ARTICLE IN PRESS

Journal of Monetary Economics 53 (2006) 22832298

www.elsevier.com/locate/jme

Risk-based pricing of interest rates

for consumer loans$
Wendy Edelberg
Received 17 June 2005; received in revised form 15 August 2005; accepted 12 September 2005

Abstract
By focusing on observable default risks role in loan terms and the subsequent consequences for
household behavior, this paper shows that lenders increasingly used risk-based pricing of interest
rates in consumer loan markets during the mid-1990s. It tests three resulting predictions: First, the
premium paid per unit of risk should have increased over this period. Second, debt levels should have
reacted accordingly. Third, fewer high-risk households should have been denied credit, further
contributing to the interest rate spread between the highest- and lowest-risk borrowers.
For people obtaining loans, the premium paid per unit of risk did indeed become signicantly
larger after the mid-1990s. For example, for a 0.01 increase in the probability of bankruptcy, the
corresponding interest-rate increase tripled for rst mortgages, doubled for automobile loans and
rose nearly six-fold for second mortgages. Additionally, changes in borrowing levels and debt access
reected these new pricing practices, particularly for secured debt. Borrowing increased most for the
low-risk households who saw their relative borrowing costs fall. Furthermore, while very high-risk
households gained expanded access to credit, the increases in their risk premiums implied that their
borrowing as a whole either rose less or, sometimes, fell.
r 2006 Elsevier B.V. All rights reserved.
JEL classification: D12; E21; E51; G21
Keywords: Consumption; Borrowing; Debt; Consumer credit; Interest rates; Banking

Federal Reserve Board, e-mail: Wendy.M.Edelberg@frb.gov. The views presented are solely those of the
author and do not necessarily represent those of the Federal Reserve Board or its staff. I would like to thank
Pierre-Andre Chiappori, Lars Hansen, Erik Hurst and Annette Vissing-Jorgensen, for their direction and advice. I
also would like to thank the University of Chicago, the National Science Foundation and the Social Science
Research Council for their nancial support. Of course, all errors are my own.
E-mail address: wendy.m.edelberg@frb.gov.
0304-3932/$ - see front matter r 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.jmoneco.2005.09.001

ARTICLE IN PRESS
2284

W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

1. Introduction
Credit industry literature suggests that by the early 1980s conventional lenders were
using credit scores and the like to automate underwriting standards, but as late as the early
1990s they still simply posted one house rate for each loan type and continued to reject
most high-risk borrowers (Johnson, 1992). As data storage costs subsequently fell and
underwriting technology improved, however, lenders began to use estimates of default risk
to price individual loans. This paper examines both the extent and consequences of this
increased use of risk-based pricing of interest rates in consumer loan markets during the
mid-1990s.
On the whole, the ndings are in keeping with the predictions that ow naturally from
these changes. First, for those obtaining loans, the premium paid per unit of risk became
signicantly larger over this time period, with the difference between high- and low-risk
borrowers interest rates nearly doubling for secured loans and increasing for most
unsecured loans as well. Second, changes in borrowing levels and access to debt reected
these new pricing practices, particularly for secured debt. While lower interest rates
generally boosted borrowing in the late 1990s, the demand for credit increased most for
low-risk households who saw lower relative borrowing costs. Third, these changes in
pricing practices led to increased credit access for very high-risk households (again,
particularly for secured debt), but the increase in the very high-risk premium also caused
their average borrowing levels to either rise less or, for some loan types, to fall. Finally,
changes in risk-based pricing may account for one- to three-quarters of the increase in debt
levels for some secured loan types and may more than account for the increase in debt use
by the highest-risk groups for secured debt.
There has not been much scrutiny of the potential for credit terms to vary by borrower
risk, let alone empirical examinations of such variance in terms. On the theoretical side,
Geanakoplos has written and co-written several papers showing the effect of default risk
on loan terms in general equilibrium (some examples are Geanakoplos (2002) and Dubey
et al. (2003)). Riley (1987) argues StiglitzWeiss style rationing will not be empirically
relevant, as he postulates that lenders should vary interest rates by risk. However, using
1983 mortgage rate data, Duca and Rosenthal (1993) nd no evidence of such interest rate
variation. My ndings are consistent with Ducas and Rosenthals, suggesting that riskbased pricing did not become a signicant factor in credit markets until more than a
decade after 1983.
Only in the 1990s did improvements in underwriting models and reductions in data
storage costs became sizeable enough to decrease the costs of risk-based pricing (Bostic,
2002).1 Certain changes in consumer credit industry practices also spurred investment in
developing new underwriting models. Canner and Passmore (1997) explain that in 1995
bank regulators began putting greater emphasis on lending in lower-income neighborhoods and to lower-income borrowers in measuring compliance with the Community
Reinvestment Act. This increased the protability of developing a technology to lend to
higher-risk households. Furthermore, Fannie Mae, which previously bought only low-risk
loans and essentially did not vary nancial terms with loan risk, introduced an improved,
1
In addition, Peter McCorkell suggests insufcient data on defaults made risk-based pricing difcult prior to
1995. He also argues that until the late 1980s, mortgage lenders simply relied on their constantly appreciating
collateral to moderate the costs of default (McCorkell, 2002).

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2285

Table 1
Interest rate data

First mortgage ratea

Second mortgage
Auto loan
Credit cardc
Other consumer loan
Education loan

Standard deviation by origination year

Observations

1989

1995

1998

All years

1.16
2.21
3.58
4.18
4.47
3.37

1.26
2.82b
4.05b
4.43b
6.07b
4.05b

1.49b
2.63
4.53b
5.01b
6.69
1.96

8,143
805
5,209
4,007
2,744
997

Only 30-year xed rate mortgages are considered.

Difference between current and preceding year is signicant with p-valueo0.1.
c
1983 is used in place of 1989.
b

automated underwriting system in 1995 and began to accept higher-risk loans subject to
some price discrimination. In 1996, Fannie Mae and Freddie Mac indicated that loan
packages must include a credit bureau score (McCorkell, 2002).
In the mid-1990s, lenders could, and did, vary interest rates and issue higher-risk
mortgages (Freeman and Hamilton, 2002). As a result even credit unions began using riskbased pricing at this time as low-risk members complained that they were able to get lower
rates at conventional banks.2 The technology of risk-based pricing made its way from
mortgage loans into other loan types, such as second mortgages, automobile loans and
credit card loans. For example, Black and Morgan (1999) nd demographic evidence that
more high-risk households gained access to the credit card market in 1995 relative to 1989.
(Indeed, this increased access appears to have been widespread: average household income
went up about 20% more than the average income of those with any debt, pre-1995 versus
post-1995. Similarly, average education rose about 40% more overall than for those with
debt.) My results show that loans easily securitized, such as those mentioned above, have
been affected the most these pricing changes, suggesting that secondary loan markets have
played a role in promoting risk-based pricing.
2. Data
This analysis uses the Surveys of Consumer Finances (SCFs) from 1983 to 1998. First
and second mortgages, automobile loans, general consumer loans, credit card loans and
education loans are considered. Loans in a category are summed and the highest interest
rate is used.3 Table 1 shows the standard deviations for three loan originations years: 1989,
1995 and 1998 (except for credit card loans, which substitutes 1983 for 1989 due to data
restrictions).4 Consistent with the increased use of risk-based pricing, interest rate variation
generally increased over time, and often signicantly. Note that standard deviations of
2

This point was made in conversations with the University of Wisconsin Center for Credit Union Research.
Credit card balances are considered loans when interest is paid on the balance. Note that Gross and Souleles
(2001) points out an underreporting of credit card debt in the SCF, which is problematic only if this
underreporting is signicantly correlated with risk, and this correlation changes over time.
4
Sampling weights are used for rst and second moments. Following Deaton (1997), the data are not weighted
in the empirical models as coefcients are assumed not to vary across the population.
3

ARTICLE IN PRESS
2286

W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

monthly prime interest rates were similar in 1989 and 1995 and actually decreased between
1995 and 1998. In addition, the table shows the total observations across the 5 years of
data for the various loan categories, anticipating some of the differences in the results
robustness. For its more extensive data on bankruptcy, the Panel Study of Income
Dynamics (PSID) is also used for the wealth supplement years of 1984, 1989 and 1994.
Total bankruptcies across all years prior to 1996 is 502, reecting a slightly lower
bankruptcy rate than in the population, a point made by Fay et al. (2002).
3. Empirical analysis
The primary goal of this empirical analysis is to estimate the role default risk plays in
interest rate determination and also to see if that role has changed over time. Assume that
a household has a reservation interest rate Ri(A,l,Pi), which is a function of a certain loan
amount, A, with collateral to ensure a recovery rate, l, of the loan balance, and household
characteristics, P. Ii(A,l,di,o,f), the interest rate offered by the lender, is a function of A, l,
default risk, d, the lenders discount rate, o, and xed costs, f.5 Because the SCF only
reports interest rate data for households who successfully secure loans, selection bias is
accounted for. We can infer that R is greater than or equal to I for those consumers who
have positive loan balances. To formalize:
Ri A; l; Pi  I i A; l; d i ; o; f H i b ui ,
ProbRi  I i 40 FH i b.
Hi is a vector of characteristics that helps predict whether the loan is observed for
household i. Ii and Ri, are subscripted i to allow for an idiosyncratic individual specic
shock, eI:
I i A; l; d i ; o; f X i g i

observed when Ri  I i 40.

Note that the linearity in the equation above assumes lenders are risk neutral or are
diversied enough to appear risk neutral. Xi is a vector of characteristics that help predict
I.6 X includes direct measures or proxies, where necessary, for the ve variables, A, l, d, o
and f. First, A is included directly.7 Second, l should be roughly constant for each type of
non-collateralized loan and hence captured by a varying constant term. For collateralized
loans, l should rise with collateral, and hence the equity in the collateral is included.8 Third,
measures of d are described in detail below. Fourth, o is assumed constant over a year and

5
Maturity does not generally vary meaningfully within a loan type and was often found to have no real
signicant effect on interest rates. For example, over one-half of mortgages have 30-year maturities, and nearly
60% of automobile loans have maturities between 4 and 5 years.
6
Note that this model essentially does not allow for a rejection by the lender. However, we can consider a loan
rejected any time RiIip0. For example, if a lender at least knows the upper bound for a households reservation
interest rate, it may choose to simply reject a loan rather than offer an interest rate above this upper bound.
7
Dollar values are deated using the CPI. For general consumer loans, current loan balances are better
predictors than original loan amounts. This may be due to the more informal nature of these loans. For example,
these loans may be renegotiated more easily so that current balance is also highly relevant for the terms.
8
The possible complication that l may be in part a function of dfor example, ex ante high-risk people in
default may be more difcult to collect from than low-risk people in defaultis not considered here.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2287

is captured by year dummies.9 Finally, inasmuch as xed costs, f, are recovered through
the interest rate, their effect should be captured by including A.
H includes both supply and demand variables that inuence whether a household holds
a loan. On the supply side, H includes variables that help predict denial such as secondorder polynomials of default risk. To account for demand, other nancial and
demographic characteristics, Pi, are included: an age polynomial, marriage status, the
number of children, whether the family has a checking account, education, log of income,
net worth, level of assets, race and variables that reect borrowing attitudes.10 The
attitudinal variables show whether households consider borrowing to be good, bad, or
okay and whether they believe borrowing is acceptable in certain circumstancessuch as
for a loss in income or to buy a house or jewelry. These variables ensure identication, as
they are excluded from X. Correlation estimates conrm that these responses are not
simple functions of borrowers debt portfolios.11

3.1. Default risk

Default risk is comprised of the risks of bankruptcy and delinquency. The SCF contains
no bankruptcy data before 1998, so the PSID is used to estimate a model of future
bankruptcy, and bankruptcy risk is imputed for SCF households.12 A household is dened
as bankrupt at time t if it declares bankruptcy during the period t to t+2. This future
measure allows for two forms of bankruptcy risk: conditional and unconditional on a
household holding debt. Conditional bankruptcy risk is relevant for lenders assessing
interest rates for loans, and thus will be included in X. Because the unconditional
bankruptcy risk measure does not include any debt measures, it is included in H, the vector
of characteristics predicting whether a household holds a certain loan. Note a full 14% of
these bankrupt households have no debt at time t.
In light of the extensive research on bankruptcy, the following variables are used in a
probit to predict bankruptcy risk: year, age, a checking account indicator, income, a self
employment indicator, home ownership status, unsecured debt, an indicator of whether the
ratio of unsecured debt to income exceeds two, net worth (with negative net worth set to
zero), unemployment indicator, race, single parent indicator, and education. For the
unconditional estimation, asset levels are used in place of debt and home ownership
status.13 Signicant time variation in coefcients is also included. Overall, the coefcients
are consistent with the bankruptcy literature. A detailed discussion of these results can be
found in Edelberg (2003).
9

This should in part reect the required rate of return to those supplying loanable funds. Ausubel (1991) nds
that credit card issuers earned possibly ve times the ordinary banking rate of return from 1983 to 1988. Here, no
specic rate of return is imposed, but it is assumed that markets are competitive.
10
Racial status may reect preferences for borrowing and potential lender bias (see Edelberg, 2002).
11
Numerous variables are included in H but not in X, such that the demands on the attitudinal variables as
appropriate instruments are less than they might be. However, robustness checks were still done. All the
signicant attitudinal variables from the probits were included in their respective regressions: For example, if an
attitudinal variable signicantly predicted mortgage use, it was included in the mortgage interest rate equation. In
all cases, most or nearly all of these coefcients in the interest rate regressions were insignicant.
12
The imputation of bankruptcy risk is based on a similar methodology in Jappelli et al. (1998).
13
Research used to identify explanatory variables includes Sullivan et al. (1989), Johnson (1992), Domowitz and
Sartain (1999), Gross and Souleles (1999), Sullivan et al. (2000), and Fay et al. (2002).

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2288

Table 2
Probability of default risk
Conditional bankruptcy

Percentiles 1%
10%
25%
50%
75%
90%
99%
Mean
Standard deviation
a

Unconditional bankruptcy

Delinquency

PSID

SCFa

PSID

SCFa

SCF

0.0
0.0
0.2
0.7
1.4
2.2
5.1
1.0
1.2

0.0
0.0
0.1
0.6
1.3
2.3
5.3
0.9
1.4

0.0
0.0
0.2
0.6
1.2
1.8
3.4
0.8
0.8

0.0
0.0
0.1
0.5
1.1
1.7
3.3
0.7
0.8

0.1
1.0
2.6
5.6
11.8
20.7
47.9
8.9
9.8

Bankruptcy risk in the SCF is imputed.

Counterparts in the SCF that are close to the bankruptcy determinants in the PSID
are used in order to impute bankruptcy risk for SCF households. The necessary
correction of the standard errors is done following Murphy and Topel (1985). Table 1
shows bankruptcy risk in the PSID and imputed risk in the SCF, which are
roughly similar. The bankruptcy risk is quite small: The 90th percentile household with
debt in the SCF still has only a 2.3% probability of declaring bankruptcy within the next 2
years.
Delinquency risk is the second measure of default risk. The SCF reports whether
respondents have been more than 60 days late on a loan payment in the previous year.14
Nearly 9% of the households report a delinquency, showing it is much more common than
bankruptcy. A probit is used to determine delinquency risk using the same determinants
that are in the conditional bankruptcy model.15 A selection model is estimated, as
delinquency data only exists for households with debt. Again, the attitudinal variables
ensure identication. And again, signicant time variation in coefcients is also included.
Delinquency riskconditional on holding debtis reported in the last column of Table 2.
Note that the correlation between bankruptcy and delinquency risk is only 0.35, showing
these are distinct measures of default risk.
3.2. Putting it all together
We now have three measures of default risk: delinquency risk, g, conditional bankruptcy
risk, fc, and unconditional bankruptcy risk, fuc. X and H are dened as follows:
X x; f c I 95 ; f c ; gI 95 ; g;

H h; f uc ; f 2uc .

14
As will be clear in the empirical analysis, a good bit of the information in late payments is indeed used by
lenders in pricing interest rates at loan origination. For every loan considered, average rates paid are higher for
those who made late payments versus those who had no late payments, with the differences ranging from a low of
0.2 percentage point for education loans to a high of nearly 2.5 percentage points for automobile loans.
15
The main results are robust to using either predicted or actual delinquencies. We might use actual
delinquencies given rational expectations, as lenders should predict correctly on average. In addition, if lenders
have superior data, their predictions of delinquency may be closer to the actual delinquencies than in this analysis.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2289

Table 3
Interest rates moments by origination year and risk class over time
High-risk versus low-risk spread

First mortgage rateb

Second mortgage
Auto loan
Credit cardd
Other consumer loan
Education loan

1989a

1995a

1998a

0.53
2.65
1.40
0.99
0.08
0.02

0.59
1.75
2.42c
1.05c
3.03c
1.30

0.69
2.84
3.94c
1.22
4.06
0.26

a
1998 spreads computed from 1998 and 1997, 1995 computed from 1995 and 1996, and 1989 computed
from 1988 and 1987 (except for credit card rates which are computed for single years).
b
Only 30-year xed rate mortgages are considered.
c
Difference between current and preceding year is signicant with p-valueo0.1.
d
1983 is used in place of 1989.

The indicator variable, I95, will determine whether the role of default risk in interest rate
determination changed after 1995. The vectors x and h contain the remaining variables in
X and H, respectively, aside from default risk. A default risk premium spread, s, measures
the difference in interest rates between the highest- and lowest-risk groups:
h
i h
i
s gf c f c;R I 95 gf c f c;R gg gR I 95 gg gR  gf c f c;L I 95 gf c f c;L gg gL I 95 gg gL ,
where the gs are the coefcients from the interest rate equation in the selection model. The
indicator function, I95, determines whether the risk premium is post- or pre-1995. R and L
dene averages of the default risk measures using conditional bankruptcy probabilities: the
highest- and lowest-risk groups are the 20% most and least likely to declare bankruptcy,
respectively.16
4. Empirical results
Table 3 shows the differences in average interest rates paid by most and least risky
groups (as dened above) for 1989, 1995 and 1998.17 The clear trend is for the difference to
rise over time, consistent with an increased use of risk-based pricing. That the difference is
not always signicant reects the value of the more careful and extensive analysis discussed
below.

16
Here, s is calculated using xed risk classespresenting an economically useful summary of the coefcients on
default risk. These premium spreads might change if we looked at the households actually using certain types of
loans pre- and post-1995, but this would make changes in pricing practices harder to isolate.
17
The spreads are taken from interest rates averaged for a 2-year period in order to have a reasonable number of
observations for each risk group. 1998 spreads are computed from 1998 and 1997, 1995 spreads are
computed from 1995 and 1996, and 1989 spreads are computed from 1988 and 1987 (except for credit card
spreads which are computed for single years and 1983 is substituted for 1989). Prime rate volatility is similar
across the time periods, though rates are a little more volatile from 1997 to 1998 then in the previous periods. For
rst mortgages, only 30-year xed-rate mortgages are used.

ARTICLE IN PRESS
2290

W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

Table 4
Default risk premium spreads

First mortgage rate*

Second mortgage rate*
Auto loan rate*
General consumer loan rate
Credit card rate*
Education loan rate

Pre-1995 risk premium spread

Post-1995 risk premium spread

0.50
0.98
1.08
1.19
0.53**
0.03**

0.98
3.97
1.94
1.08
1.30
0.41**

*Difference is signicant at a 95% condence level.

** Spread is insignicantly different from zero.

Table 4 shows the default premium spreads for pre-1995 and post-1995.18 For the three
types of secured loans, spreads at least nearly double over the sample, with the difference in
the spreads signicant at the 95% condence level in each case. The results are mixed for
unsecured loans. The spread is positive but unchanged over the sample for general
consumer loans, positive and statistically signicant for credit card loans only post-1995,
and not statistically signicant for education loans before or after 1995.19
These results are quite robust. Default premium spreads were calculated using actual
delinquencies and then using slightly different cutoff years. Overall, spreads were a little
smaller in the alternative models but still showed the same changes over time as in the base
model.20 In addition, the models were estimated using only conditional bankruptcy and
delinquency predictions, in turn. While these models generally reect the base models
results, there is value-added from using both measures of default risk. For example,
without delinquency risk, only the rst mortgage spread is truly consistent with the base
model. For example, the second mortgage spread is insignicant post-1995. In addition,
the automobile loan spread does not change over time, and the credit card spread is
signicantly negative pre-1995.
Fig. 1 shows these results graphically for rst mortgages, automobile loans, and credit
card loans. An interest rate function is plotted against conditional bankruptcy risk for preand post-1995 loan origination dates. For each loan type, interest rates are predicted by the
signicant measures of default risk, and other signicant variables are set to their mean
values for the entire sample period. The effects of year dummies pre- and post-1995 are
averaged so that the predicted zero default interest rate reects the average discount rate
over the period being considered. In total, 90% condence bands are also reported.
Consistent with Table 4, the slopes are steeper in the post-1995 period, indicating that the
default risk premium increased. A comparable gure for second mortgages is consistent
18

Overall, the explanatory variables have the expected signs and are generally signicant. A more detailed
discussion can be found in Edelberg (2003).
19
For credit card loans, state usury laws may have constrained credit card rates more than other, generally
lower, consumer loan rates in 1983. These laws were rendered ineffective by a 1978 Supreme Court decision, but
there is evidence that lenders may have been slow to adapt. Using only 1995 and 1998, the premium is 0.70% in
1995, and it signicantly rises to 1.24% in 1998.
20
For mortgages, the inclusion of maturity, xed versus exible interest rates or FHA loan guarantees does not
signicantly alter the basic default risk premium spread results.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2291

1st mortgage interest rate

11

10

Credit card loan interest rate

Automobile loan interest rate

7
0

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

16

14

12

10

20
18
16
14
12

post 1995

pre 1995

post 1995 90% CI

pre 1995 90% CI

Fig. 1. Interest rates by bankruptcy risk.

with those for rst mortgages and automobile loans. For other consumer loans, the interest
rate function is the same in both periods. The gure for education loans is similar to the
credit card loan gures, as it shows a at interest rate function pre-1995 and an upward
sloping function post-1995.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2292

The change in the slopes can be summarized by measuring how much interest rates
change with an increase of 0.01 in bankruptcy risk. This change more than doubles for rst
mortgages, going from 0.16 to 0.38 basis points.21 The change is up nearly ve times for
second mortgages and more than doubles for automobile loans. There is no change in the
slope of the interest rate curve for general consumer loans. Credit card and education loans
go from zero slopes to changes in interest rates of 0.48 and 0.30, respectively.
The results for secured loans support the hypothesis that lenders increasingly used riskbased pricing after 1995. For unsecured loans, credit card loans are the most robustly
consistent with the hypothesis. Two potential reasons may have led to this negative result for
education and general consumer loans. First, of the three unsecured loan types, credit card
loans have the highest incidence of loan securitization. As mentioned above, the secondary
market for loans may motivate risk-based pricing. Perhaps, lenders of education and other
consumer loans have yet to feel the pressures that led to risk-based pricing. Second, as lenders
keep better track of borrowers at risk of imminent bankruptcy, default losses may fall as
lenders are more aggressive in obtaining partial payments and fees (Winton, 1998). These
falling default costs would offset the forces driving the increased use of risk-based pricing.
5. Implications for borrowing
If lenders declined to charge very high-risk households sufciently high interest rates
before the mid-1990s, lending to this group may have proved signicantly unprotable,
and these households may have been rationed out of the market (Bostic, 2002). With riskbased pricing, lenders should offer these households debt with higher interest rates rather
than reject them. If at least some of these borrowers have sufciently high reservation
rates, debt use among very high-risk households should rise. Debt levels should also
change in reaction to risk-based pricing. Before the mid-1990s, low-risk borrowers paid
relatively higher rates than their default risk justied, and high-risk borrowers paid lower
rates. As premiums adjusted to better reect risk, debt levels among low-risk households
should have increased more (or decreased less) than levels for high-risk households.22 The
following selection model is used to estimate these effects across households:
_

lnB g0 g1 O95

3
X

gi2 f ic

i1

PrB40 f b0 b1 Y 95

3
X

gi3 O95 f ic u,

i1
3
X
i1

bi2 f iuc

3
X

!
bi3 Y 95

f iuc

yA ,

i1

21
The renancing boom after 1995 should not be driving the mortgage interest rate results. Loans are compared
by origination year, whether for purchase or renancing. Still, renancing booms may mean that borrowers who
receive bad shocks cannot renance to the new low rates. This may lead to less mortgage rate variation pre-1995
(i.e. if only high-risk households hold old loans), but not less variation as a function of risk.
In addition, mortgage interest rate results are not simply due to the addition of a subprime market, with little
spread within the prime and subprime markets. Instead of the implied bimodal distribution, we see a smooth bellshaped distribution of mortgage rates post-1995. For example, 50% of the post-1995 rates are between 7% and
8.5%, 20% are between 6% and 7%, and 20% are between 8% and 9%.
22
The change in the overall level of interest rates over time has a direct effect. For example, interest rates fell for
all credit card loans, and all risk classes increased credit card borrowing. However, interest rates fell more for lowrisk borrowers, and this is where we can see the effect of risk-based pricing.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2293

where B is the debt level for the various consumer loan types, in 1998 dollars, and A is the vector
of attitudinal variables. Accounting for changes in the cost of funds, O95 indicates if the loan
origination year is 1995 or later, and Y95 indicates if the survey year is 1995 or later. The third
degree polynomial in bankruptcy risk allows debt use and levels to vary with default risk, and
the interaction terms measure how debt use and levels changed across risk classes over time.
Fig. 2 shows predicted debt use and Fig. 3 shows predicted debt levels with 90% condence
bands. As the top panel of Fig. 2 shows, the very high-risk households have a higher
probability of holding a rst mortgage after 1995. Low-risk households also have a higher
probability of holding rst mortgages over time, perhaps as rates fell below some of their
reservation rates. Conversely, higher interest rates for high-risk groups lower the probability
of rst mortgage use for this group. Consistent with these effects, the increases in mortgage
levels after 1995 are predicted to fall with default risk, shown in the top panel of Fig. 2.
The increase in the use of automobile loans and credit cards loans is similar to that for rst
mortgages, as shown in the lower panels of Fig. 2. However, for both loan types, the
condence bands are wider, particularly as risk increases. In addition, the predicted debt levels
for automobile loans are quite consistent with the hypothesis, shown in the middle panel of
Fig. 3. Indeed, high-risk households (as opposed to very high-risk households), which saw no
signicant increase in access but saw relative borrowing costs rise, are predicted to hold
signicantly lower levels of automobile debt post-1995. The overall increase in popularity of
credit card borrowing overwhelms the effects of risk-based pricing, and all credit card
borrowers are predicted to increase debt levels after 1995, shown in the lower panel of Fig. 3.
Equivalent gures for the other debt types are not shown. Second mortgages are only
held by households with rst mortgages, making results on its use less informative. Other
consumer loans and education loans showed no signicant increases in their premium
spreads, so there is little reason for the hypothesis to hold in these cases, and indeed no
consistent story emerges from the graphs. In addition, a number of aggregate debt
categories were considered but are also not shown, though they bear out the hypothesis.
For example, very high-risk households have a higher probability of holding any debt,
post-1995. Low-risk borrowers increase total debt levels more than high-risk borrowers do,
and in some cases very high-risk borrowers actually decrease borrowing levels. Finally,
consistent with interest rates falling below or rising above reservation rates, low-risk (highrisk) households have a higher (lower) probability of holding any form of debt.
6. Access to debt markets
The increase in the use of debt and debt levels in the 1990s has been the subject of much
popular discussion. To isolate the role of risk-based pricing, a counterfactual of a pre-1995
world with the increased use of risk-based pricing is estimated by the following model for
the various types of consumer debt, where B is the borrowing level for each household:
4
4
X
X
^ g0 g1 O95
lnB
gi2 Q5i
gi3 O95 Q5i u,
i1

PrB40 f b0 b1 Y 95

i1
10
X
i2

bi2 Q10i

10
X

!
bi3 Y 95 Q10i

yA .

i2

For this analysis, default risk quantiles replace continuous measures of risk: (Q5)i
represents the ith of ve quantiles, and (Q10)i represents the ith of ten quantiles. Risk is

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

Probability of Credit Card Debt

Probability of Automobile Debt

Probability of 1st Mortgage Debt

2294

1
0.8
0.6
0.4
0.2
0
0

0.01

0.02
0.03
0.04
Unconditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Unconditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Unconditional Bankruptcy Probability

0.05

0.5
0.4
0.3
0.2
0.1
0

1
0.8
0.6
0.4
0.2

post 1995

pre 1995

post 1995 90% CI

pre 1995 90% CI

Fig. 2. Predicted debt use by bankruptcy risk.

measured this way since the many households with near-zero default risk should not be
represented by zero. Using zero would obscure the effects of any changes in the coefcients
for this risk group.23
23

The preceding sections analysis on the heterogeneous effects of risk-based pricing across risk groups suggests
which risk groups should be excluded in order to identify this model. Robustness checks show that the choices

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2295

Mortgage Debt Levels

200.000
150.000
100.000
50.000
0
0

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

0.01

0.02
0.03
0.04
Conditional Bankruptcy Probability

0.05

Automobile Debt Levels

20.000

15.000

10.000

Credit Card Debt Levels

5.000

2.500
2.000
1.500
1.000
500

post 1995

pre 1995

post 1995 90% CI

pre 1995 90% CI

Fig. 3. Predicted debt levels by bankruptcy risk.

Fig. 4 plots the predicted changes for borrowing levels and debt use for rst mortgages
and all debt for a pre-1995 world with and without the increased use of risk-based pricing.
Probability of debt use is plotted against bankruptcy risk, whereas predicted debt levels are
(footnote continued)
made are quite reasonable. The selection equation uses additional risk quantiles to better estimate increased debt
use among the very high-risk groups. The tenth quantile is even further divided into four ner divisions of risk.

ARTICLE IN PRESS
2296

W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

Fig. 4. Effects of risk-based pricing.

plotted against bankruptcy risk quantiles. Quantiles are used as the signicant changes
occur for households in the rst quantile, which have nearly zero variation in bankruptcy
risk.24
24

These condence intervals reect the prediction error in the coefcients and not the error associated with the
residual. These plots do not represent genuine forecasts of levels and use of debt, only the levels and use predicted
by risk-based pricing as summarized by the coefcients. Including the error associated with the residual generally
makes the condence intervals so large as to include pre- and post-1995 point estimates.

ARTICLE IN PRESS
W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

2297

Allowing for the increased use of risk-based pricing in a pre-1995 world predicts one- to
three-quarters of the actual increases in debt levels seen across the 1990s. For example, the
model predicts that risk-based pricing would have added over $7,000 to the average
mortgage amount excluding any economy-wide changes (in 1998 dollars). Actual
mortgages originated after 1995 versus those originated before 1995 increased about
$30,000. Similarly for automobile loans, the model predicts an increase of nearly $1500 in
the average loan size, whereas actual automobile loans increased over $2000. (Figures for
automobile loans are not shown for brevity, but can be seen in Edelberg (2003).) The
model predicts an increase in the average debt burden for households with debt of nearly
$6000 over the mid-1990s. The actual average rose about $14,000.
The model over-predicts the increased use of debt. For rst mortgages, the model
predicts an increase of nearly 8 percentage points of households holding mortgages from
before 1995 to after 1995. The actual increase was 3 percentage points in the SCF. For
automobile loans, the model predicts an increase of nearly 0.5 percentage point of
households holding automobile loans, and the actual increase was only 0.1 percentage
point. Note that the highest risk group saw much larger changes. For these households, the
model predicts an increase of 3.2 percentage points in those holding automobile loans. The
actual increase was 2.6 percentage points. For all debt, the model predicts an increase of
almost 7 percentage points in the number of borrowers, and the actual increase was 2
percentage points.25
7. Conclusion
Lenders increasingly used risk-based pricing of interest rates in consumer loan markets
during the mid-1990s. Risk premium spreads for secured loans rose over time by a
signicant amount. The case for unsecured loans is less clear. The premium spread for
credit card loans more than doubled, but education loan and other consumer loan
premiums are statistically unchanged. The evidence suggests that variations over time in
households debt levels and use of debt instruments are consistent with this change in
pricing practices. For example, while very high-risk and very low-risk households have
beneted from these changes, high-risk households have seen their relative premiums
increase and have changed their borrowing in response.
References
Ausubel, L., 1991. The Failure of Competition in the Credit Card Market, American Economic Review, 81(1).
Black, S.E., Morgan, D.P., 1999. Meet the New Borrowers. Current Issues in Economics and Finance. Federal
Reserve Bank of New York.
Bostic, R.W., 2002. Trends in equal access to credit products. In: Durkin, T., Staten, M. (Eds.), The Impact of
Public Policy on Consumer Credit. Kluwer Academic Publishers, Massachusetts, pp. 171202.
Canner, G., Passmore, W., 1997. The Community Reinvestment Act and the Protability of Mortgage-Oriented
Banks, Finance and Economics Discussion Series. Federal Reserve Board, Washington, DC.
Deaton, A., 1997. The Analysis of Household Surveys. Johns Hopkins University Press, Baltimore, MD, World
Bank Publication.
25

Other changes that were similarly heterogeneous across risk-classes could also account for the models
predictions. For example, if attitudes towards risk changed in the same manner that risk-based pricing affected
various risk groupsif only the lowest risk groups became exogenously more amenable to high debt levels over
the time periodsuch changes could not be rejected as alternative reasons for the predictions.

ARTICLE IN PRESS
2298

W. Edelberg / Journal of Monetary Economics 53 (2006) 22832298

Domowitz, I., Sartain, R.L., 1999. Determinants of the consumer bankruptcy decision. Journal of Finance 45 (1),
403420.
Dubey, P., Geanakoplos, J., Shubik, M., 2003. Default and punishment in general equilibrium. Cowles
Foundation Discussion Paper, 1304RRR.
Duca, J.V., Rosenthal, S.S., 1993. Borrowing constraints, household debt, and racial discrimination in loan
markets. Journal of Financial Intermediation 3, 77103.
Edelberg, W., 2002. Testing for racial discrimination in consumer loan markets. Working Paper.
Edelberg, W., 2003. Risk-based Pricing of Interest Rates in Household Loan Markets, Finance and Economics
Discussion Series 200362. Board of Governors of the Federal Reserve System, Washington.
Fay, S., Hurst, E., White, M., 2002. The bankruptcy decision. The American Economic Review 92 (3), 706718.
Freeman and Hamilton, 2002. A dream deferred or realized. In: AEA Proceedings, vol. 92(2).
Geanakoplos, J., 2002. Liquidity, default and crashes: endogenous contracts in general equilibrium. Cowles
Foundation Discussion Paper, 1316RR, June.
Gross, D., Souleles, N., 1999. Explaining the increase in bankruptcy and delinquency: stigma versus riskcomposition. Working Paper.
Gross, D., Souleles, N., 2001. Do liquidity constraints and interest rates matter for consumer behavior? Evidence
from credit card data. National Bureau of Economic Research, Working Paper 8314.
Jappelli, T., Pischke, J.-S., Souleles, N., 1998. Testing for liquidity constraints in Euler equations with
complementary data sources. Review of Economics and Statistics 80, 251262.
Johnson, R.W., 1992. Legal, social and economic issues in implementing scoring in the United States. In: Crook,
J.N., Edelman, D.B., Thomas, L.C. (Eds.), Credit Scoring and Credit Control: Based on the Proceedings of a
Conference on Credit Scoring and Credit Control, Organized by the Institute of Mathematics and Its
Applications and held at the University of Edinburgh in August 1989. Oxford University Press, New York.
McCorkell, P.L., 2002. The impact of credit scoring and automated underwriting on credit availability. In:
Durkin, T., Staten, M. (Eds.), The Impact of Public Policy on Consumer Credit. Kluwer Academic Publishers,
Massachusetts, pp. 209219.
Murphy, K., Topel, R., 1985. Estimation and inference in two-step econometric models. Journal of Business and
Economic Statistics 3 (4), 8897.
Riley, J.G., 1987. Credit rationing: a further remark. The American Economic Review 77 (1), 224227.
Sullivan, T.A., Warren, E., Westbrook, J.L., 1989. As We Forgive Our Debtors: Bankruptcy and Consumer
Credit in America. Oxford University Press, New York.
Sullivan, T.A., Warren, E., Westbrook, J.L., 2000. The Fragile Middle Class: Americans in Debt. Yale University
Press, Connecticut.
Winton, E.W., 1998. Games Creditors Play: Collecting from Overextended Consumers. Carolina Academic Press,
North Carolina.