury bonds of the
same maturity. C
urdia and Woodford (2009) consider a somewhat larger spread of 2% between interest rates for
borrowers and savers.
28
Given the heterogeneity between borrower and saver households, it would not make sense to net the financial
assets of savers against the liabilities of borrowers. However, it might be thought appropriate to net assets and
liabilities at the level of individual households, especially since a large fraction of household borrowing is to buy
27

39

model (the empirical debt ratio being based on the amount borrowed rather than the subsequent
value of loans at maturity) after adjusting for the length of one time period (T years), hence D = lT .
Using the expression for l in equation [3.16] and given a value of , the equation can be solved for
the implied value of the debt service ratio :
=

2(1 )D
.
T

[5.2]

Note that in the model, all GDP is consumed, so for consistency between the data and the models
prediction for the debt-to-GDP ratio, either the numerator of the ratio should be total gross debt
(not only household debt), or the denominator should be disposable personal income or private
consumption. Since the model is designed to represent household borrowing, and because the
implications of corporate and government debt may be different, the latter approach is taken.
In the U.S., like a number of other countries, the ratio of household debt to income has grown
significantly in recent decades. To focus on the implications of levels of debt recently experienced,
the model is calibrated to match average debt ratios during the five years from 2006 to 2010. The
sensitivity analysis considers a wide range of possible debt ratios from 0% up to 200%. Over the
period 20062010, the average ratio of gross household debt to disposable personal income was
approximately 124%, while the ratio of debt to private consumption was approximately 135%.
Taking the average of these numbers, the target chosen is a model-consistent debt-to-income ratio
of 130%, which implies using [5.2] a debt service ratio of 8.6%.29
In estimating the coefficient of relative risk aversion , one possibility would be to choose values
consistent with household portfolios of risky and safe assets. But since Mehra and Prescott (1985)
it has been known that matching the equity risk premium may require a risk aversion coefficient
above 30, while values in excess of 10 are considered by many to be highly implausible.30 Subsequent
assets (houses). If the assets held by households had the same non-contingent return as their debt liabilities then this
netting would be valid, but that is highly unlikely for assets such as housing, which experience significant fluctuations
in value. In the model, non-contingent debt is repaid only from future income, not from the sale of assets, so
the assumptions used in the calibration would be approximately correct if the value of the assets actually held by
households were positively correlated with the value of GDP and had a similar volatility. If asset returns were more
procyclical, the calibration would understate the problem of household leverage; if returns were less procyclical or
countercyclical, the calibration would overstate leverage.
29
Empirically, a direct measure of the ratio of household debt payments to disposable personal income is available,
though this is not directly comparable to the debt service ratio in the model. Between 2006 and 2010, the average
debt service ratio was 12.7%. This measure includes both interest and amortization. For conventional Tf -year fixedrate mortgages (where the borrower makes a sequence of equal repayments over the life of the loan) the share of
amortization in total repayments (over the life of the loan, or over all cohorts of borrowers at a point in time) is
approximately (1 er Tf )/(r Tf ), where r is the annual real interest rate. Taking r 5% and Tf = 30 years, this
formula implies that interest payments are approximately 48% of total debt service costs, yielding an estimate of the
interest-only debt-service ratio of around 6.1%. In the model, the debt service ratio is calculated only for borrowers,
not for all households as in the data, and is net of new borrowing (which is positive when there is positive income
growth). Using [3.15], the model implies the interest-only debt service ratio over all households (including the 50% of
+ ), which is comparable to the interest-only adjustment of
savers with a zero debt service ratio) is given by d/(1
the empirical measure. The baseline calibration yields a debt service ratio of approximately 6.5%, close to the 6.1%
in the data.
30
Large values of are also needed to generate a significant inflation risk premium. For illustration, suppose real
GDP follows a random walk, and the standard deviation of annual real GDP growth is set to 2.1% as found in the
data for the period 19722011. Under the flexible-price optimal policy of nominal GDP targeting, the inflation risk
premium would be approximately 0.044% (at an annual rate). For < 10, the inflation risk premium can be no

40

analysis of the equity risk premium puzzle has attempted to build models consistent with the large
risk premium but with much more modest degrees of risk aversion.31
Alternative approaches to estimating risk aversion have made use of laboratory experiments,
observed behaviour on game shows, and in a recent study, the choice of deductible for car insurance
policies (Cohen and Einav, 2007).32 The survey evidence presented by Barsky, Juster, Kimball and
Shapiro (1997) potentially provides a way to measure risk aversion over stakes that are large as a
fraction of lifetime income and wealth.33 The results suggest considerable risk aversion, but most
likely not in the high double-digit range for the majority of individuals. Overall, the weight of
evidence from studies suggests a coefficient of relative risk aversion above one, but not significantly
more than 10. A conservative baseline value of 5 is adopted, and the sensitivity analysis considers
values from zero up to 10.34
One approach to calibrating the discount factor elasticity parameter (from [3.2]) is to select a
value on the basis of its implications for the marginal propensity to consume from financial wealth.
Let m denote the increase in per-household (annual) consumption of savers from a marginal increase
in their financial wealth.35 It can be shown that m, , and are related as follows:
=

1 mT
.

[5.3]

Parker (1999) presents evidence to suggest that the marginal propensity to consume from wealth lies
between 4% and 5% (for a survey of the literature on wealth and consumption, see Poterba, 2000).
However, it is argued by Juster, Lupton, Smith and Stafford (2006) that the marginal propensity to
consume varies between different forms of wealth. They find the marginal propensity to consume
more than 0.44%, even though in this example inflation would be perfectly negatively correlated with real GDP
growth and have the same standard deviation (of 2.1%, not too far below the actual standard deviation of PCE
inflation of approximately 2.5% between 1972 and 2011).
31
For example, Bansal and Yaron (2004) assume a relative risk aversion coefficient of 10, while Barro (2006) chooses
a more conservative value of 4.
32
Converting the estimates of absolute risk aversion into coefficients of relative risk aversion (using average annual
after-tax income as a proxy for the relevant level of wealth) leads to a mean of 82 and a median of 0.4. The stakes
are relatively small and many individuals are not far from being risk neutral, though a minority are extremely risk
averse. As discussed in Cohen and Einav (2007), the estimated level of mean risk aversion is above that found in
other studies, which are generally consistent with single-digit coefficients of relative risk aversion.
33
Respondents to the U.S. Health and Retirement Study survey are asked a series of questions about whether they
would be willing to leave a job bringing in a secure income for another job with a chance of either a 50% increase
in income or a 50% fall. By asking a series of questions that vary the probabilities of these outcomes, the answers
can in principle be used to elicit risk preferences. One finding is that approximately 65% of individuals answers fall
in a category for which the theoretically consistent coefficient of relative risk aversion is at least 3.8. The arithmetic
mean coefficient is approximately 12, while the harmonic mean is 4.
34
The parameter is also related to the elasticity of intertemporal substitution 1 . Early estimates of intertemporal substitution suggested an elasticity somewhere between 1 and 2, such as those from the instrumental variables
method applied by Hansen and Singleton (1982). Those estimates have been criticized for bias due to time aggregation by Hall (1988), who finds elasticities as low as 0.1 and often insignificantly different from zero. Using cohort
data, Attanasio and Weber (1993) obtain values for the elasticity of intertemporal substitution in the range 0.70.8,
while Beaudry and van Wincoop (1996) find an elasticity close to one using a panel of data from U.S. states. Contrary
to these larger estimates, the survey evidence of Barsky, Juster, Kimball and Shapiro (1997) produces an estimate
of 0.18. An earlier version of the model presented here (Sheedy, 2013) has separate parameters for risk aversion and
intertemporal substitution, but quantitatively, the intertemporal substitution paramater is found to matter little for
the results. For this reason, the calibration of here focuses on its implications for risk aversion.
35
A simplifying feature of the model is that borrowers have the same marginal propensity to consume from financial
wealth as savers in the neighbourhood of the steady state.

41

is lowest for housing wealth and larger for financial wealth. Given the focus on financial wealth in
this paper, the baseline calibration assumes m 6%, which using [5.3] implies 0.993.36 The
sensitivity analysis considers marginal propensities to consume from 4% to 8%.37
The range of available evidence on the Frisch elasticity of labour supply is discussed by Hall
(2009), who concludes that a value of approximately 2/3 is reasonable. However, both real business
cycle and New Keynesian models have typically assumed Frisch elasticities significantly larger than
this, often as high as 4 (see, King and Rebelo, 1999, Rotemberg and Woodford, 1997). The baseline
calibration adopted here uses a Frisch elasticity of 2, and the sensitivity analysis considers a range
of values for from completely inelastic labour supply up to 4.38
The debt maturity parameter (which stands in for the parameter specifying the sequence
of coupon payments, given = /(1 + n
)) is set to match the average maturity of household debt
contracts. In the model, the average maturity of household debt is related to the duration of the
bond that is traded in incomplete financial markets. Formally, duration Tm refers to the average
of the maturities (in years) of each payment made by the bond weighted by its contribution to the
present value of the bond. Given the geometric sequence of nominal coupon payments parameterized
by , the bond duration (in steady state) is

X
T
`T `1
=
Tm =
,
`

(1
+
j)
1

1+
j
`=1

= (1 + j)1
using the steady-state bond price (present value of the coupon payments) Q
from [3.12].39 Let j denote the average annualized nominal interest rate on household debt, with
1 + j = ej T . In the optimal policy analysis, the steady-state rate of inflation will be zero (
= 0),
hence nominal GDP growth is n
= g, and so = /(1 + g). It follows that and can be
determined by:


T
jT
, and = eg T .
[5.4]
=e
1
Tm
Doepke and Schneider (2006) present evidence on the average duration of household nominal debt
liabilities. Their analysis takes account of refinancing and prepayment of loans. For the most recent
36
Together with the baseline calibration of , , and , the original patience parameters are b 1.006 and
s 1.012, and the implied value of is 1.009. Thus, the exogenous difference between the annual rates of time
preference of borrowers and savers is approximately 2.4%.
37
A potential alternative approach to calibrating is to use its implications for the persistence of shocks to the
wealth distribution. In the model, the impulse response of the debt-to-GDP gap is proportional to ` after ` time
periods P
have elapsed. The expected duration (in years) of the effects of shocks on the wealth distribution is thus

Td = T `=1 `(1)`1 = T /(1), which can be used to obtain given an estimate of Td . The baseline calibration
is equivalent to Td 36 years. The sensitivity analysis for the marginal propensity to consume implies a range of
values from 0.988 to 0.998, which is equivalent to considering values of Td from approximately 21 to 139 years.
38
Different Frisch elasticities for borrowers and savers are assumed to ensure that the wealth distribution has
no impact on the aggregate supply of labour. The required household-specific Frisch elasticities are b 1.6 for
borrowers and s 2.6 for savers.
39
Duration is equal to the percentage reduction in the real value of a nominal asset following a one percentage point
(annualized) permanent rise in inflation. Since = /(1 + n
), j = i, 1 + i = (1 + )(1 + g), 1 + n
= (1 +
)(1 + g),
and = (1 + g)/(1 + ), it follows that Tm = T /(1 ). A permanent rise in inflation by one percentage point at
an annualized rate is equivalent to increasing t by T in all time periods from some date onwards, and this reduces
the ex-post real return on nominal bonds by T /(1 ), confirming the interpretation of Tm .

42

year in their data (2004), the duration lies between 5 and 6 years, while the duration has not been
less than 4 years over the entire period covered by the study (19522004). This suggests a baseline
duration of Tm 5 years, which using [5.4] implies 0.967.40 The sensitivity analysis considers
the effects of having durations as short as one quarter (one-period debt), and longer durations up
to 10 years.41
There are two main strategies for calibrating the price elasticity of demand . The direct approach draws on studies estimating consumer responses to price differences within narrow consumption categories. A price elasticity of approximately three is typical of estimates at the retail level
(see, for example, Nevo, 2001), while estimates of consumer substitution across broad consumption categories suggest much lower price elasticities, typically lower than one (Blundell, Pashardes
and Weber, 1993). Indirect approaches estimate the price elasticity based on the implied markup
1/( 1), or as part of the estimation of a DSGE model. Rotemberg and Woodford (1997) estimate an elasticity of approximately 7.9 and point out this is consistent with the markups in the
range of 10%20%. Since it is the price elasticity of demand that directly matters for the welfare
consequences of inflation rather than its implications for markups as such, the direct approach is
preferred here and the baseline value of is set to 3. A range of values from the theoretical minimum
elasticity of 1 up to 10 is considered in the sensitivity analysis.
The production function is given in equation [3.6]. If e denotes the elasticity of aggregate output
with respect to hours then the elasticity of real marginal cost with respect to output can be
obtained from e using:
1e
.
e
A conventional value of e 2/3 is adopted for the baseline calibration (this would be the labour
share in a model with perfect competition), which implies 0.5. As discussed in Rotemberg
and Woodford (1999), there may be reasons to expect an elasticity of marginal cost with respect
to output higher than this (for example, if the elasticity of substitution between labour and other
factors is less than one), so the sensitivity analysis examines the effects of higher values of . An
important implication of is the strength of real rigidities (related to the term 1 + appearing
in the formula for in the Phillips curve [4.6a]), which are absent in the special case of a linear
production function ( = 0).42 The sensitivity analysis considers values of between 0 and 1.
In the model, is the probability of not changing price in a given time period. The probability
=

40
A conventional Tf -year fixed-rate mortgage has a duration of (er Tf 1 r Tf )/(r (er Tf 1)), which is approximately
11 years with Tf = 30 and r 5%. The calibrated duration may seem short given the high share of mortgage debt
in total household debt and the prevalence of 30-year fixed-rate mortgages, but refinancing shortens the duration of
debt. In the model, the frequency of refinancing is determined by 1 . The baseline calibration implies that 12.6%
of the total stock of debt is refinanced or newly issued each year.
41
The baseline value of is 0.971. The calibration method implicitly assumes is a structural parameter that
will remain invariant to changes in policy, including the change in the average rate of inflation. An alternative is to
assume is the structural parameter, in which case is calibrated by dividing from [5.4] by 1 + n
, where n
is the
average of actual nominal GDP growth. This method leads to 0.958, which is well within the range of values of
considered in the sensitivity analysis.
42
It is conventional to assume a source of real rigidities in New Keynesian models, though Bils, Klenow and Malin
(2012) present some critical evidence.

43

distribution of survival times for newly set prices is (1 )` , and hence the expected duration of a
P
`1
price spell Tp (in years) is Tp = T
= T /(1 ). With data on Tp , the parameter
`=1 `(1 )
can be inferred from:
T
[5.5]
=1 .
Tp
There is now an extensive literature measuring the frequency of price adjustment across a representative sample of goods. Using the dataset underlying the U.S. CPI index, Nakamura and Steinsson
(2008) find the median duration of a price spell is 79 months, excluding sales but including product
substitutions. Klenow and Malin (2010) survey a wide range of studies reporting median durations
in a range from 34 months to one year. The baseline duration is taken to be 8 months (Tp 8/12),
implying 0.625. The sensitivity analysis considers average durations from 3 to 15 months.43

5.2

Results

Consider an economy hit by an unexpected permanent fall in potential output. How should monetary
policy react? In the basic New Keynesian model with sticky prices but either complete financial
markets or a representative household, the optimal monetary policy response to a TFP shock is to
keep inflation on target and allow actual output to fall in line with the loss of potential output.
Using the baseline calibration from Table 1 and the solution to the optimal monetary policy problem,
t , and
Figure 1 shows the impulse responses of the debt-to-GDP gap dt , inflation t , the output gap Y
the bond yield jt under the optimal monetary policy and under a policy of strict inflation targeting
for the 30 years following a 10% fall in potential output.
With strict inflation targeting, the debt-to-GDP gap rises in line with the fall in output (10%)
because the denominator of the debt-to-GDP ratio falls, while the numerator is unchanged. The
effects of this shock on the wealth distribution and hence on consumption are long lasting (the
half-life of the debt-to-GDP impulse response is around 25 years). Under strict inflation targeting,
the output gap and bond yields are completely unchanged following the shock.
The optimal monetary policy response is in complete contrast to strict inflation targeting. Optimal policy allows inflation to rise, which stabilizes nominal GDP over time in spite of the fall in
real GDP. This helps to stabilize the debt-to-GDP ratio, moving the economy closer to the outcome
with complete financial markets where borrowers would be insured against the shock and the value
of debt liabilities would automatically move in line with income. The rise in the debt-to-GDP gap
is very small (around 1%) compared to strict inflation targeting (10%). The rise in inflation is
very persistent, lasting around two decades. The higher inflation called for is significant, but not
43

An alternative approach to calibrating the parameters , , and related to nominal and real rigidities would be
to choose values consistent with estimates of the slope of the Phillips curve. The recent literature on estimating the
New Keynesian Phillips curve studies the relationship t = Et t+1 + (1/)xt between inflation t and real marginal
cost xt , where the latter is proxied by the labour share. The baseline calibration implies 1/ 0.091. Gal and
Gertler (1999) present a range of estimates of 1/ lying between 0.02 and 0.04. Gal, Gertler and Lopez-Salido (2001)
estimate 1/ to be in the range 0.030.04, while Sbordone (2002) obtains an estimate of 0.055. The Phillips curve
implied by the baseline calibration is steeper than these estimates, but the sensitivity analysis for , , and does
allow for Phillips curve slopes in the range of econometric estimates. The maximum value of considered implies
1/ 0.021, the maximum value of implies 1/ 0.038, and the maximum value of implies 1/ 0.057.

44

Figure 1: Responses to a TFP shock, optimal monetary policy and strict inflation targeting
Debt-to-GDP gap

Inflation

10

Inflation target
Optimal policy

8
%

2
0

0
0

10

20

30

10

20

30

Bond yield

Output gap
0.2
1

0.15
%

% 0.1

0.5

0.05
0

0
0

10

20

30

10

20

30

Years

Years

Notes: The shock is an unexpected permanent TFP shock that reduces the natural level of output by 10%
relative to its trend. The debt-to-GDP gap and the output gap are reported as percentage deviations,
and inflation and bond yields are reported as annualized percentage rates. The parameters are set in
accordance with the baseline calibration from Table 1.

dramatic: for the first two years, around 23% higher, for the next decade around 12% higher,
and for the decade after that, around 01% higher. Inflation that is spread out over time is still
effective in reducing the debt-to-GDP ratio because debt liabilities have a long average maturity. It
is also significantly less costly in terms of relative-price distortions to have inflation spread out over
a longer time than the typical durations of stickiness of individual prices.
The rise in inflation does lead to a disturbance to the output gap for the first one or two years,
but this is short lived because the duration of the real effects of monetary policy through the
traditional price-stickiness channel is brief compared to the relevant time scale of decades for the
other variables. The effect is also quantitatively small because inflation is highly persistent, the
rise in expected inflation closely matching the rise in actual inflation, so the Phillips curve implies
little impact on the output gap. Finally, nominal bond yields show a persistent increase. It might
seem surprising that yields do not fall as monetary policy is loosened, but the bonds in question are
45

long-term bonds, and the effect on inflation expectations is dominant (there is a fall in real interest
rates because the rise in bond yields is less than what would be implied by higher expected inflation
alone).
The impulse response function of the debt-to-GDP gap under the optimal policy is proportional
to the impulse response under strict inflation targeting. Since the debt-to-GDP gap would be zero
if incomplete financial markets were the only concern of the policymaker, this provides a measure of
the weights attached by optimal policy to stabilizing the debt gap and to stabilizing inflation. The
response of the debt gap under the optimal policy is approximately 11.6% of the response under
strict inflation targeting, so the policy weight on debt gap stabilization is 88.4% and the policy
weight 1 on inflation stabilization is 11.6% (similarly, the inflation response under the optimal
policy is around 88.4% of what would keep the debt gap exactly at zero).
The baseline calibration implies that addressing the problem of incomplete financial markets
is quantitatively the main focus of optimal monetary policy rather than other objectives such as
inflation stabilization. What explains this, and how sensitive is this conclusion to the particular
calibration targets? Consider the exercise of varying each calibration target individually over the
ranges discussed in section 5.1, holding all other targets constant. For each new target, the implied
parameters are recalculated and the new policy weight is obtained. Figure 2 plots the values of
(the optimal policy weight on the debt-to-GDP gap) obtained for each target.
As can be seen in Figure 2, over the range of reasonable average real GDP growth rates there is
almost no effect on the optimal policy weight. The range of real interest rates is somewhat larger
(because there is less certainty about the appropriate real interest rate to assume for household
borrowing) but the optimal policy weight on incomplete financial markets changes little. Both
average real growth and average real interest rates affect the discount factor , which enters the
equations of the model in many places, but there is no intuitively obvious reason to expect it to
have a large impact on the relative benefits and costs of achieving the various objectives of policy.
The results are most sensitive to the calibration targets for the average debt-to-GDP ratio and
the coefficient of relative risk aversion. The average debt-to-GDP ratio proxies for the parameter ,
which is related to the difference in patience between borrowers and savers. It is not surprising that
an economy with less debt in relation to income has less of a concern with the incompleteness of
financial markets because it means the impact of shocks is felt more evenly by borrowers and savers.
In the limiting case of a representative-household economy, the average debt-to-GDP ratio tends to
zero, and the degree of completeness of financial markets becomes irrelevant. The debt gap receives
more than half the weight in the optimal policy as long as the calibration target for the average
debt-to-GDP ratio is not below 50%. It seems unlikely the U.S. would return to such low levels of
household debt in the foreseeable future if the levels of borrowing experienced since the 1990s do
indeed reflect the preferences and income profiles of borrowers and savers.
It is also not surprising that the results are sensitive to the coefficient of relative risk aversion.
Since the only use for complete financial markets in the model is risk sharing, if households were
risk neutral then there would be no loss from these markets being absent, as long as saving and
borrowing incomplete financial markets remained possible. The baseline coefficient of relative risk
46

Figure 2: Sensitivity analysis for optimal policy weight on incomplete financial markets

1
0.75
0.5
0.25

1
0.75
0.5
0.25
0
1
1.5
2
0.5
Real GDP growth rate (g , annual %)

5
6
7
Real interest rate (r , annual %)

1
0.75
0.5
0.25

1
0.75
0.5
0.25
0
50
150
200
100
Debt-to-GDP ratio (D, % of annual GDP)
1
0.75
0.5
0.25

6
8
0
2
4
Coefficient of relative risk aversion ()
1
0.75
0.5
0.25

4
5
6
7
8
0
1
2
3
4
Marginal propensity to consume (m, annual %) Frisch elasticity of labour supply ()
1
0.75
0.5
0.25

1
0.75
0.5
0.25
2
6
8
4
10
Duration of debt (Tm , years)

2
6
8
10
4
Price elasticity of demand ()
1
0.75
0.5
0.25

1
0.75
0.5
0.25
0
0.4
0.6
0.8
0.2
Marginal cost elasticity w.r.t. output ()

0.4
0.6
0.8
1.2
1
Duration of price stickiness (Tp , years)

Notes: The response of the debt gap under the optimal policy is 1 multiplied by its response under
strict inflation targeting. Each of the calibration targets in Table 1 is varied individually, holding all
others at their baseline values. The baseline value of is 0.884.

aversion is higher than the typical value of 2 found in many macroeconomic models (though that
number is usually relevant for intertemporal substitution in those models, not for attitudes to risk),
but it is low compared to the values often assumed in finance models that seek to match risk premia
47

(even the maximum value of 10 considered here would be insufficient to generate realistic risk premia
without adding other features to the model). The optimal policy weight on the debt gap exceeds
one half if the coefficient of relative risk aversion exceeds 1.3, so lower degrees of risk aversion do
not necessarily overturn the conclusions of this paper.
The next most important calibration target is the price elasticity of demand. A higher price
elasticity increases the welfare costs of inflation. Welfare ultimately depends on quantities not
prices, but the price elasticity determines how much quantities are distorted by dispersion of relative
prices. To reduce the optimal policy weight on the debt gap below one half it is necessary to
assume price elasticities in excess of 10. Such values would be outside the range typical in IO and
microeconomic studies of demand, with 10 itself being at the high end of the range of values used
in most macroeconomic models. The typical value of 6 often found in New Keynesian models only
reduces to approximately 71%.
The results are largely insensitive to the marginal propensity to consume from financial wealth,
which is used to determine the parameter in the specification of the endogenous discount factors.
Since this feature of the model was introduced only for technical reasons, it is reassuring that it does
not have a significant effect on the results within a wide range of reasonable parameter values. The
Frisch elasticity of labour supply has a fairly small but not insignificant effect on the results, with
the optimal policy weight on the debt gap increasing with the Frisch elasticity. A higher elasticity
increases the welfare costs of shocks to wealth distribution by distorting the labour supply decisions
of different households, as well as making it easier for monetary policy to influence the real value of
debt by changing the ex-ante real interest rate in addition to inflation. An elastic labour supply does
mean that inflation fluctuations lead to output gap fluctuations, which increases the importance of
targeting inflation, but the first two effects turn out to be more important quantitatively.
The results are somewhat more sensitive to the average duration of a price spell and the elasticity
of real marginal cost with respect to output. The first of these determines the importance of nominal
price rigidities. Greater nominal rigidity leads to more dispersion of relative prices from a given
amount of inflation, and thus reduces the optimal policy weight on the debt gap. A higher output
elasticity of marginal cost implies that the production function has greater curvature, so a given
dispersion of output levels across firms represents a more inefficient allocation of resources. However,
the range of reasonable values for the duration of price stickiness does not reduce below 65%, and
the range of marginal cost elasticities does not lead to any value below 80%.
Finally, there is the average duration of household debt, where the effects of this calibration
target are more subtle. It might be expected that the longer the maturity of household debt, the
higher is the optimal policy weight on the debt gap. This is because longer-term debt allows inflation
to be spread out further over time, reducing the welfare costs of the inflation, yet still having an
effect on the real value of debt. However, the sensitivity analysis shows the optimal policy weight is
a non-monotonic function of debt maturity: either very short-term or long-term debt maturities lead
to high values of , while debt of around 1.5 years maturity has the lowest value of (approximately
75%).
This puzzle is resolved by recalling there are two ways monetary policy can affect the real value
48

of debt: inflation to change the ex-post real return on nominal debt, and changes in the ex-ante real
interest rate (financial repression). As has been discussed, the first method is effective at a lower
cost for long debt maturities. When labour supply is inelastic, the second method is not available,
and the value of is then indeed a strictly increasing function of debt maturity (with the value of
falling to 15% for the shortest-maturity debt).
When the ex-ante real interest rate method is available, it is most effective (taking account of the
costs of using it in terms of inflation and output gap fluctuations) when debt maturities are short.
This is because monetary policy can only affect ex-ante real interest rates for a few years at most
(in line with reasonable calibrations of nominal and real rigidities). If debt is continually refinanced
each quarter or comprises floating-rate instruments, monetary policy has significant power to affect
its real value because nominal rigidities allow it to change the current real interest rate. However,
if fixed-rate debt is rarely refinanced then, holding inflation constant (that is, ignoring the first
method for affecting the real value of the debt), the real return is largely predetermined and thus
insensitive to current monetary policy. Therefore, intermediate debt maturities correspond to the
lowest optimal policy weights on stabilizing the debt gap because the maturity is too short for the
inflation method to be effective at low cost, but too long for the ex-ante real interest rate method
to work.

Conclusions

This paper has shown how a monetary policy of nominal GDP targeting facilitates efficient risk
sharing in incomplete financial markets where contracts are denominated in terms of money. In an
environment where risk derives from uncertainty about future real GDP, strict inflation targeting
would lead to a very uneven distribution of risk, with leveraged borrowers consumption highly
exposed to any unexpected change in their incomes when monetary policy prevents any adjustment
of the real value of their liabilities. Strict inflation targeting does provide savers with a risk-free real
return, but fundamentally, the economy lacks any technology that delivers risk-free real returns, so
the safety of savers portfolios is simply the flip-side of borrowers leverage and high levels of risk.
Absent any changes in the physical investment technology available to the economy, aggregate risk
cannot be annihilated, only redistributed.
That leaves the question of whether the distribution of risk is efficient. The combination of
incomplete markets and strict inflation targeting implies a particularly inefficient distribution of
risk when households are risk averse. If complete financial markets were available, borrowers would
issue state-contingent debt where the contractual repayment is lower in a recession and higher in
a boom. These securities would resemble equity shares in GDP, and they would have the effect
of reducing the leverage of borrowers and hence distributing risk more evenly. In the absence of
such financial markets, in particular because of the inability of households to sell such securities,
a monetary policy of nominal GDP targeting can effectively complete financial markets even when
only non-contingent nominal debt is available. Nominal GDP targeting operates by stabilizing the
debt-to-GDP ratio. With financial contracts specifying liabilities fixed in terms of money, a policy
49

that stabilizes the monetary value of real incomes ensures that borrowers are not forced to bear too
much aggregate risk, converting nominal debt into real equity.
While the model is far too simple to apply to the recent financial crises and deep recessions
experienced by a number of economies, one policy implication does resonate with the predicament of
several economies faced with high levels of debt combined with stagnant or falling GDPs. Nominal
GDP targeting is equivalent to a countercyclical price level, so the model suggests that higher
inflation can be optimal in recessions. In other words, while each of the stagnation and inflation
that make up the word stagflation is bad in itself, if stagnation cannot immediately be remedied,
some inflation might be a good idea to compensate for the inefficiency of incomplete financial
markets. And even if policymakers were reluctant to abandon inflation targeting, the model does
suggest that they have the strongest incentives to avoid deflation during recessions (a procyclical
price level). Deflation would raise the real value of debt, which combined with falling real incomes
would be the very opposite of the risk sharing stressed in this paper, and even worse than an
unchanging inflation rate.
It is important to stress that the policy implications of the model in recessions are matched by
equal and opposite prescriptions during an expansion. Thus, it is not just that optimal monetary
policy tolerates higher inflation in a recession it also requires lower inflation or even deflation
during a period of high growth. Pursuing higher inflation in recessions without following a symmetric
policy during an expansion is both inefficient and jeopardizes an environment of low inflation on
average. Therefore the model also argues that more should be done by central banks to take away
the punch bowl during a boom even were inflation to be stable.

References
Abel, A. B. (1990), Asset prices under habit formation and catching up with the Joneses,
American Economic Review, 80(2):3842 (May). 25
Akyol, A. (2004), Optimal monetary policy in an economy with incomplete markets and idiosyncratic risk, Journal of Monetary Economics, 51(6):12451269 (September). 6
Allen, F., Carletti, E. and Gale, D. (2011), Money, financial stability and efficiency,
Working paper, New York University. 6
s, J., Arce, O. J. and Thomas, C. (2010), Banking competition, collateral constraints
Andre
and optimal monetary policy, Working paper 1001, Bank of Spain. 6
Attanasio, O. and Davis, S. J. (1996), Relative wage movements and the distribution of
consumption, Journal of Political Economy, 104(6):12271262 (December). 22
Attanasio, O. P. and Weber, G. (1993), Consumption growth, the interest rate and aggregation, Review of Economic Studies, 60(3):631649 (July). 41
Bach, G. L. and Stephenson, J. B. (1974), Inflation and the redistribution of wealth, Review
of Economics and Statistics, 56(1):113 (February). 5

50

Bailey, S. (1837), Money and Its Vicissitudes in Value, Effingham Wilson, London. 24
Bansal, R. and Yaron, A. (2004), Risks for the long run: A potential resolution of asset pricing
puzzles, Journal of Finance, 59(4):14811509 (August). 41
Barro, R. J. (1979), On the determination of the public debt, Journal of Political Economy,
87(5):940971 (October). 16, 35
(2006), Rare disasters and asset markets in the twentieth century, Quarterly Journal of
Economics, 121(3):823866. 41
Barsky, R. B., Juster, T. F., Kimball, M. S. and Shapiro, M. D. (1997), Preference
parameters and behavioral heterogeneity: An experimental approach in the health and retirement
study, Quarterly Journal of Economics, 112(2):537579 (May). 23, 41
Bean, C. R. (1983), Targeting nominal income: An appraisal, Economic Journal, 93:806819
(December). 7
Beaudry, P. and van Wincoop, E. (1996), The intertemporal elasticity of substitution: An
exploration using a US panel of state data, Economica, 63:495512 (August). 41
Bernanke, B. and Gertler, M. (1989), Agency costs, net worth, and business fluctuations,
American Economic Review, 79(1):1431 (March). 1
Bernanke, B. S. and Gertler, M. (2001), Should central banks respond to movements in
asset prices?, American Economic Review, 91(2):253257 (May). 3
Bernanke, B. S., Gertler, M. and Gilchrist, S. (1999), The financial accelerator in a
quantitative business cycle framework, in J. B. Taylor and M. Woodford (eds.), Handbook of
Macroeconomics, vol. 1C, Elsevier, chap. 21, pp. 13411393. 1
Bernanke, B. S., Laubach, T., Mishkin, F. S. and Posen, A. S. (1999), Inflation Targeting:
Lessons from the International Experience, Princeton University Press. 1
Bils, M., Klenow, P. J. and Malin, B. A. (2012), Reset price inflation and the impact of
monetary policy shocks, American Economic Review, 102(6):27982825 (October). 43
Blundell, R., Pashardes, P. and Weber, G. (1993), What do we learn about consumer
demand patterns from micro data?, American Economic Review, 83(3):570597 (June). 43
Bohn, H. (1988), Why do we have nominal government debt?, Journal of Monetary Economics,
21(1):127140 (January). 6
Calvo, G. A. (1983), Staggered prices in a utility-maximizing framework, Journal of Monetary
Economics, 12(3):383398 (September). 27
Carlstrom, C. T., Fuerst, T. S. and Paustian, M. (2010), Optimal monetary policy in a
model with agency costs, Journal of Money, Credit and Banking, 42(s1):3770 (September). 1
Chari, V. V., Christiano, L. J. and Kehoe, P. J. (1991), Optimal fiscal and monetary policy:
Some recent results, Journal of Money, Credit and Banking, 23(3(2)):519539 (August). 6
Chari, V. V. and Kehoe, P. J. (1999), Optimal fiscal and monetary policy, in J. B. Taylor
and M. Woodford (eds.), Handbook of Monetary Economics, vol. 1C, Elsevier, chap. 26, pp.
16711745. 6
51

Christiano, L., Ilut, C. L., Motto, R. and Rostagno, M. (2010), Monetary policy and
stock market booms, Working paper 16402, NBER. 3
Christiano, L., Motto, R. and Rostagno, M. (2010), Financial factors in economic fluctuations, Working paper 1192, European Central Bank. 1
Clarida, R., Gal, J. and Gertler, M. (1999), The science of monetary policy: A New
Keynesian perspective, Journal of Economic Literature, 37(4):16611707 (December). 22
Cochrane, J. H. (1991), A simple test of consumption insurance, Journal of Political Economy,
99(5):957976 (October). 22
Cohen, A. and Einav, L. (2007), Estimating risk preferences from deductible choice, American
Economic Review, 97(3):745788 (June). 23, 41
Cole, H. L. and Kocherlakota, N. R. (2001), Efficient allocations with hidden income and
hidden storage, Review of Economic Studies, 68(3):523542 (July). 21
Cooley, T. F. and Ohanian, L. E. (1991), The cyclical behavior of prices, Journal of Monetary Economics, 28(1):2560 (August). 24
Cukierman, A., Lennan, K. and Papadia, F. (1985), Inflation-induced redistributions via
monetary assets in five European countries: 19741982, in J. Mortensen (ed.), Economic Policy
and National Accounting in Inflationary Conditions, vol. 2 of Studies in Banking and Finance,
North-Holland, Amsterdam. 5
rdia, V. and Woodford, M. (2009), Credit frictions and optimal monetary policy, Working
Cu
paper, Columbia University. 7, 39
De Fiore, F., Teles, P. and Tristani, O. (2011), Monetary policy and the financing of
firms, American Economic Journal: Macroeconomics, 3(4):112142 (October). 6
De Fiore, F. and Tristani, O. (2012), Optimal monetary policy in a model of the credit
channel, Economic Journal, forthcoming. 1
Doepke, M. and Schneider, M. (2006), Inflation and the redistribution of nominal wealth,
Journal of Political Economy, 114(6):10691097 (December). 5, 22, 42
Eggertsson, G. B. and Krugman, P. (2012), Debt, deleveraging, and the liquidity trap: A
Fisher-Minsky-Koo approach, Quarterly Journal of Economics, 127(3):14691513. 7
Erceg, C. J., Henderson, D. W. and Levin, A. T. (2000), Optimal monetary policy with
staggered wage and price contracts, Journal of Monetary Economics, 46(2):281313 (October).
27, 34
Faia, E. and Monacelli, T. (2007), Optimal interest rate rules, asset prices, and credit frictions, Journal of Economic Dynamics and Control, 31(10):32283254 (October). 1
Gal, J. and Gertler, M. (1999), Inflation dynamics: a structural econometric analysis,
Journal of Monetary Economics, 44(2):195222 (October). 44
pez-Salido, J. D. (2001), European inflation dynamics,
Gal, J., Gertler, M. and Lo
European Economic Review, 45(7):12371270 (June). 44

52


Garriga, C., Kydland, F. E. and Sustek,
R. (2013), Mortgages and monetary policy,
Discussion paper 2013-6, Centre for Macroeconomics. 26
Ghent, A. C. (2011), Securitization and mortgage renegotiation: Evidence from the Great Depression, Review of Financial Studies, 24(6):18141847. 22
Guerrieri, V. and Lorenzoni, G. (2011), Credit crises, precautionary saving, and the liquidity
trap, Working paper 17583, NBER. 7
Hall, R. E. (1988), Intertemporal substitution in consumption, Journal of Political Economy,
96(2):339357 (April). 41
(2009), Reconciling cyclical movements in the marginal value of time and the marginal
product of labor, Journal of Political Economy, 117(2):281323 (April). 42
Hall, R. E. and Mankiw, N. G. (1994), Nominal income targeting, in N. G. Mankiw (ed.),
Monetary Policy, University of Chicago Press, chap. 2, pp. 7194. 7
Hansen, L. P. and Singleton, K. J. (1982), Generalized instrumental variables estimation of
nonlinear rational expectations models, Econometrica, 50(5):12691286 (September). 41
Hayashi, F., Altonji, J. and Kotlikoff, L. (1996), Risk-sharing between and within families, Econometrica, 64(2):261294 (March). 22
Hoelle, M. and Peiris, M. U. (2013), On the efficiency of nominal GDP targeting in a large
open economy, Working paper 1273, Krannert School of Management, Purdue University. 6
Iacoviello, M. (2005), House prices, borrowing constraints, and monetary policy in the business
cycle, American Economic Review, 95(3):739764 (June). 7
Juster, F. T., Lupton, J. P., Smith, J. P. and Stafford, F. (2006), The decline in household saving and the wealth effect, Review of Economics and Statistics, 88(1):2027 (February).
41
Kehoe, T. J. and Levine, D. K. (1993), Debt-constrained asset markets, Review of Economic
Studies, 60(4):865888 (October). 21
King, R. G. and Rebelo, S. T. (1999), Resuscitating real business cycles, in J. B. Taylor and
M. Woodford (eds.), Handbook of Macroeconomics, vol. 1B, Elsevier, chap. 14, pp. 9271007. 39,
42
Kiyotaki, N. and Moore, J. (1997), Credit cycles, Journal of Political Economy, 105(2):211
248 (April). 1
Klenow, P. J. and Malin, B. A. (2010), Microeconomic evidence on price-setting, in B. M.
Friedman and M. Woodford (eds.), Handbook of Monetary Economics, vol. 3, Elsevier, chap. 6,
pp. 231284. 44
Koenig, E. F. (2013), Like a good neighbor: Monetary policy, financial stability, and the distribution of risk, International Journal of Central Banking, 9(2):5782 (June). 6
Kryvtsov, O., Shukayev, M. and Ueberfeldt, A. (2011), Optimal monetary policy under
incomplete markets and aggregate uncertainty: A long-run perspective, Journal of Economic
Dynamics and Control, 35(7):10451060 (July). 6
53

Lee, J. W. (2010), Monetary policy with heterogeneous households and imperfect risk-sharing,
Working paper, Rutgers University. 6
Lucas, R. E., Jr. and Stokey, N. L. (1983), Optimal fiscal and monetary policy in an economy
without capital, Journal of Monetary Economics, 12(1):5593. 16
Lustig, H., Sleet, C. and Yeltekin, S
. (2008), Fiscal hedging with nominal assets, Journal
of Monetary Economics, 55(4):710727 (May). 7
Meade, J. E. (1978), The meaning of internal balance, Economic Journal, 88:423435
(September). 7
Mehra, R. and Prescott, E. C. (1985), The equity premium: A puzzle, Journal of Monetary
Economics, 15(2):145161 (March). 40
Nakamura, E. and Steinsson, J. (2008), Five facts about prices: A reevaluation of menu cost
models, Quarterly Journal of Economics, 123(4):14151464 (November). 44
Nelson, J. A. (1994), On testing for full insurance using consumer expenditure survey data,
Journal of Political Economy, 102(2):384394 (April). 22
Nevo, A. (2001), Measuring market power in the ready-to-eat cereal industry, Econometrica,
69(2):307342 (March). 43
Osorio-Rodrguez, D. (2013), Foreign currency debt and optimal monetary policy: Is there a
role for the exchange rate in completing markets?, Working paper, London School of Economics.
6
Parker, J. A. (1999), Spendthrift in America? On two decades of decline in the U.S. saving
rate, NBER Macroeconomics Annual, 14:317370. 41
Persson, T. and Svensson, L. E. O. (1989), Exchange rate variability and asset trade,
Journal of Monetary Economics, 23(3):485509 (May). 6
Pescatori, A. (2007), Incomplete markets and households exposure to interest rate and inflation
risk: Implications for the monetary policy maker, Working paper 07-09, Federal Reserve Bank
of Cleveland. 6
Piskorski, T., Seru, A. and Vig, V. (2010), Securitization and distressed loan renegotiation:
Evidence from the subprime mortgage crisis, Journal of Financial Economics, 97(3):369397
(September). 22
Poterba, J. M. (2000), Stock market wealth and consumption, Journal of Economic Perspectives, 14(2):99118 (Spring). 41
Rotemberg, J. J. and Woodford, M. (1997), An optimization-based econometric framework
for the evaluation of monetary policy, NBER Macroeconomics Annual, 12:297346. 42, 43
(1999), The cyclical behavior of prices and costs, in J. B. Taylor and M. Woodford (eds.),
Handbook of Macroeconomics, vol. 1B, Elsevier, chap. 16, pp. 10511135. 43
Sbordone, A. M. (2002), Prices and unit labor costs: a new test of price stickiness, Journal of
Monetary Economics, 49(2):265292 (March). 44

54

, S. and Uribe, M. (2003), Closing small open economy models, Journal of

Schmitt-Grohe
International Economics, 61(1):163185 (October). 25
(2004), Optimal fiscal and monetary policy under sticky prices, Journal of Economic
Theory, 114(2):198230 (February). 6
Selgin, G. (1995), The productivity norm versus zero inflation in the history of economic
thought, History of Political Economy, 27(4):705735 (Winter). 6, 24
(1997), Less than zero: The case for a falling price level in a growing economy, Hobart
paper 132, Institute of Economic Affairs. 5
Sheedy, K. D. (2013), Debt and incomplete financial markets: A case for nominal GDP targeting, Discussion paper 1209, Centre for Economic Performance. 25, 41
Shiller, R. J. (1993), Macro Markets: Creating Institutions for Managing Societys Largest Economic Risks, Oxford University Press, New York. 21
(1997), Public resistance to indexation: A puzzle, Brookings Papers on Economic Activity,
1997(1):159228. 22, 24
Siu, H. E. (2004), Optimal fiscal and monetary policy with sticky prices, Journal of Monetary
Economics, 51(3):575607 (April). 6
Sumner, S. (2012), The case for nominal GDP targeting, Working paper, Mercatus Center,
George Mason University. 7
Uzawa, H. (1968), Time preference, the consumption function, and optimum asset holdings,
in J. N. Wolfe (ed.), Value, Capital and Growth: Papers in Honour of Sir John Hicks, Aldine,
Chicago. 25
Vlieghe, G. W. (2010), Imperfect credit markets: implications for monetary policy, Working
paper 385, Bank of England. 6
White, M. J. (2009a), Bankruptcy: Past puzzles, recent reforms, and the mortgage crisis,
American Law and Economics Review, 11(1):123. 22
White, W. R. (2009b), Should monetary policy lean or clean?, Working paper 34, Federal
Reserve Bank of Dallas, Globalization and Monetary Policy Institute. 3
Woodford, M. (2001), Fiscal requirements for price stability, Journal of Money, Credit and
Banking, 33(3):669728 (August). 26
(2003), Interest and Prices: Foundations of a Theory of Monetary Policy, Princeton University Press, New Jersey. 1, 27, 32
(2011), Inflation targeting and financial stability, Working paper, Columbia University. 3

Appendices

Appendices are available in the full version of the paper:

http://personal.lse.ac.uk/sheedy/papers/DebtIncompleteFinancialMarketsAppendix.pdf

55

Brookings Panel on Economic Activity

March 2021, 2014

The Wealthy Hand-to-Mouth

Greg Kaplan, Princeton University, NBER, and IFS
Giovanni L. Violante, New York University, CEPR, NBER, and IFS
Justin Weidner, Princeton University
Final conference draft

Abstract
The wealthy hand-to-mouth are households who hold little or no liquid wealth (cash, checking, and
savings accounts), despite owning sizable amounts of illiquid assets (assets that carry a transaction
cost, such as housing or retirement accounts). This portfolio configuration implies that these households
have a high marginal propensity to consume out of transitory income changesa key determinant of the
macroeconomic effects of fiscal policy.
The wealthy hand- to-mouth, therefore, behave in many respects like households with little or no net
worth, yet they escape standard definitions and empirical measurements based on the distribution of
net worth. We use survey data on household portfolios for the U.S., Canada, Australia, the U.K., Germany,
France, Italy, and Spain to document the share of such households across countries, their demographic
characteristics, the composition of their balance sheets, and the persistence of hand-to-mouth status over
the life cycle. Using PSID data, we estimate that the wealthy hand-to-mouth have a strong consumption
response to transitory income shocks. Finally, we discuss the implications of this group of consumers for
macroeconomic modeling and policy analysis.

The Wealthy Hand-to-Mouth *

Greg Kaplan
Princeton University, NBER, and IFS
gkaplan@princeton.edu
Giovanni L. Violante
New York University, CEPR, NBER, and IFS
glv2@nyu.edu
Justin Weidner
Princeton University
jweidner@princeton.edu

First version: January 2014 This version: March 2014

Abstract
The wealthy hand-to-mouth are households who hold little or no liquid wealth (cash, checking,
and savings accounts), despite owning sizable amounts of illiquid assets (assets that carry a
transaction cost, such as housing or retirement accounts). This portfolio configuration implies
that these households have a high marginal propensity to consume out of transitory income
changes a key determinant of the macroeconomic effects of fiscal policy. The wealthy handto-mouth, therefore, behave in many respects like households with little or no net worth, yet
they escape standard definitions and empirical measurements based on the distribution of
net worth. We use survey data on household portfolios for the U.S., Canada, Australia, the
U.K., Germany, France, Italy, and Spain to document the share of such households across
countries, their demographic characteristics, the composition of their balance sheets, and the
persistence of hand-to-mouth status over the life cycle. Using PSID data, we estimate that
the wealthy hand-to-mouth have a strong consumption response to transitory income shocks.
Finally, we discuss the implications of this group of consumers for macroeconomic modeling
and policy analysis.

Keywords: Consumption, Hand-to-Mouth, Household Balance Sheet, Liquidity.

JEL Classification: D31, D91, E21, H31
*

We thank Yu Zhang for outstanding research assistance, and Mark Aguiar, Karen Pence, David
Romer, and Justin Wolfers for comments. This research is supported by grant no. 1127632 from the
National Science Foundation.

Introduction

The life-cycle permanent income hypothesis (LC-PIH) is a valuable organizing framework for the analysis of both household survey and aggregate time-series data on the
joint dynamics of income and consumption. At the same time, economists have long
recognized that certain aspects of these data are at odds with some of the models most
salient predictions. This is true for both the standard version of the model (Friedman,
1957; Hall, 1978) and the more recent buffer-stock versions (Deaton, 1991; Carroll,
1997). In particular, both at the micro and macro level, it is common to estimate a
large sensitivity of consumption with respect to transitory changes in income, whereas
according to the theory these types of income dynamics should be smoothed.
At the microeconomic level, a large body of evidence finds that consumption overreacts
to predictable income growth (see Jappelli and Pistaferri, 2010, for a recent survey).
Some of the most convincing studies in this literature are based on quasi-natural experiments which measure the consumption response to the U.S. fiscal stimulus payment
episodes of 2001 and 2008.1 Johnson, Parker, and Souleles (2006), Parker, Souleles,
Johnson, and McLelland (2013), and Broda and Parker (2012) concluded that, in both
episodes, the consumption response was strong: between 20 and 40 percent of the stimulus payments were spent by households on nondurables in the quarter that they are
received. Shapiro and Slemrod (2003a, 2003b, 2009) substantiate these studies with
qualitative surveys on how consumers use their rebates and find comparable effects.
A number of additional studies based on micro data also find sizable consumption
responses to anticipated fluctuations in income. Examples are the reaction of consumption to changes in social security taxes (Parker, 1999), the analysis of federal
tax refunds (Souleles, 1999) and, most recently, the consumption response to a fiscal
stimulus episode in Singapore (Agarwal and Quian, 2013).
Similar anomalies in the joint behavior of income and consumption, relative to the
LC-PIH, have been identified from aggregate data. In a series of papers, Campbell
and Mankiw (1989, 1990, 1991) analyze U.S. and cross-country time series and show
that expected changes in aggregate consumption tend to be associated with expected
changes in aggregate income. Moreover, expected consumption growth is uncorrelated
with the real interest rate, a result that as long as the elasticity of intertemporal
1

In these two episodes, the U.S. Treasury selected the week of the payment based on the second-tolast digit of the taxpayers social security number, a digit that is effectively randomly assigned. This
randomization allows one to identify the causal effect of the fiscal transfer by simply comparing the
spending of households that received the rebate earlier with that of households who received it later.

substitution is not zero implies a break-down of the forward looking Euler equation
holding with equality.2
The most direct way to account for these facts is through the existence of a sizable
share of hand-to-mouth (HtM) consumers in the population consumers who spend all
of their available resources in every pay-period, and hence do not carry any wealth
across periods. HtM consumers have a high marginal propensity to consume (MPC)
out of transitory income changes that can account for the high correlation between consumption and the transitory component of income growth, even for anticipated income
shocks. Moreover, HtM consumers are not on their Euler equations, and thus they
are a source of misalignment between movements in the interest rate and movements
in aggregate consumption growth. The main challenge to this view is that standard
measurements using micro data on household balance sheets conclude that the fraction
of households with near zero net worth, and hence who consume all of their income
each period, is too small to quantitatively reproduce the facts discussed above.
Measuring HtM behavior using data on net worth is consistent with the vast majority
of heterogeneous-agent equilibrium macroeconomic models. These frameworks either
feature only one asset or feature two assets with different risk profiles, but with the
same degree of liquidity. Notable examples are the so called Bewley models with
uninsurable idiosyncratic risk and credit constraints, in the tradition of Huggett (1996),
Aiyagari (1994), Rios-Rull (1995), and Krusell and Smith (1998), and the so-called
spender-saver models in the tradition of Campbell and Mankiw (1989) that feature
impatient and patient consumers with complete markets. This latter class of models
has been recently revived by Gali, Lopez-Salido, and Valles (2007), Eggertson and
Krugman (2012), and Justiniano, Primiceri, and Tambalotti (2013), among others, to
analyze macroeconomic dynamics around the Great Recession. The models of Krusell
and Smith (1998) and Carroll, Slacalek, and Tokuoka (2014a, 2014b) combine the
spender-saver insight of heterogeneity in patience with an otherwise standard one-asset
incomplete-market model.
In this paper, we argue that measurements of HtM behavior inspired by this class
of models are misleading because they miss what we call the wealthy hand-to-mouth
households. These are households who hold sizable amounts of wealth in illiquid assets
(such as housing or retirement accounts), but very little or no liquid wealth, and
therefore consume all of their disposable income every period. Clearly, such households
would not be picked up by standard measurements since they own positive and often
2

Two related studies are Attanasio and Weber (1993) and Ludvigson and Michaelides (2001).

substantial amounts of net worth.

Thus to obtain a comprehensive measurement of HtM behavior using cross-sectional
survey data on household portfolios, a far better strategy is to use, as a guiding framework, a model with two assets one liquid asset and one illiquid asset that yields a
higher return, but that can only be accessed by paying a transaction cost. Recent examples of this two-asset environment are Angeletos et al. (2001), Laibson et al. (2003),
Chetty and Szeidl (2007), Alvarez, Guiso, and Lippi (2010), Huntley and Michelangeli
(2014), and Kaplan and Violante (2014a, 2014b). Through the lens of this two-asset
model, there are two types of HtM households. The poor hand-to-mouth (P-HtM) who
hold little or no liquid wealth and no illiquid wealth, and the wealthy hand-to-mouth
(W-HtM) who also hold little or no liquid wealth, but do have significant amounts of
illiquid assets on their balance sheet. Just like the P-HtM households, W-HtM households have large MPCs out of small transitory income fluctuations. However, as we
show in the paper, along many other important dimensions W-HtM households are
more similar to non HtM (N-HtM) households. As a result, the W-HtM cannot be
fully assimilated into either group, and are therefore best represented as a separate
class of households.
The goal of this paper is to investigate W-HtM behavior theoretically and empirically,
and to reflect on the implications of this peculiar, but sizable, group of households for
macroeconomic modeling and policy analysis.
First, we ask why households with significant wealth would optimally choose to consume all of their income every period, instead of using their wealth to smooth shocks.
To answer this question, in Section 2 we develop a stylized model based on Kaplan and
Violante (2014a). The model reveals that, under certain parameter configurations, a
portfolio composition with positive amounts of illiquid wealth and zero liquid wealth
is optimal. Such wealthy HtM households are better off bearing the welfare loss from
income fluctuations rather than smoothing consumption. The reason is that the latter
option requires holding large balances of cash and foregoing the high return on the illiquid asset (and, therefore, the associated higher level of long-run consumption). This
explanation is consistent with the calculations of Browning and Crossley (2001) who
showed that, in the context of a plausibly parameterized life-cycle buffer stock model,
the utility loss from setting consumption equal to income, instead of fully optimizing, is
second order. Similar calculations were performed by Cochrane (1989) and Krusell and
Smith (1998) in a representative agent environment. The model also provides useful
guidance for our empirical strategy. In Section 3 we outline this strategy in detail and
3

explain how we deal with a number of measurement issues.

Next, we ask how large the share of wealthy HtM households is in the population,
what their demographic characteristics are relative to poor and non HtM consumers,
and how their balance sheet compares with that of the N-HtM and P-HtM. Finally,
we investigate the persistence of HtM status over time to understand whether being
W-HtM is a transient or a persistent state during the life cycle of a household. This
empirical analysis is based on cross-sectional survey data on household portfolios for
a number of countries, specifically, the U.S., Canada, Australia, the U.K., Germany,
France, Italy and Spain. We describe these data in Section 4. When the literature
on household portfolios has examined these data in the past, its emphasis has been
on the allocation between risky and safe assets (see Guiso, Haliassos, and Jappelli,
2002, for a thorough cross-country comparison). Instead, our focus is on the liquidity
characteristics of the portfolio. In Section 5, we study U.S. data, for which we have
several repeated cross-sections between 1989 and 2010, as well as a two-year panel for
2007-2009. In Section 6, we present a comparative cross-country analysis with survey
data from around 2010.
The analysis of U.S. data leads to seven main findings: (i) between 25 and 40 percent
of U.S. households are HtM, with our preferred estimate at 1/3 of the population;
(ii) 1/3 of HtM households are poor HtM and 2/3 are wealthy HtM. Therefore, the
W-HtM represent the vast majority of this group; (iii) while households appear to be
most frequently P-HtM at young ages, the age profile of the W-HtM is hump-shaped
and peaks around age 40; (iv) the W-HtM typically hold sizable amounts of illiquid
wealth: for example, the median at age 40 is around $50,000; (v) W-HtM households
look a lot like the unconstrained N-HtM in terms of their age-profile of income, and in
terms of the shares of illiquid wealth held in housing and retirement account; (vi) the
W-HtM status is quite transient: we estimate that it lasts, on average, 2.5 years; (vii)
finally, estimates based on net worth miss at least half of HtM consumers.
Comparing the U.S. economy to the other countries we study, some interesting findings
emerge. In all of the eight countries, W-HtM households are a much greater share of the
population than P-HtM households, even more so than in the United States. However,
the total fraction of HtM households varies significantly across countries. As in the
U.S., it is over 30 percent in Canada, U.K., and Germany, but 20 percent or less of
the population in Australia, France, Italy, and Spain. We attribute the low number
of HtM households in Australia to the compulsory private retirement saving system
which means that even low-income young individuals have some illiquid wealth. For the
4

Euro-area countries, we observe that holdings of consumer debt are minimal suggesting
that the substantial liquid savings we observe, even among the income-poor, may act
as a buffer stock that substitutes for the expensive and limited access to credit.
In Section 7 we show that the HtM status of a household has strong predictive power
for the consumption response to transitory shocks. We apply the identification strategy
of Blundell, Pistaferri and Preston (2008) to panel data on income and consumption
from the U.S. to measure the MPC out of transitory income shocks for each type of
household. We find that W-HtM and P-HtM households have significantly stronger
responses than N-HtM households. In contrast, when we split households into HtM
groups based on net worth only, we do not find a significant difference in the consumption response between the two groups.
In Section 8, we argue that in the cast of characters of macroeconomic models, the WHtM deserve their own separate status. We use our empirical estimates of the fraction
of households in each of the three HtM groups, together with simulated MPCs from
three alternative models, to show that the W-HtM cannot be assimilated to either the
P-HtM or the N-HtM. We highlight three areas where models that feature only two
types of HtM households provide misguided intuition about the effects of fiscal stimulus
policy: the degree of cross-country dispersion in consumption responses, the degree of
non-linearity in consumption responses with respect to the size of the stimulus payment,
and the optimal degree of phasing of stimulus payments with income for maximizing
the aggregate consumption response. Section 9 summarizes and concludes the paper.

Wealthy hand-to-mouth behavior: a simple model

We start by analyzing a simple two-period model in order to illustrate the determinants of hand-to-mouth behavior. The model also offers some useful guidance for our
measurement exercise.
Consider a household that lives for two periods, t = 1, 2. Preferences over consumption
in the two periods are given by
v0 = u (c1 ) + u (c2 )
with no discounting. The household starts period 1 with an initial endowment and
a portfolio allocation decision. Two assets are available as saving instruments. First,
5

there is an illiquid asset a that pays off a gross return R before the consumption decision
in period 2, but cannot be accessed at the time of the consumption decision in period 1.
Second, there is a liquid asset m that can be accessed before the consumption decision
in both periods, but pays a return 1 < R. For now, we do not allow the agent to
borrow, i.e. take negative a position in the liquid asset, but we relax this assumption
in Section 2.4. After the initial portfolio allocation decision, households receive income
y1 and make their consumption and liquid saving decision in period 1. In the second,
and last, period, they receive income y2 and consume this endowment plus their savings
in liquid and illiquid wealth. The only two decisions to characterize are therefore the
initial portfolio allocation decision, and the consumption/saving decision at t = 1.
We make the following normalizations and parametric assumptions. Period utility
u is CRRA with intertemporal elasticity of substitution > 0. We set the initial
endowment to 1, so the initial portfolio allocation (m1 , a) has the interpretation of
shares of wealth invested in liquid and illiquid wealth. We set y2 = > 1 and we
allow two possible values for y1 , {yL , yH } where yL = 0 and yH > R + . We refer
to these two cases as low-income and high-income paths. The low income path is
increasing and the high income path is decreasing.
Our characterization of hand-to-mouth behavior concerns the asset position at the
time of the t = 1 consumption decision. We define a household as not hand-to-mouth
(N-HtM) if, after consuming at t = 1, it holds a positive amount of liquid assets, i.e.
m2 > 0 and a 0. We define a household as poor hand-to-mouth (P-HtM) if, after
consuming at t = 1, it does not hold any liquid or illiquid assets, i.e. m2 = 0 and
a = 0. We define a household as wealthy hand-to-mouth (W-HtM) if, after consuming
at t = 1, it holds a positive amount of illiquid assets but no liquid assets, i.e. m2 = 0
and a > 0.3 Therefore, the t = 1 consumption/saving decision determines whether an
agent is HtM, and the initial portfolio allocation determines whether a HtM agent is
poor or wealthy HtM.

2.1

Solution without illiquid asset

We begin by analyzing a special case where there is no illiquid asset. In this case we
refer to the liquid asset as net worth. We solve the model backwards, starting from the
3

The final case, m2 > 0 and a = 0, which is another form of N-HtM behavior, is never optimal
given the assumptions above, but could be easily accommodated.

consumption decision at t = 1. The problem faced by the household at t = 1 is

v1 (m1 ) = max u (c1 ) + u (m2 + )
c1 ,m2

s.t.
c1 + m2 = y1 + m1
m2 0
which has the solution


m2 = max


y1 + m1
,0 .
2

(1)

The interior solution for m2 implies a perfectly smooth consumption path, c1 = c2 =

(y1 + + m1 ) /2 because there is no discounting and the interest rate on the liquid
asset (the only saving vehicle available at t = 1) is 1. The corner solution m2 = 0
yields an increasing consumption path, c1 = y1 + m1 , c2 = . Since the liquid asset is
the only available asset, the initial portfolio allocation decision is trivial, and m1 = 1.
Thus there two cases, depending on the income path. Under the low income path
with yL = 0 < 1, equation (1) reveals that the constraint binds at t = 1 and the
household is P-HtM with an increasing consumption profile. Under the high income
path with yH > R + > 1, the constraint is not binding and the household is
N-HtM with a smooth consumption profile.

2.2

Solution with illiquid asset

We now turn to the general two-asset model. At t = 1 the consumption decision is

v1 (m1 , a) = max u (c1 ) + u (m2 + Ra + )
c1 ,m2

s.t.
c1 + m2 = y1 + m1
m2 0
which has the solution

m2 = max


y1 + m1 Ra
,0 .
2

(2)

The interior solution for m2 implies a smooth consumption path c1 = c2 = (y1 + + m1 + Ra) /2,
while the corner solution yields the consumption pair (c1 = y1 + m1 , c2 = + Ra).
7

Note that under the low income path yL = 0 < 1 m1 + Ra for any feasible
pair (a, m1 ). Therefore, equation (2) implies that the constraint will bind at t = 2,
regardless of the initial portfolio allocation, and m2 = 0. In this case, the household is
therefore HtM. Instead, under the high income path, yH > R + m1 + Ra for
any pair (a, m1 ). Hence equation (2) implies that the constraint will not bind at t = 2,
regardless of the initial portfolio allocation, and m2 > 0. In this case, the household is
N-HtM.
Next, consider the initial portfolio allocation decision. Under the high income path,
when the constraint is not binding, the problem is

v0 = max u
a,m1

y1 + + m1 + Ra
2

s.t.
1 = a + m1
It is immediate to see that the objective function is steeper in a than in m1 because
of the higher rate of return on the illiquid asset. Hence the household invests all of
its initial endowment in the illiquid asset and we have a corner solution with a = 1.
In this case, the household is N-HtM with a perfectly smooth consumption profile
c1 = c2 = (yH + + R) /2.
Under the low income path (y1 = yL = 0) the constraint binds at t = 1 and m2 = 0.
The problem becomes
v0 = max u (m1 ) + u (Ra + )
a,m1

s.t.
1 = a + m1
which has the solution

a = max




R
R+
, 0 , m1 = min
,1 .
R + R
R + R

(3)

Note that the portfolio allocation decision will always imply m1 > 0 since the household
needs liquidity at t = 1 for consumption. Thus, it only remains to determine when
a = 0 and when a > 0.
1

If 1 < R , equation (3) implies that a = 0 and the household is P-HtM. In this
8

case the return on the illiquid asset is not large enough for the household to tolerate
the large jump in consumption between t = 1 and t = 2 that would occur if it were to
save some of the initial endowment in illiquid wealth. Hence c1 = 1 and c2 = , and
1
therefore c2 = c1 . When R > , we instead have an interior solution for the portfolio
allocation, and the agent is W-HtM with consumption c1 = (R + ) / (R + R ) and
c2 = R c1 > c1 .
It is useful to explain the role of the three model parameters in determining wealthy
HtM behavior: the relative return on the illiquid asset R, income growth under the low
income path , and the elasticity of intertemporal substitution . A high relative return
R makes the illiquid asset more attractive and induces the agent to absorb consumption
swings across periods in order to achieve a higher overall level of consumption. Steep
income growth reduces the role of the illiquid asset as a saving instrument, since the
slope of the income profile guarantees high consumption later in life already. Finally,
the larger the elasticity of intertemporal substitution , the more the household is
willing to accept consumption fluctuations across periods, and the more likely it is to
become W-HtM rather than P-HtM.
Finally, note that since the model is deterministic, some households choose to invest in
the illiquid asset, thus diverting resources from liquid wealth, even though they know
with certainty that they will be constrained at a future point in time. Nevertheless,
they do find it optimal because the welfare gain from the rise in the overall level of
consumption more than compensates for the welfare loss from consumption fluctuations
between the two dates. In our two period model, we have abstracted from discounting.
In a multi-period model with geometric discounting, all the qualitative conclusions
remain intact, as we show in Kaplan and Violante (2014). Hyperbolic discounting
introduces an additional reason to save in illiquid assets, since illiquidity protects quasihyperbolic households from future consumption splurges (see Angeletos et al., 2001;
Laibson et al., 2003), and therefore makes it easier to generate wealthy HtM agents.

2.3

Implications for the MPC out of an unexpected income

transfer

Suppose that after the initial portfolio allocation decision, but before the consumption
decision at t = 1, the household receives an unexpected transfer from the government.
What is the households MPC out of this transfer?

A N-HtM household has an MPC of exactly 1/2, since it smooths the payment equally
across the two periods.
Next, consider the problem of a household who, in absence of the transfer would be
P-HtM, i.e., it faces y1 = yL = 0 and optimally chose the portfolio allocation (m1 =
1, a = 0):
v1 (1, 0) = max u (c1 ) + u (m2 + )
c1 ,m2

s.t.
c1 + m2 = + 1
m2 0
which has the solution


m2 = max


+1
,0 .
2

For any small payment 0 < < 1, this household remains P-HtM and has an MPC
of 1. Its consumption path is: c1 + 1 + , c2 = . If, instead, 1, the household
becomes unconstrained, consumption equals ( + + 1) /2 in both periods and its
MPC out of the transfer drops to ( + 1)/2 which approaches 1/2 as increases.
Finally, consider the problem of a household who, in absence of the transfer would be
W-HtM, i.e., it faces y1 = yL = 0 and optimally chose a = a = (R ) / (R + R ) >
0:
v1 (1 a , a ) = max u (c1 ) + u (m2 + Ra + )
c1 ,m2

s.t.
c1 + m2 = + (1 a )
m2 0
The solution to this problem is:

m2 = max


+ (1 a ) Ra
,0 .
2

This household has a MPC of 1 as long as 1 + (R + 1) a . This condition

is weaker than the condition for a P-HtM to have a MPC of 1 because the income
(and consumption) ratio between t=1 and t=1 is higher for a W-HtM compared to a

10

P-HtM.4
This finding suggests that, in a general model with more income heterogeneity (as
in Kaplan and Violante, 2014), the average MPC among W-HtM households is larger
than the average MPC among P-HtM households. We return to this point in Section 8.
Finally, note that all the results in this section carry over to the case of an anticipated
transfer, as long as the transfer is small enough that it does not change the HtM status
at t = 1.

2.4

Unsecured credit

We now extend the model and allow households to access credit to finance consumption
at t = 1. We assume that households can borrow up to a fraction 1 of their future
income and that the interest rate on borrowing is Rb > 1. Hence the credit limit is
m= /Rb . To make the exercise interesting, we impose the additional restriction that
Rb < , which ensures that a household with the low income path will always borrow
a positive amount. Indeed, the no-borrowing case studied above can be interpreted as
a model where borrowing is allowed but Rb , and credit is so expensive that no
household ever uses it. Since the role of the intertemporal elasticity of substitution is
well understood from the previous analysis, we impose = 1 (i.e., logarithmic utility)
to simplify the exposition.

2.4.1

Solution without illiquid asset

Under the high income path, the household is not constrained and chooses to save some
of its high income into the liquid asset at t = 1. Since the borrowing constraint is not
binding, the solution with borrowing is unchanged and m2 > 0.
Under the low income path, the problem is more interesting. In this case, m2 0 and
4

Put differently, the shadow value of an additional unit of income at t=1 is higher for the W-HtM
than for the P-HtM. If we let be the shadow value of a unit of income in period 1, for a P-HtM
agent we have = u0 (1 + ) u0 () and for a W-HtM agent we have = u0 ( + (1 a)) u0 (Ra + ),
which is larger.

11

at t = 1:
v1 = max log (c1 ) + log Rb m2 +

c1 ,m2

s.t.
c1 + m2 = 1
m2
which has the solution

Rb



Rb
m2 = max
.
,
2Rb
Rb

Since Rb < , the household always borrows a positive amount. Moreover, if Rb <
(1 2), then the credit limit is binding. The household is forced to choose an
increasing consumption path, c1 = 1 + /Rb , c2 = (1 ). If, instead, > Rb
(1 2), the solution for m2 is negative and interior: by borrowing, it can perfectly


smooth consumption at the level c1 = c2 = Rb + /2.
In light of the discussion in Section 2.3 about MPCs, only the household at the credit
limit has a MPC equal to 1, and only if the transfer is small enough not to change
its HtM status. For small transfers, a household with an interior negative position is
unconstrained and has a MPC equal to 1/2.

2.4.2

Solution with illiquid asset

Once again, under the high income path the household is not constrained at t = 1, so
allowing for borrowing has no effect on its decisions. Under the low-income path where
y1 = yL = 0, the household may want to borrow at t = 1. Its consumption decision at
t = 1 is:
v1 (m1 , a) = max log (c1 ) + log Rb m2 + Ra +
c1 ,m2

s.t.
c1 + m2 = m1

m2
Rb
which has the solution

m2 = max

Rb m1 Ra
,
2Rb
Rb
12

If Rb < , then for every feasible portfolio allocation (m1 , a), the first argument of the
max operator in the above equation is negative, and hence m2 < 0. The credit limit is
binding when Rb < [ (1 2) + Ra]/m1 , i.e., when borrowing is sufficiently cheap. In
this case, consumption is given by c1 = m1 + /Rb and c2 = (1 ) + Ra. When
borrowing is sufficiently expensive, i.e. Rb [ (1 2) + Ra]/m1 , the solution for
m2 is interior in the negative range and consumption is fully smoothed with c1 = c2 =


Rb m1 + + Ra /2.
We now analyze the portfolio decision at t = 0. Since we are interested in characterizing
HtM behavior, we focus on the case where the borrowing constraint binds, i.e. m2 =
/Rb . In this case, the portfolio problem is:
v0 = max u (m1 + /Rb ) + u (Ra + (1 ))
a,m1

s.t.
1 = a + m1
with solution

(
a = max

R+

)



1 + R/Rb 1
,0 .
2R

Using the restriction Rb < , it can be shown that a = 0 if R < Rb (1 ) / (1 + ) and

a is strictly positive if R > (1 ) / (1 + ). The former parameter configuration
corresponds to a P-HtM household who has borrowed up to the credit limit. The
latter parameter configuration corresponds to a household who chooses to save into
the illiquid asset and then borrows up to its credit limit. This is a W-HtM household
with negative liquid wealth (at the credit limit). Both households will have an MPC
of 1 out of a small transfer.

Identifying hand-to-mouth households in the data

The stylized model of Section 2 illustrates that there are two types of HtM households:
poor hand-to-mouth (P-HtM) who do not hold any illiquid wealth, and wealthy handto-mouth (W-HtM) who own positive amounts of illiquid wealth. For each type of HtM
household, there are two kinks in the intertemporal budget constraint where MPCs out
of small income changes can be large: the unsecured credit limit and zero liquid assets.
The unsecured credit limit is always a hard constraint. The zero liquid asset position
is a hard constraint for the subset of households who do not have access to credit,
13

and a kink for virtually all others, since the interest rates on credit cards and other
non-collateralized loans are typically much larger than the return on liquid assets.
Translating these theoretical definitions into empirical measurement poses a number
of practical challenges. In this section, we discuss the key problems that arise when
trying to identify poor and wealthy HtM households from cross-sectional survey data
on household portfolios, such as the data that we use in our empirical analysis and
that we describe in Section 4. Broadly speaking, these data contain household-level
information on income and on different categories of assets and liabilities that can be
aggregated into net liquid and illiquid wealth. We refer the reader to the next section
for the exact definition of liquid and illiquid wealth, country by country.
We now describe our strategy for identifying HtM households in the data separately
for households at the zero kink and for households at the credit limit. Let yit denote
the income of household i in period t, let mit denote holdings of liquid wealth and let
ait denote holdings of illiquid wealth. We always assume that income is paid as liquid
wealth.
First, consider a HtM household who is at the zero kink. In the model, this household
does not borrow or save into liquid assets: it consumes all its cash-on-hand for the
period and carries zero liquid wealth between t and t + 1. Suppose we have information on the length of the pay-period of the earners in the household, or equivalently
the frequency of pay (weekly, bi-weekly, monthly, etc...), and we observe the average
balance of liquid wealth over this period.
A conservative criterion to identify HtM agents on the zero kink in the data is to count
those households in the survey whose average balances of liquid wealth are positive
(to capture the fact they are not borrowing), but are equal to or less than half their
earnings per pay-period, where half presumes resources being consumed at a constant
rate. Specifically, a household is P-HtM at the zero kink if
ait = 0,

and

0 mit

yit
2

(4)

and

0 mit

yit
.
2

(5)

and W-HtM at the zero kink if

ait > 0,

Panel (a) of Figure 1 depicts a graphical representation of the dynamics of income

and average cash-on-hand m
it over a pay period for a HtM individual who starts and
14

Cash in hand

Cash in hand
m + yt

yt

t
m
t =

m
t =

yt
2

t+1

yt
2

(a) HtM at the zero kink

t+1

(b) HtM at the credit limit

Figure 1: Illustration of two cases of HtM behavior

ends the period at the zero kink. Its liquid balances (or cash in hand) peak at yit at
the beginning of the pay period, when income is paid into the liquid account, and are
depleted constantly until they reach zero at t+1. Average balances over the period are
therefore equal to half income.
This estimator provides a lower bound because, although all N-HtM households would
always hold average liquid balances above half their earnings, some HtM households
may also hold average liquid balances above half their earnings. For example, a household who starts the period with positive liquid savings in addition to earnings and ends
the period with zero liquid savings is HtM, but its average liquid balance is above half
earnings, and so it would not be counted as HtM by this criterion.
Unfortunately, most surveys do not record the frequency of pay for individual earners.
In the benchmark measurement exercise, we assume that labor income is received by
the household twice a month. In the sensitivity analysis, we report calculations under
the assumption that labor income is received weekly or monthly.
In certain surveys, instead of average balances, we observe wealth at a random point
in time during the pay-period. Our proposed measurement is still valid, as long as
the interview dates are uniformly distributed over the pay-period and are uncorrelated
with (yit , mit , ait ). In other words, this source of measurement error would tend to underestimate (overestimate) HtM behavior if the interview date is close to the beginning
(end) of the pay period, but the error averages to zero.
Next, consider a HtM household at the credit limit mit < 0. This is a household
who consumes all its cash-on-hand for the period as well as all its available credit. For
15

consistency with the strategy above, we propose to count a household as P-HtM at the
credit limit if
yit
ait = 0, mit 0 and mit
mit ,
(6)
2
and to count it as W-HtM at the credit limit if
ait > 0,

mit 0 and

mit

yit
mit .
2

(7)

Panel (b) of Figure 1 depicts a graphical representation of the dynamics of income and
average cash-on-hand mit over a pay period for a HtM individual who starts and ends
the period at the credit limit.
When, as in most surveys, individual credit limits are not reported, we assume that
the credit limit is a multiple of individual income and we experiment with a plausible
range of values. It is easy to see that this criterion is also conservative: a household
who starts the period at t with liquid wealth above its credit limit and ends the period
at t+1 having exhausted all its borrowing capacity, would carry an average balance
above the limit, and would therefore escape our criterion based on equations (6) and
(7).
We also compute the fraction of HtM agents in terms of net worth. Let nit = ait + mit
be net worth of agent i in period t. Then, a household is HtM in net worth (HtM-NW)
if

0 nit

yit
2

or,

nit 0 and

nit

yit
mit
2

(8)

A recent literature has emphasized the existence of pre-committed consumption expenditures expenditures that a household is committed to incur every pay-period, unless
it pays a transaction cost, either monetary or in terms of time, to modify its previous
commitments (see, for example, Chetty and Szeidl, 2007; Shore and Sinai, 2010;). Examples of such expenditures include rent, mortgages or other loan payments, utility
bills, school, gym, or club fees, and alimony. The key feature of committed expenditures is that they are bulk expenditures incurred at a point in time that discretely
deplete a households balance of liquid wealth. How does the presence of such expenditures affect our identification strategy? Let cit be the amount of committed
expenditures for household i at date t. If cit is incurred at the beginning of a pay
period, the criterion to identify a HtM household (say, at the zero kink) should be

16

amended as mit (yit cit ) /2, while if it occurs at the end of the period, the criterion
should be mit cit yit /2. In the first case, our baseline measurement overestimates
HtM status, and in the second case it underestimates it. Instead, if committed expenditures are incurred smoothly over the period or are paid in the middle of the pay
period, then the criterion should be mit cit /2 (yit cit ) /2 which is the same as
our baseline measurement. We verify the robustness of our estimates with respect to
those consumption commitments that we can measure in our survey data by using
these alternative assumptions about the timing of expenditures.
Finally, whenever the data allow, we also use direct survey questions as alternate
estimates of the fraction of HtM households. These questions typically ask (i) whether
expenditures over the last month have exceeded income, abstracting from purchases of
large durable goods such as housing or cars, and (ii) whether over the past month, the
household has saved a positive amount. Counts of HtM households derived from these
questions provide a useful check on the reliability of our identification strategy based
on reported liquid wealth and income.

Survey data on household portfolios

In this section we describe the cross-sectional survey data that we use to measure HtM
behavior. The countries that we study are the U.S., Canada, Australia, the U.K., and
the four largest economies in the Euro area: Germany, France, Italy, and Spain. In
order to categorize a household as W-HtM, P-HtM, or N-HtM, we need information
on labor income and on the amounts of assets and liabilities held in various categories
on the households balance sheet. We first provide background information on each
survey. Next, we describe how the definitions of the balance sheet items differ between
surveys. Finally, we present some descriptive statistics on the distribution of asset and
liabilities across countries. Appendix A contains information on how to access each of
the five data sets.

4.1

United States: SCF

Our data for the United States come from the Survey of Consumer Finances (SCF).
The SCF is sponsored by the Board of Governors of the Federal Reserve System in cooperation with the Statistics of Income Division of the Internal Revenue Service (IRS).

17

The survey has been conducted every three years and collects detailed information on
household balance sheets, income, and demographic characteristics for a representative
cross-section of U.S. households. We conduct analysis on the 1989 to 2010 surveys.5
While the surveys do not normally follow households over time, there is a panel component to the 2007 survey where a subset of households were contacted and re-surveyed
in 2009. See Bricker et. al (2011) for more information on the 2007-2009 panel of the
SCF.
The target population for the survey is all private households residing in the U.S. at
the time of data collection. The SCF uses a dual frame sample design. Households in
the first frame are intended to provide representative coverage of various characteristics
of households in the United States. Households in the second frame are drawn from
statistical records derived from tax information provided by the IRS and are intended
to disproportionately select relatively wealthy households. This oversampling design
allows the SCF to more accurately measure the distribution and composition of wealth
for the population as a whole, given the extreme right skewness in the distribution of
holdings for many asset classes.
The main interviewee is the household head. The head is defined as the core individual
in single households, the male in mixed-sex couples, and the older individual in samesex couples. In the case of couples, either member can be interviewed and the data
are rearranged after to define the household head in this way. Summary information
is then collected about all other household members. Labor market, pension, and
demographic data on the spouse or partner of the respondent are also collected. See
Kennickell (2005) for more information of the sample design of the SCF.

4.2

Canada: SFS

Our data for Canada come from the Survey of Financial Security (SFS). The SFS is
a cross-sectional survey implemented by Statistics Canada in 1999 and 2005, and is
intended to provide a comprehensive picture of net worth of Canadian households. In
our analysis, we use data from 2005. The survey asks questions on the value of all
major financial and non-financial assets and liabilities.
The surveyed households are a representative sample of all private households in Cana5

The survey started in 1983, but major technical revisions to the survey were implemented in 1989
and the structure and questions have largely been preserved since then. Since 1992, data have been
collected by the National Opinion Research Center at the University of Chicago.

18

dian provinces. Like the SCF, the SFS uses a dual frame sample design. The main
sample is a sample selected from the Labour Force Survey sampling frame. In order
to over-sample high income households, the second sample is drawn from geographic
areas in which there are a large proportion of family units with total income over a
certain threshold.
All individuals older than 15 years of age in the household are asked questions regarding
income, demographics, education, and employment. Questions regarding household
assets and liabilities are asked to the household member deemed most knowledgeable
on the subject. See Statistics Canada (2006) for more information about the 2005 SFS.

4.3

Australia: HILDA

Our data for Australia come from the Household, Income and Labour Dynamics in
Australia (HILDA) Survey. The Survey is managed by the Melbourne Institute of
Applied Economic and Social Research at the University of Melbourne. HILDA is a
broad social and economic longitudinal survey, with particular attention paid to family
and household formation, income and work. Wave One of the survey was implemented
in 2001, and households in the survey have since been interviewed annually.
The original sample for the HILDA survey was a large national probability sample
of Australian households occupying private dwellings. All members of the households
providing at least one interview in Wave 1 form the basis of the panel to be pursued
in each subsequent wave. The sample has been gradually extended to include any new
household members resulting from changes in the composition of the original household.
In addition to regular questions about economic and subjective well-being, the survey
features special modules covering specific topics. In particular, Waves Two (2002),
Six (2006), and Ten (2010) contain data from the wealth module that examines the
composition of households balance sheets.
Data for our analysis come from the Household Form and the Person Questionnaire.
The Household Form records basic information about the composition of the household. The Household Questionnaire is administered primarily to one member of the
household, and covers child-care, housing, household spending, and the wealth modules
in Waves Two, Six, and Ten. The Person Questionnaires are asked to all members of
the household aged 15 years and older, and collects information on family background,
education, employment, and income among other things. See Watson and Wooden
19

(2002) for more information on the HILDA.

4.4

United Kingdom: WAS

Our data for the United Kingdom come from the Wealth and Assets Survey (WAS).
The WAS is a longitudinal survey that is conducted by the Office of National Statistics
(ONS). The survey is intended to measure the economic well-being of households in the
U.K., by documenting the level of household savings and debt, lifecycle accumulation
of wealth, and participation in pension schemes.
For the first wave, the survey aimed to sample all persons living in private households
in Great Britain. The WAS also uses a dual frame design, using the first frame to meet
precision targets, and the second frame to over-sample the top wealth decile. The
sample for the first frame was drawn from the Royal Mails database of all addresses in
the UK. Households where at least one member was likely to have total financial wealth
above a certain threshold were flagged by Her Majestys Revenue and Customs. Flagged
households were sampled in such a way that they had two and a half times higher
probability of being sampled than non-flagged households. Wave One was conducted
from July 2006 to June 2008, and attempts were made to contact respondents for a
follow-up interview two years later for Wave Two. About two-thirds of cooperating
households completed the Wave Two interview from July 2008 to June 2010. In our
analysis, we use data from Wave Two.
The questionnaire is divided into two parts. The first part is the household questionnaire which is completed by one person in the household designated to be the household
reference person, and collects household-level information on household demographics,
as well as information about household assets and liabilities. The second part of the
questionnaire is an individual questionnaire administered to each adult aged 16 or over
in the household, and asks in-depth questions about economic status, education, employment, benefits, and individual financial assets. See Dafin (2009) and Black (2011)
for more information on the WAS.

4.5

Euro area: HFCS

Our data for Germany, France, Italy and Spain come from the Household Finance and
Consumption Survey (HFCS). The HFCS is a joint project administered by all of the

20

central banks of the Eurosystem and three National Statistical Institutes. The survey
provides detailed information on balance sheets, demographics, and other economic
variables for households in Euro area countries. Fieldwork in the various countries was
conducted between November 2008 and August 2011.
The HFCS is conducted and financed by each participating institution. For some
member countries, a previous wealth survey had already existed, and for others, an
entirely new survey had to be set up. The HFCS represents an effort towards gradual
harmonization of the content of the surveys across the member countries. The survey
will be conducted in each country every two to three years.
The core questionnaire, asked in every country, is composed of three parts. The first
comprises of questions regarding the household as a whole and contains questions regarding household assets and liabilities, transfers, and consumption-saving decisions.
This part is answered by one member of the household deemed to be the main respondent. The second part of the questionnaire is asked to all members of the household
and collects basic demographic information. The final part of the questionnaire is given
only to members of the household over 16 years of age and covers information regarding
employment, pension entitlements, and labor-market income.
There are also a set of standardized, non-core extension modules that the member
countries are allowed to include at their discretion in addition to the core questionnaire.
These non-core questions typically go into more detail on some aspect of the core
questionnaire that the member country wishes to explore. For example, Spain asks
questions that are designed to examine methods by which households pay their bills.
The target population for the survey is all private households and their current members
residing in the national territory at the time of data collection. The sampling design,
however, is chosen by each participating country. France uses a dual frame design,
exploiting individual data on taxable wealth to create the wealthy sample. The wealthy
sample is divided into four strata and sampled proportionally according to the relative
size of the strata. Germany uses regional level taxable income, and oversamples small
municipalities and, in larger municipalities, street sections with average income over a
threshold. Spain defines eight wealth strata, based on individual taxable wealth, that
are oversampled progressively at higher rates. Italy did not oversample in any way.
See Eurosystem Household Finance and Consumption Network (2013a) and (2013b)
for more information on the HFCS.

21

Survey
Years
Initial sample size
Not age 22-79
Negative income
All inc. from self empl.
Final sample size

U.S.
SCF
1989-2010
35513
2098
9
4334
29072

Canada
SFS
2005
5267
373
10

4884

Australia
HILDA
2010
7317
782
0
202
6333

U.K.
WAS
2008-10
18510
1655
0
334
18176

Germany
HFCS
2008-10
3565
246
0
228
3091

France
HFCS
2008-10
15006
1428
0
890
12688

Italy
HFCS
2008-10
7951
846
0
721
6384

Spain
HFCS
2008-10
6197
559
0
658
4980

Table 1: Summary information on the survey data used. Age selection for U.K. is
20-79 as age is provided in 5-year bins. Self employment income is not provided in the
SFS for Canada.

4.6

Sample selection and data comparability

Each individual survey is tailored to its own country, and as such, the questions asked
and the survey definition of particular asset classes will vary. Our main goal is to be
as consistent as possible in selecting the sample, and in defining income, liquid, and
illiquid wealth across surveys.
Sample selection.
In all surveys, we restrict our analysis to households in which
the head of is between 22 and 79 years of age, and drop households only if their
income is negative or if all of their income originates from self-employment.6 Table 1
summarizes the survey years we use for each country, the sample selection, and the
final sample sizes.
Income.
In choosing our definition of income, we make an attempt to include all
labor income plus government transfers that are regular inflows of liquid wealth. We
exclude interests, dividends, and other capital income because they are realized more
infrequently. Income in the SCF is gross wages and salaries, self-employment income,
unemployment benefits, workers compensation, regular private transfers such as child
support and alimony, regular public transfers such as food stamps and Social Security
Income (SSI), and regular income from other sources excluding investment income.
Income in the Canadian SFS is after-tax total income. There is no distinction between
labor, capital, and self-employment income. In HILDA, income is wages and salaries,
self-employment income, regular private transfers such as child support and alimony,
and public benefits such as the Australian Government Parenting Payment. For the
U.K. WAS, we define income as net employee earnings, net self-employment income,
plus any public benefits such as the Jobseekers Allowance and Maternity Allowance.
6

The only exception to our age range is for the U.K. WAS which provides ages in 5 year age bins,
so we include households with heads between 20 and 79 years of age.

22

Income in the HFCS is gross income from wages, salaries, and self-employment, unemployment benefits, regular private transfers such as child support and alimony, and
regular public transfers.7
The main discrepancy in income measurement across surveys is that in Canada income
is after taxes, whereas in all other countries the surveys ask for gross income before
taxes. For most households, except the self-employed, taxes are withheld at the source
and hence the amount paid into the liquid account and available for spending is net
of taxes. Thus, using income before taxes does somewhat overstate the fraction of HtM
households by inflating the liquid wealth threshold. Whenever possible, we verify the
robustness of our results to an adjustment for the individual tax liability.
Liquid wealth.
In the U.S. SCF, we consider liquid assets to be checking, saving,
money market and call accounts plus directly held mutual funds, stocks, corporate
bonds and government bonds. Liquid assets in the Canadian SFS are deposits in
financial institutions plus holdings in mutual funds, other investment funds, stocks
and bonds. In the HILDA, liquid assets include balances in bank accounts, equity
investments, and cash investments (bonds). In the U.K. WAS, liquid assets include
bank accounts, Individual Savings Accounts (ISAs), and holdings of shares, corporate
bonds, and government bonds.8 For the Euro area HFCS, liquid assets are cash, sight
(also called current, draft, or checking) accounts, mutual fund holdings, shares in
publicly traded companies, and corporate or government bond holdings.
The main shortcoming in the definition of liquid wealth is the absence of information
on holdings of cash. To address this problem, we resort to an imputation procedure.
We impute cash holdings for all surveys based on data from the Survey of Consumer
Payment Choice (SCPC) administered by the Federal Reserve Bank of Boston (see
Foster et al., 2011). We impute cash holdings by taking the ratio of average cash
holdings in the SCPC in 2010 to the median value of checking, saving, money market
7
The reference period for the income questions differs between surveys. For income variables in the
SCF, the survey asks for annual income in the previous year. For example, the 2010 SCF uses 2009 as
its reference period for income. The income reference period differs by country in the HFCS. France
and Germany both use 2009 as a reference period, Spain uses 2007, and Italy uses 2010. Wave Two of
the WAS (2008-2010) asks questions regarding the usual amounts for monthly income and benefits.
The 2005 SFS uses 2004 as its reference period, and gave its respondents the option of skipping the
income questions and using linked data from their 2004 tax return. Wave Ten of the HILDA uses the
2009-2010 financial year for its reference period for income which runs from July 1, 2009 to June 30,
2010.
8
ISAs are accounts designed for the purpose of saving with a favorable tax status. A broad range
of asset categories, including cash, can be held in ISAs. There are no restrictions to how much and
when funds can be withdrawn.

23

and call accounts from the 2010 SCF. We then inflate the value of each households
checking, saving, money market and call accounts by this ratio in all surveys9 .
We define liquid debt in the SCF as the sum of all credit card balances that, after
the most recent payment, accrue interest.10 Liquid debt in the SFS is credit card and
installment debt. Liquid debt in the HILDA is credit card debt. In the U.K. WAS,
liquid debt is credit card debt, plus any balances on store cards, hire purchases, and
mail orders. In the HFCS, liquid debts are considered to be the balance on credit cards,
after the most recent payment, which accrue interest, and any balances on credit lines
or bank overdrafts which also accrue interest.
The measure of liquid wealth that we use to compute HtM status is net liquid wealth,
or liquid assets minus liquid debt. We also examine a narrower definition of net liquid
wealth that excludes directly held mutual funds, stocks and bonds from liquid assets,
and a broader one that includes outstanding debt in home-equity lines of credit as
liquid debt.
Illiquid wealth.
Net illiquid wealth in the SCF includes the value of housing,
residential and non-residential real estate net of mortgages and home equity loans, private retirement accounts (such as 401(k)s, IRAs, thrift accounts, and future pensions),
cash value of life insurance policies, certificate of deposits, and saving bonds. Illiquid
wealth in the Canadian SFS is the value of the principal residence and other real estate
investment less mortgages on the properties and lines of credit using the property as
collateral. It also includes retirement savings such as Registered Retirement Savings
Plans, Registered Retirement Income Funds, employer pension plans, and other retirement funds. In the HILDA, illiquid wealth is net equity in home and other real-estate
properties plus life insurance policies and superannuation (government-supported, compulsory private retirement funds).11 In the U.K. WAS, we take illiquid wealth to include the value of the main residence, other houses, and land less mortgage and land
9
Average cash holdings, excluding large-value holdings in 2010 was $138. Median checking, saving,
money market and call accounts in the 2010 SCF is $2500, making the ratio about 5.5%. In the HFCS,
information on cash holdings is available for Spain from a non-core module. We check the median
ratio of cash to checking, saving, money market and call accounts and find it to be about 6% in Spain.
10
As most credit cards in the U.S. feature a one month grace period on purchases which makes them
a close substitute for cash in the very short term, we remain conservative and restrict our measure of
credit card debt only to debt for those households that do not regularly pay off their balances in full
each month. A specific question in the SCF allows us to identify such households.
11
Superannuation has some features of private retirement accounts, such as 401(k) accounts in the
U.S., which we include into illiquid wealth, and some features of public pensions (e.g., the compulsory
nature of a minimum contribution) which we exclude from illiquid wealth. As a result, we also offer
a sensitivity analysis where we exclude superannuation wealth from illiquid assets.

24

debt, plus occupational and personal pensions, insurance products, and National Savings products. The definition of net illiquid wealth in the HFCS is the value of the
household main residence and other properties net of mortgages and unsecured loans
specifically taken out to purchase the home, plus occupational and voluntary pension
plans, cash value of life insurance policies, certificate of deposits, and saving bonds.
We also explore broader definitions of illiquid wealth that include the value of businesses
for the self-employed, the resale value of vehicles net of the loans taken up to purchase
them, and other non-financial wealth not included in our baseline, such as antiques,
artwork, jewels, gold, etc.12 Note that changing the definition of illiquid wealth only
affects the split between poor and wealthy HtM, but not the total number of HtM
households.
The reference period for the wealth questions varies across surveys. In the SCF, for
most assets it is the interview date, but for some, such as checking and saving accounts,
when the respondent was unsure, the interview could prompt for an average balance
over the month. The SFS asks for information on assets and debts for a time as close
as possible to the date of the interview. Both the WAS and HILDA ask for current
balances or values of assets and liabilities. In the HCFS, France, Germany, and Spain
use the date of the interview, and Italy uses December 31, 2010.

4.7

Descriptive statistics

Table 2 reports some basic descriptive statistics on household income, liquid and illiquid
wealth holdings, and portfolio composition, for each country in the sample.
In all countries, the typical household portfolio structure is rather simple. It comprises
a small amount of liquid wealth in the form of bank accounts, some housing equity,
and a private retirement account. In particular, the median holdings of other financial
assets such as directly held stocks, bonds, mutual funds, and life insurance are zero
everywhere. This is a well known fact in the empirical study of household portfolios
(see Guiso, Halassios, and Jappelli, 2002). There are, however, some interesting crosscountry differences. With respect to net liquid wealth, consumer credit appears a lot
less frequent in the Euro area: less than 10 percent of households have credit card debt
in France, Italy, and Spain, compared to 30 to 40 percent in the Anglo-Saxon countries.
Figure 2, which plots the distribution of net liquid wealth to monthly income for the
12

In our robustness with respect to business equity we include all households whose income is entirely
from self-employment as long as they had non-negative income from their business.

25

26

Median
35444
46798
1319
1154
0
0
0
39306
0
0
0

Table 2: Data for the U.S. are from the 2010 survey only. All figures are in local currency units. Data for Canada is adjusted
to 2010 CA$ using the Canadian CPI. From the Federal Reserve Boards G.5 release, the average exchange rates in the survey
years are 1.2 CA$, 1.1 AU$, 0.6 British pounds, and 0.7 euros per U.S. dollar.

Income (age 22-59)

Net Worth
Net liquid wealth
Cash, checking, saving, MM accounts
Directly held stocks
Directly held bonds
Revolving credit card debt
Net illiquid wealth
Housing net of mortgages
Retirement accounts
Life insurance

Income (age 22-59)

Net Worth
Net liquid wealth
Cash, checking, saving, MM accounts
Directly held stocks
Directly held bonds
Revolving credit card debt
Net illiquid wealth
Housing net of mortgages
Retirement accounts
Life insurance

Median
47040
56944
1950
2640
0
0
0
52000
29000
1508
0

US
CA
AU
UK
Frac. Pos. Median Frac. Pos. Median Frac. Pos. Median Frac. Pos.
0.984
49905
1.000
79555
0.993
29340
0.979
0.887
113565
0.897
380889
0.984
187157
0.880
0.767
3194
0.769
12551
0.884
2111
0.632
0.923
2873
0.864
8709
0.978
2639
0.766
0.142
0
0.109
0
0.351
0
0.160
0.014
0
0.106
0
0.015
0
0.154
0.297
0
0.232
0
0.257
0
0.405
0.761
100713
0.752
347500
0.939
174999
0.843
0.629
64238
0.648
250000
0.714
81400
0.677
0.526
871
0.518
61000
0.863
58560
0.766
0.186
0
0.033
0
0.064
0
0.110
DE
FR
IT
ES
Frac. Pos. Median Frac. Pos. Median Frac. Pos. Median Frac. Pos.
0.994
31518
0.999
26116
0.987
26961
0.991
0.949
108976
0.966
165420
0.919
178925
0.967
0.853
1453
0.925
5226
0.769
2685
0.890
0.876
1255
0.953
4181
0.769
2261
0.908
0.110
0
0.151
0
0.043
0
0.106
0.050
0
0.015
0
0.146
0
0.014
0.225
0
0.076
0
0.049
0
0.086
0.876
104214
0.922
148524
0.803
171161
0.885
0.476
86372
0.607
148524
0.716
162491
0.847
0.245
0
0.039
0
0.088
0
0.037
0.493
0
0.378
0
0.193
0
0.245

eight countries, reinforces this observation.

Housing equity forms the majority of illiquid wealth for households in every country,
with the exception of Germany where median housing wealth is zero, since only 48
percent of the population are homeowners. This homeownership rate is at least 10
percentage points less than in all other countries (see also Eymann and Borsch-Supan,
2002). The median value of housing equity relative to median annual income is especially remarkable in Italy and Spain, where this ratio exceeds six. There are also
large differences in the fraction of households with positive private retirement wealth:
in the Anglo-Saxon countries, at least half of all households hold a personal retirement account, whereas in France, Italy and Spain less than one in ten do. Surely, a
big part of the explanation is in the generosity of the PAYG pension system in these
countries: according to the OECD, replacement rates for the median earner are between 60 and 70 percent in these countries, compared to 40 percent in the U.K. and
the United States. The size of private retirement wealth in Australia and the U.K.
is astonishing. In Australia, this is partly due to the superannuation regulations
that require all employers to generously contribute to tax-deferred retirement accounts
on behalf of their employees.13 In the U.K., the Pension Schemes Act of 1993 created
tax-free employer-sponsored (defined benefits) occupational pensions and (defined contributions) personal pensions. The Pension Act of 2008 established that workers must
choose to opt out of an occupational pension plan of their employer, rather than opt in
(see Banks and Tanner, 2002, for more details). Finally, the proportion of households
with life insurance in their portfolio is a lot higher in the Euro area than in the AngloSaxon countries. We conjecture that solid intergenerational family ties, and a stronger
precautionary savings motive linked to the lower female participation rate may account
for these differences.

United States

In this section, we report the main findings for the United States, using data from the
1989-2010 waves of the SCF. We begin by estimating the fraction of HtM households
and assessing the robustness of our estimates to a variety of aspects of the definition
adopted in Section 3. We then analyze the key demographic characteristics of N-HtM,
P-HtM and W-HtM households, and we examine their portfolio composition in more
13

In the survey years, the compulsory minimum employer contribution rate was 9 percent of the
employee salary.

27

.3
0

.05

Fraction of households
.1
.15
.2

.25

.3
.25
Fraction of households
.1
.15
.2
.05
0
10

5
0
5
Net liquid wealth to monthly labor income ratio

10

10

10

.25
Fraction of households
.1
.15
.2
.05
0

.05

Fraction of households
.1
.15
.2

.25

.3

(b) Canada

.3

(a) United States

5
0
5
Net liquid wealth to monthly labor income ratio

10

5
0
5
Net liquid wealth to monthly labor income ratio

10

10

10

.25
Fraction of households
.1
.15
.2
.05
0

.05

Fraction of households
.1
.15
.2

.25

.3

(d) United Kingdom

.3

(c) Australia

5
0
5
Net liquid wealth to monthly labor income ratio

10

5
0
5
Net liquid wealth to monthly labor income ratio

10

10

10

.25
Fraction of households
.1
.15
.2
.05
0

.05

Fraction of households
.1
.15
.2

.25

.3

(f) France

.3

(e) Germany

5
0
5
Net liquid wealth to monthly labor income ratio

10

5
0
5
Net liquid wealth to monthly labor income ratio

10

10

(g) Italy

5
0
5
Net liquid wealth to monthly labor income ratio

10

(h) Spain

Figure 2: Distribution of liquid wealth to monthly income ratios by country.

28

detail. Lastly, we exploit the longitudinal dimension of the 2007-2009 waves of the SCF
to study the persistence of HtM status over the life cycle.

5.1

The share of HtM households

Panel (a) of Figure 3 plots the fraction of HtM households in the U.S. population over
the period 1989-2010 and their split between wealthy and poor HtM. Recall that our
benchmark definition sets the pay frequency to two weeks, and uses equations (4) to
(8) with the credit limit mit set to one month of income.
Our estimates indicate that, on average, 31% of U.S. households are HtM over this
period. Of these, roughly 1/3 are poor HtM and 2/3 are wealthy HtM. This is the first
main result of our paper: the vast majority of hand-to-mouth households are not poor,
but rather own illiquid assets.14
Looking at changes over time across the two decades our data cover, the fraction of
HtM households remains fairly stable and the split between poor and wealthy does not
change significantly. Panel (b) plots the share of W-HtM households who each own
housing, retirement wealth, or both. About a half of W-HtM have both, about a third
have positive housing and no retirement wealth, and a sixth have positive retirement
wealth and no housing.
The first line of Table 3 also reports the share of U.S. households that are HtM when
using net worth as an index of wealth. We find that less than 14% of households are
HtM in terms of net worth, and thus looking at the wealth distribution through the
eyes of net worth misses over half of the HtM households in the United States.

5.1.1

Robustness

Figure 4 and Table 3 summarize our sensitivity analysis.

Panel (a) of Figure 4 provides an estimate of HtM behavior using a combination of
sequential questions in the SCF aimed at assessing whether over the past year, [household] spending exceeded, or was about the same as, income, and such expenditures did
included purchases of a home or automobile or spending for any investments.15 Based
14

A small number of HtM households (0.5 percent of the population over the whole sample period)
have negative illiquid wealth because they have negative equity on their house. Currently, they are
included among the P-HtM.
15
These questions (X7510, X7509, X7508) were included in the survey starting from 1992.

29

.5

.5

Other illiquid but no housing wealth

Only housing wealth
Both other and housing wealth

.3
.2
.1
0

.1

.2

.3

.4

PHtM

.4

WHtM

1989

1992

1995

1998

2001

2004

2007

1989

2010

(a) Share of total, wealthy, and poor HtM

1992

1995

1998

2001

2004

2007

2010

(b) W-HtM by portfolio composition

Figure 3: Time-series of fraction of HtM households in the U.S.

on this definition, the share of HtM households is around 40-45 percent. W-HtM households account for 3/4 of the total, and fluctuations in this measure over time follow
very closely those in the baseline definition of Figure 4(a). The third row of Table 3
also reports results for another sequence of direct questions in the SCF. The first question asks households Which of the following statements comes closest to describing
your saving habits? We label a household as HtM if it responds Dont save - usually
spend more than (or as much as) income. Just under 25 percent of households are
HTM according to this definition.
It is very reassuring that the baseline count of HtM households sits in between the
counts based on these two direct questions. Our baseline calculations refer to the
current HtM status for the households. The first set of direct questions asks about the
past year, so if there were periods when the household spent more than its income,
even if now it is no longer HtM, it would answer that question positively. Conversely,
the second set of direct questions asks about the usual HtM status, and therefore if the
current HtM status is perceived as transitory, the household would answer negatively.
In panel (b) of Figure 4, we verify the robustness of our estimates with respect to
the tightness of the credit limit. When we use the self-reported credit limit in the
SCF, the fraction of HtM households drops by 5 percentage point, with all the drop
being accounted for by a lower number of W-HtM households. Panel (c) plots HtM
shares when the pay-period is set to a month instead of two weeks. The fraction of
HtM households increases by 9 percentage points and W-HtM account for most of the
difference with the baseline. Symmetrically, the sixth line of Table 3 shows that when
the pay-period is set to one week, the share of W-HtM drops by 5 percentage points.
30

.5

.5

.3
.2
.1
0

.1

.2

.3

.4

PHtM

.4

WHtM

1989

1992

1995

1998

2001

2004

2007

2010

1989

1995

1998

2001

2004

2007

2010

.4
.3
.2
.1
0

.1

.2

.3

.4

.5

(b) Reported credit limit

.5

(a) HtM (past year c > y)

1992

1989

1992

1995

1998

2001

2004

2007

2010

1989

(c) Pay-period of 1 month

1992

1995

1998

2001

2004

2007

2010

(d) Vehicles in illiquid wealth

Figure 4: Time series of fraction of HtM households in the U.S., alternate definitions.

Panel (d) in Figure 3 shows that by including vehicles as illiquid wealth we move
roughly half of the P-HtM into the W-HtM group, but, by construction, the total
share of HtM households in the population is unchanged. Table 3 shows that using
a higher illiquid wealth threshold in the definition of W-HtM ($1,000 instead of $1)
has no impact on our findings. Including business equity, or directly held stocks and
bonds, or other valuables (artwork, antiques, jewels, etc.) among illiquid assets has
small effects relative to the baseline.
Table 3 also contains other sensitivity analyses. Changing the definition of liquid
debt by including used up HELOCs while simultaneously increasing the credit limit
by the total available line of credit increases the fraction of HtM households by 1
percentage point. The SCF collects data on a households normal, or usual, income as
well as on their actual income. This alternate definition of income has no effect on our
calculations. Recall that our definition of income is gross income before taxes and tax
credits. Through the NBER TAXSIM, we have constructed, household by household,
31

a measure of after tax income.16 As expected, under this income measure, the total
fraction of HtM households declines, but quantitatively this effect is very small.
As explained in Section 3, accounting for committed expenditures has an ambiguous
effect on the share of HtM agents, depending on whether the expenditures occur mostly
at the beginning or at the end of the pay-period. Table 3 shows that these two opposite timing assumptions bound the share of total HtM households between 27 and 42
percent.
To summarize, we estimate that, depending on the exact definition, between 20 and
40 percent of U.S. households are HtM, but most importantly around 2/3 of them
are wealthy HtM, i.e. they own positive illiquid wealth in spite of their low, or even
negative, holdings of liquid wealth. In contrast, estimates of HtM status in terms of net
worth vary between 6 and 17 percent, and therefore miss more than half of the HtM
households in the United States. Overall, these findings are consistent with related,
but different, measurements by Lusardi, Schneider and Tufano (2012), who document
that nearly one half of U.S. households would probably be unable to come up with
$2,000 in 30 days.

5.2

The demographics of HtM groups

We now turn to the demographic characteristics of the three groups of HtM households.
For this analysis, we pool the 1989 to 2010 waves of the SCF in order to maximize
sample size, and we use the baseline definition from Figure 3.
Figure 5 plots the share of the population that is W-HtM and P-HtM by age.17 Not
surprisingly, the bulk of P-HtM behavior is observed in the early stages of the life-cycle.
The fraction of P-HtM households drops sharply until age 30, and keeps falling steadily
over the life cycle until reaching roughly 5 percent in retirement. The age profile of the
fraction of W-HtM households is instead markedly hump shaped: it peaks at around
age 40 when over 20 percent of U.S. households are W-HtM, and remains above 10
percent throughout the lifecycle. The share in the residual group of N-HtM individuals
16

The variables we used in TAXSIM are year, marital status, the number of children, and the
breakdown of income into its parts (wages, UI benefits, etc.). We deducted from gross income federal
income taxes. We assumed each household files their actual marital status and claims all their children
as dependent. As an upper bound, we have also computed the case where they all file as single without
dependents.
17
These plots are based on pooled data from all surveys and do not control for time or cohort effects.
We verified that age profiles are similar in both cases, but become more somewhat noisy.

32

P-HtM
Baseline
0.121
In past year, c > y
0.130
0.089
Usually, c > y
Reported credit limit
0.115
1 year income credit limit
0.104
Weekly pay period
0.106
Monthly pay period
0.142
Vehicles as illiquid assets
0.052
Higher illiquid wealth cutoff
0.121
0.114
Businesses as illiquid assets
Direct as illiquid assets
0.120
Other valuables as illiquid assets 0.117
Includes usually paid off cc debt
0.121
0.120
HELOCs as liquid debt
Usual income
0.120
Disposable income - Reported
0.121
0.120
Disposable income - Single
Comm. cons. - beg. of period
0.101
Comm. cons. - end of period
0.149

W-HtM
0.188
0.309
0.156
0.147
0.120
0.147
0.258
0.257
0.188
0.189
0.212
0.192
0.192
0.178
0.194
0.184
0.183
0.162
0.269

N-HtM
0.691
0.561
0.756
0.738
0.776
0.748
0.600
0.691
0.691
0.697
0.668
0.691
0.688
0.702
0.687
0.695
0.698
0.737
0.581

HtM HtM-NW
0.309
0.137
0.439

0.244

0.262
0.127
0.224
0.109
0.252
0.118
0.400
0.164
0.309
0.061
0.309
0.137
0.303
0.129
0.332
0.137
0.309
0.132
0.312
0.137
0.298
0.135
0.313
0.136
0.305
0.137
0.302
0.135
0.263
0.115
0.419
0.173

Table 3: Robustness results for fraction HtM in each category in the SCF pooled
1989-2010. Higher illiquid wealth cutoff requires households to have above $1,000 in
illiquid assets to be considered W-HtM. Businesses as illiquid assets drops the self
employment income sample selection and adds business assets to illiquid wealth and
self employment income to income. Direct as illiquid assets classifies directly held
mutual funds, stocks, corporate and government bonds as illiquid assets. Disposable
income subtracts federal income taxes estimated from NBERs TAXSIM from income.
Disposable income - Reported assumes that each household files their actual marital
status and number of children as dependents. Disposable income - Single assumes that
every household files as single with no dependents. Comm. cons. - beg. of period
assumes the households committed consumption is incurred at the beginning of the
period. Comm. cons. - end of period assumes the household incurs it at the end of the
period.
increases steadily from 50 percent at age 22 to 80 percent in retirement.
The first three panels of Figure 6 report some demographic characteristics of the three
HtM groups by age. N-HtM households have on average one year of education more
than the W-HtM who, in turn, have one more year of education than the P-HtM. In
terms of marital status, H-HtM and W-HtM households are indistinguishable, whereas
the figure shows that the P-HtM households are 30 percent less likely to be married.
In contrast, P-HtM and W-HtM are both more likely to have children than are N-HtM

33

.4

PHtM

.1

.2

.3

WHtM

20

40

60

80

Age

Figure 5: Age profile of fraction of HtM households in the U.S., pooled 1989-2010.

households.
Figure 6(d) shows that P-HtM households are income-poor, with median annual income
around $20,000 (in $2010) during the working years, while the N-HtM are high-income
households who earn on average $70,000 at their life-cycle peak. The most surprising
finding is that the W-HtM look a lot like the N-HtM in terms of their income path. The
same conclusion holds for the incidence of unemployment and for the likelihood of receiving welfare benefits, which are both much lower for N-HtM and W-HtM households
than for the P-HtM.

5.3

The portfolio composition of HtM groups

Figure 7 digs deeper into the composition of the balance sheets of the three groups of
HtM households. Panel (a) shows that median liquid wealth holdings are zero for the PHtM and, perhaps unexpectedly, slightly negative for the W-HtM. N-HtM households
have substantial holdings of liquid wealth, peaking at around $15,000 before retirement.
Panel (b) reveals that the W-HtM hold significant amounts of illiquid wealth: for
example, median holdings at age 40 exceed $50,000. Hence, W-HtM households are
not just P-HtM households with small amounts of savings in less liquid assets.
The next two panels of Figure 7 articulate this observation further. Panels (c) and (d)
plot age profiles of the average fraction of illiquid wealth held in housing and retirement
accounts for W-HtM and N-HtM households. The conclusion is striking: the lines are
34

.8

16

PHtM

.2

10

.4

12

.6

14

WHtM
NHtM

20

40

60

80

20

40

Age

60

80

Age

(b) Fraction married

.5

20000

40000

1.5

60000

80000

(a) Average years of education

20

40

60

80

20

40

Age

60

80

Age

(d) Median income

.2

.1

.4

.6

.2

.8

.3

(c) Average number of children

20

40

60

80

20

Age

40

60

80

Age

(e) Frac. with at least one unemp. member

(f) Frac. of income from gov. benefits

Figure 6: Age profile of demographic characteristics of the HtM in the U.S., pooled
1989-2010.

35

300000

25000

PHtM

5000

100000

10000

15000

200000

20000

WHtM
NHtM

20

40

60

80

20

40

Age

60

80

Age

.8
.6
.4
.2
0

.2

.4

.6

.8

(b) Median net illiquid wealth

(a) Median net liquid wealth

20

40

60

80

20

Age

(c) Mean frac. of ill. wealth in housing

40

60

80

Age

(d) Mean frac. of ill. wealth in ret. account

Figure 7: Age profile of the portfolio composition of the HtM in the U.S., pooled 19892010. To reduce the sensitivity to outliers, means are computed after trimming the
overall top and bottom 0.1 percent of that statistics distribution.

on top of each other, indicating that the portfolio allocation of these two groups is
nearly identical. The key difference is that N-HtM are income and wealth richer, and
that they hold a lot more liquid wealth relative to their income.

5.4

Persistence of HTM status

How persistent is the HtM status of a household? The 2007-2009 panel of the SCF
allows us to address this question. Table 4 reports the 2-year transition matrix across
the three HtM statuses for U.S. households. P-HtM households have a 52% chance of
still being P-HtM two years later, but also a 37% chance of becoming N-HtM households
within the same period. The W-HtM status appears to be the most transitory: the
probability of remaining W-HtM two years later is approximately 17%. To put it
36

07 09
P
P
0.520
W
0.058
0.028
N

W
0.114
0.169
0.052

N
0.366
0.773
0.92

Table 4: Transition matrix for the 2007-2009 panel of the SCF. Fraction of households
classified as the row status in 2007 and the column status in 2009.
differently, the expected length of W-HtM status is around 29 months. Finally, the
N-HtM state is the least transient of the three, and it is almost an absorbing state.

Cross-country evidence

The previous section showed that around 30% of households in the United States are
HtM, with 1/3 of which are P-HtM and 2/3 of which are W-HtM. In this section we use
household portfolio data from seven other developed economies to assess whether the
prevalence of W-HtM households is a common feature of the wealth distribution across
countries and, if so, whether the characteristics of W-HtM in terms of demographics,
income, and balance sheets are similar to those in the U.S.
As discussed in Section 4, we focus our attention on three other Anglo-Saxon countries
Canada, Australia, and the U.K. and the four most populated European countries,
Germany, France, Italy, and Spain. While data is available for more than one point
in time for most of these countries including panel data for Australia and the U.K.
in order to keep the discussion manageable we focus on the most recent single crosssection in each country. For Australia and the European countries this is 2010, for the
U.K. it is 2009, and for Canada it is 2005. For the sake of comparability, we use only
the 2010 wave of the SCF for the United States.
Figure 8(a) shows the fraction of poor and wealthy HtM households in each country.
There is a striking similarity in the overall fraction of HtM households, as well as in
their breakdown between poor and wealthy, between the U.S., Canada, and the U.K.
These three countries have a large share of HtM households, exceeding 30 percent.
Australia is an outlier among the Anglo-Saxon countries in two ways: first, the total
fraction of HtM is roughly half the fraction in the U.S., the U.K., and Canada; second,
90 percent of HtM households in Australia are W-HtM. Among the European countries,
France, Italy, and Spain have smaller shares of HtM households than the U.S., U.K.,
and Canada around 20 percent whereas in Germany this share is closer to 30 percent.
37

.5

.5

Other illiquid but no housing wealth

Only housing wealth
Both other and housing wealth

.3
.2
.1
0

.1

.2

.3

.4

PHtM

.4

WHtM

US

CA

AU

UK

DE

FR

IT

US

ES

(a) Share of total, wealthy, and poor HtM

CA

AU

UK

DE

FR

IT

ES

(b) W-HtM by portfolio composition

Figure 8: Fraction of HtM households across countries

Even for the Euro area countries, the fraction of W-HtM among the HtM households
exceeds 2/3. For all eight countries, Figure 8(a) shows there are W-HtM households
than P-HtM. Thus a wide-spread feature of international household portfolios is that
a complete characterization of the fraction of the population that is likely to exhibit
HtM behavior requires going beyond those with just low net worth.
Figure 8(b) reveals that there are significant differences in the portfolio composition for
the W-HtM across countries. In Italy and Spain, virtually all the W-HtM own some
housing wealth. Homeowners are also dominant among the group of W-HtM in the
U.S. and Canada. In contrast, around half of the W-HtM in Australia, Germany, and
Canada have no housing wealth. Rather, the majority of their illiquid assets are held
in private retirement accounts.
What explains the fact that Australia and the Euro area countries, on the other, have
a smaller fraction of HtM households than in the U.S.? Figure 8(a), shows that this
difference is largely accounted for by differences in the fraction of P-HtM households.
Table 5 reveals that for Australia this discrepancy can be traced to the very high share
of the population that owns private retirement wealth. Over 86% of N-HtM households
in Australia have assets in a retirement account compared with 62% in the U.S. As
explained in Section 4, the high ownership rate rate of retirement accounts in Australia
is largely due to the superannuation regulations. When we exclude superannuation
accounts as a component of wealth, the fraction of P-HtM in Australia rises from 3 to
9 percent, closing more than half of the gap with the United States.

38

39

US
0.003
-1.172
0.362
0.000
0.839
0.610
0.782
0.034
0.316
0.533

Median liquid wealth / income

Mean liquid wealth / income
Frac. neg. liquid wealth
Frac. neg. illiquid wealth

Median liquid wealth / income

Mean liquid wealth / income
Frac. neg. liquid wealth
Frac. neg. illiquid wealth
Median housing / illiquid wealth
Mean housing / illiquid wealth
Frac. pos. housing equity
Median retire / illiquid wealth
Mean retire / illiquid wealth
Frac. pos. retirement account
UK
4.690
38.504
0.149
0.004
0.432
0.459
0.785
0.545
0.524
0.849

UK
0.000
-1.902
0.464
0.000
0.312
0.398
0.667
0.659
0.583
0.862

UK
0.000
-1.332
0.289
0.037

DE
1.133
6.144
0.081
0.006
0.534
0.446
0.533
0.000
0.052
0.259

DE
0.059
-0.019
0.082
0.000
0.000
0.369
0.464
0.000
0.093
0.271

DE
0.009
-0.012
0.051
0.039

FR
0.878
4.730
0.042
0.007
0.839
0.604
0.654
0.029
0.081
0.045

FR
0.122
-0.206
0.105
0.000
0.389
0.471
0.505
0.066
0.175
0.020

FR
0.077
-0.092
0.072
0.071

IT
3.701
8.918
0.016
0.001
1.000
0.965
0.770
0.036
0.174
0.107

IT
0.000
-0.067
0.054
0.000
1.000
0.968
0.834
0.044
0.390
0.041

IT
0.000
-0.002
0.007
0.001

ES
2.129
6.977
0.017
0.003
1.000
0.948
0.877
0.021
0.083
0.042

ES
0.011
-0.266
0.090
0.000
1.000
0.985
0.933
0.023
0.036
0.021

ES
0.000
-0.228
0.102
0.164

Table 5: Portfolio characteristics by country and HtM status. To reduce the sensitivity outliers, means are computed after
trimming the overall top and bottom 0.1 percent of that statistics distribution.

US
Median liquid wealth / income
1.737
Mean liquid wealth / income
15.144
Frac. neg. liquid wealth
0.102
Frac. neg. illiquid wealth
0.025
Median housing / illiquid wealth 0.651
Mean housing / illiquid wealth
0.593
Frac. pos. housing equity
0.714
Median retire / illiquid wealth
0.237
Mean retire / illiquid wealth
0.329
Frac. pos. retirement account
0.620

US
0.000
-0.631
0.167
0.140

P-HtM
CA
AU
0.000 0.013
-0.486 -0.399
0.180 0.122
0.114 0.043
W-HtM
CA
AU
0.000 0.021
-0.893 -0.783
0.369 0.305
0.000 0.000
0.733 0.377
0.587 0.406
0.800 0.547
0.253 0.563
0.403 0.583
0.655 0.924
N-HtM
CA
AU
2.324 3.404
12.440 34.667
0.086 0.061
0.035 0.004
0.774 0.709
0.647 0.598
0.722 0.771
0.204 0.273
0.343 0.387
0.570 0.878

In the Euro area countries, the lower fraction of P-HtM implies that households hold
more liquid wealth relative to their income. Table 5 shows that this difference can
be in part attributed to differences in liquid debt, a fact that we also highlighted
in the discussion of Figure 2. The fraction of P-HtM households in the Euro area
countries with negative liquid wealth is 2 to 4 times smaller than in the Anglo-Saxon
countries. Presumably, lower access to unsecured credit in Europe means that there are
more incentives for households to hold liquid wealth for transaction and precautionary
reasons. For example, Vandone (2009) documents that, in 2006, the total value of
consumer credit amounted to 25 percent of disposable income in the U.K., 15 percent
in Germany and Spain, 12 percent in France, and only 10 percent in Italy.

6.1

Robustness

Table 6 contains an extensive sensitivity analysis on our definitions of P-HtM and

W-HtM households that parallels that displayed in Table 3.
Questions on whether household spending exceeded income in the past year are present
in all surveys. Like in the U.S., we find larger shares of both P-HtM and W-HtM households when we use these direct questions to measure the incidence of HtM behavior.
The difference is especially marked for Italy and Spain where, according to this criterion, over 60 percent of households and hence three times the baseline estimate
are HtM. Extending the credit limit from one month of income to one year of income
has a substantial effect for the Anglo-Saxon countries, but virtually no impact for the
Euro area countries. This finding is consistent with the empirical distribution of liquid
assets documented in Figure 2, which showed that households with negative net liquid
wealth are extremely rare in the Euro area countries.18
Shortening the pay-period to a week and extending it to a month, from the bi-weekly
baseline, has a very small impact on the fraction of P-HtM households, but decreases
(increases, respectively) the fraction of W-HtM households by 5 percentage points on
average. Including vehicles as illiquid wealth shifts HtM households from poor to
wealthy in every country, but to a lesser extent than in the United States. In two
countries, Canada and Italy, adding other non-financial assets (valuables, collectibles,
jewels, etc.) in the definition of illiquid wealth definition shifts 12 and 5 percent of
households from poor to wealthy HtM, respectively. Including HELOCs among liquid
18

Recall that, based on the definitions of Section 3, changing the credit limit affects HtM status
only for households with negative liquid debt.

40

US
Baseline
0.138
In past year, c > y
0.157
1 year income credit limit
0.118
Weekly pay period
0.117
0.164
Monthly pay period
Vehicles as illiquid assets
0.059
Businesses as illiquid assets
0.131
0.137
Direct as illiquid assets
Other valuables as illiquid assets 0.133
HELOCs as liquid debt
0.134
Disposable income
0.136
Comm. cons. - beg. of period
0.114
0.174
Comm. cons. - end of period

Baseline
In past year, c > y
1 year income credit limit
Weekly pay period
Monthly pay period
Vehicles as illiquid assets
Businesses as illiquid assets
Direct as illiquid assets
Other valuables as illiquid assets
HELOCs as liquid debt
Disposable income
Comm. cons. - beg. of period
Comm. cons. - end of period

US
0.198
0.327
0.131
0.151
0.269
0.276
0.201
0.214
0.203
0.188
0.195
0.169
0.280

P-HtM
CA
AU
0.123 0.027
0.181 0.020
0.102 0.024
0.105 0.022
0.150 0.033
0.083 0.012
0.117 0.027
0.121 0.027
0.008 0.025
0.127

W-HtM
CA
AU
0.184 0.161
0.409 0.189
0.117 0.115
0.146 0.115
0.250 0.229
0.224 0.177
0.190 0.163
0.219 0.193
0.300 0.164
0.113

UK
DE
FR
0.103 0.074 0.032
0.093 0.090

0.078 0.070 0.030

0.098 0.058 0.021
0.111 0.086 0.048
0.065 0.052 0.002
0.102 0.071 0.031
0.102 0.074 0.032
0.099 0.071

0.103 0.074 0.032

0.103

0.066 0.025

0.092 0.050

IT
0.083
0.156
0.083
0.080
0.091
0.028
0.076
0.083
0.034
0.083
0.080
0.076
0.090

ES
0.044
0.091
0.040
0.036
0.061
0.024
0.043
0.045
0.044
0.044

0.036
0.064

UK
DE
FR
0.232 0.248 0.173
0.251 0.392

0.135 0.229 0.157

0.211 0.161 0.087
0.276 0.370 0.354
0.269 0.270 0.204
0.232 0.251 0.173
0.246 0.303 0.198
0.235 0.252

0.154 0.238 0.166

0.237

0.219 0.127

0.344 0.336

IT
0.155
0.474
0.147
0.142
0.188
0.211
0.158
0.165
0.204
0.147
0.149
0.148
0.173

ES
0.152
0.596
0.141
0.119
0.220
0.173
0.154
0.162
0.153
0.140

0.138
0.199

Table 6: Robustness results for fraction P-HtM and W-HtM in each category. Higher
illiquid wealth cutoff requires households to have above 1,000 local currency units in
illiquid assets to be considered W-HtM. Vehicles as illiquid assets includes the value
of other valuables for France as the value of vehicles combined with other valuables.
Businesses as illiquid assets drops the self employment income sample selection and
adds business assets to illiquid wealth and self employment income to labor income.
Direct as illiquid assets classifies directly held mutual funds, stocks, corporate and government bonds as illiquid assets. Disposable income removes taxes from gross income.
Taxes for the U.S. are estimated from NBERs TAXSIM assuming all households file
as single with no dependents. Comm. cons. - beg. of period assumes the households
committed consumption is incurred at the beginning of the period. Comm. cons. end of period assumes the household incurs it at the end of the period.

41

debt has no effect, except in Canada, where the share of HtM increases by 8 percentage
points.
Our baseline measure of income is after transfers but before taxes, except for Canada
where it is disposable income. For three countries, the U.S., the U.K., and Italy, we
can analyze the effect of netting taxes at the source for every household.19 In these
three countries, the effect is minor.

6.2

Demographic characteristics and persistence of HtM behavior across countries

Age profiles of the fraction of poor and wealthy HtM households in each country are
shown in Figure 9. For most countries, the fraction of P-HtM households declines
monotonically with age. The exceptions are Australia and France, where the age
profiles of the P-HtM is flat. There are some marked differences in the age profiles of
the W-HtM which are explained by differences in portfolio holdings across countries.
In countries where housing wealth is a substantial part of household portfolios, such
as the U.S., Canada, and the U.K., the age profile is hump shaped with a peak in the
early 40s. In contrast, in Australia and Germany, where a high fraction of W-HtM
households hold retirement accounts, the share of W-HtM is decreasing with age.
An important caveat to these results is that because we infer age profiles from a single
cross-section, we necessarily confound age, cohort, and time effects. This could explain
why in Spain, in spite of the large share of homeowners among the W-HtM, their share
falls with age. This pattern may reflect time effects, since 25-35 year olds have faced
much harsher economic conditions upon entry into the labor market than 35-45 and
45-55 year olds faced over the past two decades.
Figure 10, which shows age-income profiles for each country by HtM status, confirms
our findings from Section 5.2. The age-income profile for W-HtM households is much
more similar to the profile of the N-HtM than to the profile for P-HtM. The only two
exceptions are Italy and Spain, where the three age paths are all very similar.
19

For the U.S., we resort to an imputation based on TAXSIM as explained in Section 5.1.1. The
U.K. and Italian surveys ask households about their tax liabilities.

42

.4
.1
0

.2

.1
0

.2

.3

.4

PHtM

.3

WHtM

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

.3
.1
0

.2

.1
0

.2

.3

.4

(b) Canada

.4

(a) United States

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2024 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

.3
.1
0

.2

.1
0

.2

.3

.4

(d) United Kingdom

.4

(c) Australia

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

.3
.1
0

.2

.1
0

.2

.3

.4

(f) France

.4

(e) Germany

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

(g) Italy

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

(h) Spain

Figure 9: Age profile of fraction of HtM households by country.

43

80000
20000
0

40000

60000

80000
60000
40000
20000
0
2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

(b) Canada

30000
10000
0

20000

25000
0

50000

75000

40000

100000

(a) United States

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2024 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

20000
0

40000

20000
0

40000

60000

(d) United Kingdom

60000

(c) Australia

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

20000
0

40000

20000
0

40000

60000

(f) France

60000

(e) Germany

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

(g) Italy

2224 2529 3034 3539 4044 4549 5054 5559 6064 6569 7074 7579
Age

(h) Spain

Figure 10: Age profile of median income by HtM status by country.

44

The consumption response of the wealthy handto-mouth to transitory income shocks

In the previous sections we documented a sizable presence of wealthy HtM households

across a number of countries. In this section we show evidence that, as predicted by
the theory presented in Section 2, these households have a large MPC with respect to
transitory income shocks. From the Panel Study of Income Dynamics (PSID), we estimate the consumption response to transitory changes in income using the methodology
proposed by Blundell, Pistaferri, and Preston (2008, hereafter BPP), and further examined in Kaplan and Violante (2010). The novelties of our empirical analysis, relative
to BPP, are that we use a more recent sample period with enriched data and, most
importantly, we estimate transmission coefficients of income shocks to consumption
separately for different types of HtM households.
Data source and sample selection.
Estimating the consumption response to
income shocks for households with different types of HtM status requires a longitudinal dataset with information on income, consumption, and wealth at the household
level. Starting from the 1999 wave, the PSID contains all this data. The PSID started
collecting information on a sample of roughly 5,000 households in 1968. Thereafter,
both the original families and their split-offs (children of the original family forming a
family of their own) have been followed. The survey was annual until 1996 and became
biennial starting in 1997. In 1999, the survey augmented the consumption information
available to researchers, which now covers over 70% of all consumption items available in the Consumer Expenditure Survey (CEX), and also asked a set of additional
questions on the household balance sheet in every wave.20 This addition makes the
PSID the only large scale representative U.S. panel to include income, consumption,
and assets data.
From the PSID Core Sample, we drop households with missing information on race,
education, or state of residence, and those whose income grows more than 500%, falls
by more than 80%, or is below $100. We drop households who have top-coded income
or consumption. We also drop households that appear in the sample fewer than three
consecutive times, because identification of the coefficients of interest requires a minimum of three periods. In our baseline calculations, we keep households where the
20

Until 1999, the Wealth Files supplemented the annual survey every five years. Starting from 1999,
they became biannual, like the survey itself. In 2009 and 2011, the wealth questions were enriched
further with the Housing, Mortgage Distress, and Wealth Data Supplements.

45

head is 25-55 years old. Our final sample has 39,772 observations over the pooled years
1999-2011 (seven sample years).
Definitions.
The construction of our consumption measure follows Blundell,
Pistaferri, and Saporta-Eksten (2013). We include food at home and food away from
home, utilities, gasoline, car maintenance, public transportation, child care, health
expenditures, and education. Our definition of household income is labor earnings
of the households plus government transfers. Liquid assets in the PSID include the
value of checking and savings accounts, money market funds, certificates of deposit,
savings bonds, and Treasury bills plus directly held shares of stock in publicly held
corporations, mutual funds, or investment trusts. Before 2011, liquid debt is the value
of debts other than mortgages, such as credit cards, student loans, medical or legal
bills, and personal loans. In 2011, liquid debt includes only credit card debt. Net
liquid wealth is liquid assets minus liquid debt. Net illiquid wealth is the value of home
equity plus the net value of other real estate plus the value of private annuities or IRAs
and the value of other investments in trusts or estates, bond funds, and life insurance
policies.21 . Net worth is the sum of net illiquid and net liquid wealth. Given these
definitions of income and wealth, the HtM status indicators are constructed exactly as
outlined in Section 3, where the pay-period is assumed to be two weeks. In our PSID
sample, 25 percent of households are W-HtM, in line with the SCF estimates. The
share of the P-HtM is 21 percent, and hence is somewhat larger than in the SCF.
The BPP methodology.
We refer the reader to BPP and to Kaplan and Violante
(2010) for a thorough description of the methodology. Here, we only sketch the key
steps. As in BPP, we first regress log income and log consumption expenditures on
year and cohort dummies, education, race, family structure, employment, geographic
variables, and interactions of year dummies with education, race, employment, and
region. We then construct the first differenced-residuals of log consumption cit and
log income yit . Recall that, since the survey is biannual, a period is two years.
The income process yit is represented as an error component model which comprises
orthogonal permanent components and transitory I.I.D. component. Hence, income
growth is given by
yit = it + it ,
(9)
where it is the permanent shock and it is the transitory shock. This is a common
income process in the empirical labor literature, at least since MaCurdy (1982) and
21

The main discrepancies with the SCF definitions are that (i) we do not attempt a cash imputation
and (ii) CDs and saving bonds are in liquid, instead of illiquid, wealth.

46

Abowd and Card (1989) who showed that this specification is parsimonious and fits
income data well. The BPP estimator of the transmission coefficient of transitory
income shocks to consumption, the MPC, is given by
cov(cit , yi,t+1 )
\
M
P Ct =
.
cov(yi,t , yi,t+1 )

(10)

Under the assumption that consumption growth in period t is uncorrelated with the
transitory and permanent shocks at t + 1, i.e.:
cov (cit , i,t+1 ) = cov (cit , i,t+1 ) = 0,

(11)

this estimator uncovers the marginal propensity to consume out of a transitory shock,
i.e.:
cov (cit , it )
(12)
M P Ct =
var(it )
Condition (11) means that the household has no foresight, or no advanced information,
about future shocks. Under this condition, (10) is a consistent estimator of the passthrough coefficient of transitory shocks into consumption in (12). The estimator is
implemented by an IV regression of cit on yit , instrumented by yi,t+1 . Kaplan and
Violante (2010) show that the presence of tight borrowing constraints does not bias
the estimate of the transmission coefficient for transitory shocks. This is an important
finding in light of the fact that we are interested in the differential response of HtM
households, who may be close to a constraint, and non HtM households.
Results.
Table 7 summarizes our results. In our baseline specification, the MPC
of the W-HtM group is the highest, around 30 percent. In other words, in the first
two years, the W-HtM households consume 30 percent of an unexpected change in
income whose effect dissipates entirely within the period. The point estimate of the
MPC for the P-HtM is 24 percent, and for the N-HtM is less than 13 percent. Given
the well known measurement error present in survey data, especially for consumption
expenditures, and the small sample size, it is not surprising that these estimates are
somewhat imprecise. However, the difference between the MPC for the W-HtM and
the N-HtM is statistically significant.
When the sample is split between HtM and non HtM based on net worth, the estimated transmission coefficients are very similar across the two groups. The group of
HtM-NW is essentially the same as the P-HtM, and in fact their estimated MPCs are
similar. However, among the N-HtM-NW there are many W-HtM households with high
47

P-HtM
Baseline
0.243
(0.065)
0.131
Pre-tax earnings
(0.043)
Include food stamps
0.217
(0.059)
Cont. married households
0.095
(0.194)
Stable marital status
0.239
(0.085)
Households with male heads 0.186
(0.080)
Monthly income
0.229
(0.068)

W-HtM
0.301
(0.048)
0.223
(0.035)
0.264
(0.045)
0.193
(0.079)
0.282
(0.054)
0.193
(0.058)
0.288
(0.053)

N-HtM HtM-NW
0.127
0.229
(0.036)
(0.054)
0.122
0.143
(0.027)
(0.036)
0.105
0.203
(0.035)
(0.050)
0.079
-0.048
(0.043)
(0.129)
0.110
0.190
(0.038)
(0.070)
0.073
0.150
(0.040)
(0.064)
0.159
0.236
(0.034)
(0.057)

N-HtM-NW
0.201
(0.030)
0.164
(0.023)
0.171
(0.029)
0.157
(0.042)
0.195
(0.033)
0.129
(0.035)
0.199
(0.030)

Table 7: MPC out of transitory income shocks for different types of HtM households.
Bootstrapped standard errors based on 250 replications in parenthesis. Pre-tax earnings: transfers excluded. Include food stamps: food stamps are included among transfers. Cont married households: sample restricted to continuously married households.
Stable marital status: sample restricted to households with no change in marital status.
Households with male heads: households with female heads (mostly single) excluded
from the sample. Monthly earnings: pay-period set to one moth instead of two weeks.
MPCs. Based on this latter household classification, one would conclude that there is
no evidence of a differential response of consumption to income shocks based on HtM
status. A classification based on liquid and illiquid wealth, instead, finds economically
significant differences.
The remaining rows in Table 7 offer a robustness analysis with respect to the definition
of income and consumption, household composition, and the assumed pay-period. The
ranking of MPCs between wealthy, poor, and non HtM is always as in the baseline
specification, and as predicted by the theory, the gap between HtM households based
on the net worth criterion is always very small or is not statistically significant.

7.1

Additional evidence

Although none of the existing empirical investigations of the consumption reaction to

income shocks has explicitly tried to separate wealthy and poor HtM, as we did here,
some of them offer, indirectly, evidence on wealthy HtM behavior.

48

.8
Share of HtM among Homeowners
.2
.4
.6
0
=0

0.3

.3.6

.6.9

.91.2 1.21.5 1.51.8 1.82.1 2.12.4 2.42.7 2.73

Loan to Value Ratio

Figure 11: Share of W-HtM households among homeowners by loan-to-value ratio.

U.S. SCF 1989-2010.

In some of these studies, the sample is split between homeowners with large mortgages
and those with small leverage ratios. With respect to hand-to-mouth behavior, it is
intuitive to conjecture that the former group is more likely to be wealthy HtM because
the regular mortgage payments absorb a significant fraction of disposable income. Figure 11 shows that, in the U.S., this is indeed the case: among homeowners, the fraction
of W-HtM, according to our definition, grows steadily with the loan to value ratio.
A number of recent papers find evidence of consumption responses to income shocks
whose strength increases in the degree of indebtedness of the homeowner. Misra and
Surico (2013) expand on the research of Johnson, Parker and Souleles (2006) and
Parker, Souleles, Johnson and McLelland (2014) on the U.S. fiscal stimulus payment
episodes of 2001 and 2008. They conclude that, for both stimulus programmes, the
largest propensity to consume out of the tax rebate is found among households who
own real estate but have high levels of mortgage debt. Cloyne and Surico (2013)
exploit a long span of expenditure survey data for the U.K. and a narrative measure
of exogenous income tax changes. Their key finding is that homeowners with high
mortgage debt exhibit large and persistent consumption responses to tax shocks. Baker
(2013) combines several novel sources of household data on consumption expenditures,
income, and household balance sheets to investigate the comovement of income and
consumption, at the micro level, around the Great Recession. He finds that highlyindebted households with illiquid assets are more sensitive to income fluctuations, and
attributes this result largely to borrowing and liquidity constraints.
A different sort of evidence on wealthy HtM behavior comes from the Danish fiscal

49

stimulus of 2009. At the end of 2009, the Danish government allowed households to
withdraw from their own pension funds accounts that would otherwise be illiquid until
age 65 and transform them into liquid resources available for consumption. Kreiner,
Lassen, and Pedersen (2012) measured household consumption expenditures before and
after the policy as well as their marginal interest rate on loans and credit cards. They
conclude that households with a high borrowing cost consumed the largest share of
their withdrawals.

Implications of the wealthy hand-to-mouth for

fiscal policy

What does the existence of W-HtM households, together with their large propensities
to consume out of transitory income shocks, imply for how one should think about
fiscal policy? In this section we use a series of policy simulations from three alternative
models to argue that in the cast of hand-to-mouth characters, W-HtM households
require their own unique place. Viewing the household portfolio data through the lens
of a model with only two types of households -the non HtM and the HtM leads to
a distorted view of the effects of fiscal stimulus policies on household consumption,
compared with the effects one obtains when viewing the data though the lens of a
model with all three types of households.
Our approach is to use simulated marginal propensities to consume (MPCs) from three
alternative structural life-cycle models, in conjunction with the empirical estimates of
the shares of HtM households from Sections 5 and 6. We use calibrated versions of
each model to predict average MPCs for households with each type of HtM status. We
then use the estimated fractions of the population in each of these groups to compare
the predicted overall average MPCs for each country, under each of the three models.
The first model that we use is the two-asset incomplete markets model from Kaplan and
Violante (2014a,2014b, KV thereafter). We label this model SIM-2, since it extends
the standard incomplete markets (SIM) life-cycle economy by adding a second illiquid
asset which pays a higher return through both a financial component and a housing
services component but is subject to a transaction cost. For the reasons explained in
Section 2, the illiquidity due to the transaction cost means that the model generates
households of all three HtM types. We refer the reader to KV for a full description of
the model, its calibration, and a comparison of the predictions of the model with lifecycle data, and with the aggregate consumption response to the 2001 and 2008 fiscal
50

stimulus payments as estimated by Johnson, Parker and Souleles (2006), and Parker,
Souleles, Johnson, and McLelland (2013), respectively. Here, it suffices to say that the
version of the model we use here does not allow borrowing and has a transaction cost
of $1,000.
The second model, which we label SIM-1, is a standard one asset incomplete markets life
cycle model. The version that we adopt is the same as in KV, but with the transaction
cost set to zero, and recalibrated to data on net worth alone, rather than data on
illiquid and liquid assets separately. Since this is a one asset model, it generates only
P-HtM and N-HtM households. When using this model to compute aggregate MPCs,
we impute the model implied MPCs for N-HtM households to the W-HtM households
in the data. This is because, when the data is viewed through the lens of this model,
net worth is the relevant measure of wealth, and by this measure the wealthy and non
HtM are equivalent.
The third model, which we label SPS, is a spender-saver model in the spirit of Campbell
and Mankiw (1989) and, more recently, Eggertson and Krugman (2012), and Justiniano, Primiceri and Tambalotti (2013). In the SPS model, one group of households (the
savers) act as forward-looking optimizing consumers who can save in a single risk-free
asset. The remaining households (the spenders) follow the rule-of-thumb consumption
policy of consuming their income in every period. This class of models is typically
calibrated so that the distinction between the spenders and savers is based on their
holdings of liquid wealth rather than net worth. Thus, in the SPS model, the W-HtM
and the P-HtM households are considered to be the spenders, while the N-HtM households are considered to be the savers. When we compute the aggregate MPC for the
SPS economy we impute the MPC for the N-HtM households in the SIM-1 model to
the N-HtM households in the data, and we impute an MPC of one to the W-HtM and
P-HtM households in the data.
These three alternative economies can thus be thought of as follows. SIM-2 is an
economy in which all the W-HtM households are explicitly modeled as a distinct group.
SIM-1 is a net worth economy where the W-HtM households are treated as if they
were N-HtM households, and the only households with a high MPC are the P-HtM.
SPS is a liquid wealth economy where both the W-HtM and the P-HtM are treated
identically as HtM households, who have an MPC that is always equal to 1.
For each of the three models, we first simulate a cohort of households. For each
simulated household we compute their consumption response to a one-time unexpected

51

Average simulated MPC under SIM2 model

0
.2
.4
.6
<=40 4060

>60

low income

<=40 4060

>60

middle income
NHtM

PHtM

<=40 4060

>60

high income
NHtM

Figure 12: Average simulated MPC under SIM2-model by age group, income tercile
and HtM status.

cash windfall or cash loss of different amounts ($50, $500, $2,000). We then divide the
simulated population into 27 bins, based on three income terciles, three age groups
(22-40, 40-60, 60+) and the three HtM groups. For each of these bins we compute the
average consumption response from the model.
Figure 12 shows the average MPCs out of a $500 windfall in the SIM-2 economy for
each of the 27 bins. The MPCs are close to zero for N-HtM households, except for
households who are both income poor and very young. For high-income households
who are N-HtM, the average MPC is slightly negative. The intuition for this finding is
discussed in detail in KV. It arises because the receipt of a $500 windfall may trigger a
household who has already accumulated lots of liquid wealth, and is close to its planned
date of deposit, to pay the transaction cost and make an earlier deposit into the illiquid
account. Since such a household can effectively save at the rate of return on the illiquid
asset, it chooses to consume less and save more than it would have in the absence of the
income windfall. This example illustrates how explicitly modeling W-HtM behavior
through transaction costs may alter the MPC even for N-HtM households.
The MPCs for both the W-HtM and P-HtM households in the SIM-2 economy are
substantial. They are slightly larger for the W-HtM than the P-HtM, particularly for
households with a high level of income. The reason is that, as explained in Section 2,
52

Model
SIM-2 SIM-1

SPS

$500 windfall

0.18

0.04

0.35

Size asymmetry
$50 windfall
$2000 windfall

0.29
0.05

0.05
0.03

0.35
0.35

Sign asymmetry
$500 loss

0.42

0.14

0.36

Income targeting
$500 windfall, bottom tercile
$500 windfall, top tercile

0.26
0.20

0.07
0.03

0.50
0.34

Table 8: Estimated aggregate consumption responses for US using demographics and

HtM composition from 2010 SCF.
since the W-HtM have higher lifetime incomes than the P-HtM households, they have
higher desired consumption and hence spend more out of a moderately sized payment.22
The simulated MPCs in the other two models are similar to those in Figure 7.1. In the
SIM-1 model, the MPCs for P-HtM households are almost identical to those for P-HtM
households in the SIM-2 model, and the MPCs for N-HtM households and hence for
the W-HtM as well are only slightly larger than those for N-HtM households in the
SIM-2 model (and are never negative). For the SPS model, the MPCs for the N-HtM
households are the same as in the SIM-1 model and are equal to one for both groups
of HtM households.
Implied cross-country variation in MPCs
We compute an aggregate MPC
for each of the eight countries in our data. To do this, we estimate the fraction of
households in each country who fall into each of the 27 bins, and then apply these
country-specific group weightings to the model-generated MPCs. The implied aggregate quarterly consumption responses for the U.S. using the SCF data from 2010 are
shown in Table 8. For a $500 windfall, the aggregate MPC of the SIM-2 economy
is 0.18. This is substantially larger than for the MPC of the SIM-1 economy (0.04)
22

In the model there are no simulated P-HtM households who are in the top tercile of the income
distribution and are over 60 years old. This is almost true in the data too. For example, in the U.S.
in 2010, less than 0.1% of households fall in this category. For these households we impute the MPC
as the average between the the MPC for P-HtM aged 40-60 in the top income tercile, and the MPC
for P-HtM aged 60+ in the middle income tercile.

53

.4
.05 .1 .15 .2 .25 .3 .35
Estimated MPC under SPS model

.4
Estimated MPC under SIM1 model
.05 .1 .15 .2 .25 .3 .35

us

uk

ca
de

it

fr
au

it

de

ca

us uk

es

aufr

es

.1

.125
.15
.175
Estimated MPC under SIM2 model
SIM1 model

SPS model

.2

45 degree line

Figure 13: Estimated aggregate consumption response by country under SIM-2 model
(x axis), SIM-1model (triangles, left y axis) and SPS model (circles, right y axis).

because the SIM-1 economy treats the W-HtM households as N-HtM and so misses
a large part of the population who have a high MPC. The MPC is highest for the
SPS economy (0.35), because the SPS model implicitly assumes that the P-HtM and
W-HtM households all spend the entire $500, while we saw in Figure 12 that this is
an exaggeration: in the SIM-2 economy, they only spend 30%-40% of the payment on
average in the quarter they receive it.
The differences in the level of the MPCs predicted by the three models are driven by
the differences in the fraction of households to whom the models assign high MPCs.
Since there is substantial variation in the fraction of HtM households across countries,
this implies that the three models also yield very different predictions for the degree
of cross-country dispersion in the MPC to a $500 fiscal stimulus check. To illustrate
these differences, Figure 13 plots the estimated aggregate MPC under the SIM-2 model
against the corresponding MPC under the SIM-1 model (triangles) and the SPS model
(circles).
The figures show striking differences in the amount cross-country dispersion in the
54

MPC across the three models. There is much less dispersion under the SIM-1 model
than under the SIM-2 model. Cross-country dispersion arises from differences in the
fraction of HtM households. By treating the W-HtM as N-HtM, the SIM-1 model
misses most of the cross-country variation in HtM behavior. In contrast, there is more
dispersion under the SPS model than under the SIM-2 model. This is because, by
assigning an MPC of 1 to all the W-HtM households, compared with around 0.3-0.4 in
the SIM-2 model, the SPS model exhibits too much of a difference between the MPC of
HtM and N-HtM households than does the SIM-2 model. Cross-country heterogeneity
in the fraction of HtM households is thus more strongly reflected in the MPC under
the SPS model than under the SIM-2 model.
Other implications
We now highlight three additional areas where the three
classes of models yield different predictions because of the different types of consumption behavior that they assign to the P-HtM, W-HtM, and N-HtM.
First, Table 8 shows that the degree of size asymmetry in the aggregate MPC differs
remarkably across the three models. Under the SIM-2 model, the consumption response
to a $50 windfall is 0.29 while the response to a $2,000 windfall is only 0.05. The reason
for this large asymmetry is the availability of an illiquid savings instrument subject to
a transaction cost. For small windfalls, households in SIM-2 face an inter-temporal
tradeoff that is governed by the (low) return on the liquid asset and thus have a large
incentive to consume. However, for large enough windfalls, many HtM households may
find it optimal to pay the transaction cost and make a deposit into the illiquid asset.
This size asymmetry is absent from both the SIM-1 and SPS models. In the SPS model
it is absent because of the assumed rule-of-thumb behavior: the HtM households in
the SPS model always consume their entire windfall, regardless of the size. In the
SIM-1 model there is some decline in the MPC with the size of the payment, but it is
much more modest than in the SIM-2 model because households always face the same
inter-temporal trade off when making their consumption decisions. These experiments
clearly illustrate why it is important to think deeply about the response of W-HtM
households when considering the design of stimulus policies. Viewing the data through
the lens of either the SIM-1 or SPS model would lead one to conclude that there is far
more scope for stimulating consumption by increasing the size of stimulus payments,
than is predicted by the SIM-2 model.
Second, the degree of sign asymmetry differs across the three models. Under the SIM-1
and SIM-2 models, the response to a $500 negative income shock is substantially larger

55

than the response to a $500 positive income shock. The reasons is that even HtM
households, who by definition are at a kink in their budget constraints, desire to save
some part of a positive windfall if the windfall is large enough to push them off the kink.
Negative income shocks, however, cannot be smoothed for households at the constraint
and withdrawing from the illiquid account is too expensive to be optimal recall that
in the calibrated SIM-2 model, the transaction cost is $1,000. Under the SPS model,
the response to a positive and negative shock are essentially the same, since the MPC
is driven by the HtM households who have an MPC of 1 to both positive and negative
shocks.
Third, Table 8 reveals different implications of income targeting across the three
models. A widely held view is that the aggregate consumption response to a fiscal
stimulus policy, per dollar paid out, is larger when the payments are made to households
with lower income, i.e. stimulus payments should be phased out for middle- and highincome households for maximum effect. This view is based on the observation that PHtM households have very low income, as illustrated in Figure 6d. However, Figure 6d
also showed that W-HtM households have much higher income than P-HtM households,
yet we saw in Section 7, as well as in Figure 12, that the W-HtM are the households
who are most likely to spend their stimulus payments. For this reason, the SIM-2
model generates only a very modest decline (0.26 to 0.20) between the average MPC
out of a $500 windfall for households in the lowest income tercile and households in
the middle income tercile. The corresponding relative declines across income terciles
are much larger under the SIM-1 and SPS models. In the case of the SIM-1 model this
occurs because the only high MPC households are the P-HtM; in the case of the SPS
model this occurs because all HtM households are assumed to have the same MPC,
while under the SIM-2 model we saw in Figure 12 that among W-HtM households,
MPCs are increasing in income. We thus conclude that explicitly considering the
consumption response of W-HtM households may lead one to prefer stimulus policies
with substantially less phasing out at middle-income levels.

Concluding remarks

We set out in this paper to investigate, theoretically and empirically, the behavior of an
often overlooked, but highly relevant part of the population wealthy hand-to-mouth
households and to reflect on their implications for macroeconomic modeling and the
design of fiscal policy. We will conclude by taking stock of what we have learnt.
56

Theoretically, we showed that W-HtM behavior can arise when households face a tradeoff between the long-run gain from investing in illiquid assets (i.e. assets that require
the payment of a transaction cost for making unplanned deposits or withdrawals),
and the short-run cost of having fewer liquid assets available to smooth consumption.
Because of a wedge between the interest rates at which households can borrow and the
return that households earn on liquid savings, access to credit does not qualitatively
change the conclusion that W-HtM typically have high propensities to consume out of
small transitory income shocks.
Empirically, we found that around 30% to 40% of households in the US are HtM,
and that this fraction has been relatively constant over the past two decades. We
showed that the majority of HtM households, around two-thirds, are W-HtM, not PHtM. Although the total fraction of HtM households varies somewhat across the eight
countries that we have studied, from less than 20% in Australia and Spain to above
30% in the US and the UK, in all countries the W-HtM constitute the vast majority
of HtM households.
Who are the W-HtM? We highlight two key findings. First, unlike P-HtM households,
the W-HtM are not predominantly young households with low income. Rather they
have a humped-shaped age profile that peaks in the early forties, and an income profile
that mirrors strongly the income of the N-HtM. Second, the W-HtM are not simply
P-HtM with very small holdings of illiquid assets. Rather they hold substantial wealth
in housing and retirement accounts in the same proportions as N-HtM households.
So what? Why does this group of households deserve the attention of economists and
policy makers?
W-HtM households are important because they tend to have a large consumption
response to transitory income shocks, which is a key determinant of the efficacy of
many types of fiscal interventions, such as the fiscal stimulus payments that were
used in the last two recent recessions. To demonstrate this, in Section 7 we utilized
the identification strategy of Blundell, Pistaferri and Preston (2008), to show that
the consumption response to transitory income shocks is significantly larger for WHtM and P-HtM households than for N-HtM households. But splitting households
according to whether they were HtM based on their holdings of total assets yields
identical consumption responses for HtM and non-HtM households.
The W-HtM thus have consumption responses that are similar to the P-HtM, yet have
demographic and financial characteristics that resemble the N-HtM. This suggests that
57

the three types of HtM households each need their own unique place in frameworks that
are to be used for understanding and forecasting the effects of fiscal policy. Macroeconomists need to move beyond one-asset models, i.e. those in the spirit of Aiyagari
(1994), Huggett (1995) and Rios-Rull (1996), since these models assume W-HtM households act the same way as N-HtM households. Spender-saver models., i.e those in the
spirit of Campbell and Mankiw (1989) and Eggertsson and Krugman (2012), are not
a panacea, since these models assume W-HtM households act identically to P-HtM
households. Moreover, spender-saver models do not allow for the fact that the W-HtM
households hold illiquid wealth to affect their behavior.
In Section 8, we used an alternative framework from Kaplan and Violante (2014), which
explicitly models W-HtM behavior through the presence of both liquid and illiquid
assets, to illustrate two examples where misleading inferences would be obtained by
using either of the two simpler models of HtM households. First we showed that
the drop in the consumption response to a fiscal stimulus payment as the size of the
payment increases, is much steeper in the model that allows for W-HtM behavior
than in the models that do not. Second, we showed that the model that allows for
W-HtM behavior implies that to maximize the aggregate consumption response to
fiscal stimulus payments, the payments should feature more moderate phasing out
with household income.

58

References
Abowd, J. M., and D. Card (1989): On the Covariance Structure of Earnings and Hours
Changes, Econometrica, 57(2), 411445.
Agarwal, S., and W. Qian (2013): Consumption and Debt Response to Unanticipated
Income Shocks: Evidence from a Natural Experiment in Singapore, Working paper.
Aiyagari, S. R. (1994): Uninsured Idiosyncratic Risk and Aggregate Saving, The Quarterly Journal of Economics, 109(3), 65984.
Alvarez, F., L. Guiso, and F. Lippi (2010): Durable Consumption and Asset Management with Transaction and Observation Costs, NBER Working Paper, 15835.
Angeletos, G., D. Laibson, A. Repetto, J. Tobacman, and S. Weinberg (2001):
The Hyperbolic Consumption Model: Calibration, Simulation, and Empirical Evaluation, Journal of Economic Perspectives, 15(3), 4768.
Attanasio, O. P., and G. Weber (1993): Consumption growth, the interest rate and
aggregation, The Review of Economic Studies, 60(3), 631649.
Baker, S. R. (2013): Debt and the Consumption Response to Household Income Shocks,
Mimeo.
Banks, J., and S. Tanner (2002): Household Portfolios in the United Kingdom, in
Household Portfolios, ed. by L. Guiso, M. Haliassos, and T. Jappelli, pp. 219250. MIT
Press.
Black, O. (2011): Wealth in Great Britain, Main Results from the Wealth and Assets
Survey 2008/10, Office for National Statistics.
Blundell, R., L. Pistaferri, and I. Preston (2008): Consumption Inequality and
Partial Insurance, American Economic Review, 98(5), 18871921.
Borsch-Supan, A., and A. Eymann (2002): Household Portfolios in Germany, in Household Portfolios, ed. by L. Guiso, M. Haliassos, and T. Jappelli, pp. 291340. MIT Press.
Bricker, J., B. Bucks, A. Kennickell, T. Mach, and K. Moore (2011): Surveying
the aftermath of the storm: changes in family finances from 2007 to 2009, Discussion
paper, Board of Governors of the Federal Reserve System (US).
Broda, C., and J. Parker (2014): The Economic Stimulus Payments of 2008 and the
Aggregate Demand for Consumption, Mimeo.
Browning, M., and T. F. Crossley (2001): The Life-Cycle Model of Consumption and
Saving, Journal of Economic Perspectives, 15(3), 322.
Campbell, J. Y., and N. G. Mankiw (1989): Consumption, Income and Interest Rates:
Reinterpreting the Time Series Evidence, in NBER Macroeconomics Annual 1989, Volume
4, NBER Chapters, pp. 185246. National Bureau of Economic Research, Inc.
(1990): Permanent income, current income, and consumption, Journal of Business
& Economic Statistics, 8(3), 265279.
(1991): The response of consumption to income: A cross-country investigation,
European Economic Review, 35(4), 723756.

59

Carroll, C. D. (1997): Buffer-Stock Saving and the Life Cycle/Permanent Income Hypothesis, Quarterly Journal of Economics, 112(1), 155.
Carroll, C. D., J. Slacalek, and K. Tokuoka (2014a): Digestible Microfoundations:
Buffer Stock Saving in a Krusell-Smith World, Mimeo.
(2014b): The Distribution of Wealth and the Marginal Propensity to Consume,
Mimeo.
Chetty, R., and A. Szeidl (2007): Consumption Commitments and Risk Preferences,
The Quarterly Journal of Economics, 122(2), 831877.
Cloyne, J., and P. Surico (2013): Household Debt and the Dynamic Effects of Income
Tax Changes, CEPR Discussion Papers, (9649).
Cochrane, J. H. (1989): The Sensitivity of Tests of the Intertemporal Allocation of Consumption to Near-Rational Alternatives, American Economic Review, 79(3), 31937.
Daffin, C. (2009): Wealth in Great Britain, Main Results from the Wealth and Assets
Survey 2006/08, Office for National Statistics.
Deaton, A. (1991): Saving and Liquidity Constraints, Econometrica, 59(5), 12211248.
Eggertsson, G. B., and P. Krugman (2012): Debt, Deleveraging, and the Liquidity
Trap: A Fisher-Minsky-Koo Approach, The Quarterly Journal of Economics, 127(3),
14691513.
Eurosystem Household Finance and Consumption Network (2013a): The Eurosystem Finance and Consumption Survey: methodological report for the first wave,
ECB Statistics Paper Series, 1.
(2013b): The Eurosystem Finance and Consumption Survey: results from the first
wave, ECB Statistics Paper Series, 2.
Foster, K., E. Meijer, S. Schuh, and M. Zabek (2011): The 2009 Survey of Consumer
Payment Choice, Federal Reserve Bank of Boston Public Policy Discussion Paper, 11-1.
Friedman, M. (1957): A Theory of the Consumption Function. Oxford and IBH Publishing
Company.
pez-Salido, and J. Valle
s (2007): Understanding the Effects of
Gal, J., J. D. Lo
Government Spending on Consumption, Journal of the European Economic Association,
5(1), 227270.
Guiso, L., M. Haliassos, and T. Jappelli (2002): Household Portfolios. MIT Press.
Hall, R. E. (1978): Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence, Journal of Political Economy, 86(6), 97187.
Huggett, M. (1996): Wealth Distribution in Life-Cycle Economies, Journal of Monetary
Economics, 38(3), 469494.
Huntley, J., and V. Michelangeli (2014): Can Tax Rebates Stimulate Spending in a
Life-Cycle Model?, American Economic Journal: Macroeconomics, 6(1), 162189.
Jappelli, T., and L. Pistaferri (2010): The consumption response to income changes,
Discussion paper, National Bureau of Economic Research.

60

Johnson, D., J. Parker, and N. Souleles (2006): Household Expenditure and the
Income Tax Rebates of 2001, The American Economic Review, 96(5), 15891610.
Justiniano, A., G. E. Primiceri, and A. Tambalotti (2013): Household Leveraging
and Deleveraging, NBER Working Papers, 18941.
Kaplan, G., and G. Violante (2010): How Much Consumption Insurance Beyond SelfInsurance?, American Economic Journals: Macroeconomics, 2, 5387.
(2014a): A Model of the Consumption Response to Fiscal Stimulus Payments,
Econometrica, forthcoming.
(2014b): A Tale of Two Stimulus Payments: 2001 vs 2008, American Economic
Review: Papers and Proceedings.
Kennickell, A. B. (2005): The Good Shepherd: Sample Design and Control for Wealth
Measurement in the Survey of Consumer Finances, Mimeo.
Kreiner, C. T., D. D. Lassen, and S. Leth-Petersen (2012): Consumer Responses
to Fiscal Stimulus Policy and Householdsa Cost of Liquidity, CEPR Discussion Papers,
(9161).
Krusell, P., and A. Smith, Jr. (1998): Income and Wealth Heterogeneity in the Macroeconomy, Journal of Political Economy, 106(5), 867896.
Laibson, D., A. Repetto, and J. Tobacman (2003): A Debt Puzzle, in Knowledge,
Information, and Expectations in Modern Economics: In Honor of Edmund S. Phelps, ed.
by P. Aghion, R. Frydman, J. Stiglitz, and M. Woodford, pp. 228266. Princeton University
Press.
Ludvigson, S. C., and A. Michaelides (2001): Does buffer-stock saving explain the
smoothness and excess sensitivity of consumption?, The American Economic Review,
91(3), 631647.
Lusardi, A., D. Schneider, and P. Tufano (2011): Financially Fragile Households:
Evidence and Implications, Brookings Papers on Economic Activity, 42(1), 83150.
MaCurdy, T. E. (1982): The use of time series processes to model the error structure of
earnings in a longitudinal data analysis, Journal of econometrics, 18(1), 83114.
Misra, K., and P. Surico (2013): Consumption, income changes and heterogeneity:
Evidence from two fiscal stimulus programmes, CEPR Discussion Papers, (9530).
Parker, J. A. (1999): The reaction of household consumption to predictable changes in
social security taxes, The American Economic Review, 89(4), 959973.
Parker, J. A., N. S. Souleles, D. S. Johnson, and R. McClelland (2011): Consumer spending and the economic stimulus payments of 2008, Discussion paper, National
Bureau of Economic Research.
Ros-Rull, J.V. (1995): Models with Heterogeneous Agents, in Frontiers of Business
Cycles Research, ed. by T. F. Cooley, pp. 98125. Princeton University Press.
Shapiro, M., and J. Slemrod (2003a): Consumer Response to Tax Rebates, American
Economic Review, 93(1), 381396.
(2003b): Did the 2001 Tax Rebate Stimulate Spending? Evidence from Taxpayer
Surveys, Tax Policy and the Economy, 17, 83109.

61

(2009): Did the 2008 Tax Rebates Stimulate Spending?, American Economic
Review, 99(2), 374379.
Shore, S. H., and T. Sinai (2010): Commitment, Risk, and Consumption: Do Birds of
a Feather Have Bigger Nests?, The Review of Economics and Statistics, 92(2), 408424.
Souleles, N. (1999): The Response of Household Consumption to Income Tax Refunds,
The American Economic Review, 89(4), 947958.
Statistics Canada (2006): The Wealth of Canadians: An Overview of the Results of the
Survey of Financial Security, 2005, Pension and Wealth Research Paper Series.
Vandone, D. (2009): Consumer Credit in Europe: Risks and Opportunities of a Dynamic
Industry. Springer Verlag Berlin Heidelberg.
Watson, N., and M. Wooden (2002): The Household, Income and Labour Dynamics
in Australia (HILDA) Survey: Wave 1 Survey Methodology, HILDA Project Technical
Paper Series, 1.

62

Appendix
A

Data sources

The datasets can be accessed at the following websites:

SCF: http://www.federalreserve.gov/econresdata/scf/scfindex.htm
SFS: http://www23.statcan.gc.ca/imdb/p2SV.pl?Function=getSurvey&SurvId=1706&InstaId=
8244
HILDA: http://www.melbourneinstitute.com/hilda/
WAS: http://discover.ukdataservice.ac.uk/catalogue/?sn=7215&type=Data%20catalogue
HFCS: http://www.ecb.europa.eu/home/html/researcher_hfcn.en.html
The SCF, SFS, and WAS datasets are open to public use. Applications for access to
the HILDA and HFCS datasets are available at their respective websites.

63

Brookings Panel on Economic Activity

March 2021, 2014

Effects of Unconventional Monetary

Policy on Financial Institutions
Gabriel Chodorow-Reich, Harvard University
Final conference draft

Abstract
Unconventional monetary policy affects financial institutions through their exposure to real project risk,
the value of their legacy assets, their temptation to reach for yield, and their choice of leverage. I use high
frequency event studies to show the introduction of unconventional policy in the winter of 2008-09 had a
strong, beneficial impact on banks and especially on life insurance companies, consistent with the positive
effect on legacy asset prices dominating any impulse for additional risk taking.
Subsequent policy announcements had minor effects on these institutions. The interaction of low nominal
interest rates and administrative costs led money market funds to waive fees, producing a possible
incentive to reach for higher returns to reduce waivers. I find some evidence of high cost money market
funds reaching for yield in 2009-11, but little thereafter. Private defined benefit pension funds with worse
funding status or shorter liability duration also seem to have reached for higher returns beginning in
2009, but again the evidence suggests such behavior dissipated by 2012.
Overall, in the present environment there does not seem to be a trade-off between expansionary policy
and the health or stability of the financial institutions studied.

EFFECTS OF UNCONVENTIONAL MONETARY

POLICY ON FINANCIAL INSTITUTIONS
Gabriel Chodorow-Reich
Harvard University

March 2014
Abstract

Unconventional monetary policy affects financial institutions through their exposure to real project risk, the value of their legacy assets, their temptation to reach
for yield, and their choice of leverage. I use high frequency event studies to show the
introduction of unconventional policy in the winter of 2008-09 had a strong, beneficial
impact on banks and especially on life insurance companies, consistent with the positive effect on legacy asset prices dominating any impulse for additional risk taking.
Subsequent policy announcements had minor effects on these institutions. The interaction of low nominal interest rates and administrative costs led money market funds
to waive fees, producing a possible incentive to reach for higher returns to reduce
waivers. I find some evidence of high cost money market funds reaching for yield in
2009-11, but little thereafter. Private defined benefit pension funds with worse funding status or shorter liability duration also seem to have reached for higher returns
beginning in 2009, but again the evidence suggests such behavior dissipated by 2012.
Overall, in the present environment there does not seem to be a trade-off between
expansionary policy and the health or stability of the financial institutions studied.

Harvard University Department of Economics, Littauer Center, Cambridge, MA 02138 (e-mail: chodorowreich@fas.harvard.edu). Preliminary conference draft prepared for the spring 2014 Brookings Panel on Economic
Activity. I am grateful to Markus Brunnermeier, Joshua Hausman, Nellie Liang, Atif Mian, Hyun Song Shin,
David Sraer, and especially to the editors David Romer and Justin Wolfers and my discussant Annette VissingJorgensen for discussions and comments. Matthieu Gomez provided superb research assistance. This work was
completed while I was visiting the Julis-Rabinowitz Center at Princeton University. I also acknowledge financial
support from the Brookings Institution.

1.

Introduction
In the winter of 2008, the Federal Reserves Federal Open Market Committee (FOMC) began

using a mix of policy instruments unprecedented in its history. These expanded to include a
target federal funds rate of essentially zero, purchases of Treasury bonds, Agency (Fannie
Mae, Freddie Mac, and Ginnie Mae) mortgage-backed securities, Agency bonds, and explicit
guidance concerning the future path of the federal funds rate. I refer to these instruments
as unconventional monetary policy. The FOMC introduced these policies with the intention
of reducing long-term real interest rates, which it believed would lead to a stronger economic
recovery.1 A number of studies have confirmed the policies success in reducing long-term
rates (Gagnon et al., 2010; Krishnamurthy and Vissing-Jorgensen, 2011; Campbell et al., 2012;
Wright, 2012).
The introduction of new treatments raises the concern of side effects. In the aftermath
of one of the worst financial crises in history, a particularly acute one involves the effect of
unconventional policy on the health and stability of the financial sector. Indeed, numerous
FOMC participants have cited increased risk taking by financial institutions as a potential
constraint on their policy choices (Bernanke, 2013; Stein, 2013a; Fisher, 2014). Yet, relatively
little evidence exists of whether or how much unconventional policy has impacted risk taking
or overall stability of the financial sector.
I discuss four effects of unconventional monetary policy on financial institutions. First,
reducing the risk free rate lowers the hurdle rate for risky investment projects. This leads
to increased new spending on projects with both lower mean returns and higher variances.
1

See for example the FOMC statement announcing a new round of asset purchases in September 2012 (http:
//www.federalreserve.gov/newsevents/press/monetary/20120913a.htm).

Depending on the distribution of newly-funded projects, the optimal level of real risk in the
economy may change. Financial institutions carry exposure to real project risk through their
role as intermediaries between borrowers and savers. Second, unconventional policy has general equilibrium effects which lower delinquency and default rates, raise profits, and possibly
lower risk aversion. Higher probability of payoff, higher profits, and less risk aversion together
raise the price of legacy assets held by financial institutions, improving their balance sheets.
Third, unconventional policy may lead some financial institutions to seek higher returns due
to institutional dissatisfaction with low yields. By definition, such reaching for yield constitutes a deviation of risk taking from the first-best level. Fourth, low interest rates reduce the
opportunity cost of holding capital or reserves. This generates higher leverage at institutions
facing binding collateral or reserve requirements, such as banks.
I then turn to evidence on four classes of financial institutions. I use high frequency event
studies to measure the response of credit default swap (CDS) spreads, bond yields, and equity
prices of life insurers and bank holding companies in narrow windows surrounding surprise
announcements of policy changes. I study bank holding companies because of their importance
to the financial system. For life insurers, the combination of long-term fixed income liabilities
and shorter duration assets may generate a compressed or even negative interest spread at low
rates and prompt reaching for yield behavior. On the other hand, many life insurers faced
solvency crises in early 2009, and the corporate and mortgage bonds on their balance sheets
would have benefited from expansionary policy. In the event, the initial round of expansionary
policy in the winter of 2008-09 lowered CDS spreads and bond yields and raised stock prices,
with a particularly pronounced effect on life insurers. Subsequent policy announcements had

smaller or neutral effects.

Many money market funds engaged in reaching for yield in late 2007 and 2008, with disastrous consequences following the Lehman Brothers bankruptcy. The low interest rate environment could again provoke such behavior by squeezing funds ability to cover administrative
costs. Beginning in 2009, funds passed higher gross yields through into higher charged expenses
nearly one-for-one, with almost no effect on the net yield received by investors. The high pass
through rate suggests funds operators understood the cost of waiving fees and the potential to
avoid such costs by generating higher gross returns reaching for yield. I exploit cross-sectional
differences in administrative costs to determine whether such costs pushed funds to increase
their risk-taking. I find some evidence of high cost funds pursuing higher gross returns and
accepting greater return variance in 2009-11. Such behavior vanishes by 2013, however, as the
compression in yields across asset classes left little room for funds to reach for yield.
Finally, I analyze risk-taking by private defined benefit pension funds. Like life insurers,
pension funds have long-term fixed income liabilities that often exceed the duration of available
assets. I ask whether funds with a shorter duration of liabilities or worse funding status engaged
in riskier behavior beginning in 2009. The strongest evidence of pension funds reaching for
yield comes in 2009. Funds with shorter liability duration or worse funding status had higher
loadings on the market excess return in 2009, and experienced a higher variance of returns over
2009-12. Reaching for yield appears to decline thereafter. Here again, the general equilibrium
effects of unconventional monetary policy on the stock market improved the solvency position
of defined benefit pension funds, helping to counteract any deleterious effect of low interest
rates on reaching for yield.

In interpreting these results, it helps to review why policy makers might care about the
health of the financial sector, and to clarify a distinction between financial institution risk
taking and financial sector stability. Financial institution risk taking involves an active decision
by managers to change their risk profile. Low interest rates may spur increased risk taking
through the hurdle rate effect, through reaching for yield, or through their effect on leverage.
The combined gross assets of life insurers, private defined benefit pension funds, money market
funds, and regulated banks exceeded $24 trillion at the end of 2013. Their attitude toward risk
has the potential to affect the market price of risk in the economy. Policy makers may care
directly about limiting reaching for yield if it causes risk premia to fall below their first-best
level.2 On the other hand, from the theory of the second best, an increase in reaching for yield
by some institutions may improve welfare if other distortions (for example, insufficient capital
in the financial sector) have resulted in too little risk taking in the economy.
The effect of monetary policy on financial sector stability combines both the change in risk
taking and the general equilibrium effects on legacy asset prices and the economic environment.
The recent theoretical literature has emphasized the sharply non-linear effect of financial sector
capital on risk premia and lending (He and Krishnamurthy, 2013; Brunnermeier and Sannikov,
2014). In these models, large contractions in financial sector capital cause capital constraints
to bind or set off adverse feedback loops. One such feedback loop relates to the practice of
marking assets to market and the possibility of fire sale externalities. In a fire sale, an institutions need to delever puts downward pressure on the price of assets held at other institutions,
which may then cause further deleveraging (Shleifer and Vishny, 2011). The rapid economic
2

See also Borio and Zhu (2012) and Hanson and Stein (2012) for discussion of risk taking and risk premia
channels of monetary policy.

collapse following the Lehman Brothers bankruptcy highlights these non-linear dynamics, and
also the existence of systemically important financial institutions. The role of the financial
sector in intermediating credit to the real economy provides a direct link between the health of
the financial sector and the Federal Reserves employment mandate (Chodorow-Reich, 2014).
Importantly, here higher risk matters regardless of whether it stems from first-best reallocation
of resources toward riskier projects or from reaching for yield or higher leverage, particularly
if the additional risk concentrates in systemically important financial firms. Conversely, policy
makers might worry less about low interest rates spurring reaching for yield behavior if they
simultaneously reduce financial institution risk by improving legacy asset values.
With this guidance in mind, I draw three conclusions regarding the effects of unconventional
policy on financial institutions. First, the expansionary policy of the winter of 2008-09 had an
important component of financial rescue, particularly for life insurers. The subsequent designation of a large life insurer (Prudential) as a systemically important institution accentuates the
value of helping life insurers balance sheets to recover. Second, a low interest rate environment
does pose challenges to money market funds and some pension funds. Some of these funds will
reach for yield given the opportunity. Third, both life insurers and banks continued to benefit
from expansionary monetary policy in 2013, and I find no evidence of money market funds or
pension funds reaching for yield in 2012 or 2013. In the present environment, there does not
seem to be a trade-off between expansionary policy and the health or stability of the financial
institutions studied.
The paper proceeds as follows. The next section characterizes the theoretical effects of
unconventional monetary policy on financial institutions. Section 3 presents the evidence from

the event studies of life insurance companies and bank holding companies. Section 4 discusses
the response of money market funds to unconventional policy. Section 5 analyzes the response
of pension funds. Section 6 concludes.

2.

Theoretical effects of unconventional policy on financial

institutions
The policy instruments of a sustained low federal funds rate, forward guidance, and large

scale asset purchases or quantitative easing (QE) affect the economy by lowering long-term
real interest rates. It is useful to distinguish an expectations channel from a portfolio balance
channel. Each of the three instruments may trigger the expectations channel by lowering the
publics expectation of the path of policy rates. Through the expectations hypothesis of the
term structure, long-term nominal interest rates then fall. Additionally, forward guidance and
QE may generate an expectation of a lower policy rate after the zero lower bound no longer
binds, causing consumption to boom in that future state and raising spending immediately
through the Euler equation. These two effects generate a third effect of the expectations channel, a rise in expected inflation. Lower long-term nominal interest rates and higher inflation
together imply a decline in long-term real interest rates. The portfolio balance channel arises
only with QE. Here investors value certain types of securities beyond their risk-adjusted payoff
structure (for example, satisfying regulatory requirements or due to habit or market segmentation). The central bank can then affect interest rates other than the short-term policy rate
by changing the portfolio of assets private investors must hold in equilibrium.
Low interest rates in turn affect financial institutions in four ways.

2.1.

Real spending

Lowering long-term real interest rates stimulates riskless and risky real spending. Riskless
spending rises because households substitute intertemporally, firms discount future profits at a
lower rate, and a commitment to future expansionary policy generates a positive wealth effect.
These effects form the textbook channel of monetary policy (Werning, 2012).
In a world with project risk, reducing the risk free interest rate also stimulates spending
on risky projects. Project risk comes from uncertainty over, inter alia, consumer taste shocks,
future technology, future tax policy, and the regulatory environment. In corporate finance, investment decisions depend on whether the expected return from a project exceeds the projects
hurdle rate. The hurdle rate depends on both the cost of capital and the projects riskiness.
When the risk free interest rate declines, newly viable projects have either lower expected returns or higher risk. If projects on the margin of funding mostly have higher risk, total real
project risk in the economy may rise. Conversely, if the marginal projects have lower expected
returns and lower variance than the average project, total risk may fall.
A small model illustrates the channel. Consider a two period economy consisting of a representative consumer, a producer, and a monetary authority. The producer passively provides
output in exchange for money, at a price normalized to 1. The consumer enters period 0 with
real money balances Y0 , and may purchase output from the producer or make deposits at the
monetary authority at a safe gross real interest rate Rf . Purchased output can be consumed
immediately or invested in a project with risky return. The space of projects indexed by expected return and variance 2 characterize the investment opportunities. There are K (, )


(independent) projects with the same , 2 , each of which can receive either one unit of in7

vestment or no investment. Let A (, ) = 1 if the projects receive investment and 0 otherwise.3

The consumer chooses allocations to maximize utility from consumption C0 and C1 ,

U0 = u (C0 ) + E0 [u (C1 )] ,

(1)

subject to the period budget constraints

Y 0 = C0 + A f + A p

(2)

C1 = Rf Af + Rp Ap ,

(3)

where Af denotes the allocation to the safe deposit, Ap the portfolio of risky assets, and Rp
the (endogenous) return on the risky portfolio.
The model has the following interpretation. C0 and Ap require purchased output, and hence
constitute real spending. Investment projects do not have scalability increasing investment
requires accepting either a lower mean return or higher project variance. I do not model the
economys production side explicitly, but could allow for demand constrained producers as in
a standard New Keynesian model. Deposits Af do not utilize real resources. One can think of
Rf as interest paid on reserves at the monetary authority, with two caveats. First, Rf is a real
rather than nominal return. Implicitly, agents have perfect foresight over the path of inflation.
Second, in this simple two period economy, Rf encompasses all of the policies discussed above
to affect the real interest rate.
I make parametric assumptions that yield a closed-form solution to the consumers problem.
The return on a risky project with mean and variance 2 , denoted R (, ), is normally
3

Because all projects with the same (, ) are ex-ante identical, I restrict to solutions where either all or none
receive funding.

distributed and independent of the return on other projects. The space of possible projects has
a maximum return H . Utility takes the exponential (CARA) form, u (C) = exp (C).4
Finally, I assume initial real money balances Y0 large enough that the household makes strictly
positive riskless deposits, Af > 0.
Under these assumptions, Appendix B shows the consumers problem simplifies to a simple
mean variance tradeoff:
Z

max

{A(,)}

Rf

h


i
2
R K (, ) A (, ) dd .
2
f

(4)

Projects receive funding if their expected excess return Rf exceeds a multiple /2 of

their variance. This gives rise to a mean-variance frontier, on and below which every asset


receives funding. The frontier has slope 2/ in , 2 space, and has a right bound at H .


If the consumer makes strictly positive riskless deposits, Af > 0, then Rf , 0 must lie on the


frontier. Figure 1 illustrates the set of funded projects in , 2 space.
Figure 1: Set of funded projects

H , 2 H Rf



Rf

A decline in Rf generates a parallel shift out of the mean-variance frontier. It follows

4

I do not require C > 0.

2
immediately that the maximum riskiness of any funded project max
and the size of the portfolio

of risky assets Ap both increase.

In this economy, a risk-taking channel of monetary policy operates through the effect on
investment Ap . The decline in Rf increases spending on riskier projects, in the sense that at
every mean return projects with higher variance now receive funding.5 The increased spending
on risky projects does not, however, necessarily imply an increase in the total riskiness in the
economy, because spending on projects with low returns and low risk also rises. The sign of
the effect on the variance of the risky asset portfolio ( p )2 depends on the relative densities of
high risk, high return and low risk, low return projects on the margin of funding, i.e. on the
distribution of K (, ).6
The risk-taking channel of monetary policy occurs because reducing the safe interest rate
lowers the hurdle rate on investment projects, which leads to increased real spending on risky
projects. While the model has abstracted from the role of the financial sector to focus on the
effect on real spending, in reality the financial system will have exposure to the additional risky
projects through its role in intermediating between savers and borrowers.7 Notably, the change
in risk-taking in the model does not stem from any departure from the first-best allocation of
resources. As a corollary, one cannot simply examine the risk quantity or pricing of new lending
to determine optimality. At least some increase in the asset risk of financial institutions may
constitute an intended channel of unconventional monetary policy.
5

The assumptions of CARA utility and independent normal returns make mean-variance the investors sole
trade-off. More generally, spending might also increase on projects with higher CAPM s due to more procyclical
returns.
2
6
For example, one can show that if K (, ) = 1 , , both p and ( p ) unambiguously decline. Even here,
risk in the financial system might still increase if not all institutions hold the same portfolio.
7
See Adrian and Shin (2010) for an exposition of the risk-taking channel in a model with financial intermediaries.

10

2.2.

General equilibrium effects

Low real interest rates have general equilibrium effects through the increase in aggregate
demand. Higher real spending raises firm profits. It also reduces unemployment, which in
turn leads to lower loan delinquency and charge-off rates. Higher profits and lower default
probabilities raise state-contingent payoffs and hence raise asset values. Asset values will further
increase if the discount rate used to discount risky future profits falls, as would happen for
example in consumption-based asset pricing models.
Higher asset prices raise the value of legacy assets held by financial institutions, a phenomenon Brunnermeier and Sannikov (2011) refer to as stealth recapitalization. The resulting increase in net worth increases the distance to default, lowering the risk of bankruptcy. As
well, if proximity to default encourages risk-shifting behavior, the increase in net worth will
also reduce the amount of sub-optimal risk-taking by financial institutions.

2.3.

Reaching for yield

Low interest rates may also spur risk-taking by financial institutions beyond the first-best
level. Investment management poses a classic principle-agent problem, in which the incentives
of managers may not align perfectly with the objectives of shareholders and debt holders (Jensen
and Meckling, 1976; Rajan, 2005). Chairman Bernanke referenced these concerns in his May
2013 Congressional testimony:
the Committee is aware that a long period of low interest rates has costs and risks...
one that we take very seriously, is the possibility that very low interest rates, if maintained too long, could undermine financial stability. For example, investors or portfolio
managers dissatisfied with low returns may reach for yield by taking on more credit
risk, duration risk, or leverage.8
8

Testimony to the Joint Economic Committee (http://www.federalreserve.gov/newsevents/testimony/

bernanke20130522a.htm).

11

The definition of reaching for yield varies across authors. I use it to mean increases in risktaking for reasons other than the end-holders risk preferences. In the language of the model in
section 2.1, reaching for yield is an increase in the slope of the mean-variance frontier without a
change in the risk aversion coefficient . The additional risk may come from shifting investments
into higher risk asset classes (i.e. equities instead of investment grade debt), choosing higher
yield investments within an asset class, or by increasing leverage. In the following sections I give
explicit examples of why certain classes of financial institutions might reach for yield (see also
Stein, 2013a).

2.4.

Leverage

Finally, the corporate finance literature has highlighted reasons why low interest rates may
affect leverage decisions apart from reaching for yield motivations. The channel mostly concerns
banks. A decline in the safe interest rate reduces the cost of holding required reserves. For banks
facing binding collateral constraints or reserve requirements, the decline in the opportunity cost of
holding reserves leads to larger total portfolios and higher leverage (Stein, 2012; DellAriccia et al.,
2014; Drechsler et al., 2013).9 Brunnermeier and Sannikov (2014) describe a related phenomenon
stemming from the low volatility environment induced by low interest rates. As low realized
volatility feeds into banks value at risk models, banks respond by increasing leverage.

2.5.

Summary

Unconventional monetary policy affects the risk held by financial institutions by changing
the hurdle rate for risky projects, through general equilibrium effects on asset values, and by
possibly causing some institutions to reach for yield or expand leverage. The combination of
these effects implies an ambiguous overall effect of policy on the stability of the financial sector.
I thus turn to an empirical assessment of the effects of unconventional policy on four classes of
9

The increase in leverage through this channel can also exceed the first-best increase. For example, in Stein
(2012) a fire sale externality makes one banks expanded leverage negatively affect the asset values of another bank
in the event of a shock that requires deleveraging. The fire sale externality can lead to an increase in leverage
beyond the social optimum when banks do not internalize the effect of their own leverage decision on the collateral
constraints of other banks.

12

financial institutions: life insurance companies, bank holding companies, money market funds,
and pension funds.

3.

Life insurance companies and bank holding companies

I study the effect of unconventional policy on life insurance companies and bank holding

companies using event studies. Event studies permit the identification of the causal effect of a
monetary policy surprise on asset prices.10 Identification requires (1) a window narrow enough
that aggregate shocks other than the monetary policy shock do not affect the asset price during
the window, and (2) a window wide enough to allow markets to process the new information.
I compile a data set of CDS, bond, and equity prices of life insurers and bank holding companies, and Treasury prices. The single label CDS data come from Bloomberg generic quotes
for contracts with a five year tenor (contract horizon), written in U.S. dollars, and with an MR
restructuring clause.11 I obtain end of day quotes for Tokyo, London, and New York, each corresponding to the last trade before 5:15pm local time. The bond price data come from the FINRA
TRACE database, which contains transaction-level data for over-the-counter bond transactions
as collected by FINRA per the rule 6200 series. The equity data come from the TAQ database,
which contains prices from the consolidated tape of stocks traded on the NYSE, tick-by-tick, and
from Bloomberg for the most recent months. GovPx provides tick-by-tick indicative bid and ask
prices on Treasury securities from five inter-dealer brokers. The life insurer sample contains an
unbalanced panel all life insurers in the top 30 by assets, excluding AIG, and with outstanding
equity (13 firms), debt (11 firms), or CDS (4 firms) securities. The 13 life insurers in the equity
sample together held 45% of total life insurer assets at the end of 2012. The bank holding company sample contains an unbalanced panel of all bank holding companies with publicly-traded
10

See Nakamura and Steinsson (2013) for the effect of conventional and Gagnon et al. (2010) and Krishnamurthy
and Vissing-Jorgensen (2011) for the effect of unconventional policy on interest rates. English et al. (2012) and
International Monetary Fund (2013, chapter 3) extend the methodology to the study of the effect on commercial
bank stock prices, and Kiley (2013) and Gilchrist and Zakrajsek (2012) examine the effect on corporate borrowing
rates and default risk, respectively.
11
MR stands for modified restructuring. Contracts with an MR clause give the protection buyer some recourse
in the event of a credit event other than outright default. The five year tenor and the MR clause correspond to
the most liquid contracts for the reference entities studied.

13

equity (305 firms) or bonds (46 firms), and the CDS spreads for eight of the largest bank holding
companies.
Single label CDS spreads provide a direct market-based measure of an institutions likelihood
of default. When quoted in basis points, the CDS spread, also called a premium, gives the
required annual payment for a contract that will pay $10,000 if the reference institution triggers
a default clause during the contract horizon. Holding the price of risk fixed, an increase in the
spread thus indicates an increase in the default probability of the reference entity. For bonds, the
spread between the yield and the Treasury yield measures the bond risk premium. The effect on
equity prices answers the question of whether the stock market perceives the surprise as having
a positive or negative effect on the institutions future net income, suitably discounted. The
liquidity in equity markets permits the narrowest windows of the assets studied and the inclusion
of the largest number of firms. For all three measures, the response reflects the combination of
the effects described in section 2, as well as any change in the economy wide price of risk induced
by the monetary policy action. That is, I am not able to separate movements in the bond or CDS
premium into the change in default probability (the quantity of risk) and the excess premium
(the price of risk).12
Table 1 lists the unconventional monetary policy surprise dates used in the study. I classify
the dates according to four policy programs. QE1 consists of the initial round of asset purchases
announced in late 2008 and lasting through 2009. QE2 corresponds to purchases announced in
November 2010. QE3 consists of purchases begun in September 2012 and still ongoing as of
March 2014. FG contains dates related to forward guidance. Through 2012, the dates form
a subset of the announcement dates listed in International Monetary Fund (2013). With the
exception of the initial announcement on November 25, 2008, which occurred before normal
trading hours, I keep all dates used in Krishnamurthy and Vissing-Jorgensen (2011). Because
my interest lies in identifying the conditional response of financial institution asset prices to
monetary policy surprises, however, I exclude some other announcements which do not move the
five-year Treasury. For example, the International Monetary Fund (2013) list includes November
3, 2010, when the FOMC announced a new round of asset purchases. Market participants widely
12

See Gilchrist and Zakrajsek (2012) for further discussion.

14

expected this announcement, and the Treasury barely reacted on impact. As a result, I include two
dates, August 10 and September 21, 2010, when FOMC statements raised expectations of future
purchases and Treasury prices rose, but exclude the November 3 announcement itself.13 Each of
the included announcements from 2013 had an identifiable Federal Reserve communication and
a discrete change in the Treasury yield at the time of the announcement.
Table 1: Unconventional monetary policy announcement dates
Episode

Date

Time

Event

QE1
QE1
QE1
QE1
QE1
QE2
QE2
FG
FG
QE3
QE3
QE3
QE3
QE3

December 1, 2008
December 16, 2008
January 28, 2009
March 18, 2009
September 23, 2009
August 10, 2010
September 21, 2010
August 9, 2011
January 25, 2012
September 13, 2012
May 22, 2013
June 19, 2013
July 10, 2013
September 18, 2013

1:45pm
2:21pm
2:15pm
2:17pm
2:16pm
2:14pm
2:14pm
2:18pm
12:28pm
12:31pm
10:30am
2:00pm
4:45pm
2:00pm

Bernanke speech
FOMC statement
FOMC statement
FOMC statement
FOMC statement
FOMC statement
FOMC statement
FOMC statement
FOMC statement
FOMC statement
Bernanke testimony
FOMC statement
Bernanke speech
FOMC statement

Effect on 5yr Treasury notea

(Basis points)
9.2
16.8
3.1
22.8
8.9
5.8
1.8
14.4
6.3
6.4
6.6
7.8
7.3
14

a. Change in the yield to maturity of the on-the-run five year Treasury note from the 5 minute window ending
2 minutes before the announcement to the 5 minute window beginning 18 minutes after the announcement. The
yield to maturity is based on the mean of the indicative bid and ask prices in each window.

For equity and bond prices, I use tick-by-tick data to construct the difference in the five
minute average trading price from seven to two minutes before the monetary policy announcement
to eighteen to twenty-three minutes after. The narrow intraday window virtually ensures that
other aggregate shocks do not contaminate the results (Nakamura and Steinsson, 2013; Kiley,
2013). Furthermore, both the TRACE bond data and the TAQ equity data contain prices from
actual trades, and the constructed windows contain only securities with valid transactions both
before and after the announcement. This should help to alleviate concerns of market illiquidity
biasing price changes toward zero. For CDS spreads, I use data from the market close in Tokyo,
13

As in Krishnamurthy and Vissing-Jorgensen (2011), my focus on conditional responses to monetary surprise

shocks implies a loss of power but no bias from omitting valid policy surprises.

15

London, and New York to construct quasi-intraday windows corresponding to the shortest window
surrounding each announcement.14 While the CDS data have a greater concern of incorporating
information other than the monetary policy surprise, the quasi-intraday refinement substantially
improves on using only daily frequency.15
I use Treasury prices to validate the second identification assumption. Appendix figure C.1
shows minute-by-minute yields of the on-the-run five year Treasury note on each date. The dashed
lines define the period from two minutes before to 18 minutes after the announcement. Treasuries
adjust almost instantaneously to the news, with the window sufficient to capture nearly all of the
movement.16

3.1.

Life insurer results

Life insurers liabilities consist of roughly $4.5 trillion in life insurance policies and annuities
(American Council of Life Insurers, 2013). Life insurers divide their asset holdings between a
general account backing essentially all of their life insurance policies as well as fixed rate annuities,
and a separate account backing pass-through products such as variable rate annuities and pension
products. State law regulates asset allocation of the general account but not the separate account.
Both annuities and life insurance contracts often have longer duration than available assets,
leading to balance sheet duration mismatch. When interest rates fall, life insurers face a compressed or even negative interest spread as they roll over assets at the lower rates. The long-term
liability structure provides insulation from the threat of runs. Nonetheless, a compressed spread
reduces operating profits. This has led some industry analysts to postulate a negative effect of
low interest rates on life insurers, and especially so if the low interest rates persist for a long
period of time (Moodys Investors Services, 2012; McKinsey Global Institute, 2013). Life insur14

For announcements before 12:15pm, the window covers 3:15am to 12:15pm. For announcements after 12:15pm,
the window covers 12:15pm to 5:15pm. All times eastern.
15
For example, the Markit end of day spreads for March 17, 2009 and March 18, 2009 indicate an increase of 51
basis points in the value-weighted mean life insurer spread. To preview a result to come, using the change between
the London and New York close instead gives a decline of 25 basis points.
16
The on-the-run five-year note is the mostly recently issued five-year note, and generally has the greatest
liquidity. In some cases the surprise announcement came in the form of an FOMC statement, which was later
followed by a Chairmans press conference. Although Treasury prices also respond to news in the press conference,
I limit the window to include only effects from the statement itself. The narrow window does assume market
participants interpret the effect of an announcement on bond or equity prices as quickly as they do the effect on
Treasuries.

16

ers can try to offset the lower interest rates by reaching for yield, implying greater risk to their
policy holders, shareholders, and the state guarantee funds backing their policies in the event of
bankruptcy.17 On the other hand, life insurers hold roughly one-quarter of their general account
assets in mortgage-backed securities or directly held mortgages, and nearly half in corporate securities. These assets lost value during the 2008-09 crisis, rendering some life insurers nearly
insolvent. Also, before the crisis many life insurers had sold variable annuities with minimum
return guarantees, on which they would need to make good if the stock market did not recover
sufficiently. As a result, the positive general equilibrium effects of Federal Reserve policy on asset
prices may have benefited life insurers equity values and reduced the likelihood of insolvency.
Before presenting the event study results, it helps to view the broader time series of CDS
spreads, the most direct measure of riskiness. These time series cannot establish the causal effect
of low interest rates, because they do not control for other factors. However, they do provide
context for the event studies as well as helping to assess whether the post-2009 environment
cum unconventional monetary policy has resulted in an unusual concentration of risk. Figure 2
presents CDS spreads for the six insurance companies with large life insurance components and
publicly-traded CDS.18
As noted, insurance companies faced significant financial challenges during the 2008-09 crisis.
A contemporary account put their estimated losses from assets related to subprime at greater
than $180 billion (Harrington and Frye, 2009), resulting in a series of downgrades by rating
agencies. According to Koijen and Yogo (2013), many life insurers sold policies at deep discounts
during the period to exploit an accounting loophole and avoid further downgrades. The spikes in
the CDS spreads in early 2009 reflect this distress. For example, the annual premium required
for a payoff of $10 million in the event of a default by Lincoln Financial (ticker: LNC) reached
the equivalent of $3 million per year, while CDS on MetLife (ticker: MET), Prudential (ticker:
PRU) and Hartford Financial (ticker: HIG) all had premiums go as high as $1 million per year.19
17

Becker and Ivashina (forthcoming) find evidence of life insurers reaching for yield before the crisis in the sense
of choosing riskier fixed income securities within regulatory asset classes. However, they find the phenomenon
diminished following the financial crisis.
18
These data come from Markit end of day quotes. The Bloomberg generic quotes used to construct the quasiintraday windows lack data for Aflac (ticker: AFL) and Hartford Financial Services (ticker: HIG).
19
Because premium payments cease following a default event, especially distressed entities trade with an upfront
payment along with fixed coupons. For example, on April 6, 2009, purchasing a five year contract for $10 million

17

Figure 2: Insurance company CDS

5 year tenor, basis points
3000
AFL
HIG
MET

ALL
LNC
PRU

2000

1000

0
Jan08

Jan09

Jan10

Jan11

Jan12

Jan13

Spreads began to decline in March and April of 2009, roughly coincident with the stabilization
of financial markets generally and the beginning of the recovery in asset prices. The timing thus
appears consistent with general equilibrium effects of Federal Reserve policy during the winter of
2008-09 having a large benefit on life insurers. As of the end of the sample CDS spreads have
returned to historically low levels.
Turning to the event studies, figures 3 to 5 report the responses of CDS, bond, and equity
prices of life insurers to the monetary policy surprises. Each figure shows a scatterplot of the asset
price change and the change in the five year Treasury, with the announcement date labeled on the
lower horizontal axis. Table 2 reports the value-weighted mean for each announcement date, using
the sample average market capitalization as the weight for all three asset categories.20 Shaded
rows indicate contractionary surprises, defined as a positive response of the five-year Treasury
yield during the announcement window. Significance thresholds in the table use the larger of
the conventional or robust standard error from a firm-level regression of the change in the asset
of protection against Lincoln Financial required an upfront payment of $4.85 million and annual coupon payments
of $500,000 thereafter. The spreads reported in figure 2 and the Bloomberg generic quotes reported below use
the ISDA CDS standard model to amortize the upfront payment and report as if the only required payment were
annual premiums.
20
The results in the table change little using an unweighted mean or median as the measure of central tendency.

18

price on a constant on each date.21 The table also reports the log change in the on-the-run
North American IG CDX and the value-weighted mean stock price change of all companies in
the S&P 500, excluding banks and life insurers.22 The comparison with nonfinancial firms helps
qualitatively in distinguishing effects stemming from changes in the economy-wide price of risk
from effects specific to the life insurance sector.
Consistent with strong general equilibrium effects on legacy asset prices, the introduction of
near zero interest rates and quantitative easing in the winter of 2008-09 had a clear, beneficial
impact on life insurers. Using the response of the five-year Treasury yield as a guide, the two
most important announcements occurred on December 16, 2008, when the FOMC announced a
75 basis point reduction in the federal funds rate to a new target of 0-25 basis points, and on
March 18, 2009, when it announced a balance sheet expansion of up to $1.15 trillion. Summing
over the changes in the two announcement windows gives a cumulative impact of 40 basis points
on the five-year Treasury. From figure 5, every life insurer in the sample experienced an increase
in their stock price during each announcement window, with a combined value-weighted change
of 7.6 percent. In fact, equity prices of life insurers benefited far more from the announcements
than did the average bank or the average nonfinancial firm in the S&P 500. The cost of insuring
against default and bond yields also fell. The value-weighted 5 year CDS spread fell 32 basis
points over both announcements. The bond yield fell a total of 73 basis points, suggesting a
decline in the risk premium.23
The general equilibrium effects on life insurers legacy assets appear to have unfolded in line
with the markets expectation. For example, MetLifes 2010 annual report begins its discussion of
financial condition and result of operations: As the U.S. and global financial markets continue to
recover, we have experienced a significant improvement in net investment income and favorable
21

Some of the entries in table 2 have very few observations, in which case robust standard errors can have
large upward bias (Angrist, Joshua and Jorn-Steffen Pischke, 2009). Unlike some previous studies, I do not use
movements on non-event days in constructing the standard errors. Under the identifying assumption that no other
aggregate shocks occur during the event window, the standard errors used here inform whether the monetary policy
shock has a statistically significant systematic effect on the asset prices studied.
22
The North American IG CDX, published by Markit, follows 125 North American single label investment grade
entities, chosen based on liquidity. The index rolls every six months, meaning on March 20 and September 20
of each year Markit launches a new index tracking CDS with a maturity date five years hence. The on-the-run
CDX is the mostly recently launched index and the most liquid.
23
The maturity of bonds in the sample may differ from the five-year maturity of Treasuries. On a value-weighted
basis, the median remaining time to maturity of life insurer bonds in the sample is 5.8 years.

19

Figure 3: Insurance company CDS price response to unconventional monetary policy surprises
Change in ontherun 5 year Treasury, basis points
20
10
0

30

10

Change in CDS, basis points

10

20

30
ALL
MET

LNC
PRU
9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

7/10/13
1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

3/18/09

12/16/08

40

Announcement date

Notes: The change in CDS is the change in the 5 year spread, Tokyo close to London close (announcement before
12:15pm) or London close to New York close (announcement after 12:15pm), on the announcement date. The
change in the on-the-run 5 year Treasury is the change in the yield to maturity from 2 minutes before to 18
minutes after the announcement.

Figure 4: Insurance company bond price response to unconventional monetary policy surprises
Change in ontherun 5 year Treasury, basis points
20
10
0

Change in bond yield, basis points

30
150
100

AEG
AMP
LNC
PFG

AFL
GNW
MET
PRU

10

ALL
HIG
MFC

50
0
50
100

9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

12/16/08

3/18/09

150

Announcement date

Notes: The change in bond yield is over the shortest window containing at least 2 minutes before to 18 minutes
after the announcement. The change in the on-the-run 5 year Treasury is the change in the yield to maturity
from 2 minutes before to 18 minutes after the announcement. To enhance readability, the figure omits: 12/1/2008
(HIG[+346]), 12/16/2008 (GNW[-232]), 1/28/2009 (GNW[-466]), 8/9/2011 (PRU[-298]), 9/13/2012 (PRU[-239]).

20

Table 2: Event study effects of unconditional monetary policy

Treasurya

Life insurers
CDSb

12/1/08
12/16/08
1/28/09
3/18/09
9/23/09
8/10/10
9/21/10
8/9/11
1/25/12
9/13/12
5/22/13
6/19/13
7/10/13
9/18/13

9.2
16.8
3.1
22.8
8.9
5.8
1.8
14.4
6.3
6.4
6.6
7.8
7.3
14.0

Initial QEg 39.7

Taperh
14.4
Sample endi 21.4

7.2
3.1
24.5+
0.2
0.8
2.5
1.6
0.9
2.7+
1.1
1.6
1.5
2.1
31.7
0.5
3.6

Bondc
52.6
42.9
7.2
33.9
3.3
25.1
10.5
96.3
15.4
39.4
5.5
13.5+

Bank holding companies

8.3

Stockd
0.4
3.6
1.2
4.0
0.6
0.8
0.6
2.0
0.6
1.3
0.4
0.1
0.3
0.4

72.8
18.6
8.3

7.6
0.3
0.5+

CDSb

Market

0.7
9.4
2.6
0.2
0.2
0.2
1.3
1.1
4.1
1.8
2.9
0.4
0.9

Bondc
18.1+
9.1
0.9
16.6
12.5
1.5
16.1
5.5
2.6
10.5
5.5
24.4
5.2
33.5+

Stockd
0.6
2.2
0.3
2.5
0.6
0.9
0.7
1.7
0.0
1.0
0.5
0.2

CDXe

0.9

3.9
0.4
3.7
2.5
0.5
1.1
2.8
2.3
7.0
2.8
7.4
0.5
5.7

3.3
0.8
0.5

25.0
28.7
37.0

4.5
0.4
0.9

7.6
4.6
5.2

Stockf
0.6
1.3
0.3
1.5
0.6
0.7
0.5
1.4
0.4
0.6
0.5
0.2
0.2
1.0
2.8
0.6

Notes: +,*,** indicate significance at the 0.1, 0.05, 0.01 levels, respectively, based on the larger of the conventional
or robust standard error from a regression of the change in the asset price on a constant on the date indicated.
a. Change in the on-the-run five year Treasury note during announcement window covering two minutes before to
eighteen minutes after announcement, basis points.
b. Value-weighted mean change in the CDS spread, Tokyo close to London close (announcement before 12:15pm)
or London close to New York close (announcement after 12:15pm), 5 year tenor, basis points.
c. Value-weighted mean change in the bond yield during announcement window covering at least two minutes
before to eighteen minutes after announcement, basis points.
d. Value-weighted mean change in the log stock price during announcement window covering two minutes before
to eighteen minutes after announcement, log points.
e. Log change in the on-the-run North American IG CDX index, Tokyo close to London close (announcement
before 12:15pm) or London close to New York close (announcement after 12:15pm). Significance thresholds not
available.
f. Value-weighted mean change in the log stock price of companies, excluding life insurers and bank holding
companies, in the S&P 500, or change in the log of the S&P 500 index (September 2013) during announcement
window covering two minutes before to eighteen minutes after announcement, log points.
g. 12/16/2008 and 3/18/2009. Totals may differ due to rounding or sample composition.
h. 5/22/2013 and 6/19/2013. Totals may differ due to rounding or sample composition.
i. 7/10/2013 and 9/18/2013. Totals may differ due to rounding or sample composition.

21

Figure 5: Insurance company stock price response to unconventional monetary policy surprises
Change in ontherun 5 year Treasury, basis points
20
10
0

Log change in stock price

30
.1

AB
ALL
HIG
MFC
SLF

.05

10
AEG
AMP
LNC
PFG

AFL
GNW
MET
PRU

9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

7/10/13
1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

12/16/08

3/18/09

.05

Announcement date

Notes: The log change in stock price is from 2 minutes before to 18 minutes after the announcement. The change
in the on-the-run 5 year Treasury is the change in the yield to maturity during the same window.

changes in net investment and net derivative gains (MetLife, Inc., 2010, p. 6). The report goes
on to attribute the investment gain to a decrease in impairments and a decrease in the provision
for credit losses on mortgage loans. The improvement in legacy assets accounted for fully half
of the increase in pre-tax operating income MetLife experienced in 2010.
Subsequent announcements concerning unconventional policy had a more muted effect. To
help quantify the difference, table 3 reports regressions of the value-weighted mean asset price
response on the change in the 5 year Treasury, allowing for separate slope coefficients during
winter 2008-09 and thereafter.24 The coefficients are signed such that a contractionary surprise
corresponds to a positive realization of the right hand side variable. As foreshadowed by the
previous discussion, during winter 2008-09 a 10 basis point expansionary surprise results in an 8
basis point decline in both the CDS and the bond yield, and a stock price decline of 1.7 log points,
with both the CDS change and the stock price change highly statistically significant. In contrast,
the CDS and stock price coefficients for announcements after winter 2008-09 fall to essentially
zero, and none of the asset prices exhibits a response statistically distinguishable from zero. To be
24

Results are essentially unchanged using the unweighted mean or median price response instead. Winter 2008-09
includes all event dates in December 2008-March 2009.

22

Table 3: Value-weighted mean price response regressions

Life insurers

Bank holding companies

Dependent variable: change in

CDS

Bond

Stock

CDS

Bond

Stock

(1)

(2)

(3)

(4)

(5)

(6)

Right hand side variable:

10 b.p. Treasury X winter 2008-09
10 b.p. Treasury X post winter 2008-09
P value of coefficient equality
R2
Observations

7.90 7.73 1.68

(1.36) (13.70) (0.38)
0.26
17.53
0.28
(1.43) (14.79) (0.41)
0.00
0.61
0.00
0.77
0.13
0.66
13
13
14

0.21 0.34
(1.21) (4.84)
1.04
1.75
(1.26) (5.18)
0.62
0.75
0.06
0.01
13
14

0.91
(0.35)
0.11
(0.38)
0.06
0.42
13

Notes: The dependent variable is the value-weighted mean change in the log stock price during announcement
window covering two minutes before to eighteen minutes after announcement, in log points, the value-weighted
mean change in the CDS spread, Tokyo close to London close (announcement before 12:15pm) or London close to
New York close (announcement after 12:15pm), 5 year tenor, in basis points, or the value-weighted mean change
in the log bond price during announcement window covering at least two minutes before to eighteen minutes after
announcement, in log points, as indicated by the table header. The variable 10 b.p. Treasury is the change in the
yield to maturity of the on-the-run 5 year Treasury from 2 minutes before to 18 minutes after the announcement,
in 10 basis point increments. Winter 2008-09 includes all announcements in December 2008-March 2009. Standard
errors in parentheses. +,*,** indicate significance at the 0.1, 0.05, 0.01 levels respectively.

sure, these regressions have only a handful of observations, and it is possible market participants
had greater difficulty interpreting later announcements during the narrow windows.25 Indeed,
table 2 shows many later individual announcements still had significant, if small, effects.
The announcement surprises in 2013 merit special mention. If market participants underwent
a learning process of the effects of unconventional policy on financial institutions, these dates
should reflect the maturation of that process and hence contain more precise signals of the actual
effect of the policies. As well, during the spring of 2013 financial markets shifted their beliefs
toward an earlier taper of the Federal Reserves most recent round of asset purchases. The
response to the contractionary policy surprises contained in Chairman Bernankes congressional
testimony in May and in the FOMC statement in June of 2013 thus provide information on the
25

The post period includes two dates where the immediate response of the S&P 500 differed from that of the
five-year Treasury yield. The first occurred on August 9, 2011, when the FOMC introduced calendar-based forward
guidance. While Treasury yields fell immediately, the S&P 500 initially fell as well before reversing and ending
the day higher. The opposite occurred on September 13, 2012, with the introduction of a new round of QE. While
the S&P 500 rose immediately, Treasury yields also rose during the announcement window only to reverse the
increase by the end of the day.

23

symmetry of the markets response.

Beginning with the taper surprises, the Treasury yield rose a combined 14.4 basis points during
the two announcement windows.26 CDS spreads rose a statistically insignificant 0.5 basis point
over the two dates, while bond yields rose a statistically significant 19 basis points and stock
prices fell 0.3 log point.
I identify two expansionary policy surprises from the summer of 2013, a speech by Chairman
Bernanke at the NBER and the September FOMC statement. Summing over the two events gives
a decline in the five-year Treasury of 21 basis points. The value-weighted mean CDS spread of
life insurers declined on both dates. Of the ten life insurers with bond transactions surrounding
the September FOMC statement, nine experienced a decline in their yield, with a value-weighted
mean decline of 8 basis points.27 The mean stock price rose on both dates.
To summarize, the event studies appear consistent with general equilibrium effects strong
enough to ensure a positive, or at worst neutral, effect of unconventional monetary policy on life
insurers. Many life insurers faced solvency concerns in early 2009, and the expansionary policy in
the winter of 2008-09 appears to have had a substantial beneficial effect. This conclusion bears
resemblance to previous work finding that lower grade corporate bond prices reacted positively
to the initial round of QE due to a decline in default risk (Krishnamurthy and Vissing-Jorgensen,
2011), and that unconventional policy lowered downside tail risk in a broad class of asset prices
(Hattori et al., 2013; Roache and Rousset, 2013). Subsequent policy announcements had smaller
or neutral effects. Even in late 2013, however, market participants continued to view expansionary
monetary policy as beneficial to life insurers.

3.2.

Bank holding company results

The regulated banking sector remains at the core of the financial system in the United States.
It also contains nearly all of the U.S. institutions that have received the systemically important
26

The 5 year Treasury constant maturity yield rose by 95 basis points between May 1 and July 5, 2013. However,
much of the movement came in response to stronger than expected economic data releases rather than to policy
announcements. For example, the yield rose 21 basis points on July 5 following the release of the June employment
report by the Bureau of Labor Statistics on that date.
27
I do not have bond yield observations for the NBER speech because the speech took place after normal trading
hours.

24

Figure 6: Bank CDS

5 year tenor, basis points
800
ALLY
C
JPM
USB

BOFA
GS
MS
WFC

600

400

200

0
Jan08

Jan09

Jan10

Jan11

Jan12

Jan13

financial institution designation. Bank holding companies stood to benefit from the general
equilibrium effects of unconventional monetary policy on loan repayment and recovery rates, as
well as on the price of legacy securities on their balance sheets. Discussed already, a decline in
the safe interest rate may also lead to higher leverage.
Figure 6 presents CDS spreads for the sample of eight bank holding companies. Spreads rose
sharply during the crisis, and again in late 2011 amid concerns over the U.S. debt ceiling and
sovereign defaults in Europe. Like life insurers, as of the end of 2013 spreads had returned to
pre-crisis levels.
Figures 7 to 9 show scatterplots of the event study results for bank holding companies. Bank
holding companies also benefited from the introduction of unconventional monetary policy. Summing over the December 16, 2008 and March 18, 2009 announcements, seven of the eight banks
in the CDS sample (all except Citigroup [ticker: C]) experienced a decline in their CDS spread.28
The value-weighted stock price increase of 4.5 log points exceeded that of the average nonfinancial
28

This result contrasts somewhat with the small effects on banks CDS found in Gilchrist and Zakrajsek (2013).
The quasi-intraday windows may explain the difference. For example, using the quasi-intraday spread, table 2
shows a statistically significant decline of 3.3 basis points in response to the announcements on December 16, 2008
and March 18, 2009. Using the end of day spreads reported by Markit instead gives a statistically significant
increase of 5.3 basis points on the two announcement days.

25

Figure 7: Bank CDS price response to unconventional monetary policy surprises

Change in ontherun 5 year Treasury, basis points
20
10
0

Change in CDS, basis points

30
20
ALLY
C
JPM
USB

10

10

BOFA
GS
MS
WFC

10

9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

7/10/13
1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

12/16/08

3/18/09

20

Announcement date

Notes: The change in CDS is the change in the 5 year spread, Tokyo close to London close (announcement before
12:15pm) or London close to New York close (announcement after 12:15pm), on the announcement date. The
change in the on-the-run 5 year Treasury is the change in the yield to maturity from 2 minutes before to 18
minutes after the announcement.

Figure 8: Bank bond price response to unconventional monetary policy surprises

Change in ontherun 5 year Treasury, basis points
20
10
0

Change in bond yield, basis points

30
150

10

Median
100
50
0
50
100

9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

7/10/13
1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

12/16/08

3/18/09

150

Announcement date

Notes: The change in bond yield is over the shortest window containing at least 2 minutes before to 18 minutes
after the announcement. The change in the on-the-run 5 year Treasury is the change in the yield to maturity
from 2 minutes before to 18 minutes after the announcement. To enhance readability, the figure omits: 12/1/2008
(CIT[+168]), 12/16/2008 (CIT[-370], BPOP[+688]), 1/28/2009 (CIT[-279], COF[-210]), 3/18/2009 (CIT[-576],
COF[-300]), 9/21/2010 (BK[-326]), 5/22/2013 (ASBC[+307]), 6/19/2013 (BAC[-203], TRV[+171]).

26

Figure 9: Bank stock price response to unconventional monetary policy surprises

Change in ontherun 5 year Treasury, basis points
20
10
0

30

10

Log change in stock price

.05

.05

Median
9/13/12
5/22/13
6/19/13

1/28/09

9/21/10

1/25/12
8/10/10

12/1/08
9/23/09

8/9/11
9/18/13

12/16/08

3/18/09

.1

Announcement date

Notes: The log change in stock price is from 2 minutes before to 18 minutes after the announcement. The change
in the on-the-run 5 year Treasury is the change in the yield to maturity during the same window.

firm in the S&P 500. The average bond yield fell 25 basis points.
Once again, subsequent announcements had smaller effects on banks. And like life insurers,
the sign of the response remains unchanged even in 2013: CDS prices rise on Taper days and fall
on non-Taper days, and the opposite for stock prices. Unconventional monetary policy does not
appear to raise concerns about the health or riskiness of regulated banks.

4.

Money market funds

Money market funds provide liquidity services to institutional and retail clients. The interac-

tion of three features of money market funds in the United States makes them a potential concern
for financial stability at low nominal interest rates. First, money market funds maintain a stable
net asset value of one dollar per share. They do so by valuing assets at amortized cost and providing daily dividends as securities progress toward their maturity date.29 Investors can redeem
29

An example, adapted from Investment Company Institute (2011), helps to clarify. Treasury bills sell as
discount securities, meaning a 91 day Tbill with a face value of $100 and an interest rate of 1.2% will pay no
1
coupons and sell at auction for $99.70 = $100 (1 0.012) 4 . A money market fund that acquired the security on

27

shares at the par net asset value even if the shadow market value has fallen below. This feature
makes funds subject to runs.30 Second, SEC rule 2a-7 imposes duration, risk, and concentration
limits on a funds asset holdings. However, funds choose investments subject to these limitations.
Third, money market funds charge fees, also called expense ratios, typically on a pro rata basis.
The expense ratios do not affect the net asset value calculation, which depends solely on the
amortized value of the funds security holdings. However, they do affect the nominal net return
of investing in the fund.
In normal conditions, the spread between the return on funds assets and the interest rate
on checking accounts easily accommodates the expense ratios. When nominal interest rates
approach zero, however, the gross yield on funds assets may fall short of their normal charged
expenses. An after-fee return below zero would prompt investors to move into hard currency or
bank deposit accounts that do not charge fees. A funds sponsor can suspend the fees, implying
an operational loss to the sponsor of keeping the fund open. Alternatively, funds may seek higher
yield investments within the allowed asset classes reach for yield to avoid having to waive
fees. If the additional risk causes some of a funds securities to lose even a small amount of value,
the fund may have to break the buck, causing a broad run on money market funds similar
to what ensued following the Reserve Primary Funds breaking the buck in September 2008.31
Importantly, a single fund will not internalize the social costs of a broad run in the event its
the auction date would book the security at $99.70. Under straight line amortization, on each day until maturity
the booked value of the security would rise by $100$99.70
= $0.0033. The fund would balance the increase in the
91
value of its assets by increasing its daily dividend by the same $0.0033, thereby maintaining the stable net asset
value of $1 dollar per share.
30
The so-called penny rounding rule requires a funds board of directors to consider repricing the funds shares
(break the buck) if the shadow market value falls 0.5% below the par net asset value. Funds must calculate
their shadow market value on a periodic basis, with the interval determined by the board of directors. Thus the
shadow price can fall well below 0.995 before the fund suspends redemptions at the par value. Furthermore, many
of the assets held by funds, especially prime funds, do not have liquid secondary markets. If the fund must sell
assets to satisfy redemptions, the market value may fall further as fire sale prices generate additional capital losses.
See Securities and Exchange Commission (2013, pp. 14-19) for further discussion of the mechanics of a run on a
money market fund.
31
McCabe (2010) and Kacperczyk and Schnabl (2013) study the behavior of prime money market funds between
the onset of the subprime crisis and the Lehman Brothers bankruptcy. The subprime crisis sparked a revaluation
of risk and an opening of yield differentials among eligible AAA securities. Some funds responded by investing in
higher yield securities within the AAA class, prompting institutional investors to reallocate their investments to
higher yield funds and further increasing the incentive for a fund to reach for yield. For example, beginning in
August 2007 the Reserve Primary Fund offered a yield of roughly 20 basis points higher than competitor funds, in
part by purchasing large quantities of Lehman Brothers commercial paper, generating large inflows into the fund
until the Lehman bankruptcy forced it to break the buck (Kacperczyk and Schnabl, 2013, figure III).

28

Table 4: Money market fund summary statistics

Statistic
Mean

P90

Obs.

0.54
0.14

Charged expense ratio, annual average

0.27
0.20
0.53
0.93
0.06
0.07
0.13
0.22

685
469

0.65
0.57

Incurred expense ratio, annual average

0.33
0.27
0.62
1.03
0.35
0.23
0.52
0.98

685
469

2006
2013

4.40
0.16

7 day gross simple yield, annual average

0.82
3.39
4.92
5.06
0.06
0.09
0.14
0.25

685
469

2006
2013

0.33
0.03

7 day gross simple yield, standard deviation

0.12
0.23
0.35
0.39
0.01
0.02
0.03
0.04

685
469

3.94
0.02

7 day net compound yield, annual average

0.90
2.73
4.30
4.90
0.02
0.00
0.01
0.04

685
469

2006
2013

2006
2013

2006
2013

Std. dev.

P10

P50

Notes: The table reports cross-sectional statistics, by year, of time-series properties at the fund level.

additional risk exposure causes its assets to lose value.

I begin by describing how money market funds have adjusted their fees. Data on asset holdings,
yields, and administrative costs come from iMoneyNet. These data cover the universe of domestic
money market funds.32 I follow Kacperczyk and Schnabl (2013) and aggregate asset holdings,
yields, and expenses up to the fund level using share class asset shares as weights. Table 4 displays
summary statistics for 2006 and 2013.
Figure 10 shows a scatterplot of incurred expenses (horizontal axis) and charged expenses
(vertical axis). Incurred expenses are meant to reflect the cost of running the fund, including
management fees and advertising, while charged expenses are the fees actually paid by investors.
Data points below the 45 degree line indicate funds that have waived part of their fees. The red
dots show the relationship in 2006. Most lie on the 45 degree line or slightly below. The data also
32

See figure C.2 for a comparison to the Financial Accounts of the United States.

29

Figure 10: Money market fund expense ratios, 2006 and 2013
Charged expenses as percent of assets
2
2006
2013
1.5

.5

0
0

.5

1
1.5
Incurred expenses as percent of assets

2.5

Notes: To enhance readability, the figure omits one 2013 fund with $6 million under management, incurred expenses
of 4.6 percent, and charged expenses of 0.04 percent.

indicate substantial dispersion in fees charged, with a mean of 0.54 and a standard deviation of
0.27 (table 4 top panel). The blue triangles show the relationship in 2013 and almost all lie well
below the 45 degree line. They also show a substantial decline and compression in fees charged
in 2013, with the mean falling to 0.14 and standard deviation to 0.06. The bottom panel of
table 4, labeled 7 day net compound yield, sheds further light. The panel reports summary
statistics for the average net (after-fee) yield, annualized, an investor would earn if she reinvested
the dividends received each week. In 2013, the average yield was two basis points, with a median
of 1 and a 90th percentile of 4 basis points.
The compression of net yields and ubiquity of fee waivers suggest funds responded to the
low interest rate environment by waiving exactly enough of their fees to ensure their investors
received a non-negative nominal return. I confirm this interpretation by estimating the following
regression by fund share class and at a weekly frequency:





Charged expensesi,t = t + t Gross yieldi,t + t Incurred expensesi,t + ei,t .

30

(5)

Figure 11: Determinants of money market fund charged expenses

Loading
1

.8

Gross yield

.6

.4
Incurred expenses
.2

0
Jan06

Jan07

Jan08

Jan09

Jan10

Jan11

Jan12

Jan13

Notes: The solid lines plot the weekly coefficients from an OLS regression of charged expenses on gross yield
and incurred expenses. Equation (5) provides the estimating equation. The dotted lines plot 95% confidence
interval bands based on standard errors clustered by fund sponsor. The regression winsorizes observations with
the smallest and largest 0.01 percent of incurred expenses.

Equation (5) allows non-parametrically for time-varying loadings of charged expenses on the gross
(before fee) yield (the {t }) and on the costs of running the fund (the {t }). Inclusion of weekly
fixed effects limits identification to coming from variation across funds in a given week.33 Figure 11
reports the estimated {t } in blue and {t } in red. Prior to late 2008, an additional marginal basis
point of gross yield has essentially no effect on the charged fee. In contrast, a percentage point
increase in incurred fees corresponds to an increase of roughly 80 basis points in fees charged.
The relationship reverses completely after gross yields fall to close to zero. Throughout 2010 to
2013, an additional percentage point in gross yield corresponds to an increased in charged fees of
90 basis points, while the marginal basis point of incurred expenses has no effect on charged fees.
The almost complete pass-through of higher gross returns to higher charged fees suggests
funds operators had cognizance of the cost of waiving fees and the potential to avoid such costs
33

I have also estimated a specification replacing the weekly fixed effects with fundXyear fixed effects, such that
identification comes from variation within a single fund over the course of the year. I obtain coefficients of 0.86,
0.84, and 0.90 for the loadings on gross yield in 2011, 2012, and 2013, respectively, up from 0.00 in 2007. These
loadings mirror almost exactly those shown in figure 11. Absence of within fund variation in incurred expenses,
however, renders the loadings on incurred expenses very imprecisely estimated with the fund fixed effects.

31

by generating higher gross returns reaching for yield.34 I therefore test in the cross section of
funds whether funds with higher structural administrative costs reached for yield.
I examine four measures of reaching for yield: the gross yield, the ex-post realized standard
deviation of monthly excess returns, the share of holdings in foreign bank obligations net of
repurchase agreements (repo), and the average asset maturity. The gross yield captures directly
whether a fund has successfully reached for yield. The ex-post standard deviation provides a
measure of risk. Figure 12 below shows foreign bank obligations to be the highest yield asset
class during most of the unconventional monetary policy period, and repo the lowest. The average
asset maturity provides a measure of reaching for yield through reaching for maturity.35
I assume a data generating process for each measure of the form

yi,t = i + t +

2013
X

I {t Y } Y [Administrative costs]i,t +

Y0 xi

+ ei,t .

(6)

Y =2006

The identifying assumption in equation (6) is that funds with high and low management costs do
not differ along other features that would make them respond differently to low nominal interest
rates. The fund fixed effect i absorbs time-invariant unobserved fund characteristics such as
managerial skill. The time effect t controls for variation in the macroeconomic environment such
as short term interest rates. The process (6) allows for potentially time-varying loadings Y on the
observed fund characteristics of fund category (tax-free, prime, or U.S. government and agency
securities and backed repo) and whether the fund has any institutional shares. Time-varying
loadings may matter if, for example, prime funds perform relatively better in certain years because
of higher yields on the assets available to prime funds relative to government securities in those
years. Finally, because incurred expenses may proxy imperfectly for administrative costs and
because funds may endogenously adjust incurred expenses in the low interest rate environment,
I instrument for a funds incurred fees using the funds 2005 incurred fees.
Table 5 reports results using weekly data for a balanced panel of funds over 2006-2013. I
34

The apparent negative operational profits raises the question of why funds remain in business. In fact, the
industry has experienced significant exit since the crisis, as seen both in the decline in the number of funds by
year in table 4 and by the decline in total assets under management in figure C.2. Many of the remaining funds
receive cross-subsidization from other lines of business of their sponsors.
35
Kacperczyk and Schnabl (2013) also use the gross yield and average maturity measures. I adapt their measure
of risky asset holdings based on the asset class loadings during my sample period.

32

cluster standard errors at the fund sponsor level, thus allowing for both arbitrary serial correlation and arbitrary correlation across funds with the same sponsor. The table reports the first
stage F statistics for 2013 expenses, showing 2005 incurred expenses to be a strong instrument.
Identification of the unobserved characteristics {i } requires imposing Y = 0 for at least one
year, and I choose to set 2006 = 0. The near zero coefficients across specifications for Y in the
normal interest rate years of 2007 and 2008 both justify this restriction and serve as a useful
placebo check.
Table 5 provides some evidence of money market funds reaching for yield in 2009-2011. However, the effects appear economically quite small. For example, the 2010 coefficient for the gross
yield measure achieves statistical significance at the 1 percent level, but has the interpretation of a
one percentage point (or roughly three standard deviation) increase in administrative costs resulting in an additional five basis points of annualized gross yield. Likewise, 2011 incurred expenses
have a statistically strong effect on the standard deviation of returns, but a one percentage point
increase in expenses still results in an increase of the standard deviation of just 1.7 basis points.
Both the gross yield and standard deviation measures indicate a precise zero effect of incurred
expenses on reaching for yield in 2013, the most recent data available. The asset allocation and
maturity measures indicate no differences in allocations among funds with high and low costs.
What explains the relative absence of reaching for yield compared to during the subprime
crisis, and the apparent complete dissipation in 2012 and 2013? Figure 12 displays 3-month
centered moving averages of coefficients from a weekly regression of gross yield on the allocation
to each asset class. As stressed by Kacperczyk and Schnabl (2013), the subprime crisis created
large return differentials from investing in different asset classes. Many prime funds responded
by concentrating holdings in higher yield classes. These differentials compressed substantially
beginning in mid 2009, and by 2013 had reached historically low levels. Such small differentials
provide little opportunity for prime funds to reach for yield through asset class allocation. Of
course, reaching for yield could return if spreads open up again, for example due to renewed
sovereign risk concerns in Europe. Even so, the first-best policy response should involve further
reform of the money market sector to remove the threat of runs, rather than changing monetary
policy. Indeed, a second explanation of the 2012 and 2013 results stems from an initial set of
33

Table 5: Money market fund reaching for yield regressions

Dependent variable:
Gross
yield
(1)

Std. dev.
return
(2)

Risky asset
allocation
(3)

Average
maturity
(4)

0.006
(0.011)
0.049
(0.044)
0.171
(0.076)
0.053
(0.016)
0.036
(0.022)
0.018
(0.019)
0.001
(0.012)
2006-13
Yes
Yes
No
81.2
379
76
151,600

0.007
(0.011)
0.013
(0.013)
0.047
(0.042)
0.011
(0.008)
0.017
(0.005)
0.010
(0.009)
0.001
(0.005)
2006-13
Yes
No
Yes
156.8
379
76
3,032

1.036
(2.603)
6.197
(3.977)
1.075
(6.910)
3.168
(7.615)
8.447
(9.400)
12.192
(9.586)
8.513
(9.384)
2006-13
Yes
Yes
No
80.2
135
65
54,000

1.148
(1.829)
0.285
(2.215)
0.688
(2.976)
0.793
(2.025)
0.970
(2.886)
4.609
(3.473)
3.567
(3.401)
2006-13
Yes
Yes
No
98.8
379
76
151,533

Right hand side variables:

2007 incurred expenses (IV: 2005 value)
2008 incurred expenses (IV: 2005 value)
2009 incurred expenses (IV: 2005 value)
2010 incurred expenses (IV: 2005 value)
2011 incurred expenses (IV: 2005 value)
2012 incurred expenses (IV: 2005 value)
2013 incurred expenses (IV: 2005 value)
Sample period
Fund FE
Week FE
Year FE
2013 first stage F statistic
Unique funds
Fund sponsor clusters
Observations

Notes: The gross yield is the weekly gross simple yield, annualized. The standard deviation is the annual
standard deviation of the monthly excess return, defined as the gross return less the 1 month TBill. The risky
asset allocation equals the asset share in foreign bank obligations less the share in repo, and for this regression the
sample excludes non-Prime funds. The average maturity refers to the average of the maturity of a funds securities,
in days. Regressions also include categorical variables, interacted with year, for fund category (tax free, prime, or
U.S. government and agency securities and backed repo) and for whether the fund has any institutional shares.
Regressions exclude observations between September 1 and December 31, 2008. Standard errors in parentheses
and clustered by fund sponsor. +,*,** indicate significance at the 0.1, 0.05, 0.01 levels respectively.

reforms implemented by the SEC in mid-2010. These included reducing the allowable fraction
of assets in illiquid securities, the weighted-average maturity of assets, the fraction of assets in
second tier securities, and the concentration limit for securities issued by any single issuer.
If the link between charged fees and gross yields has not produced much reaching for yield
by money market funds, it does still have a potentially important implication for constraints on
34

Figure 12: Money market fund yields by category

Excess return over U.S. Treasuries, basis points
200
USOther
Repos
Time Deposits
Domestic Bank Obligations
Foreign Bank Obligations
FRNS
Non ABCP Commercial Paper
ABCP

150

100

50

50
Jan06

Jan07

Jan08

Jan09

Jan10

Jan11

Jan12

Jan13

Notes: The figure plots the 3-month centered moving average of coefficients from a weekly regression of gross yield
on the asset allocation in the categories shown. U.S. Treasuries are the omitted category, and the sample includes
all taxable money market funds. The regression also includes a fund fixed effect allowed to vary by calendar year.

Federal Reserve policy. Even with the introduction of the extraordinary policy measures described
above, the FOMC since 2009 has maintained a target federal funds rate of between 0 and 25 basis
points and paid interest on excess reserves of 25 basis points. One justification for not reducing
both of these rates further has centered on the risk of disruptive outflows from money market
funds if a drop in rates to zero forced a negative net yield. Yet, funds already subsidize investors
by waiving fees. The evidence in figure 11 suggests funds might respond to a further decline
in yields by simply waiving fees completely, leaving investors with the same net yield of zero
they currently receive. Such a move might prompt further consolidation of the industry, but
consolidation need not be disorderly if it involves funds proactively announcing their withdrawal
from an unprofitable business rather than investors running on funds.

35

5.

Pension funds
Private defined benefit pension funds manage roughly $3 trillion in retirement assets. Similar

to life insurers, promised benefit payments may have longer duration than available assets, leading
to balance sheet duration mismatch. As well, funds with large numbers of current beneficiaries
have less time to smooth the decline in asset values during the crisis. Figure 13 shows the
funded status of defined benefit plans using data from the Federal Reserve Financial Accounts of
the United States. The fraction of defined benefit liabilities unfunded by asset holdings gives a
measure of the solvency concerns of pension funds.
Figure 13: Unfunded liabilities as share of total, defined benefit plans
Percent
40
Private
State and local

20

20

40
1990

1995

2000

2005

2010

2015

Notes: Negative value indicates over funding.

Source: Federal Reserve Financial Accounts of the United States (tables L117.b. and L118).

I collect data on private pension funds from the form 5500 reports funds must file with the
IRS. Plans with 100 or more participants must file a schedule H, containing a detailed asset and
income statement, and a schedule B (or SB or MB after 2009) showing their funding status.
Appendix A contains further details of the sample construction and filters.
I construct two measures of fund risk taking, similar to those used in the analysis of money
market funds. The first uses the funds reported investment earnings to construct an annual
36

Table 6: Defined benefit pension fund summary statistics

S&P 500
Year
2006
2007
2008
2009
2010
2011
2012

return
13.6
3.5
38.5
23.5
12.8
0
13.4

Fund return
Mean
11.2
7.2
22.1
18.3
11.6
1.7
11.3

St. dev.
2.6
2.5
7.3
6.1
2.8
3.9
2.9

Expenses/Assets
Mean
0.06
0.06
0.06
0.08
0.07
0.06
0.07

St. dev.
0.03
0.03
0.03
0.04
0.03
0.03
0.04

Benefits NPV/Assets
Mean
1.13
1.1

St. dev.
0.22
0.22

1.16
1.19
1.24
1.09

0.18
0.16
0.17
0.16

Obs.
4225
4225
4225
4225
4225
4225
4225

Notes: The table reports cross-sectional statistics, by year. Data for funding status in 2008 are not available from
the Department of Labor.

return on assets, defined as total earnings on investments (including unrealized capital gains)
divided by the sum of beginning of period assets plus one-half of net contributions. The second
measure divides the sample into two periods, 2004-08 and 2009-12, and constructs the standard
deviation of the funds return in each period. Table 6 reports summary statistics for the balanced
panel. Of note, the mean return tracks the return on the S&P 500 reasonably closely.
I investigate two motivations for pension funds to reach for yield. First, if funds reduce risk
exposure as duration decreases, then funds with large current liabilities should on average exhibit
less risk taking. However, having large current liabilities in a period of low interest rates following
a market downturn may lead to increased risk taking as these funds have less time to make up
any funding shortfall. I measure liability maturity horizon using the ratio of a funds benefits
expenses to its total assets. Letting ri,t denote fund is return in period t, equation (7) describes
a difference-in-difference specification for estimating the effect of liability timing on reaching for
yield:

ri,t =

t +t0 Zi +





 e  Benefit payments
 e  Benefit payments
rm,t
+I {t > 2006} t rm,t
+ei,t ,
Assets
Assets
i,t
i,t
(7)

e
where rm,t
is the Fama-French excess return of the stock market over the risk free rate. The

vector of controls Zi includes fund size and age.

Equation (7) defines reaching for yield as a funds increasing its loading on the market excess

37

return.36 gives the estimated loading per unit of benefits/expenses for 2006. < 0 if plans with
more current liabilities allocate to safer assets. 2007 2012 then ask whether plans with a shorter
liability horizon allocated relatively less to safer assets in the low interest rate environment.
The difference-in-difference specification for the standard deviation of returns has the form

(ri,t ) =

0
tsd +tsd Zi + sd

Benefit payments
(rm,t )
Assets

sd

+0912
i,t

Benefit payments
(rm,t )
Assets

+sd
i,t .
i,09

(8)
Equation (8) implements a difference-in-difference specification of the ex post standard deviation
of fund returns, over 2004-08 and 2009-12, and for funds with varying liability horizon as of the
sd
gives the additional standard
first year of the period (i.e. 2004 or 2009). The coefficient 09

deviation of returns, scaled by the standard deviation of the market excess return, of funds with
a shorter horizon in 2009.
The first two columns of table 7 report the results corresponding to equations (7) and (8). The
coefficients and sd are both negative and highly statistically significant, confirming a negative
relationship between near-term liabilities and risk taking. The coefficients 2007 and 2008 cannot
reject a constant effect of liability duration on return before the onset of the low interest rate
environment.
The regressions provide some evidence of reaching for yield beginning in 2009. The effect of
liability duration on return falls by one-half to two-thirds in 2009 and 2010, and the differences
relative to 2006 are statistically significant at the 1 and 5 percent levels, respectively. Similarly,
the additional loading on the standard deviation of the market excess return falls by about twothirds for 2009-12 relative to 2004-08. While statistically significant, these differences do not
translate into particularly large economic effects. For example, the difference in the loading on
the market excess return in 2005 and 2009 for a fund one standard deviation (0.04) above the
mean of expenses/assets is 0.51*0.04=0.02. This difference corresponds to an additional return
of about 50 basis points on the 2009 excess return of 28 percent. The effect on return loading
36

According to table L.117.b of the Federal Reserves Financial Accounts of the United States, in aggregate
private defined benefit pension funds held between 50 and 70 percent of their assets in equities between 2005
and 2013. These plans held an additional 15 percent of their assets in the safe asset categories of deposits,
money-market funds, repo, open market paper, treasury securities, and agency securities.

38

Table 7: Defined benefit pension fund reaching for yield regressions

Funding measure:
Benefits expenses / Assets

Benefits NPV / Assets

Dependent variable (p.p.):

Return

(Return)

Return

(Return)

(1)

(2)

(3)

(4)

Right hand side variables:

e
X (Funding measure)
rm

0.67
(0.12)
0.08
(0.03)
0.56
(0.54)

e
X (Funding measure)
2006 X rm
e
X (Funding measure)
2007 X rm
e
2008 X rm
X (Funding measure)
e
X (Funding measure)
2009 X rm
e
2010 X rm
X (Funding measure)
e
2011 X rm
X (Funding measure)
e
X (Funding measure)
2012 X rm

0.31
(1.13)
0.10
(0.14)
0.52
(0.13)
0.32
(0.14)
0.41
(3.47)
0.08
(0.17)

e
(rm
) X (Funding measure)
e
2009 X (rm
) X (Funding measure)

Year FE
Size, age controls
Fund FE
e
Fund-specific rm
loading
Unique funds
Fund sponsor clusters
Observations

Yes
Yes
No
No
4,225
3,719
29,575

0.14
(0.02)
0.04+
(0.03)
5.20
(1.32)
0.05+
(0.03)
0.36
(0.11)
0.24
(0.11)
Yes
Yes
No
No
3,580
3,186
7,160

Yes
No
Yes
Yes
3,746
3,263
22,475

0.01
(0.02)
0.18
(0.03)
Yes
Yes
No
No
3,177
2,806
6,353

Notes: The pension return is the annual earnings on investments divided by the sum of beginning of year assets
plus one half net transfers and contributions. (Return) is the standard deviation of the pension return over
e
2004-08 or 2009-12. rm
is the Fama-French stock market excess return. Benefits NPV/assets are not available for
2008, and in all years specifications including this variable remove multiemployer (MB) plans. If included, size
(log assets) and age controls interacted with year. Observations with a distance to the median larger than five
times the interquartile range are winsorized. Standard errors in parentheses and clustered by fund sponsor. +,*,**
indicate significance at the 0.1, 0.05, 0.01 levels respectively.

39

loses statistical significance in 2011 and 2012.37

The second motivation for reaching for yield stems from pension funds with funding shortfalls
requiring a high return on assets to avoid raising contributions or insolvency. A low level of
interest rates exacerbates this problem by lowering the expected return absent additional risk
taking, suggesting underfunded pension funds may have reached for yield beginning in 2009. Low
interest rates also contribute directly to underfunded status by lowering the discount rate used
to discount future liabilities.38 I measure funding shortfall as the ratio of the net present value
of benefits to total assets. To isolate within-fund variation in funding status, I assume the data
generating process

e
e
ri,t = i + t + i rm,t
+ t rm,t
[Funded status]i,t1 + i,t .

(9)

Under the null hypothesis t = 0 t, and if t = rtrisk free , the process (9) collapses to a
single factor pricing model, with the market excess return the single factor. The factor pricing
model allows funds to have different permanent risk profiles through the fund-specific loading
i , and fund-specific managerial skill through the fixed effect i . The time effect t controls for
macroeconomic conditions such as the risk free rate, as well as reporting changes in the measure
of funding status that occurred throughout the sample period. Identification of requires zero
serial correlation of the residuals i,t .39 Finally, the time-varying coefficient on t allows funding
37

The large standard error in 2011 reflects the near zero Fama-French excess return (0.44) in that year.
Under the Pension Protection Act of 2006 (PPA), single employer plans must use the two year average of the
yield on investment grade corporate bonds to discount future liabilities, with separate maturities of bonds used
for liabilities due in 0 to 5 years, 5 to 15 years, and longer than 15 years. A 2012 law temporarily changed the
horizon for averaging yields to 25 years, resulting mechanically in improved funding status.
The PPA modified four other aspects of the regulation of defined benefit plans relevant for the effect of funding
status on risk taking. First, it requires 100% funding of liabilities. Underfunded plans have seven years to
amortize their unfunded liability through higher contributions. Second, the PPA applies the penalty premium for
contributions to the Pension Benefit Guaranty Corporation to any underfunded plan. Third, it created the category
of at risk plans, containing plans in particularly poor funding status and subjecting them to additional required
contributions. Fourth, the PPA narrows the permitted difference between the actuarial value of a plans assets,
which may smooth investment returns, and the assets fair market value. These changes make comparisons of
funded status reported on the 5500 forms across years inappropriate, providing additional motivation for inclusion
of time fixed effects in the regressions reported in table 7.
39
Intuitively, if there is positive serial correlation in the error term, then a fund that does poorly one year
resulting in poor funding status will also do poorly the next, biasing down t , and vice versa if the serial correlation
is negative. Inclusion of the fund fixed effect i may exacerbate the identification problem for the same reason
panel data models cannot include both a fixed effect and lagged dependent variable. However, results remain
qualitatively unchanged with the fixed effect removed.
38

40

status to affect risk-taking differentially before and during the low interest rate environment.
The fund fixed effect and fund specific loading make equation (9) infeasible for estimating the
effect of funding status on return standard deviation with standard deviations computed over
only two periods. Instead, the specification for the effect of funded status on return standard
deviation mirrors the specification described in equation (8).
Columns 3 and 4 of table 7 report results for the effect of funded status.40 The differencein-difference specification for the return standard deviation in column 4 indicates no effect of
funding shortfall on return standard deviation during 2004-08, but a strong positive relationship
in 2009-12, consistent with some reaching for yield. Results for equation (9) are more mixed. I
find small positive effects of funding status on return in 2009, 2010, and 2012, but a large negative
effect in 2011.
In sum, the strongest evidence of pension funds reaching for yield comes in 2009, and occurs
robustly for both measures of fund status and for both the level and standard deviation of returns.
The effect on the loading on the market excess return declines thereafter or loses statistical
significance. Returning to figure 13, the explanation may lie in the improvement in funding status
as the stock market recovered. The motivation to reach for yield should diminish as funding gaps
improve, even if low interest rates persist. In aggregate, private defined benefit pension funds hold
more than half of their assets in equities, making the positive effects of unconventional monetary
policy on the stock market at least as important as any deleterious effects from low interest rates
spurring reaching for yield.

6.

Conclusion
The paper has investigated the effects of unconventional monetary policy on financial insti-

tutions. Using high frequency event studies, I find the introduction of unconventional policy in
the winter of 2008-09 had a strong, beneficial impact on banks and especially on life insurance
companies, consistent with the positive effect on legacy asset prices dominating any impulse for
40

Data for funding status in 2008 are not available. Data are missing in select other years for a few plans. Results
remain essentially unchanged when restricting to a balanced panel of funds not missing data in any non-2008 year.

41

additional risk taking. The interaction of low nominal interest rates and administrative costs
led money market funds to waive fees, producing a possible incentive to reach for higher returns
to reduce waivers. I find some evidence of money market funds reaching for yield in 2009-11,
but little thereafter. Private defined benefit pension funds with worse funding status or shorter
liability duration also seem to have reached for higher returns beginning in 2009, but again the
evidence suggests such behavior dissipated by 2012.
I conclude with some caveats to the analysis presented so far. First, if unconventional policy
has benefited life insurers and banks, then withdrawing monetary stimulus may have some detrimental impact. In the extreme, tightening might undo the increase in the value of legacy assets,
leaving only the deleterious effect of additional risk taking. The 2013 taper event dates provide
some reassurance here. Tightening adversely affected life insurers and banks, but the magnitudes
do not appear especially large or asymmetric. Extrapolating linearly from those event dates, an
increase of 150 basis points in the five year Treasury would cause CDS spreads of life insurers and
banks to rise by 5-10 basis points, and stock prices to fall roughly 3-4 percent.
A second concern relates to financial instability originating outside of the institutions studied.
Feroli et al. (2014) discuss the possibility of a small tightening of monetary policy generating a
coordinated withdrawal by asset fund managers from riskier assets, resulting in an increase in risk
premia. This mechanism has resemblance to the metaphor in Stein (2013b) of the Federal Reserve
having a looser grip on the steering wheel than it would prefer. Third, some emerging market
economies appear vulnerable to the effects of tightening, and history teaches that instability from
emerging markets can have global effects. Finally, money market fund reaching for yield could
again become a vulnerability if credit spreads widen.
In sum, in the present environment, if either unconventional monetary policy or exit from the
same carries the risk of fostering instability in the financial sector, it would appear such instability
would have to come from a channel other than through the effect on the financial institutions
studied here.

42

References
Adrian, Tobias and Hyun Song Shin, Financial Intermediaries and Monetary Economics, in
Benjamin M. Friedman and Michael Woodford, eds., Benjamin M. Friedman and Michael
Woodford, eds., Vol. 3 of Handbook of Monetary Economics, Elsevier, 2010, pp. 601 650.
American Council of Life Insurers, 2013 Life Insurers Fact Book, Technical Report 2013.
Angrist, Joshua and Jorn-Steffen Pischke, Mostly Harmless Econometrics, Princeton: Princeton
University Press, 2009.
Becker, Bo and Victoria Ivashina, Reaching for Yield in the Bond Market, Journal of Finance,
forthcoming.
Bernanke, Ben, Testimony before the Joint Economic Committee May 22, 2013, 2013.
Borio, Claudio and Haibin Zhu, Capital regulation, risk-taking and monetary policy: A missing
link in the transmission mechanism?, Journal of Financial Stability, 2012, 8 (4), 236 251.
Brunnermeier, Markus K. and Yuliy Sannikov, The I Theory of Money, 2011.
and

, A Macroeconomic Model with a Financial Sector, American Economic Review, 2014,

104 (2), 379421.

Campbell, Jeffrey, Charles Evans, Jonas Fisher, and Alejandro Justiniano, Macroeconomic Effects of Federal Reserve Forward Guidance, Brookings Papers on Economic Activity, 2012,
Spring, 180.
Chodorow-Reich, Gabriel, The Employment Effects of Credit Market Disruptions: Firm-level
Evidence from the 2008-9 Financial Crisis, The Quarterly Journal of Economics, 2014, 129
(1), 159.
DellAriccia, Giovanni, Luc Laeven, and Robert Marquez, Real interest rates, leverage, and bank
risk-taking, Journal of Economic Theory, 2014, 149, 65 99.
Drechsler, Itamar, Alexi Savov, and Philipp Schnabl, A Model of Monetary Policy and Risk
Premia, 2013.
English, William B., Skander J. Van den Heuvel, and Egon Zakrajsek, Interest Rate Risk and
Bank Equity Valuations, Finance and Economics Discussion Series 2012-26, Board of Governors of the Federal Reserve System 2012.
43

Feroli, Michael, Anil Kashyap, Kermit Schoenholtz, and Hyun Song Shin, Market Tantrums and
Monetary Policy, 2014. http://research.chicagobooth.edu/igm/usmpf/2014-file.aspx.
Fisher, Richard, Beer Goggles, Monetary Camels, the Eye of the Needle and the First Law of
Holes, Remarks before the National Association of Corporate Directors January 2014.
Gagnon, Joseph, Matthew Raskin, Julie Remache, and Brian Sack, Large-Scale Asset Purchases
by the Federal Reserve: Did They Work?, Staff report 441, Federal Reserve Bank of New York
2010.
Gilchrist, Simon and Egon Zakrajsek, Credit Spreads and Business Cycle Fluctuations, American Economics Review, 2012, 102 (4), 16921720.
and

, The Impact of the Federal Reserves Large-Scale Asset Purchase Programs on Cor-

porate Credit Risk, 2013.

Hanson, Samuel and Jeremy Stein, Monetary Policy and Long-Term Real Rates, 2012.
Harrington, Shannon and Andrew Frye, Lincoln Leads Increase in Insurer Credit Swaps After Downgrade, Bloomberg.com March 2009. http://www.bloomberg.com/apps/news?pid=
newsarchive&sid=aurl7R1EMtEA.
Hattori, Masazumi, Andreas Schrimpf, and Vladyslav Sushko, The Response of Tail Risk Perceptions to Unconventional Monetary Policy, 2013.
He, Zhiguo and Arvind Krishnamurthy, Intermediary Asset Pricing, American Economic Review, 2013, 103 (2), 73270.
International Monetary Fund, Global Financial and Stability Report, Technical Report April
2013.
Investment Company Institute, Pricing of U.S. Money Market Funds, Technical Report 2011.
Jensen, Michael C. and William H. Meckling, Theory of the firm: Managerial behavior, agency
costs and ownership structure, Journal of Financial Economics, 1976, 3 (4), 305 360.
Kacperczyk, Marcin and Philipp Schnabl, How Safe Are Money Market Funds?, The Quarterly
Journal of Economics, 2013, 128 (3), 10731122.
Kiley, Michael, Monetary Policy Statements, Treasury Yields, and Private Yields: Before and
After the Zero Lower Bound, 2013. Federal Reserve Finance and Economics Discussion Series
2013-16.

44

Koijen, Ralph and Motohiro Yogo, The Cost of Financial Frictions for Life Insurers, 2013. Avaiable at SSRN: http://ssrn.com/abstract=2031993orhttp://dx.doi.org/10.2139/ssrn.
2031993.
Krishnamurthy, Arvind and Annette Vissing-Jorgensen, The Effects of Quantitative Easing on
Interest Rates: Channels and Implications for Policy, Brookings Papers on Economic Activity,
2011, Fall, 215287.
McCabe, Patrick, The Cross Section of Money Market Fund Risks and Financial Crises, Finance
and Economics Discussion Series 2010-51 2010.
McKinsey Global Institute, QE and ultra-low interest rates, Technical Report November 2013.
MetLife, Inc., Annual Report, 2010.
Moodys Investors Services, US life insurers vulnerable to protracted low interest rates, Technical Report December 2012.
Nakamura, Emi and Jon Steinsson, High Frequency Identification of Monetary Non-neutrality,
2013. Unpublished manuscript: http://www.columbia.edu/~en2198/papers/realrate.pdf.
Rajan, Raghuram, Has Financial Development Made the World Riskier?, in Proceedings of
the Jackson Hole Conference 2005.
Roache, Shaun and Marina Rousset, Unconventional Monetary Policy and Asset Price Risk,
2013. IMF Working Paper 13/190.
Securities and Exchange Commission, Money Market Fund Reform Proposed Rule, 2013. http:
//www.sec.gov/rules/proposed/2013/33-9408.pdf.
Shleifer, Andrei and Robert Vishny, Fire Sales in Finance and Macroeconomics, Journal of
Economic Perspectives, 2011, 25 (1), 2948.
Stein, Jeremy, Monetary Policy as Financial Stability Regulation, The Quarterly Journal of
Economics, 2012, 127 (1), 5795.
, Overheating in Credit Markets: Origins, Measurement, and Policy Responses, Remarks at
Restoring Household Financial Stability after the Great Recession: Why Household Balance
Sheets Matter February 2013.
, Yield-Oriented Investors and the Monetary Transmission Mechanism, Remarks at Banking,
Liquidity and Monetary Policy, a Symposium Sponsored by the Center for Financial Studies
in Honor of Raghuram Rajan September 2013.
45

Werning, Ivan, Managing a Liquidity Trap: Monetary and Fiscal Policy, 2012.
Wright, Jonathan H., What does Monetary Policy do to Long-term Interest Rates at the Zero
Lower Bound?, The Economic Journal, 2012, 122 (564), F447F466.

46

Appendix

A.

Pension fund data description

The Department of Labor collects the form 5500 series annual filings by private pension plans
and provides them in machine readable format.41 The main form reports the filing date and the
plan filing id. The filing id allows within year linking of schedule H (financial information), schedule B (actuarial information, pre-2008) and schedule SB (actuarial information, single employer
plans post-2008). Schedule SB is not provided for 2008.
The sample in the paper starts from the universe of all observations for defined benefit plans
with a plan year between 2004 and 2012, and with a non-missing filing id. I drop observations
where:
the reported total asset value differs by more than 1% from the sum of individual asset
categories,
reported total income is not the sum of income categories,
reported total assets at the end of the filling period differ by more than 1% from the sum
of reported total assets at the beginning of the period plus total flows,
the reporting period covers fewer than 300 days or more than 400 days and doesnt start
January 1st,
plan year expenses or income exceed one third of initial assets.
In general, the filing id also longitudinally matches plans across years. Plans that change filing
id report their previous filing id. A complication emerges when multiple plans consolidate into
one. In that event, I drop observations where total assets at the beginning of the year differ by
more than 5% or $1000 from the sum of end of year assets across filings in the previous year.
Finally, before 2009 the raw data may contain multiple submissions (amended or restated
filings) by the same plan. To remove duplicates, I sequentially
41

http://www.dol.gov/ebsa/foia/foia-5500.html.

47

 drop duplicate filing id-year observations with a filing date before the most recent filing,
drop duplicates with incorrect signature or filing status,
drop duplicates reported as non amended filings,
keep the filing id-year with beginning of year assets closest to the end of year assets in the
previous year,
drop duplicates with missing funding measures,
drop duplicates with missing signature or entity code.
If duplicate filing id-years remain I randomly select one for inclusion.

B.

Derivations for the model in section 2.1

First, for a risk averse consumer, the portfolio of risky projects Ap satisfies

A =

K (, ) A (, ) dd.
Rf

(B.1)

With independent normally distributed returns, the risky portfolio return Rp has distribution
Rp N (p , p ), where
Z H Z
1
= p
K (, ) A (, ) dd,
A Rf 0
Z H Z
1
p 2
( ) =
2 K (, ) A (, ) dd.
2
p
f
(A ) R
0
p

(B.2)
(B.3)

The assumption Af > 0 means that Rf prices the consumption Euler equation,

exp (C0 ) = Rf E0 [exp (C1 )] .

(B.4)

Substituting into the Euler equation (B.4) using the budget constraints in equations (2) and (3),

48

taking logs, and solving for Af as a function of Ap , p , ( p )2 , and parameters, gives







p p 2
f
p
p
f 1 1
ln + ln R + Y0 A (1 + ) + (A ) .
A = 1+R

2
f

(B.5)

Substituting equations (B.5), (2) and (3) into the maximization problem (1), making use of
the parametric assumptions of exponential utility and normally-distributed returns, and ignoring
constants, the problem becomes

max

p p 2o
A (A ) .
2
p

(B.6)

Substituting equations (B.2) and (B.3) into the maximization (B.6) gives expression (4) in the
text.

C.

Appendix figures

49

Figure C.1: Five year note yield to maturity, percent

December 1, 2008

1.7

1.9

1.6

1.8

1.5

1.7

1.4

1.6

1.3

1.5
08:00

1.9

10:00

12:00

14:00

1.2
08:00

16:00

January 28, 2009

1.8

1.9

1.7

1.8

1.6

1.7

1.5

1.6

1.4
08:00

2.7

10:00

12:00

14:00

1.5
08:00

16:00

September 23, 2009

1.8

2.6

1.7

2.5

1.6

2.4

1.5

2.3

1.4

2.2
08:00

1.6

10:00

12:00

14:00

1.3
08:00

16:00

September 21, 2010

1.3

1.5

1.2

1.4

1.1

1.3

1.2

.9

1.1
08:00

10:00

12:00

14:00

.8
08:00

16:00

50

December 16, 2008

10:00

12:00

14:00

16:00

March 18, 2009

10:00

12:00

14:00

16:00

August 10, 2010

10:00

12:00

14:00

16:00

August 9, 2011

10:00

12:00

14:00

16:00

Figure C.1: Five year note yield to maturity, percent (continued)

January 25, 2012

1.1
1

.8

.9

.7

.8

.6

.7

.5

.6
08:00

10:00

12:00

14:00

.4
08:00

16:00

May 22, 2013

1.1

September 13, 2012

.9

1.3

.9

1.2

.8

1.1

.7

1
10:00

12:00

14:00

.9
08:00

16:00

July 10, 2013

1.7

1.8

1.6

1.7

1.5

1.6

1.4

1.5

1.3

1.4

1.2
08:00

10:00

12:00

14:00

1.3
08:00

16:00

51

12:00

14:00

16:00

June 19, 2013

1.4

.6
08:00

10:00

10:00

12:00

14:00

16:00

September 18, 2013

10:00

12:00

14:00

16:00

Figure C.2: Money market fund total financial assets

Billions of dollars
4000
iMoneyNet
FFA

3000

2000

1000

0
95:Q1

00:Q1

05:Q1

52

10:Q1

15:Q1

Brookings Panel on Economic Activity

March 2021, 2014

Are the Long-Term Unemployed on the

Margins of the Labor Market?
Alan B. Krueger, Princeton University and NBER
Judd Cramer, Princeton University
David Cho, Princeton University
Final conference draft

Are the Long-Term Unemployed on the Margins of the Labor Market?

Alan B. Krueger, Princeton University & NBER

Judd Cramer, Princeton University
David Cho, Princeton University

March 10, 2014

*We thank Katharine Abraham, Hank Farber, Larry Katz, Scott Kostyshak, Chris Nekarda,
David Romer, Rob Shimer, Thomas Winberry and Justin Wolfers for helpful comments, and Tito
Boeri for providing data for Italy. The authors are solely responsible for the views expressed in
the paper and any errors.

Introduction
A number of observers have noted that in recent years conventional Phillips Curve and
Beveridge Curve models predicted greater price deflation, greater real wage declines and fewer
vacancies as a result of the high rate of unemployment experienced during the Great Recession
and its aftermath than actually occurred. Several economists have provided possible explanations
for the missed predictions of the Phillips Curve, based on anchoring of expectations (Bernanke,
2007 and 2010) or changes in the distribution of price increases and interactions in the Phillips
Curve (Ball and Mazumder, 2011). Others have shown that the price Phillips Curve relationship
is stable if the short-term unemployment rate (defined as the number of job seekers unemployed
for 26 weeks or less relative to the labor force) is used instead of the total unemployment rate
(Gordon, 2013 and Watson, 2014), while others have shown that the Beveridge Curve
relationship is stable if short-term unemployment rate is used instead of the overall
unemployment rate (see Ghayad and Dickens, 2012).
This paper explores the plausibility of a unified explanation for the recent shifts in the
price and real wage Phillips Curves and Beveridge Curve in the U.S.: namely, that the long-term
unemployed, whose share of overall unemployment rose to an unprecedented level after the
Great Recession, are on the margins of the labor force and therefore exert very little pressure on
the job market and economy. The hypothesis we seek to test is that the longer workers are
unemployed the less they become tied to the job market, either because, on the supply side, they
grow discouraged and search for a job less intensively (e.g., Krueger and Mueller, 2011) or
because, on the demand side, employers discriminate against the long-term unemployed, based
on the (rational or irrational) expectation that there is a productivity-related reason that accounts
for their long jobless spell (e.g., Kroft, Lange and Notowidigdo, 2013 and Ghayad, 2013). Either

of these explanations would imply that the long-term unemployed are on the margins of the labor
market, and have a different effect on the macroeconomy than the short-term unemployed.
Moreover, the demand-side and supply-side effects of long-term unemployment can be viewed
as complementary and reinforcing of each other as opposed to competing explanations, as
statistical discrimination against the long-term unemployed could lead to discouragement, and
skill erosion that accompanies long-term unemployment could induce employers to discriminate
against the long-term unemployed.
Motivated by the apparent stability of the Phillips and Beveridge Curves when the shortterm unemployment rate is used to measure labor market slack, we assemble varied evidence to
assess the hypothesis that the long-term unemployed are on the margins of the labor market. To
preview our main findings, we tentatively conclude that the long-term unemployed exert
relatively little pressure on the economy, although the international evidence that we have been
able to assemble to this point is more mixed than the evidence for the U.S., and suggests that
long-term unemployment means different things in different countries and contexts.
We first briefly review evidence from the U.S. showing that the price Phillips Curve,
expected real wage Phillips Curve and Beveridge Curve are all stable if the short-term
unemployment rate is used to measure labor market slack, and that the long-term unemployment
rate has a comparatively modest effect when it is included in regression models. This result is
consistent with Llaudess (2005) conclusion that the long-term unemployment rate was a much
less significant determinant of price inflation and wage growth than the short-term
unemployment rate in many OECD countries prior to the Great Recession. We also create two
new measures of the unemployment rate, one in which the duration of unemployment is
weighted by a measure of search intensity and another in which duration is weighted by the

callback rate from audit studies. Both alternative measures have greater predictive power than
the total unemployment rate. We also extend this analysis to estimate the Beveridge Curve in the
U.K., which saw a sharp rise in long-term unemployment in the early 1980s, and Sweden, which
saw a sharp rise in long-term unemployment in the early 1990s, and then a gradual decline.
Sweden and the U.K. in these periods were selected because, compared to other countries, their
pattern of long-term unemployment as a share of the unemployed more closely resembles that of
the U.S. over the last decade.
Next we provide a detailed profile of the long-term unemployed in 2012, and examine
how the composition of the long-term unemployed has varied over time. While some notable
industries (e.g, construction) and demographic groups (e.g., African Americans) are over
represented among the long-term unemployed, the long-term unemployed are ubiquitous, spread
throughout all corners of the economy. Fully 36 percent of the long-term unemployed last
worked in the sales and service sector, suggesting that weak aggregate demand was the driver of
long-term unemployment. Using the relationship between workers characteristics and wages
from 2004 to 2006 to project earnings for the long-term unemployed, we find modest changes in
the composition of the long-term unemployed over the business cycle, with workers with more
highly rewarded characteristics more likely to be represented among the long-term unemployed
in recessions, although the differences are small.
We next examine the rates at which unemployed workers find employment or exit the
labor force, by duration of unemployment. Importantly, we examine transition rates both over a
month and over a year or longer. Longer durations of unemployment are associated with a lower
transition rate into employment, and available evidence suggests that observed durationdependent transition rates are not primarily a result of heterogeneous job searchers (e.g.,

Heckman and Singer, 1984). From 2008 to 2012, only 11 percent of those who were long-term
unemployed in a given month returned to steady, full-time employment a year later.1 This lowtransition rate into steady employment is considerably lower than what would be predicted from
monthly transition rates if such rates were independent over time and groups, highlighting that
the long-term unemployed frequently are displaced soon after they gain reemployment.
The long-term unemployed normally have a higher rate of labor force withdrawal than
the short-term unemployed, although following a recession the labor force withdrawal rates for
all duration groups tend to collapse to a common, relatively low level. We explore whether the
process of labor force withdrawal rates gradually returning to their historical normwith higher
exit rates for the long-term unemployed as well as a lower match rate for the long-term
unemployed, can cause the Beveridge Curve to loop around a stable path following a sharp
downturn. Specifically, we extend the calibration-type model of Kroft, et al. (2013) to allow for
varying labor force exit rates and differential match efficiency for the long-term unemployed to
project the path of the Beveridge Curve under a stable matching function. The results suggest
that from 2002-07 the long-term unemployed were about 60 percent as efficient in job matching
as the short-term unemployed. Using the matching function estimated for the 2002-07 period,
the calibrated model does a reasonably good job capturing the rise in unemployment and shift of
the Beveridge Curve in the 2008-13 period, as well as the rise in the share of unemployed
workers who are long-term unemployed. Future projections predict a gradual return to the
original Beveridge Curve as the share of long-term unemployment declines due to labor force
exits or (less likely) transitions to employment.

Steady employment in this context means that someone who was unemployed for 27 weeks or longer in month t
was employed full-time for four consecutive months starting in month t+12.

An alternative explanation for the Beveridge Curve gradually returning to its original
position is that a stronger labor market enables more of the long-term unemployed to transition
into employment. To explore this possibility, we compare trends in long-term and short-term
unemployment in different regions of the U.S. Our preliminary analysis suggests that the longterm unemployment rate has remained elevated even in low-unemployment rate states (defined
as the 13 states with the lowest unemployment rates in the U.S. as of October 2013). In addition,
we do not find evidence that the long-term unemployed are faring better in terms of transitioning
to employment in the low-unemployment states than in the high-unemployment states. Indeed,
the long-term unemployed appear to be following a similar path of transition rates both into
employment and out of the labor force in both the low- and high-unemployment states. These
findings suggest that the long-term unemployed will continue to encounter difficulty finding
employment even if the unemployment rate continues to fall, although a stronger economy
would undoubtedly raise the prospects of the long-term unemployed.
We conclude the paper by briefly considering some of the policy implications of the
hypothesis that the long-term unemployed are on the margins of the labor market. Because the
short-term unemployment rate has returned to its pre-recession average, one important
implicationif the hypothesis that the long-term unemployed are largely on the margins of the
labor market is correctis that further declines in short-term unemployment would be expected
to be associated with rising inflation and stronger real wage growth. More importantly, our
findings also suggest that a concerted effort will be needed to raise the employment prospects of
the long-term unemployed, especially as they are likely to withdraw from the job market at an
increasing rate if they follow the same path as in the previous recovery.

The Duration of Unemployment and Inflation, Wage Growth and Vacancies

This section summarizes movements in the price Phillips Curve, expected real wage
growth (which we call the wage Phillips Curve), and the Beveridge Curve since the Great
Recession, and provides econometric evidence suggesting that the rise in long-term
unemployment has caused these historical relationships to shift. We start with evidence from the
United States and then turn to two European countries that had previously experienced
substantial rises in long-term unemployment: the United Kingdom and Sweden.

The Price Phillips Curve

We start by estimating a simple short-run expectations-augmented Phillips Curve for core
price inflation using the following specification:
t t-1 = + t + t
where t denotes the annual average percent change in the consumption expenditures chain price
index excluding food and energy items (core PCEPI) in year t and t denotes the average
annual unemployment rate.2 Expectations are captured by lagged inflation. We restrict our
sample to the period from 1976 through 2013 to avoid complications from the oil shock and
stagflation of the early 1970s. Nevertheless, the results are robust to the inclusion of data back to
1960, the first year for which core PCEPI inflation can be officially estimated. Despite the
parsimonious nature of this model, our specification of the price Phillips Curve illustrates the
previously noted shift in the relationship between changes in price inflation and the
unemployment rate since the Great Recession.3

That is, t is the percent change in the average price level from year t-1 to year t. Similar results were obtained
with the Q4 to Q4 percent change in inflation.
3
For alternative specifications of the price Phillips Curve, see Gordon (2013) and Staiger, Stock, and Watson
(1997).
2

As shown in Figure 1(a), based on the fit of this Phillips Curve from 1976 to 2008, the
unemployment rates that have been observed since 2009 would have been expected to
correspond to declines in core consumer price inflation during this period. Given that the
unemployment rate averaged 8.7 percent from 2009 to 2013, a linear price Phillips Curve would
have predicted an average decline in core inflation of 1.0 percentage point per year during this
period. By contrast, the annual rate of inflation in the core personal consumption expenditures
price index has fallen an average of 0.2 percentage point per year since 2009.
A more detailed look at the composition of unemployed workers since the Great
Recession suggests that the unprecedented rise in long-term unemployment may have been
responsible for dampening the traditional relationship between price inflation and labor market
slack. In light of our hypothesis that a workers labor market attachment wanes as the duration of
unemployment lengthens, we estimate a short-term unemployment rate by calculating the
share of the civilian labor force that has been unemployed for less than 26 weeks. We then reestimate the core price inflation Phillips Curve relationship using this short-term unemployment
rate rather than the traditional measure. As shown in Figure 1(b), this modified price Phillips
Curve relationship appears to be more accurate in predicting the observed changes in the rate of
core inflation since 2009. Indeed, the average error between the observed and predicted price
Phillips Curve estimates since 2009 declines from 0.8 percentage point per year using the overall
unemployment rate to 0.1 percentage point per year using the short-term measure.4

These results were also robust to our attempts to account for the effects of the 1994 redesign of the Current
Population Survey, which found a greater proportion of unemployed workers who had long spells of joblessness.
See Polivka and Miller (1994). In particular, if we adjust the short-term and long-term unemployment rate series to
be more consistent using Polivka and Millers factors, the results are qualitatively similar. For simplicity, we use the
official data in the paper.

Figure 1(b). Change in Core Consumer Price

Inflation vs. Short-Term Unemployment Rate

Figure 1(a). Change in Core Consumer Price

Inflation vs. Unemployment Rate
Change in Core Inflation (Percentage Points)
3

Change in Core Inflation (Percentage Points)

3

1976-2008

1
2012

2012

2011

2011 2010

2010

2009
2013

-1

1976-2008

2013

-1 -1

2009

-2 -2

-2

-2

-3 -3

-3
3

5
6
7
8
9
Unemployment Rate (Percent)

10

-3
3

11

-1

4
5
6
7
8
9
Unemployment Rate: 26 Weeks & Less (Percent)

Note: Core consumer price inflation is defined as the annual average percent change in the personal consumption expenditures chain price
index excluding food and energy items.
Source: Bureau of Economic Analysis (National Income and Product Accounts), Bureau of Labor Statistics (Current Population Survey).

We also construct two alternative measures of unemployment to more directly account

for aspects of the supply-side and demand-side factors that appear to affect unemployed workers
job-finding prospects. First, we leverage previous research on the amount of time that jobless
workers typically spend looking for a job (Krueger and Mueller, 2011) to adjust the overall
unemployment rate by the estimated search intensity of the unemployed. In their longitudinal
analysis of unemployed workers in New Jersey in 2009 and 2010, Krueger and Mueller find that
the amount of time that jobless workers devoted to job search declined by approximately 1.5
minutes with each additional week of unemployment.5 Following Davis (2011), we apply the
coefficients from Krueger and Muellers fixed effects specification to the overall mean duration
of unemployment as published in the Current Population Survey in order to estimate a searchtime adjusted unemployment rate.6 We then index this measure of search intensity relative to

Similarly, Wanberg et al. (2013) find that an unemployed workers job search declines from roughly 18 hours per
week during the first week of joblessness to about 11 hours per week during the 20 th week of joblessness.
6
Because Krueger and Muellers specification is linear, results are unchanged if we use the average duration within
bins and assign the corresponding effects estimated by Krueger and Mueller.

2001 and apply it to scale the overall unemployment rate in an attempt to reflect the degree of
search intensity among jobless workers. One important caveat to this exercise, however, is that
Krueger and Mueller find much sharper within-worker declines in job search than across workers
with varying durations of unemployment. Although they provide some evidence suggesting that
the person fixed effects estimates more accurately reflect worker search behavior, it is also
possible that repeated questioning causes workers to reduce their reported job search activities
(time spent in search activities, applications submitted, etc.), but not their actual search behavior.
Second, we incorporate previous research on the probability that a job applicant receives
an interview after submitting an application for a job opening in order to adjust the overall
unemployment rate by the expected callback rate experienced by unemployed workers of
various durations. According to Kroft, Lange and Notowidigdos (2013) analysis, which
randomly assigned durations of unemployment to resumes that had been submitted to job
postings in 100 U.S. cities between 2011 and 2012, the likelihood of receiving a callback in
response to an application declines with time spent unemployed.
Specifically, Kroft, Lange, and Notowidigdo estimate the equation:
yi,c = 0.047 0.011 log(Di,c) 0.020 Ei,c + Xi,c + i,c
where yi,c denotes the probability of being called back for an interview for applicant i in
location c, log(D) denotes the logarithm of the duration of the workers spell of unemployment,
E is a dummy variable indicating whether or not the applicant was already employed, and X is
applicant and location characteristics. The average callback rate was 4.7 percent in this sample.
Using the coefficients from this regression model, we estimate the probability that the median
worker in each category of unemployment durationless than 5 weeks, 5 to 14 weeks, 15 to 26
weeks, and at least 27 weekswould receive a callback. Then, we convert these callback

probabilities into weights for each category of unemployment duration in order to adjust the
overall unemployment rate in an attempt to reflect the degree to which employers are considering
job applicants given the duration distribution of the unemployed.7 Table 1 reports estimates of
the simple price Phillips Curve using the conventional unemployment rate (column 1), the shortterm and long-term unemployment rates as separate regressors (column 2), and the two
alternative measures that reweight the unemployment rate to reflect declining search intensity
and callback probabilities with duration of unemployment (columns 3, 4, and 5). Although the
standard errors are relatively large, the short-term unemployment rate is a significant predictor of
inflation, and the long-term rate is not in the model shown in column 2; a joint test, however,
does not reject equality of the coefficients at conventional significance levels. We also find that
the alternative measures of labor market slack appear to have more predictive power for changes
in price inflation than does the traditional unemployment rate. Because the two measures are
highly collinear, we present results with only one of the alternative unemployment measures in
columns 3 and 4, and then with both measures simultaneously in column 5.

We normalize the weights relative to the lowest category, those unemployed for less than 5 weeks. Under this
framework, the number unemployed for less than 5 weeks is unadjusted, while the number unemployed 5 to 14
weeks is adjusted by a factor of 0.79, the number unemployed 15 to 26 weeks is adjusted by a factor of 0.58, and the
number unemployed for 27 weeks or more is adjusted by a factor of 0.37.

10

Table 1: Estimated Core Price Phillips Curve

Dependent Variable: Annual Average Percent Change in Core Inflation Less
Previous Year's Annual Average Percent Change in Core Inflation
(1)
(2)
(3)
(4)
Intercept
1.106
1.545
1.311
1.466
(0.568)
(0.573) *
(0.505) *
(0.551) *
Unemployment Rate
-0.200
(0.095) *
Unemployment Rate: 26 Weeks or Less
-0.321
(0.133) *
Unemployment Rate: 27 Weeks or More
-0.054
(0.141)
Unemployment Rate: Search-Weighted
-0.261
(0.098) *
Unemployment Rate: Callback-Weighted
-0.340
(0.121) **
Wald Test for Equal Unemployment Variables: p-value
0.266
Adjusted R-Squared
0.153
0.176
0.183
0.185
Note: Annual data from 1976 to 2013 (38 observations). Newey-West standard errors with 3 lags shown in parentheses.
Levels of Significance: *** = 0.01, ** = 0.05, * = 0.10
Source: Bureau of Economic Analysis (National Income and Product Accounts), Bureau of Labor Statistics (Current Population
Survey), authors' calculations.

(5)
1.505
(0.532) **

-0.137
(0.148)
-0.186
(0.164)
0.869
0.177

The Real Wage Phillips Curve

We next estimate an expected real wage growth Phillips Curve using the following
specification:
t t-1 = + t + t
where denotes the annual average percent change in average hourly earnings of production and
nonsupervisory employees, is the annual average percent change in the core PCEPI, and
denotes the average annual unemployment rate.8 Again, we limit our analysis to the period from
1976 through 2013 in order to avoid measurement issues surrounding the oil shock and
stagflation of the early 1970s. Nevertheless, the main conclusions are robust to the inclusion of
data back to 1965, the first year for which growth in average hourly earnings can be officially
estimated. Furthermore, despite the relatively modest structure of this model, our specification of
8

This follows the specification, for example, of Katz and Krueger (1999). The dependent variable subtracts lagged
inflation from wage growth to reflect expected real wage growth. Results are qualitatively similar if
contemporaneous inflation is used instead.

11

the real wage Phillips Curve effectively illustrates the shift in the relationship between changes
in expected real wage growth and the unemployment rate since the Great Recession.
Figure 2(b). Change in Real Expected Average Hourly
Earnings vs. Short-Term Unemployment Rate

Figure 2(a). Change in Real Expected Average

Hourly Earnings vs. Unemployment Rate
Change in Real Expected Earnings (Percentage Points)
3

Change in Real Expected Earnings (Percentage Points)

3

2009

2010

2010

2012 2011

2013

2009

2011
2013

2012

-1 -1

-1

-1
1976-2008

1976-2008

-2

-2 -2

-2

-3

-3 -3

-3

-4 -4

-4
3

5
6
7
8
9
Unemployment Rate (Percent)

10

-4
3

11

4
5
6
7
8
9
Unemployment Rate: 26 Weeks & Less (Percent)

Note: Change in real expected earnings is defined as the annual average percent change in average hourly earnings of production and
nonsupervisory employees less the previous year's annual average percent change in the personal consumption expenditures chain price index
excluding food and energy items.
Source: Bureau of Economic Analysis (National Income and Product Accounts), Bureau of Labor Statistics (Current Employment Statistics
and Current Population Survey).

The scatter diagram in Figure 2(a) shows the fitted expected real wage Phillips Curve
using data from 1976 to 2008, and the observed data since 2008. It is clear that the historical
relationship between real wage growth and unemployment would have predicted sharp real wage
declines over the past five years. Based on the average unemployment rate of 8.7 percent from
2009 to 2013, the real wage Phillips Curve would have predicted an average decline in expected
earnings of 1.9 percentage points per year. Instead, however, the annual rate of change in real
expected earnings has risen an average of 0.6 percentage point per year since 2009.
Comparable to our analysis of the price Phillips Curve, Figure 2(b) reports the real wage
Phillips Curve relationship using the short-term unemployment rate instead of the total
unemployment rate. Again, we find a better and more stable fit. The average error between the
observed and predicted real wage Phillips Curve declines from 2.6 percentage points per year
12

using the overall unemployment rate to 0.4 percentage point per year using the short-term
measure. Furthermore, as shown in Table 2, the alternative measures of labor market slack
appear to have considerably larger effects on changes in expected real wages than does the
traditional unemployment rate, and a Wald test rejects an equal effect of short-term and longterm unemployment. When both alternative measures are included in the regression (column 5),
the callback-weighted unemployment rate has a statistically significant negative effect, although
one cannot reject the hypothesis that the search-weighted and callback-weighted unemployment
rates have equal coefficients.
Table 2: Estimated Expected Real Wage Phillips Curve
Dependent Variable: Annual Average Percent Change in Average Hourly Earnings Less
Previous Year's Annual Average Percent Change in Core Inflation
(1)
(2)
(3)
(4)
Intercept
3.936
5.718
4.723
5.459
(1.322) **
(0.715) ***
(1.041) ***
(0.793) ***
Unemployment Rate
-0.550
(0.224) *
Unemployment Rate: 26 Weeks or Less
-1.045
(0.169) ***
Unemployment Rate: 27 Weeks or More
0.041
(0.230)
Unemployment Rate: Search-Weighted
-0.759
(0.181) ***
Unemployment Rate: Callback-Weighted
-1.046
(0.184) ***
Wald Test for Equal Unemployment Variables: p-value
0.003
Adjusted R-Squared
0.371
0.582
0.490
0.557
Note: Annual data from 1976 to 2013 (38 observations). Newey-West standard errors with 3 lags shown in parentheses.
Levels of Significance: *** = 0.01, ** = 0.05, * = 0.10
Source: Bureau of Economic Analysis (National Income and Product Accounts), Bureau of Labor Statistics (Current Employment
Statistics and Current Population Survey), authors' calculations.

(5)
5.529
(0.750) ***

-0.245
(0.222)
-0.772
(0.319) *
0.318
0.558

The Beveridge Curve

Figures 3(a) and 3(b) graphically depict the Beveridge Curve, or relationship between job
openings and the unemployment rate. It has been well documented that, since the Great
Recession ended, the job openings rate has shifted rightward relative to the previous relationship
with the unemployment rate that had prevailed from 2001 to 2007 (Hobijn and Sahin 2012), and
13

this is apparent in Figure 3(a). In Figure 3(b), which displays the Beveridge Curve using the
short-term unemployment rate, however, the relationship appears notably more stable, with little
evidence of a shift.
Figure 3(b). Job Openings Rate vs.
Short-Term Unemployment Rate

Figure 3(a). Job Openings Rate vs.

Unemployment Rate

Job Openings Rate (Percent)

3.8

Job Openings Rate (Percent)

3.8
Dec-00

3.6

Dec-00 to
Feb-01

Dec-00

3.6

Dec-00 to
Feb-01

3.4

3.4

3.2

3.2
Mar-01 to
Nov-01

3.0

Dec-13

2.8

2.8

Jul-09 to
Dec-13

2.6

Dec-13

2.6

Dec-01 to
Nov-07

2.4

Mar-01 to
Nov-01

3.0

Dec-01 to
Nov-07

2.4

Dec-07 to
Jun-09

2.2

2.2
Dec-07 to
Jun-09

2.0

2.0

Jul-09 to
Dec-13

1.8

1.8

1.6

1.6

5
6
7
8
9
Unemployment Rate (Percent)

10

11

3.0

3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5

Unemployment Rate: 26 Weeks & Less (Percent)

Note: Job openings rate is defined as job openings as a percentage of the sum of job openings and total nonfarm payroll employment.
Source: Bureau of Labor Statistics (Job Openings and Labor Turnover Survey and Current Population Survey).

Table 3 summarizes estimates of a Beveridge Curve using the following specification:

t = + t + t
where denotes monthly job openings as a percentage of job openings plus total nonfarm payroll
employment and denotes the monthly unemployment rate. We begin our analysis in December
2000, the first year available from the Job Openings and Labor Turnover Survey (JOLTS). Given
that the unemployment rate averaged 8.6 percent from July 2009 to December 2013, a Beveridge
Curve relationship estimated from 2001 to 2007 would have predicted an average job openings
rate of 0.9 percent from mid-2009 to the end of 2013. By contrast, the job vacancy rate has
averaged 2.4 percent during this period.
The results in Table 3 indicate a large role for the short-term unemployment rate in
determining job openings, but the long-term rate is statistically insignificant (column 2).
14

Moreover, both the search-weighted unemployment rate and the callback-weighted

unemployment rate have statistically significant, negative effects on job openings in column 5.
Table 3: Estimated Beveridge Curve
Dependent Variable: Job Openings as a Percent of Total Nonfarm Payroll Employment Plus Job Openings
(1)
(2)
(3)
(4)
(5)
Intercept
3.876
5.077
4.708
4.550
4.835
(0.184) ***
(0.234) ***
(0.237) ***
(0.191) ***
(0.184) ***
Unemployment Rate
-0.188
(0.028) ***
Unemployment Rate: 26 Weeks or Less
-0.514
(0.057) ***
Unemployment Rate: 27 Weeks or More
-0.026
(0.029)
Unemployment Rate: Search-Weighted
-0.395
-0.224
(0.041) ***
(0.037) ***
Unemployment Rate: Callback-Weighted
-0.420
-0.225
(0.040) ***
(0.045) ***
Wald Test for Equal Unemployment Variables: p-value
0.000
Adjusted R-Squared
0.653
0.833
0.769
0.759
Note: Monthly data from December 2000 to December 2013 (157 observations). Newey-West standard errors with 12 lags
shown in parentheses.
Levels of Significance: *** = 0.01, ** = 0.05, * = 0.10
Source: Bureau of Labor Statistics (Job Openings and Labor Turnover Survey and Current Population Survey), authors' calculations.

0.994
0.842

Further Evidence from the United Kingdom and Sweden

While high long-term unemployment had been an unusual phenomenon in the United
States in the post-World War II era, it has been much more prevalent in Europe. In a prescient
empirical study, Llaudes (2005) argued that the rise in long-term joblessness distorted the
determination of prices and wages in many European countries. Llaudes employed alternative
measures of unemployment that were re-weighted to account for the duration of joblessness, and
found that measures that down-weighted the long-term unemployed tended to produce more
accurate predictions of changes in prices and wages than the traditional unemployment rate.
For example, consistent with our findings for the United States since the Great Recession,
Llaudes finds that the long-term unemployed had a relatively smaller effect on changes in prices
and wages in the United Kingdom from 1973 to 2002 than did those who had shorter durations
15

of joblessness. A modified unemployment rate in which workers who have been unemployed for
at least a year are weighted by roughly 20 percent as much as the short-term unemployed does a
better job of predicting price changes in the United Kingdom than does the overall aggregate
measure in which all unemployed workers are weighted equally. Likewise, an adjusted
unemployment rate in which workers who have been unemployed for at least a year are weighted
by 17 percent as much as the short-term unemployed does a better job of predicting wage
changes in the United Kingdom than does the overall aggregate measure.
Nevertheless, the degree to which the long-term unemployed affect changes in prices and
wages can vary considerably across countries. For instance, in his specification of the price
Phillips Curve for Sweden from 1971 to 2002, Llaudes estimates a weight of about 50 percent
for workers who have been unemployed for at least a year in his adjusted measure of labor
market slack. Similarly, Llaudes calculates a weight of about half for the long-term unemployed
in his alternative measure of the unemployment rate when estimating the wage Phillips Curve for
Sweden.
More recently, Hobijn and Sahin (2012) provide a careful examining of the Beveridge
Curve in 14 OECD countries, both since the Great Recession and during earlier periods. They
find that the Beveridge Curve shifted out notably in the U.S. and four countries since the Great
Recession: Portugal, Spain, the U.K., and Sweden. While they do not explore the role of
unemployment duration, they suggest (p. 26-27): The common policy response to reduce the
burden of displacement for the unemployed by increasing the generosity and duration of
unemployment insurance further contributes to the rightward shift of the Beveridge [C]urve.
They also highlight the potential effects of house lock and skills mismatch for explaining
rightward shifts of the Beveridge Curve since the Great Recession.

16

Rather than replicate Llaudess and Hobijn and Sahins work for a large set of countries,
we examine whether long-term unemployment had a differential effect on the Beveridge Curve
in the U.K. and Sweden. We selected the United Kingdom and Sweden because those countries
experienced sharp run-ups in overall and long-term unemployment and then returned to a more
normal situation in the 1980s and 1990s; as a result, they may yield more useful insights for the
United States than is the case in countries that experienced persistently high long-term
unemployment. In addition, we examine their experience up to the most recent period to probe
the robustness of our findings.
Tables 4 and 5 summarize estimates of the Beveridge Curve for the United Kingdom and
Sweden. Columns 1 and 2 incorporate data from 1983 to 2005reflecting the period during
which long-term unemployment rose sharply and then declined. As in the analysis for the United
States, we decompose the overall unemployment rate into short-term and long-term measures. In
both countries, the short-term unemployment rate is a more important predictor of vacancies than
the long-term rate through 2005. Although these results are similar to the findings for the United
States, the results are not robust if we extend the data through the Great Recession (columns 3
and 4). In particular, in the longer sample, long-term unemployment appears to have at least as
great an effect on vacancies as short-term unemployment. Thus, we have not found robust
evidence that a differential matching efficiency of the short-term and long-term unemployed can
consistently account for movements in the Beveridge Curve in the U.K. and Sweden.

17

Table 4: United Kingdom: Estimated Beveridge Curve

Dependent Variable: Job Openings as a Percent of Total Employment Plus Job Openings
(1)
(2)
(3)
Intercept
2.727
3.492
3.011
(0.480) ***
(0.542) ***
(0.386) ***
Unemployment Rate
-0.210
-0.233
(0.055) **
(0.045) ***
Unemployment Rate: Less Than 6 Months
-0.484
(0.144) **
Unemployment Rate: 6 Months or More
-0.170
(0.066) *

(4)
2.851
(0.917) **

-0.178
(0.305)
-0.241
(0.062) ***

Wald Test for Equal Unemployment Variables: p-value

0.089
0.855
Adjusted R-Squared
0.637
0.641
0.632
0.619
Number of Observations
23
23
29
29
Note: Annual data from 1983 to 2005 in columns 1 and 2. Annual data from 1983 to 2011 in columns 3 and 4. Newey-West
standard errors with 3 lags shown in parentheses.
Levels of Significance: *** = 0.01, ** = 0.05, * = 0.10
Source: Organisation for Economic Cooperation and Development, authors' calculations.

Table 5: Sweden: Estimated Beveridge Curve

Dependent Variable: Job Openings as a Percent of Total Employment Plus Job Openings
(1)
(2)
(3)
Intercept
1.061
1.247
1.035
(0.106) ***
(0.104) ***
(0.110) ***
Unemployment Rate
-0.077
-0.059
(0.014) ***
(0.018) **
Unemployment Rate: Less Than 6 Months
-0.211
(0.063) **
Unemployment Rate: 6 Months or More
0.048
(0.057)

(4)
0.974
(0.140) ***

-0.015
(0.053)
-0.108
(0.041) *

Wald Test for Equal Unemployment Variables: p-value

0.039
0.310
Adjusted R-Squared
0.596
0.693
0.327
0.331
Number of Observations
23
23
30
30
Note: Annual data from 1983 to 2005 in columns 1 and 2. Annual data from 1983 to 2012 in columns 3 and 4. Newey-West
standard errors with 3 lags shown in parentheses.
Levels of Significance: *** = 0.01, ** = 0.05, * = 0.10
Source: Organisation for Economic Cooperation and Development, authors' calculations.

A Profile of the Unemployed

This section provides a detailed portrait of the long-term unemployed in comparison to
employed workers and short-term unemployed workers. We begin by reviewing trends in the
incidence of long-term unemployment, both in the U.S. and in other economically advanced

18

countries, then summarize characteristics of the long-term unemployed, and then examine how a
summary measure of the composition of the long-term unemployed (based on earnings
prospects) has varied over time.
Figure 4 reports duration-specific unemployment rates in the U.S. based on published
seasonally adjusted monthly data from BLS from January 1948 through the end of 2013. The
red line indicates the long-term unemployment rate (defined as the number unemployed for 27
weeks or longer divided by the labor force), whereas the blue line is the similarly defined
unemployment rate for those unemployed for 14 weeks or less, and the green line is the rate for
the intermediate group unemployed for 15 to 26 weeks. Notice that the long-term
unemployment rate, which tends to rise during periods of recession and peak shortly afterwards,
jumped to record heights during the Great Recession, and peaked in early 2010 before starting to
decline. Despite declining over the last four years, the long-term unemployment rate still
exceeds its previous peak, reached in the aftermath of the deep 1981-82 recession, and is well
above its average in the last recovery. The two measures of short-term unemployment, however,
are close to their average rates experienced during the last recovery. Thus, as an accounting
matter, unemployment remains elevated because of the large number of people who have been
unemployed for more than half a year.

19

Figure 4. Unemployment Rates by Duration

Percent of Civilian Labor Force (Seasonally Adjusted)
7
6
Unemployed for
14 Weeks & Less

5
4

2001-2007
Historical Average
Unemployed for
27 Weeks & Over

3
2

2001-2007
Historical Average

Unemployed for
15-26 Weeks

2001-2007
Historical Average

Dec-13

1
0
Jan-48

Jan-58

Jan-68

Jan-78

Jan-88

Jan-98

Jan-08

Note: Shading denotes recession.

Source: Bureau of Labor Statistics (Current Population Survey), National Bureau of Economic
Research.

As Figure 5 shows, for most of the four decades prior to the Great Recession, the share of
the unemployed in the U.S. who were out of work for more than half a year oscillated between
10 and 20 percent during recoveries and recessions.9 By contrast, at least since the 1980s, a
much higher share of the unemployed have been on long-term unemployment spells in the U.K.,
France, Germany and, especially, Italy. The share of the unemployed who were long-term
jumped from 20 percent to 50 percent in Sweden after its severe financial crisis in the early
1990s, and slowly fell to near its pre-crisis levels, before rising again in the latest recession. The

See Abraham and Shimer (2002) for a careful analysis of why the duration of unemployment in the U.S. rose
relative to the unemployment rate in the 1980s and 1990s.

20

long-term unemployed typically made-up a higher share of the unemployed in Canada than in the
U.S., although its share of long-term unemployed workers trended down from the early 1990s to
the Great Recession, and is now almost half the U.S. share. These disparate trends, particularly
the high-share of the unemployed who have been out of work for longer than half a year in
France, Germany, Italy, and the U.K., often exceeding 60 percent, suggest that long-term
unemployment affects different segments of the workforce in different countries, and can be a
more or less persistent phenomenon depending on a nations institutions and social benefits.

Figure 5. Long-Term Unemployment in Selected Countries

Percent of Total Unemployed in Each Country
90
80
Italy

2012

70
Germany

60

United
Kingdom

50

France

2012
Dec-13

40

Sweden

30

2012

Canada

20

2012
United States

10
0
Jan-70

Jan-80

Jan-90

Jan-00

Jan-10

Note: Data for the United States are at monthly frequency and represent those who have been unemployed for
at least 27 weeks. Data for all other countries are at annual frequency and represent those who have been
unemployed for more than 6 months.
Source: Bureau of Labor Statistics (Current Population Survey); Organization for Economic Cooperation and
Development.

Table 6 reports the distribution of employees, short-term unemployed workers and longterm unemployed workers along several dimensions for the U.S. using the 2012 Current

21

Population Survey (CPS) survey. For example, the table indicates that 34 percent of employed
individuals are age 16 to 34, 33 percent are age 35 to 49, and 33 percent are age 50 and older.
Compared to their share of employment, young people are notably over represented among the
short-term unemployed, while the middle age group is under represented. Compared to their
share of the short-term unemployed, the oldest group is over represented among the long-term
unemployed, although their share of the long-term unemployed roughly matches their share of
unemployment.
If the unemployed as a whole are compared to the employed, notably larger shares of the
unemployed are younger, unmarried, and less well educated. For example, although about one
third of employed workers have earned a bachelors degree, less than 20 percent of the
unemployed have done so. By contrast, nearly 20 percent of the unemployed lack a high school
diploma, which is twice the rate for the employed. African Americans and Hispanics are also
over represented among the ranks of the unemployed compared with the employed. Not
surprisingly given the housing bubble, a higher proportion of the unemployed previously worked
in the construction industry than the share of workers currently employed as construction
workers; nonetheless, only 11 percent of the unemployed are former construction workers.
If the long-term unemployed are compared to the short-term unemployed, a larger
proportion of the long-term unemployed are over age 50 and do not have a spouse. Fully 44
percent of the long-term unemployed were never married, while nearly 20 percent are either
widowed, separated, or divorced. In addition, African Americans comprise 22 percent of the
long-term unemployed, compared with just 10 percent of the employed population.
Among many other dimensions, however, the long-term unemployed appear similar to
the short-term unemployed. Other than high school dropouts, the educational achievement of the

22

two groups is comparable, and both the industry distribution and occupational distribution are
similar. Differences across regions and between urban and rural areas are also typically small.
Table 6: Profile of the Employed, Short-Term Unemployed and
Long-Term Unemployed, 2012
Percent of
Employed

Percent of
Short-Term
Unemployed
(< 14 Weeks)

Percent of
Long-Term
Unemployed
(> 26 Weeks)

Gender
Male
Female

53
47

54
46

55
45

Age
16-34
35-49
50+

34
33
33

57
23
19

40
29
31

Marital Status
Married
Widowed/Divorced/Separated
Never Married

56
15
29

33
15
52

37
19
44

Race
White, Nonhispanic
African American
Hispanic
Asian/Pacific Islanders
Other

67
10
15
6
2

55
16
22
4
3

51
22
19
5
3

Education
Less than High School
High School
Some College
Associate's Degree
Bachelor Degree or Higher

9
27
19
11
34

23
33
20
8
17

18
36
20
9
18

Industry
Construction
Manufacturing
Wholesale and Retail Trade
Finance and Real Estate
Prof. and Business Services
Education and Health Care
Leisure and Hospitality
All Other

6
10
14
7
12
23
9
19

12
9
15
4
14
16
15
16

11
11
16
5
14
15
12
15

13
7
12
39
29

12
10
14
36
28

Occupation
Professional and Technical
22
Managerial and Financial
16
Administrative
12
Sales and Service
32
Blue Collar
18
Source: Authors' calculations from the Current Population Survey.

23

A majority of the long-term unemployed last held jobs in just two occupational
categories, blue-collar jobs (28 percent) and sales and service jobs (36 percent). The former
category tends to be dominated by men, while the latter is divided roughly equally between men
and women. Professional and technical workers, administrative workers, and managerial and
financial workers, each comprise notably smaller shares of the long-term unemployed.
We compared tabulations analogous to those in Table 6 for Italian employees in
2013Q1.10 Italy provides an interesting contrast because long-term unemployment is persistently
a much higher share of the unemployed in Italy than in the U.S. One striking difference is that
young workers were vastly over represented among the long-term unemployed in Italy,
especially in comparison to their share of employment. Just under a quarter of all workers are
age 15-34 in Italy, while nearly half of the long-term unemployed are age 15-34. Only 15
percent of the long-term unemployed in Italy are age 50 or older, compared with 31 percent in
the U.S. Additionally, a notably high percentage (56 percent) of Italian long-term unemployed
workers had less than a high school education, and nearly half of the long-term unemployed in
Italy were in the south and islands, almost twice their share of national employment. In the U.S.,
by contrast, long-term unemployment is distributed more evenly across regions.
The high concentration of the long-term unemployed among the young, less educated and
southern region in Italy suggest that the nature of long-term unemployment is different there than
in the U.S. In particular, rather than a sharp rise in long-term unemployment that hit all
segments to a considerable extent, the Italian predicament of long-term unemployment appears
to be more of a persistent, structural phenomenon. The disproportionate share of unemployment
borne by the young in Italy is consistent with an insider-outsider model, where older workers are

10

We thank Tito Boeri for providing these tabulations to us. The tabulations were based on the ISTA RCFL (Labor
Force Survey).

24

the insiders, and where they retire if they lose their jobs. The higher share of young workers
among the long-term unemployed in Italy suggests that the long-term unemployed may have a
longer term commitment to staying in the job market and finding a job in Italy than in the U.S.

Composition of the Long-Term Unemployed over Time in U.S.

Table 6 summarizes the characteristics of the unemployed at a point in time. To create a
single summary measure of the characteristics of the long-term unemployed that can be tracked
over time we used the following procedure. We first estimated a wage regression using data
from 2004-06, which was a more or less normal period for the labor market, and then we
combined the coefficients from this regression with the characteristics of the long-term
unemployed each year to track the earnings potential of long-term unemployed workers each
year from 1995 to 2013. Specifically, the wage regression related the log hourly wage of
workers to their education, experience, industry, occupation , race, gender and marital status.
The estimated coefficients from this regression were then combined with the characteristics of
the long-term unemployed (defined as those unemployed for longer than 26 weeks at the time of
the survey) each year to derive a simple summary of the composition of the long-term
unemployed with respect to their earnings prospects.11
Figure 6 contains the results of this exercise. There appears to be both a mild secular
trend and a mild cyclical pattern in the composition of the long-term unemployed, at least as far
as their characteristics that predict earnings are concerned. The composition of the unemployed
11

We use CPS data from 1995 forward because we limit the sample to the period after the 1994 redesign of the CPS,
which affected the share of long-term unemployed workers and improved the ability to track individuals over time.
Industry and occupation were measured on a consistent two-digit basis. The description in the text ignores one
complication. For the minority of unemployed workers who are new entrants, information on occupation and
industry is lacking. To include these workers, we estimated the wage regression a second time, but omitted industry
and occupation from the model. The coefficients from this second regression were used to predict earnings for those
lacking industry and occupation data. The results are qualitatively unchanged if the new entrants are excluded from
Figure 6, however.

25

has tilted towards those with characteristics associated with higher earnings, such as more
education, since the mid 1990s. In addition, the mix of the long-term unemployed with
characteristics associated with higher earnings tends to rise during downturns. Predicted
earnings of the long-term unemployed rose by about 5 log points in the past two recessions. This
pattern is consistent with Andreas Muellers (2012) finding that in recessions the pool of the
unemployed tends to shift towards those with higher earnings in their previous job because such
workers are more likely to be displaced in recessions. In the section on transition rates below,
we perform a similar exercise to examine changes in the composition of the long-term
unemployed with respect to measured characteristics that predict job finding and labor force
withdrawal.
Figure 6. Mean Predicted Log Wages for the Long-Term Unemployed Based on Their
Characteristics
Log 2013 Dollars
2.85

2.80

2.75
2013

2.70

2.65

2.60
1995

1997

1999

2001

2003

2005

2007

2009

2011

2013

Note: Annual averages. Shading denotes recession.

Source: Bureau of Labor Statistics; National Bureau of Economic Research; Authors calculations using Current
Population Survey Longitudinal Population Database (see Nekarda, 2009).

26

Industry and Occupation Distributions

Table 7 reports the previous and new occupations and industries, respectively, of shortterm and long-term unemployed workers who regained employment in 2012 based on CPS data.
For a benchmark, the share of employed workers in each occupation and industry is also
reported. The results are striking in that the new job distribution closely replicates the preunemployment job distribution for those who were fortunate enough to be reemployed in 2012.
For example, 27 percent of long-term unemployed workers who were reemployed in 2012 had
previously worked in blue collar jobs, and 26 percent of the short-term workers who found work
in 2012 were employed in blue collar jobs. The corresponding figures for short-term
unemployed workers who regained employment are that 33 percent came from blue collar jobs
and 32 percent were reemployed in blue collar jobs.
A similar pattern holds for the industry distributions: the distribution of the industries that
both short-term and long-term unemployed workers regained employment in closely replicate the
industries from which they were displaced. There is no tendency in these data for reemployed
workers to gravitate to growing industries or occupations relative to the industries from which
they were displaced. These results suggest that assisting unemployed workers to transition to
expanding sectors of the economy, such as health care, professional and business services, and
management, is a major challenge. Instead, unemployed workers who do return to work tend to
return to jobs in their previous sectors.

27

Table 7: Previous and New Occupations and Industries of Short-term and Long-term
Unemployed Workers who Regained Employment in 2012
Occupations of Workers Who Were Employed or Found Work in 2012
(Percent of Total in Each Category of Employed Workers)
Employees Who Had
Employees Who Had
Previously Been
Previously Been
Unemployed
Unemployed
Short-Term But
Long-Term But
Total
Found Work
Found Work
Occupation
Employed
Old Work New Work
Old Work New Work
Blue-Collar Occupations
17.5 %
33.4 %
31.9 %
26.8 %
25.6 %
Sales & Services
32.2
37.1
38.2
37.4
40.1
Administrative
12.4
9.1
9.5
12.9
12.9
Professional & Technical
22.0
14.7
14.8
12.1
12.8
Managerial & Financial
15.9
5.7
5.6
10.8
8.5
Note: Short-term unemployed defined as 26 weeks or fewer of unemployment. Long-term unemployed defined
as 27 weeks or more of unemployment.
Source: Estimates from the 2012 Current Population Survey.

Industries of Workers Who Were Employed or Found Work in 2012

(Percent of Total in Each Category of Employed Workers)
Employees Who Had
Employees Who Had
Previously Been
Previously Been
Unemployed
Unemployed
Short-Term But
Long-Term But
Total
Found Work
Found Work
Industry
Employed
Old Work New Work
Old Work New Work
Construction
6.3 %
14.3 %
13.6 %
10.9 %
11.5 %
Manufacturing
10.3
8.1
7.7
9.5
6.5
Wholesale & Retail Trade
14.0
13.2
12.9
15.0
15.7
Financial Activities
6.7
3.1
2.9
5.4
3.7
Professional & Business Services
11.6
13.2
13.4
15.5
16.9
Educational Services
9.1
9.1
8.6
5.2
6.0
Health Care Services
13.6
8.2
8.7
10.6
10.4
Leisure & Hospitality
9.3
14.6
16.2
12.2
12.6
All Other
19.1
16.3
16.1
15.7
16.6
Note: Short-term unemployed defined as 26 weeks or fewer of unemployment. Long-term unemployed defined
as 27 weeks or more of unemployment.
Source: Estimates from the 2012 Current Population Survey.

28

Transition Rates
This section explores labor market activity of the long-term unemployed over time.
Specifically, we use longitudinally linked CPS data (see Nekarda, 2009) to study how the longterm unemployed fare in later survey months. In particular, we investigate the factors underlying
whether the long-term unemployed move into employment, continue actively searching for
employment, or transition from unemployment to not in the labor force. As others have shown
(e.g., Valetta, 2011), the long-term unemployed have disparate labor market flows compared to
short-term unemployed workers. We begin by reviewing basic trends in labor force flows by
duration of unemployment. We then discuss some explanations for the observed patterns, and
explore the implications for the future labor market. One channel that is most important in our
view for the future path of long-term unemployment in the U.S. is how the decision to continue
searching for a job conditional on not having become employed evolves. We will later embed
different assumptions about movement from unemployed to out of the labor force into a
calibration model along the lines of Kroft, et al. 2013 to see how outcomes differ, focusing
especially on the evolution of the Beveridge Curve, the macroeconomic relationship which
arguably most directly reflects the operation of the job market. Our findings suggest that we
could see a return to the original Beveridge Curve following the labor force exit of many of the
long-term unemployed.
Figure 7 displays annual averages of monthly transition rates from unemployment to
employment each year since 1994, based on BLSs published transition rates for five duration of
unemployment categories. Although many researchers have documented that CPS data can
severely misstate gross labor market flows because of classification errors, which require caution
in interpreting the data, the series nonetheless convey some signal and reflect movements in the

29

official unemployment rate (Abowd and Zellner, 1985, Summers and Poterba, 1986 and Shimer,
2012). A few patterns are clear. First, the job-finding rate is lower for those with a longer
duration of unemployment, with the long-term unemployed finding jobs at less than half the rate
of those very short-term unemployed. Second, the cyclicality of job finding is clear in these
data, with all rates declining during the recession of the early 2000s, and declining more
dramatically during the Great Recession. Third, job finding rates for all groups remain well
below their pre-Great Recession averages. Fourth, the job finding rate has risen for each group
in the last four years, although it has barely increased for those unemployed longer than a year.
In 2013, just under 10 percent of those who had been unemployed for more than one year
transitioned into employment in the average month. This rate, though higher than in many
European countries (Elsby, et al. 2011), might overstate how well the long-term unemployed are
faring due to measurement error and the fact that the long-term unemployed are more likely to
take low-paying, part-time jobs and temporary jobs; a point we revisit below.
Observed duration-dependence in job finding rates could reflect worker heterogeneity
(i.e., as those with the most marketable skills tend to find jobs more quickly), or an effect of
discouragement, skill erosion and employer statistical discrimination against the long-term
unemployed. Available evidence suggests that observed duration-dependent transition rates are
not primarily a result of heterogeneous job searchers (e.g., Heckman and Singer, 1984), although
econometric evidence on the respective roles of heterogeneity and duration dependence remains
unsettled.

30

Figure 7. Probability of Transitioning From Unemployment to Employment by

Duration of Unemployment
Percent of Each Category of Unemployment Duration
45
Unemployed Less
Than 5 Weeks

40
35

2013
Unemployed for
5 to 14 Weeks

30

Unemployed for
15 to 26 Weeks

25
20
Unemployed for
27 to 52 Weeks

15

Unemployed for
53 Weeks & Over

10
5
1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

Note: Dotted lines represent 1994-2007 averages. Shading denotes recession.

Source: Bureau of Labor Statistics (Current Population Survey); National Bureau of Economic Research.

Figure 8, which uses the same scale for the y-axis as Figure 7, suggests that any effect of
changing heterogeneity on the pattern of job finding rates over the business cycle for the longterm unemployed is small. To construct this figure, we first used the same characteristics that
were used to predict wages in the last section to estimate a logistic model where the dependent
variable was one if a worker who was unemployed in month t was classified as employed in
month t+1, and zero otherwise (i.e., if the worker remained unemployed or exited the labor
force). The model was estimated for the years 2004-06. We then used the coefficients from this
model to predict the job finding rate of the long-term unemployed (27 weeks or longer) based on
their characteristics each year.

31

Figure 8. Mean Predicted Probability of Transitioning From Unemployment to

Employment for the Long-Term Unemployed Based on Their Characteristics
Percent
45
40
35
30
25

2013

20

15
10
5
1995

1997

1999

2001

2003

2005

2007

2009

2011

2013

Note: Annual averages. Shading denotes recession. Predicted transition rate was derived by estimating a logistic
model where the dependent variable was one if a worker who was unemployed in month t was classified as
employed in month t+1, and zero otherwise (i.e., if the worker remained unemployed or exited the labor
force). Explanatory variables were: education, experience, industry, occupation, race, gender and marital status. The
model was estimated for the years 2004-06. The estimated coefficients from this model were then combined with the
characteristics of the long-term unemployed (defined as those unemployed for longer than 26 weeks at the time of
the survey) each year to predict the probability of transitioning to employment in the next month.
Source: National Bureau of Economic Research; Authors calculations using Current Population Survey
Longitudinal Population Database (see Nekarda, 2009).

The cyclical pattern suggests that there is a very slight shift in the characteristics of the
long-term unemployed in recessionary periods toward those that are more favorable for finding a
job, but the shift in the composition is very modest, predicting a rise in the job finding rate of
only about 1-2 percentage points. This is in contrast to the roughly 5 percentage point fall in the
job finding rate for the long-term unemployed in the past two recessions.
Notice also that the predicted job finding rate for the long-term unemployed based on
their characteristics is consistently around 25 percent according to Figure 8. However, Figure 7
shows that the job finding-rate for the long-term unemployed is consistently below that rate,

32

even in periods of a relatively strong job market. This contrast is consistent with the view that
the long-term unemployed face discrimination in the job market or become discouraged, or that
they possess unobserved characteristics that lead to lower job finding prospects or some
combination of all three.
Figure 9 displays the monthly labor-force withdrawal rates for the unemployed in each of
the duration groups from 1994 to 2013. A few patterns are noteworthy. First, the long-term
unemployed tend to have a higher rate of labor force exit than the short-term unemployed,
perhaps partly reflecting discouragement on the part of the long-term unemployed. Second,
labor force exit rates tend to drop in a recession, especially for the long-term unemployed.
Indeed, in the mild recession in the early 2000s, the labor force exit rate for the long-term
unemployed almost fell around 10 percentage points, to about the same level as the rate for
recently unemployed workers, and in the deep recession in 2008-09, the labor force withdrawal
rate force exit rate for the long-term unemployed again fell by around 10 percentage points, to
virtually the same level as that of the short-term unemployed. Third, the labor force exit rate
gradually rises for all duration groups after a recovery takes hold, and the rate rises more for the
long-term unemployed. In other words, after labor force exit rates collapse in recession to about
the same level for all duration groups, the exit rate tends to move towards its historical norm in
the recovery, with a higher exit rate for the long-term unemployed.

33

Figure 9. Probability of Transitioning From Unemployment to Out of the Labor Force

by Duration of Unemployment
Percent of Each Category of Unemployment Duration
34
32

Unemployed for
27 to 52 Weeks

30

Unemployed for
53 Weeks & Over

Unemployed Less
Than 5 Weeks

28
26
2013

24
22
20

18

Unemployed for
15 to 26 Weeks

Unemployed for
5 to 14 Weeks

16
1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

Note: Dotted lines represent 1994-2007 averages. Shading denotes recession.

Source: Bureau of Labor Statistics (Current Population Survey); National Bureau of Economic Research.

Figure 10 suggests that a relatively small part of the cyclical pattern in the labor force exit
rate for the long-term unemployed is due to compositional shifts. The figure shows the predicted
unemployment-to-out-of-the-labor-force transition rate based on the same characteristics and
approach used to construct Figure 8. Again, the scale is the same as in Figure 9. Although there
is a cyclical pattern in the composition of the unemployed, with those with a stronger attachment
to the labor force becoming long-term unemployed during a recession, movements in
composition would predict only about a 2 percentage point decline in the labor force withdrawal
rate in a recession, in contrast to the roughly 10 percentage point drop observed in the last two
recessions.

34

Figure 10. Mean Predicted Probability of Transitioning From Unemployment to Out of

the Labor Force for the Long-Term Unemployed Based on Their Characteristics
Percent
34
32
30
28
26
24

2013

22
20
18
16
1995

1997

1999

2001

2003

2005

2007

2009

2011

2013

Note: Annual averages. Shading denotes recession. Predicted transition rate was derived by estimating a logistic
model where the dependent variable was one if a worker who was unemployed in month t was classified as out of
the labor force in month t+1, and zero otherwise (i.e., if the worker remained unemployed or was classified as
employed). Explanatory variables were: education, experience, industry, occupation, race, gender and marital
status. The model was estimated for the years 2004-06. The estimated coefficients from this model were then
combined with the characteristics of the long-term unemployed (defined as those unemployed for longer than 26
weeks at the time of the survey) each year to predict the probability of transitioning to employment in the next
month.
Source: National Bureau of Economic Research; Authors calculations using Current Population Survey
Longitudinal Population Database (see Nekarda, 2009).

Labor force exits and their effect on the unemployment rate have been neglected in much
past research, although recent work by Elsby et al. (2013) suggests that changes in the
participation margin account for 33 percent of the cyclical variation in the unemployment rate.
For now, we focus on these flows in isolation, but it is important to bear in mind Shimers (2013)
observation that a decrease in the job-finding rate will indirectly raise the measured transition
rate from employment to unemployment, so the flows are best considered as part of a system.
As mentioned, during the most recent recession, and similarly to the recession of the
early 2000s, the rate of labor force withdrawal dropped for all durations of unemployment, but
35

most markedly for the long-term unemployed, and only a small part of this drop was a result of
compositional shifts. This phenomenon probably reflects, in part, the extension of
unemployment insurance benefits, which require workers to search for a job and has been shown
to induce unemployed workers to stay in the labor force, thus elevating the measured
unemployment rate (see Rothstein 2011, Farber and Valletta 2013). Many commenters have
predicted that as these benefits are exhausted or scaled back, the withdrawal rate for the longterm unemployed may begin to rise towards its historical average. By 2013, it appears that this
process has begun to take place for those who have been unemployed for over one year, but it is
less apparent for those who have been unemployed between 26 and 52 weeks.
As shown in the simulations below, the process of the labor force withdrawal rates of the
long-term unemployed moving toward their historical averages has important implications for
the unemployment rate, and, relatedly, the Beveridge Curve. Nevertheless, barring an
extraordinarily fast rebound in the labor force exit rates of the long-term unemployed relative to
their short-term unemployed counterparts, it appears likely that the long-term unemployment rate
will remain persistently high for a considerable amount of time, as was the case in Sweden and
the United Kingdom after their long-term unemployment rates spiked up.

Longer-Term Transitions
To investigate more fully whether the long-term unemployed are on the margins of the
labor market, we also look at transition rates for the long-term unemployed over longer periods
of time using matched data from the CPS. The CPSs rotation group design (interviewed four
consecutive months, out of the survey for eight months, and interviewed four more months)
makes it possible to examine transitions over a 15 month interval. The monthly transitions could

36

overstate the prospects of the unemployed if there are classification errors or if the jobs to which
the unemployed gain reemployment tend to be transitory.
A useful benchmark with which to compare longer term job finding rates is the implied
fraction of the unemployed who would be employed 15 months hence if the monthly job-finding
rate is constant for all workers and employment were an absorbing state. Then, given a flow into
employment of approximately 10 percent for the long-term unemployed, one would expect
nearly 80 percent of those who were long-term unemployed in a given month to be employed 15
months later, and 20 percent to remain jobless (either unemployed or out of the labor force).12
Of course, this calculation exaggerates the actual reemployment rate because workers can lose
their job soon after finding one or withdraw from looking for a job if they dont find one after a
period of time, but it provides a handy benchmark for thinking about the monthly labor force
transition rates, and would suggest that a job finding rate as low as 10 percent implies that the
long-term unemployed still have a reasonable attachment to employment.
In fact, however, actual long-term transition rates are considerably lower than those
implied by the monthly data. Figure 11 shows that since the beginning of the Great Recession,
36 percent of those who were long-term unemployed in a given month were employed 15
months later. Another 34 percent were not in the labor force, and 30 percent were unemployed
15 months later. Furthermore, of the 36 percent who were employed 15 months later, less than
one third had been employed full-time for four consecutive months. As a result, from 2008 to
2012, only 11 percent of those who were long-term unemployed in a given month returned to
steady, full-time employment a year later.13

12

Assuming independence and a constant 0.10 probability of finding a job in any given month, the proportion of
unemployed workers who gained employment within 15 months would be 1-(1-0.10)15 = 0.794.
13
Steady employment in this context means that someone who was unemployed for 27 weeks or longer in month t
was employed full-time for four consecutive months starting in month t+12.

37

Figure 11. Longitudinal Transition Rates for the Long-Term Unemployed, 2008-2013
Already Unemployed
> 27 Weeks

Month 1

Unemployed
(30.4%)

Employed
(35.9%)

29.7%

38.5%

Not in Labor
Force (33.7%)
31.8%

Month 16
Employed Full-Time
During Months 13-16
(10.7%)

Employed in Months
13-16, But Part-Time
at Least 1 Month
(13.8%)

Employed in 1-3
Months During
Months 13-16
(11.4%)

Source: Authors calculations using Current Population Survey Longitudinal Population Database (see Nekarda, 2009).
Note: Chart reflects the experience of those who were long-term unemployed in their fist Current Population Survey
interview (2008-2012) and their labor force status 15 months later (2009-2013).

If we compare these rates to an earlier period, we find that the longer term transition rates
for the long-term unemployed who were first surveyed in the CPS from 2008 to 2012 are not
very different from those exhibited by those who became long-term unemployed before the
Great Recession. In particular, the chance of a long-term unemployed worker transitioning to
employment 15 months later was 39 percent in 2005-07, as compared with 36 percent in the
period starting with the Great Recession. Likewise, only 12 percent of the long-term
unemployed were continuously reemployed in full-time jobs in month 13-16 in the 2005-07
period, versus 11 percent in the 2008-13 period. These results suggest that the major difference
between the 2008-13 period and the end of the last recovery is that there were many more long-

38

term unemployed workers in the latter period, not that they fared dramatically differently in the
labor market once they became long-term unemployed.
We should also note that the longer term employment transition rates for those who were
initially unemployed for less than 27 weeks when they entered the survey are substantially below
what might be suggested by their monthly transition rates. Only 50 percent of the short-term
unemployed in 2008-13 were employed 15 months later, which is higher than the 36 percent rate
for the long-term unemployed but still relatively low, and suggestive of the high risk of longterm unemployment facing the newly unemployed in this period.
Figure 12 provides a further disaggregated look at the transitions of the long-term
unemployed, which highlights the transitory nature of their employment opportunities. In
particular, the diagram divides the data by labor market status in the second, third, and fourth
months in the sample. The first notable observation from this figure is that only 23 percent of
the long-term unemployed in month one report being employed for one month or more in months
two through four. This compares to 11 percent who report being out of the labor force in months
two, three and four, and 67 percent who report having been unemployed in one month between
months two and four without moving to employment. Once the long-term unemployed leave the
labor force for three straight months, they were likely to stay out of the labor force, with only 32
percent reentering: a slightly larger share move into employment than unemployment, consistent
with Barnichon and Figura (2013).

39

Figure 12. Longitudinal Transition Rates for the Long-Term Unemployed, 2008-2013
Already Unemployed
> 27 Weeks

Month 1

Not Employed
at All During
Months 2-4
(66.5%)

Months 2-4

37.7%

Unmployed
(25.1%)

26.6%

29.1%

Employed
(19.4%)

37.0%

Not in Labor
Force During
Months 2-4
(10.5%)

Employed in at
Least 1 Month
(23.0%)

33.2%

17.0%

Not in
Labor
Force
(22.1%)

Unmployed
(3.9%)

36.4%

36.7%

65.3%

Employed
(15.0%)

41.1%

17.7%

13.8%

Not in
Labor
Force
(4.1%)

Unmployed
(1.4%)

22.3%

11.9%

18.1%

Employed
(1.9%)

36.0%

68.2%

Not in
Labor
Force
(7.2%)

52.1%

Month 16
Employed
Full-Time
During
Months 1316 (5.2%)

Employed
in Months
13-16, But
Part-Time
at Least 1
Month
(7.2%)

Employed
in 1-3
Months
During
Months 1316 (7.1%)

Employed
Full-Time
During
Months 1316 (5.5%)

Employed
in Months
13-16, But
Part-Time
at Least 1
Month
(6.2%)

Employed
in 1-3
Months
During
Months 1316 (3.3%)

Employed
Full-Time
During
Months 1316 (0.2%)

Employed
in Months
13-16, But
Part-Time
at Least 1
Month
(0.7%)

Employed
in 1-3
Months
During
Months 1316 (1.0%)

Source: Authors calculations using Current Population Survey Longitudinal Population Database (see Nekarda, 2009).
Note: Chart reflects the experience of those who were long-term unemployed in their fist Current Population Survey
interview (2008-2012) and their labor force status 15 months later (2009-2013).

Those who did move from not in the labor force for three straight months to employment
were mostly employed intermittently in the previous four months. Of the 23 percent who were
employed at least one of months 2 through 4 after having been long-term unemployed in period
1, 65 percent were also employed in month 16. But, even for this group, steady employment is
not as prevalent as may be expected. Those who stay unemployed in one of months 2 through 4
without moving into a job, displayed similar behavior to those who were initially long-term
unemployed as described in Figure 11, though with slightly greater movement to unemployment
40

than employment and not in the labor force. All of these results underscore that the long-term
unemployed face difficulty regaining full-time, steady work over the longest period we can
observe in CPS data. It appears that reemployment does not fully reset the clock for the longterm unemployed.

The Meaning of Not in the Labor Force for the Long-Term Unemployed
We next explore the reasons the long-term unemployed state when they leave the labor
force. This provides some purchase on whether the long-term unemployed are on the margins of
the labor force, and whether there will be continued movement toward the historical labor force
exit rate. Furthermore, it is important to know whether those who leave the labor force are likely
to be classified as marginally attached, since the marginally attached are more likely to reenter the labor market than others who are out of the labor force (see Barnichon and Figura, 2013
and Krusell, Mukoyama, Rogerson, and Sahin, 2011). Currently, the official BLS measures of
discouraged workers and marginally attached are relatively low. The U-5 measure of labor
underutilization, which includes all marginally attached workers and has as its denominator the
civilian labor force plus all persons marginally attached stood at 8.1 percent in December, just
1.4 percentage points above the headline unemployment rate.
Using linked CPS data, we tabulated responses by those who had been long-term
unemployed but then left the labor force to the following question, (Do / Does) (name/you)
currently want a job, either full or part time? This question is critical for the BLSs
classification scheme. Someone who is out of the labor force but indicates that he or she wants a
job is asked follow-up questions to determine their potential degree of discouragement.

41

Conversely, someone who is out of the labor force but indicates that he or she does not want a
job is precluded from being classified as marginally attached to the labor force.
Since the Great Recession, fully 73 percent of those who had been long-term unemployed
in month 1 and then left the labor force by month 16 indicated that they did not want a job in
month 16 of the survey. This number had been trending up over time (Barnichon and Figura,
2013). If the long-term unemployed were leaving the labor force for "economic reasons," it
would seem reasonable for such workers to indicate that they might still want a job. The
apparently high rate of permanent exits is consistent with the view that many of the long-term
unemployed were induced to search for a job and remain in the labor force longer than they
otherwise desired in order to qualify for extended UI benefits, and then left the labor force once
benefits expired (Rothstein, 2011 and Faber and Valletta, 2013). The large number of long-term
unemployed who say they do not want a job once they leave the labor force is consistent with the
hypothesis that the long-term unemployed are, to a large extent, on the margins of being in the
labor force.
This analysis included all respondents who classified themselves as long-term
unemployed in their first month of the survey, and thus, it is possible that a substantial portion
of these workers may have already been effectively out of the labor force by the time they were
surveyed in the CPS in month 16. To test whether or not this result essentially reflects a
misclassification of long-term unemployed workers in month one of the survey, we also
looked at those who reported being long-term unemployed every single month during months
one-four and months 13-15, and then left the labor force in month 16. Even with this severe
restriction on the consistency of reported long-term unemployment, more than 40 percent of

42

these long-term unemployed workers indicated that they did not want a job in the first month that
they left the labor force.
A follow-up question for those who report they do not want a job is, What best
describes (name's/your) situation at this time? For example, (are / is) (you/he/she) disabled, ill, in
school, taking care of house or family, or something ELSE? Typically, those who leave the
labor force because they no longer want a job report that they are either taking care of house or
family or in school (Hotchkiss, Pitts, and Avila, 2012).
Since the Great Recession, those who had been long-term unemployed in the initial
interview and then left the labor force by month 16 of the survey and reported that they no longer
wanted a job indicated that they were currently taking care of house or family (56 percent),
engaged in other unspecified activities (19 percent), and in school (16 percent).14 Other
possible responses, including retirement (2 percent), disability (3 percent), and illness (1
percent), had modest response rates.
The low rate of long-term unemployed workers who withdraw from the labor force and
report a disability as their reason for not wanting a job suggests that the Disability Insurance (DI)
program plays, at most, a minor role in incentivizing the long-term unemployed to withdraw
from the labor force, or in supporting them once they do withdraw from the labor force. This
observation is also consistent with Mueller, Rothstein and von Wachters (2013) conclusion that
Unemployment Insurance exhaustions and DI take up are unrelated.

14

In comparison to the long-term unemployed, those who had been unemployed for less than 27 weeks in month 1
and then left the labor force by month 16 and reported that they no longer wanted a job were: (1) twice as likely to
report that they were currently in school (32 percent versus 16 percent), and (2) less likely to report that they were
currently taking care of house or family (43 percent versus 56 percent).

43

Calibration Model
We previously presented time-series evidence suggesting that the shift in the Beveridge
Curve is due to the increase in long-term unemployment during the Great Recession. The
relationship between vacancies and unemployment appears to be stable if one uses the short-term
unemployment rate. One possibility is that, after a severe shock, the Beveridge Curve shifts out
because of slow job growth, a rise in long-term unemployment, and a decline in labor force exits,
particularly among the long-term unemployed. Moreover, the path of unemployment and
vacancies could eventually loop back to the original Beveridge Curve position because many of
the long-term unemployed exit the labor force or (less likely) find a job, and the unemployment
rate primarily reflects the historical share of short-term unemployed workers after a time.15 Our
goal in this section is to explore these two hypotheses with a calibrated model of labor force
flows and job matching in which the participation rate by duration eventually moves back to its
historical norm.
Specifically, we extend the calibration model in Kroft, et al. (2013). The approach has at
its core a straightforward method of accounting for the disparate labor market flows of the
unemployed by duration of unemployment. Kroft, et al.s partial equilibrium model focuses on a
search and matching framework that is a simplification of Mortensen and Pissarides (1994) and
Shimer (2005), ignoring the firm side, with slight tweaks to help fit the data (e.g., to adjust for
population growth and inconsistencies in flow data and reported durations). They focus on
workers age 25 to 54 to avoid issues concerning the aging of the baby boom and increased
school attendance. Their model, which was calibrated using data from 2002 to 2007, generates
predictions about the labor force movements of the unemployed by duration of unemployment.
Kroft, et al.s model captures most of the rise in the share of long-term unemployed workers as a
15

See Blanchard and Diamond for a discussion of loops around the Beveridge Curve.

44

result of the slowdown in job vacancies that accompanied the Great Recession. Their model
does not appear to generate a loop around the Beveridge Curve, however.16
We extend their model in two important respects. First, using data from 2002-07, we
estimate a matching function of the form:
= ( + + ) ()
where J is the number of jobs being filled by the nonemployed, is the number of short-term
unemployed workers, is the number of long-term unemployed workers, N is the number of
nonparticipants, and V is the number of vacancies. The parameters and reflect the lower
match efficiency for the long-term unemployed and nonparticipants. The short-term unemployed
are defined as those with less than 27 weeks of unemployment, while the long-term unemployed
are those with 27 weeks or more of unemployment. We estimate coefficients of =0.60 and
= 0.29. Kroft, et al. did not allow for a differential match parameter for the short-term and
long-term unemployed; their estimate of was similar to ours.
Kroft, et al. also assumed the same labor force withdrawal rate for the short-term and
long-term unemployed. While this is plausible in the immediate aftermath of a recession, over
time the labor force withdrawal rate tends to rise, especially for the long-term unemployed (see
Figure 7). Consequently, we allow for differential labor force withdrawal rates by duration of
unemployment.
We follow Kroft, et al. in using the observed number of vacancies, labor force
withdrawal rates, and transitions into unemployment as forcing variables in the model. We also

16

This observation is based on our replication of their model. In particular, when we replicated their exact model
we also found a rise in the share of the long-term unemployed that mirrored the observed data. When we projected
the Beveridge Curve using their model there was no shift in the curve after the Great Recession. Intuitively, this
finding resulted because their matching function generated job growth, and a consequent drop in unemployment,
that was stronger than observed in the recovery because it did not allow for a lower match rate of the long-term
unemployed.

45

follow Kroft, et al. and assign a duration of unemployment to those who initially transition from
nonparticipant to unemployed, and from employed to unemployed based on the observed
distributions in that calendar year. For those who remain unemployed from one period to the
next, we increment their duration of unemployment by one month.
Figure 13 uses the 2002-07 matching function to project the Beveridge Curve from 2008
to 2013. The projection seems to match the broader trends in the data reasonably well. The
calibrated model predicts an outward shift in the Beveridge Curve similar to what has been
observed. Moving into 2012 and 2013, the projection begins to move back towards the original
Beveridge Curve. There are some notable points where the simulation deviates from the data.
The projection initially under predicts the 5 percentage point rise in the unemployment rate from
January 2008 to October 2009 by one percentage point. This finding is consistent with Hall and
Schulhofer-Wohl (2013), who found that there has been a steady downward drift in matching
efficiency in the 2000s.17 After initially under predicting the actual unemployment rate, as the
number of long-term unemployed grows and the unemployed as a whole thus have a lower
matching efficiency, the job finding rate falls and the unemployment rate rises above the level
actually observed. As the labor force exit rate of both the long- and short-term unemployed
began to rise, however, the projection began to move back towards the original Beveridge Curve.
As of December 2013, the model predicts that the unemployment rate would be 0.7 percentage
point lower than the actual rate. As a whole, however, the calibrated model seems to capture the
broad outlines of the shift of the Beveridge Curve. The root mean square error between the
actual unemployment rate and projection based on the calibrated model is 0.8 percentage point.

17

Hall and Schulhofer-Wohls analysis focuses on eight different types of job seekers to control for heterogeneity.

46

Figure 13. Calibrated Beveridge Curve

Vacancy Rate
(As a share of 25-54 year old Labor Force)
6%

5%

Dec. 2016

Dec. 2015
Dec. 2014
Dec. 2013
4%

3%

2%

Actual Data
Projection from Dec. 2013
Calibrated Model

1%

0%
0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

Unemployment Rate for 25-54 year olds

To probe what the calibration exercise implies going forward, we extended the
calibration projections through 2016 starting with the data observed for December 2013.
Vacancies are treated as exogenous in the model; we assume that vacancies grow at the same rate
they had over calendar year 2013 (32,000 per month). The labor force exit rates of the long- and
short-term unemployed are assumed to linearly return to their 2006 averages by 2016, an
assumption that appears consistent with the 2002-07 recovery and current trends (see Figure 7;
notice also that the short-term exit rate is close to its pre-recession level). Under these
assumptions, the figure shows that by December of 2016 the labor market would almost have
returned to the original Beveridge curve. This implies that the combination of rising labor force

47

withdrawal rates and lower match efficiency for the long-term unemployed can account for a
loop around the Beveridge Curve.
Figure 14 shows the share of prime-aged workers who were predicted by the model to be
long-term unemployed. Although the projection does not match the actual data as closely as
Kroft, et al., the matching function does a relatively good job of capturing the rise in the share of
long-term unemployment from 2009 to 2010.18 An extension of the calibrated model implies
that as vacancies and matches rise, coupled with labor force withdrawal rates returning to their
earlier, higher levels, long-term unemployment is expected to decline gradually; though by the
end of 2016, however, the share remains well above pre-recession levels.
To assess the extent to which the fit of the calibrated model from 2008-2013 is driven by
the varying labor force withdrawal rates for the short-term and long-term unemployed, we also
conducted a counterfactual calibration exercise using the 2008-2009 overall average flow rate
from unemployed to out of the labor force for all durations each month instead of the durationspecific rate each month. The projections from this exercise fit the actual data much less well.
Although there is still a loop in the projected Beveridge Curve in this counterfactual projection
due to the fact that the matching function places a lower weight on the long-term unemployed,
the projection under predicts peak unemployment by nearly 2.5 percentage points. In total, the
omission of the actual unemployed to out of the labor force flows by duration results in a root
mean square error that is nearly twice as large as our base model: 0.8 percentage point compared
to 1.5 percentage points.

18

The rises in the share of long-term unemployed at the beginning of 2011 and 2012 are due to a feature of the
calibration model: at the beginning of the year the distribution of unemployment spells for those who entered
unemployment from employment or out of the labor force is updated to correspond to the actual distribution for such
workers in that calendar year.

48

Figure 14. Long-Term Share of Unemployment From Calibration

Share of Unemployed
Greater than 26 weeks
60%

50%

40%

calibration
30%

actual
projected from calibration

20%

10%

0%
Feb-08 Oct-08 Jun-09 Feb-10 Oct-10 Jun-11 Feb-12 Oct-12 Jun-13 Feb-14 Oct-14 Jun-15 Feb-16 Oct-16

We conclude from this exercise that the varying pattern of labor force withdrawal by
unemployment duration is an important feature of the job market. Moreover, the fact that a
similar pattern was observed in the past recovery suggests that labor force withdrawal of the
long-term unemployment rate is historically an important (but unfortunate) mechanism by which
the labor market returns to equilibrium.

Regional Differences within the U.S.

Some states are further along than others in recovering from the Great Recession than
others. We divide the states into those with high and low unemployment, based on their statewide unemployment rate in October 2013 and examine trends in long- and short-term
49

unemployment in those groups of states. Our analysis suggests that the long-term unemployment
rate has remained elevated even in low-unemployment rate states, defined as the 13 states with
the lowest unemployment rates in the U.S. in October 2013. The low unemployment rate states
many of which have benefited from energy production and a rebound in the agricultural sector
are Hawaii, Kansas, Minnesota, Montana, Nebraska, New Hampshire, North Dakota,
Oklahoma, South Dakota, Utah, Vermont, Virginia, and Wyoming; the average unemployment
rate in these states in December 2013 was 4.3 percent, compared to 7.0 in the rest of the country.
The comparative paths of short-term and long-term unemployment in the two sets of
states are shown in Figure 15(a) and 15(b), where the ratio of the short- and long-term
unemployment rates to their respective 1996-2007 averages are displayed. The states currently
with low unemployment were also on average low-unemployment states prior to the recession;
as a consequence, the historical averages we use for each group of states are from those states.
The figures show that even in states with low unemployment rates in October 2013, long-term
unemployment grew dramatically during the recession, reaching 4.5 times its historical average.
This is close to the peak ratio of just below 5.0 in the high-unemployment states. The long-term
unemployment rate does appear to be falling faster in the lower unemployment states, but it
remains at a historically high level, more than double its historical average as of December 2013.

50

Figure 15(a). Low Unemployment States:

Unemployment Rates by Duration Relative to
1996-2007 Averages

Figure 15(b). High Unemployment States:

Unemployment Rates by Duration Relative to
1996-2007 Averages

Ratio Relative to Historical Average

5.0

Ratio Relative to Historical Average

5.0 5.0

5.0

4.5

4.5 4.5

4.5

4.0

Unemployed for
27 Weeks or More

4.0 4.0

3.5

4.0

3.5 3.5

3.0

3.5
2013

3.0 3.0

2.5
2013

2.0
1.5
1.0
0.5

Unemployed for
27 Weeks or More

2.5 2.5

2.5

2.0 2.0

2.0

1.5 1.5

1.5

1.0 1.0
Unemployed for
26 Weeks & Less

0.0
1996 1998 2000 2002 2004 2006 2008 2010 2012

3.0

0.5 0.5

1.0
Unemployed for
26 Weeks & Less

0.0 0.0
1996 1998 2000 2002 2004 2006 2008 2010 2012

0.5
0.0

Note: Low unemployment states had a 2013 average unemployment rate below 6.3 percent as of October 2013. High unemployment states had
a 2013 average unemployment rate above 6.3 percent as of October 2013.
Source: Authors' calculations from the Current Population Survey.

Lastly, we examine transition rates from unemployment to employment and out of the
labor force by duration of unemployment for both sets of states. We do not find evidence that
the long-term unemployed are faring notably better in terms of transitioning to employment in
the low-unemployment states than in the high-unemployment states, as is shown in Figure 16(a).
The long-term unemployed appear to be following similar trends in transition rates into
employment in both groups of states, although the low-unemployment states have exhibited
slightly higher job-finding rates for the long-term unemployed since 1996. The job-finding rates
dropped in both groups of states during the Great Recession and have remained comparably low
during the recovery. The job-finding rates for the short term unemployed (not shown) also
exhibited similar trends.

51

Figure 16(a). Transition Rates From Long-Term

Unemployment to Employment
Percent
28%

Figure 16(b). Transition Rates From Long-Term

Unemployment to Out of the Labor Force
Percent
34%

26%

32%

24%
30%

22%

Low Unemployment
States

20%

High Unemployment
States

28%

18%

26%
2012

16%

24%

14%
22%

High Unemployment
States

12%

Low Unemployment
States

20%

10%

2012

8%
1996 1998 2000 2002 2004 2006 2008 2010 2012

18%
1996 1998 2000 2002 2004 2006 2008 2010 2012

Note: Low unemployment states had a 2013 average unemployment rate below 6.3 percent as of October 2013. High unemployment states had
a 2013 average unemployment rate above 6.3 percent as of October 2013.
Source: Authors' calculations from the Current Population Survey.

Figure 16(b) shows the flows from unemployment to out of the labor force by duration of
unemployment for high- and low-unemployment states. The graph shows that although the lowunemployment states do have slightly higher labor-force exit rates than the high-unemployment
states in the post-recession period, this is consistent with the historical pattern and the exit rates
are currently similar.
Overall, there is little evidence in this comparison to suggest that the long-term
unemployed fare substantially better in the states with the lowest unemployment rates, consistent
with the idea that the long-term unemployed are on the margins of the labor force, even where
the economy is stronger.

52

Conclusion
The Great Recession and subsequent modest recovery have been distinguished by an
exceptionally high rate of long-term unemployment. Our calibration exercise suggests that longterm unemployment rose for a simple reason: because job vacancies grew slowly in the economy
as a whole. The decline in job vacancies during the Great Recession set in motion a dynamic
that led to unprecedented long-term unemployment and a rise in the unemployment rate. Indeed,
using parameters of a job matching function that allows for lower match efficiency for the longterm unemployed and that was estimated prior to the recession, we are able to capture most of
the shift of the Beveridge Curve by the path of vacancies.
Historically, the long-term unemployed have faced great difficulty regaining steady, fulltime employment. The long-term unemployed also tend to withdraw from the labor force at a
higher rate than the short-term unemployed, although labor force withdrawal rates tend to
collapse during a recession. Our calibration exercise suggests that the historical dynamic of the
labor force participation rate of the long-term unemployed declining and then returning to its prerecession level, coupled with lower match efficiency for the long-term unemployed, can account
for loops around the Beveridge Curve. The Beveridge Curve tends to return to its original
position as the share of long-term unemployed falls.
Although the long-term unemployed have about a one in ten chance of moving into
employment in any given month, when they do return to work their new jobs are often transitory.
After 15 months, the long-term unemployed are more than twice as likely to have withdrawn
from the labor force than to have settled into steady, full-time employment. And when they exit
the labor force, the long-term unemployed tend to say that they no longer want a job, suggesting
that many labor force exits could be enduring. The subset of the long-term unemployed who do

53

regain employment tend to return to jobs in the same occupations and industries from which they
were displaced, suggesting that significant challenges exist for helping the long-term
unemployed to transition to growing sectors of the economy. A stronger macroeconomy helps
the long-term unemployed in part because it raises demand in their previous sectors. But even in
good times, the long-term unemployed are often on the margins of the labor market, with
diminished employment prospects and relatively high labor force withdrawal rates.
The portrait of the long-term unemployed in the U.S. that emerges here suggests that, to a
considerable extent, they are an unlucky subset of the unemployed. Their diverse and varied set
of characteristics implies that a broad array of policies will be needed to substantially lower the
long-term unemployment rate and stem labor force withdrawal, as concentrating on any single
occupation, industry, demographic group or region is unlikely to have a substantial impact
reducing long-term unemployment by itself. Understanding the labor market and personal
hurdles faced by the long-term unemployed should be a priority for future research in order to
craft solutions to reduce long-term unemployment.
Some may wish to draw macroeconomic policy implications from our findings. Only
time will tell if inflation and real wage growth are more dependent on the short-term
unemployment rate than total unemployment rate. To us, the most important policy challenges
involve designing effective interventions to prevent the long-term unemployed from receding
into the margins of the labor market or withdrawing from the labor force altogether, and
supporting those who have left the labor force to engage in productive activities. Overcoming
the obstacles that prevent many of the long-term unemployed from finding gainful employment,
even in good times, will likely require a concerted effort by policy makers, social organizations,
communities and families, in addition to appropriate monetary policy.

54

References
Abowd, John M., and Arnold Zellner. 1985. Estimating Gross Labor-Force Flows. Journal of
Business & Economic Statistics 3, no. 3: 254-283.
Abraham, Katharine, and Robert Shimer. 2002. Changes in Unemployment and Labor Force
Attachment. The Roaring Nineties, edited by Alan Krueger and Robert Solow (New
York: Russell Sage Foundation), pp. 367-420.
Ball, Laurence, and Sandeep Mazumder. 2011. The Evolution of Inflation Dynamics and the
Great Recession. Brookings Papers on Economic Activity Spring 2011: 337-381.
Barnichon, Regis, and Andrew Figura. 2013. Declining Labor Force Attachment and
Downward Trends in Unemployment and Participation. Finance and Economics
Discussion Series Working Paper 2013-88. Board of Governors of the Federal Reserve
System.
Barnichon, Regis, and Christopher J. Nekarda. 2012. The Ins and Outs of Forecasting
Unemployment: Using Labor Force Flows to Forecast the Labor Market. Brookings
Papers on Economic Activity Fall 2012: 83-131.
Blanchard, Olivier Jean and Peter Diamond. 1989. The Beveridge Curve. Brookings Papers on
Economic Activity 1989:1: 1-76.
Davis, Steven J. Comment on Krueger and Mueller. 2011. Brookings Papers on Economic
Activity Spring 2011, 58-70.
Elsby, Michael W.L., Bart Hobijn, Aysegl Sahin, and Robert G. Valletta. 2011. The Labor
Market in the Great RecessionAn Update to September 2011. Brookings Papers on
Economic Activity Fall 2011: 353-384.
Elsby, Michael W.L., Bart Hobijn, and Aysegl Sahin. 2013. On the Importance of the
Participation Margin for Labor Market Fluctuations. Federal Reserve Bank of San
Francisco Working Paper Series 2013-05.
Farber, Henry S., and Robert G. Valletta. 2013. Do Extended Unemployment Benefits Lengthen
Unemployment Spells? Evidence from Recent Cycles in the US Labor Market. Working
Paper no. 19048. National Bureau of Economic Research.
Ghayad, Rand. 2013. The Jobless Trap. Working Paper. http://media.wix.com/ugd/
576e9a_f6cf3b6661e44621ad26547112f66691.pdf.
Gordon, Robert. 2013. The Phillips Curve is Alive and Well: Inflation and the NAIRU During
the Slow Recovery. Working Paper no. 19390. National Bureau of Economic Research.

55

Hall, Robert E. and Sam Schulhofer-Wohl. 2013. Measuring Matching Efficiency with
Heterogeneous Jobseekers, mimeo, Stanford University.
Heckman, James, and Burton Singer. 1984. A Method for Minimizing the Impact of
Distributional Assumptions in Econometric Models for Duration Data." Econometrica:
Journal of the Econometric Society 52, no. 2: 271-320.
Hobijn, Bart and Aysegl Sahin. 2012. Beveridge Curve Shifts across Countries since the Great
Recession. 13th Jacques Polak Annual Research Conference.
Hotchkiss, Julie, M., Melinda Pitts, and Fernando RiosAvila. 2012. A Closer Look at
Nonparticipants During and After the Great Recession. Federal Reserve Bank of Atlanta
Working Paper Series 2012-10.
Katz, Lawrence F. and Alan B. Krueger. 1999. The High-Pressure U.S. Labor Market of the
1990s. Brookings Papers on Economic Activity 1999:1, 1-65.
Kroft, Kory, Fabian Lange, and Matthew J. Notowidigdo. 2013. Duration Dependence and
Labor Market Conditions. Quarterly Journal of Economics 128, no. 3: 1123-1167.
Kroft, Kory, Fabian Lange, Matthew J. Notowidigdo, and Lawrence F. Katz. 2013. Long-Term
Unemployment and the Great Recession: The Role of Composition, Duration
Dependence, and Non-Participation. Working Paper. http://scholar.harvard.edu/files/
lkatz/files/klnk_ltu_and_great_recession_sep16_2013.pdf
Krueger, Alan B. and Andreas Mueller. 2011. Job Search, Emotional Well-Being, and Job
Finding in a Period of Mass Unemployment: Evidence from High-Frequency
Longitudinal Data. Brookings Papers on Economic Activity Spring 2011, 1-57.
Krusell, Per, Toshihiko Mukoyama, Richard Rogerson, and Aysegl Sahin. 2011. A Three State
Model of Worker Flows in General Equilibrium. Journal of Economic Theory 146, no.
3: 1107-1133.
Llaudes, Ricardo. 2005. The Phillips Curve and Long-Term Unemployment. European
Central Bank Working Paper Series no. 441.
Mortensen, Dale T., and Christopher A. Pissarides. 1994. Job Creation and Job Destruction in
the Theory of Unemployment. The Review of Economic Studies 61, no. 3: 397-415.
Mueller, Andreas. Separations, Sorting and Cyclical Unemployment. IZA Discussion Paper
No. 6849, 2012.
Mueller, Andreas, Jesse Rothstein, Till von Wachter. Unemployment Insurance and Disability
Insurance in the Great Recession, NBER Working Paper No. 19672, November 2013.

56

Nekarda, Christopher J. 2009. A Longitudinal Analysis of the Current Population Survey:

Assessing the Cyclical Bias of Geographic Mobility. Working Paper.
http://chrisnekarda.com/papers/geographic-mobility.pdf. Data were updated through
December 2013.
Polivka, Anne E. and Stephen M. Miller. 1994. The CPS After the Redesign: Refocusing the
Economic Lens. CRIW Labor Statistics Measurement Issues Conference.
Poterba, James M. and Lawrence H. Summers. 1986. Reporting Errors and Labor Market
Dynamics. Econometrica: Journal of the Econometric Society 54, no. 6: 1319-1338.
Rothstein, Jesse. 2011. Unemployment Insurance and Job Search in the Great Recession.
Working Paper no. 17534. National Bureau of Economic Research.
Shimer, Robert. 2012. Reassessing the Ins and Outs of Unemployment. Review of Economic
Dynamics 15, no. 2: 127-148.
. 2005. The Cyclical Behavior of Equilibrium Unemployment and Vacancies. American
Economic Review 95, no. 1: 25-49.
Staiger, Douglas, James H. Stock, and Mark W. Watson. 1997. The NAIRU, Unemployment,
and Monetary Policy. Journal of Economic Perspectives 11, no. 1: 33-49.
Wanberg, Connie R., Jing Zhu, Ruth Kanfer, and Zhen Zhang. 2012. After the Pink Slip:
Applying Dynamic Motivation Frameworks to the Job Search Experience. Academy of
Management Journal 55: 261-264.

57

Brookings Panel on Economic Activity

March 2021, 2014

Abenomics: Preliminary Analysis and

Outlook
Joshua K. Hausman, University of Michigan
Johannes F. Wieland, University of California, San Diego
Final conference draft

Abstract
In early 2013, Japan enacted a monetary regime change. The Bank of Japan set a two percent inflation
target and specified concrete actions to achieve this goal by 2015. Shinzo Abes government is supporting
this change with fiscal policy and structural reforms. We show that Abenomics ended deflation in 2013
and raised long-run inflation expectations.
Our estimates suggest that Abenomics also raised 2013 output growth by 0.9 to 1.7 percentage points.
Monetary policy alone accounted for up to a percentage point of growth, largely through positive effects
on consumption. In the medium and long-run, Abenomics will likely continue to be stimulative.
But the size of this effect, while highly uncertain, thus far appears likely to fall short of Japans large
output gap. In part this is because the Bank of Japans two percent inflation target is not yet fully credible.
We conclude by outlining how to interpret future data releases in light of our results.

Abenomics: Preliminary Analysis and

Outlook
Joshua K. Hausman
University of Michigan

Johannes F. Wieland
University of California, San Diego

March 2014 Conference Draft

Abstract
In early 2013, Japan enacted a monetary regime change. The Bank of Japan set a two
percent inflation target and specified concrete actions to achieve this goal by 2015. Shinzo
Abes government is supporting this change with fiscal policy and structural reforms. We
show that Abenomics ended deflation in 2013 and raised long-run inflation expectations.
Our estimates suggest that Abenomics also raised 2013 output growth by 0.9 to 1.7 percentage points. Monetary policy alone accounted for up to a percentage point of growth,
largely through positive effects on consumption. In the medium and long-run, Abenomics
will likely continue to be stimulative. But the size of this effect, while highly uncertain,
thus far appears likely to fall short of Japans large output gap. In part this is because
the Bank of Japans two percent inflation target is not yet fully credible. We conclude by
outlining how to interpret future data releases in light of our results.

Hausman: Ford School of Public Policy and Department of Economics, University of Michigan. 735 S. State
St. #3309, Ann Arbor, MI 48109. Email: hausmanj@umich.edu. Phone: (734) 763-3479. Wieland: University
of California, San Diego, Department of Economics, 9500 Gilman Dr. #0508 La Jolla, CA 92093-0508. Email:
jfwieland@ucsd.edu. Phone: (510) 388-2785. We are grateful for extensive comments from the editors, David
Romer and Justin Wolfers, and our discussants, Ben Bernanke and Paul Krugman. We have benefitted from
conversations and assistance from Gabriel Chodorow-Reich, Catherine Hausman, David Hausman, Koichiro Ito,
Stefan Nagel, and Mu-Jeung Yang, as well as from comments from seminar participants at UCSD. We also
thank Walid Badawi, Matthew Haarer and Ben Meiselman for superb research assistance, and the International
Policy Center at the Ford School of Public Policy for financial support.

1 Introduction
A great monetary experiment is taking place in Japan today. In early 2013, the Bank of
Japan announced a monetary policy regime change. Along with this monetary expansion, the
government is enacting complementary fiscal policy and structural reforms. The hope is to end
two decades of stagnation and deflation. In this paper, we provide a preliminary evaluation of
these policies. We argue that Abenomics ended deflation in 2013 and raised output growth by
0.9 to 1.7 percentage points. Going forward, Abenomics will continue to have benefits, but so
far appears unlikely to close Japans large output gap. In part this is because the monetary
regime change is not (yet) fully credible.
That Japan needs some new policies is clear. The Japanese economy has stagnated since
1992. Between 1993 and 2012, real GDP growth averaged just 0.8 percent.1 Prices have fallen
most years since 1998. Economists have blamed Japanese policymakers for an insufficiently
aggressive response to these trends.2 No longer. Shinzo Abe became prime minister on December 26th , 2012. A member of the Liberal Democratic Party, Abe campaigned on a platform
of radical action to end economic stagnation. His economic program (dubbed Abenomics)
consists of monetary expansion, fiscal stimulus, and structural reforms. In a reference to a
Japanese legend, these components are referred to as three arrows (Eichengreen, 2013).
The first arrow is a monetary policy regime change. Beginning in November 2012, thencandidate Abe argued that that the Bank of Japan should increase its inflation target and
engage in unlimited easing. After his election on December 16th , Abe threatened to revise the
law granting the Bank of Japan independence if it did not agree to a higher inflation target.3
The Bank of Japan acceded to Abes demand, announcing a two percent inflation target at
its meeting on January 22nd , 2013. While hardly extreme, two percent inflation would be the
highest year-on-year inflation rate in Japan since 1991. In what we show was a more significant
announcement, on April 4th the new Bank of Japan governor, Haruhiko Kuroda, promised
to reach this target in two years through open-ended asset purchases and a doubling of the
1

The data appendix describes our sources for series cited throughout the text.
See Bernanke (2000) and Ito and Mishkin (2006) among many others.
3
See Reuters and Financial Times on unlimited easing and Reuters on Abes threat.
2

monetary base (Bank of Japan, 2013b).

The second arrow is fiscal policy. In February 2013, the Diet passed a two percent of GDP
supplementary budget (Ito, 2013), although the actual stimulus being carried out is much
smaller than this headline number suggests.4 Our preferred measure of stimulus size compares
the cyclically adjusted primary budget balance forecast by the IMF before Abes fiscal measures
were announced with that forecast in late 2013. This suggests actual stimulus in 2013 of one
percent of GDP.5 This stimulus is dwarfed by upcoming tax increases. Consumption taxes will
rise from five to eight percent in April 2014 and by a further two percentage points in October
2015 (Ito, 2013). Thus the IMF projects that the cyclically adjusted primary budget deficit
will fall from 8.5 percent of potential GDP in 2013 to 6.0 percent in 2014 and 4.8 percent in
2015. While the fiscal consolidation package precedes Abeit passed the lower house of the
Diet on June 26th 2012it is now often treated as part of Abenomics. Unless otherwise noted,
however, we report the effects of Abenomics excluding the consumption tax, although we show
that including its effects does not change our qualitative conclusions.
The third arrow consists of structural reforms to increase Japans potential GDP growth.
The proposed reforms are thus far mostly vague, but include relaxations of labor market rigidities, less protection for farmers, and utility deregulation.6 These reforms may be made more
credible by the Japanese governments part in the Trans-Pacific Partnership, which suggests a
willingness to take on special interests.7
For substantive and pragmatic reasons, we focus primarily on the first arrow of Abenomics,
the monetary policy regime change. Monetary policy is the most clearly defined element of
Abenomics. It is also the newest and most radical. Abes government is not the first to try fiscal
stimulus or fiscal consolidation.8 And Abes government is not the first to push for structural

Throughout the 1990s, actual fiscal stimulus also usually fell short of headline numbers (Posen, 1998).
Abes fiscal expansion was originally conceived of as a one-time program, but in October 2013 the Abe
government announced another one percent of GDP supplementary budget to offset adverse effects from mediumrun fiscal consolidation. (See Davies (Financial Times), and IMF.) Like the previous expansionary measures,
we suspect that the headline number is an overestimate of the actual spending that will occur.
6
Economist.
7
See Deardorff (2013) and Financial Times.
8
Posen (1998) and Kuttner and Posen (2001) analyze Japanese fiscal policy in the 1990s.
5

reforms.9 By contrast, the Bank of Japans current policies are a clear break from previous
attempts at unconventional monetary policy. Japans monetary policy is also likely to be the
element of Abenomics of most interest to other countries. There are many recent examples of
fiscal stimulus programs and attempts at structural reform. Monetary policy regime changes
are less common and less well understood.
We begin our analysis in section 2 by considering Japans recent macro history and the size
of its current output gap. In order for monetary policy to be an effective stimulus, there must
be substantial unused resources in the economy. Using three different approaches, we find that
the current output gap is in fact quite large, in the range of 4.5 percent to 10 percent. Thus it
can at least be hoped that demand side policies will have large effects.
In section 3, we consider Abenomics effects in 2013. We start by examining the financial
market response. Over the year, the yen fell 21 percent against the dollar, and the Nikkei 225
stock market index rose 57 percent. Measured by inflation swaps and surveys of professional
forecasters, long-run inflation expectations have risen to between 1 and 1.4 percent. Combined
with a small decline in nominal interest rates, this means ten-year real interest rates have fallen
nearly a percentage point. Inflation also responded as hoped. The twelve-month percent change
in the CPI rose from negative 0.1 percent in December 2012 to positive 1.6 percent in December
2013. Much of this increase was driven by the effects of the weaker yen on imported energy
and food, but even the CPI excluding food and energy rose over the year.
Growth was also decent. Year-over-year, 2013 output growth was 1.5 percent, 0.9 percentage
points better than expected by professional forecasters in December 2012, before Abe took office.
Statistical (VAR) forecast counterfactuals imply possibly larger effects of Abenomics on 2013
growth, on the order of 0.9 to 1.7 percentage points.
That Abenomics contributed to a good year in 2013 is clear. Less clear is how much
expansionary monetary policy as opposed to one-off fiscal stimulus drove that growth. A clue
comes from the behavior of consumption. As we discuss in section 3.6, both the time path and
composition of consumption strongly suggest that expansionary monetary policy accounts for
9

Hoshi and Kashyap (2011) discuss reforms made under the Koizumi administration.

a significant part of 2013 consumption growth, perhaps contributing as much as a percentage

point to overall GDP growth.
In the medium and long-run, Abenomics, and Japans new monetary policy in particular, is
likely to continue to pass a cost-benefit test. The costs are few or none, and the benefits both
to output and the government budget are potentially large. But, as of yet, it appears unlikely
that Abenomics will fully close the output gap.
In section 4, we look to four sources of evidence on the future effects of Abenomics: (1)
forecasts from professional forecasters; (2) the stock market; (3) the analogy to the U.S. in
1933; and (4) new and old Keynesian models. Professional forecasts suggest that Abenomics
has raised both the level and the growth rate of GDP. But the forecast gains are modest
relative to the output gap. Excluding negative effects from the consumption tax increase, by
2022, GDP is forecast to be 3.1 percent above the no Abenomics baseline. The stock markets
2013 boom might seem to forecast larger gains. But a Campbell-Shiller decomposition shows
that, historically, the stock market has been a poor predictor of dividend growth and thus
likely an even worse predictor of GDP growth. Also initially suggesting optimism is the success
of Franklin Roosevelts monetary regime change in spring 1933. This is often taken as an
example of the large potential effects of a monetary regime shift. So far, however, the effects of
Abenomics have been an order of magnitude smaller. Likewise, new Keynesian models suggest
that higher expected inflation and a lower real interest rate can be highly stimulative. But
again this prediction is at odds with Japans more modest experience in 2013.
An important caveat is that we treat Abenomics as is. That is, we take as given the effect of
Abes policies on expected inflation and expected future output. This means we are analyzing
a policy that has a stated goal of two percent inflation but has not yet convinced the public
that the goal will be achieved. Market- and survey-based inflation forecasts suggest that the
two percent inflation target is not (yet) credible, presumably because there is some probability
that Abe or the Bank of Japan will change course. Thus we are measuring the effects of a two
percent inflation target times the probability that the target is achieved. We argue in section 5
that the Bank of Japans credibility problems are likely a product of its past actions and the
4

interaction of Japans demographics and non-inflation-indexed pensions.

The current lack of credibility means that if the Bank of Japan does succeed in raising
inflation expectations to two percent, output effects are likely to be much larger than current
indicators suggest. Holding nominal rates fixed, full credibility would double the decline in
the real interest rate that has occurred thus far. We would therefore expect another boost
to GDP commensurate with our estimate of the monetary policy contribution to 2013 GDP
growth. In the medium-run the gains are likely to be even larger. But with current data it is
difficult to produce an estimate of this effect. For instance, long-run forecast revisions do not
allow us to distinguish monetary effects from structural reform expectations. In section 6 we
do, however, offer a brief roadmap to interpret future data releases and forecasts revisions. We
provide guidance to help distinguish between the effects of the monetary and structural reform
channel going forward.

2 Context
2.1 A brief history In the 1980s, Japans economy was the envy of other nations: GDP had
been growing rapidly for decades, the stock and property markets were booming, and Japanese
production techniques were widely regarded as superior.10 This all changed in the 1990s. From
December 1989 to August 1992, the Nikkei 225 fell almost 60 percent. Land prices in six large
Japanese cities fell by 50 percent from 1991 to 1996 and continued to decline thereafter. The
fallout from the bursting of these asset bubbles dragged down the economy. Economists have
often criticized the Bank of Japan for not responding more quickly to these asset price declines
and the resulting economic slowdown. By 1996, the Bank of Japan found itself effectively stuck
at the zero lower bound on nominal interest rates. Thus the Japanese experience in the 1990s
bears remarkable similarities to the period since 2008 in the U.S. and Europe.
Panel A of table 1 summarizes this story.11

After growing rapidly until 1991, annual

Japanese real GDP growth slowed to roughly one percent. Inflation fell to near zero, with
prices falling most years after 1998. Nominal interest rates fell more than inflation. To U.S.
10
11

For background on the Japanese economy before 1990, see Ito (1992).
This table is a deliberate updating of Krugman (1998), table 4.

Table 1 Macro summary statistics

Panel A: Raw data
Year

1974-1992 average
1993-2007 average
2008-2012 average

Real GDP
growth
(% change)

Unemployment
rate
(%)

CPI
inflation
(%)

Money market
interest rate
(%)

4.0
1.1
-0.2

2.3
4.1
4.6

4.8
0.1
-0.2

6.8
0.6
0.2

Panel B: Adjusted for working age population

Year

1974-1992 average
1993-2007 average
2008-2012 average

Real GDP per

person age 15-64
(% change)
Japan
U.S.
3.1
1.4
0.4

1.5
1.9
0.2

Multifactor
productivity
(% change)
Japan
U.S.
0.9
0.3*

1.2
0.8*

Employment /
pop. age 15-64
(%)
Japan
U.S.
67.4
69.3
70.8

68.1
72.5
67.8

Sources for Japanese and U.S. data: See data appendix. *2008-2011 average.

observers, the Japanese unemployment rate remained puzzlingly low. Its peak during the 200809 recession was 5.5 percent. But this reflects particular Japanese labor market institutions.
Many Japanese workers have de facto lifetime tenure in their jobs (Kato, 2001). Thus when
output falls, they remain employed.
Japans economy was hard hit by the 2008 recession. From its peak in the first quarter of
2008 to its trough in the first quarter of 2009, output fell 9.2 percent. Since then, recovery has
been incomplete. GDP in the last quarter of 2013 was still below that in the first quarter of
2008. In part this is because a nascent recovery in 2010 was reversed in 2011: GDP growth
swung from +4.7 percent in 2010 to -0.5 percent in 2011. A major culprit was the devastating
March 11th , 2011 Tohoku earthquake and tsunami. This disaster killed almost 20,000 people
and destroyed 275,000 homes (Schnell and Weinstein, 2012).
Poor headline GDP numbers in part reflect Japans low birth rate. The Japanese National
Institute of Population and Social Security Research projects that the Japanese population
peaked in 2010 and has since begun declining.12 Thus GDP per capita growth now exceeds
GDP growth. Furthermore, as Japans population has aged, the growth rate of the working
12

See http://www.ipss.go.jp/p-info/e/psj2012/PSJ2012.asp.

age population has generally been below that of the population as a whole. The number of
people age 15 to 64 peaked in 1995 at 87 million; by 2010 there were only 81 million. Given
these demographic trends, it is unreasonable to expect Japanese GDP growth to match that in
countries with growing populations, like the U.S.
A crude way to correct for the growth consequences of Japans demographics is to measure
the growth of GDP per person age 15 to 64. Columns 1 and 2 of panel B in table 1 do this
for Japan and the U.S. since 1974. With this adjustment, the Japanese growth slowdown is
less pronounced. And the Japanese economys performance relative to the U.S. is better than
some popular accounts would suggest. This is not to say that Japan did well. From 1992 to
2007 GDP per working age person grew 0.5 percentage points more slowly in Japan than in
the U.S. A comparison of multifactor productivity has similar implications: from 1992 to 2007
multifactor productivity grew 0.3 percentage points more slowly in Japan. Thus as of 2012,
Japans PPP-adjusted real GDP per person age 15 to 64 was only 71 percent of that in the U.S.
Columns 5 and 6 of panel B in table 1 compare the employment to population ratio for people
age 15 to 64 in the two countries. This ratio was lower in Japan than in the U.S. prior to the
2008 recession. But during the recession, the employment to population ratio fell sharply in the
U.S. while remaining fairly stable in Japan. So as of 2012, more of the working age population
in Japan is employed.
Slow growth, an enormous natural disaster, and a shrinking population are not the only
headwinds facing the Japanese economy. Japan also has a large sovereign debt burden. The
gross debt to GDP ratio rose from 66 percent in 1991 to 244 percent today. Net debt in 2013 is
an only somewhat less staggering 140 percent of GDP. This debt burden has thus far coexisted
with extraordinarily low nominal and real interest rates. As of February 19th , 2014 the nominal
interest rate on ten-year Japanese government bonds was 0.6 percent while expected ten-year
inflation (from inflation swaps) was 1.1 percent. Yet some economists fear that Japan is close
to the point at which investors will no longer tolerate an increasing debt to GDP ratio (e.g.,
Hoshi and Ito, 2012). This debt burden adds to the sense of many that the past twenty years
were two lost decades.
7

2.2 The output gap Japans two decade stagnation was likely a product of both demand and
supply side problems. A clear separation is difficult. Financial sector problems, for instance,
affect both investment demand and firm marginal costs. But some decomposition is important;
if Japans problems were entirely the result of changes to potential output, then a monetary
regime change could have little real effect. In order to know what a demand side policy can
achieve, we need to know the size of Japans output gap.
Official estimates of Japans output gap are small. In its October 2013 World Economic Outlook, the IMF estimated that Japans 2013 output gap would be -0.9 percent. In its November
2013 Economic Outlook, the OECD estimated that Japans 2013 output gap would be positive
0.1 percent. These measures suggest that there is little role for demand side policy in Japan:
neither fiscal stimulus nor lower real interest rates will raise output if supply constraints are
binding. In this case, the only arrow of Abenomics that matters is the third.
The IMF and OECD measures, however, almost certainly have difficulty with Japans two
decade stagnation. Prolonged slumps can mean that conventional estimates of potential output
which use, for instance, a Hodrick-Prescott filter, interpret the slowdown in actual growth to
have been primarily a slowdown in potential output growth (Krugman, 1998). While the
IMF currently uses a production function approach, not a Hodrick-Prescott filter, to measure
Japanese potential output, its methods suffer from a similar problem: the IMFs potential
output measure closely tracks forecast and actual output (figure 1(a)).13 Between 2008 and
2013, the IMFs estimate of 2009 potential output growth fell from 1.7 percent to 0.3 percent.
The decline in the IMFs estimate of Japanese potential output during the recent recession is
analogous to the Congressional Budget Offices revision of its estimate of U.S. potential output.
In 2007, the Congressional Budget Office believed that potential output would grow at a 2.7
percent annual rate from 2007 to 2013. Now, in 2014, the Congressional Budget Office believes
it grew at a 1.7 percent annual rate.14 Whether this is desirable depends in part on the time
horizon of interest. A large part of the CBOs revision to U.S. potentialand, we presume, to
13
The IMF publishes a forecast for actual GDP and the output gap as a percent of potential in the World
actual
Economic Outlook. We computed the implied level of potential output as Ypotential = Y1+gap
.
14
See http://www.cbo.gov/publication/45068.

115

2007

2009

Actual real GDP

Potential: 1998-2006 GDP/working pop. trend
Potential: 10/2007 forecast
Potential: Okun's Law (Krugman,1998)

Actual and potential real GDP, 2007=100

95
100
105
110
115

95

Actual and potential real GDP, 2007=100

100
105
110

Actual real GDP

WEO 10/2013: Potential output forecast
WEO 4/2008: Output forecast
WEO 4/2008: Potential output forecast

2011

2013

(a) IMF forecasts of actual and potential output

2007

2009

2011

2013

(b) Trend and forecast potential output estimates

Figure 1 Japanese actual and potential output.

the IMFs revision to Japanese potentialcomes from lower investment (Jacobson and Occhino,
2013). A smaller capital stock certainly diminishes the level of output achievable at a given
point in time, but over several years, investment shortfalls can be made up. So it is not obvious
how much a period of lower investment should affect ones view of what a demand side policy
can achieve in the medium-run. A historical example may clarify. Investment was low in the
U.S. throughout the 1930s. But in part because of stimulative monetary policy after 1933 and
stimulative fiscal policy during World War Two, the Great Depression does not appear to have
lowered potential or actual U.S. output in 1950.
Hysteresis effects in the labor market, such as high long-term unemployment and early
retirement, may also lower potential. But it seems likely that a prolonged boom could reverse
many of these effects. Ball (2009) finds that large demand expansions are associated with
declines in the natural rate of unemployment and, by implication, increases in potential output.
The CBO and the IMF measure potential output conditional on all resources being employed
at their current capacity. But for the purpose of evaluating the potential effects of a prolonged
demand side policy, we believe the correct measure of potential output conditions on resources
having been close to fully-employed for some time. Thus we do not believe that the IMF
measurean output gap of -0.9 percentis a good indicator of what sustained expansionary
monetary policy should aim for in Japan. We suspect that if Japanese output growth is high
9

for several years, the IMF estimate of potential output will rise. If one disagrees with this
assessment, one may find the official, small measures of Japans output gap unproblematic. In
that case, one would expect little from monetary policy other than price effects. Presumably,
like us, Japanese officials do not hold this view; if they did, it is unclear why they would have
gone to the effort of staging a monetary regime shift.
We are not the first to struggle with estimating Japans output gap. Krugman (1998) also
argued that official estimates of Japans output gap were too small. Instead of relying on official
figures, he combines an estimate of Japans natural rate of unemployment with an estimate of
the Okuns law coefficient in Japan. Krugman estimates Japans natural rate of unemployment
as 2.5 percent. This was Japans average unemployment rate in the decade from 1982 through
1991. Again using data from the 1980s, Krugman argues that Japans Okuns law coefficient
is roughly six. These two estimates, combined with Japans 2013 unemployment rate (4.0
percent), suggest that the output gap in 2013 was roughly 6 (4 2.5) = 9 percent.
Of course the unemployment rate in the 1980s may no longer be a good indicator of Japans
natural rate of unemployment. Even ignoring hysteresis effects, demographic changes, for
instance, have likely changed the natural rate.15 Another approach to estimating potential
output is to assume that output continues to grow at its pre-depression trend. This is the
approach taken by Romer (1999) for the Great Depression. We use Japans trend growth rate
of output per working age person from 1998 to 2006 to obtain a measure of Japans current
output gap. 1998 and 2006 are sensible start and end dates since the unemployment rate was
the same in these two years. Normalizing the trend by the working age population incorporates
Japans changing demographics since the 1990s. Figure 1(b) shows this trend along with actual
output. We assume that output was at potential in 2007. This suggests a 2013 output gap of
4.6 percent. Many observers would, however, argue that output in 2007 was below potential.
15

Unfortunately, recent data are not well suited to estimating Japans natural rate: Japans inflation rate
over the past several years has been negative but not falling. This has unclear implications for the natural
rate. For instance, Ball (2006) and Blanchard (2000) suggest that the Phillips curve relationship in Japan may
be between the level of inflation and slack rather than the more conventional relationship between changes in
the inflation rate and slack. The challenges to estimating the natural rate of unemployment in Japan now are
similar to those that plague any such calculation for the U.S. in the 1930s. Just as in Japan now, low output
and employment in the 1930s U.S. coexisted with increases in the inflation rate (Romer, 1999).

10

Okuns law implies that the 2007 output gap was in fact 8 percent (figure 1(b)). Thus our
trend-based measure is biased towards underestimating Japans current output gap. Bias in
the other direction comes from effects of the 2008 recession and the 2011 tsunami on potential
output. As discussed above, it is not obvious that a potential output measure that responds
strongly to recessions is correct for our purposes. But the 2011 tsunami raises separate issues.
This disaster not only caused enormous physical destruction, it also led to the shutdown of all
Japanese nuclear power plants. This raised energy prices, with economy-wide negative effects
on aggregate supply (Schnell and Weinstein, 2012). Unfortunately, quantifying the effect on
potential output is difficult. A hint that it was small comes from the IMFs October 2013
World Economic Outlook, which shows no slowdown in potential output growth between 2010
and 2011.
A final way to measure the output gap is to see what professional forecasters expected longrun growth to be in 2007. Absent any large and expected demand side policies, these forecasts
are likely to measure observers best guess of trend growth. Like the previous measure, this
ignores any effects of the 2008 recession or the 2011 tsunami on potential output, but it may
also suffer from the opposite bias, in that forecasters may not have believed that the economy
would return to potential in 2013. In any case, this forecast is also shown in figure 1(b). It
suggests a 2013 gap of 10 percent, similar to that using the Okuns law method. Therefore
these three measures suggest a large output gap in Japan, in the range of 4.5 to 10 percent.

3 2013 Impact
In this section, we examine the effect of Abenomics on the Japanese economy in 2013. First,
we consider how financial markets reacted. We focus on interest rates and the exchange rate.
Since models suggest that Abenomics will work by lowering the real interest rate,16 and to a
lesser extent by weakening the yen, these are relevant intermediate indicators of success. We
next examine whether actual inflation and output growth have responded.

16

This is the main channel emphasized in old and new Keynesian models (Romer, 2012; Woodford, 2003).

11

3.1 Financial markets Figure 2(a) shows nominal interest rates in Japan since 2007. Twoyear government bond yields were near zero before Abe took office, and they have changed
little since. By contrast, both ten-year and thirty-year government bond yields have fallen. As
of February 19th , 2014, the yield on ten-year Japanese government bonds is 0.61 percent, 25
basis points below the average yield in 2012. Thus any expectations of inflation induced by the
Bank of Japan have not led to higher nominal interest rates, even over quite long horizons.
Given small changes in nominal interest rates, expected inflation has been the primary
determinant of movements in real interest rates. Unfortunately, there is no ideal measure of
Japanese expected inflation. The market for inflation-linked Japanese government bonds is
too thin for these prices to be reliable (Mandel and Barnes, 2013). As a market measure of
Japanese inflation expectations, we instead use inflation swap rates. These are also not ideal,
since they are illiquid, and they incorporate potentially time-varying risk premia (Mandel and
Barnes, 2013). But we are reassured by the similarity of this measure of inflation expectations
with that derived from surveys of professional forecasters. In October 2012, the Consensus
Economics survey of professional forecasters showed average annual expected inflation in 2013
and 2014 of 0.65 percent. In the same month, the yield on two-year inflation swaps averaged
0.8 percent. A year later, in October 2013, two-year inflation expectations from Consensus
Economics had risen to 1.8 percent, while the yield on two-year inflation swaps had risen to 1.7
percent.17
An alternative approach to measuring Japanese inflation expectations, suggested by Krugman (2013) (following Mandel and Barnes, 2013), is to use uncovered real interest rate parity
and the purchasing power parity (PPP) condition to estimate Japanese inflation expectations
using U.S. inflation-linked bonds (TIPS).18 This calculation requires taking a stand on when the
17

At longer horizons, there is a larger difference between the level of inflation expectations from Consensus
Economics and that from inflation swaps. But the change has been similar.
18
Uncovered real interest rate parity implies that rjapan = rus + %e, where r is the real interest rate, e is
the real exchange rate, the price of U.S. goods in terms of Japanese goods, and %e is the depreciation required
for purchasing power parity to hold. Expressing the real interest rate as the difference between the nominal
e
e
interest rate (i) and expected inflation, this implies that japan
= ijapan ius + us
%e. Following Krugman
(2013), we assume that the real exchange rate will take ten years to get back to its January 2010 value. We

calculate the real exchange rate as EP

where E is the yen-dollar exchange rate, P is the seasonally adjusted
P
U.S. CPI and P is the seasonally adjusted Japanese CPI.

12

Expected inflation (risk-neutral, %)

0
2

2.5
2
Nominal bond yields (%)
1
1.5
.5
0

2009m1

-2

2-year
10-year
30-year

2007m1

2011m1

2013m1

2007m1

2009m1

2011m1

2013m1

2011m1

2013m1

2007m1

Nominal exchange rate

Real exchange rate
2009m1

2011m1

2013m1

(d) Exchange rates

40

Index (1/2007=100)
60
80

100

(c) Real bond yields

Yen/$ exchange rate and real exchange rate (1/2007 Yen/$)

80
100
120
140

2
Real bond yields (%)
0
1
-1

2-year
10-year

2007m1

2009m1

(b) Market inflation forecasts

(a) Nominal bond yields

-2

2-year inflation swap rate

10-year inflation swap rate
10-year UIP-PPP (Krugman, 2013)

2007m1

Nikkei
Topix
2009m1

2011m1

2013m1

(e) The stock market

Figure 2 Abenomics financial market effects. Abenomics begins at the red line, November 2012.
Real bond yields are calculated as the difference between nominal bond yields and inflation swap rates.
Source: See data appendix.

13

yen-dollar real exchange rate was consistent with purchasing power parity. We follow Krugman
(2013) in assuming the equilibrium exchange rate was that in January 2010. Uncertainty about
this leads to uncertainty about the level of expected inflation. But we are primarily interested
in the change after Abenomics began.
Figure 2(b) shows the behavior of inflation swaps and this alternative measure of inflation
expectations. All measures of inflation expectations were rising before Abenomics and rose
rapidly in the first half of 2013. Between October 2012 and April 2013, 10-year inflation
swap rates rose from 0.3 percent to 1.1 percent. The interest rate parity measure of inflation
expectations rose somewhat more. In October 2012 it was 0.8 percent; in January 2014, it
was 1.9 percent. Since, as we describe below, long-run inflation expectations of professional
forecasters are well below two percent, the interest rate parity measure1.9 percentmay
overestimate the level of expected inflation. But we see it as confirming the basic insight
from our other measures: that Abenomics raised long-term inflation expectations in Japan by
roughly a percentage point.
The combination of steady or falling nominal interest rates with increasing inflation expectations has meant a precipitous decline in real interest rates (figure 2(c)). We measure the real
interest rate as the difference between the nominal government bond yield and the inflation
swap yield. Between October 2012 and February 2014, the two-year real interest rate fell 1.7
percentage points, and the ten-year real interest rate fell 0.9 percentage points. The larger
decline in the two-year real interest rate was in part driven by the upcoming consumption tax
increases, which increased expected inflation.
The Bank of Japans higher inflation target is likely to affect the economy primarily through
its effect on real interest rates. But an important, secondary channel is the exchange rate.
Figure 2(d) shows the behavior of the yen-dollar exchange rate since 2007. After averaging
79 yen per dollar in the first ten months of 2012, the yen weakened rapidly, reaching 103 yen
per dollar in May. Prices changed little over this short period in Japan or its major trading
partners, so the Bank for International Settlements broad effective real exchange rate index
moved essentially one-for-one with the nominal rate (figure 2(d)).
14

Abenomics also had dramatic effects on the stock market. Figure 2(e) shows the Nikkei 225
and the broader Topix index. Between October 2012 and May 2013, the Nikkei rose 65 percent
and the Topix 63 percent. Higher asset prices are one channel through which Abenomics may
help the Japanese economy. Most obviously, higher stock prices lead to more consumption and
investment. The stock price increase has also been taken by some to be an implicit forecast of
large long-run effects of Abenomics. We consider this argument in section 4.2.
We would like to know how much these changes reflect the effects of Abenomics and monetary policy in particular. One source of evidence is announcement effects: if the Bank of
Japans new-found resolve to raise inflation drove the 2013 movements of interest rates, exchange rates, and the stock market, we should see large immediate reactions when the Bank of
Japan announced its new policies.
Table 2 shows the change in several financial market indicators in a 24-hour window around
dates of news about the Bank of Japans new monetary policy.19 On the first date, November
15th , 2012, then-candidate Abe argued that the Bank of Japan should conduct unlimited
easing. 20 There was little change in nominal bond yields or inflation swap rates, but the
yen weakened by more than one percent and stock prices rose two percent. On the second
date, January 22nd , 2013, the Bank of Japan committed itself to a two percent inflation target
(Bank of Japan, 2013a). Since Abe had already called on the Bank of Japan to raise its inflation
target, this announcement had little new information. And despite Abes support of the Bank
of Japans new target, there was a widespread view that it lacked credibility. The Wall Street
Journal quoted the economist Joseph Gagnon saying: Its meaningless . . . The Bank of Japan
is very good at telling politicians its going to do a lot, and then doing nothing. 21 Consistent
with this interpretation, the yen strengthened and the Topix stock market index fell nearly half
a percent. Yields on government bonds and inflation swaps were essentially unchanged.
The third date is February 5th , 2013, when the Bank of Japan governor, Masaaki Shirakawa,

19

Note that though inflation swap yields are shown in the table, their general lack of liquidity makes any
non-responses difficult to interpret. Inflation swap yields often only change once every few days.
20
Financial Times.
21
Wall Street Journal.

15

Table 2 Announcement effects

Date

JGB
2-yr

11/15/12
1/22/13
2/5/13
4/4/13

-0.5
-0.1
-1.2
0.3

10-yr
bp change
-0.9
-0.8
-2.9
-11.4

Inflation swap
30-yr
2.8
-0.1
-2.2
-21.7

2-yr
10-yr
bp change
0.5
0.0
-0.7
3.2
5.3
15.9
0.0
0.0

Yen / $

TOPIX

% change
1.14
2.07
-1.00
-0.44
1.35
1.36
3.49
2.67

Notes: One-day changes in Japanese government bond (JGB) yields, inflation swaps, log nominal yen-dollar
exchange rate, and the Topix index following announcements or events on a given date. The dates are selected
as follows: on 11/15/12 then-candidate Shinzo Abe calls for unlimited easing; 1/22/13 is the Bank of Japans
inflation target announcement; 2/5/13 is the resignation announcement of the Bank of Japan governor Masaaki
Shirakawa (we use a 2-day change since it is unclear whether markets had already closed); 4/4/13 is the
announcement date of the Bank of Japans new quantitative and qualitative monetary easing policy.

announced that he would resign early. This was interpreted as evidence that the Bank of Japan
would soon be led by someone more sympathetic to bold monetary actions.22 Government bond
yields fell, inflation swap yields rose, the yen weakened, and stock prices rose.23
The Bank of Japan finally specified actions to reach its two percent inflation target on April
4th , 2013, the last date in table 2. These actions were dubbed quantitative and qualitative
monetary easing (QQME). They included a commitment to double the monetary base and the
Bank of Japans holdings of Japanese government bonds (JGBs) in two years. The statement
quite deliberately emphasized the newness of the policies; it referred to a new phase of monetary
easing in terms of quantity and quality (Bank of Japan, 2013b).
Markets were jubilant. Ten-year government bond yields fell 11 basis points, and thirtyyear yields fell 22 basis points. This was the largest one-day decline in the thirty-year yield
in almost a decade. Inflation swap yields were unchanged that day, presumably reflecting a
lack of trading volume. But in their next quote, both the two-year and ten-year inflation swap
yields rose 4 basis points. The April 4th announcement also had sizable effects on the yen and
the stock market. The 3.5 percent decline in the yens value against the dollar was the fourth
largest one-day depreciation since 1978. Thus a comparison of the January 22nd and April 4th
announcements strongly suggests that the mere announcement of an inflation target had much
22
23

Financial Times.
It is unclear whether this announcement came while markets were still open, so we use a two-day window.

16

smaller effects than the announcement of actions which could make the target credible.24 In
any case, the announcement effects suggest that monetary policy was an important driver of
financial market behavior in 2013.
3.2 Inflation We now turn to an examination of actual inflation in 2013. Figure 3(a) shows
three measures of Japanese inflation, the CPI, the CPI excluding fresh food and energy, and
the GDP deflator. The overall CPI has risen since March 2013. The twelve-month percent
change from December 2012 to December 2013 was 1.6 percent. This was the highest inflation
rate in Japan since the 2008 energy price shock. Year-over-year CPI inflation was 0.3 percent,
0.5 percentage points above the December 2012 Consensus Forecast. Thus Abenomics ended
deflation in 2013. But a concern is that much of this inflation may have been driven by a

Consensus Economics: 1 year-ahead CPI forecast

Consensus Economics: 10 year-ahead CPI forecast
BoJ Opinion Survey: 1 year-ahead CPI forecast

92

94

Price indices
96
98

Expected inflation (%)

4
6

100

102

10

weaker yen and hence may not continue.

2007m1

CPI (SA), 2/2007 = 100

CPI ex. fresh food and energy (SA)
GDP deflator (SA), 2007:Q1 = 100

2007m1
2008m1

2009m1

2010m1

2011m1

2012m1

2013m1

(a) CPI data

2014m1

2009m1

2011m1
Survey date

2013m1

(b) Survey inflation forecasts

Figure 3 Japanese actual and expected inflation. Abenomics begins at the vertical line, November
2012. Source: See data appendix.

Energy and fresh food prices are especially sensitive to the yens value. The so-called corecore CPI strips out both these categories. Figure 3(a) shows that this measure of the CPI has
also risen since March 2013, but by much less. In the twelve months from December 2012 to
December 2013, it rose 0.7 percent. A final measure of inflation, which has the advantage of
excluding all imports, is the GDP deflator. After years of steady decline, the GDP deflator was
24

See Kuttner and Posen (2001) for a discussion of the announcement effects of earlier Bank of Japan
announcements. For more discussion of the effects of Abenomics announcements, see Ueda (2013).

17

flat through most of 2013 (figure 3(a)). Notably, all these price measures have risen more in
2013 than they did between early 2005 to early 2007 when the yen depreciated by nearly 20
percent. While the comparison is imperfect because the 2013 depreciation was more rapid, it
does suggest that Abenomics effect on prices goes beyond pass-through.
Positive inflation is not yet firmly established, however. In January 2014, the last month
for which we have data, both the overall CPI and the CPI excluding food and energy fell. Consistent positive inflation likely requires increases in nominal wages. Here the news is somewhat
encouraging. Including bonuses, nominal hourly earnings rose 0.4 percent from December 2012
through December 2013.25 Many workers wages are set in annual spring negotiations between
unions and employers.26 As this paper was being finalized, these negotiations were ongoing.
The extent to which firms grant workers wage increases depends in large part on their inflation expectations. We consider two proxies for firm inflation expectations in Japan: (1) the
expectations of professional forecasters and (2) the expectations of households. Professional
forecasters are likely to provide a useful proxy for the inflation expectations of large multinational firms. Such firms have the resources either to employ forecasters themselves or to
seek outside professional opinions. By contrast, Coibion and Gorodnichenko (2013) argue that
household inflation expectations are a good proxy for smaller firms expectations.
Figure 3(b) shows one-year-ahead and ten-year ahead inflation expectations of professional
forecasters from Consensus Economics, as well as one-year-ahead household inflation expectations from the Bank of Japan opinion survey.27 A surprising feature of these data is the level of
inflation forecast by Japanese households. Respondents to the opinion survey generally forecast
more than four percent inflation over the next year. Households also often say that inflation
over the past year exceeded four percent, even though Japan has not experienced four percent
year-on-year inflation since 1981. For our purposes, however, what matters most is the change

25

Total cash earnings rose 0.5 percent while hours rose 0.1 percent.
See Olivei and Tenreyro (2010) for a discussion of how synchronized wage setting affects monetary policy
effectiveness in Japan.
27
The Bank of Japan opinion survey is a poll of approximately 4000 randomly-sampled individuals above
age 20. Individuals are sent a questionnaire in the mail, and the response rate, though variable, tends to be
around 55 percent. See Bank of Japan Opinion Survey.
26

18

in these expectations.
Both household and professional inflation expectations have risen since late 2012 (figure
3(b)). The one-year inflation forecast from professional forecasters rose particularly fast in
early 2013, but this likely reflects the April 2014 consumption tax hike. Respondents to the
Bank of Japan opinion survey were instructed to ignore the consumption tax increase in forming
their inflation expectations, but professional forecasters received no such instruction.28
While the change in inflation expectations is encouraging, the level of inflation forecast by
professional forecasters, like inflation swap rates, suggests that the Bank of Japans target is not
(yet) fully credible. As of October 2013, professional forecasters expected 1.4 percent annual
inflation over the next ten years. (Recall that ten-year inflation swap yields are roughly 1.1
percent, figure 2(b).) This implies that the output effects we discuss below are the effects of an
imperfectly credible monetary policy change.
3.3 Comparison to Quantitative Easing We have argued that Japans recent monetary
policy announcements and actions had large effects on financial markets and both actual and
expected inflation. These effects are strikingly different from those of the Bank of Japans 2001
to 2006 experiment with quantitative easing (QE). What is the Bank of Japan doing now that it
did not do in 2001? On March 19th , 2001 the Bank of Japan announced its quantitative easing
policy. The announcement had three key features (Ugai, 2007): (1) the Bank of Japan would no
longer target an interest rate, but instead would target the excess reserves of commercial banks
held at the Bank of Japanthe Banks so-called current account balance. (2) Quantitative
easing would be explicitly state-dependent. It would continue until deflation ended. (3) The
Bank of Japan would purchase long-term government bonds.
The effects of this policy are clearly visible in figure 4(a), which shows a rapid acceleration
in narrow money (M0) growth after March 2001. Despite this massive increase in liquidity
provision, broad money growth changed little (figure 4(b)). And in the year following March
28

To get a better sense of how Abenomics changed the inflation expectations of professional forecasters, one
can look at inflation expectations in October 2012 and October 2013. Since the consumption tax increase was
passed in June 2012, the difference between the earlier and later forecasts provides an estimate of the effect of
Abes new policies on inflation expectations. This measure of inflation expectations rose at all time horizons.

19

2001, professional forecasters inflation expectations fell and the real exchange rate depreciated
only slightly. The real effects of quantitative easing appear to ultimately have been small (Ugai,
2007). Krugman (1998, 2000) and Eggertsson and Woodford (2003) argue that this is exactly
what one would expect of a policy that merely temporarily increases monetary aggregates
without changing expectations about inflation or future nominal rates. For quantitative easing
to be effective it must either lower nominal interest rates or raise expected inflation or both. In
general, this means it must be expected to persist even after the economy exits the zero lower
bound. By contrast, if quantitative easing is expected to be temporary, then it will likely have
little or no effect on expected future real interest rates and the broad money supply; consumers

40

will simply substitute cash for deposits while banks hold excess reserves.

Broad money 12-month growth (SA), Abenomics (11/2012)

Broad money 12-month growth (SA), QE (3/2001)

12-month growth rate (%)

2
3

12-month growth rate (%)

10
20
30

Narrow money 12-month growth (SA), Abenomics (11/2012)

Narrow money 12-month growth (SA), QE (3/2001)

-24

-18

-12

-6
0
6
12
Months before / after policy change

18

24

30

-24

(a) Narrow money (M0)

-18

-12

-6
0
6
12
Months before / after policy change

18

24

30

(b) Broad money (M3)

Figure 4 Comparison of money growth during quantitative easing (QE) and Abenomics. Month 0
is March 2001 for quantitative easing and November 2012 for Abenomics. Source: See data appendix.

This appears to be what happened in the early 2000s. Quantitative easing was indeed
temporary. In early 2006 the Bank of Japan mopped up most excess reserves (see Blinder,
2010, figure 8); later that year it raised the uncollateralized call rate to 0.25 percent. In
absolute terms, the narrow money supply declined only slightly. But relative to its 1990s trend,
all of quantitative easing was reversed.
Figure 4 shows that Abenomics is different. Narrow money has grown much less during
Abenomics, but the broad money supply has grown much more. This is exactly what one
20

would expect from a more credible monetary policy change. If, unlike in 2001, people now
expect lower future real interest rates, the resulting increase in credit demand will also lead
to money creation in the banking system. Thus the increase in the broad money supply. The
12-month growth rate of broad money is now (11/2013) at its highest level since August 1999.
Taken together, figure 4 is a striking confirmation of model-based predictions. Exactly as
predicted by Krugman (1998) and Eggertsson and Woodford (2003), a credible commitment to
future expansion is having effects that temporary changes in the narrow money supply did not.

3.4 Output Ending deflation and increasing inflation expectations are intermediate goals.
The ultimate target of Japans new monetary policy is higher output. Here we document the
turnaround of Japanese output in 2013. We then turn to the harder task of determining what
part of recent Japanese growth is due to Abenomics and to monetary policy in particular.
Figure 5(a) shows quarterly GDP growth at an annual rate in Japan since 2007. Performance
in 2013 was notably better than that at the end of 2012. Whereas quarter-on-quarter growth
was negative in the second and third quarters of 2012, output grew during each quarter of
2013, and particularly so in the first half of the year. Fourth quarter over fourth quarter, the
Japanese economy grew 2.6 percent in 2013; year-on-year it grew 1.5 percent.
However, the trend is negative: since the first quarter of 2013, growth has declined each
quarter. The principal problem was net exports. Figure 5(b) shows that the contribution of
exports to real GDP growth fell from 2.4 percentage points (at an annual rate) in the first
quarter of 2013 to 0.2 percentage points in the final quarter of the year. At the same time,
the contribution of imports fell from -0.7 percentage points to -2.4 percentage points. Put
differently, had the contribution of net exports been the same in the fourth quarter as it was in
the first, annualized fourth quarter growth would have been 4.6 percent rather than 0.7 percent.
We will return to the puzzling behavior of net exports below.
Figure 5(c) provides a different way of seeing recent performance. It plots average annual
GDP growth as well as the contributions of major components in 2013. For comparison, we also
show average growth rates for two subsamples: the lost decades excluding the 2008 recession
21

(1995-200729 ) and the 2008 recession and recovery (2008-2012). 2013 growth has been strong

-15

Real GDP growth (%, annualized)

-10
-5
0
5

10

Contributions to quarterly GDP growth (p.p., annualized)

-2
-1
0
1
2
3

relative to these prior periods.

2007q3

2009q1

2010q3
Date

2012q1

2013q3

Consumption
Gov. spending
Imports(-)

Investment
Exports

Quarter

(a) Quarterly growth

(b) Contributions to growth in 2013

-1

Avg. year-on-year growth rate (p.p.)

0
1
2

1995-2007
2008-2012
Abenomics (2013)

GDP

Consumption

Investment
Exports
Gov. spending

Imports(-)

Contribution to GDP growth

GDP

Per working age

(c) Contributions to growth

Figure 5 Panel (a) shows annualized quarter-on-quarter GDP growth since 2007. Quarters since
Abenomics began are marked in red. Panel (b) shows contributions to GDP growth by component for
each quarter during 2013. Panel (c) provides a comparison of contributions to GDP by component
during Abenomics (2012-2013), the lost decade excluding the Great Recession (1994-2007), and the
Great Recession (2007-2012). Contributions are calculated as in Japans national accounts. Panel (c)
also displays working-age adjusted GDP growth. Source: see data appendix.

As we noted in section 2, when making historical (and international) comparisons it is

important to adjust for Japans changing demographics. The right-most bars in figure 5(c)
show growth rates of output per person age 15 to 64. Since the 15 to 64 year old population
29

We start this comparison in 1995, since official national accounts data for the level of GDP become available
in 1994.

22

has been shrinking since 1995, this raises all growth rates. Population decline has, however,
been particularly rapid since 2006, so this adjustment makes performance under Abenomics
look more impressive: 2013 growth per working-age person was 3.1 percent, compared to 1.6
percent from 1995 to 2007 and 0.4 percent from 2008 to 2012.
Output growth in 2013 is almost entirely accounted for by consumption, which contributed
1.2 percentage points, and government spending, which contributed 0.9 percentage points.
Residential investment added another 0.3 percentage points. Non-residential investment, the
change in inventories, and net exports subtracted a total of 0.8 percentage points from growth.
3.5 Forecasts Without a counterfactual, it is difficult to say how much of 2013 growth was
due to Abenomics. We need to know what output would have been without Abenomics to
know the effect of Abenomics. As an estimate of this counterfactual, we first look to forecasts
from Consensus Economics made in late 2012, before Abes policies were fully known.
What complicates our analysis is that news of the 2012 recession arrived almost simultaneously with Abes first policy speeches. The first news of actual contraction came on November
12th , 2012, when preliminary GDP data showed that the economy contracted in the third quarter of 2012 at a 3.5 percent annual rate.30 This news came just days before Abes unlimited

2013 Forecast
2014 Forecast
2012 Forecast

.5

.5

GDP Growth
1.5

Consumption Growth
1
1.5

2013 Forecast
2014 Forecast
2012 Forecast

2.5

2.5

easing speech on November 15th .

2012m1

2012m7

2013m1
Survey Date

2013m7

2014m1

(a) GDP growth

2012m1

2012m7

2013m1
Survey Date

2013m7

2014m1

(b) Consumption growth

Figure 6 2012, 2013 and 2014 forecasts by survey date for GDP growth (panel a), and consumption
growth (panel b). The vertical line is the first announcement of Abenomics, November 2012.
30

See http://www.esri.cao.go.jp/jp/sna/data/data_list/sokuhou/files/2012/qe123/pdf/jikei_1.pdf.

23

Forecasters interpreted bad news about growth in 2012 to also be bad news about growth
in 2013. The GDP growth forecast for 2013 was revised down from 1.3 percent in October 2012
to 0.8 percent in November 2012. This is shown in figure 6, where we plot the evolution of
Consensus forecasts for GDP and consumption by survey date. Even in November the depth
of the 2012 recession was not fully appreciated. On November 28th , data for October retail
sales showed a larger than predicted 1.2 percent decline.31 While the 2012 growth forecast in
figure 6(a) did not decline in December, this reflects single-digit rounding error in the mean
forecast. The unrounded 2012 mean growth forecast fell from 1.84 percent in November to 1.80
in December. Forecasters read this as a negative signal for future growth: the 2013 growth
forecast deteriorated from 0.8 percent in November 2012 to 0.6 percent in December. Given
that so much was already known about actual 2012 GDP at this time, it is not surprising that
2013 forecasts moved more than 2012 forecasts.
Thus it is not obvious what forecasts provide the best counterfactual. Forecasts made
in early fall 2012, before Abe announced his polices, suffer from mistaken optimism about
growth in 2013. This means these forecasts are an over optimistic counterfactual. By contrast,
later forecasts, in December 2012, more accurately reflect the 2012 recessions effect on the
2013 growth outlook but were made when then-candidate Abe was widely expected to win
the election and after some of Abes initial economic policy announcements. In particular,
by December 2012, Abe had forcefully called for more expansionary monetary policy. To the
extent that these announcements had a positive effect on the December Consensus forecasts,
we will underestimate the effect of Abenomics.
Evidence that the December forecasts include some positive announcement effects comes
from out-of-sample statistical forecasts. In table 3 we tabulate 2013 (year-on-year) growth
forecasts from autoregressive and vector-autoregressive (VAR) models estimated with data
through the fourth quarter of 2012. Columns 2 and 3 report the forecasts from an autoregressive
model of quarterly GDP growth. These two columns differ in that column 2 estimates the
model based on final data releases, whereas column 3 estimates the model based on professional
31

Bloomberg

24

forecasters December 2012 view of growth in 2011 and 2012. Columns 4 and 5 conduct the same
exercise using a VAR with quarterly GDP growth, CPI inflation, and changes in the ten-year
bond yield. We estimate all models with one to four lags (AIC and BIC criteria consistently
favor one lag) starting in 1996 when the zero lower bound began to bind.
The consistent result is that 2013 growth is forecast to be lower than thought by professional forecasters in December 2012. This is true even when one uses the data available to
forecasters at the time (columns 3 and 5). This suggests that professional forecasts include
positive expectations about the effects of Abenomics. While admittedly the standard errors for
the VAR forecasts are large (as they are for the Consensus forecasts), we are encouraged that
the forecasts do much better in 2012 than in 2013. Estimating the models through 2011 and
forecasting 2012 growth yields estimates ranging from 1.3 percent to 1.9 percent; actual GDP
growth was 1.45 percent.32

Table 3 2013 Growth forecasts from autoregressive models

Lags (=p)

1
2
3
4

AR(p)
Final release
data
-0.11
0.01
0.05
0.04

VAR(p)

Forecast data
-0.17
-0.06
-0.04
-0.23

Final release
data
-0.10
0.39
0.59
0.47

Forecast data
-0.09
0.34
0.46
0.17

Notes: 2013 growth forecasts in percent based on autoregressive and vector-autoregressive models. Column
1 displays the number of autoregressive lags. Columns 2 and 3 display forecasts from autoregressive models of GDP growth estimated from 1996Q1 to 2012Q4. Column 2 uses actual data, whereas column 3 uses
professional forecasts made in December 2012 for 2011 and 2012. Columns 4 and 5 display forecasts from
vector-autoregressive models of GDP growth, CPI inflation, and changes in the 10-year bond yield estimated
from 1996Q1 to 2012Q4. Column 4 uses actual data, whereas column 5 uses forecast data for 2011 and 2012.

The VAR and forecast-based counterfactuals imply that Abenomics raised 2013 growth by
roughly a percentage point. Actual year-over-year growth was 1.5 percent compared to the 0.6
percent forecast by Consensus Economics in December 2012. This implies a 0.9 percentage point
effect of Abenomics on growth. As we argued above, we believe this estimate is conservative.
32

The 2011 out-of-sample forecasts are much higher than the data, although this likely reflects the effect of
the Great East Japan Earthquake and tsunami.

25

For instance, our most pessimistic VAR forecast suggests that 2013 growth would have been -0.2
percent. That would imply a much larger 1.7 percentage point contribution to 2013 growth from
Abenomics. Further confirming that Abenomics added at least a percentage point to growth is
an estimate from the IMF. In October 2013, the IMF argued that the entire package of policies
would add 1.3 percentage points to 2013 growth (International Monetary Fund (2013c), p. 49).
The IMF did not describe how it obtained this estimate, and it may now (in early 2014) be out
of date. Still, we are encouraged that it is close to our figures.
Figure 6 shows that the primary improvement in output relative to the December 2012
forecast was in consumption. The December 2012 Consensus forecast was for 0.3 percent
consumption growth; in factas indicated by the rapid improvement in the forecast over 2013
consumption grew 1.9 percent. This difference alone accounts for a percentage point of output
growth. While not provided by Consensus Economics, a comparison to the OECD Economic
Outlook forecast suggests that government consumption growth exceeded expectations due
to Abes fiscal stimulus. At the same time, business investment (not shown) underperformed
relative to expectations. The December 2012 forecast was for 0.4 percent growth; actual growth
was -1.6 percent. This underperformance is mysterious since one would expect stimulative
monetary policy to have large effects on business investment. In our view, the most likely
explanation is that forecasters were simply over optimistic in December 201233 and/or already
incorporated positive announcement effects from Abenomics.
3.6 Which Arrow? While a conservative read of the evidence suggests that Abenomics increased 2013 output by one percent, it is less clear whether this should be ascribed to the
monetary, fiscal, or structural arrow. In our view there are at least two compelling reasons
to believe that the third arrow has not yet been a major contributor. First, there has been
little if any structural reform passed, let alone implemented. Thus the third arrow would have
to work through anticipation effects, but it is not clear what consumers and firms anticipate.
Second, the improvement in growth and growth forecasts has coincided with a rise in prices
33

For instance, if low 2012 growth reduced firm cash flows and these are needed to finance investment
because of credit frictions (Fazzari, Hubbard, and Petersen, 1988), then continuous downward revisions in the
2012 output forecast could trigger downward revisions in the 2013 investment forecast.

26

and inflation expectations (figure 3). If anticipation of structural reforms were the dominant
driver of real effects, we should observe falling prices and inflation forecasts.
We therefore focus on disentangling the first and second arrows: how much of Abenomics
boost to 2013 growth was due to monetary policy and how much to fiscal policy? This question
is as challenging as it is important. Fiscal policy was first discussed in December 2012 and was
formally announced in January 2013,34 the same month that the Bank of Japan announced its
two percent inflation target. Thus it is difficult to use timing to separate the effects of the two
policies. Instead, we focus on movements in different components of GDP.
To have any hope of disentangling the effects of these two policies, we need to say something
about the size of Abes fiscal stimulus. We compare the current IMF estimate of the cyclicallyadjusted primary budget balance to the IMF forecast before Abes fiscal stimulus was revealed.
In October 2012, the IMF forecast that Japans cyclically adjusted primary budget deficit as a
share of potential GDP would be 7.5 percent (International Monetary Fund, 2012). In October
2013, the IMF estimated it would be 8.5 percent (International Monetary Fund, 2013b). This
implies an unexpected fiscal stimulus equal to 1.0 percent of GDP in 2013.35
A case can be made, therefore, that there is little growth for monetary policy to explain in
2013. Suppose that Abenomics as a whole added one percentage point to 2013 output growth.
If there was one percent of GDP worth of fiscal stimulus, a multiplier of one would explain all
of this excess growth. Nonetheless, the behavior of consumption in 2013 strongly suggests that
monetary policy also contributed to growth.
Recall that 2013 consumption growth exceeded forecasts by 1.6 percentage points, enough to
contribute an entire percentage point to output growth. Consumptions strength is difficult to
explain by fiscal policy. Unlike the U.S. stimulus in 2009 and 2010 (the American Recovery and
Reinvestment Act), little, if any, of Abes stimulus consisted of transfer payments or lower taxes
for households. Forty-eight percent of Abes supplementary budget was public investment, 32

34

See SCMP and Bloomberg.

See also, International Monetary Fund (2013a), p. 1. To be precise, this is 1.0 percent of potential GDP.
But since the IMF believes that actual Japanese GDP is close to potential, the difference is small.
35

27

percent was private investment subsidies, and 14 percent was transfers to local governments.36
The remaining seven percent was medical and educational spending. So there was no direct
mechanism through which a larger budget deficit translated into higher consumption.
Evidence that the second arrow is not solely responsible for strong 2013 consumption growth
comes from the time-series pattern of consumption and government spending in figure 5(b).
Across quarters in 2013, there is a slight negative correlation between consumption growth
and government spending growth. If government spending were raising private consumption
through an old Keynesian multiplier effect, one would expect to see the opposite pattern.37
In addition to evidence against fiscal stimulus explaining consumption growth, there is evidence for monetary policy explaining this growth. Consider the spending patterns documented
in Japans Family Income and Expenditure Survey. Among so-called worker households,
households with a member employed, real disposable income fell 1.3 percent between 2012 and
2013. This fits with the more rapid increase in prices than in wages documented in section 3.2
and with the lack of any substantial new government transfer payments. At the same time as
incomes fell, real consumption expenditures rose 0.9 percent, exactly as one would expect if a
lower real interest rate was inducing households to spend sooner rather than later. Furthermore,
from 2012 to 2013, the value of new loans taken rose nine percent.
The composition of consumer spending also suggests a role for monetary policy. Among
worker households, the largest positive contribution to consumer spending was in the transportation category, especially spending on private vehicles. A similar pattern is visible in the
aggregate national accounts data. Consumption spending on durable goods rose 15.7 percent
from the fourth quarter of 2012 to the fourth quarter of 2013, while spending on semidurables
rose 3.9 percent and spending on services rose 1.9 percent. Spending on nondurables fell 0.5
percent.
Along with consumer durables, residential investment is likely to be one of the components

36

These figures are based on Abes supplementary budget released in January 2013. See budget.
Government spending could raise consumption through intertemporal substitution by raising marginal
production costs and inflation expectations and thus lowering real interest rates. However, evidence from
Wieland (2014) and Dupor and Li (2013) suggests that this mechanism is unlikely to be quantitatively important.
37

28

of GDP most responsive to the real interest rate (Bernanke and Gertler, 1995). Here too, one
sees large effects. Fourth quarter over fourth quarter, real residential investment grew 10.4
percent. This is particularly remarkable given that Japans declining population presumably
reduces demand for new housing.
Some of the increase in durables purchases and private residential investment may be driven
by the upcoming consumption tax increase. In April 2014, Japans consumption tax will rise
from five percent to eight percent. The tax increase is widely thought to be leading consumers
to pull forward purchases. The tax increase applies to houses as well consumption goods, so its
effects are also likely to show up in residential investment. Furthermore, a similar consumption
tax increase occurred in April 1997, and it appears to have pulled forward durables purchases.38
We doubt, however, that the upcoming tax increase is a major contributor to the increase
in household spending. Most obviously, it has been expected since 2012, yet forecasters in
December 2012 did not expect significant consumption growth in 2013. Therefore it seems
implausible that the consumption tax increase can explain why year-over-year consumption
growth exceeded December 2012 forecasts by 1.6 percentage points. Furthermore, consumption
growth was particularly strong in the first two quarters of 2013 (figure 5(b)), but intertemporal
substitution in advance of the tax increase ought to have had larger effects later in the year.
In our view, it is more plausible that the consumption tax increase explains some part of
the residential investment boom. But even here, there is a strong hint that monetary policy
mattered: from December 2012 to June 2013, the percent of borrowers choosing fixed rather
than variable rate mortgages increased by almost 10 percentage points.39
Overall we believe there are plausible reasons to attribute much of the increase in consumption to monetary policy. If we simply assign all of the consumption increase relative to the
forecast counterfactual to the monetary arrow, then Japans new monetary policy raised 2013
output growth by roughly a percentage point. This might be an overestimate, insofar as some
excess consumption growth came from other causes and was spent on imports. But it might
38

Fourth quarter over fourth quarter, in 1996 real GDP rose 3.4 percent, consumer durables 12.0 percent,
and residential investment 17.4 percent.
39
See Wall Street Journal.

29

also be an underestimate, insofar as it ignores the likely small but positive effects of monetary
policy on residential investment and net exports.
Take one percentage point as our estimate of the contribution of monetary policy to 2013
growth. Together with fiscal stimulus of one percent of GDP and a multiplier of one, this would
imply a total Abenomics effect of two percentage points.40 That is much larger than the 0.9
percentage point difference between actual growth and the Consensus Economics forecast in
December 2012. But, as we noted above, this is a conservative counterfactual. If instead we
use the most pessimistic forecast from the autoregressive models in table 3, -0.2 percent, then
Abenomics raised GDP growth by 1.7 percentage points, close to the above calculation. In
short, while one can make a case that monetary policy added little to 2013 growth, one can
also argue that it added roughly a percentage point to GDP. This large confidence interval
reflects the difficulty of disentangling simultaneous monetary and fiscal changes. Nevertheless,
we find compelling evidence in favor of some positive effects of monetary policy.
3.7 Discussion The above analysis implies that the effects of Japans new monetary policy
in 2013 were modest relative to the output gap. An important reason why the effects were
not larger is the behavior of net exports. Despite the 20 percent depreciation of the yen in
spring 2013, real net exports deteriorated over the year (figure 5(b)). Of course, that nominal
net exports fell is unsurprising. Whether a currency depreciation raises nominal net exports
depends on the Marshall-Lerner condition. Assuming that net exports are initially zero, the
sum of import and export elasticities must exceed one in order for depreciation to raise nominal
net exports. This condition is unlikely to be satisfied immediately, but is generally assumed
to hold in the long-run. Initially there is little response of quantities, so price effects dominate
and the trade balance deteriorates. But long-run elasticities are larger, so eventually quantities
adjust and nominal net exports rise above their pre-depreciation value. This is the so-called
J-curve (Magee, 1973; Bahmani-Oskooee and Ratha, 2004). Japan is still on the downward
slope. Nominal net exports fell from minus 2.4 percent of GDP in the first quarter of 2013 to
40

Auerbach and Gorodnichenko (2014) provide estimates of the government spending multiplier in Japan.
Their estimates suggest large uncertainty about its current value.

30

minus 3.7 percent of GDP in the fourth quarter of 2013.

In the real national accounts data, separate price deflators are used for each component of
real GDP, including imports. So the price effects that have driven down Japans nominal trade
balance are taken out. Yen depreciation cannot explain why real net exports have subtracted
from Japanese real GDP growth in 2013. Quite the opposite. As long as the elasticities of export
and import volumes with respect to the real exchange rate are positive, a real depreciation
should lead to some increase in export volumes and some decrease in import volumes.
In fact, year-over-year in 2013, export volumes (real exports) rose only 1.6 percent while
import volumes rose 3.4 percent. Part of the explanation is that Japanese trade quantities appear to be quite inelastic with respect to yen movements. Japanese exports are overwhelmingly
manufactures. In 2013, 73 percent of goods exports were in the categories of manufactured
goods, machinery, electrical machinery, and transport equipment (e.g. motor vehicles).
These are sectors in which so-called pricing to market is likely. Exporters of manufactures are
likely to set prices in foreign currency. These prices will be set for a variety of competitive
reasons (e.g. to maintain market share) and will thus be relatively unresponsive to short-run
changes in the yens value.41 An older literature suggests that such pricing to market is more
prevalent among Japanese exporters than it is among other countries exporters (Dominguez,
1999; Gagnon and Knetter, 1995).
Consider one of Japans largest exports: automobiles. Gagnon and Knetter (1995) (table
6) find that between 1978 and 1983, when the real value of the yen fell 39 percent against the
dollar, the real retail price of a Honda Civic fell only 7 percent. Casual observation suggests
a similar phenomena today. Japanese cars sold in the U.S. are not 25 percent cheaper now
than they were in 2012. Such pricing to market can explain why Japanese exports have not
risen more. The number of Japanese cars exported, for instance, rose only 1.3 percent between
December 2012 and December 2013.
Short-run elasticities may if anything be smaller for imports. Hooper, Johnson, and Marquez
(2000) estimate that the short-run elasticity of Japanese imports with respect to the exchange
41

There is a large literature on pricing to market. See Krugman (1987) and Atkeson and Burstein (2008).

31

rate is only -0.1. This estimate is based on historical data, but it is unlikely that recent
changes have improved matters. In 2013, fossil fuels accounted for 34 percent of Japans
imports, equivalent to 5.7 percent of Japanese GDP. By contrast, in 2004, energy imports
were only two percent of GDP. The change has been driven by the rise in the world price
of oil and the 2011 tsunami. Prior to the tsunami, Japan produced roughly a quarter of its
electricity from nuclear power. All nuclear reactors are now shut down, with fossil fuel imports
substituted. Unfortunately for the trade balance, demand for importsincluding fossil fuels
is quite inelastic. From December 2012 to December 2013, for instance, the quantity of oil
imported rose 0.2 percent despite a 19 percent increase in the yen price of oil.
Aside from net exports, the effects of a weaker yen on real GDP are ambiguous. On the
one hand, as documented in section 3.2, a weaker yen is a major reason why deflation ended in
2013. By raising actual inflation, yen weakness likely contributed to higher expected inflation
and thus lower real interest rates. On the other hand, a weaker currency has direct negative
effects on aggregate demand. More resources spent on imported goods mean fewer resources
spent on domestic goods.
Thus, at least in Japan, currency depreciations appear to be a less effective short-run tool
to jump-start an economy at the zero lower bound than originally anticipated.42 If the yens
current level proves persistent, however, it is likely that firms and consumers will change their
behavior in ways that significantly lower imports and raise exports. Thus as long as yen
weakness persists, it is reasonable to expect net exports to contribute to medium-run growth.

4 Medium-Run to Long-Run Outlook

The previous section discussed the behavior of Japanese output and inflation in 2013. We
argued that Abenomics as a whole likely added between 0.9 and 1.7 percentage points to 2013
growth, with strong, though circumstantial evidence, that monetary policy explains part of
this gain. But proponents of Abenomics hope for more than one good year. Presumably the
ultimate goal is to close Japans output gap.
42

For instance, McCallum (2000) and Svensson (2003) argue for currency depreciations to jump-start an
economy at the zero lower bound, along with other measures such as price level targeting.

32

There are two distinct questions here. First, does Abenomics, and the monetary policy
regime change in particular, pass a cost-benefit test? Do expected benefits, even if small,
exceed costs? To that our answer is a likely yes. A different question is whether Japans
monetary policy is likely to have large output effects, perhaps large enough to close the large
(and, as we shall see, possibly growing) output gap of 4.5 to 10 percent that we found in section
2.2. To that, our answer is that it is too soon to tell, but that there is of yet little evidence of
effects this large.
We look at four sources of evidence on the medium to long-run effects of Abenomics: professional forecasts, the stock market, a comparison with Franklin Roosevelts regime change,
and Keynesian models. Unfortunately, each of these sources of evidence has problems, and
they do not paint a consistent picture. In particular, current data and professional forecasts
are inconsistent with the larger effects suggested by the analogy to the U.S. in 1933 and new
Keynesian models.
It is worth emphasizing again that we take Abenomics effect on inflation expectations as
given. Inflation expectations have not (yet) risen to two percent (section 3). As we discuss
further in section 5, if Japans new inflation target becomes more credible, its positive effects
will likely be larger.

4.1 Forecasts We first consider an obvious source: professional forecasts. In particular, we

examine long-run forecasts from Consensus Economics. Figure 7(a) shows that between October
2012 and October 2013, professional forecasters not only raised their level forecasts for GDP,
they also raised their long-run growth forecast from 0.9 percent to 1.1 percent. Output in 2022
is predicted to be 3.1 percent above the pre-Abenomics forecast. Importantly, in October 2012,
Japan had already legislated the 2014 and 2015 consumption tax increases, so the forecast
change from October 2012 to October 2013 measures only the effects of Abes new policies.
Whether this forecasts revision is good or bad news depends on ones perspective. It suggests
a greater than three percent output gain in perpetuity. That is a large gain for any economic
policy, particularly one with few obvious costs. On the other hand, these forecasts suggest little
33

120

120

115

115

Consumption
110

110

105

GDP
105

2005

2010

2015

2020

2025

Data (2007=100)
Forecast 2012m10
Forecast 2013m10
Forecast 2007m10

100

100
95

Data (2007=100)
Forecast 2012m10
Forecast 2013m10
Forecast 2007m10

2005

(a) GDP

2010

2015

2020

2025

(b) Consumption

Figure 7 Actual and forecast output and consumption. Forecasts are from Consensus Economics.
Abenomics begins at vertical red line (2012).

prospect of closing Japans output gap. As a benchmark, we plot forecasts made in late 2007.
In section 2.2 we argued that this provides a reasonable measure of the output gap. With this
metric, output was 10 percent below potential in 2012. As of late 2012, this gap was projected
to grow to 13 percent in 2017. While Abenomics is projected to raise long-run growth, current
forecasts still show this gap widening to 11 percent in 2017. Put differently, forecasters do
not expect Abenomics to close the gap with their 2007 view of Japanese economic prospects.
If one had hoped that Japans monetary regime change would eliminate Japans demand side
problems, this is disappointing news.43
The forecasts for consumption, non-residential investment (not shown), and industrial production (not shown) have similar implications. A one percent level gain is projected for consumption, but that is not large enough to close the gap with the late 2007 projection. Consumption is forecast to grow less than output, presumably because forecasters expect a weaker
yen to raise real net exports. Business investment in 2012 is more than 20 percent below the
2007 projection, and, while Abenomics has led to upward forecast revisions, the change is too
small to close this gap.
The forecast revisions we document here appear to be consistent with those of the general
43

Similar implications obtain from the Okuns law calculation of the output gap discussed in section 2.2. As
of October 2013, the IMF forecasts that Japanese unemployment will be above 4.2 percent through 2018 (the
end of the forecast horizon), implying a persistent 10 percent output gap (International Monetary Fund, 2013c).

34

public. We examine the output expectations of individuals in the Bank of Japan opinion
survey and firms in the Tankan business survey.44 Respondents to the Bank of Japan survey
are pessimistic both before and after Abenomics. Pre-Abenomics, in September 2012, only
5.3 percent of respondents expected their income to rise over the next year, and 45.8 percent
expected it to fall. In December 2013, 8.1 of respondents expected their income to rise, and
37.8 percent expected it to fall. So there is improvement. But since 80.9 percent of respondents
in December 2013 expect prices to go up (either slightly or substantially), at least 72.8 percent
of the polled population expect their real income to fall. By comparison, only 47 percent of
respondents to the August 2013 U.S. Michigan consumer survey expect their real income to fall
over the next year.
In the December 2013 Tankan survey, we examine whether firms describe business conditions
as favorable or unfavorable, and whether firms say there is excess demand or excess supply for
products and services. Net favorability ratings are now positive for the first time since the
financial crisis. They are at their early-2007 level. But unsurprisingly, the current figures are
some distance away from the very positive levels of the late 1980s. Responses to the excess
supply question show that in 2013, instances of excess supply fell and those of excess demand
rose. On net, excess supply still exceeds excess demand, but on this question responses have
improved over their early 2007 levels.
These three sources of forecasts suggest cautious optimism. But at least so far, there are not
signals of a dramatic break with the growth rates of the 1990s and 2000s. Of course, a limitation
of this evidence is that it is difficult to know what part of Abenomics is driving forecast revisions.
Upward revisions could reflect forecasters belief that growth-enhancing structural reforms will
take place. Or perhaps forecasters think that monetary policy will increase growth, even in the
long-run. The fact that long-run inflation forecasts have risen certainly points to an important
role for monetary policy, since structural reforms would presumably lower expected future
prices. But we cannot rule out the possibility that the real side is (in the long-run) primarily

44

The Tankan survey polls a sample of 10,000 private enterprises with more than U20 million in capital for
quantitative and qualitative data.

35

determined by structural reforms while the nominal side is primarily determined by monetary
policy.
Even as an indicator of the effects of Abenomics as a whole, these forecasts are not definitive.
Economic forecasters have no crystal ball, and their past experience could be a misleading guide
to the effects of a large-scale policy change. Previous work has shown that forecasters may be
slow in updating to news (Coibion and Gorodnichenko, 2012). In appendix tables 5 and 6 we
show that short- and long-run professional forecast revisions have historically been reasonable
in the sense that they were unbiased. In other words, when GDP growth forecasts are revised
from 0.5 percent to 1 percent, our best guess is that actual GDP growth will be 1 percent. This
makes us sufficiently confident to present the forecast data here, but it does not establish that
the forecasts will be good guides to Japans future.

4.2 Stock market The small changes in forecasts of professionals and households may seem
surprising given Japans dramatic stock market boom. Between October 2012 and December
2013, the Nikkei 225 rose by 77 percent and the broader Topix index by 70 percent. Does this
not suggest much larger real effects of Abenomics than implied by professional forecasts? To
answer this question, we investigate how well stock prices forecast dividend growth.
To provide intuition for our exercise, we use the Gordon growth formula: when dividends
D grow at a constant rate g and are discounted at a constant rate r, then the stock price is
given by
P =

D
.
rg

(1)

Thus, prices should be proportional to dividends, holding r and g constant.

In 2013, stock prices rose much faster than dividends. The MSCI Japan index, which has
a long history of dividend data, has risen by 76.9 percent from October 2012 to December
2013 but nominal net dividends only rose by 9.1 percent since the third quarter of 2012. This
suggests that either r has fallen or g has risen. But only an increase in growth rates would
imply large real output effects. If discount rates (r) fall, for instance, because investors become
36

less risk averse or irrationally exuberant, then stock prices may not forecast real growth.45
One should keep in mind the limits of this exercise: the stock market, at best, forecasts
dividend growth and not necessarily GDP growth. These series differ notably: dividends are
much more volatile than GDP. Since 1980 the standard deviation of annual dividend growth
(22.3 percent) is nine times that of GDP growth (2.5 percent). For instance, in 2009 real GDP
fell by 5.5 percent while dividends declined by 36.2 percent. The overall correlation between
year-on-year GDP growth and dividend growth is quite lowonly 0.14 since 1980. Thus our
exercise allows us only to make qualitative statements about forecasts of future GDP growth.
We follow Campbell and Shiller (1988) to determine how well the stock market has historically forecasted dividend growth in Japan.46 We define the (ex-post) discount rate Rt+1 as the
value that makes current prices equal to discounted future prices plus dividends, Pt

Pt+1 +Dt+1
.
Rt+1

This implies that the discount rate is identical to the ex-post return. Log-linearizing this equation and solving forward for h periods yields the Campbell-Shiller decomposition
dt pt =

h1
X

rt+j+1

j=0

h1
X

j dt+j+1 + h1 (dt+h1 pt+h1 ).

(2)

j=0

Here dt pt is the log dividend-price ratio, rt+j+1 are future log discount rates, dt+j+1 is log
dividend growth, and =

1
1+exp{dp}

< 1 is a constant. For large h, and in the absence of

bubbles, the last term will be close to zero. Thus, as a matter of accounting, low dividend-price
ratios must be followed by either lower future discount rates (i.e., lower future returns) or higher
dividend growth. One can determine the importance of each component with the regressions,
h1
X

j rt+j+1 = ahr + bhr (dt pt ) + hr,t

(3)

j dt+j+1 = ahd + bhd (dt pt ) + hd,t .

(4)

j=0
h1
X
j=0

These regressions answer the following question: Given that stock prices have risen faster than
45

Note that our use of the term discount rates encompasses both rational and behavioral elements as in
Cochrane (2011). It should thus be thought of as a residual: any price movements we cannot explain with
dividend levels or dividend growth will necessarily show up in discount rates.
46
The first study we are aware of that tests for predictability of returns using the dividend-price ratio in
Japan is Campbell and Hamao (1992).

37

dividends in 2013, should we expect higher growth (higher g in (1)) or lower discount rates
in the future (lower r in (1))? In other words, we are asking how well the stock market has
historically forecasted growth. Note that for large enough h (and absent bubbles), we should
find that bhr bhd = 1. Further, the size of the coefficients determines the relative importance of
discount rates versus dividend growth: If bhr 1 then the dividend-price ratio forecasts discount
rates (returns); if bhd = 1 then the dividend-price ratio forecasts future dividend growth. In
their seminal work, Campbell and Shiller (1988) show that in the U.S., the dividend-price ratio
almost exclusively forecasts discount rates and not dividend growth, i.e., bhr 1 and bhd 0.
Table 4 tabulates the coefficients bhr and bhd at horizons of 10, 15, and 20 years for gross and
net returns. Like Campbell and Shiller (1988), we find that the dividend-price ratio strongly
forecasts discount rates (returns), with bhr 1. By contrast, estimates of bhd are always close to
zero and typically insignificant. If anything, the positive estimates of bhd imply that high prices
relative to dividends forecast lower future dividend growth. What does this mean? If bhd = 0,
then prices have historically reverted to a level consistent with dividends. Since dividends rose
by approximately 9.1 percent in 2013, our best (long-run) forecast for future prices is that they
will fall until they are 9.1 percent above those in November 2012. Thus, this exercise suggests
caution in interpreting high stock prices (relative to dividends) as an implicit forecast of future
dividend growth, even leaving aside the issues of linking dividend growth to GDP growth.

Table 4 Campbell-Shiller decomposition

h = 10 years
Net returns
Gross returns
N

bhr
0.97
(0.17)
0.95
(0.16)
133

bhd
0.03
(0.16)
0.07
(0.12)
133

h = 15 years
bhr
1.23
(0.10)
1.19
(0.11)
113

bhd
0.08
(0.09)
0.13
(0.06)
113

h = 20 years
bhr
1.13
(0.10)
1.10
(0.10)
93

bhd
-0.03
(0.05)
0.04
(0.04)
93

Ph1
Notes: bhr is the slope estimate of a regression of future realized returns, j=0 j rt+j+1 , on the log dividendPh1
price ratio dt pt . bhd is the slope estimate of a regression of future realized dividend growth, j=0 j dt+j+1 ,
on the log dividend-price ratio dt pt . The net returns row shows results for dividends net of taxes. The
gross returns row shows results for gross dividends. = 0.9874 based on historical mean of dividend-price
ratio. Newey-West standard errors are in parentheses (bandwidth = h). + p<0.10, p<0.05, p<0.010.

38

The stock market, at least unconditionally, gives no reason to expect future rapid growth
in Japan. It is more likely that Japan will see falling stock prices than higher dividend growth.
Or so the history suggests. To argue that the recent behavior of the stock market forecasts
future growth, one needs to argue that this robust historical relationship does not hold today.
4.3 The Roosevelt analogy Given uncertain predictions from professional forecasts and the
stock market, it is natural to turn to history as a guide to the prospects for Japans monetary
regime change. The analogy most often discussed is to Franklin Roosevelts monetary policy
regime change in spring 1933. This is far from the only possible analogy: for instance, one might
fruitfully compare current Japanese policies to disinflation efforts in the U.S. and Europe in
the early 1980s and to other countries efforts to reflate in the 1930s. But given its prominence,
we focus on the analogy to the U.S. in 1933.
The U.S. recovery after Franklin Roosevelts inauguration in March 1933 is widely interpreted as evidence for the effectiveness of monetary policy regime changeswith direct implications for Japan (Temin and Wigmore, 1990; Eggertsson, 2008, Romer, 2013).47 And it is
not only outside observers who have found the 1933 analogy useful. In a speech on December
25th , 2013, the Bank of Japan governor, Haruhiko Kuroda, argued that the U.S. economys
response to Roosevelts actions shows that monetary policy can quickly raise inflation expectations (Kuroda, 2013). In the rest of this subsection we explore the extent to which this
comparison is warranted: does 1933 indeed have lessons for Japan today?
Two obvious similarities have motivated comparisons. First, in 1933 in the U.S., as in Japan
in 2012, the economy was suffering from deflation, the output gap was large, and monetary
policy was constrained by the zero lower bound. Second, Franklin Roosevelt, like Shinzo Abe,
attempted a monetary policy regime change. He combined actions and words to convince the
public that deflation would be replaced by moderate inflation. But of course the scale of the
economic problems that confronted each leader was quite different. Roosevelts inauguration
followed three years of continuous large declines in output and prices. In 1932 alone, real GDP
fell 13 percent and the CPI fell 10 percent. By contrast, over 14 years, from 1998 to 2012, the
47

Also see this op-ed by Barry Eichengreen (http://english.caixin.com/2013-02-18/100491854.html?p2).

39

cumulative decline in the Japanese CPI was 4 percent.

Figure 8 shows the path of industrial production before and after Roosevelt and Abe took
office. It makes clear why the Roosevelt example inspires optimism about Abenomics. In
the four months following Roosevelts inauguration, seasonally adjusted industrial production
rose 57 percent. This initial expansion persisted: real GDP growth from 1934 through 1936
averaged 11 percent. Many economists have argued that this growth was primarily explained
by the effects of Roosevelts monetary policy on inflation and output expectations (Temin and

Industrial production (SA)

100
120
140

160

Wigmore (1990), Romer (1992), Eggertsson (2008)).

80

Abe, 12/2012
FDR, 3/1933
-24

-18

-12
-6
0
6
12
Months before / after regime change

18

24

Figure 8 Industrial production in the U.S. and Japan after Franklin Roosevelt took office in March
1933 and Shinzo Abe took office in December 2012. Industrial production is indexed to be 100 in the
month that each leader took office.

The lesson usually drawn from this episode is thus that regime changes can have large,
sustained effects on output. It suggests that Japans new monetary policy could have large
output effects. But it also raises a puzzle: why did Roosevelts actions have large, immediate
effects in a way that Abes policies have not? One answer is that Roosevelts policy change
was much larger. In appendix B, we provide a detailed comparison of the change in the real
interest rate in Japan now with that in the U.S. between 1932 and 1934. Using inflation swaps
as a measure of expected inflation, ex ante ten-year real interest rates fell 0.9 percentage points
between October 2012 and February 2014 in Japan. An analogous calculation, detailed in
appendix B, implies that real interest rates fell by two to four percentage points in the U.S.
between 1932 and 1934.
40

This exercise suggests that the change in the real interest rate in the U.S. in 1933 was
between two and four times as large as that in Japan in 2013. Therefore all else equal, one
might expect the effects of Abenomics to be only a quarter to one half as large as those of
Roosevelts actions. But this only resolves part of the puzzle. Abenomics has achieved far less
than half or even a quarter of U.S. growth after 1933. The optimistic view is that this puzzle
will be resolved by future rapid growth in Japan; perhaps the difference between Japan now
and the U.S. in 1933 is simply in the lags with which a monetary regime change affects the
economy. We are more inclined towards the pessimistic view that 2013 data reflect fundamental
differences between the current situation in Japan and that in the U.S. in 1933.
4.4 Keynesian models Like history, models are a natural place to turn for evidence on the
effects of a new macro policy. Both old and new Keynesian models suggest that monetary
policy can have large effects even in any economy with nominal interest rates at the zero lower
bound. For instance, Krugman (1998) and Eggertsson and Woodford (2003) use old and new
Keynesian models to argue in favor of raising inflation expectations when nominal interest rates
are constrained by the zero lower bound.48
We first use the baseline new Keynesian model from Woodford (2003, Ch. 4). It consists
of an Euler equation, a new Keynesian Phillips curve, and an interest-rate rule. It implies the
conventional IS relationship
yt = Et

rt+s + y ,

(5)

s=0

where yt is the deviation of output from steady-state, rt+s is the ex-ante one-year real interest
rate from t+s to t+s+1, and is the intertemporal elasticity of substitution. In this model, the
(peak) change in output from reducing the annualized ten-year real rate by r is

dY
dr

= 10.

Since typical calibrations set between 0.5 and 1 (e.g., Eggertsson and Woodford (2003);
Christiano, Eichenbaum, and Evans, 2005; Smets and Wouters, 2007), the 0.9 percentage point

48

Here we focus entirely on the models predictions for monetary policy. To the extent that fiscal multipliers
are large and that structural reforms have expansionary effects through wealth and credit effects (e.g., FernndezVillaverde, Guerrn-Quintana, and Rubio-Ramrez (2011) and Wieland (2014)), these calculations should be
viewed as conservative model-implied estimates of Abenomics total effects.

41

decline in the ten-year real interest rate49 should raise output by 4.5 to 9 percent. Since there
are no sources of persistence in this model, the output effect should be immediate.
As a second example, consider the Smets and Wouters (2007) model. This is a medium-scale
new Keynesian model that incorporates, among other features, capital, habits, sticky-wages,
price- and wage-indexation, and interest-rate smoothing. We take the estimated parameters in
Smets and Wouters (2007) as given and conduct the following experiment. First, we subject
the model to a discount factor shock such that the zero lower bound binds for eight quarters
the maximum duration for which a unique equilibrium exists in the model. Next we add a
one-standard-deviation monetary policy shock and calculate the difference in outcomes with
and without the monetary shock. We are particularly interested in two numbers: the change in
10-year real interest rates, dr, and the maximum change in output due to the monetary shock,
dY . We then calculate the peak interest-rate semi-elasticity of output,

dY
dr

. The result is -7.5

after five quarters. Based on the 0.9 percentage point decline in 10-year real interest rates, this
model predicts an increase in output of 7.5 0.9 = 6.75 percent after five quarters.
Unlike the baseline new Keynesian model in Woodford (2003, Ch. 4), the effects of a real
interest rate change in the Smets-Wouters model are not literally immediate. But they happen
more quickly than is consistent with Japanese data. Most of the decline in Japanese real
interest rates had occurred by late spring 2013, but output does not look as if it will be 6
percent higher in summer 2014. Thus conventionally calibrated new Keynesian models suggest
much largeror at least fastergains than are currently apparent in the data or in professional
forecasts.
Like the new Keynesian IS curve, an old Keynesian IS curve suggests large effects of a
change in the real interest rate. But since (some) old Keynesian models are purely backward
looking, they can do a better job of matching an initially small output gain in 2013. Consider
the old Keynesian IS curve suggested by Ball (1999): yt = yt1 rt1 , where yt is the natural
log of output in year t. Ball (1999) calibrates to be 0.8.50 Assuming that the reduction in
49

This is the decline in figure 2(c) between October 2012 and February 2014.
Ball (2006) argues that is 0.6 in Japan. For comparison with the new Keynesian models calibrated with
U.S. data, we use equal to 0.8.
50

42

ten-year bond yields is spread out equally over time, this implies an output gain of 4.3 percent
after ten years.51 Because the IS curve is purely backward looking, the dynamics are slow.
Output gains in the first year after the real interest rate change are only equal to the change in
the real interest rate. Thus this model fits with positive but small output gains from monetary
policy in 2013. It suggests that the gains from monetary policy will be more visible in 2014.
While old and new Keynesian models disagree about the speed of adjustment to a monetary
regime shift, they both suggest that the medium-run effects of monetary policy are larger than
professional forecasts imply. Furthermore, throughout this analysis we have assumed that the
monetary regime shift will remain imperfectly credible and thus will only lower future real
interest rates by 0.9 percentage points. Assuming that the policy becomes more credible over
time would imply even larger effects. In short, Keynesian models imply large real effects from
the first arrow of Abenomics. That we do not (yet) observe such rapid growth suggests either
that real-world dynamics are primarily backward-looking (as in old Keynesian models) or that
these models overestimate the effect of monetary policy.
4.5 The Consumption Tax Increases We have primarily focussed on monetary policy,
since it is the most radical element of Abenomics. But it is worthwhile comparing the future
effects of Japans expansionary monetary policy with those of the upcoming consumption tax
increases. As discussed in section 3, fiscal policy will become contractionary in the medium-run
due to the sales tax increase scheduled for April 2014 (from 5 to 8 percent) and October 2015
(from 8 to 10 percent). Here we confine ourselves to an examination of how these legislated tax
increases affect long-run forecasts.52
The original tax bill was passed on June 26th 2012. We compare long-term forecasts for 2014
and 2015 made in April 2012 and October 2012. 2014 growth forecasts were revised down by 0.5
percentage points and 2015 forecasts by 0.1 percentage points. This may be an upper bound on
the negative effects of the consumption tax increase. From April 2012 to October 2012, growth
forecasts for all years were generally revised down by 0.1 to 0.2 percentage points, suggesting
51
52

The effect in year t is given by yt = 1 0.8t/1 0.8 0.9 for t less than or equal 10.
Of course, just as with monetary policy, one could explore other sources of evidence on their effects.

43

that during these six months other bad news was revealed. This forecast-based estimate of
the consumption taxes effect can be compared to model-based estimates from the IMF that
imply a 1.0 percent cumulative contraction.53 The forecast revisions after the consumption
tax increase are much smaller than those that occurred after Abes new policies were revealed
even if we focus solely on the years 2014 and 2015. As discussed above, and shown in figure 7,
between October 2012 and October 2013, forecasters revised up their estimate of 2014 growth
by one percentage point and 2015 growth by 0.2 percentage points.
It is in some sense a semantic matter whether one wishes to consider the consumption tax
increases to be part of Abenomics. They are not new policies. But Abe enthusiastically adopted
them, and it is conventional to consider them an integral part of his second arrow. What our
discussion suggests is that even with these tax increases, Abenomics will likely have net positive
effects on output.
4.6 Discussion The preceding pages may appear needlessly complicated: why not simply
look to a model or to forecasters for insight into Japans future prospects? Our discussion
of several sources has, however, been deliberate. We do not believe that any single source
is a reliable guide. Here we return to the two motivating questions of this section: (1) will
Abenomics future benefits exceed costs? And (2) will Abenomics close Japans output gap?
We start with the latter question. In section 2.2, we estimated the 2013 output gap in
Japan to be 4.5 to 10 percent. It appears unlikely that Abenomics will close a gap this large;
current professional forecasts suggest the gap will actually widen. The historical behavior of the
Japanese stock market suggests that lower stock prices are more likely to lie in Japans future
than are higher profits and dividends. And both the 1933 analogy and new Keynesian models
have difficulty explaining slow growth in Japan in 2013, suggesting that these two sources may
be poor guides to the future.
Of course, this conclusion is enormously uncertain. Professional forecasters could be wrong.
Japans economy may simply be responding more slowly than the U.S economy did in 1933 or
53

Kang, Keen, and Pradhan (2011) show that a VAT increase of 1 percent of GDP reduces GDP by 0.4
percent in the IMFs GIMF model for Japan (their figure 4). We combine this with the estimated budget
impact of the consumption tax increase of 2.5 percent of GDP to obtain the 1.0 percent figure in the text.

44

new Keynesian models predict. And an old Keynesian IS curve, which thus far is consistent
with Japanese data, suggests large gains to come.
To the first question, we have both a more certain and more positive answer: the benefits
of Japans new monetary policy, and Abenomics as a whole, appear almost certain to exceed
the costs. In part, this is because all the data point to at least some positive output effect. But
even if monetary policy has no effect on output, the real interest rate decline has benefits for
Japans fiscal situation. Here we perform an illustrative calculation that compares the effects
of a change in the real interest rate with the effects of other policies on Japans budget outlook.
We take 2013 net debt of 140 percent of GDP from the October 2013 World Economic
Outlook and the following values from Doi, Hoshi, and Okimoto (2011): tax rates and social
security contributions are 30 percent of GDP, general government expenditure (including social
security benefits) are 39 percent of GDP, and annual real GDP growth is 1 percent. We further
assume that the real interest rate on debt for the next ten years is initially 0.4 percent per year
the yield on ten-year government bonds minus the ten-year inflation swap rate in October 2012.
We then compare the evolution of debt over the next ten years with and without Abes fiscal
and monetary policy. Abenomics impact consists of raising expenditures temporarily by 1
percent of GDP in year one and 0.5 percent in year two, raising tax income permanently by 1.5
percent of GDP in year two and another 1 percent in year three, and lowering the real interest
rate on debt by -0.9 percentage points per year.54 We make the conservative assumption that
any future deficits must be financed with debt carrying the old 0.5 percent real interest rate.
In the baseline scenario without Abenomics, net debt rises to 219.5 percent of GDP after
ten years. With Abenomics net debt rises to 186.8 percent. Thus Abenomics has shaved
32.7 percentage points off the future debt burden. Most of the decline is accounted for by
the consumption tax increase (21 percentage points) though the fall in real interest rates is a
non-trivial second (12.1 percentage points). The temporary stimulus packages, on the other
hand, play only a very small role over the ten-year horizon. Note that to be conservative, this
54

We implicitly assume that the government refinances all debt into ten-year bond yields. In practice,
since most of the decline in real interest rates comes from a rise in expected inflation (which does not require
refinancing), this assumption is likely not particularly important.

45

simulation excludes any effects of Abenomics on real GDP. If Abenomics increases real GDP,
it will further reduce the debt to GDP ratio. This effect would come both through a larger
denominator and through the effect of higher tax revenue on the numerator.
In short, Abenomics has improved the fiscal outlook through both the tax hikes and the
monetary arrow. If improving the fiscal outlook is a good thing in and of itself, then Abenomics,
and more expansionary monetary policy in particular, looks like good policy, even if one believes
output effects will be small or nonexistent.
Against the benefits of Japans new monetary policy, one must tally the costs. But to our
mind, costs if any, are likely small. One might think that the most obvious cost is higher
inflation. This is not necessarily true. Costs of higher price dispersion and deviations from the
Friedman rule need to be balanced against the benefit of higher steady-state nominal interest
rates, which allow the economy to avoid the zero lower bound. Coibion, Gorodnichenko, and
Wieland (2012) suggest that in an economy like Japans, the optimal inflation rate is more
likely to be two percent than it is to be zero percent.55
Another possible concern is that higher inflation will at some point drive down nominal
bond prices through the Fisher effect. Relative to a no-Abenomics baseline, this would have no
adverse effects on the government budget (unless real interest rates rose with nominal interest
rates). But it might cause problems for the banking sector, which holds large amounts of
government debt on its balance sheet (Hoshi, 2013). As of early 2014, this scenario is entirely
speculative: Abenomics has thus far driven up the price of government bonds, improving the
position of banks who hold these bonds.
Thus we see the evidence as strongly in favor of positive net benefits from Abenomics
whether considered as an entire policy package or as monetary policy alone. This is the more
remarkable since we have not considered the third arrowstructural reforms. Abes structural
reforms are as yet imprecisely specified, so we have left a detailed treatment to further work.
But the right reforms could undoubtedly raise potential and actual GDP.56
55

Ball (2013) argues for even higher inflation targets.

Perhaps most obviously, womens labor force participation in Japans lags behind that in the U.S. The
activity rate for women age 15 to 64 (number of women employed plus unemployed divided by the female
56

46

5 Credibility
We have thus far taken the effects of Abenomics on inflation expectations as given. We
have evaluated a policy that has not yet convinced markets or professional forecasters that
the two percent target will be reached. Even the Japanese government appears to be of two
minds about the credibility of this target: in December 2013, Takahiro Mitani, the head of the
semi-autonomous Japanese Government Pension Investment Fund, said that he expects CPI
inflation to remain between 0.1 and 1 percent.57
This raises two questions. First, why has the Bank of Japan been unable to make its target
fully credible? Second, is there a way the Bank of Japan could make its target more credible?
Start with the first question. The Bank of Japans lack of credibility is in some ways odd.
Two percent inflation is not high by international standards. And in standard new Keynesian
models if an inflation target is optimal then it is also a feasible (time-consistent) policy for a
central bank that cannot commit to future actions (Woodford, 2003, Ch. 7). These points of
reference suggest that the Bank of Japan ought to be able to raise inflation expectations to two
percent (or more).
Despite this, there are two good reasons to doubt the Bank of Japans commitment. First,
the Bank of Japan now has a record of announcing that it will achieve higher inflation and
failing to do so. In February 1999, it began its zero interest rate policy. It emphasized that it
would maintain this policy until deflation concerns subside (qtd. in Ito and Mishkin (2006), p.
145). In fact, it raised the interest rate to 0.25 percent in August 2000, a year in which the the
price level fell 0.7 percent (Ito and Mishkin, 2006). Likewise, when the Bank of Japan began
its quantitative easing program in March 2001, it said it would continue the policy until CPI
inflation excluding fresh food was positive. In this case, it followed the letter of its commitment,
ending the policy in March 2006 after several months of positive inflation (Ugai, 2007). But
prices quickly began falling again: in the twelve months from March 2006 to March 2007, the
CPI excluding fresh food fell 0.3 percent.
population) was 65 percent in Japan in the third quarter of 2013 versus 67.2 percent in the U.S. Some of Abes
proposals, such as more funding for childcare, could help eliminate this gap.
57
Bloomberg

47

Demographics and the structure of Japans pension system compound this credibility problem. Japans large retired population has benefited from deflation. In theory, Japanese pensions
are indexed to CPI inflation. But in practice, when CPI inflation has been negative, this indexation has been incomplete or has occurred only with a long lag.58 Furthermore, the Japanese
population holds over half its financial wealth in bank deposits and currency (compared to 13
percent in the U.S.). Inflation imposes losses for which (at the very least) those who are retired
cannot be compensated for with labor market improvement.59
The combination of benefits to the retired population and this groups outsized political
power leads Feldman, Berner, Caron, Mutkin, and Minack (2010) to argue that Japan made
a social decision to entrench deflation (p. 3) in the 2000s. Regardless of whether one entirely
agrees, the presence of a large constituency that benefits from deflation naturally leads to
questions about the governments commitment to two percent inflation. There is at least a hint
of evidence that forecasters fear that Japans next government might be less committed to this
target. After the Liberal Democratic party, led by Abe, won seats in the House of Councillors
on July 21, 2013, both 5 and 10 year inflation swap rates rose, albeit only slightly.
Much of this discussion suggests that credibility is outside the control of the Bank of Japan:
it cannot change Japans demography or pension system. But its actions matter. One lesson
from the financial market response to monetary announcements (section 3.1) is that for the
Bank of Japan, actions speak louder than words. No doubt in part because it had not lived
up to past commitments, merely committing to higher inflation, as it did in January 2013, had
a much smaller effect on markets than the announcement of concrete actions in April. This
suggests that further large-scale asset purchases could be beneficial.
If the Bank of Japan does manage to convince the public of its two percent inflation target,
this will likely further stimulate the economy. By raising current long-term inflation expectations from 1.1 percent to two percent, long-term real interest rates would decline a further
0.9 percentage points, doubling the decline in real interest rates that has occurred thus far. A
58

See Hosen (2010) and Ministry of Finance slides, p. 27. Incidentally, this lack of indexation is another
reason why inflation might improve Japans fiscal situation.
59
Bank of Japan, p.2.

48

linear model suggests this would double the effects of the policy, which we found added up to
a percentage point to 2013 GDP growth. As an illustrative calculation of the long-run effects
of full credibility, consider professional forecasts of output in 2022. Abenomics (excluding the
consumption tax increases) led this forecast to be revised up by 3.1 percent (section 4.1). If this
is entirely a monetary policy effect, then full credibility could double it, leaving 2022 output
as much as 6.2 percent above its no Abenomics baseline. Although illustrative, this is likely
an overestimate of what full credibility would achieve. The currently observed 3.1 percent upward forecast revision may reflect expectations of structural reforms, not only monetary policy
effects. Still, this calculation shows that making the two percent inflation target credible could
have large benefits and ought to be a high priority for the Japanese government.

6 Conclusion and Outlook

We have provided a preliminary evaluation of Abenomics and Japans monetary regime
change in particular. Our analysis suggests that Abenomics as a whole raised 2013 GDP growth
by 0.9 to 1.7 percentage points, with monetary policy accounting for up to one percentage point
of this gain. This suggests large net welfare gains from Abenomics. But there is as yet little
evidence that the policy will close Japans output gap, which we estimated to be 4.5 to 10
percent in 2013 (section 2.2). For instance, professional forecasts suggest an output gain from
Abenomics (excluding the consumption tax increases) of 3.1 percent by 2022. Since the output
gap implied by professional forecasts was projected to widen to 13 percent in 2017 absent
Abenomics, this gain is far from enough to close the gap. Of course, future performance is
highly uncertain, and it is quite possible that Abenomics may exceed (or underperform) our
and professional forecasters expectations. We therefore conclude with a brief discussion of
what future data releases would point to larger effects of Abenomics.
In part the modest effects of Abenomics relative to the output gap reflect the imperfect
credibility of the policy. Thus expected inflation is a crucial indicator. If our analysis is correct,
future upward movements in inflation expectations whether measured by inflation swaps or
surveys of professional forecasterswill be accompanied both by higher actual growth and
49

higher expectations of future growth. Concretely, as of January 2014, the Consensus forecasts
for 2014 growth and 2015 growth are 1.7 and 1.2 percent. If inflation expectations rise, we
expect the Japanese economy to beat these forecasts, perhaps significantly.
Sustained monetary stimulus likely depends on the outcome of future wage negotiations.
Assuming the marginal propensity to consume out of profits is low, real consumption cannot
grow indefinitely if CPI inflation continues to outstrip money wage growth. Quite simply, this
would violate individual budget constraints. Furthermore, a two percent inflation target is
unlikely to be politically sustainable if, in addition to lowering retirement incomes, inflation
lowers the incomes of working Japanese. Much more will be known about the future course of
nominal wages once the spring 2014 negotiations are complete.
In the short-run, the contractionary effects of fiscal policy are also an important unknown.
With the consumption tax to be implemented in April, there will be uneven growth: as of
the December 2013 quarterly consensus forecasts, output is expected to grow at an annualized
rate of 4.9 percent in the first quarter and -4.7 percent in the second quarter. If the Japanese
economy can exceed these forecasts, and resume growth in the second half of 2014, this would
be very good news for Abenomics. It would mean avoiding a repeat of the 1997 recession, which
some have attributed to the 1997 consumption tax hike.
Perhaps the largest long-run unknown about Abenomics is the third arrow, structural reforms. No one yet knows what reforms will be enacted or what the effect of these reforms will
be. If this arrow is to have large effects, we would expect to see it first in rising forecasts of
real future growth and lower inflation expectations. If inflation expectations fall at the same
time as growth forecasts rise, this would suggest that forecasters expect positive future supply
shocks.
Developments in Japan should be of wide interest to macroeconomists. The outcome of
Abenomics will determine whether Japan experiences another lost decade or resumes healthy
growth. Abenomics success may also influence policy in Europe and the U.S. As this was going
to press, in March 2014, both the Federal Reserve and the European Central Bank were up
against the zero lower bound, with inflation near one percent on both sides of the Atlantic.
50

Thus far, neither the Federal Reserve nor the European Central Bank has considered a radical
regime shift. If Abenomics succeeds, that may change.

References
Atkeson, Andrew and Ariel Burstein, Pricing-to-Market, Trade Costs, and International Relative Prices, American Economic Review, 2008, 98 (5), 19982031.
Auerbach, Alan J and Yuriy Gorodnichenko, Fiscal Multipliers in Japan, 2014.
Bahmani-Oskooee, Mohsen and Artatrana Ratha, The J-Curve: A Literature Review, Applied
Economics, 2004, 36 (13), 13771398.
Ball, Laurence, Efficient Rules for Monetary Policy, International Finance, 1999, 2 (1), 6383.
, The Case for Four Percent Inflation, Central Bank Review, 2013, 13 (2).
Ball, Laurence M, Fiscal Remedies for Japans Slump, in Takatoshi Ito and Andrew K. Rose,
eds., Monetary Policy under Very Low Inflation in the Pacific Rim, NBER-EASE, Volume
15, University of Chicago Press, 2006, pp. 279306.
, Hysteresis in Unemployment: Old and New Evidence, 2009.
Bank of Japan, Introduction of the Price Stability Target and the Open-Ended Asset Purchasing Method, 2013.
, Introduction of the Quantitative and Qualitative Monetary Easing, 2013.
Bernanke, Ben, Japanese Monetary Policy: A Case of Self-Induced Paralysis?, in Ryoichi
Mikitani and Adam Posen, eds., Japans Financial Crisis and Its Parallels to U.S. Experience,
Institute for International Economics, September 2000, pp. 149166.
Bernanke, Ben S and Mark Gertler, Inside the Black Box: The Credit Channel of Monetary
Policy, Journal of Economic Perspectives, 1995, 9 (4), 2748.
Blanchard, Olivier, Bubbles, Liquidity traps, and Monetary Policy: Comments on Jinushi et
al, and on Bernanke, in Ryoichi Mikitani and Adam Posen, eds., Japans Financial Crisis
and Its Parallels to U.S. Experience, Institute for International Economics, September 2000,
pp. 185193.
Blinder, Alan S, Quantitative Easing: Entrance and Exit Strategies, Federal Reserve Bank of
St. Louis Review, 2010, 92 (November/December 2010).
Campbell, John Y and Robert J Shiller, The Dividend-Price Ratio and Expectations of Future
Dividends and Discount Factors, Review of Financial Studies, 1988, 1 (3), 195228.
and Yasushi Hamao, Predictable Stock Returns in the United States and Japan: A Study
of Long-Term Capital Market Integration, Journal of Finance, 1992, 47 (1), 4369.
Cecchetti, Stephen G, Prices During the Great Depression: Was the Deflation of 1930-1932
Really Unanticipated?, American Economic Review, 1992, 82 (1), 141156.
Christiano, Lawrence J, Martin Eichenbaum, and Charles L Evans, Nominal Rigidities and
the Dynamic Effects of a Shock to Monetary Policy, Journal of Political Economy, 2005,
113 (1), 145.
Cochrane, John H, Presidential Address: Discount Rates, Journal of Finance, 2011, 66 (4),
10471108.
Coibion, Olivier and Yuriy Gorodnichenko, Information Rigidity and the Expectations Formation Process: A Simple Framework and New Facts, 2012.
and , Is The Phillips Curve Alive and Well After All? Inflation Expectations and the
Missing Disinflation, 2013.
, , and Johannes Wieland, The Optimal Inflation Rate in New Keynesian Models, Review
of Economic Studies, 2012, 79 (4), 13711406.
Deardorff, Alan, Trade Implications of the Trans-Pacific Partnership for Asean and Other
Asian Countries, 2013. Research Seminar in International Economics, Discussion Paper No.
51

638.
Doi, Takero, Takeo Hoshi, and Tatsuyoshi Okimoto, Japanese Government Debt and Sustainability of Fiscal Policy, Journal of the Japanese and International Economies, 2011, 25 (4),
414433.
Dominguez, Kathryn, The Role of the Yen, International Capital Flows, 1999, pp. 133171.
Dupor, William and Rong Li, The 2009 Recovery Act and the Expected Inflation Channel of
Government Spending, 2013.
Eggertsson, Gauti B, Great Expectations and the End of the Depression, American Economic
Review, 2008, 98 (4), 14761516.
and Michael Woodford, The Zero Bound on Interest Rates and Optimal Monetary Policy,
Brookings Papers on Economic Activity, 2003, (1), 139233.
Eichengreen, Barry, Japan Rising?, Milken Institute Review, October 2013, pp. 1525.
Fazzari, Steven M, R Glenn Hubbard, and Bruce C Petersen, Financing Constraints and
Corporate Investment, Brookings Papers on Economic Activity, 1988, pp. 141206.
Feldman, Robert Alan, Richard Berner, James Caron, Laurence Mutkin, and Gerard Minack,
Deflation: Will America and Europe Follow Japan?, 2010. Morgan Stanley MUFG Global
Economics.
Fernndez-Villaverde, Jess, Pablo A Guerrn-Quintana, and Juan Rubio-Ramrez, SupplySide Policies and the Zero Lower Bound, 2011.
Gagnon, Joseph E and Michael M Knetter, Markup Adjustment and Exchange Rate Fluctuations: Evidence from Panel Data on Automobile Exports, Journal of International Money
and Finance, 1995, 14 (2), 289310.
Hamilton, James D, Was the Deflation during the Great Depression Anticipated? Evidence
from the Commodity Futures Market, American Economic Review, 1992, pp. 157178.
Hooper, Peter, Karen Johnson, and Jaime Marquez, Trade Elasticities for the G-7 Countries
number 87. In Princeton Studies in International Economics., International Economics Section, Department of Economics, Princeton University, 2000.
Hosen, Mitsuo, Japans Public Pension: The Great Vulnerability to Deflation, 2010. Economic
and Social Research Institute, Cabinet Office, Government of Japan, ESRI Discussion Paper
No. 253.
Hoshi, Takeo, Will Abenomics Restore Japans Growth?, 2013.
and Anil Kashyap, Why Did Japan Stop Growing?, NIRA Report, 2011.
and Takatoshi Ito, Defying Gravity: How Long Will Japanese Government Bond Prices
Remain High?, 2012.
International Monetary Fund, Fiscal Monitor: Taking Stock, A Progress Report on Fiscal
Adjustment, Technical Report, Washington: International Monetary Fund 2012.
, Fiscal Monitor: Fiscal Adjustment in an Uncertain World, Technical Report, Washington:
International Monetary Fund April 2013.
, Fiscal Monitor: Taxing Times, Technical Report, Washington: International Monetary
Fund October 2013.
, World Economic Outlook: Transitions and Tensions, Technical Report, Washington: International Monetary Fund 2013.
Ito, Takatoshi, The Japanese Economy, Vol. 10, MIT press, 1992.
, Abenomics: Early Success and Prospects, Economy, Culture, and History: Japan Spotlight, September/October 2013, pp. 47.
and Frederic S Mishkin, Two Decades of Japanese Monetary Policy and the Deflation
Problem, in Takatoshi Ito and Andrew K. Rose, eds., Monetary Policy under Very Low
Inflation in the Pacific Rim, NBER-EASE, Volume 15, University of Chicago Press, 2006,
pp. 131202.
52

Jacobson, Margaret and Filippo Occhino, Behind the Slowdown of Potential GDP, 2013.
Federal Reserve Bank of Cleveland.
Kang, Kenneth, Michael Keen, and Mahmood Pradhan, Raising the Consumption Tax in Japan:
Why, When, How?, International Monetary Fund, 2011.
Kato, Takao, The End of Lifetime Employment in Japan?: Evidence from National Surveys
and Field Research, Journal of the Japanese and International Economies, 2001, 15 (4),
489514.
Kennedy, David M, Freedom from Fear: The American People in Depression and War, 19291945, Oxford University Press, 1999.
Krugman, Paul, Pricing to Market when the Exchange Rate Changes, 1987.
, Its Baaack: Japans Slump and the Return of the Liquidity Trap, Brookings Papers on
Economic Activity, 1998, (2), 137205.
, Thinking about the Liquidity Trap, Journal of the Japanese and International Economies,
2000, 14 (4), 221237.
, PPP and Japanese Inflation Expectations, 2013.
Kuroda, Haruhiko, Overcoming Deflation and After, 2013.
Kuttner, Kenneth N and Adam Simon Posen, The Great Recession: Lessons for Macroeconomic Policy from Japan, Brookings Papers on Economic Activity, 2001, (2), 93185.
Magee, Stephen P, Currency Contracts, Pass-Through, and Devaluation, Brookings Papers
on Economic Activity, 1973, pp. 303325.
Mandel, Benjamin R. and Geoffrey Barnes, Japanese Inflation Expectations, Revisited, 2013.
McCallum, Bennett T, Theoretical Analysis Regarding a Zero Lower Bound on Nominal Interest Rates, 2000.
Olivei, Giovanni and Silvana Tenreyro, Wage-Setting Patterns and Monetary Policy: International Evidence, Journal of Monetary Economics, 2010, 57 (7), 785802.
Posen, Adam Simon, Restoring Japans Economic Growth, Vol. 57, Peterson Institute, 1998.
Romer, Christina, It Takes a Regime Shift: Recent Developments in Japanese Monetary Policy
Through the Lens of the Great Depression, 2013.
Romer, Christina D, The Prewar Business Cycle Reconsidered: New Estimates of Gross National Product, 1869-1908, Journal of Political Economy, 1989, 97 (1), 137.
, What Ended the Great Depression?, Journal of Economic History, 1992, 52 (04), 757784.
, Why Did Prices Rise in the 1930s?, Journal of Economic History, 1999, 59 (1), 167199.
Romer, David H, Short-Run Fluctuations, University of California, 2012.
Schnell, Molly K and David E Weinstein, Evaluating the Economic Response to Japans Earthquake, 2012.
Smets, Frank and Rafael Wouters, Shocks and Frictions in US Business Cycles: A Bayesian
DSGE Approach, American Economic Review, 2007, 97 (3), 586606.
Svensson, Lars E O, Escaping from a Liquidity Trap and Deflation: The Foolproof Way and
Others, 2003.
Temin, Peter and Barrie A Wigmore, The End of One Big Deflation, Explorations in Economic
History, 1990, 27 (4), 483502.
Ueda, Kazuo, The Response of Asset Prices to Monetary Policy under Abenomics, Asian
Economic Policy Review, 2013, 8 (2), 252269.
Ugai, Hiroshi, Effects of the Quantitative Easing Policy: A Survey of Empirical Analyses,
Monetary and Economic Studies-Bank of Japan, 2007, 25 (1), 147.
Wieland, Johannes F, Are Negative Supply Shocks Expansionary at the Zero Lower Bound?,
2014. University of CaliforniaSan Diego, Working Paper.
Woodford, Michael, Interest and Prices: Foundations of a Theory of Monetary Policy, Princeton University Press, 2003.
53

APPENDIX

54

A Forecast unbiasedness
Table 5 Bias of Quarterly Consensus Forecasts
yt yt|t2
yt|t1 yt|t2

yt yt|t3
yt|t2 yt|t3

yt yt|t4
yt|t3 yt|t4

yt yt|t5
yt|t4 yt|t5

Panel A: CPI
1.06
1
(0.10)
N
82
2
R
0.521

1.19
(0.14)
80
0.462

1.05
(0.17)
80
0.285

1.23
(0.21)
79
0.284

1.11
(0.30)
78
0.143

Panel B: GDP
0.72+
1
(0.36)
N
69
R2
0.144

1.27
(0.36)
70
0.138

0.19
(0.36)
70
0.004

0.68
(0.26)
70
0.091

0.81
(0.22)
68
0.089

Panel C: Consumption
0.82
1
(0.35)
N
64
2
R
0.161

0.38
(0.34)
63
0.016

1.32
(0.49)
63
0.196

0.56+
(0.30)
63
0.062

0.14
(0.26)
61
0.003

Panel D: Industrial Production

1.23
1.31
1
(0.10)
(0.17)
N
81
79
2
R
0.847
0.532

1.38
(0.24)
80
0.378

1.59
(0.30)
79
0.283

1.53
(0.58)
77
0.079

LHS
RHS

yt yt|t1
yt|t yt|t1

Note: Estimated coefficient 1 from the regression yt yt|tl1 = 0 + 1 (yt|tl yt|tl1 ) + t . yt|tl are
forecasts for variable yt at date t l. Data are from Consensus forecasts at quarterly frequency. Newey-West
standard errors in parentheses. + p<0.10, p<0.05, p<0.010.

55

Table 6 Bias of Annual Consensus Forecasts

LHS
yt yt|t1
RHS
yt|t yt|t1
Panel A: CPI
0.99
1
(0.05)
N
30
2
R
0.852

yt yt|t2
yt|t1 yt|t2

yt yt|t3
yt|t2 yt|t3

yt yt|t4
yt|t3 yt|t4

yt yt|t5
yt|t4 yt|t5

0.90
(0.27)
28
0.351

0.54+
(0.28)
26
0.114

0.48
(0.20)
24
0.119

1.07
(0.35)
21
0.209

Panel B: GDP
1.01
1
(0.08)
N
30
R2
0.815

0.58
(0.60)
28
0.039

-0.14
(0.61)
26
0.001

0.43
(0.72)
24
0.006

2.43+
(1.19)
21
0.156

Panel C: Consumption
0.86
1
(0.14)
N
30
R2
0.627

0.35
(0.37)
28
0.036

0.42
(0.31)
26
0.061

0.11
(0.23)
24
0.003

0.87
(0.62)
21
0.129

Panel D: Industrial Production

0.91
1.31
1
(0.06)
(0.53)
N
30
28
2
R
0.904
0.206

-0.62
(0.38)
26
0.020

0.40
(1.17)
24
0.003

2.48
(1.72)
21
0.072

Note: Estimated coefficient 1 from the regression yt yt|tl1 = 0 + 1 (yt|tl yt|tl1 ) + t . yt|tl are
forecasts for variable yt at date t l. Data are from Consensus forecasts at annual frequency. Newey-West
standard errors in parentheses. + p<0.10, p<0.05, p<0.010.

B Franklin Roosevelt and Shinzo Abe

Table 7 compares the change in the real interest rate between October 2012 and February
2014 in Japan with that in the U.S. between 1932 and 1934. The first line shows expected
inflation before the monetary policy change. For Japan, this is straightforward to measure; one
can use either a market measure from inflation swaps or the forecast from Consensus Economics.
In the table, we use the value from inflation swaps. Unfortunately, there is no analog to these
measures of inflation expectations for the U.S. in 1933.60 We consider two possible scenarios.
In scenario A, we assume that expected ten-year inflation in 1932 was zero. This is a sensible
baseline for two reasons: (1) inflation in the 1920s was near zero. (2) It is a reasonable midpoint
given the plausibility of both further deflation and inflation.61
In scenarios B and C we apply a simple version of the method used by Cecchetti (1992),
60

Cecchetti (1992) and Hamilton (1992) do provide estimates of expected inflation in the 1930s. But these
estimates are for one-quarter ahead inflation and are thus not ideal for our purposes.
61
Further deflation was plausible given that 1932 was the third consecutive year of large price declines. Many
people must have expected these to continue. Inflation was plausible given strong political pressure to raise
prices to pre-depression levels (Kennedy, 1999).

56

Table 7 Two regime changes

Japan

U.S., A

U.S., B

U.S., C

Expected 10-year inflation pre-policy (%)

Nominal 10-year rate pre-policy (%)
Ex ante 10-year r pre-policy (%)

0.3
0.8
0.4

0.0
3.7
3.7

-2.3
3.7
6.0

-2.3
3.7
6.0

Expected 10-year inflation post-policy (%)

Nominal 10-year rate post-policy (%)
Ex ante 10-year r post-policy (%)

1.1
0.6
-0.5

1.2
3.1
1.9

1.2
3.1
1.9

-0.5
3.1
3.7

dr (percentage points)

-0.9

-1.8

-4.1

-2.4

Notes: Comparison of real interest rate behavior during Abenomics with the U.S. regime change in 1933. In
U.S. scenario A, pre-policy expected inflation is assumed to be zero; post-policy expected inflation is assumed to
be equal to actual inflation from 1934 to 1939. In scenario B, pre-policy expected inflation comes from the VAR
forecast described in the text; post-policy expected inflation is again assumed to be equal to actual inflation
from 1934 to 1939. In scenario C, both pre-policy and post-policy expected inflation comes from the VAR.
Data Sources: See data appendix.

Hamilton (1992), and Romer (1992) to measure expected inflation. We estimate a three variable
VAR with annual data from 1920 to 1940 for real output growth, the change in long-term
nominal government bond yields, and CPI inflation. We use one lag. The predicted values for
CPI growth are our measure of expected inflation: they represent the best guess of inflation
in the future given this model of the economy. Column 3 shows the resulting 1932 forecast of
ten-year inflation. With this model of the economy, one would have forecast continued deflation
of -2.3 percent per year after 1932.
Post-policy expected inflation, line 6, is again straightforward to measure for Japan with
inflation swaps. In the U.S. case, we consider two possibilities. Scenarios A and B assume
that expected inflation equaled actual inflation from 1934 to 1939. In scenario C, we use the
ten-year inflation rate predicted by the above-described VAR. Although actual CPI inflation
from 1934 to 1939 was just above one percent, the VAR suggests that expected inflation, even
in 1934, was slightly negative (-0.5 percent per year).

57

C Data appendix
Activity rates: OECD series LRAC64FE.
CPI: We use the official data from the Japanese Statistics Bureau
(http://www.stat.go.jp/english/data/cpi/).
Employed population age 15-64: OECD series LFEM64TT.
Exchange rate nominal: We use noon buying rates in New York, as reported by the
Federal Reserve (FRED series DEXJPUS). For the announcement table we use the Bloomberg
series, USDJPY.
Exchange rate real: Bank for International Settlements trade-weighted index. We use
the broad index including 61 countries.
See http://www.bis.org/statistics/eer/.
Government bond yields: We use the Bloomberg series GJGBX where X is the years to
maturity. The last day of data is 2/19/14.
Gross and net debt: October, 2013 IMF World Economic Outlook.
Inflation swaps: We use the Bloomberg series JYSWIT2, JYSWIT10. The last day of
data is 2/19/14.
Land price index in six large Japanese cities: Statistics Bureaus Urban Land Price
Index. The six cities are Tokyo, Yokohama, Nagoya, Kyoto, Osaka, and Kobe.
Money market interest rate: IMF, International Financial Statistics.
Money supply: Narrow is OECD series MANMNM01; broad is OECD series MABMM301.
Multifactor productivity: OECD StatExtracts, multifactor productivity table.
Nominal hourly earnings: We compute hourly earnings as the ratio of total monthly
wages to hours worked. Data are from the final report of the Monthly Labor Survey from the
Ministry of Health, Labor and Welfare
(http://www.mhlw.go.jp/english/database/db-l/25/2512re/2512re.html).
Real GDP and GDP deflator: Official Japanese real GDP are available from the
Japanese Cabinet office (http://www.esri.cao.go.jp/en/sna/menu.html). These data begin in
1994. Prior to 1994, we use the real GDP index from the IMFs International Financial Statistics. We ratio splice this series to the official series in 1994.
Real GDP PPP-adjusted: The OECDs measure of constant-PPP, constant-prices
GDP.
Share of electricity from nuclear power: See
http://www.eia.gov/countries/cab.cfm?fips=ja.
Stock Market: We use Nikkei 225 data from Yahoo finance and Topix data from the
Bloomberg series TPX. Topix data ends on 2/19/14.
Trade shares and quantities: Trade shares are from the Ministry of Finance, Trade
Statistics of Japan. This is also the source for data on changes in trade quantities and values.
See http://www.customs.go.jp/toukei/shinbun/happyou_e.htm.
Unemployment rate: We use the OECD series LRUNTTTT.
Working age population: We use the OECD series for the population age 15 to 64,
LFWA64TT. This series ends in the third quarter of 2013. We impute the value in the fourth
quarter of 2013 using the average growth from the third quarter of 2012 to the third quarter of
2013.
U.S. data in Table 1: Real GDP: Federal Reserve Bank of St. Louis, FRED database
series GDPCA. Other series as above.
U.S. 1930s comparison (table 7): CPI: Federal Reserve Bank of St. Louis, FRED
58

series CPIAUCNS. Real output growth: prior to 1929, this is measured as the percent change
in the real GNP series constructed by Romer (1989). In 1929, this series is spliced to the
Bureau of Economic Analysis series for real GDP measured in chained 1937 dollars (NIPA
table 1.1.6A). Long-term government bond yields: From 1919 to 1925, this is the average yield
on U.S. government bonds due or callable after eight years. After 1925, it is the average yield
on U.S. government bonds due or callable after 12 years. The series is presented and described
in the August 1938 Federal Reserve Bulletin, p. 1045. We take the data from the NBER
macrohistory database series m13033.

59

Brookings Panel on Economic Activity

March 2021, 2014

The Political Economy of Discretionary Spending:

Evidence from the American Recovery and
Reinvestment Act
Christopher Boone, Columbia University
Arindraji Dube, University of Massachusetts Amherst
Ethan Kaplan, University of Maryland at College Park
Final conference draft

Abstract
We study the spatial allocation of contract, grant and loan awards in the ARRA, one of the largest
discretionary funding bills in the history of the United States. Contrary to both evidence from previous
fiscal stimulus and standard theories of legislative politics, we do not find evidence of substantial political
targeting.
Party leaders did not receive more funds than rank-and file legislators. Democrats in swing districts
did not receive more than Republicans in such districts. Ideological moderates and pivotal members of
Congress received slightly less than average. While Democratic districts overall received slightly more
than Republican districts per resident, this differential mostly disappears when we consider award per
worker in the district.
At the same time, we find no relationship between extent of award and measures of the severity of the
downturn in the local economy, or the pace at which funds were spent, while we do find more funds owing
to areas with more economic activity and a greater incidence of poverty.
Overall, most of the variation in spending across congressional districts remains unexplained. The results
are consistent with the discretionary component of ARRA being allocated based on project characteristics
other than countercyclical efficacy or political expediencyboth of which stand in contrast to evidence
from fiscal stimulus in the New Deal. One explanation suggests that legislative norms have reduced the
scope of discretionwith attendant benefits and costs.

The Political Economy of Discretionary

Spending: Evidence from the American
Recovery and Reinvestment Act
Christopher Boone, Arindrajit Dube, and Ethan Kaplan
March 12, 2014

Abstract
We study the spatial allocation of contract, grant and loan awards
in the ARRA, one of the largest discretionary funding bills in the history of the United States. Contrary to both evidence from previous
fiscal stimulus and standard theories of legislative politics, we do not
find evidence of substantial political targeting. Party leaders did not
receive more funds than rank-and-file legislators. Democrats in swing
districts did not receive more than Republicans in such districts. Ideological moderates and pivotal members of Congress received slightly
less than average. While Demoratic districts overall received slightly
more than Republican districts per resident, this differential mostly
disappears when we consider award per worker in the district. At the

We thank Bryan Hardy, Kyle Meng, Ruchita Tiwary, Sebastien Turban and Laurence Wilse-Samson for excellent research assistance and the Institute for New Economic
Thinking for generous financial support. Any mistakes are our own.

Columbia University

University of Massachusetts Amherst

University of Maryland at College Park

same time, we find no relationship between extent of award and measures of the severity of the downturn in the local economy, or the pace
at which funds were spent, while we do find more funds flowing to
areas with more economic activity and a greater incidence of poverty.
Overall, most of the variation in spending across congressional districts
remains unexplained. The results are consistent with the discretionary
component of ARRA being allocated based on project charactersitics
other than countercyclical efficacy or political expediencyboth of
which stand in contrast to evidence from fiscal stimulus in the New
Deal. One explanation suggests that legislative norms have reduced
the scope of discretionwith attendant benefits and costs.

Introduction

The American Recovery and Reinvestment Act (ARRA), otherwise known

as the Obama stimulus, was one of the largest discretionary spending bills
in US history. At the time of passage, it allocated a total of $787 billion,
consisting of: $212 billion in tax cuts; $267 billion in entitlement programs,
which was largely channeled through state governments; and $308 billion in
discretionary projects awarded through contracts, grants and loans. In this
paper, we study the allocation of the funds in the ARRA, with an eye towards
several objectives. First, given the size and significance of this bill, understanding how and where the money was spent is important in and of itself.
Second, we use this bill to test theories from political economy in order to
learn about the legislative process and the distribution of government spending in general. While the tax and entitlement components of the spending
were largely formulaic, the allocation of the discretionary component affords
a window into assessing a variety of economic and political considerations
that may have played some role. Finally, the ARRA provides an opportunity to examine the political economy of a particular type of government
2

spendingnamely, fiscal policy for the purpose of macroeconomic stabilization. In doing so, we consider the implications for improving the design of
such policies.
Passed during the the Great Recession, the stated aim of the ARRA bill
was to stimulate the economy, in large part by saving and creating jobs.
Recent evidence suggests that state and local multipliers are larger in areas
with greater excess capacity (Dube, Kaplan and Zipperer, 2014; Nakamura
and Steinsson 2013; Shoag 2010). Thus we might expect areas with larger
increases in unemployment to receive more money: in other words, Sacramento, California would be a larger recipient than, say, Flint, Michigan or
Fargo, North Dakota. Secondly, we might expect lower income areas to receive more funds. Johnson et al. (2001) document that individuals with low
levels of income and low levels of liquidity respond more strongly to stimulus.
Finally, we might expect primacy placed upon projects which are more labor
intensive or projects which are more rapidly implementable (shovel ready).
Contrary to such expectations, however, there was no significant or consistent relationship between the amount spent and a number of different
measures of local unemployment, including the unemployment rate in January 2009, the change in the unemployment rate between 2007 and 2009.
We also find no evidence that more funds went to districts that were able to
spend the money more quickly (i.e., those with more shovel ready projects).
While our proxies (especially of shovel readiness) are admittedly imperfect,
targeting of areas with high unemployment, excess capacity, or shovel readiness did not appear to play a major role in the allocation of the discretionary
funds across Congressional Districts. We do find that districts with greater
employment-per-resident and districts with greater incidence of poverty received more but these characteristics are not directly linked to countercyclical objectives. To be sure, some of the formula-based components of the
ARRA bill, such as Medicaid funding or unemployment insurance benefits,
explicitly targeted areas with higher unemployment. However, such targeting

did not occur for the most discretionary component of the bill.
To political economists, the lack of evidence in favor of optimal economic
targeting is not surprising, given that the actual allocation of funds occurs
through the legislative process. Political economy models make other predictions about the cross-sectional variation in discretionary spending, based
on the interests of individual politicians; instead of simply maximizing social
welfare, politicians may seek to maximize party success, individual success,
or some combination of the three. In this sense, the ARRA affords a window
into testing theories of positive political economy, a key goal of which is to
understand in whose interests politicians act.
One view is that politicians act in the interests of the party. However,
there is a longstanding literature in political science suggesting that the
United States, having a majoritarian political system, has weak political
parties (Baron, 1991; Shor, 2006a; McCarty, Poole and Rosenthal, 2001).
Our evidence supports the weak party hypothesis. In our baseline sample
that excludes state capitals, Democratic districts in the House recieved, on
average, only $95 more per capita in discretionary funding than their Republican counterparts, despite the fact that Democrats had large majorities in
both chambers of Congress and held the presidency at the time of passage
of the ARRA bill. Controlling for district characteristics that were relevant
for some of the funding formulas, this differential falls to $34 per capita and
is not statistically significant. In addition, we do no find any differential by
party when we consider award per worker (as opposed to per resident) in the
districtswith or without controls. We do find that strongly Democratic districts (with between 80 and 90 percent Democratic Party vote share) received
more funds. However, this differential is at least in part driven by a small
number of very dense urban districts with high levels of employmentand
the differential is smaller when we consider the stimulus award per worker
in the district.
Importantly, neither of the two parties appear to have targeted extra

funds towards marginal districts, i.e., those with close outcomes in the previous election. We find no significant impact of being a marginal Democrat
versus a marginal Republican. This is an important result for two reasons.
First, legislators working in the interests of the party might choose to target
funds towards marginal districts because these are the districts that are most
vulnerable to switching parties in the next election. Second, close Democratic
districts and close Republican districts are likely to be similar along many
other dimensions as well, and so this comparison offers the cleanest empirical
test of whether districts were targeted based on their explicit party affiliation. Our results here echo those of Albouy (2009) who, using a regression
discontinuity design, finds that states with a majority of congressional districts aligned with the party of the president receive a relatively small and
statistically insignificant 2.6% more in intergovernmental transfers from the
Federal Government. Shor (2006a; 2006b) finds similar results for the allocation of funds in state legislatures. Using less well-identified panel methods,
Berry, Burden and Howell (2010) do find that counties with a representative
aligned with the party of the president receive more funds. In contrast to the
literature, we focus on a single piece of legislation. However, the legislation
we analyze is sizeable with $308 billion in discretionary funds. Albouy (2009)
finds his null results during a time of divided government when the president
had to bargain with Congress in order to get legislation passed. Interestingly,
we find the same null effect during a period of strong unified government.
The voting on the ARRA bill was highly polarized with only seven Democrats
voting against the bill and zero Republicans voting for it.1 Party defiers who
either voted with the opposition or abstained from voting received a somewhat smaller average amount of contract/grant/loan money conditional upon
covariates compared to those who voted with their party. However, the magnitudes are not very large, and the results are not statistically significant in
1

The seven Democrats includes Parker Griffith, a Democrat from Alabamas fifth Congressional district, who switched parties 10 months after the vote.

most specifications. Overall, there is not much that the discretionary ARRA
spending was aimed at maximizing the success of either political party.
A second view posits that politicians are self-interested and maximize
their own return without strong regard to either general welfare or party
interests. In this case, districts with powerful politicians should receive more
discretionary funds. Both Knight (2005) and Cox, Kousser and McCubbins
(2010) provide suggestive empirical evidence that those with agenda setting
power are able to gain substantial rents. In contrast, Berry, Burden and
Howell (2010) use a 23-year panel of county-level disbursements of federal
government expenditures and find that committee leaders and members of
important comittees do not receive greater amounts of federal dollars per
capita than other districts. Since we consider a single large discretionary
spending bill, our statistical power is more limited; however, the funds we
consider are all discretionary, which is an attractive feature of our analysis.
We find that the Democratic leadership did not receive more funds than other
legislators. The Republican leadership did receive somewhat more, depending on specifications, but these results are often not statistically significant
and driven by two outliers.. Overall, these findings are not strongly consistent with an explanation of funding allocation based on the behavior of
self-interested politicians.
Similarly, theory suggests that political moderates who can credibly threaten
to vote for either side should need to be bribed to vote. The most commonly
used model in political economy, the probabilistic voting model (Lindbeck
and Weibull, 1987), captures this intuition. It finds that policy and rents
are apportioned to individuals in rough proportion to their probability of being swing voters. Rent accrual to ideological moderates has been verified in
the Swedish context (Dahlberg and Johansson, 2002). However, in contrast
to Dahlberg and Johansson (2002), we find that ideological moderates were
actually less likely to receive ARRA funds. Moreover, swing districts also
receive a lower proportion of ARRA funds than more partisan districts. In

the Senate, where the bill necessarily relied on the support of 3 Republicans
and one conservative Democrat to overcome a Republican fillibuster, we find
no evidence that pivotal states siding with the majority received more funds.
A related theory posits that the president targets swing states in the electoral college. Similar to Menchaca (2012), we find no evidence for strategic
targeting of likely swing states by the president. Moreover, in contrast to
Larcinese et al. (2006) but in concordance with Inman (2010), we find little
evidence of greater funding for states marginally aligned with the party of
the president.
Overall, when we consider the variation in spending across congressional
districts, we do not find much evidence of targeting based on both countercyclical or political economic considerations. It is possible that Democrats
felt constrained in their ability to explicitly target districts based on partisan
affiliation, yet they still engaged in a more subtle form of targeting, whereby
funds were targeted based on characteristics such as employment and poverty
because those characteristics tend to be associated with Democratic districts.
Another possibility is that the policy preferences of Democratic legislators
support the types of projects that tend to be located in Democratic constituencies. While we can rule out more explicit measures of partisan targeting, we are not able to identify or distinguish between these more subtle and
limited forms.
We do find that districts with greater employment-per-resident (i.e., more
urban and economically more active areas) and districts with greater incidence of poverty received more funding. This could reflect a form of economic
targeting, or reflect a greater weight on Democratic party priorities. However,
in general the evidence suggests that both economic and political targeting
was limited. We note that we are only able to account for less than 30% of
the variation in our baseline model with 14 variables and 334 observations.
Most of the variation in ARRA receipts across districts is unexplained. At
the same time, much of our evidence is consistent with the burgeoning empir-

ical political economy literature in both economics and in political science,

which document similar failings of the theories.
The relative lack of political and counter-cyclical targeting in the most
discretionary components of the ARRA contrasts with evidence from the
New Deal legislation. Fishback, Kantor and Wallis (2003) argue that grants
under the New Deal were targeted both to high unempoyment areas and
to swing district supporters of President Roosevent in the prior presidential
campaign. In the section where we interpret our results, we consider some
possible explanations for this difference. These include the relatively speedy
enactment of the legislation; the non-cyclical (i.e., reinvestment) goals of
the bill focusing on public goods provision; and a greater reliance on rules
over discretion in the current legislative envrionment. We also discuss some
policy implications of our findings on targeting.
The remainder of the paper is structured as follows. In section 2, we
summarize how decisions were made to allocate ARRA funds. In section
3, we describe the data that we use. Section 4 presents our results, and in
section 5 we interpret our findings and discuss policy implications. Section
6 concludes.

How Congress Budgeted the ARRA

Budget items are normally sent through finance committees in both the
House and the Senate. In the case of the ARRA, most of the details of
the contract, grants and loans (hereafter CGL) portion of the bill were sent
to the 12 appropriations subcommittees in both the House and the Senate.
Since the Obama administration and leaders in the Republican Party both
had argued strongly against expenditures which were explicitly targeted to
particular districts (earmarks), no earmarks were incorporated into the bill.
Before the bill was passed, each of the 12 appropriations subcommittees separately in both the House and the Senate came up with proposed budgets.
These budgets were then reconciled across the House and the Senate and ul-

timately written into the bill. Since no Republican House members actually
voted for the bill, Republicans in the House of Representatives played little
role in formulating the budget. In the Senate, all of the Democrats who were
present voted in favor of the bill; they were joined by two independents, Joe
Lieberman from Connecticut and Bernie Sanders from Vermont2 . In addition, one Democratic Senator, Ben Nelson of Nebraska publicly announced
his ambivalence towards voting in favor of the bill in the weeks leading up to
the vote. The Democrats and independents were joined by three Republican
Senators Susan Collins and Olympia Snowe from Maine, and Arlen Specter
from Pennsylvania3 . So, in the Senate, Republicans played a larger role than
in the House, but the bill was still predominantly formulated by staffers on
the Democratic side.
The bill was crafted with strong time constraints. Thus most of the appropriations committees doled out money based upon pre-existing formulas
for allocating funds. However, due to the new administrations interest in
selecting at least some projects based solely on project quality, many committees allocated a minority of their funds through competitive grants. Thus,
almost all of the CGL money was spent using either funding formulas or
competitive grants.
The competitive grant money was allocated to the departments in the
Executive Branch and was then allocated to recipients based upon merit using
criteria described in the bill. The criteria used by the various departments
differed by type of grant and were written by both legislators and staffers.
The formula money was given to the governors. However, gubernatorial
discretion was limited in terms of how it could be spent, where it could be
spent and how quickly it should be spent. For example, highway money had
to be used to build and repair highways and 50% of it had to be spent within
2

Ted Kennedy did not vote due to illness and Al Franken was not allowed to sit for the
Senate seat until the summer of 2009 due to litigation following a very close election.
3
Arlen Specter was a Republican until May 2009 when he switched to the Democratic
Party.

6 months, all of it within a year. The money that remained unspent was to be
redistributed to other states by the Federal Department of Transportation.
Other portions of the bill mandated that a certain percent of the funds be
spent in rural or urban areas.
Despite the pervasive use of funding formulas and competitive grants,
there was still room for political influence over the distribution of expenditures. The Obama administration pushed certain projects of interest to
the administration (money for alternative energy, high speed rail, local public transportation). Similarly, whereas pushing for district-specific projects
was not allowed, members of Congress could have influenced the amount of
money spent through different programs. For example, members of Congress
from rural areas might have tried to push the subcommittees to increase allotments to funding formulas based upon highway miles while representatives
from urban districts might have pushed for increased funds through funding
formulas for public transportation systems. Overall, despite the pervasive use
of funding formulas and competitive bidding, there was indeed substantial
scope for politicians to affect the geographical distribution of funds.

3
3.1

Data Description
American Recovery and Reinvestment Act of 2009

Our data on ARRA spending comes from the website www.recovery.gov,

which was created to provide taxpayers with information on how the ARRA
funds are spent, as mandated by the Act itself. Spending under the ARRA
can be divided into three major categories: tax benefits; entitlements; and
contracts, grants and loans. For the last category, Recovery.gov provides
information on each individual recipient, including the recipients address
along with the primary zip code and congressional district where the activity
was to be carried out. For the tax benefits and entitlements categories,
only state-level statistics are available. Most of the analysis in this paper
is therefore focused exclusively on the CGL funds. As of January 2014, the
10

total estimated expenditure under the ARRA was $840 billion (an increase
from the original estimate of $787 billion). Of this, CGL funds accounted for
around $267 billion. We aggregate the amount of disbursement to the House
district level and, for some specifications, to the state level.
We exclude from our baseline analysis all congressional districts containing any part of a state capital. We do this because a large portion of the
money going to state capital districts was disbursed to the state governments
and subsequently redistributed across the state. For example, educational
grants are predominantly sent to state capitals and then distributed by governors and state legislatures throughout the state. This type of spending
is recorded in the ARRA data as going to the congressional district where
the headquarters of the state governmental agency is located, and does not
accurately reflect the distribution of spending across congressional districts.
After dropping congressional districts containing any part of a state capital,
we are left with 334 out of 435 congressional districts, implying that the average capital city is contained within two congressional districts. We discuss
this issue in more detail below. Overall, $106 billion of the $267 billion in
CGL funds remains after excluding state capitals from the sample.

3.2

Voting and Congress Data

We use vote returns from Congressional Quarterly for the November 2008
general election for the U.S. House of Representatives. The two-party Democratic vote share in each congressional district is computed as the number of
votes for the Democratic candidate divided by the total votes for the Democratic and Republican candidates. The median Democratic vote share in
our sample is 57.7%, reflecting the Democratic majority in the House at the
time. On February 13, 2009 the date of the vote on the ARRA conference
report there were 255 Democrats, 178 Republicans, and 2 vacant seats in
the House of Representatives.
Data on the tenure for each member of the House comes from the Office
of the Clerk of the House of Representatives. We use data from the website
11

Govtrack.us to find committee assignments for the 111th Congress and each
representatives vote on the ARRA. DW-Nominate scores for individual legislators were downloaded from voteview.com. For district-level presidential
vote returns, we use data from the Swing State Project.

3.3

Other Data

We also use data for a number of demographic and economic characteristics at the state and district levels. Data on state-level Medicaid and
county-level unemployment insurance expenditure is from the Bureau of Economic Analysis. District-level information on population, poverty, and land
area comes from the U.S. Census Bureau. We use data on county-level unemployment rates from the Local Area Unemployment Statistics program at
BLS. County-level information on home loans is from Home Mortgage Disclosure Act (HMDA) data made available by the Federal Financial Institutions
Examination Council. District-level characteristics for unemployment rate,
home loans, highway miles, and unemployment insurance are derived from
the county-level information using geographic correspondences provided by
the Missouri Census Data Center. For data on total employment, we use
the 2008 LEHD Origin-Destination Employment Statistics (LODES) data
available from the U.S. Census Bureau, and aggregate the census block-level
data to the district level.4

Results

4.1

What Was the Money Spent On?

In this section, we give an overview of what types of projects were funded

with the ARRA money. In figures A1 and A2a and A2b, we display a map
of the United States where darker areas received more funding per capita.
Figure A1 shows amounts per capita aggregated to the state level. Figures
4

LODES data is not available for Massachusetts, so we instead use the ZIP Code
Business Patterns from the U.S. Census Bureau, and derive district-level employment
using geographic correspondences from the Missouri Census Data Center.

12

A2a and A2b show amounts at the congressional district level, with Figure
A2b omitting districts continaing portions of state capitals. Figure A2b
displays the distribution of funding we use in our main district-level analysis.
There are no obvious patterns from the maps.
The mean amount of stimulus funds received per resident in a district
in the form of contracts, grants and loans was $469 per capita, or $900
per capita including state capitals. This reflects the exclusion of all CGL
funds to state governments, particularly education funds which constituted
a very large proportion of total funds. We show descriptive statistics in
Table 1. There is sizeable variation in expenditure across districts. The
standard deviation of per capita expenditure is $542 per person and $1786
per person including state capitals. State-level amounts per resident are
slightly higher than district-level amounts ($156 excluding money going to
state capitals and $48 including money going to state capitals). This reflects
small states getting slightly larger amounts. Additionally, since the House
of Representatives was highly Democratic at the time, the mean Democratic
vote share was 58%. The vote share was identical for districts containing at
least some portion of a state capital.
Focusing only on districts not located in state capitals, we begin by showing a kernel density estimate of the per capita expenditure.5 The distribution,
shown in Figure 1a, is heavily skewed to the right. The lowest amount of
money obtained by any district was Anthony Weiners (Democratic) district
in New York City, which received around $7 per person of CGL funding. The
most received by any congressional district was around $3750 per person for
Doc Hastings (Republican) district containing Spokane, Washington. Hastings district received over 500 times more money per person than Weiners
district. In Figure 1b, we show a histogram and kernel density estimates of
the Democratic vote share. The mean vote share was approximately 57% for
5

We use an Epanechnikov kernel with an optimal bandwidth minimizing mean squared

integrable error relative to a fitted Gaussian distribution.

13

the Democrats, reflecting that the House was highly Democratic at the time.
There are two modes of the distribution of two-party vote share, one Republican mode around 40% and another Democratic mode around 70%. There
are also a few districts with uncontested Republican winners and almost 3
times as many Democratic ones.
The amount of variation across congressional districts was substantial. If
most districts received a similar amount of funds, then variation across districts in levels of funding would constitute a small fraction of total funding.
However, when we compute the de-meaned variation relative to total variaI
P

tion,

(yi
y )2

i=1
I
P
i=1

, de-meaned variation accounts for 79.9% of total variation and

yi2

the mean payment to a district accounts for a mere 20.1% of total variation.
This shows that ARRA expenditures were not distributed equally.
A substantial majority of the variation in funds allocated across noncapital districts is due to a small number of districts that received a large
amount of funds. Out of 334 districts that dont contain state capitals, we
identify 17 outliers, using a cutoff of $1300 per resident. Of these districts,
12 are Democratic districts and 5 are Republican districts. Out of the 10
districts receiving more than $2000 per resident, 5 are Republican and 5 are
Democratic. If we take a simple model with only a dummy for a Democratic
district, adding in dummies for these 17 districts increases the R2 from 0.007
to 0.808. When we look at the distribution of funds on a per worker basis,
we find 5 Republican districts which get more than $3500 per worker and
2 Democratic districts which received more than that amount. In the per
worker regressions, a simple model with a constant term and a Democratic
dummy yields an R2 of zero to three significant digits. However, adding
7 dummies for the outlier districts increases the R2 to 76.4%. Interestingly,
adding in dummies for outlier districts does not qualitatively change the signs
and magnitudes of the other coefficients in more fully specified models. We
show these results in Appendix Figure A3. The large amounts allocated to

14

the outlier districts are due to sizable individual grants rather than a large
number of grants. The two top Republican recipients, Doc Hastings (WA)
and Buck McKeon (CA) received 22% and 42% of their funds respectively
in their top award. In both cases, these were large competitive DOE grants
or loans (water reclamation at the Hanford nuclear site and a solar energy
loan, respectively). The two top Democratic recipients, Elijah Cummings
(MD) and Barbara Lee (CA), received 29% and 11% of their funds from
their top grant respectively. In the case of Elijah Cummings district, the
funds were Maryland Department of Education funds which would usually
go to the state capital (Annapolis) but in this case went to Baltimore6 .
In the case of Barbara Lees district, the large grant was to the California
Department of Transportation for the construction of a local highway tunnel.
More generally, the top 10 recipients all received over $2000 per recipient
and in all cases, a majority of the money was contained in the top 5% of
grants. The top 4 recipients all received over $3000 per resident and more
than 80% of their funds were in the top 5% of grants. In general, as seen
in Appendix Figure A3, there is a strong positive correlation between the
amount a district received and the percent of funds received in the top 5%
of grants. This suggests that outlier districts were outliers largely because
they had a particular large project that they were awarded, often through a
competitive grant.
The money in the ARRA bill was distributed through 207 different federal funding agencies. However, the top four of these account for 55% of the
total CGL funds distributed. These four agencies are (in order): the Office of Elementary and Secondary Education ($64.6 billion), the Department
of Energy ($39.7 billion), the Federal Highway Administration ($27.7 billion), and the Office of Special Education and Rehabilitation Services ($13.6
billion). The other top-10 granting agencies were the National Institute of
6
This is the only case we found where state funds went to a district that did not contain
the state capital.

15

Health, the Department of Housing and Urban Development, the Federal

Transit Administration, the Federal Railroad Administration, the Environmental Protection Agency and the Rural Utility Service. These 10 agencies
distributed 74% of the $267 Billion in our data. The amounts for these top
ten funding agencies are listed in Appendix Table A1. The top 5 granting agencies in the data excluding the state capitals were the Department
of Energy, the Federal Highway Administration, the National Institute of
Health, the Federal Transit Administration, and the Department of Housing
and Human Services (Appendix Table A2).

4.2

States and Funding Formulas

The variation in ARRA expenditures across states is well explained by

simple formulas: largely Medicaid and highway miles. This has already been
pointed out by Inman (2010). Using state-level data on ARRA expenditures
from Federal Funds Information for States (FFIS), he shows that 55% of
state-level expenditures are explained by simple funding formulas incorporating highway miles, prior Medicaid expenditures and state budget gaps.
Sixty-seven percent of the variation in highway funding can be explained
by 7 variables including highway mileage per capita. However, Inman finds
that his simple cross-state funding formulas do not work well in explaining
either stability aid or other aid, where they explain only 6% and 10% of the
cross-state variation, respectively.
We replicate Inmans results with our data at the state level and report
it in Table 2. In particular, we regress:
ARRAs = + unemp09 s + deficit s + medicaid2005 s + statepop s +

(1)

interestatemiles2005 s + senFiscalChair s + closeObamawinner s

where unemp09 is the February 2009 state unemployment rate as reported
by the Bureau of Labor Statistics, deficit s is the 2009 state budget gap,
statepop s is the state population in July 2009 as reported by the American
16

Community Survey, interstatemiles2005 s is the number of miles on interstate

highways per 10,000 residents in 2005 using data from the 2008 Statistical
Abstract, senFiscalChair is a dummy that takes a value of 1 if a senator from
the state is the chair or ranking member of one of six budget-related Senate
committees,7 and closeObamawinner s is a dummy variable which takes on a
value of 1 if the state two-party vote share for the Democratic Party in the
2008 presidential general election was between 50 and 52%.
Our results are largely consistent with Inman (2010). Column 1 of Table
2 uses the same data source for the dependent variable (ARRA expenditures
from FFIS) as Inmans regressions.The data sources for the explanatory variables are not identical to Inmans, and so we are unable to exactly replicate
his results.8 Columns 3-6 use our data source for ARRA spending the contracts, grants and loans from Recovery.gov at the state and district level.
Column 7 also uses data from Recovery.gov, but in this case we use the more
comprehensive agency-reported data that is available at the state level; this
data set includes information for the entitlement portion of the bill, as well
as the contracts, grants and loans, for a total of $470 billion. Using the FFIS
data and Inman specifications, our regressions explain over 81% of the variation in state-level ARRA allocations.9 Estimating the same specification but
now using the CGL funds data from Recovery.gov aggregated to the state
7
The six committees are: Appropriations, Banking and Urban Affairs, Budget, Commerce and Transportation, Environment and Public Works, and Finance.
8

In particular, our results for the state budget gap do not correspond to Inmans. He
finds a positive correlation between the amount received and the state budget gap. We
also find this in some specifications, depending on how the state budget gap is defined.
However, the specification displayed in Table 2 uses the budget gap as a percent of the
general fund budget, instead of the budget per capita. We do this in order to reduce the
influence of an outlier, and we find no relationship between our measure of the budget gap
and amount received.
9

In Appendix Table A3, we show that Inmans model explains only 7% of the crossstate variation in education spending, but 67% of the cross-state variation in highway
spending.

17

level, we are still able to explain about 72% of the variation in spending.
This is shown in column 3. However, when we remove the funding going to
state capitals, and aggregate the remaining money to the state level, the explanatory power of the regression drops dramatically. The R2 falls to a value
less than 0.14. The money going to state capitals is primarily formula-based
funding for education and infrastructure. Exclusion of the funds going to
state capitals removes 60% of the total dollar amount awarded. Moreover,
as shown in columns 5 and 6, when we run the Inman regressions at the
congressional district level, we explain less than 4% of the total variation.
Since the vast majority of money sent to state capitals is actually then redistributed across all districts within a state, we conclude that simple funding
formulas do not explain the allocation of discretionary funds from the ARRA
bill across congressional districts. Finally, column 7 displays the state-level
results using the agency-reported data, which includes data for state capitals,
as well as entitlement spending. Here the explanatory power is much lower
than in column 3, suggesting that the entitlement portion of the spending is
not as well-explained by the funding formulas.
The state-level regressions show three interesting facts about the distribution of ARRA funds. First, simple funding formulas do not explain any of the
state- or district-level variation in the CGL portion of ARRA funding. Second, if anything swing states that voted narrowly in favor of Obama received
lower amounts of discretionary stimulus funding than other states. States
that narrowly voted in favor of Obama in 2008 but within a 2% vote margin received $81 less per capita in discretionary funds; omitting the money
going to state capitals, the magnitude of the differential increases to $202
less per capita, and becomes statistically significant at the 1% level. Moreover, looking at Appendix Figure A1, we see that swing states in general
including those voting narrowly for McCain as well as those voting narrowly
for Obama did not receive significantly more funds than other states. The
third interesting finding is that a small number of variables fit cross-state

18

expenditure patterns quite well for total expenditures. However, the same
models fit quite poorly when explaining the money that is spent outside of
state capitals, where the awards largely exclude the formula-based allocations
for education and infrastructure. Since this is true even at the state level
where spatial measurement error is unlikely to to be large, it suggests that
measurement error is probably not the main reason for the low explanatory
power of political models in explaining the cross-district patterns of ARRA
expenditure.

4.3

Economic Targeting

Given the stated countercyclical objective of the legislation, we now consider whether financial need or high excess capacity was predictive of how
much CGL funding a state received. In Table 3, we show coefficients for the
change in the unemployment rate between January 2009 and January 2007,
We do this using the agency data as well as with the CGL data and we show
specifications with and without controls using the agency data. In the specifications with controls, we control for state employment and the poverty rate.
The specification is rather sparse because of the limited degrees of freedom
with only 50 observations. This obviously also limits our ability to interpret
our results.
The coefficient on change in the unemployment rate is relatively small
and the sign of the coefficient flips when we add controls. This is surprising
since some of the state-level money was explicitly set aside for high unemployment states. For example, 35% of the $87 billion allocated for Medicaid
was set aside for high unemployment states. We also look to see whether the
amounts given across states varied systematically with employment levels in
the states. Certain states such as New York have a high level of employment
per resident and thus might be expected to receive larger amounts of funds
to stimulate employment. However, we find that amount received per resident and state-level employment is negatively correlated in all specifications
though the coefficients are far from statistically significant. Finally, because
19

unemployment has higher incidences in poorer areas, we look at how statelevel poverty rates correlate with overall funds in a state and again find a
negative relationship. States with higher poverty levels on average received
less funding. An increase in the poverty rate by 1 percent is correlated with
a decline in funding of $46 per resident in our baseline specification with
controls. The poverty coefficients are significant at a 10% level when we use
the agency data. The coefficient on poverty is roughly 1/3 the size and not
statistically significant in the contracts, grants and loans data.
Turning to the district level, in Appendix Figure A5, we non-parametrically
regress the amount received per resident on our two measures of excess capacity. In both cases, we see that districts with higher unemployment received
slightly less funding per district resident. These results reinforce similar
findings by Gimpel et al. (2013) and Inman (2010). In the district-level
regressions in Table 4, we report coefficients on 7 different covariates for 5
different specfications, reflecting results from 35 different regressions. First
of all, in the first three panels (rows), we show conditional correlations of
amount of CGL funds per district resident with our covariates. In the bottom two panels, we report conditional correlations of amount of CGL funds
per worker in the district. We look at per worker specifications because we
do find that, in contrast to the cross-state variation, more money did go to
highly urbanized areas. This is not surprising given that much of the formula
money given to the states stipulated that a certain percentage of the funds be
spent in urban areas. Most people who work in Wyomings congressional district also live in it. However, a much lower percentage of people who work in
the congressional districts in Manhattan live there. Moreover, neighboring
Brooklyn and Queens are much more places of consumption than production. This, in part, explains the low amounts of funds received by Anthony
Weiners district. Interestingly, the amount of funds received per resident is
much more strongly correlated with employment in a district than with the
percent of the district that is urban. In panels A and D we show per resident

20

and per worker results for covariates without other controls. In panel B, we
control for log 2008 employment, the poverty rate, highway miles and percent
urban as well as 9 vote share bin dummies for the vote share of the House
Represenative in the district. Our results adding in all these covariates are
generally similar to the unconditional results. In panel C, we also control
for land area. In some specifications, this leads to different results. All of
these covariates are covariates which were explicitly or effectively incorporated into funding formulas. Finally, in panel E, we report the per worker
variant of panel C; panel E differs only in that we omit log employment since
we already incorporate employment in our dependent variable.
We find no statistically significant correlation in any of the specifications
between amount per capita and either of our measures of unemployment or
the amount of UI spent in the district per capita. We also find no statistically significant correlation with percent urban in any of the specifications.
We also find very strong positive correlations between employment in a district and amounts per resident. The estimates are very tightly estimated.
Districts with 100,000 more workers on average get $200 more per resident.
In all three per-resident specifications, the t-statistics are above 5 and the
point estimates vary by less than 0.06 across specifications. Moreover, employment is not only highly correlated with funding, it is quantitatively important. Adding just employment increases the R2 by over 13 percentage
points. Also, adding in other controls does not change the marginal R2 of
adding in employment. The estimates are surprisingly invariant to controlling for percent urban. Unsurprisingly, employment is not correlated with
amount received per worker in the district.
The second most important of our covariates is poverty. Poor districts
received more money. Adding in poverty increases the R2 by approximately
0.03. Again, this is not surprising given that some of the formula money
set aside portions specifically for historically poor areas.10 The results on
10

For example, some of the money disbursed by the Department of Labor used definitions

21

poverty are less robust than those on employment. Adding in our other
controls lowers the coefficient estimates slightly but also almost doubles the
standard errors. Unconditionally, places with a one percentage point higher
poverty rate received $16 more in funds per resident and $42 more per worker.
Note that since around 1/3 of the residents in a district work,11 the per worker
coefficients are usually approximately 3 times as high. Unconditionally, both
coefficients are significant at the 1% level. With the full set of controls, the
t-statistics in the per resident specification are just below significance at the
10% level with a t-statistic of 1.57 and the per worker coefficient is significant
at between the 10% and 5% level with a t-statistic of 1.87.
It is possible that the bill didnt target unemployment because it targeted shovel ready projects and places with more shovel ready projects were
places with lower unemployment (or lower increases in unemployment). The
Obama administration said that shovel-ready projects would be made high
priority and this was reflected in the bill. Much of the formula grant money
had stipulations that money would have to be returned if not spent quickly
enough. While it is difficult to measure shovel readiness using an ex-ante
measure, we do have recipient reported information on the pace at which
the funds were disbursed and spent. To assess shovel readiness, we construct
measures of what fraction of the funds in a district were in projects that
were completed within one year. Using this measure, we show in Online
Figure A7 that places which were allocated more money were on average
somewhat slower in completing projects. This is possibly because districts
which received more money received money for large infrastructure projects
and these on average took longer. However, Online Figure A7 shows that
uniformly places that received more money never completed their projects
more quickly. Across all specifications, the coefficient on percent completed
within one year is negative and significant at the 10% level or less. The specof poor areas defined in a bill in 1965.
11
Note that this ratio is different from the employment to population ratio as standardly
defined in that it includes the elderly and children.

22

ifications with controls are more significant and the per worker coefficients
are significant at a 1% level. Interestingly, the negative results on completion
within one year contrast with null results on percent of funds spent within 2
years despite the fact that there is slightly more variation in the percent completed within two years variable. Moreover, in reported regressions, we note
that controlling for percent of funds spent within one year does not impact
the coefficients on unemployment. Though shovel readiness may have played
a role in the selection of projects, we find no evidence that it influenced the
how money was allocated across congressional districts.

4.4

Group Targeting: Partisanship

We next investigate whether members of Congress acted in the interests

of their party. At the time of passage of the ARRA, Democrats had a strong
majority in both the House and the Senate and they also held the presidency.
Therefore, they were able to pass the legislation without any support from
Republicans in the House, and they passed the bill in the Senate, overcoming
a potential Republican fillibuster, with the help of only three Republicans.
Did, then, the Democrats get a large majority of the funds, as might be
expected from simple and standard political economy models?
As De Rugy (2010) has shown, districts represented by Democrats received substantially larger sums of ARRA money than those represented
by Republicans. In our database of contracts, grants and loans, districts
represented by Democrats received 55% more than those represented by Republicans. The mean Republican district received $684 in funds whereas
the mean Democratic district received $1,057. However, as pointed out by
Nate Silver (2010), the state capitals received funds that, while nominally
allocated to the capital, were in turn allocated across the state.12 Education
funds are usually allocated in this manner. The top 17 recipient districts are
all part of state capitals, and 26 of the top 30 are state capital districts as
12

Albouy (2009) also points out the state capitals problem as part of his justification
for running state-level regressions.

23

well. The probability of either of these events occuring by chance is below

1 in 1013 . However, the pattern of funds is slightly more nuanced than Silvers account. He argues that state capitals tend to be heavily Democratic
and therefore these transfers to state agencies are more likely to be counted
as going to Democratic districts. In fact, on average, state capital disticts
have almost the same Democratic vote share as districts not located in state
capitals. In Table 1, we show that the Democratic vote share for the House
seat is 58% whether we include the 101 out of 435 districts which contain
portions of a state capital or not. 13 While urban areas tend to vote more
Democratic, there are many districts containing portions of state capitals
and either surrounding suburban land or even surrounding rural land (e.g.,
Wyoming is itself one district). Republican districts containing state capitals
are slightly larger and less dense. More importantly, state capital districts
in larger states do tend to be more Democratic than average, while capital
districts in smaller states tend to be more Republican. The state capitals
in the largest states received disproportionately larger sums because most of
the education and Medicaid spending for the entire state is often given to
the capital, and states with large populations received more total education
and Medicaid funds. The districts of the top ten recipients are, in order,
the capitals of California, New York, Illinois, Florida, Texas, Ohio, Michigan, Georgia, Pennsylvania, and Massachussetts. Only two of the top 17
districts, and four of the top 26 districts were Republican.
Since we do not know how the funds nominally allocated to state capitals
were actually dispersed within the states, we exclude all districts that include
state capitals from our analysis. This eliminates 13 states which do not have
a congressional district without some part of a state capital from our sample.
In our revised sample, the partisan gap is substantially lower, with Democratic districts receiving 23% more than Republican districts. The average
13

Our set of capital districts differs slightly from Silvers. We identifiy 101 congressional
districts containing some portion of a state capital city or its surrounding county. Silver
defines 78 districts as containing all or part of a states capital city.

24

Republican district receives $416 per capita as compared to $510 per capita
in the average Democratic district. While this $95 differential is statistically
significant (see column 2 of Table 5), it becomes $19 and statistically insignificant once percent living in poverty is introduced as a control. We then
add our controls. None of the controls change the coefficient on the Democrat dummy by much with the exception of the poverty variable. Column
5 shows that conditional on the four other funding formula controls, including employment, the coefficient rises slightly to $109 and remains significant
at a 5% level. However, introducing poverty reduces the coefficient back to
$33 and makes it insignificant. In the per worker regressions, reported in
columns 7 and 8, the Democrats dummy is insignificant and small with or
without controls. In fact, conditional upon controls, the coefficient on the
Democrat dummy is -$85. Ultimately, it is not clear whether Democrats funneled money to their districts through funding formulas that targeted high
poverty areas and high employment areas or whether these areas got more
because they were trying to create jobs in poor areas and centers of employment. Nonetheless, it seems unlikely that poor areas benefited from explicit
targeting of Democratic areas. In fact, the poverty is slighlty more strongly
correlated with funds per resident in Republican areas than Democratic ones.
We now look more closely at how Representatives did as a function of the
two-party Democratic vote share. Figure 2a plots the CGL funds per resident
against the Democratic vote share. Five Republican outliers who received
substantially larger amounts stand out and roughly triple the number of
Democrats also stand out. Figure 2c shows that the Democratic outliers
are all in districts which are 100% or very close to 100% urban. Thus, it
is not surprising that in Figure 2b, we see only two Democratic outliers
though we still see the same five Republican outliers. In Figure 2d, we plot
worker-per-resident ratio nonparametrically on vote share. Interestingly, we
find a similar hump for districts with 80-90 percent Democratic vote share.
This suggests that much of the hump in funding for Democrats in the 80-90

25

percent range is attributable to those districts being centers of employment.

In all of the scatterplots in Figure 2, we can see that Democrats receiving
between 80 and 90 percent of the vote share seem to have received a higher
amount of funds per capita. To assess this further, we non-parametrically
regress per-capita amount of CGL funds on the Democratic Party vote share
and report the results in Figure 3a.14 We also use a (semi-parametric) partial
linear model to regress the per-capita CGL funds on the Democratic Party
vote share, controlling parametrically for the five funding formula controls:
percent living in poverty, percent living in an urban area, land area, road
miles, and employment.We estimate:15
Ai = f (vi ) + X + t
where Ai the the per capita amount of ARRA funding received in district
i, X is the set of demographic and economic controls and f (vi ) is a nonparametric function of the two-party Democratic vote share.16 The results
are shown in Figure 3b.
Figures 3a confirms the evidence from the simple scatterplot that strongly
Democratic districts with around 80 percent of the vote share receive well
above the mean amount of CGL funds. Figure 3b shows that this relationship
continues to hold after we account for covariates. Figure 3a also suggests
that highly Republican districts get a modest amount more than the average
14

We use the STATA command lpoly, using a rule of thumb plug-in bandwidth that
minimizes the conditional weighted mean integrated square error. Standard errors are
obtained by bootstrapping with 100 replications.
15
Due to concerns that very urban districts got substantially more money and are also
highly Democratic, we ran specifications controlling for a dummy variable which takes on
a value of one for districts with 100% of their area in urban land. We also tried controlling
for a dummy which takes on a value of one if a district has 90% or more of its land in
urban areas. Our results are highly robust to these alternative specifications.
16
We use the Yatchew method to difference out the parametric component X, and use
local polynomial regression (STATA command lpoly) to estimate the f (vi ) component
nonparametrically; the bandwidth selection is based on the rule of thumb plugin method.
Standard errors are obtained by bootstrapping with 100 replications.

26

recipient. However, this difference is not statistically significant and does

not survive the inclusion of covariates in Figure 3b. Figure 4 shows the
same results with amount per worker as the left hand side variable. In
Figure 4a, we see that, whereas Democrats in the 80-90 percent vote share
range do not get as much more as in the per resident figure, they still do get
more than Democrats in less safe districts. However, close Republican highly
Republican districts also now get substantial amounts. The bootstrapped
confidence intervals are much wider for the very Republican districts because
they are driven by one outlier who received a substantial amount. Putting in
our five controls also increases the amount received per capita for marginal
Republican districts though not for marginal Democratic ones.
Interestingly, when we compare close Democrats to close Republicans,
we find that these districts receive about the same amount (if anything the
marginal Democratic districts obtained less CGL funds). In Table 5, the
point estimates for Republican leaning districts with 40-50% (Democratic)
vote share are positive, but generally not statistically signficant. Democratic
and Republican districts with around 50% Democratic vote share are likely
to be similar in terms of other characteristics, so comparing these districts
offers another way to test whether the partisan representation of the district
affects the spending allocation. These results lend support to the argument
that the higher average level of spending in Democratic districts is driven
more by district characteristics than by party affiliation per se. However, the
substantially larger amount of CGL funds going to districts where Democrats
received between 80 and 90 percent of vote share suggests that there may be
a partisan gap in funds, or some other factor that may be different in these
districts.

4.5

Individual Targeting: Party Elites

Did legislators use their individual positions of power to their own benefit?
We break down per capita ARRA funds by whether members had leadership
positions in Congress, whether they were legislative swing voters and how
27

long they had been members in Congress. Our findings are predominantly
negative.
Congressional leaders have influence both because they have agenda setting powers as well as because they are socially well-connected and potentially
persuasive allies. From column 4 of Table 5, we can see that the Democratic
Speaker of the House, Nancy Pelosi, received about $250 more per capita
than the average House member, though this difference is not statistically
significant when we include our full set of controls. John Boehner, the Minority Leader, by contrast, received approximately $114 less per capita. Out
of our sample of 334 representatives who do not represent any portion of
state capitals, there are 20 Democratic committee chairs and 19 Republican
ranking members. The Democratic committee chairs received on average
$133 less than similar members. However, the Republican ranking members
received almost $375 more on average, and this is significant at a 10% level
of confidence. However, the results for the Republican leadership, as can be
seen in Figure 2c, are driven by two strong outliers amongst the Republican
ranking members.
As noted earlier, the top 2 recipients out of these 334 representatives were
both Republican ranking members. They were Doc Hastings of Spokane,
Washington, ranking member of the House Committee on Natural Resources,
and Buck McKeon of Santa Clarita, California, ranking member of the House
Armed Services Committee. The probability that 2 or more out 19 Republican Ranking Members would be in the top 1% of recipients by random chance
would happen with a probability of 1.5%. However, besides these two top
recipients, Republican ranking members and party leaders did not receive
more than average members of Congress. Ten out the nineteen Republican
ranking members in our sample receive above the median amount of funding.
The Minority Leader, John Boehner, was in the bottom 11% of recipients.
The top Democratic committee chairs did not receive more CGL funds
than average. Out of 20 Democratic committee chairs in our sample, only 2

28

were in the top 10% of recipients. The top ARRA fund recipient in the Democratic Party Leadership went to the Speaker of the House, Nancy Pelosi, who
represented San Franciso, California. She received the 11th largest amount of
funds of the 334 representatives in our sample. The second largest recipient
in the Democratic leadership was Patrick Murphy of Pennsylvanias District
8, which serves Levittown and surrounding areas. He received the 28th highest amount. Similar to Republican leaders, the median Democratic leader
did not receive a significantly different amount from the median Representative. Eleven out of the 20 Democratic committee chairs got more than the
median amount received in our sample. The fact that Democratic leaders
received significantly less on average than the average congressperson is due
to the fact that 6 out of the 20 leaders received below the 25th percentile of
funds and three, two representatives from New York and one from California,
were in the bottom 4%. All three of these representatives received less than
$100 per capita for their districts.
In Table 7, we regress amount per resident and per worker on dummies
for the Democratic leadership and for the Republican leadership separately.
We do this with and without our baseline controls. We show that Democratic
Congressional leaders on average get the same or less than other members
of Congress. The coefficient on the Democratic leadership is not statistically significant in any of the specifications. The coefficient is zero in the
per resident specification without controls and negative in all others. In Appendix Table A6, we report the dummies separating out the Speaker of the
House (Pelosi) on the Democratic side and the Minority Leader (Boehner)
on the Republican side. Pelosi is an outlier and the speaker dummy is statistically significant in at the one percent level in all specifications except
the per worker specification with controls. In the per resident specification
with controls, she gets almost $600 more per resident than other similar
members. Conditional on the speaker dummy, the rest of the Democratic
leadership gets significantly (at the 5% level) less controling for funding for-

29

mula amounts in both per worker and per resident specifications. They get
$126 less in the per resident specification. On the other hand, returning
to Table 7, we see that in all specifications, the Republican leadership gets
more than average. With controls, the amounts are statistically significant
at the 10% level. However, the minority leader on the Republican side is
also an outlier. Unconditionally, his district gets $338 less per resident and
$150 less conditional upon controls. Despite being at the 11th percentile in
the distribution of funds, the Boehner results are significant themselves at
the 1% level without controls, suggesting that these estimates are plagued by
small sample sizes and skewed data. Accounting for Boehner, the rest of the
Republican leadership gets $409 more on average conditional upon controls
and this is significant at the 10% level. However, again, these results are
driven by two outliers, Buck Mckeon and Doc Hastings, who on average, get
$3325 more funds per resident. Taking out these two outliers, we do not find
evidence that the rest of the Republican leadership received more than other
Congressional members.
Because of the small numbers of party leaders, we also calculate small
sample tests to assess whether these individuals received a disproportionate
amount of CGL funds. Intuitively, by comparing those in power during the
ARRA bill to those in power in the following and prior Congresses, we estimate the impact of being in power precisely when the ARRA bill passed. We
matched the districts of current congressional leaders in the 111th Congress
to the districts of their predecessors who held the same positions in the 110th
Congress. There were 6 Republican leaders who we could match due to a
number of exits of Republicans in the leadership of the 110th Congress. In
5 out of 6 cases, the district with a leader in power during the ARRA vote
received more money. This would happen by random chance 4% (one-sided
test) of the time. We also look at leadership changes between the 111th and
112th Congress. For Republicans, this raises our sample size to 10. We find
that in 7 out of 10 cases, the district with a leader in power during the ARRA

30

vote received more, which would happen by random chance with a probability of 8.9%. We run the same comparison for Democratic leaders as well. In
the case of Democrats, there is no evidence that committee chairs received
more money17 ; in only 5 out of 11 cases did the leader in power during the
ARRA vote get more. These results are reported in Appendix Table A5.
Legislative tenure can be thought of as a proxy for social connections or
institutional knowledge of the functioning of Congress. For this reason, it is
possible that higher tenure legislators are able to procure a greater amount
of funds. Rows 5 and 7 of Table 7 show that tenure is positively correlated
with amount received per resident and per worker. Both are significant at
the 10% level. However, Figure 5b shows that many of the higher tenure
(shaded more heavily) observations are in the 80-90 percent range. Conditional upon baseline controls, coefficients drop by a half to two third. An
additional Congressional session is associated with $2.5 more in funds per
resident and is not statistically significant at conventional levels.We show,
in Appendix Figure A7, non-parametric and semi-parametric regressions of
amount received per capita on tenure. We also break them down by party.
For Republicans, we can see that very high tenure individuals actually receive less and for Democrats, results, increases in funding with tenure are
solely due to a few outliers with more than 50 years in the House18 . Overall,
we see very little evidence that powerful individuals targeted ARRA funds
to their own districts.
17

Our negative results on the importance of committee chairs is in contrast to recent

work by Cohen et al. (2011) who use committee chairs in a district as an instrument
for the impact of public spending. Berry et al. (2010) find the opposite. Our own work
suggests that these estimates may suffer from small sample sizes and non-normal errors.
18
Feyrer and Sacerdote (2011) use average tenure of a Congressional delegation in a
state as an instrument for amount of ARRA money. Our results suggest that tenure is
not an effective instrument at the district level and that high tenure districts are different
from low tenure districts in ways other than just the tenure of the Congressional members.

31

4.6

Individual Targeting: Party Discipline and Pivotal

Members of Congress

Another measure of whether legislators act in the interests of their party

is whether they punish party members who deviate from the party line. For
the ARRA bill, there were a few legislators who voted against the majority
of their party: Republicans who abstained from voting in the House or even
voted for it in the Senate, and Democrats who abstained or voted against
the ARRA bill. No House Republicans voted for the bill; however, two did
not vote. Additionally, seven Democrats voted against the bill, one voted
present and one did not vote. Party defectors did get less on average than
those voting with their parties both in the Democratic and the Republican
party. Coefficients are not significant at conventional levels and amount to
-$86 pooled across parties or -$87 for Democrat defiers. These results are
shown in columns 5 and 6 of Table 8. Putting in baseline controls increases
the magnitudes of the estimates. With the full set of baseline controls, defiers
do get less and results are statistically significant in per worker or per resident
specifications at the 5% level or less. For Democrats, the amounts are -$144.
However, looking at Figure 5D, we see that with the exception of one outlier,
Democratic defiers get exactly the expected amount conditional upon vote
share. Dan Lipinski, a Democrat who represents the 3rd congressional district
in the suburbs of Chicago, was a strong outlier, and his district recieved only
$75 per resident. Lipinski voted present on the bill, and his district received
the 17th lowest amount per capita in Congress. In a median regression, which
reduces the influence of outliers, the coefficient on voting against party is -$49
per capita and is not statistically significant at conventional levels (results
not shown).
Legislators can also use their individual power to gain political rents is
by claiming to be undecided. Legislators in the middle of the ideological
distribution can more credibly claim to be undecided and thus need persuading in terms of rents for their district. To assess this possibility, we look at
32

how Congress member ideology correlates with per capita receipts of ARRA
funding. We use the first dimension of McCarty, Poole and Rosenthals
DW-nominate, which is the most commonly used measure of congressional
ideology. In the modern era, more than 85% of legislative voting behavior
can be explained by the first dimension alone.
We are interested in how ideology correlates with funding for two reasons. First, we want to make sure that returns to vote share are not simply
proxying returns to ideology. Second, theories of congressional politics suggest that moderate politicians may be able to capture larger amounts of
rents because they are swing voters on bills and thus able to demand compensation in order to vote in favor of a bill. In Appendix Figure A8, we
semi-parametrically regress ARRA receipts per capita at the district level
on DW-nominate, controlling for baseline funding formula controls in both
panels A and B but including linear vote share controls in panel B. Liberal
members do get more funds. This is also seen in the first column of Table
8. However, vote share controls reduce and take away statistical significance
of DW-nominate. However, the ARRA bill passed by a large margin in the
House of Representatives and so noone was pivotal. By contrast, in the Senate, the administration had to negotiate to get three Republicans (Susan
Collins of Maine, Olympia Snowe of Maine, and Arlen Spector of Pennsylvania) to break a potential fillibuster. Moreover, one conservative Democrat,
Ben Nelson of Nebraska, publicly aired his ambivalence towards voting in favor of the bill. In column 4 of Table 8, we put in a dummy for Pennsylvania
and Nebraska. Unfortunately, we cannot include the two Maine districts in
our analysis because both contain part of the state capital. Conditional upon
controls, we do find that pivotal districts got $120 more and the results are
significant at the 5% level. However, in all other specifications, the results
are not significant and in the per worker specification, they are negative.
More important, in our state level analysis in Table 3, the estimates range
from -$265 to -$365 across CGL and agency data sets and all specifications

33

are significant at the 5% level. We take this as evidence that pivotal states
did not get more money. If anything, they got less. Adding in Maine does
not qualitiatively change these results except in the agency reported data
without controls. This is show in column 7 of Table 3.
Finally, we also look at whether Governors of Democratic states received
more funds. The party of the Governor is used as an instrument for fiscal
expenditure in Conley and Dupor (2013). Similar to Conley and Dupor,
we find that using CGL data at the district level, a positive correlation
between a Congressional District lying in a state with a Democratic Governor
and the amount that the district receives per capita. However, in our four
specifications, only one is statistically significant and at only the 10% level.
Moreover, however, in Table 3, we see that, in the agency data, the sign
of the coefficient on Democratic Governor depends upon whether controls
are included. Using the CGL data, the sign is negative and statistically
significant at the 10% level. We interpret our findings as showing that neither
Democratic nor Republican Governors got systematically more ARRA funds
per capita.

Interpretations and implications for future

policy

To summarize our findings, we find little to no evidence of positive targeting of funds towards powerful individuals. We also find limited targeting
in favor of the party that held the majority in both chambers of Congress
and held the Presidency. Additionally, we find no targeting of funds towards
districts that had large increases in unemployment. Nonetheless, we do find
that there was substantial variation across electoral districts in the allocation
of funds.
At a certain level, our finding that party leaders, swing districts, and
ideological moderates did not recieve more funding, and that defiers did not
suffer a clear penalty, may not be surprising to close political observers.
34

This is especially so given that the Obama administration had mandated

that contracts, grants and loans be allocated based on either formulas or
competitive grants. Funding formulas are intentionally coarse instruments
for the purposes of limiting politically motivated geographical targeting. At
the same time, since formula money is allocated by state governors, it is
somewhat possible to target particular districts in small statesfor example,
single district states. Similarly, while the competitive bidding process limits
the ability to target specific areas, the choice of projects still allows a degree of
discretion and thus influence. Moreover, while targeting individual districts
would have been difficult, the bill could have easily used funding formulas
and even competitive grants to tilt money strongly towards more Democratic
states. The spatial outcome for the ARRA bill stands in stark contrast to
findings about the other large fiscal stimulus in U.S. history: the New Deal.
Fishback, Kantor and Wallis (2003) document that although loans were not
politically targeted or targeted to high unemployment areas in the New Deal
Programs, grants were targeted both to high unempoyment areas and to
swing-district supporters of Roosevelt in the prior presidential campaign. In
this section, we discuss some possible explanations for the limited geographic
targeting in the ARRA bill, and why this may have been different during the
New Deal era.
One possible explanation for our results is that the spatial allocation
reflected the desires of the administration and Congress. As its name suggests, the bill had two componentsrecovery and reinvestment. While some
parts of the bill (Unemployment Insurance, Supplemental Nutrition Assistance Program, Medicaid) were focused on recovery (or at least relief), the
CGL component was more focused on reinvestment. Certainly automatic
stabilizers, for example, are well targeted to those with a high propensity to
consume out of income and those in (or at risk of of falling into) poverty.
Maybe, then, the spatial allocation of CGL funds reflected the variation in
local needs for public goods rather than jobs.

35

However, even though other parts of the bill were better targeted, we find
it unlikely that the allocation of CGL funds was what the administration
would have preferred in the absence of political constraints. In January of
2009, the unemployment rate in California was 9.7% and rising rapidly; the
unemployment rate in Michigan was 11.4%; however, the unemployment rate
in North Dakota was 3.9%. Many schools and highways needed to be rebuilt,
and it would not have been particularly difficult to use the severity of the
business cycle in an area as a factor in allocating funds. We think a more
plausible argument is that there are certain tradeoffs that arise when greater
targeting in fiscal policies is allowed, and the administrations relatively nontargeted approach reflected constraints and not preferences.
There are a number of costs that may arise from finer grained geographic
targeting. First, there may be increased delays in project selection due to
bargaining between legislators, something which runs counter to a rapid enactment in face of a major economic crisis. For example, when Congress
tried to set aside Medicaid money for high unemployment areas, the Senate opted for a lower percentage to be allocated towards high unemployment areas and this caused delay in bargaining in the conference committee.
Second, such discretion may lead to precisely the type of political opportunism as both predicted by theories of legislative politics and at least partly
confirmed by the New Deal experience. Such opportunism, however, may
have deligitimized a piece of legislation that was quite controversial and was
under substantial scrutiny. For these reasons political constraintsreal or
perceivedmay have prevented the administration from engaging in a more
targeted approach, instead favoring the use of funding formulas and competitive bidding.
In a majoritarian system where politicians gain re-election in part through
providing funds to their districts, funding formulas limit the degree to which
politicians can steer funds to their own districts. Starting with the FederalAid Highway Act of 1916, Congress increasingly used formulas to allocate fed-

36

eral funds. This was done in large part to reduce graft by allocating across
states according to established empirically-verifiable criteria that reflected
need. Moreover, using formulas to allocate funds streamlined bargaining by
allowing Congress to decide on the level of funding and then delegate the dispersion of funds. It allows for Congress to retain control over what types of
projects are funded while bypassing bargaining over the spatial distribution
of the funding in which politicians may be particularly interested. Over time,
the use of formulas to allocate funds has become more prevalent. Given that
the Obama administration advocated a quick passage and expressed an interest in limiting pork, a reliance on funding formulas and competitive grants
may have been the means of acheiving those goals. In addition to allowing for
timely passage of the legislation, the channeling of money through existing
programs also provided a method to ensure that funds were spent quickly,
which was a priority for this particular piece of legislation.
However, while a relatively non-targeted approach has merits in terms of
reducing the possibility of graft, and possibly expediting the pace of legislation, it comes with costs as well. In particular, it prevents an ability to target
based on economic as well as political considerations. It is noteworthy that
while the fiscal stimulus in the New Deal may have provided greater award
to swing districts, it also channeled more money to higher unemployment
areas. The good news is that political constraints appear to have made fiscal
policy more politically neutral; the bad news is those constraints may have
also reduced the countercyclical efficacy.
This has a number of implicaitons for future stimulus bills. First, this
means that the components of the policy which are better targetede.g., automatic stabilizersare quite important and it may be useful to put greater
weight on these. At the same time, besides automatic stabilizers, there are
economically sensible reasons to increase public goods provision as part of
a countercyclical policy, as discussed in Delong and Summers (2013). The
relevant hurdle rate for such projects effectively falls, and an optimal fiscal

37

policy should likely include an expansion in the funding of such projects.

However, it should be possible to include local area unemployment rates,
or other transparent measures of excess capacity, as a factor in the funding
allocation for contracts, grants and loans. Such an approach would combine
the virtues of a more rules-based policy regime with some of the gains from
a more targeted variant of stimulus policy. If we had found that the political
process generated sufficient demand for added funding in harder hit areas,
the need for such an explicit rule would not arise. However, that is not what
we found, suggesting that the inclusion of local excess capacity measures in
funding formulas is likely to have a substantial payoff. Moreover, we note
that the government relied on already existing funding formulas to disperse
the funds, which means that the time for updating the formulasin order
to better accomodate the objectives of countercyclical fiscal policyis now,
before the next crisis hits.

Conclusion

In this paper, we have looked at the spatial distribution of ARRA funds to

assess how it was targeted economically and politically, and use it as a window
towards evaluating theories of political economy. Our main finding is that
much of the variation across congressional districts in discretionary spending
remains unexplained. Though simple formula-based theories do quite well
at the state level, this is almost entirely due to the allocation of Medicaid,
education and highway money based upon simple funding formulas. Once
state capitals are excluded, it is difficult to find systematic and significant
explanatory variables capable of explaining the variation in spending.
Parties did not appear to act in their own collective interest by targeting marginal districts. For the most part, politicians did not exploit their
individual positions of power within the Congress to grab funds. At the
same time, discretionary funds were not well targeted to high unemployment
districts, and differences in shovel readiness as measured by the pace of

38

spending do not appear to explain the cross-district variation in expenditures. We do find, however, that more funding went to districts with more
employment per resident, which likely reflects greater economic activity. Additionally, districts with higher incidence of poverty also received somewhat
greater funding. Overall, even in our most saturated regressions models, the
percentage of the unexplained variation remains above 70%. This suggests
that factors other than cyclical targeting and political power explains the
choice of projects. For example, we cannot rule out that some of the variation that we fail to explain is accounted for by highly stimulative projects
in lower unemployment areas, projects with greater long-term benefits (such
as school construction or weatherization) or projects with particular interest to the administration (renewable energy or high speed rail projects, for
example).
A comparison of the targeting of discretionary funds under ARRA with
New Deal legislation is also illustrative. Grants under the New Deal appear to
have gone both to higher unemployment areas, as well as swing district supporters of President Roosevelt. The reduced political and cyclical targeting
in the ARRA may reflect a more rules-based environment today compared to
the 1930s. At the same time, future stimulus policies can increase countercyclical targeting by explicitly using measures of unemployment and excess
capacity as part of the allocation of contracts, grants, and loans.
We see our findings as a call for continued empirical research on the distribution of funds. What explains the targeting of funds? Is the surprising
lack of explanatory power of measures of party or individual legislator influence due to the presence of repeated interactions in Congress? Is it different
from in the past (Levitt and Snyder, 1995), perhaps due to greater public
scrutiny in the presence of modern media? These are questions that empirical
political economists will hopefully address.

39

References
[1] Albouy, David (2013), Partisan Representation in Congress and the
Geographic Distribution of Federal Funds, Review of Economics and
Statistics 95(1), pp. 127-141.
[2] Baron, David P. (1991), Majoritarian Incentives, Pork Barrel Programs, and Procedural Control, American Journal of Political Science
35, pp. 57-90.
[3] Berry, C. R., B. C. Burden, and W. G. Howell (2010), The President
and the Distribution of Federal Spending, American Political Science
Review 104, pp. 783799.
[4] Cohen, Lauren, Joshua Coval, and Christopher Malloy (2011), Do
Powerful Politicians Cause Corporate Downsizing, Journal of Political Economy 119(6), pp. 1015-1060.
[5] Cox, Gary W., Thad Kousser, and Matthew D. McCubbins (2010),
Party Power or Preferences? Quasi-Experimental Evidence from American State Legislatures, The Journal of Politics 72(2), pp. 799-811.
[6] Conley, Tim and Bill Dupor (2013), The American Recovery and Reinvestment Act: Solely a Government Jobs Program?, Journal of Monetary Economics 60 (5), pp. 535-549.
[7] Dahlberg M. and E. Johansson (2002), On the Vote-Purchasing Behavior of Incumbent Governments, American Political Science Review 96
(1), pp. 27-40.
[8] De Rugy, Veronique (2010), Stimulus Facts - Period 2, Mercatus
Working Paper.

40

[9] DeLong, J. Bradford and Lawrence H. Summers (2012), Fiscal Policy in

a Depressed Economy, Brookings Papers on Economic Activity (2012),
pp. 233-297.
[10] Dube, Arindrajit, Ethan Kaplan and Ben Zipperer (2014), Heterogeneous Effects of Fiscal Expenditure: Local Evidence from the ARRA
Bill, Working Paper.
[11] Feyrer, Jim and Bruce Sacerdote (2011), Did The Stimulus Stimulate?
Real Time Estimates of the Effects of the American Recovery and Reinvestment Act, NBER Working Paper 16759.
[12] Fishback, Price, Shawn Kantor and John Wallis (2003), Can the New
Deals three Rs be Rehabilitated? A Program-by-Program, County-byCounty Analysis, Explorations in Economic History 40, pp. 278-307.
[13] Gimpel, James, Frances Lee, and Rebecca Thorpe (2013), Geographic
Distribution of the Federal Stimulus of 2009, Political Science Quarterly 127:4, pp. 567-595.
[14] Grunwald, Michael (2012), The New New Deal: The Hidden Story
of Change in the Obama Era, Simon and Schuster.
[15] Inman, Robert (2010), States in Fiscal Distress, NBER Working Paper 16086.
[16] Johnson, David S., Jonathan A. Parker, and Nicholas S. Souleles (2006),
Household Expenditure and the Income Tax Rebates of 2001. American Economic Review 96(5): pp. 1589-1610.
[17] Knight, Brian (2005), Estimating the Value of Proposal Power, American Economic Review 95: pp. 16391652.

41

[18] Larcinese, Valention, Leonzio Rizzo, and Celia Testa (2006), Allocating
the U.S. Federal Budget to the States: The Impact of the President,
Journal of Politics 68 (2): pp.447-456.
[19] Levitt, Steven D. and Snyder, James M. (1995), Political Parties and
the Distribution of Federal Outlays, American Journal of Political Science 39, pp. 958-980.
[20] Lindbeck, Assar and Jurgen Weibull (1987),Balanced Budget Redistribution and the Outcome of Political Competition, Public Choice 52:
27397.
[21] McCarty, Nolan, Keith T. Poole, and Howard Rosenthal, (2001) The
Hunt for Party Discipline in Congress, American Political Science Review 95: 67387.
[22] Menchaca, Marcos (2012), Swing Voter Theory in American Presidential Elections, working paper.
[23] Nakamura, Emi and Jon Steinsson (forthcoming), Fiscal Stimulus in
a Monetary Union: Evidence from U.S. Regions, American Economic
Review.
[24] Serrato-Suarez, Juan Carlos and Philippe Wingender (2011), Estimating Local Fiscal Multipliers, working paper.
[25] Silver, Nate (2010), http://www.fivethirtyeight.com/2010/04/studyclaiming-link-between-stimulus.html
[26] Shor, Boris, (2006a) Presidential Power and Distributive Politics: Federal Expenditures in the 50 States, 19832001, Working Paper.
[27] Shor, Boris (2006b), The Weak Leviathan? Testing Partisan Theories
of Political Influence on Defense Procurement in Congressional Districts:
98th - 102nd Congress, Working Paper.
42

Figure 1. Histograms of stimulus amount and Democratic vote share

Notes: Sample includes the 334 districts not containing state capitals. Stimulus data includes only contract, grants, and loans reported on
Recovery.gov. Democratic vote share is the two-party vote share: the percentage of Democratic votes out of total votes for the Democratic and
Republican candidate in the 2008 congressional election.

Figure 2: Compare stimulus amount per resident to amount per worker.

Notes: For the scatter plots, each dot represents one of the 334 districts not containing state capitals. Amount per resident refers to people
living in the congressional district; Amount per worker refers to people employed within the congressional district, though they may reside
elsewhere. Panel (d) displays the results of a non-parametric regression of the worker-to-resident ratio on the Democratic vote share.

Figure 3: Non-parametric and semi-parametric regressions of amount per resident on Democratic vote share.

Notes: Panel (a) is a non-parametric regression (using kernel-weighted local polynomial smoothing) of stimulus amount per resident on
Democratic vote share. Panel (b) reports the results of semi-parametric regression that includes (parametric) controls for employment, poverty
rate, percent urban, land area, and road miles. Sample includes 334 congressional districts not located in state capitals.

Figure 4: Non-parametric and semi-parametric regressions of amount per worker on Democratic vote share.

Notes: Corresponds to Figure 3, except that the dependent variable in this figure is amount per worker, and employment is not included in the
set of controls in Panel (b).

Figure 5: Congressional district-level stimulus amount per resident.

Notes: Each dot represents one of the 334 districts not containing state capitals. Percent urban is the percentage of the district population
living in urban areas. Tenure refers to the number of terms served by the congressional representative. Leadership refers to party leaders and
committee chairs (for Democrats) or ranking members (for Republicans).

Table 1: Descriptive Statistics

Variable

With Capitals
Mean Std. Dev.

Without Capitals
Mean Std. Dev.

5334
1056
8.7
7.6
976
.00029
6.1
.24
.06

5705
450
6.6
1.8
326
.00035
6.8
.43
.24

2112
517

3133
262

613
900
1941
13
79
.81
852
297
33
6.2
58
-.034

1213
1786
3623
5.6
20
3.1
975
100
11
4.6
23
.44

316
469
1101
14
80
.54
762
289
31
6.5
58
-.033

370
543
1308
5.8
20
1
864
103
10
4.6
24
.44

State level
Stimulus amount (millions)
Stimulus amount per capita
State budget gap (%)
State unemployment rate
Medicaid per capita
Interstate miles per capita
State population (millions)
Senate fiscal Chair or Ranking Member
Obama vote share 50-52%
District level
Stimulus amount (millions)
Stimulus amount per capita
Stimulus amount per worker
Percent below poverty line
Percent urban
Land Area (square miles)
Total highway miles
Employment (1000s of workers)
% spent in 1 year
Tenure (number of 2-year terms)
Democratic vote share
DW-Nominate score
Total stimulus amount

$267 billion

$106 billion

Table 2: State- and district-level regressions: stimulus amount per resident vs. state-level characteristics.

Constant
State budget gap (%)
State unemployment rate
Medicaid per capita
Interstate miles
State population
Senate fiscal Chair or RM
Obama 50-52%
Observations
R2
State or district level
Excludes state capitals

(1)
FFIS Total

(2)
FFIS Other

(3)
CGL

(4)
CGL

(5)
CGL

(6)
CGL

(7)
Agency

363.2
(50.31)
-1.105
(1.178)
-0.615
(5.107)
0.330
(0.0388)
19.08
(2.388)
0.881
(1.363)
11.25
(19.54)
-13.31
(16.82)

107.2
(18.69)
-1.518
(0.557)
2.602
(2.547)
0.0738
(0.0100)
8.239
(1.200)
-0.683
(0.365)
9.912
(8.140)
-13.21
(7.574)

336.9
(254.0)
27.95
(8.506)
-15.94
(22.94)
0.369
(0.136)
93.21
(29.20)
-6.743
(4.721)
45.39
(94.29)
-81.11
(73.75)

648.7
(580.1)
8.774
(7.385)
-28.65
(38.63)
0.0344
(0.102)
21.24
(114.6)
-2.901
(6.927)
-47.10
(66.89)
-202.0
(59.60)

420.9
(518.7)
18.74
(15.50)
4.029
(56.58)
0.166
(0.295)
95.37
(33.47)
-2.920
(10.33)
34.06
(184.0)
-140.0
(275.9)

561.1
(345.8)
13.48
(4.745)
-15.13
(19.24)
-0.00501
(0.0911)
-9.019
(89.25)
-4.082
(5.237)
40.99
(71.58)
-211.8
(61.41)

1103.4
(456.6)
9.857
(7.005)
-21.74
(51.47)
0.569
(0.102)
27.35
(18.32)
-4.628
(5.386)
-41.22
(126.7)
-63.84
(83.25)

50
0.817
state

50
0.775
state

50
0.718
state

37
0.136
state
Y

435
0.010
district

334
0.033
district
Y

50
0.269
state

Notes: Columns (1)-(4) and (7) are state-level regressions, and columns (5) and (6) are district-level regressions. The dependent variable is a
measure of stimulus receipts per resident and varies across columns. Column (1) uses estimated state-level total receipts from Federal Funds
Information for States. Column (2) uses the same data source but excludes stability, medicaid, and highway aid. Columns (3)-(6) use contracts,
grants and loans from Recovery.gov. Column (7) uses state-level agency-reported data from Recovery.gov. The units for the explanatory
variables are as follows: state budget gaps are estimates for FY 2009, expressed as percentage of general fund budget; state unemployment
rates are for February 2009; medicaid expenditures for 2005 are in dollars per capita; interstate miles are in miles per 10,000 persons; and state
populations are in millions. Robust standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

Table 3: State-level stimulus amount per resident.

Independent Variable:

Change in
unemployment

State
Employment

Percent
Poverty

Democratic
Governor

Average Cong
Tenure

NE, PA
ME

NE, PA

Panel (a): Amount per resident Agency-reported

No controls

R2

-42.60
( 40.38)

-0.0254
( 0.0163)

-48.28*
( 25.20)

43.78
( 120.7)

15.91**
( 7.712)

-56.43
( 211.8)

-287.5**
( 134.4)

0.015

0.027

0.110

0.003

0.043

0.001

0.018

Panel (b): Amount per resident Agency-reported

With controls

R2

29.41
( 38.03)

-0.0117
( 0.0131)

-46.03*
( 24.61)

-3.873
( 93.85)

10.10*
( 5.414)

-305.0**
( 120.1)

-365.4***
( 115.9)

0.351

0.346

0.346

0.346

0.363

0.373

0.374

Panel (c): Amount per resident Contracts, Grants, Loans

With controls

R2
Observations

39.77
( 26.35)

-0.0141
( 0.0157)

-18.38
( 13.73)

-160.5*
( 92.30)

13.87
( 8.992)

-281.0***
( 76.78)

-266.2***
( 53.03)

0.605

0.596

0.596

0.627

0.625

0.617

0.609

50

50

50

50

50

50

50

Notes: The dependent variable is amount per state resident, using the agency-reported data from Recovery.gov for panels (a) and (b), and the
contracts, grants, and loans data for Panel (c). Controls include medicaid expenditures per capita, interstate highway miles, total employment
in the state, and the poverty rate. Robust standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

Table 4: District-level regressions: stimulus amount vs. economic characteristics

Independent Var:

Unemp
Rate

Change in
Unemp

UI per
capita

% spent
in 1 year

Employment
(thousands)

Poverty
Rate

% Urban

Panel (a): Amount per resident

No controls

R2

-8.840
( 11.62)

-32.68
( 25.83)

-47.76
( 310.2)

-4.966*
( 2.762)

1.962***
( 0.366)

16.14***
( 5.911)

1.562
( 1.015)

0.001

0.007

0.000

0.009

0.139

0.030

0.003

Panel (b): Amount per resident

All controls except
land area
R2

8.523
( 14.47)

-2.242
( 31.37)

418.2
( 475.1)

-6.573**
( 2.618)

1.970***
( 0.321)

14.05*
( 8.172)

0.533
( 2.004)

0.257

0.256

0.259

0.271

0.256

0.256

0.256

Panel (c): Amount per resident

All controls

R2

6.824
( 15.24)

-6.252
( 33.04)

443.5
( 481.8)

-5.619**
( 2.550)

2.019***
( 0.315)

12.87
( 8.219)

-1.579
( 1.434)

0.277

0.277

0.280

0.287

0.277

0.277

0.277

Panel (d): Amount per worker

No controls

R2

31.09
( 31.62)

-32.07
( 66.99)

4.744
( 698.3)

-20.05***
( 6.598)

-0.207
( 0.477)

42.03***
( 11.52)

-2.963
( 2.594)

0.002

0.001

0.000

0.026

0.000

0.035

0.002

Panel (e): Amount per worker

All controls

R2
Observations

22.41
( 35.70)

-6.855
( 74.86)

928.7
( 1022.3)

-17.76***
( 6.21)

0.052
( 0.341)

30.76*
( 16.45)

-2.544
( 3.847)

0.102

0.101

0.104

0.120

0.101

0.101

0.101

334

334

334

334

334

334

334

Notes: The dependent variable is district-level stimulus receipts as contracts, grants and loans for 334 districts not in state capitals; panels
(a)-(c) use stimulus amount per resident, and panels (d)-(e) use amount per worker in the district. All controls include the vote share
group dummies shown in Table 6, as well as employmemt, poverty rate, percent urban, land area, and road miles (though employment is not
included in the per worker specifications). UI per capita indicates year 2008 total district-level unemployment insurance receipts in thousands
of dollars per capita. The district unemployment rate is the value for January 2009. The change in unemployment is the difference between
the unemployment rates in January 2009 and January 2007. Robust standard errors in parentheses, adjusted for clustering at the state level.
* p < 0.10, ** p < 0.05, *** p < 0.01.

Table 5: District-level regressions: stimulus amount vs. political party and district characteristics

Democrat

(1)

(2)

(3)

(4)

(5)

(6)

(7)
per wkr

(8)
per wkr

373.4
(122.6)

94.96
(45.44)

92.88
(51.08)

19.00
(61.58)

109.3
(50.07)

33.88
(64.45)

34.00
(158.4)

-85.12
(181.4)

1.953
(0.366)

2.172
(0.395)

2.097
(0.322)

2.247
(0.359)

Employment (thousands)

22.19
(7.437)

Percent Poverty

20.49
(7.859)

39.85
(14.80)

Land Area

108.0
(54.13)

103.6
(54.72)

327.6
(193.5)

Total Highway miles

-0.0171
(0.0498)

-0.0399
(0.0596)

-0.152
(0.193)

Percent Urban

0.183
(1.391)

0.198
(1.441)

1.576
(3.494)

Constant
Observations
R2
Includes state capitals

683.6
(78.84)

415.5
(48.56)

-148.8
(103.3)

-472.1
(148.1)

-260.2
(129.6)

-521.1
(185.9)

1085.1
(162.7)

423.0
(347.5)

433
0.011
Y

332
0.007

332
0.145

332
0.195

332
0.176

332
0.217

332
0.000

332
0.071

Notes: The dependent variable in columns (1)-(6) is district-level stimulus receipts as contracts, grants and loans, per resident; in column
(7)-(8), the dependent variable is stimulus receipts per worker employed in the district. Column (1) includes data for all congressional districts;
columns (2)-(8) include data for only those districts not located in state capitals. Two House seats were vacant at the time of the ARRA
vote, and they are coded as missing party affiliation, so the number of observations is 2 less than in other tables. Robust standard errors in
parentheses, adjusted for clustering at the state level. * p < 0.10, ** p < 0.05, *** p < 0.01.

10

Table 6: District-level regressions: stimulus amount vs. vote share blocks

(1)
per res

(2)
per res

(3)
per res

(4)
per worker

(5)
per worker

Unopposed Republican (13)

161.5
(130.9)

242.1
(115.7)

123.8
(106.5)

437.9
(278.7)

150.7
(290.3)

Dem vote share 1-30 (16)

70.13
(133.8)

147.9
(124.8)

-95.42
(174.0)

77.35
(343.1)

-579.6
(537.4)

Dem vote share 30-40 (63)

-6.304
(100.3)

60.67
(105.8)

26.38
(91.28)

-31.88
(321.4)

-117.2
(287.7)

Dem vote share 50-60 (42)

-32.82
(66.14)

13.01
(89.64)

-56.16
(87.81)

-138.5
(285.7)

-330.1
(298.4)

Dem vote share 60-70 (44)

12.87
(67.85)

86.87
(75.33)

15.65
(71.92)

-39.66
(263.1)

-228.8
(254.6)

Dem vote share 70-80 (51)

105.4
(40.03)

146.4
(47.52)

100.1
(64.82)

34.42
(219.4)

-101.7
(235.6)

Dem vote share 80-90 (20)

792.9
(144.0)

685.3
(147.3)

603.7
(189.1)

837.3
(390.6)

658.6
(460.0)

Dem vote share 90-99 (5)

-21.47
(188.0)

261.3
(135.5)

106.6
(206.7)

405.7
(395.0)

-100.5
(551.4)

Unopposed Democrat (34)

47.45
(103.5)

125.2
(115.0)

51.49
(139.3)

-16.47
(353.5)

-215.5
(374.0)

1.741
(0.334)

2.019
(0.315)

334
0.220

334
0.277
X

334
0.031

334
0.101
X

Employment (thousands)
Observations
R2
Additional controls

334
0.121

Notes: The dependent variable in columns (1)-(3) is stimulus receipts per district resident; the dependent variable in columns (4)-(5) is stimulus
receipts per worker (i.e., per person employed in the district). The vote share blocks are dummy variables that equal 1 if the Democratic vote
share for that representative falls in the specified range. The number of representatives in each group is indicated in parentheses. The omitted
category is 40-50% Democratic vote share, and there are 46 representatives in this group. Additional controls include poverty rate, percent
urban, land area, and road miles. Robust standard errors in parentheses, adjusted for clustering at the state level. * p < 0.10, ** p < 0.05,
*** p < 0.01.

11

Table 7: District-level regressions: stimulus amount vs. party elites

(1)
per res

(2)
per res

(3)
per worker

(4)
per worker

Democratic leader

0.142
(76.50)

-86.95
(68.03)

-19.91
(156.3)

-249.7
(157.5)

Republican leader

244.0
(203.4)

383.3
(206.2)

1055.9
(742.5)

1175.4
(684.2)

Number of terms
Observations
R2
Vote share controls
Additional controls

334
0.011

334
0.301
X
X

334
0.035

334
0.141
X
X

(5)
per res

(6)
per res

(7)
per worker

(8)
per worker

7.021
(3.506)

2.430
(3.079)

14.06
(7.780)

7.567
(7.331)

332
0.004

332
0.276
X
X

332
0.002

332
0.101
X
X

Notes: Additional controls include employmemt, poverty rate, percent urban, land area, and road miles. Democratic leaders include committee
chairs as well as the Speaker of the House (Pelosi); Republican leaders include ranking minority members of committees as well as the Minority
Leader (Boehner); the Majority Leader and both Whips represented districts in state capitals, and so are not included in the sample. Per res
indicates that the dependent variable is amount per resident. Robust standard errors in parentheses, adjusted for clustering at the state level.
* p < 0.10, ** p < 0.05, *** p < 0.01.

12

Table 8: District-level regressions: stimulus amount vs. other political characteristics

Independent Variable:

DW-Nominate

abs(DW-N)

Democratic
Governor

PA, NE

Vote Against
Party or Abstain

Democrats
voting against

-145.4***
( 51.83)

308.3*
( 158.9)

52.96
( 75.73)

-6.932
( 38.24)

-86.49
( 57.93)

-87.65
( 68.22)

0.014

0.008

0.002

0.000

0.001

0.000

-109.6
( 154.1)

7.897
( 161.8)

77.94
( 76.13)

40.70
( 44.25)

-82.79
( 49.96)

-66.96
( 60.11)

0.256

0.255

0.261

0.257

0.255

0.257

-120.2
( 158.4)

15.32
( 166.4)

88.46
( 71.85)

70.76
( 52.22)

-132.1***
( 46.71)

-143.9**
( 57.75)

0.276

0.275

0.282

0.277

0.277

0.278

-76.10
( 167.5)

386.1
( 358.8)

24.53
( 190.7)

-89.36
( 99.54)

-230.5
( 145.9)

-150.5
( 167.6)

0.001

0.002

0.000

0.000

0.001

0.000

-29.50
( 366.7)

134.4
( 412.2)

140.3
( 161.6)

33.91
( 143.5)

-323.7**
( 126.3)

-334.8**
( 159.2)

0.101

0.101

0.104

0.101

0.102

0.102

332

332

334

334

332

334

Panel (a): Amount per resident

No controls

R2

Panel (b): Amount per resident

All controls except
land area
R2

Panel (c): Amount per resident

All controls

R2

Panel (d): Amount per worker

No controls

R2

Panel (e): Amount per worker

All controls

R2
Observations

Notes: The dependent variable is district-level stimulus receipts as contracts, grants and loans for 334 districts not in state capitals; panels
(a)-(c) use stimulus amount per resident, and panels (d)-(e) use amount per worker in the district. All controls include the vote share group
dummies shown in Table 6, as well as employmemt, poverty rate, percent urban, land area, and road miles (though employment is not included
in the per worker specifications). Vote against party or abstain includes 5 Democrats who voted against the bill, 1 Democrat who voted
present, and 2 Republicans who did not vote. Democrats voting against include 5 Democrats who voted against the bill and 1 who voted
present. Robust standard errors in parentheses, adjusted for clustering at the state level. * p < 0.10, ** p < 0.05, *** p < 0.01.

13

A1

Appendix

A1

Figure A1. State-level map: stimulus per resident.

Notes: Stimulus funds paid out as of February 2014, by state. Source is Federal agency-reported data from Recovery.gov. Total amount for 50
states is $460 billion, and includes spending on entitlements as well as contracts, grants, and loans.

A2

Figure A2a. District-level amount per resident, with capitals.

Figure A2b. District-level amount per resident, no capitals.

A3

Figure A3. Proportion of total district amount accounted for by largest 5% of awards in that district.

A4

Figure A4. Outliers.

A5

Figure A5. District-level: Amount per resident vs. unemployment rate and excess capacity.

A6

Figure A6. District-level: Shovel readiness.

A7

Figure A7. District-level: Amount per resident vs. Congressional tenure.

Notes: Semi-parametric controls include total employment, poverty rate, percent urban, land area, and road miles.

A8

Figure A8: Semi-parametric regression of amount per resident on DW-NOMINATE score. Controlling for Democratic
vote share.

Notes: Semi-parametric regression of stimulus amount per resident on DW-NOMINATE score for 334 congressional districts. The set of semiparametric controls includes total employment, poverty rate, percent urban, land area, and road miles. Panel (b) adds the Democratic vote
share as an additional parametric control. The horizontal axis is reversed so that left-wing representatives appear on the right (to correspond
more closely to the previous figures).

A9

Table A1. Top funding agencies. Includes money going to state capitals.

Rank
w/ caps
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Funding Agency

Office of Elementary and Secondary Education

Department of Energy
Federal Highway Administration
Office of Special Education and Rehabilitative Services
Department of Housing and Urban Development
National Institutes of Health
Federal Transit Administration
Federal Railroad Administration
Environmental Protection Agency
Rural Utilities Service
Administration for Children and Families
U.S. Army Corps of Engineers - civil program financing only
Department of Labor
Public Buildings Service
National Telecommunication and Information Administration
Department of Justice
Department of Education
National Science Foundation
Health Resources and Services Administration
Department of Health and Human Services
Federal Financing Bank
Department of the Army
Department of the Air Force
Rural Housing Service
Department of Defense (except military departments)

Rank
no caps
9
1
2
44
5
3
4
21
28
6
14
7
31
8
11
10
46
13
12
16
15
20
23
19
26

Total Amount (million $)

w/ caps
64,678
38,352
27,892
13,585
10,869
10,462
9,277
7,515
7,354
6,741
5,015
4,638
4,248
4,242
4,218
4,086
2,898
2,690
2,468
2,357
2,037
1,943
1,541
1,538
1,529

A10

no caps
1,922
20,952
19,196
309
5,831
7,384
6,692
866
699
4,811
1,568
3,456
529
2,696
1,850
1,909
290
1,599
1,603
1,125
1,202
904
806
934
753

% in caps
97%
45%
31%
98%
46%
29%
28%
88%
90%
29%
69%
25%
88%
36%
56%
53%
90%
41%
35%
52%
41%
53%
48%
39%
51%

Number
of awards
w/ caps
756
4316
13879
735
6411
16891
974
86
903
2148
2992
4583
707
1487
294
4728
1929
4980
3680
899
3
1973
1708
1753
283

Median
amount ($)
w/ caps
809,162
649,350
569,170
316,278
317,515
257,773
2,000,000
8,315,000
612,000
1,295,500
261,244
132,302
946,034
167,267
5,660,544
157,102
58,017
330,935
333,469
367,801
692,000,000
359,613
314,774
95,000
680,144

Max award
size ($)
w/ caps
4,880,000,000
1,360,000,000
261,000,000
1,230,000,000
326,000,000
157,000,000
423,000,000
2,550,000,000
433,000,000
83,100,000
220,000,000
62,300,000
489,000,000
148,000,000
155,000,000
136,000,000
1,120,000,000
148,000,000
89,700,000
62,500,000
1,200,000,000
32,700,000
51,700,000
54,000,000
531,000,000

Table A2. Top funding agencies. Excludes money going to state capitals.

Rank
no caps
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Funding Agency

Department of Energy
Federal Highway Administration
National Institutes of Health
Federal Transit Administration
Department of Housing and Urban Development
Rural Utilities Service
U.S. Army Corps of Engineers - civil program financing only
Public Buildings Service
Office of Elementary and Secondary Education
Department of Justice
National Telecommunication and Information Administration
Health Resources and Services Administration
National Science Foundation
Administration for Children and Families
Federal Financing Bank
Department of Health and Human Services
Department of Veterans Affairs
National Aeronautics and Space Administration
Rural Housing Service
Department of the Army
Federal Railroad Administration
Federal Aviation Administration
Department of the Air Force
Department of the Navy
Office of Science

Rank
w/ caps
2
3
6
7
5
10
12
14
1
16
15
19
18
11
21
20
26
30
24
22
8
27
23
28
32

Total Amount (million $)

no caps
20,952
19,196
7,384
6,692
5,831
4,811
3,456
2,696
1,922
1,909
1,850
1,603
1,599
1,568
1,202
1,125
1,000
999
934
904
866
829
806
801
762

A11

w/ caps
38,352
27,892
10,462
9,277
10,869
6,741
4,638
4,242
64,678
4,086
4,218
2,468
2,690
5,015
2,037
2,357
1,501
1,104
1,538
1,943
7,515
1,287
1,541
1,261
961

% in caps
45%
31%
29%
28%
46%
29%
25%
36%
97%
53%
56%
35%
41%
69%
41%
52%
33%
10%
39%
53%
88%
36%
48%
36%
21%

Number
of awards
no caps
2668
9057
12061
681
4753
1441
3156
967
183
3050
125
2471
3191
1895
1
359
1275
375
1305
907
26
267
960
255
287

Median
amount ($)
no caps
675,381
595,605
253,194
1,770,192
297,457
1,402,000
156,606
190,648
287,500
143,288
6,162,554
341,595
328,505
269,838
1,200,000,000
457,844
280,417
500,000
89,081
382,000
5,002,500
1,811,658
285,235
1,513,770
600,000

Max award
size ($)
no caps
1,190,000,000
261,000,000
157,000,000
423,000,000
144,000,000
83,100,000
62,300,000
127,000,000
720,000,000
50,200,000
155,000,000
12,000,000
18,500,000
27,700,000
1,200,000,000
62,500,000
29,600,000
166,000,000
44,700,000
32,700,000
400,000,000
30,300,000
23,700,000
64,800,000
65,000,000

Table A3. State-level regressions: additional specifications

Constant
State unemployment rate

(1)
Total

(2)
Total

(3)
Total

(4)
education

(5)
highway

(6)
Total

1433.4
(223.1)
-53.51
(24.43)

913.4
(103.0)

1565.5
(278.7)
-102.2
(46.35)
2709.6
(1760.0)

291.6
(34.89)
-3.621
(6.107)
189.5
(282.0)

98.81
(25.53)
0.842
(3.118)
22.51
(87.99)

-0.677
(1.174)
-18.08
(13.54)
-6.643
(13.39)

18.44
(2.060)
1.047
(1.031)
-3.354
(12.53)
-21.67
(17.36)

295.7
(251.8)
-13.43
(23.67)
2597.3
(845.4)
0.390
(0.135)
89.42
(28.83)
-6.921
(4.603)
56.39
(93.28)
-80.82
(68.87)

50
0.066

50
0.669

50
0.710

State budget gap

1304.2
(1397.2)

Medicaid per capita

Interstate miles
State population
Senate fiscal Chair or RM
Obama 50-52%
Observations
R2

50
0.050

50
0.039

50
0.176

Notes: Dependent variable is stimulus amount per resident (from Recovery.gov) aggregated to the state level. Columns (4) and (5) only contain
those contracts, grants, or loans that we were able to classify as education or highway spending, respectively. Specifications in this table
correspond to similar specifications in Table 3 of Inman (2010). See notes to Table 2 for explanation of variables. Robust standard errors in
parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

A12

Table A4. District-level stimulus amount: other controls

(1)
per res
Percent Poverty

(2)
per res

(3)
per res

(4)
per res

16.14
(5.911)

Percent Urban

1.562
(1.015)

(7)
per wkr

(8)
per wkr

-2.963
(2.594)
52.88
(25.24)

257.3
(99.41)

Total Highway miles

Observations
R2

(6)
per wkr

42.03
(11.52)

Land Area

Constant

(5)
per wkr

0.154
(0.0677)

0.00556
(0.0257)
249.1
(74.59)

344.3
(77.96)

440.6
(31.69)

464.9
(41.63)

528.4
(143.6)

1338.1
(211.2)

962.4
(69.30)

983.6
(103.4)

334
0.030

334
0.003

334
0.010

334
0.000

334
0.035

334
0.002

334
0.040

334
0.010

Notes: The table shows coefficients on the additional controls that were included in several of the regressions discussed in the paper. The
dependent variable is district-level stimulus amount per resident (per res) or stimulus amount per worker (per wkr) received as contracts,
grants and loans. Robust standard errors in parentheses, adjusted for clustering at the state level. * p < 0.10, ** p < 0.05, *** p < 0.01.

A13

Table A5. Small sample committee tests

Compare district of 111th committee leaders to
districts represented by committee leaders in...
110th Congress 112th Congress 110th or 112th
Democrats
No. of cases
No. of times 111th leader receives more
Percentage
p-value:
1-sided test
2-sided test

3
1
0.33

8
4
0.50

11
5
0.45

0.50
1.00

0.69
1.00

0.50
1.00

6
5
0.83

4
2
0.50

10
7
0.70

0.04
0.08

0.76
1.00

0.09
0.18

Republicans
No. of cases
No. of times 111th leader receives more
Percentage
p-value:
1-sided test
2-sided test

Notes: This table shows the results of small sample tests (Fisher exact tests) for whether committee leaders receive more stimulus funds.
Leaders here includes both chairpersons and ranking members. We compare the districts whose representatives are leaders during the 111th
Congress (which passed the ARRA) to districts whose representatives are leaders of the same committee during the previous or following
Congress. This test only works for those committees with a change in leadership.

A14

Table A6: District-level regressions: stimulus amount vs. party elites

(1)
per res

(2)
per res

(3)
per worker

(4)
per worker

(5)
per res

(6)
per res

(7)
per worker

(8)
per worker

Committee chair (D)

-83.38
(70.50)

-126.2
(57.57)

-96.81
(173.4)

-309.7
(149.7)

Ranking member (R)

271.0
(212.8)

409.1
(212.6)

1149.8
(774.7)

1255.3
(698.1)

Pelosi

1485.9
(31.28)

581.8
(169.3)

1311.6
(68.70)

649.5
(398.0)

Boehner

-338.4
(31.28)

-149.8
(63.52)

-757.8
(68.70)

-422.9
(216.2)

Democratic leader

0.142
(76.62)

-84.22
(65.41)

-19.91
(156.5)

-240.9
(149.3)

Republican leader

-106.0
(56.18)

29.58
(43.46)

-139.1
(99.73)

-16.14
(113.4)

McKeon, Hastings

3325.0
(70.06)

3291.4
(78.48)

11352.2
(2104.0)

11057.7
(2140.0)

334
0.212

334
0.491
X
X

334
0.440

334
0.511
X
X

Observations
R2
Vote share controls
Additional controls

324
0.038

324
0.308
X
X

324
0.044

324
0.147
X
X

Notes: Additional controls include employmemt, poverty rate, percent urban, land area, and road miles. Democratic leaders include committee
chairs as well as the Speaker of the House (Pelosi); Republican leaders include ranking minority members of committees as well as the Minority
Leader (Boehner); the Majority Leader and both Whips represented districts in state capitals, and so are not included in the sample. Per res
indicates that the dependent variable is amount per resident. Robust standard errors in parentheses, adjusted for clustering at the state level.
* p < 0.10, ** p < 0.05, *** p < 0.01.

A15