CHAPTER

C OII\TTE GRATI ON AI\TD ERROR-

CORRECTIOT{ MODELS

This chapter explores an exciting new development in econometrics: the estimation

of a structural equation or a VAR containing nonstationary variables. In univariate
models, we have seen that a stochastic trend can be removed by differencing. The
resulting stationary series can be estimated using univariate Box-Jenkins techniques.
At one time, the conventional wisdom was to generalize this idea and difference all
nonstationary variables used in a regression analysis. However, the appropriate way
to treat nonstationary variables is not so straightforward in a multivariate context. It
is quite possible for there to be a linear combination of integrated variables that is sta-
tionary; such variables are said to be cointegrated. Many economic models entail
such cointegrating relationships. This chapter has three aims:
1. Introduce the basic concept ofcointegration and show that it applies in a
variety of economic models. Any equilibrium relationship among a set of
nonstationary variables implies that their stochastic trends must be linked.
After all, the equilibrium relationship means that the variables cannot
move independently of each other. This linkage among the stochastic
trends necessitates that the variables be cointegrated.
2. Consider the dynamic paths of cointegrated variables. Since the trends of
cointegrated variables are linked, the dynamic paths of such variables must
bear some relation to the current deviation from the equilibrium relation-
ship. This connection between the change in a variable and the deviation
from equilibrium is examined in detail. It is shown that the dl,namics of a
cointegrated system are such that the conventional wisdom was incorrect.
After all, if the linear relationship is already stationary, differencing the
relationship entails a misspecifi cation error.
3. Study the alternative ways to test for cointegration. The econometric meth-
ods underlying the test procedures stem from the theory of simultaneous
difference equations. The theory is explained and used to develop the two
most popular cointegration tests. The proper way to estimate a system of
cointegrated variables is examined. Several illustrations ofeach testing
methodology are provided. Moreover, the two methods are compared by
applying each to the same data set.

319
ENmt2sepda

32O cHAprER 6 coTNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS
1. LINEAR COMBINATIONS OF
INTEGRATED VARIABLES
Since money demand studies stimulated much of the cointegration literature, we
begin by considering a simple model of money demand. Theory suggests that indi-
viduals want to hold a real quantity of money balances, so that the demand for nom-
inal money holdings should be proportional to the price level. Moreover, as real
income and the associated number of transactions increase, individuals will want to
hold increased money balances. Finally, since the interest rate is the opportunity cost
of holding money, money demand should be negatively related to the interest rate. In
logarithms, an econometric specification for such an equation can be written as
mt: 0o+ /rpt+ $y,+ Pjr,+ e, (6.1)
where: mr: demand for money
pr: price leve1
yr: real income
rr: interest rate
er: stationary disturbance term
Bi-- parameters to be estirnated
and all variables but the interest rate are expressed in logarithms.
The hypothesis that the money market is in equilibrium allows the researcher to col-
lect time-series data of the money supply (: money demand if the money market always
clears), the price level, real income (possibly measured using real GDP), and an appro-
priate short-term interest rate. The behavioral assumptions require that p1: l, &., 0, *d
&. O; a researcher conducting such a study would certainly want to test these parame-
ter restrictions. Be aware that the properties of the unexplained portion of the demand for
money (i.e., the {er} sequence) are an integral part of the theory. If the ttreory is to make
any sense at all, any deviation in the demand for money must necessarily be ternporary
in nature. Clearly, if e, has a stochastic trend, the errors in the model will be cumulative
so that deviations from money market equilibrium will not be eliminated. Hence, a key
assumption of the theory is that the {er} sequence is stationary.
The problem confronting the researcher is that real GDP, the money supply, price
level, and interest rate can all be characterized as nonstationary (l) variables. As
such, each variable can meander without any tendency to return to a long-run level.
However, the theory expressed in (6.1) asserts that there exists a linear combination
of these nonstationary variables that is stationary! Solving for the error term, we can
rewrite (6.1) as
r: ffit- 0o- /tpt- hy,- &rt (6.2)
Since {er} must bb stationary, it follows that the linear combination of integrated
variables given by the right-hand side of (6.2) must also be stationary. Thus, the the-
ory necessitates that the time paths of the four nonstationary variables {m,}, {pr},
$,,,\, and {r,} be linked. This example illustrates the crucial insight that has domi-
nated much of the macroeconometric literature in recent years: Equilibrium theories
involving nonstationary variables require the existence of a combination of the vari-
ables that is stationary.
*l
LINEAR COMBINATIONS OF INTEGRATED VARIABLES 321
I The money demand function is just one example of a stationary combination of
I
nonstationary variables. Within any equilibrium framework, the deviations from
equilibrium must be temporary. Other important economic examples involving sta-
tionary combinations of nonstationary variables include:
1. Consuntption function theory. A simple version of the permanent income
hypothesis maintains that total consumption (cr) is the sum of permanent
consumption (cf) and transitory consumption (ci). Since permanent con-
sumption is proportional to permanent income (yl), we can let Bbe the
constant of proportionality and write cr: ful + cf . Transitory consumption
is necessarily a stationary variable, and both consumption and permanent
income are reasonably characterized as (1) variables. As such, the perma-
nent income hypothesis requires that the linear combination of rwo (1)
variables given by 4- ful be stationary.
.,
Unbiasedfonuard rate hypothesis. One form of the effrcient market hypothe-
sis asserts that the forward (or futures) price of an asset should equal the
expected value of that asset's spot price in the future. The argument that the
foreign exchange market is efficient requires that the one-period forward
exchange rate equal the expectation of the spot rate in the next period. Letting
f, derote the log ofthe one-period price offorward exchange in / and s, the log
fr.If
of the spot price of foreign exchange in r, the theory asserts that E,s ,*1 --
this relationship fails, speculators can expect to make a pure profit on their
trades in the foreign exchange market. Ifagents'expectations are rational, the
forecast error for the spot rate in / + 1 will have a conditional mean equal to
zero, so that sral Ets t+l: 6y+1 where Erpt:0. Combining the two equa-
-
tions yieldss*t:ft* 6r+1. Since {s,} and {f,) are (1) variables, the unbiased
forward rate hypothesis necessitates that there be a linear combination of
nonstationary spot and forward exchange rates that is stationary.
3. Commodity market arbitrage and purchasing power parity. Theories of spa-
tial competition suggest that in the short run, prices of similar products in
varied markets might differ. However, arbiters will prevent the various prices
from moving too far apart even if the prices are nonstationary. Similarly, the
prices of Apple computers and PCs have exhibited sustained declines.
Economic theory suggests that these simultaneous declines are related to
each other since a price discrepancy between these similar products cannot
continually widen. Also, as we saw in Chapter 4, purchasing power parity
places restrictions on the movements of nonstationary price levels and
exchange rates. Ife, denotes the log ofthe price offoreign exchange andp,
arrd p7 denote, respectively, the logs of domestic and foreign price levels,
long-run PPP requires that the linear combination e, + py - pt be stationary.
All of these examples illustrate the concept of cointegration as introduced by
Engle and Granger (1987). Their formal analysis begins by considering a set of eco-
nomic variables in long-run equilibrium when
/fiu + 0f,2, + ... + Bnx,,: 0
322 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRBEcloN MoDELS
Letting B and x, denote the vectors (frv &., ..., B) and (xy, x2t, ...,xrr)', the sys-
tem is in long-run equilibrium when Bxr: 0. The deviation from long-run equilib-
rium----called the equilibrium error-is el, so that
et: &t
If
the equilibrium is meaningful, it must be the case that the equilibrium error
process is stationary. In a sense, the use of the term equilibrium is unforlunate because
economic theorists and econometricians use the term in different ways. Economic the-
orists usually use the term to refer to an equality between desired and actual transac-
tions. The econometric use of the term makes reference to any long-run relationship
among nonstationary variables. Cointegration does not require that the long-run rela-
tionship be generated by market forces or by the behavioral rules of individuals. In
Engle and Granger's use of the term, the equilibrium relationship may be causal,
behavioral, or simply a reduced-form relationship among simitarly trending variables.
Engle and Granger (1987) provide the following definition of cointegration:
The components of the vector ,/ : (xb xzt, . . ., xn)' are said to be cointegrated
oforder d, b, denotedby *,- CI(d, b) if
1. All components of x, are integrated of order d
2. There exists a vector 0: (0y h, ..., 0) such that the linear combination
fu1: |py + /fzt+ ... + Bnxr, is integrated of order (d - b)where b > 0.
The vector pis called the cointegrating vector.l
In terms of equation (6.1), if the money supply, price level, real income, and
interest rate are all (1) and the linear combinationmt- 0o- /flt-
&.yr- p3rr: e,is
stationary, then the variables are cointegrated of order (1, 1). The vector*, is (m1, 1,
po lo rt)' and the cointegrating vector B is (1,
-00, -4, -02, -ft). The deviation from
long-run money market equilibrium is er; since {er} is stationary, this deviation is
temporary in nature.
There are four important points to note about the definition:
1. Cointegration refers to a linear combination of nonstationary variables.
Theoretically, it is quite possible that nonlinear long-run relationships exist
among a set of integrated variables. However, the current state of econo-
metric practice is just beginning to allow for tests of nonlinear cointegrat-
ing relationships. Also note that the cointegrating vector is not unique. If
(4, h, ..., 0) is a cointegrating vector, then for any nonzero value of ),
()fi, )&., ..., is also a cointegrating vector. Typically, one of the
^P")
variables is used to normalize the cointegrating vector by fixing its coeffi-
cient at unity. To normalize the cointegrating vector with respect to x1r,
simplyselect):1/0r
, From Engle and Granger's original definition, cointegration refers to vari-
ables that are integrated of the same order. Of course, this does not imply
that all integrated variables are cointegrated; usually, a set of(@ variables
is nor cointegrated. Such a lack ofcointegration implies no long-run equi-
librium among the variables, so that they can wander arbitrarily far from
each other. Iftwo variables are integrated ofdifferent orders, they cannot
LINEAR COMBINATIONS OF INTEGRATEDVARIABLES 323
be cointegrated. Suppose x1,is I(d) ar,dx2,is I(d) where d2> d1.
Question 6 at the end of this chapter asks you to prove that any linear com-
bination of x1, and x2, is I(d2).
Nevertheless, it is possible to find equilibrium relationships alnong
groups ofvariables that are integrated ofdifferent orders. Suppose thatxl,
and x2, are I(2) and that the other variables under consideration are {1). As
such, there cannot be a cointegrating relationship between x1, (or x2) arrd
x3l. However, if x1, and x21 arta CI(2,1), there exists a linear combination of
the form !fitt+ fux2,which is (1). It is possible that this combination of
x1, and x2, is cointegrated with the (1) variables. Lee and Granger (1990)
use the term multicointegration to refer to this type of circumstance.
3. If x, has n nonstationary components, there may be as many as n - 1 lin-
early independent cointegrating vectors. Clearly, if x, contains only two
variables, there can be at most one independent cointegrating vector. The
number ofcointegrating vectors is called the cointegrating rank ofxr. For
example, suppose that the monetary authorities followed a feedback rule
such that they decreased the money supply when nominal GDp was high
and increased the nominal money supply when nominal GDp was low
This feedback rule might be represented by
mt: yt(lt + p) + eu
: 7o-- th- l Pt* ett
7o
(6.3)
where {e1r} : a stationary error in the money supply feedback rule"
Given the money demand function in (6.1), there are two cointegrating
vectors for the money supply, price level, real income, and the interest rate.
Let Abe the (5 ' 2) matrix:
-lr -gs -gt
, -1, -gz -gt
" -10 tt 1r o
The two linear combinations given by fu, are stationary. As such, the
cointegrating rank of r, is 2. As a practical matter, if multiple cointegrating
vectors are found, it may not be possible to identi$z the behavioral rela-
tionships from what may be reduced-form relationships.
4. Most of the cointegration literature focuses on the case in which each variable
contains a single unit root. The reason is that traditional regression or time-
series analysis applies when variables are (0) and few economic variables are
integrated of an order higher than unity.2 When it is unambiguous, mary
authors use the term cointegrationto refer to the case in which variables are
CI(|, 1). The remainder of the text follows this convention" Of course, rnany
other possibilities arise. For example, a set of (2) variables may be cointe-
grated of order CI(2, 1) so that there exists a linear combination which is (1).
Worksheet 6.1 illustrates some of the important properties of cointegration rela-
tionships. In Case l, both the 1,,,\ and {zr} sequences were constructed so as to be
random walk plus noise processes. Although the 20 realizations shown generally
324 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRRECTIoN MoDELS COINTEGRATIONANDCOMMONTRENDS 325

in the second graph, there exists a stationary linear combination ofthe three such that
e1 : lt * zt - wr Thus, it follows that the dynamic behavior of at least one variable
ILLU S TRATIN G C OINTE GRATE D SYS TE MS must be restricted by the values of the other variables in the system.
Figure 6.1 displays the information of Case 1 in a scatter plot of {yr} against the
CASE 1: The series {tt} is a random walk process and {e;,r} and {e}
are white noise. associated value of {2,\; each of the 20 points represents the ordered pairs (y1, z), (!2,
Hence, the lv1) and {21} sequences are both random walk plus noise processes. Although zz), . . ., (!zo, zzd. Comparing Worksheet 6. I and Figure 6. 1 , you can see that low val-
each is nonstationary, the two sequences have the same stochastic trend; hence they are
ues in the fyr) sequence are associated with low values in lhe {2,} sequence and that
cointegrated such that the linear combination Qt, - z,) is stationary. The equilibrium error
values near zero irt one series are associated with values rrear zero in the other. Since
term (9, +r) is an .I(0) process.
- both series move together over time, there is a positive relationship between the two.
Itt + yt zt = + zt The equilibriu m ertoti yt - zt
Yt = l-Lt The least-squares line in the scatter plot reveals this to be a strong positive association.
0 10 t5 20 In fact, this line is the "long-run" equilibrium relationship between the series, and the
0r deviations frorn the liue are the stationary deviations from long-run equilibrium.
-1 For comparison purposes, Panel (a) in Worksheet 6.2 shows the time paths of
two random walk plus noise processes that are not cointegrated. Each seems to mean-
-Yt '-2
--- zt
der without any tendency to approach the other. The scatter plot shown in Panel (b)
conhrms the irnpression of no long-run relationship between the variables. The devi-
-3
ations from the straight line showing the regression of z, on y, are substantial. Plotting
-4 the regression residuals against time, [see Panel (c)] suggests that the regression
t residuals are not stationary.
CASE 2: A11 three sequences are random walk plus noise processes. As constructed, no
two are cointegrated. However, the linear combination $t, + z1- w) is stationary; hence, 2. GOINTEGRATION AND COMMON TRENDS
the three variables are cointegrated. The equ.ilibrium error is an (0) process.
yt+ Stock and Watson's (1988) observation that cointegrated variables share common sto-
Yt= layt* y7, Zt= [Lzt+ e I W1= Fwt* ewt The equilibrium error: 4- wt
chastic trends provides a very useful way to understand cointegration relationships.3

---wt -0.5
--- yt
N -1
o
3 -1.5
6

decline, extending the sample would eliminate this tendency. In any event, neither
series shows any tendency to return to a long-run level, and formal Dickey-Fuller
tests are not able to reject the null hypothesis of a unit root in either series. Although
each series is nonstationary you can see that they do move together. In fact, the dif- -3
-3 -2.5 -2 -1.5 -1 -0.5 0
ference between the series $t, - z,)-shown in the second graph-is stationary; the Values of y
equilibriunt error term et : (!t - zr) has a zero mear and a constant variance. The scatter plot was drawn using the {h} and {2,}
Case 2 illustrates cointegration among three random walk plus noise processes. sequences from Case 1 ofWorksheet 6.1. Since
As in Case 1, no series exhibits a tendency to retum to a long-run level, and formal both series decline over time, there appears to
Dickey-Fuller tests are not able to reject the null hypothesis of a unit root in any of be a positive relationship between the two.The
equilibrium regression line is shown.
the three. In contrast to the previous case, no two ofthe series appear to be cointe-
grated; each series seems to "meander" away from the other two. However, as shown FIGURE 6.1 Scatter Plot of Cointegrated Variables
326 cHAprER 6 corNrEGRATroN AND ERRoR-coRREcroN MoDELS COINTEGRATIONANDCOMMONTRENDS 327

If {yr} and {zr} arc cointegrated of order (1, 1), there must be nonzero values of
01 and B2for which the linear combination llt+ hzt is stationary. Consider the sum
N ONCOINTERATED VA NABLES
?tt+ 0*t: 0{t\,,+ "y,) + hQ",t+ e,)
The {yr} and {2,} sequences are constlucted as independent random walk plus noise : (?t\,t+ &.t + (Ofyt+ &"o)
) (6.6)
processes. There is no cointegrating relationship between the two variables. As shown
For Bp,+ p2z,to be stationary, the term (A$,t+ 821l,r,) m:ust vanish. After all, if
in Panel (a), both seem to meander without any tendency to come together. Panel (6)
either of the two trends appears in (6.6), the linear combination 0h+ /Ft will also
shows the scatter plot of the two sequences and the regression line z1 : h + &yt.
have a trend. Since the second term in parentheses is stationary, the necessary and
However, this regression line is spurious. As shown in Panel (c), the regression residu-
als are nonstationary.
sufficient condition for {y,} and {2,} to be C(1, 1) is
Yt= l-tyt+ er, zt= lJ,zt + e4 Regression oI zron y, 1t4y+ 0z.l1,t:0 (6.7)
4 Clearly, 15ratd /-tt are variables whose realized values will be continually chang-
ing over time. Since we preclude both fiand B2 from being equal to zero, it follows
2 that(6.7) holds for all I if and only if
/5t: -hh/ 0t
--- yt For nonzero values of B1 and the only way to ensure the equality is for the
P"z,
stochastic trends to be identical up to a scalar. Thus, up to the scalar -p2/p1, two I(1)
-2 stochastic processes $t,\ and {2,} must have the same stochastic trend if they are
cointegrated of order (1, 1).
-4 Retum your attention to Worksheet 6.1. In Case 1, the b/) and {2,} sequences
t -4 were constructed so as to satisfy
Panel (a)

Regression residuals
!t: Ht+ yt
zt: Ht * zt
and /4: llt-t * t
where ero zh and e, are independently distributed white-noise disturbances.
By construction, 7-2, is a pure random walk process representing the same stochas-
tic trend for both the $) and {2,} sequences. The value of p6 was initialized to zero,
and three sets of 20 random numbers were drawn to represent the {er}, {er,}, and {er)
sequences. Using these realizations and the initial value of pn, the {yr} , {rr} , and {pr,}
sequences were constructed. As you can clearly determine, subtracting the realized
value ofz, fromy, results in a stationary sequence:
lt - zt: g, + 04 + ,) : 1,t - rt
ey,) -
Panel (c)
To state the point using Engle and Granger's terminology, multiplying the cointe-
grating yector B: (1, -1) by the vector h: (!o zr)'yields the stationary sequence 6/:
\,t - zr Indeed, the equilibrium elror term shown in the second graph of Worksheet
6.1 has all the hallmarks of a stationaryprocess. The essential insight of Stock and
For ease of exposition, retum to the case in which the vector x, contains only two vari-
Watson (1988) is that the parameters of the cointegrating vector must be such that they
ables so that x,: (!o z)'.Ignoring cyclical and seasonal terms, we can write each vari-
purge the trend from the linear combination. Any other linear combination of the two
able as a random walk plus an irregular (but not necessarily a white noise).component:
variables contains a trend so that the cointegrating vector is unique up to a normaliz-
!t: lt7rt+ eYt (6.4) ing scalar. Hence, &yt+ Qqzt cannot be stationary lurtless B3lpa: P1/&.
Recall that Case 2 illustrates cointegration between three random walk plus
21: l1r7* r1 (6 5)
noise processes. Each process is (1), and Dickey-Fuller unit root tests would not be
where: p,i,: a rardom walk process representing the trend in variable i able to reject the null hypothesis that each contains a unit root. As you can see in the
ei,: the stationary (irregular) component of variable i lower portion of Worksheet 6.1, no pairwise combination of the series appears to be
328 cHAprER 6 cotNTEGRATIoN AND ERRoR-coRREcrloN MoDELS
cointegrated. Each series seems to meander, but, as opposed to Case 1, no one sin-
gle series appears to remain close to any other series. However, by construction, the
trend in w, is the simple summation of the trends in y, and zr:
tt*t: /\,t+ llrt
Here, the vector.x/ : (!o zt, wr)' has the cointegrating vector (1, 1, -1),
linear combinationy,l zt-wt is stationary. Consider:
rt * zt - *, -9,:.rnl:@,, + e,) - (w, + e*)
'yl -zt '-wt
This example illustrates the general point that cointegration will occur u'henever
the trend in one variable can be expressed as a linear combination of the trends in the
other variable(s). In such circumstances it is always possible to find a vector / such
that the linear combination Bp, + 0z?, + fiw, does not contain a ftend. The result eas-
ily generalizes to the case ofn variables. Consider the vector representation:
xt: /at+ er
where: xr: the vector (x1u x2p ..., xn)'
1-l,: the vector of stochastic trends (p1n lzo ..., Pn)'
et: an m . 1 vector of stationary components
If one trend can be expressed as a linear combination of the other trends in the
system, it means that there exists a vector B such that
/l-ttt + Pr1ur, + '.. + p,lrrt: 0
Premultiply (6.8) by this set of B;'s to obtain
fut:0P,+ &t
Si:nce 0, it follows that fu,: &1. Hence, the linear combination Bx,is sta-
Bpr:
tionary. As shown in Section 8, this argument easily generalizes to the case of multi-
ple cointegrating vectors.
3. COINTEGRATION AND ERROR CORRECTION
A principal feature of cointegrated variables is that their time paths are influenced by
the extent of any deviation from long-run equilibrium. After all, if the system is to
retum to long-run equilibrium, the movements of at least some of the variables must
respond to the magnitude of the disequilibrium. For example, theories of the term
structure ofinterest rates imply a long-run relationship between long- and short-term
rates. Ifthe gap between the long- and short-term rates is "large" relative to the long-
run relationship, the short-term rate must ultimately rise relative to the long-term rate.
Ofcourse, the gap can be closed by (1) an increase in the short-term rate and/or a
decrease in the long-term rate, (2) an increase in the long-term rate but a commensu-
rately larger rise in the shorl-term rate, or (3) a fall in the long-term rate but a smaller
fall in the short-term rate. Without a full dynamic specification of the model, it is not
COINTEGHATION AND ERROH CORHECTION 329
possible to determine which of the possibilities will occur.'Nevertheless, the short-run
dynamics must be influenced by the deviation from the long-run relationship.
The dynamic model implied by this discussion is one of error correction. In an
error-correction model, the short-term dynamics of the variables in the system are influ-
enced by the deviation from equilibrium. If we assume that both interest rates are -I(1), a
so that the simple error-correction model that could apply to the term structure of interest rates is4
Lrsr: as(rtt-t - Qrsr-t) + es, a^s >0 (6.e)
LrLr: -at(rrt-t - Br5,-y) + ey dt> o (6.10)
where e5, and e11are white-noise disturbance terms which may be correlated, ry afid
rg7 ?t@ the long- and short-term interest rates, and e9 e1; and B are parameters"
As specified, the short- and long-term interest rates change in response to stochas-
tic shocks (represented by 6,, and e1) and in response to the previous period's devia-
tion from long-run equilibrium. Everything else being equal, if this deviation happened
to be positive (so that rLt-t - !rsr-t> 0), the short-term interest rate would rise and the
(6.8)
long-term rate would fall. Long-run equilibrium is attained when ry: /rsr
Here you can see the relationship between error-correcting models and cointe-
grated variables. By assumption, Arg, is stationary so that the left-hand side of (6.9)
is (0). For (6.9) to be sensible, the right-hand side must be (0) as well. Given that
e5, is stationary, it follows that the linear combination ry-1 - /rs,-t must also be sta-
tionary; hence, the two interest rates must be cointegrated with the cointegrating vec-
ror (1, -B). Of course, the identical argument applies to (6.10). The essential point to
note is that the error-correction representation necessitates that the two variables be
cointegrated of order CI(l,l). This result is unaltered if we formulate a more general
model by introducing the lagged changes of each rate into both equations:6
Arsr: arc+ as(rLt-t - 1rst) + Xa11(l) Ar5,-, + Dap(i) Lr1,-i* es1 (6.11)
Lrn-- a2s - a1(r71-t - grsr) + Da21(i) Ars,-, * Da22(i) L,ry-;* e17 6.12)
Again, Sr, Lt, and all terrns involving Arsl-i and L,r1,-iare stationary. Thus, the
linear combination of interest rates rLt-t - prg;1 mu t also be stationary.
Inspection of (6.11) ard (6.12) reveals a striking similarity to the VAR models of
the previous chapter. This two-variable error-correction model is a bivariate VAR in
first differences augmented by the error-correction terms as(rrrt - Brg,-1) and
-at(rt t_t - /rs,-t). Notice that ag and ayhave the interpretation of speed of adjust-
nxent parafireters. The larger a5 is, the greater the response ofto the previous period's
deviation from long-run equilibrium. At the opposite extreme, very small values of
a5 imply that the short-term interest rate is unresponsive to last period's equilibrium
error. For the {Ar5,} sequence to be unaffected by the long-term interest rate
sequence, ag and all the ap (i) coeflicients must be equal to zero. Of course, at least
one of the speed of adjustment terms in (6.11) and (6.12) must be nonzero. If both a5
and alare equal to zero, the long-run equilibrium relationship does not appear and
the model is not one of error correction or cointegration.
330 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRHECTIoN MoDELS
COINTEGRATION AND ERBOR CORRECTION 331

The result can easily be generalized to the n-variable model. Formally, the (n . 1)
vector,/ : (xv x21; ...; xn)' has an error-correction representation if it can be
expressed in the form
L,x,: rgt ,rxr1 + t1Au-7-1 + r2L,x;2+ ... + noNr-o+ e, (6.1 3)
where: To-- trt (n . 1) vector of intercept terms with elements zr;6
Ti: (n. r) coeffrcient matrices with elements ri{i)
z': a matrix with elements rr.7. such that one or more of the tri1, * 0
q : at (n . 1) vector with elements e;,
Note that the disturbance terms are such that ei, may be correlated with eit.
Let all variables in x, be (1). Now, if there is an error-correction representation
of these variables as in (6.13), there is necessarily a linear combination of the (l)
variables that is stationary. Solving (6.13) for zar_1 yields
ffit-1 : N,- r,g-Dr;Lxr-;- e,
Since each expression on the right-hand side is stationary zry_1 must also be sta-
tionary. Since z'contains only constants, each row of z' is a cointegrating vector of
x, For example, the first row can be written as (ryx1r_1 * r12x2t_t + ... + \r{nt_t).
Since each series is I(l), (r11, Ttz, .. ., z-1n) must be a cointegrating vector for xr.
The key feature in (6.13) is the presence of the matrix z-. There are two impor-
tant points to note:
1. If all elements of z'equal zero, (6.13) is a traditional VAR in first differ-
ences. In such circumstances there is no error-correction representation
since Ax, does not respond to the previous peribd's deviation from long-
run equilibrium.
2. If one or more of the rip differs from zero, A;r, responds to the previous
period's deviation from long-run equilibrium. Hence, estimating x, as a
VAR in first dffirences is inappropriate if x, has an eruor-correction repre-
sentation. The omission of the expression zxr_1 entails a misspecification
error ifx, has an error-correction representation as in (6.13).
A good way to examine the relationship between cointegration and error correc-
tion is to study the properties of the simple VAR model:
!t: attlrt + anzrt + 1,r (6.14)
zt : a2ltt * a22zr_1 * er, (6. 15)
Using Cramer's Rule or matrix inversion, we can obtain the solutions for y, and z, as
(l- a22L)ey + a12Le,,
v-= (6.16)
(l - ayL)(l - azzL) - a12a21 L2
ayLey + (1 - a1lL)e ,,
zr=@ (6.r1)
We have converted the two-variable frrst-order system represented by (6.la)
and (6.15) into two univariate second-order difference equations of the type exam-
ined in Chapter 2. Note that both variables have the same inverse characteristic equa-
tion: ( 1 - a ,L)(l - azzL) - a pa21L2. Setting (1 - a 1 {)Q - a zzL) - a pa21L2 : 0 and
solving for Z yields the two roots ofthe inverse characteristic equation. In order to
work with the characteristic roots (as opposed to the inverse characteristic roots),
define ) : llL and write the characteristic equation as
)2 - (at + azz)) + (anazz - apa21) : 0 (6.18)
Since the two variables have the same characteristic equation, the characteristic
roots of (6.18) determine the time paths of both variables. The following remarks
summarize the time paths of 5,,,\ and dz,j:
1. Ifboth characteristic roots (,\1, )2) lie inside the unit circle, (6.16) and
(6.17) yield stable solutions for {yr} and {2,}.If r is sufficiently large or if
the initial conditions are such that the homogenous solution is zero, the
stability condition guarantees that the variables are stationary. The vari-
ables cannot be cointegrated oforder (1, 1) since each is stationary.
2. If either root lies outside the unit circle, the solutions are explosive. Neither
variable is difference stationary so they cannot be C(1, l). In the same way,
ifboth characteristic roots are unity, the second difference ofeach variable
will be stationary. Since each is I(2), the variables cannot be C(1, 1)"
3. As you can see from (6.14) and (6.15), if ap: azt:0, the solution is
trivial. For {yr} and {2,} to be unit root processes, it is necessary for all:
azz: l. It follows that )1 - A2 1 and that the two variables evolve
without any long-run equilibrium relationship; hence, the variables cannot
be cointegrated.

whete ert and er, are white-noise disturbances that may be correlated with each other
4. For {y,} and {2,} to be C(l, 1.), it is necessary for one characteristic root
and, for simplicity, intercept terms have been ignored. Using lag operators, we can
to be unity and the other to be less than unity in absolute value. In this
write (6.14) and (6.15) as instance, each variable will have the same stochastic trend and the first dif-
ference of each variable will be stationary. For example, if ) 1 : I , (6. l6)
(l - ayL)y,- apLz,: 7, will have the form:
-aztLyt+ (l - a22L)2,: e,,
yt: l0 - a22L)eyt + apLe,,ltl(t - L)(t -
The next step is to solve fory, and 2,. Writing the system in matrix form, we obtain ^2L)l
or, multiplying by (l - Z), we get
[(r -airr.) - onL]lr,l_l'r,l (1- L)n: Lyt: lQ - a22L)er,+ apLe,)/(t - )zL)
| -orrL - (t a22D)lr,l- lrr) which is stationary if l,\21 < 1.
 corNTEGBAroN AND ERRoR coRREcloN 333
332 cHAprER 6 cotNTEGBATIoN AND ERBoR-coRREcrloN MoDELS

Thus, to ensure that the variables are CI(\, 1), we must set one of the character-
adjustment coefficient a, :
0. In this case, yt changes only in response to y as
istic roots equal to unity and the other to a value that is less than unity in absolute
Lyr: The {zr} sequence does all of the correction to eliminate any deviation
eyt.1

value. For the larger of the two roots to equal unity, it must be the case that
from long-run equilibrium.
To highlight some of the important implications of this simple model, we
0.5(a11 + a22)+O.5 Glr+ + 4apa21 =l have shown:
"fr\ -2a1p22 The restrictions necessary to ensure that the variables are CI(\, 7) guaran-
1.
so that after some simplification, the coefficients are seen to satis$rs tee that an error-correction model exists.In our example, both fur) and
arr : [(l - azz) - ap a2ll(l - a22) (6.1 e)
{zr} are unit root processes but the linear combinatioty,- Ar,is station-
Now consider the second characteristic root. Since ap andlot a21 must differ ary; the normalized cointegrating vector is [1, -(1 - a22)la2l. The vari-
from zero ifthe variables are cointegrated, the condition l)z I ' I requires ables have an error-correction representation with speed of adjustment
(6.20) coefficients a, : -apa21/0 - az) and o, : a21. It was also shown that an
a22> -l error-correction model for (1) variables necessarily implies cointegration.
and This finding illustrates the Granger representation theorem stating that
atza2t t (ozz)2 .1 (6.21) for any set of(l) variables, error correction and cointegration are equiva-
lent representations.
Equations (6. 1 9), (6.20), and, (6.21) are restrictions we must place on the coeffi-
) A cointegration necessitates cofficient restrictions in a VAR model. It rs
cients of (6.t4) and (6.15) if we want to ensure that the variables are cointegrated of
important to realize that a cointegrated system can be viewed as a
order (1, l). To see how these coefficient restrictions bear on the nature ofthe solu-
restricted form of a general VAR model. Let x, : (/ o z )' and e, : (q,u ,r)'
tion, write (6.14) and (6.15) as
so that we can write (6.22) in the form

=
[";:1 l^ :;' ";;'-,)l';, -i]. [ ;, ] (6.22)
Lxr: rxr;* e, (6.2s)
Clearly, it is inappropriate to estimate a VAR of cointegrated variables
Now, (6. 19) implies that ar - | : -apa21/(l - a2) so that after a bit of manip- using only first differences. Estimating (6.25) without the expressiot Mrl
ulation, (6.22) can be written in the form would eliminate the error-correction portion of the model. It is also impor-
L,y,: -lapax/0 - azz))y; + anzFt + fl (6'23) tant to note that the rows of 7T are not linearly independent if the variables
are cointegrated. Multiplying each element in row 1 by {1 - a22)lap
Lzr: a21y;1 - (1 - a22)2,-y + er, (6.24)
yields the corresponding element in row 2. Thus, the determinant of z-is
Equations (6.23) and (6.24) form an effor-correction model. If both a12arld a21 equal to zero, and y, and z,have the error-correction representation given
differ from zero, we can normalize the cointegrating vector with respect to either vari- by (6.23) and (6.24).
able. Normalizing with respect to yb we get This two-variable example illustrates the very important insights of
Ly,: oy(y,-t &,-) + 1,, Johansen (1988) and Stock and Watson (1988) thatwe can use the rank of
-
r to determine whether or not two variables.0J and {2,} are cointe-
L.zr: a"(yr_1 - &'_t) + rt grated. Compare the determinant of zrto the characteristic equation given
where; a.,:
I -a9a21l(1
\- Lr ' - azz) by (6.18). Ifthe largest characteristic root equals unity, ()1 : 1), it follows
P-- (l - a22)1a21 that the determinant of zr is zero and that z'has a rank equal to unity. If z
ar: a21 were to have a rank of zero, it would be necessary for all: \, a22: 1, and
You Jan r"" thuty, and z, change in response to the previous period's deviation alz: a2t: 0. The VAR represented by (6.14) and (6.15) would be nothing
from the long-run equi1ib.irr.., ltq - Azrt. If yt: ?zrt, lt and z, change only in more than Ly1: yt and L,z,: err. In this case, both the S,t,| and {2,}
response to en, and q, shocks. Moreover, if %,. O and ar' 9.,.!r.decreases and z, sequences are unit root processes without any cointegrating vector. Finally,
increases in response to a positive deviation frilm long-run equilibrium. if the rank of z-is fuIl, then neither characteristic root can be uniry so the
You can easily convince yourself that conditions (6.20) and (6.21) ensure that {y,} and {zr} sequences are jointly stationary.
(i.e., ay and o2) is
B = O and that at least one of the speed of adjustment parameters 3. In general, both variables in a cointegrated system will respond to a devia-
not equal to zero. Now, refer to (6.9) and (6.10); you can see this model is in exactly
presented in the beginning of this section. tion from long-run equilibrium. However, it is possible that one (but not
the same form as the interest rate example
zero, interesting special case arises both) of the speed of adjustment parameters is zero. For example, it g: 0,
Although apand a21 cvtrrotboth equal arr
if we set at2: 0, the speed of [yr] does not respond to the discrepancy from long-run equilibrium and
if one of these ioefficients ls zero. For example,
334 cHAprER 6 coTNTEGRATToN AND ERRoR-coRRECTroN MoDELS TESTTNG FoR coTNTEGRATIoN: THE ENGLE-cRANGER METHoDoLoGy 335

{2,} does all of the adjustment. In this circumstance, {y,} is said to be
weakly exogenous. As such, an econometric model for {2,} can be esti-
mated and hypothesis testing can be conducted without reference to a spe-
cific model for 1,,,\ . Section l0 and Appendix 6. 1 consider modeling in a
cointegrated system when a variable is weakly exogenous.
Also, if is necessary to reintet?ret Granger causality in a cointegrated
system. In a cointegrated system, pr) does not Granger cause {zr} if lagged
values Ayr_; do not enter the Lz, equation and if z, does not respond to the
deviation from long-run equilibrium. Hence, {zr} must be weakly exoge-
nous. If a21 '- 0 in (6.2$, {zr} is weakly exogenous and is not Granger
caused by {y,) . Similarly, in the cointegrated system of (6. I I ) and (6. l2),
Each ofthese n equations is an independent restriction on the long-run solution
of the variables; the n variables in the system face n long-run constraints. In this case,
each of the n variables contained in the vector x, must be stationary with the long-run
values given by (6.28). The variables cannot be cI(|,1) since all are stationary.
In intermediate cases, in which the rank of z- is equal to r < n, there are r coir.-
tegrating vectors. with r independent equations and, n variables, there are n r sto-
-
chastic trends in the system. rf r:1, there is a single cointegrating vector given by
any row of the matrix r. Each {&rr} sequence can be written in error-correction
form. For example, we can write Axr, as
Lrl, = Tllxlt_l * 7Tpx21_1 I ... * 1r1,rXr7_1 -t 1,

{r1,} does not Granger cause {r5r} if all a12(i) : 0 and if a5 : 0. or, normalizing with respect to we can set a1 : gy: nlilry
x17_1, zr11 and to obtain
The n-Variable Gase L*1, = a1(xy_1 + 0 + ... + p lnxtn_r ) + er, (6.2e)
t*2,_t
Little is altered in the r-variable case. The relationship between cointegration, error In the long run, the {x,,} will satisfu the relationship
correction and the rank of the matrix zr is invariant to adding additional variables to
the system. The interesting feature introduced in the n-variable case is the possibility xU + lnxZt + ... + /nxrt: O
of multiple cointegrating vectors. Now consider a more general version of (6.25): Hence, the normalized cointegrating vector is (1, fi2, gn, .. ., 0t) and the speed
xr: Alxr_1 * e, (6.26) of adjustmentparameter is o1. In the same way, with two cointegration vectors the
: long-run values of the variables will satisSr the two relationships
where: the (n . l) vector (xu, xzx ..., xrt)'
x,
er: the (n . l) vector (et, ezu ..., nt)' 1f 17xb + 1T12x21 -t ... * rhx n: 0
A1: an (n ' n) matrix of parameters 1fzlxtu + 1T22x2, -t ... * r2nx :
n 0
Subtracting xr_1 from each side of (6.26) and letting 1be an (n. r) identity matrix, which can be appropriately normalized.
we get The main point here is that there are two important ways to test for cointegration.
N,: -(I - A1)x,-1 + e, The Engle-Granger methodology seeks to determine whether the residuals of the
: TTXr-1 * 6, (6.27) equilibrium relationship are stationary. The Johansen (198g) and Stock-watson
(1988) methodologies determine the rank of zr. The Engle-Granger approach is the
where is the (n ' n ) matrix -(I - A) and ry denotes the element in row I and col-
zr
subject ofthe next three sections. Sections 7 through 10 examine the Johansen (l9gg)
umnT of z-. As you can see, (6.27)is a special case of (6.13) such that all ri:0.
and Stock-Watson (I 988) methodologies.
Again, the crucial issue for cointegration concems the rank of the (n - n) matrix
a'. If the rank of this matrix is zero, each element of r must equal zero. In this
instance, (6.27) is equivalent to an r-variable VAR in first differences: 4. TESTING FOR COINTEGRATION:
Lx,: e,
THE ENGLE-GRANGER METHODOLOGY
Here, each N,,: ei, so that the first difference of each variable in the vector x, To explain the Engle-Granger testing procedure, Iet's begin with the type of problem
is (0). Since each xit: xit_t * eir, all the {x;r} sequences are unit root processes and likely to be encountered in applied studies. Suppose that two variables-say y, and, zr-
there is no linear combination of the variables that is stationary. are believed to be integrated of order I and we want to determine whether there exists
At the other extreme, suppose that r is of full rank. The long-run solution to an equilibrium relationship between the two. Engle and Granger (1997) propose a four-
(6.27) is given by the n independent equations step procedure to determine if two (1) variables are cointegrated of order cl(l, l).
Tlllxlt+ Tl2x2t+ Tl3x3t* ... * rlrxnr:0 srEP 1: Pretest the variables for their order of integration. By definition, cointegra-
7r2lx1t+ T22xZt+ T23x3t * ... t lT2nxnt : 0 tion necessitates that two variables be integrated of the same order. Thus,
the first step in the analysis is to pretest each variable to determine its order
of integration. The augmented Dickey-Fuller test discussed in chapter 4
can be used to infer the number ofunit roots (ifany) in each ofthe variables.
Tnlxl.t + rn2x2t * Tn3x3t + ... t TnnXnt: 0 (6.28) Ifboth variables are stationary it is not necessary to proceed since standard
336 cHAprER 6 cotNTEGRATIoN AND ERHoR-coRBECTloN MoDELS TESTING FOR COINTEGRATION: THE ENGLE-GRANGER METHODOLOGY 337

time-series methods apply to stationary variables. If the variables are inte- test statistic used to test the rnagnitude of a1 must reflect this fact. Only if
grated ofdifferent orders, it is possible to conclude they are not cointegrated Bg and B1 were known in advance and used to construct the true {er}
in the usual sense of the term. However, as detailed in Section 5, if some sequence would an ordinary Dickey-Fuller table be appropriate. When you
variables are I(l) and some are I(2), you may want to determine whether the estimate the cointegrating vector, use the critical values provided in Table C
variables are multicointegrated. at the end of the text. These critical values depend on sample size and the
STEP 2: i3stimate the long-run equilibrium relationship. If the results of Step I indi- number of variables used in the analysis. For example, to test for cointegra-
cate that both {y,} arrd {2,\ are I(1), the next step is to estimate the long-run tion between two variables using a sample size of 100, the critical value at
equilibrium relationship in the form the 5 percent significance level is -3.398.
If the residuals of (6.31) do not appear to be white noise, an augmented
!t:00* BP,+ e, (6.30)
form of the test can be used instead of (6.31)' Suppose that diagnostic
If the variables are cointegrated, an OLS regression yields a "super- checks indicate that the {er} sequence of (6.31) exhibits serial correlalion.
consistent" estimator of the cointegrating parameters 0g and B1. Stock Instead of using the results from (6.31), estimate the autoregression:
(1987) proves that the OLS estimates of Bs and p1 converge faster than in
OLS models using stationary variables. To explain, reexamine the scatter L,a, = a12r-1 *fo,*, L,ar-, + e, (6.32)
plot shown in Figure 6.1. You can see that the effect of the common trend i=1
dominates the effect of the stationary component; both variables seem to :0, we can conchrde that the
Again, if we reject the null hypothesis at
rise and fall in tandem. Hence, there is a strong linear relationship as shown are cointegrated.
residual sequence is stationary and that the variables
by the regression line drawn in the figure. cointegrated (i.e., if
STEP 3: Estimate the error-correction model. If the variables are
In order to determine ifthe variables are actually cointegrated, denote
the null hypothesis of no cointegration is rejected), the residuals from the
the residual sequence from this equation by {2,}. Thus, the {2,} series are the
equilibrium regression can be used to estimate the error-correction model.
estimated values of the deviations from the long-run relationship. If these
deviations are found to be stationary, the l,t,j and {2,} sequences are coin-
If {yr} and {zr} are CI(\, l), the variables have the error-correction form
tegrated of order (1, 1). It would be convenient if we could perform a Ly,= 11 + c"rL! t-r_ 0tzt-tl+ tCI1(r)A y r-i + Dtz(i)L'2,-; + er1 (6.33)
Dickey-Fuller test on these residuals to determine their order of integration. i=l i=l
Consider the autoregression ofthe residuals:
Lzt= dz + arl! 1-t_ 0tzt-tl + )]o21(,)A y,-i + Ddzz(i)Lz,-i + er, (6'34)
L2.r:a12,-1 +e, (6.31) ;-l
l=1

Since the {2,} sequence is a residual from a regression equation, there where B;: the parameters of the cointegrating vector given bv (6.30), yar.d
is no need to include an intercept term; the parameter of interest in (6.31) is er, : white-noise disturbances (which may be correlated with each other),
a1. If we cannot reject the null hypothesis a1 :0, we can conclude that the and 41, d.2, cty, d, all(i), ap(i), a21(i), and a22(i) are all parameters.
residual series contains a unit root. Hence, we conclude that the lttrj and Engle and Granger (1987) propose a clever way to circumvent the
{zr} sequences are not cointegrated. The more precise wording is awkward cross-equation restrictions involved in the direct estimation of (6.33) and
because of a triple negative, but to be technically correct, if it is not possi- (6.34). The magnitude of the residual 2,-1 is the deviation from long-run
ble to reject the null hypothesis lall:0, we cannot reiect the hypothesis that equilibrium in period (/ - 1). Hence, it is possible to use the saved residuals
the variables are not cointegrated. Instead, the rejection ofthe null hypoth- {2r-,} obtained in Step 2 as an instrument for the expression yr-1 - Bprl in
esis implies that the residual sequence is stationary.8 Given that {yr} and (6.33) and (6.34). Thus, using the saved residuals from the estirnation of the
{zr} were both found to be (l) and that the residuals are stationary, we can long-run equilibrium relationship, estimate the error-correcting model as
conclude that the series are cointegrated oforder (1, 1).
Ly r= ot + a, a,-1+ lar,1;11y,-i + Datz(i)L'2,-i + ey (6'35)
In most applied studies it is not possible to use the Dickey-Fuller tables
i=1 i=l
themselves. The problem is that the {2,} sequence is generated from a
regression equation; the researcher does not know the actual error efi only Lz, = az + a, a,-1+ y,-i + Dzz(i)Ll.-i* ,1 (6'36)
the estimate of the error A,.The methodology of fitting the regression in
lart {r)A e

i=1
(6.30) selects values of B6 ar.d B1 that minimize the sum of squared residu-
als. Since the residual variance is made as small as possible, the procedure Other than the error-correction term At1,(6.35) and (6.36) constitute VAR
is prejudiced toward finding a stationary error process in (6.31). Hence, the in first differences. This Vz\R can be estimated using the same methodology
338 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRBECTloN MoDELS
developed in Chapter 5. All of the procedures developed for a VIA,R apply to
the system represented by the error-correction equations. Notably:
1. OLS is an efficient estimation strategy since each equation contains
the same set of regressors.
2. Since all terms in (6.35) and (6.36) are stationary li.e., Ly, and its lags,
Lz, and its lags, and A,-1are (0)], the test statistics used in traditional
VAR analysis are appropriate for (6.35) and (6.36). For exatrple, lag
lengths can be determined using a f-test, and the restriction that all
%t(i):0 can be checked using an F-test. If there is a single cointegrating
vector, restrictions conceming oi or a, canbe conducted using a l-test.
STEP 4: Assess model adequacy. There are several procedures that can help deter-
mine whether the error-correction estimated model is appropriate.
1. You should be careful to assess the adequacy of the model by per-
forming diagnostic checks to determine whether the residuals of the
error-correction equations approximate white noise. If the residuals
are serially correlated, lag lengths may be too short. Reestimate the
model using lag lengths that yield serially uncorrelated errors. It may
be that you need to allow longer lags ofsome variables than ofothers.
If so, you can gain efficiency by estimating the near-VAR using the
seemingly unrelated regressions (SUR) method.
2. The speed of adjustment coefficients a, and a, are of particular interest
in that they have important implications for the dlmamics of the
system.g If we focus on (6.36) it is clear that for any given value of 2,-1,
alarge value of a2 is associated with a large value of &r. If aris zero,
the change in z, does not at ali respond to the deviation from long-run
equilibrium in (r-1). If a, is zero and if all a21(i): 0, then it can be said
that {Ayr} does not Granger cause {&r}. We know that a, andlor a,
should be significantly different from zero if the variables are cointe-
grated. After all, ifboth a1 and a, are zero, there is no error correction
and (6.35) and (6.36) comprise nothing more than a VAR in f,rrst differ-
ences. Moreover, the absolute values of these speed of adjustment coef-
ficients must not be too large. The point estimates should imply that Ay,
and N, converge to the long-run equilibrium relationship.l0
If all but one variable is weakly exogenous, you may want to esti-
mate that variable using the general-to-specific technique described in
Section 10.
3. As in a traditional VAR analysis, Lutkepohl and Reimers (1992) show
that innovation accounting (i.e., impulse responses and variance
decomposition analysis) can be used to obtain information concerning
the interactions among the variables. As a practical matter, the two
innovations er, and ztmay be contemporaneously correlated if y,has
a contemporaneous effect on z, andlor if z,has a contemporaneous
effect on yr. In obtaining impulse response functions and variance
decompositions, some method-such as a Choleski Decomposition-
must be used to orthogonalize the innovations.
ILLUSTRATINGTHEENGLE-GRANGEHMETHoDoLoGY 339
The shape of the impulse response functions and the results of the
variance decompositions can indicate whether the dynamic responses
of the variables conform to theory. Since ali variables in (6.35) and
(6.36) are .I(0), the impulse responses of Ly, and A.z, should converge
to zero. You should reexamine your results from each step if you
obtain a nondecaying or explosive impulse response function.
Before closing this section, a word of warning is in order. It is very tempting to
use /-statistics to perform significance tests on the cointegrating vector. Howevel you
must avoid this temptation since the coefficients have an asymptotic t-distribution
only in one special circumstance. Suppose that the cointegration relationship between
t6,,r| and {zr} is such that
j1^r* 1Ft+ 1t
azl
^1, - E2t
where Ee1,e2,:0.
The notation is designed to illustrate the point that the residuals from both equa-
tions are uncorrelated white-noise disturbances. The set of assumptions is fairly
restrictive in that the residuals from both equations must be serially uncorrelated and
the cross-correlations mustbe zero. If these conditions hold, the oLS estimates of B6
and p1 can be tested using r-tests and.F-tests. If the disturbances are not normally dis-
tributed, the asymptotic results are such that l-tests and F-tests are appropriate. Be
aware that both conditions are necessary to perform such tests. If Ee1re2,
= O, {zr} is
not exogenous since shocks to e1, affect zr. Moreoveq as in a standard regression, if
the residuals ofthe cointegrating vector are serially correlated, inference ofthe coef-
ficients is inappropriate. Phillips and Hansen (1990) develop a procedure you can use
in such circumstances. The details are outlined in Appendix I of this chapter.
5. ILLUSTRATING THE
ENGLE.GRANGER METHODOLOGY
Figure 6.2 shows three simulated variables that can be used to illustrate the
Engle-Granger procedure. Inspection ofthe figure suggests that each is nonstation-
ary, and there is no visual evidence that any pair is cointegrated. As detailed in Table
6.1, each series is constructed as the sum of a stochastic trend component plus an
autoregressive irregular component.
The first column of the table contains the formulas used to construct the $,,,\
sequence. First, 150 realizations of a white-noise process were drawn to represent the
{g,r} t"qrr"nce. Initializing bo:0, we constructed 150 values of the random walk
process {4,r} using the formula /bt:
h,rt + er, (see the first cell of the table).
Another 150 realizations of a white-noise process were drawn to represent the {r7rr}
sequence; given the initial condition :
{,0 0, these realizations were used to construct
{{,r} ut 5r1:0.56r;f ?lyt(see the next lower cell). Adding the two constructed series
yields 150 realizations for 5,,,\. To help ensure randomness, only the last 100 obser-
vations are used in the simulated study. Hence, {yr} is the sum of a stochastic trend
and a stationary (i.e., irregular) component.
34O cHAprER 6 coTNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS TLLUSTRATTNGTHE ENGLE-GRANGER METHoDoLoGy 341

If the data happened to be quarterly, it would be natural to perform the aug-

mented Dickey-Fuller tests using lag lengths that are multiples of 4 (i.e., n: 4,8, . ..).
For each series, the results of the Dickey-Fuller test and the augmented test using 4
lags are reported hTable 6.2.
With 100 observations and a constant, the 5 percent critical value of the
-2 Dickey-Fuller test is -2.89. Since the absolute values of all f-statistics are well below
this critical value, we cannot reject the null hypothesis of a unit root in any of the
series. Of course, ifthere were any serious doubt about the presence of a unit root,
we could use the procedures in Chapter 4 to test for the presence of a trend or drift.
-6 If various lag lengths yield different results, we would want to test for the most appro-
priate lag length.
The luxury of using simulated data is that we can avoid these potentially sticky
problems and move on to Step 2. Since all three variables are presumed to be jointly
determined, the long-run equilibrium regression can be estimated lsing either yr, 2,,
or wt as the "1eft-hand-side" variable. The three estimates of the long-run relationship
(with r-values in parentheses) are
-12
20 30 40 50 60 70 80 90
y,: 4.4843 - 0.92732, + 0.97687w, + er,
(-0.s7s 1) (-38.0es) (s3.462)
'Y -"'z ----w
FIGURE 6,2 Three Cointegrated Series z,:0.0589 - 1.0108y, + 7.02549w,+ e",
(0.670e) (-38.Oes) (6s.323)
Table 6.1 The Simulated Series
w,:0.0852 + 0.9901y, + 0.953472t+ eu,t
(vl lztl lw) (-1.008e) (s2.462) (6s.462)
Trend Fyt= ltyt-,t + ey1 lazt= !zt-1 1n Pwt= lly1+ 1tn where erx ern and err,: the residuals from the three equilibrium regressions.
Pure irregular 6vt= O.56vt_.t + tivt rlzt
6zt= 0,562t-'t + 6wt= O'56wt-t + Twt The essence of the test is to determine whether the residuals from the equilibrium
Series yt= ltyt+ 611 zt= lla+ 6rr+0.56yt w1= lts.l 6*+ 0.56n+ 0.56r, regression are stationary. Again, in performing the test, there is no presumption that
any one ofthe three residual series is preferable to any ofthe others. Ifwe use each
of the three series to estimate an equation in the form of (6.31) for (6.32)], the esti-
The {2,} sequence was constructed in a similar fashion; the {err} and {r7r,\ mated values of a1 are given in Table 6.3.
sequences are each represented by two different sets of 150 random numbers. The From Table C, you can see that the critical values of the t-statistic as -3.398. Hence,
trend {pr,} and the autoregressive irregular term {62) were constmcted as shown in using any one ofthe three equilibrium regressions, we can conclude that the series are
the second column of the table. The {6r,) sequence can be thought of as a pure irreg- cointegrated of order (1, 1). Fortunately, all three equilibrium regressions yield this
ular component in the {zr} sequence. In order to introduce correlation between the same conclusion. We should be very wary of a result indicating that the variables are
$,r\ and {zr} sequences, the irregular component in {zr} was constructed as the sum:
6rt+ O.56yt.In the third column you can see that the trend in {w,} is the simple sum- Table 6.2 Estimated a1 Table 6.3 Estimated al and
mation of the trends in the other two series. As such, the three series have the cointe- and the Associated t-statistic the Associated t-statistic
grating vector (1, 1, -1). The irregular component in {wr} is the sum of pure
innovation 6nu, ar,d 50 percent of the innovations {,1 and 6",. No Lags 4 Lags No Lags 4 Lags
Now pretend that we do not know the data-generating process. The issue is Lyt -0.01995 -0.02691 L"r, -0.44301 -0.59525
whether the Engle-Granger methodology can uncover the essential details of the l-0.741571, (-1.0465) (-5.17489) (4.O741l,
process. The first step is to pretest the variables in order to determine their order of Lz, Len -0.45195 -0.59344
-0.02069 -o.25841
integration. Consider the augmented Dickey-Fuller regression equation for pr): (-0.99213) (-1.14371 (-5.37882) l-4.2263i
Lwt -0.03501 -0.03747 Le*t -0.45525 -0.60711
Ay, = ,O * dl!t-1+ !a;+rAlr -i + t (-1.9078) (-1.e335) (-5.3896) (4.2247\
i=l
342 cHAprER 6 cotNTEGRATIoN AND ERHoR-coRRECTtoN MoDELS ILLUSTRATINGTHE ENGLE-GRANGER METHODOLOGY 343

cointegrated using one variable for the normalization but are not cointegrated using However, a richer set of possibilities is given by the stationary relationship:
another variable for the normalization. In such circumstances, it is possible that only a
x1r:82x2,+ yN2,+ ar2,
subset ofthe variables are cointegrated. Suppose that x1n x21, arld x3, are three (1) vari
ables and that x1, Na x21 a;re cointegrated such that x1, - 82x2, is stationary. A regression This specification allows for the possibility that the linear combin ation x
1, - 82x2,
of x1, on the other two variables should yield the stationary relationship \t: Pix% + is (l) and cointegrated with the other (l) independent variables in the system: &1,
0x3r. Similarly, a regression ofx2, on the other variables should yield the stationary rela- and 2,. To make sure you understand the issue, ask yourself if it is possibl e for
B2lo
tionship x2r: Q/p)x1t* 0x3, However, a regression ofx3y on xllar,dx2l cannot reveal be zero. The answer is a resoundin gno.rf B2:0,the l(2)variable x1, cannot, by it;lt
the cointegrating relationship. Nevertheless, the possibility of a contradictory result is a be cointegrated with the (1) variables.
weakness of the test; other methods can be hied if mixed results are found. In principle, it is possible to check for multicointegration using a two-step pro-
You must be careful in conducting significance tests on the estimated equilibrium cedure. First, search for a cointegrating relationship among the l(2)variables and then
regressions. The coefficients do not have an asymptotic f-distribution unless the right- use this relationship to check for a possible cointegrating relationship with the
hand-side variables are actually independent and the error terms are serially uncorrelated. remaining (1) variables. Engsted, Gonzalo and Haldrup (lgg7) show that this proce-
Step 3 entails estimating the error-correction model. Consider the first-order sys- dure is effective only if the cointegrating vector for the first step is known. Otherwise,
tem shown with r-statistics in parentheses: the second step is contaminated with the errors generated in the first step. In the most
general form ofthe one-step procedure, you estimate an equation in the form
Ayr: 0.009 + 0.44leur-t + 0.190A/rr + 0.332A,2,_1- 0.380Aw,_1 + g,1 6.3i)
(0.2er) (2.e4) (1.1s) (2.0s) (-2.3s) x1t: ao+ a(+ a2P + prx2r+ &*1,+ 71N2,+ nL,4,* alzr* e, (6.40)
where x1p x2n and x31 a,ra (2) variables , z, is a vector of (1) variables, and the deter-
Mt: -0.0q2 + 0.054ewt-t + 0.139Ly,-1+ 0.253L2,-1- 0.304Lw,_1* e,, (6.38) ministic regressors can include a quadratic time trend.
(-1.11) (0.304) (0.711) (t.32) (-l.s9) Hence, the test allows you to include up to two I(2) variables and an unrestricted
: 0.065 e.,_1 + 0.1 57 L!
number of (1) variables as regressors. You might want to include the quadratic time
Lw, -g.OO, - ; + 0.302 Lz tt - 0.421 Lw,_1 + .t t (6.39)
(-0.31)(-0.e07) (1.3s) trend if A2x1, contains a drift. Since the key issue is the stationarity of the
(0.688) (-1.88) {er} series,
estimate a regression of the form
where e*r-1: wr-r* 0.0852 - 0.9901yr4 -0.953472r-l sothat e.r-y is the lagged value
p
ofthe residual from the equilibrium relationship using w/ as the dependent variable.
L,2, : pA,q+lc;Li-,_i +v,
Equations (6.37) through (6.39) comprise a first-order VAR augmented with the
i:l
single error-correction term e-r4. Again, there is aL area of ambiguity since the resid-
uals from any of the "equilibrium" relationships could have been used in the estima- where {2,} are the regression residuals from (6.40).
tion. The signs of the speed of adjustment coefficients are in accord with convergence If it is possible to reject the null hypothesis p: 0, it is possible to conclude
toward the long-run equilibrium. In response to a positive discrepancy it e.;1, both that there is multicointegration. In addition to sample size, the critical values of
y, and z, tend to increase while w, tends to decrease. The error-correction term, how- the f-statistic for the null hypothesis p: 0 depend on the number of I(2) regressors
ever, is significant only in (6.37). (*z: 1 or 2), the number of (1) regressors (*t: 0 to 4), and the form of the deter-
Finally, the diagnostic methods discussed in the last section should be applied to ministic regressors. The critical values are shown in Table D at the end of the text.
(6.37) through (6.39) in order to assess the model's adequacy. Using actual data, lag- consider the u.K. money demand equations for the sample period 1963:el to
length tests and the properties of the residuals need to be considered. Moreover, innova- 1989:Q2 estimated by Haldrup Q99g:
tion accounting could help determine whether the model is adequate. These tests are not mt : ao + 0.68p, + 1.57!t - 2.67 r, - 2.55 L,p, (6.4t)
performed here since there is no economic theory associated with the simulated data.
and

The Engle-Granger Procedure with I(2) Variables ffit: ao * ai + 0.89p, + 2.39y, - 2.69r, - 3.25Lp, (6.42)
Pretesting the variables indicated thatm, (as measured bv Ml) andp,(the implicit
Multicointegration refers to a situation in which a linear combination of I(2) and (1)
price deflator) were I(2) and that y, (total final expenditure) and r, (a measure of the
variables is integrated of order zero. For example, suppose that x1, and x27 are I(2) arrd
interest rate differential) were (1). The only variable needing explanation is the pres-
thatz,is (1). It is possible that a linear combination of x1, and x2, is (1) and that this
ence of a'p,in the money demand function. The idea is to allow forthe demand for
combination is cointegrated with zr. Hence, it is possible to have a long-run equilib-
money to depend on the inflation rate (i.e., change in the 1og of the price level) since
rium relationship of the forml1
high inflation should reduce the desire to hold money balances. Since there is a total of
x1,: Bix2r+ 712, 105 observations, one(2) regressor (so that mZ= l), and three(l) regressors, the
344 cHAprEB 6 colNTEGRArtoN AND ERBoR-coRREcrloN MoDELS COINTEGRATION AND PURCHASING POWER PARITY 345

5 percent critical values for models without and with ttre linear trend are 4.56 and4.97,
respectively. Using the residuals from the money demand equations given by (6.a1) and
(6.a2), Haldrup found that the r-statistics for the null hypothesis p: 6 were -2.35 and
are spuri-
-2.66, respectively. Hence, it is possible to conclude that the two regressions
ous (i.e., it is not possible to reject the null hypothesis of no multicointegration).
Even though multicointegration fails, Haldrup goes on to experiment with vari-
ous estimates of the error-correction mechanism. One interesting model (with stan-
dard errors in parentheses) is
L2m, : -O.O4AF1* stationary regressors
(0.02)
where the stationary regressors can include lagged values of L2m, as well as current
and lagged values of L2pt, L,yr, LPt, and L,r,. The point estimate is such that Lzm,
is expected to decline in response to a positive discrepancy from the long-run rela-
tionship. The t-statistic of -0.04/0.02:2 suggests that the effect is just significant at
the 5 percent level.
an intercept term in the equilibrium regression. In fact, Engle and Granger's (1987)
Monte Carlo simulations all include intercept terms.
The estimated values of B1 arrd their associated standard errors are reported in
Table 6.4. Note that five of the six values are estimated to be quite a bit below unity.
Be especially careful not to make too much of these findings. It is not appropriate to
conclude that each value of pl is significantly different from unity simply because the
values of (1 - B1) exceed two or three standard deviations. It is hard to overstate the
point that the assumptions underlying this type of /-test are not applicable because
there is no presumption that p, is the exogenous variable while I is the dependent
variable, or that {p,} is white noise.13
The residuals from each regression equation, calted {7D,}, were checked for unit
roots. The unit root tests are straightforward because the residuals from a regression
equation have a zero mean and do not have a time trend. The following two equations
were estimated using the residuals from each long-run equilibrium relationship:
L[Lt -- a1[tr4+ e, (6.44)
and

6. COINTEGRATION AND PURCHASING Lh : a1P,, 1 + Da;*rL,P.r-i + e, (6.4s)

POWER PARITY
To illustrate the Engle-Granger methodology using "real world" data, reconsider the
theory of purchasing power parity (PPP). Respectively, if e, p7, and p, denote the
logarithms of the price of foreign exchange, the foreign price level, and the domestic
price level, long-run PPP requires lhat e, + pp - ptbe stationary. The unit root tests
reportedinChapter4indicatethatrealexchange rates-definedasrr:et+Pft-pt-
appear to be nonstationary. Cointegration offers an alternative method to check the
theory; if PPP holds, the sequence formed by the sum {e/ + p7} should be cointe-
grated with the {pr} sequence. Call the constructed dollar value of the foreign price
levelfT;thatis,f,: et* Pft. Long-run PPP asserts that there exists a linear combina-
tion of the formfr: go +' gpr + p, such that {p,r} is stationary and the cointegrating
vector is such that 0, -- l.
As reported in Chapter 4, in Enders (1988), I used price and exchange rate data for
Germany, Japan, Canada, and the United States for both the Bretton Woods
(1960-1971) and post-Brenon Woods (1973-1988) periods.12 Pretesting the data indi-
cated that for each period, the U.S. price level {pr} and the dollar values of the foreign
price levels {er+ pfi\ both contained a single unit root. With differing orders of integra-
tion, it would have been possible to immediately conclude that long-run PPP had failed.
The next step was to estimate the long-run equilibrium relation by regressing
eachf,: e, + py, on p,'.
f,: 0o+ ltPt+ ttt (6.43)
Absolute PPP assertsf : po so this version of the theory requires 0o:0 atrd
Table 6.5 reports the estimated values of a1 from (6.44) and from (6.45) using a
lag length of four. It bears repeating that failure to reject the null hypothesis at: 0
means we cannot reject the null of no cointegration. Altematively, if -2 < a1 < 0, it is
possible to conclude that the {p,} sequence does not have a unit root and that the ffi}
ffid {p) sequences are cointegrated. Also note that it is not appropriate to use the con-
fidence intervals reported in Dickey and Fuller. The Dickey-Fuller statistics are inap-
propriate because the residuals used in (6.44) afi (6.45) are not the actual error terms.
Rather, these residuals are estimated error terms that are obtained from the estimate
of the equilibrium regression. If we knew the magnitudes of the actual errors in each
period, we could use the Dickey-Fuller tables.
Under the null hypothesis at: O, the critical values for the f-statistic depend
on sample size. Comparing the results reported in Table 6.5 with the critical values
provided by Table C indicates that only for Japan during the fixed exchange rate
period is it possible to reject the null hypothesis of no cointegration. With four lags,
Table 6,4 The Equilibrium Regressions
Germany Japan Canada
1973-1986
Estimated B1 0.5374 0.8938 0.7749
Standard error 5) (0.0316) (O.OO77)
(0.041

&: l. The intercept Bg is consistent with the relativeversion of PPR requiring only 1960-1971
that domestic and foreign price levels are proportional to each other. Unless there Estimated B1 0.6660 0.7361 1.0809

are compelling reasons to omit the constant, the recommended practice is to include Standard error (0.0262) (0.0154) (o.o2oo)
346 cHAprER 6 coTNTEGRATIoN AND ERBoH-coRRECTroN MoDELS CHARACTERISTIC ROOTS, RANK, AND COINTEGRATION 347

Table 6.5 Dickey-FullerTests of the Residuals from long-run PPP in period r - 1, the Japanese price level falls by 0. 10548 units and the
U.S. price level rises by 0.01114 units. Both ofthese price changes in period I act to elim-
Germany Japan Canada
inate the positive discrepancy from long-run PPP present in period t - l.
1973-1986 Notice the discrepancy between the magnitudes of the two speed of adjustment
No lags coefficients; in absolute value, the Japanese coefficient is approximately ten times
Estimated a1 -0.0225 -0.0151 -0.1001 that of the U.S. coefficient. As compared to the Japanese price level, the U.S. price
Standard error (0,0169) (0.0236) (0.0360) level responded only slightly to a deviation from PPP. Moreover, the error-correction
t-statistic for a1 - 0 -1.331 -0.640 -2.781 term is about'/: of a standard deviation from zero for the U.S. (0.01114/0.03175:
4 lags 0.3509) and approximately 2.5 standard deviations from zero for Japan
Estimated a, -0.0316 -0.0522 -0.0983 (0.10548/0.04184:2.5210). Hence, at the 5 percent significance level, we can con-
Standard error (0.0170) (0.0236) (0.0388) clude that the speed of adjustment term is insignificantly different from zero for the
f-statistic for a1 - 0 -1.859 -2.212 -2.533 United States but not for Japan. This result is consistent with the idea that the United
1960-1971 States was a large country relative to Japan-movements in U.S. prices evolved inde-
No lags pendently of events in Japan, but movements in exchange tate adjusted Japanese
Estimated a1 -0.0189 -0.1137 -0.0528 prices responded to events in the United States.
Standard error (0.0196) (0.044s) (0.0286) You can update the study using the data contained on the file COINT-PPP.XLS. The
f-statistic for a1 = I -0.966 -2.535 -1.846 file contains quarterly values of German, Japanese, and Canadian wholesale prices and
4 lags bilateral exchange rates with the United States. The file also contains the U.S. wholesale
Estimated a1 -o.o294 -0j821 -0.0509 price level. Question 3 at the. end ofthe chapter guides you through the process.
Standard error (0.01s8) (0.0530) (o.o3o6)
f-statistic for a., = 6 -1.468 -3.437 -1.663 7. CHARACTERISTIG ROOTS, RANK,
AND COINTEGRATION
the f-statistic for the null hypothesis a1 : 0 is calculated to be -3.437. At the 5 per- Although the Engle and Granger (1987) procedure is easily implemented, it does
cent significance level, the critical value of I is -3.398 for two variables and 100 have several important defects. The estimation of the long-run equilibrium regression
observations. Hence, at the 5 percent significance level we can reject the null of no requires that the researcher place one variable on the left-hand side and use the oth-
cointegration (i.e., we accept the alternative that the variables are cointegrated) and ers as regressors. For example, in the case of two variables, it is possible to run the
find in favor of PPP. For the other countries in each time period, we cannot reject Engle-Granger test for cointegration by using the residuals from either of the follow-
the null hypothesis ofno cointegration and must conclude that PPP generally failed. ing two "equilibrium" regressions:
The third step in the methodology entails estimation of the error-correction
model. Only the Japan/U.S. model needs estimation since it is the sole case for which
lt: 1rc+ 0sz7* e1 (6.48)

cointegration holds. The final error-correction models for Japanese and U.S. price
levels during the 1960 to 1.971period were estimated to be zt: &.0 + &.tyt + ezt (6.4e)
Ai : 0.001 t 9 - 0.r0548ilFl (6.46) As the sample size grows infinitely large, asymptotic theory indicates that the test
(0.00044) (0.04r84) for a unit root in the {e1r} sequence becomes equivalent to the test for a unit root in
the {e2r} sequence. Unfortunately, the large sample properties on which this result is
Lp t : 0.00156 + 0.01114 it-r (6.47) derived may not be applicable to the sample sizes usually available to economists. In
practice, it is possible to find that one regression indicates that the variables are coin-
(0.00033) (0.03175)
tegrated, whereas reversing the order indicates no cointegration. This is a very unde-
where p,-, is the lagged residual from the long-run equilibrium regression. Note that pr_, sirable feature ofthe procedure because the test for cointegration should be invariant
is the estimated value of fr_t - 0o - /pr_t and that standard errors are in parentheses. to the choice of the variable selected for normalization. The problem is obviously com-
Lag length tests (see the discussion of y2 andF-tests for lag length in Chapter 5) pounded using three or more variables since any ofthe variables can be selected as the
indicated that lagged values of Af_ior Lpr_i did not need to be included in the error- left-hand-side variable. Moreover, in tests using three or more variables, we know that
correction equations. Note that the point estimates in (6.46) and $.a\ indicate a direct there may be more than one cointegrating vector. The method has no systematic pro-
convergence to long-run equilibrium. For example, in the presence of a one-unit deviation cedure for the separate estimation of the multiple cointegrating vectors.
344 cHAprER 6 cotNTEGRAloN AND ERHoR-coRRECloN MoDELS
Another defect of the Engle-Granger procedure is that it relies on a two-step esli-
mator. The first step is to generate the residual series {2,}, and the second step uses
these generated errors to estimate a regression of the form L,2,: ap,_1+ . . .. Thus, the
coefficient a, is obtained by estimating a regression using the residuals from another
regression. Hence, any error introduced by the researcher in Step I is carried into Step
2. Fortunately, several methods have been developed that avoid these problems. The
Johansen (1988) and the Stock and Watson (1988) maximum likelihood estimators
circumvent the use of two-step estimators and carl estimate and test for the presence
of multiple cointegrating vectors. Moreover, these tests allow the researcher to test
restricted versions of the cointegrating vector(s) and the speed of adjustment param-
eters. Often, we want to determine whether it is possible to veritr a theory by testing
restrictions on the magnitudes of the estimated coefficients.
Both the Johansen (1988) and the Stock and Watson (1988) procedures rely
heavily on the relationship between the rank of a matrix and its characteristic roots.
Appendix 6.2 to this chapter reviews the essentials of these concepts; those of you
wanting more details should review this appendix. For those wanting an intuitive
explanation, notice that the Johansen procedure is nothing more than a multivariate
generalization of the Dickey-Fuller test. In the univariate case, it is possible to view
the stationarity of {yr} as being dependent on the magnitude (a1 - 1); that is,
yt: atyFl + Et
Ly,: (ot - l)y; + et
If (a1- l) : pr) process has a unit root. Ruling out the case in which fur)
0, the
is explosive, if (a1- l)
= 0 we can conclude that the {yr} sequence is stationary. The
Dickey-Fuller tables provide the appropriate statistics to formally test the null
hypothesis (or - 1): 0. Now consider the simple generalizationto nvariables; as in
(6.26),let
xr: Alxr_1 * e,
so that
A,x,-- Ap,_1 - xt_t I et
:(A1-I)x,_1 + e,
:Mrl*t (6.s0)
where: x, and e, are (n . l) vectors
: elements of As it is possible to include a constant in the cointegrating vector(s) with-
A1 an (n . n) matrix of parameters
I: an (n. n) identity matrix
z'is defined to be (A1 - I)
As indicated in the discussion surrounding (6.27), the rank of (A1 - I) equals the
number of cointegrating vectors. By analogy to the univariate case, if (Ar 1) con-
- Tlx1-l*
sists of all zeroes-so that rank(a') : 0-all of the {x;r} sequences are unit root
processes. Since there is no linear combination of the {xrr} processes that is station-
ary, the variables are not cointegrated. If we rule out characteristic roots that are
greater than unity and if rank(z') : n, (6.50) represents a convergent system of differ-
ence equations, so that all variables are stationary.
CHARACTERISTIC ROOTS, RANK, AND COINTEGRATION 349
There are several ways to generalize (6.50). The equation is easily modified to
allow for the presence of a drift term; simply let
A.xr:Ao* rxrt* t (6.s 1)
where lg : the (n . 1) vector of constants (ag1, ag2, . . ., aon)' .
The effect of including the various a6; is to allow for the possibility of a linear
time trend in the data-generating process. You would want to include the drift term if
the variables exhibited a decided tendency to increase or decrease. Here, the rank of z-
can be viewed as the number of cointegrating relationships existing in the "detrended"
data. In the long rurt, 7tx;1: 0 so that each {Axrr} sequence has an expected value of
aig. Aggregating all such changes over r yields the deterministic expression ai6l.
Figure 6.3 illustrates the effects of including a drift in the data-generating
process. Two random sequences with 100 observations each were generated; denote
these sequences as {err} ard {err\.Initializing y0: zo:0, we constructed the next
100 values of the {y,} and {2,} sequences as
lav-l l-o.z
1".,1=[ o,
so that the cointegrating relationship is
-0.2v.,+0.22,,:0
It: zt
In Panel (a) of Figure 6.3, you can see that each sequence resembles a random
walk process and that neither wanders too far from the other. Panel (b) adds drift coef-
ficients such that arc: azo: 0.1; now each series tends to increase by 0.1 units in
each period. In addition to the fact that each sequence shares the same stochastic
trend, note that each also has the same deterministic time trend. The fact that each has
the same deterministic trend is not a restlt of the equivalence between ap and a2g;
the general solution to (6.51) necessitates that each have the same linear trend. For
verification, Panel (c) sets 410:0.1 and a2s:0.4. Again, the sequences have the
same stochastic and deterministic trends. As an aside, note that increasing a2g and
decreasing a1g would have an ambiguous effect on the slope of the deterministic
trend. This point will be important in a moment; by appropriately manipulating the
out imparting a deterministic time trend to the system.
One way to include a constant in the cointegrating relationships is to restrict the
values of the various a;6. For example, if rank(zi'): 1, the rows of zrcan differ by only
a scalar, so that it is possible to write each {&;r} sequence in (6.51) as
+... + Tlnxnt-l* a16* e1,
A.:/-y: 1Tl2x2rl
Lx21: s2(zrgxy-1 I 7T12x2t-t + ... + Thxrrl) * a2g -t e2,
Lxnr: sr(r1py-1 * 7\2x2t-t + ... + r1nxn1-1) * ono * n,
where s; : scalars such that s;ry: zrii.
 'ifi!
'. l
350 cHAprER 6 cotNTEGRATIoN AND ERBoR-coHBECTIoN MoDELS CHARACTERISTIC ROOTS, RANK, AND COINTEGRATION 351

or in compact form,

Lxr: i xl4+ e, (6.s2)

where: xt: @r x21, ..., xn;)'
xi-- : (x1;1, x2t-1, ..., xn;y, l)'
-*-l l
" - | Tl! Tl2 Tl, atOl

ln , T22 T2n orol

20 40 60
Panel (a): No drift or intercept
80 1oo
tltl
a,tl= a2o=0.1 ato=O.1 and aro=9.4
lr"1 Tn2 T,rn anO
l
30
The interesting feature of (6.52) is that the Iinear trend is purged from the sys-
tem. In essence, the various arghave been altered in such a way that the general solu-
tion for each {r;r} does not contain a time trend. The solution to the set of difference
10 equations represented by (6.52) is such that all Lx, are expected to equal zero when
TTllxlrl* l\Zx1rt +... + Tlnxnr-l +419:0.
To highlight the difference between (6.51) and (6.52), Panel (d) of Figure 6.3
illustrates the consequences of settinE arc:0.1 and a2g: *0.1. You can see that nei-
_RL
"0 20 40 60 80 ther sequence contains a deterministic trend. In fact, for the data shown in the figure,
1oo 20 40 60 80 100
the trend will vanish so long as we select values of the drift terms maintaining the
Panel (b): Drift terms Panel (c): Drift terms
relationship alo: -azo. (You are asked to demonstrate this result in the Questions and
aro = 0.1 and a2s - -0.1 Exercises section at the end ofthis chapter.)
Some econometricians prefer to include an intercept term in the cointegrating
1
vector along with a drift term. This makes sense if the variables contain a drift and if
economic theory suggests that the cointegrating vector contains an intercept.
"0 However, it should be clear that the intercept in the cointegrating vector is not iden-
--- zt tified in the presence of a drift term. After all, some portion of the unrestricted drift
can always be included in the cointegration vector. In terms of the example above, the
-1
system can always be written as
-2 o
20 40 60 80 100
N1r: (rnxtrt * Tl2x4t-t + ... + 'trtnxnt-l+ btO) + bn + eU
t
Panel (d): Constant in the cointegrating vector Lxil-- sr(r1x17-1 * Ttyx2t-t + ... + Tlnxnt-l+ btO) + bil * nt
FIGURE 6.3 Drifts and lntercepts in Cointegrating Relationships where D;1 is defined to the value that satisfies ,rrb 10 + bt: ato.
All that was done is to divide a1s into two parts and to place one part inside the
cointegrating relationship. As such, some identif,rcation strategy is necessary since the
proportion of the drift to include in the cointegrating vector is arbitrary. The popular
software package E-Views, for example, identifies the portion belonging in the cointe-
If
the a;s can be restricted such that ars:.sia10, it follows that the all of the
grating vector as the amount necessary to force the error-correction term to have a sam-
sequences can be written with the constant included in the cointegrating
{&,r} ple mean of zero. Nevertheless, as you can see from Figure 6.3, a drift term outside of
vector: the cointegrating relationship is necessary to capture the effects ofa sustained tendency
Mt,: Q\fitrt * i\2x2r_t+ ... + Ttnxrt_t + arc) * ett for the variables to increase (or decrease). Most researchers include drift terms if the
&,2,: s2(rlxy-1 * 7r12x2rt + ... + Ttnxnt-l +-irc) +"ezt data match Panels (b) or (c) of Figure 6.3. Otherwise, they include intercepts in the
cointegrating vector or exclude the deterministic regressors altogether. If you are
Lxn,: sn(zfixl-l * Ttnx2rt + ... + Tlnxnt_l * arc)-l er, unsure, you can use the methods described in the next section to test whether the drifts
 352 cHAprER 6 cotNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS CHARACTERISTIC ROOTS, RANK, AND COINTEGRATION 353

can be appropriately restricted. Some software packages allow you to include a deter-
ministic time trend in the model. However, it is best to avoid the use of a trend as an
\^*(r,r +l) = - I In (l - i,*1) (6.56)

explanatory variable unless you have a good reason to include it in the model. Johansen where: j; : the estimated values of the characteristic roots (also called eigenvalues)
(1994) discusses the role ofthe deterministic regressors in a cointegrating relationship. obtained from the estimated r matrix
As with the augmented Dickey-Fuller test, the rnultivariate model can also be :
the number of usable observations
T
generalized to allow for a higher-order autoregressive process. consider: Wlren the appropriate values of r arc clear, these statistics are simply referted to as

xr: Apr-1 + A2xr1+ ... + Arx,_, * e, (6.s3)

)ou"" and ),,,u*.
The first statistic tests the null hypothesis that the number of distinct cointegrat-
where: xr: the (n . 1) vector (*tt, *zu ..., xn), ing vectors is less than or equal to r against a general alternative. From the previous
q: an independently and identically distributed z-dimensional vector with discussion, it should be clear that ),.u"" equals zero when all ); : 0. The further the
estimated characteristic roots are from zero, the more negative is ln(l - j;) and the
zero meafi and variance matrix X..
Equation (6.53) can be put in a more usable form by adding and subtracting
larger is the .l ru"" statistic. The second statistic tests the null that the number of coin-
Apxt_p+lto the right-hand side to obtain
tegrating vectors is r against the alternative of r f 1 cointegrating vectors" Again, if
xr: Apr-1 + Azxr2 + A3xt_l + . ..* Ao_zxtp+z + (Ap_t + Ao)xtp+t AoLx,_o*1 +
- e, the estimated value of the characteristic root is close to zero, )^*will be small.
Next, add and subtract Critical values of the )rru"" and the ).u* statistics are obtained using the Monte
Q4p_t + Ap)xrp+z to obtain
Carlo approach. The critical values are reproduced in Table E at the end ofthis text.
x,: Ap,-1 t A2xt_z + A3xt1+ ... - (Ap_r + Ao)Nrp+z- AoLx,_r*1 * e, The distribution of these statistics depends on two things:
Just as in the augmented Dickey-Fuller test developed in chapter 4, we can con- 1. The number of nonstationary components under the null hypothesis (i.e',
tinue in this fashion to obtain
n-r).
p-1 2. The form of the vector 16. Use the top portion of Table E if you do not
N, = zrxr-1 +lrrL,xr_; + e, (6.s4) include either a constant in the cointegrating vector or a drift term. Use the
i=1 middle portion of the table if you include a drift term As. Use the bottom
Pp portion ofthe table ifyou include a constant in the cointegrating vector.
where z: - (, -Dt) and ri _ : D e, . Using quarterly data for Denmark over the sample period 1974:1 to 1987:3,
i:l j:i+1
Johansen and Juselius (1990) let the x, vector be represented by
Again, the key feature to note in (6.54) is rank of the matrix z-, the rank of z. is
equal to the number of independent cointegrating vectors. clearly, if rank(zi.):0, the xr : Qn2s, yr, * , if )'
matrix is null and (6.5a) is the usual VAR model in hrst differences. Instead, if zr is
where: m2: logof the real money supply as measured by Att) deflated by a price index
of rank z, the vector process is stationary. In intermediate cases, if rank(2.) : l, there
is a single cointegrating vector and the expression ta,-1 is the error-correction term. y:
log of real income
For other cases in which 1 < rank(a) < n, there are multiple cointegrating vectors. /:
deposit rate on money representing a direct retum on money holding
jD :
bond rate representing the opportunity cost of holding money
As detailed inAppendix 6.2,the number of distinct cointegrating vectors can be
obtained by checking the significance of the characteristic roots of z'. We know that Including a constant in the cointegrating relationship (i.e., augmenting x,-1 with
the rank of a matrix is equal to the number of its characteristic roots that differ from a constant), they report that the residuals from (6.54) appear to be serially uncofre-
zero. Suppose we obtained the matrix zr and ordered the n characteristic roots such lated. The four characteristic roots of the estimated zr matrix are given in the first col-
that,\1, )2u..., )n. Ifthevariablesinx, arenotcointegrated,therankofz-iszero umn of the following table:14
and all of these characteristic roots will equal zero. Since ln(l) : 0, each of the ).,r* )t."".
expressions 1n(1 -,\) will equal zero if the variables are not cointegrated. Simitarly, -Iln(l - i.+r) -Z.E ln (1 - i)
if the rank of z'is unity, 0 < )1 < 1 so the first expression ln(l - )1f wilt be negative 0.4332 30.09 49.t4
and all the other ); : 0 so rhat ln(1 )) -- ln(l- )3) ln(l - ,lr; : 6.
- ^t-- 10.36 19.05
In practice, we can obtain only estimates of z-and its characteristic roots. The test
for the number ofcharacteristic roots that are insignif,rcantly different from unity can
i3:0.1r28
^2:0.1776 6.34 8.69

be conducted using the following two test statistics:

)o: o.oqzq 2.3s 2.35

The second column reports the various )*u* statistics as the number of usable
observations (7: 53) multiplied by 1n(1- i,.*1). For example, -53 ln(l - 0.0434)
:
,\tra"e(r) --r D ln(1-i;) (6.55) :
2.35 and- 53 ln(l - 0.1128) 6.34. The last column reports the )ou". statistics as the
i=r+l
 354 cHAprER 6 cotNTEGRATtoN AND ERRoR-coBRECTtoN MoDELS
summation of the )-u* statistics. Simple arithmetic reveals that 8.69 : 2.35 + 6.34
and 19.05 :2.35 + 6.34 + t0.36.
To test the null hypothesis r : 0 against the general alternative r : l, 2, 3, or 4,
use the ),.u"" statistic. Since the null hypothesis is r: 0 and there are four variables
(i.e., n : 4), the summation in (6.55) runs from 1 to 4. If we sum over the four val-
ues, the calculated value of.lor." is 49.14. Since Johansen and Juselius (1990)
include the constant in the cointegrating vector, this calculated value of 49.14 is com-
pared to the critical values reported in the bottom portion of rable E. For n r : 4,
the critical values of ),.u
-
are 49.65,53.12, and 60.16 at the 10, 5, and 1 percent sig-
""
nificance levels, respectively. Thus, at the I 0 percent level, the restriction is n ot bind-
ing, so that the variables are not cointegrated using this test.
To make a point and to give you practice in using the table, suppose you want to test
the null hypothesis r ( 1 against the alternative r : 2,3, or 4. Under this null hypothe-
sis, the sunmation in (6.55) runs from 2 to 4 so that the calculated value of )ou"" is 1 9.05.
For n - r: 3,the critical values of )ou". are 32.00,34.91, and,4l.07 at the i0, 5, and 1
percent significance levels, respectively. The restriction r: 0 or r: I is not binding.
In contrast to the ),.u"" statistic, the )-u* statistic has a specific altemative
hypothesis. To test the null hypothesis r: 0 agiinst the specific altirnative r: 1, use
equation (6.56). The calculated value of the ).u*(0, statistic is -53 ln (l - 0.4332)
: 30.09. For n - r : 4, the critical values of ).u* are1)25.56,2g.74,30.32, and 33.24
at the 10, 5,2.5, and 1 percent significance levels, respectively. Hence, it is possible
to reject the null hypothesis r:0 at the 5 percent significance level (but not the 2.5
percent level) and conclude that there is only one cointegrating vector (i.e., r: l).
Before reading on, you should take a moment to examine the data and convince your-
self that the null hypothesis r: 1 against the alternative r: z cannot be rejected at
conventional levels. You should find that the calcuiated value of the ,l-u, statistic for
r: 1 is 10.36 and that the critical value at the l0 percent level is 19.77. Hence, there
is no significant evidence of more than one cointegrating vector.
The example illustrates the important point that the results of the x,.u.. and \-u*
tests can conflict. The ).u, test has the sharper arternative hypothesis. li-is
preferred for trying to pin down the number of cointegrating vectors. "r"uity
8. HYPOTHESIS TESTING
In the Dickey-Fuller tests discussed in Chapter 4, it was important to correctly ascer-
tain the form of the deterministic regressors. A similar situation applies in the
Johansen procedure. As you can see in Table E, the critical varues of ihe .\,.u"" and
)rnu* statistics tend to be smallest without any deterministic regresso.s ana laigest
with an intercept term included in the cointegrating vector. Instead of cavalierly posit-
ing the form of lg, it is possible to test restricted forms of the vector.
One of the most interesting aspects of the Johansen procedure is that it allows for
testing restricted forms of the cointegrating vector(s). In a money demand study, you
might want to test restrictions conceming the long-run proportionality between
money and prices, or the sizes of the income and interest rate elasticities of demand
for money. In terms of equation (6.1) (i.e., mt: go+ gflt+
restrictions of interest are p1 : 1,
hyr+ p3rt+ e,),the
&., 0, atd, p3 < O.
HyporHEStsrESTtNG 355
The key insight to all such hypothesis tests is that if there are r cointegrating vec-
tors, only these r linear combinations of the variables are stationaty. All other linear
combinations are nonstationary. Thus, suppose you reestimate the rnodel restricting
the parameters of zr. If the restrictions are not binding, you should find that the num-
ber of cointegrating vectors has not diminished.
To test for the presence of an intercept in the cointegrating vector as opposed to
the unrestricted driftlg, estimate the two forms of the model. Denote the ordered char-
acteristic roots of the unrestricted tr matrix by i, ir, . . ., i, and the characteristic roots
of the model with the intercept(s) in the cointegrating vector(s) by i;, ij, , i;
Suppose that the unrestricted form of the model has r nonzero characteristic roots.
Asymptotically, the statistic
n
t rh (r -.{t - rnlr -iy1
-r i=r+1 (6.57)
has a yz distribution with (r - r) degrees of freedom.
The intuition behind the test is that all values of 1n(1 - i;) and ln(l - i,) should
be equivalent if the restriction is not binding. Hence, small values for the test statis-
tic imply that it is permissible to include the intercept in the cointegrating vector.
However, the likelihood of finding a stationary linear combination of the n variables
is greater with the intercept in the cointegrating vector than if the intercept is absent
from the cointegrating vector. Thus, a large value of il*, [and a corresponding large
value of -71n(1 - ii*r)l implies that the restriction artificially inflates the number of
cointegrating vectors. Thus, as proven by Johansen (1991), ifthe test statistic is suf-
ficiently large, it is possible to reject the null hypothesis ofan intercept in the cointe-
grating vector(s) and conclude that there is a linear trend in the variables. This is
precisely the case represented by the middle portion of Figure 6.3.
Johansen and Juselius (1990) test the restriction that their estimated Danish
money demand function does not have a drift. Since they found only one cointegrat-
ing vector among m2, y, id, and ib, set n: 4 and r: 1. The calculated value of the a2
statistic in (6.57) is 1.99. With three degrees of freedom, this is insignificant at con-
ventional levels; they conclude that the variables do not have a linear time trend and
find it appropriate to include the constant in the cointegrating vector.
In order to test other restrictions on the cointegrating vector, Johansen defines the
two matrices a and B, both of dimension (n ' r) where r is the rank of z. The proper-
ties of a and B are such that
r: a0'
Note that p is the matrix of cointegrating parameters and a is the matrix of
weights with which each cointegrating vector enters the n equations of the VAR. In a
sense, a can be viewed as the matrix of the speed of adjustment parameters. Due to
the cross-equation restrictions, it is not possible to estimate a and Busing OLS.15
However, using maximum likelihood estimation, it is possible to (1) estimate (6.5a)
as an error-correction model; (2) determine the rank of a (3) use the r most signifi-
cant cointegrating vectors to form B' ; arrd (4) select a such that r: qd . Question 5
at the end of this chapter asks you to find several such a and B' matrices.
 HYPOTHESISTESTING 357
356 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRRECTIoN MoDELS

It is easy to understand the process in the case of a single cointegrating vector. test statistic has a y2 distribution with I degree of freedom. A yz table indicates that
Given that rank( zr.) : 1, the rows of r are all linear multiples of each other. Hence, the 0.05 is not significant; hence, they conclude that the restriction is not binding.
equations in (6.5a) have the fonn Restrictions on 6l can be tested in the same way. The procedure is to restrict a and
+ T12X2t-1+ ... + TTnxnt-l* "' + 611 compare the r most significant characteristic roots for the restricted and unrestricted
aLlt-
A__
tr11x1p1
(r1(y;1 t 7\zx2tt + .. . + Thxntt) + ... + e2r models using (6.59). If the calculated value of (6.59) exceeds that from a 12 table,with
A,x2,: s2
degrees of freedom equal to the number of restrictions placed on a, the restrictions can
belejected. For example, Johansen and Juselius (1990) test the restriction that only
Lxn,:sn(ryx1;1 * 7\2x2r1+... + Thxnt-l)* "'+ nt
m2) responds to the deviation from long-run equilibrium.
-o."y demand (i.e.,
where the s; are scalars and, for notational simplicity, the matrices r;Lx,-i have not
formatty, they test the restriction that a2: c\: aq: O. Restricting the three values of
been written out. a; to equal zero, they find the largest characteristic root in the restricted model is such
Now define ai: siltlj and r5, so that each equation can be written
Bi: r1y' as
,n", f irff - ii:'_23.42. Since the unrestricted model is such that f h(l - i):
N;r: a;(xrr-r + /z*zr-t + ... + frxnrl)+ ... + il (i:1, ...,n) (-30.09) : 7 .67 . The y2 statistic with three
-30.09, equation (6.59) becom es -23.42 -
or in matrix forrn, degrees oi freedom is 7.81 at the 5 percent significance level. Hence, they hnd mild
p-l support for the hypothesis that the restriction is not binding'
&, =D r;L'xr-i+aB'xr-1 ]_e, (6.58) If there is a iingle cointegrating vector, the Engle-Granger and Johansen meth-
i=7 ods have the same asymptotic distribution. If it can be determined that only one
where the single cointegrating vector is p: (1, h, 02, . . ., 0n)' and the speed of adjust- cointegrating vector exists, it is common to rely on the estimated error-correction
ment parameters are given by a : (ab a2, . . ., @11)' . modello test restrictions on a. If r: 7, and a single value of a is being tested, the
Once a and B are determined, testing various restrictions on a and B is straight- usual ,-statistic is asymptotically equivalent to the Johansen test'
forward if you remember the fundamental point that if there are r cointegrating vec-
tors, only these r linear combinations of the variables are stationary. Thus, the test Lag Length and CausalitY Tests
statistics involve comparing the number of cointegrating vectors under the null and
The simplest way to understand lag length tests is to consider the system in the form
alternative hypotheses. Again, let i, i2, ..., i, and ii, ii, ..., i) denote the ordered
of (6.5a)
characteristic roots of the unrestricted and restricted models, respectively. To test
p-1
restrictions on B, form the test statistic:
r^ L*, : Mt; ll niL'xr_-; I e,
rf
. trntr-iil-rn (1-i)l (6.se) i:l
i=7
Regardless of the rank of r, all of the Axr-, are stationary variables. Hence, we
Asymptotically, this statistic has a y2 distribution with degrees of freedom equal .rn or" Rul" I of Sims, Stock, and Watson (1990). Recall that the rule implies that
to the number of restrictions placed on B. Small values of ii relative to i (for i < r) the coeffrcients of interest on zero-mean stationary variables can be tested using a
imply a reduced number of cointegrating vectors. Hence, the restriction embedded in normal distribution. Since lag length depends solely on the values of the rr, a y2 dis'
the null hypothesis is binding ifthe calculated value ofthe test statistic exceeds that tribution is appropriate to test any restriction concerning lag length. As in the case of
in a y2 table. For example, Johansen and Juselius test the restriction that money and any VAR, lei D, ana X,. be the variance/covariance matrices of the unrestricted and
incomemoveproport"T),':'rnlrr',:::7,;:";:;:n;:':'**rerationshipis restricted systems, respectively. As in Chapter 5, let c denote the maximum number
ofregressors contained in the longest equation. The test statistic
(I-cxlogl D,-l - logl X,l)
They restrict the coefficient of income to be unity and find the restricted values
can be compared to a y2 distribution with degrees of freedom equal to the number
of
of the ii to be such that
restrictions in the system. Alternatively, you can use the multivariate AIC or SBC to
ii rro(r - i) determine the lag length. If you want to test the lag lengths for a single equation, an
i:1 0.433 -30.04 F-test is approPriate.
i:2 0.172 -10.01 The rule also means that you cannot perform Granger causalify tests in a cointe-
i:3 0.044 1.36 :
grated system using a standard F-test. First, suppose that rank (r.) 0 so that
i:4 0.006 -0.32 p-1
Given that the unrestricted model has r : I and -T ln(l - ir) :
30.09, (6.59) L*, :l r;N,-i -t e,

becomes -30.04 + 30.09:0.05. Since there is only 1 restriction imposed on B, lhe i:l
358 cHAprER 6 cotNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS
HYPOTHESISTESTING 359

As such, Granger causality involves only stationary variables. yet, this was pre- want to extend the number of lags if you suspect a substantial amount of seasonaliry For
cisely the case discussed in Chapter 5 when the variables in a VIA,R are not cointegrated.
example, with 100 observations of two variables using quarterly data, you might want to
Hence, Granger causality tests can be conducted using a standard F distribution. begin with 12 lags even though I1l3 is approximately five. Select the most appropriate
However, if the variables are cointegrated, a Granger causality test involves the coeffr-
lag length and then perform a cointegration test. Ifthe variables are not cointegrated, esti-
cients of zr. Since these coefficients multiply nonstationary variables, it is not appropri-
mate the system in first differences. If the variables are cointegrated, you can work with
ate to use a F-statistic to test for Granger causality. Block exogeneity tests are also ruled
the error-correction model. Since the error-correction term and all values of N,_rare sta-
out. If w, is cointegrated withy, or zt,yov cannot use a standard f testto determine tionary, you can inference on any variable (except those appearing within the cointegrat-
whether w, belongs in the equations for y, and 2,. ing vectors) using the usual test statistics. Impulse responses and variance
decompositions will yield consistent estirnates of the actual values.
To Difference or Not to Difference
We have reached a point where it is possible to address the issue of differencing the
The Test in the Presence of ll2l Variables
nonstationary variables in an unrestricted VAR. There is no question that differencing
It is also possible to test for multicointegration using Johansen's methodology.
leads to a misspecification error if the variables are cointegrated. Suppose that the
Consider the VAR system:
acfial data-generating process is given by the cointegrated system of(6.5a) but you
p-2
estimate the following VAR in first differences: :
L2 x, lrxr-1 *lA.x, 1 -f I riL2xt-i * t (6.60)
p-1 i:l
Axr:fn,N,-i+e, The issue of multicointegration concerns the ranks of both z' and l. In principle, it
i:1
of cointegration for the variables in the system.
is possible to consider all possible orders
The system is misspecified since it excludes the long-run equilibrium relation- However, to illustrate the procedure, it is useful to begin with a three-vaiable system
ships among the variables that are contained in rx,-1. Given the misspecification consisting of the three {2) variables xtb x2t, au,:td 4, that are multicointegrated such that
erroq all of the coeffrcient estimates, l-tests, -E-tests, tests of cross-equation restric- :
tions, impulse responses and variance decompositions are not representative of the
1t ix h + 1T
t2x2t + ir 1;.x3r + 11 1 An1, + I DLx2t + I 1 3A-x3, 0

true process. Hence, there is a substantial penalty to pay if you estimate a vAR in first
differences when the data are actually cointegrated; differencing "throws away"
information contained in the cointegrating relationship(s).
Why not simply estimate all VARs in levels? The answer is that it is preferable to
use the first differences if the (l) variables are not cointegrated. There are three conse-
quences if the (1) variables are not cointegrated and you estimate the vAR in levels:
1.
Tests lose power because you estimate n2 more parameters (one extra lag
ofeach variable in each equation).
2. For a VAR in levels, tests for Granger causality conducted on the (l) vari-
ables do not have a standard F distribution. Ifyou use first differences, you
can use the standard F distribution to test for Granger causality.
3. When the VAR has (l) variables, the impulse responses at long forecast
horizons are inconsistent estimates of the true responses. Since the impulse
responses need not decay, any imprecision in the coefficient estimates will
have a permanent effect on the impulse responses. If the VAR is estimated
in first differences, the impulse responses decay to zero and so the esti-
mated responses are consistent.
The suggestion is that it is impofiant to properly determine whether the (1) vari-
ables are cointegrated. You can perform lag length tests regardless ofwhether the vari-
ables are cointegrated. As such, the suggested methodology is to estimate an unrestricted
VrA,R. Most researchers would begin with a lag length of approximately fl/3. you may
Let r denote the rank ofz-and 11 denote the rank off so that (6.60) is such that
r : rt: 1. Clearly, if r : 0, multicointegration fails since there is no linear combina-
tion of the three I(2) variables that forms an equilibrium relationship. If r: I and
rt: 0, the equilibrium relationship has the form z-11xr, + 1Tt2x2t+ 1T l3x3t:0. As such,
L2xr-- lTxFt + (0) variables so that T1,xtt+ 1Tt2x2t+ lt ..x31 is necessarily a station-
ary relationship-the variables are CI(2,2). All of this may seem straightforward, but
there is a complicating factor when the ranks of zr and I have to be estimated. To
illustrate the point, suppose that the.I(2) variables are cointegrated such that
1f txU+ 1T12x2t+ r34r- I(l)
where - (@ indicates the order of integration.
zi-11Ax1, + nr2Lx2,+ r;;L4,- I(0).
If you take the first difference, it follows that
You should be able to figure out the problem. For any cointegrating vector in z-, it is
possible to estimate an identical cointegration vector for the first differences of the
variables. Yet a linear combination of the two relationships is not stationary. Consider
the result obtained by subtracting the (0) relationship from the (1) relationship:
1Tlfitt + 1Tt2x2t + Tt3x3t- n1Lx11- rpLx2, - n1L4,: rllxtt-l * 7Tt2x2t-7 * r13x3t-l
Since z-11x1r_ r * Tt2x2t-r * T Bx3t-t is (1), all that has been done is to change the
time subscript for the variables in the cointegrating relationship. The point is that it is nec-
essary to find cointegrating vectors in f that are not linear combinations of those inr.rl
Ifwe take the more general case considered by Johansen (1995), let rank(zr) : r
and let s denote the number of cointegrating vectors in I that are orthogonal to those
 360 CHAPTER 6 coINTEGRATIoN AND ERRoR-CoRBECTIoN MoDELS
HYPoTHESTsTESTTNG 361
in z'. In an n-variabre system such that some of the variables
are l(2),you should be three-variable system, there is a cointegrating vector for each bilateral pair of the vari-
able to veriff that:
ables (2 : n - r + 1). To demonstrate the point, let xr-- (xtt, xzu hb x4)' and suppose
1. rf r: 0, there is no relationship among the variabres
that is stationary. there are two cointegrating vectors for these four variables. If we normalize each vec-
2. In an variable system, if r + s : n T,there is a
- unique multicointegrating tor with respect to rll, we can write that condition p'x,: 0 as
vector. The number of I(2) stochastic trends in a
n-viriabre system is given
byn-r-s.
3' The value of s must be such that s < n r.
- For the anarysis of I(2) vari- r -otz -g* -e. l[il,l-r.l
ables to be appropriate, t}te values of r and
s : n - r, then x, contains no I(2) variables.
., must be such that s + r < n.If t -gzz -gzt -rr-]l,r,l - Lol

cointegration test with (2) variables is actualry

- the
In .Johansen's
first step, you
a two-step procedure.
estimate a model as in (6.60) to determine the Consider the 2 . n matrix B' consisting of the cointegrating parameters. Subtract
Determine the value of r using the )r.r"" and rank of 2.. row I from row 2 to obtain
)-"* statistics in the usual way. In the
second step, you determine the varue o?,i
hypothesis be s : s6 and consider:
conditibnal on the varue of ,.ra Let the null It
t'-l -gr, -gn -1rq I

L0 -0zz + hz -?zt -r 0n -?zq + 0Al

.n
Q,..,:-r f mft-)i) (6.61) Noq renormalize row 2 by dividing each of its elements by (0n- 822)to obtain
i:so *l
Hence, 8l', is constructed in the same fashion Ir -a,. -At -Aq]
differences are that you test the rank ofr
obtain the number of cointegrating vectors
as a .\r.""" statistic. The principal
conditionut on itii"rurue ofr and that you
orthogonal to those in tr.ts As such, the
lo 1'' -g): -p;^l
critical values needed to determine the value of where -g)t : (gn - 7z)l(gp - 822) and
: (0v - gzil@p - 0z).Hence, x21,
value of r, if the sample value of ol,, exceeds
s have to be modified. Given the
x37, and x41 &ra cointegrated such that x21 :-aio
gLt\1 + g\cxqr Similarly, add B12 times
the criticar value calcurated by
Johansen, reject the nulr hypothesis s : sg
in favor ofthe alternative s > s6. For r: l, row 2 to row I to obtain
the critical values at the 10 percent, 5 peicent,
and I percent significance Ievers
r*l o -gL
[r
are
Critical Values for Oi -ol^1

s:0 s:l
"
fo t -a)z -P;^)
t0% 31.88 17.79 where 0i; : gy + gng\i.
5% 34.80 19.99 i3, arrd x4 are cointegrated such that xy: gizxtt + Biaxar. Since the label-
Thus, x1,
lyo 40.84 24.74 ing ofthe variables is irrelevant, it follows that there exists a cointegrating vector for
each subset of three variables. More generally, B' will be ar ' n matrix of cointegrating
For example, let r :1and suppose that the sample value of 0i, is
found to be 35.00. parameters, and each subset of n - r + I variables will be cointegrated. From the pre-
As such, the null hypothesis,: o .un be rejected
ut tt" s p".""r,'rigrrr"ance rever. ceding discussion, it should be clear that standard row and column operations ot B' do
Tests on Muttiple Cointegrating Vectors not entail restrictions on the cointegrating vectors. Such operations merely result in
additional cointegrating vectors that are linear combinations ofthe original vectors.
Ifthe rank ofzr exceeds one, it is not straightforward to interpret
the cointegrating vec- EXAMPLE 1: VARIABLE EXCLUSION WITHIN AN EOUATION With multi-
tors. when there are multiple cointegratirg
vectors, any linear combination of these ple cointegrating vectors, you cannot test whether any one particular Ey: 0 since this
vectors is also a cointegrating vector.horrr.irately,
it is often porriut" to iaentifu sepa_ assumption does not restrict the cointegrating space. In the general case where B' is an
rate behavioral relationships by appropriatery
restricting the individual cointegrating
vectors' The only complication is that you r . n matrix, a testable exclusion restriction entails the exclusion of r or more variables
r""d to be clear about the number of restric- from a cointegrating vector. Hence, excluding r variables from a cointegrating vector
tions you impose on the system. It is important
to note that if there are r cointegration
relationships in an n-variabre system, tiere entails only one restriction. If the sample value of the 12 statistic with one degree of
exists a cointegrating vectorfor each sub_
set of (n - r + 1) variabres. For example, freedom (since there is only one restriction involved) exceeds a critical value, reject
if there are two cointegrating vectors in a the null hypothesis that this set of variables contains a cointegrating relationship.
 362 cHAprER 6 cotNTEGRATtoN AND ERRoR_coBRECTtoN MoDELS
TLLUSTRATTNGTHEJoHANSENMETHoDoLoGy 363
EXAMPLE 2: VAR|ABLE ExcLUStoN AcRoss EouATtoNs Next, data-generating process. In most instances you will have t ariables that are inte-
suppose that you want to test whether x4t cal be excluded from the set of
cointegrat- grated of the same order. In other cases, you can check for multicointegration.
gA:
ing relationships. The restriction gzq: 0 entails only one restriction on the coin- The results of the test can be quite sensitive to the lag length, so it is
r.
tegrating space. In the general case where gy: /zi:
B' is an nmatrix,the test ... important to be careful. The most conlmon procedure is to estimate a vector
j 1ri:0_still involvesonly one restriction. This follows sincex;, can be eliminated autoregression using the undffirenced data. Then use the same lag-length tests
from r - I equations using simple row and column operations.
as in a traditional VAR. Begin with the longest lag length deemed reasonable
and test whether it can be shortened. For example, if we want to test whether
EXAMPLE 3: coNDrrtoNAL RESTRtcrloNs It is also possibre to Iags 2 through 4 are important, we can estimate the following two VARs:
restrict one cointegrating vector conditional on the values ofall other cointegrating
vec-
tors. For example, you might want to determine if (1, xt: AO * Appl-t A2x;2-t A3x;3 * A4x,4-t e1,
822,0, 7zc),is a cointegrating vec_ * APr-1 * e2,
tor for the fixed normalized values of gn, gn,and xt: Ao
B,a. thus, you fix the varues ir prr,
0n, and 0s and determine whether you canixcrude x3, from ihe second vector. cutler, where: x,:
the (n . 1) vector of variables
Davis, and Smith (1999) consider the identification issue in considerable
detail. They Ao: . 1) matrix of intercept terms
(n
examine the following four behavioral relationships in a seven variable
system: At: @ . n) matrices of coefficients
tn,: dg + dlt + d2r, + drp, + sr, ey and ezt: (n . 1) vector of error terms
Ct: aO* alyr* a2rr* e2, Estimate the first system with four lags of each variable in each equa-
ir: bg+ b1y7+ b2r,* e3, tion and call the variance/covariance matrix of residuals X4. Now estirnate
int,: gg+ gtyt+ g2rt* e4t the second equation using only one lag ofeach variable in each equation and
call the variance/covariance matrix of residuals X1. Even though we are
where: m,: log of norninal money holdings
working with nonstationary variables, we can perform lag length tests using
yt:lo1 of real income
the likelihood ratio test statistic recommended by Sims (1980):
r, : real interest rate
c,: log of real consumption (r-c)(loglxl | - loglxal)
i, : log of real investment
where: I:number of observations
pt: lo1 of the price level c :
number of parameters in the unrestricted system
im,: log of real imports
log X; | :
natural logarithm of the determinant of X;
eq are stationary error terms.
e1r, e2t, e37, and I

The first equation is the money demand equation. The next three equations Following Sims, use the X2 distribution with degrees of freedom equal
are a to the number of coefficient restrictions. Since each Ai has 12 coefFrcients,
simple consumption function, an investment function, and an import
demand func- constraining I z: At: Aq: 0 entails 3r2 restrictions. Alternatively, you can
tion, respectively. consumption, investment, and imports ur" .u.h assumed
to be select lag lengthp using the multivariate generalizations of theAIC or SBC.
functions of only income and the interest rate. The issue is to determine whether
it is STEP 2: Estimate the model and determine the rank of zr. Many time-series statistical
possible to identi$, these four equations from a seven-variable
systern. Toward this software packages contain a routine to estimate the model. Here, it suffices
end, they obtained estimates of a (7 x 7) r matrix over a numbeiof
sample periods. to say that OLS is not appropriate because it is necessary to impose cross-
There were at least four cointegrating vectors in every case considered.
over the equation restrictions on the zr matrix. In most circumstances, you may choose
entire sample,1960:2 to r99o:4, cutler, Davis, and Smith (1999)
found that they to estimate the model in three forms: (1) with all elements of l6 set equal to
could not reject the restrictions at conventional significance levels (the
prob-value zero, (2) with a drift, or (3) with a constant term in the cointegrating vector.
was 16 percent).
For example, we can use the simulated data shown in Figure 6.2 so
9. ILLUSTRATING THE JOHANSEN METHODOLOGY that xr: (j r z t w )' . If we pretend that we do not know the form of the data-
generating process, we might want to include an intercept term in the coin-
An interesting way to illustrate the Johansen methodology is to use exactly tegrating vector(s). As we saw in the last section, it is possible to test for
the same
data shown in Figure 6.2. Although the Engre-Grangeiiechnique the presence ofthe intercept. Lag-length tests indicate settingp:2 so that
did find that the
simulated data were cointegrated, a comparison of the two pro""du.",
is useful. Use the estimated form of the model is
the following four steps when implementing the Johansen piocedure.
Nt: Ao a rx7_1 + rrLx,_1 -t e, (6.62)
STEP 1: It is good practice to pretest all variables to assess their order
of integra- wherc Ag was constrained so as to force the intercept to appear in the corn-
tion. Plot the data to see if a linear time trend is rikely to be present in the
tegrating vector.
364 CHAPTER 6 COINTEGRATION AND ERROR-CORRECTION MODELS ILLUSTRATINGTHEJOHANSEN METHODOLOGY 365

As always, carefully aflalyze the properties of the residuals of the esti- Table 6.6 The ,\.r, and ),r"""Tests
mated model. Any evidence that the errors are not white noise usually means
that lag lengths are too short. Figure 6.4 shows deviations ofy, from the long-
Alternative 5%Critical 10% Critical
NullHypothesis Hypothesis Value Value
run relationship (fu -- -0.0133 1 - y1 - 1 .03502t + I .0162w) and one of the
elror sequences (i.e., the {el,/} sequence that equals the residuals from they, ,\trace tests A,,r"" value
equation in (6.62)). Both sequences conform to their theoretical properties in r=0 r>0 56.786 34.91 32.00
that the residuals from the long-run equilibrium appear to be stationary and r<1 r>1 18.123 19.96 17.85
the estimated values of the {ery} series approximate a white-noise process. r <2 r>2 3.306 9.24 7.52
The estimated values ofthe characteristic roots ofthe zr matrix in(6.62) are lmax tests )-r, value
:0.32600; ),2:0.14032; and )3:0.033168 r=0 r= 1 38.663 22.OO 19.77
11
r= 1 r=2 14.817 15.67 13.75
Since 7:98 (100 observations less the two lost as a result of using 2 r=3 3.306 9.24 7.52
r =2
lags), the calculated values of .\-r* and )ou"" for the various possible val-
ues of r are reported in the center column of Table 6.6.
Consider the hypothesis that the variables are not cointegrated (so that Since 56.786 exceeds the 5 percent critical value of the .\ou"" statistic
the rank 7r: 0). Depending on the altemative hypothesis, we have a choice (in the bottom panel of Table E, the critical value is 34.91), it is possible to
of two possible test statistics. If we are interested in the hypothesis that the reject the null hypothesis ofno cointegrating vectors and accept the alterna-
variables are not cointegrated (r:0) against the alternative ofone or more tive of one or more cointegrating vectors. Next, we can use the ,\,.r""(l) sta-
cointegrating vectors (, > 0), we can calculate the )or""(0) statistic: tistic to test the null of r I I against the alternative of two or three
)*u""(0) : -7 [n(l - )r) + 1n(1 - .\2) + 1n(1 - .\3)] cointegrating vectors. In this case, the .\o"".(1) statistic is
: -98 [n(l - 0.326) + ln (l- 0.14032) + ln(1 - 0.033168)] lt u."(1)
: -7 [ln(l - ]z) + ln(l - .\3)l
:56.786 : -98 [ln (1- 0.14032) + ln(l - 0.033168)]
:18.123
Since 18.123 is less than the 5 percent critical value of 19.96, we can-
not reject the null hypothesis at this significance level. However, 18.123
0.8 does exceed the 10 percent critical value of 17.85; some researchers might
reject the null and accept the alternative offwo or three cointegrating vec-
0.6 tors. The )t u""(2) statistic indicates no more than two cointegrating vectors
at the 10 percent significance level.
0.4
The .\.r* statistic does not help to clariS the issue. The null hypothesis
of no cointegrating vectors (r : 0) against the specific alternative r : I is
clearly rejected. The calculated value )-u*(0, 1) : -98 ln(l - 0.326) :
0.2

0 38.663 exceeds the 5 percent critical value of 22.00. Note that the test of the
null hypothesis r : I against the specific alternative r:2 cannot be rejected
-o.2 at the 5 percent, but can be rejected the 10 percent, significance level. The
calculated value of ).u*(1, 2) is -98 1n(1 - 0.14032): 14.817, whereas the
-0.4 critical values at the 5 and 10 percent significance levels are 15 .67 and 13 .1 5,
respectively. Even though the actual data-generating process contains only
-0.6
one cointegrating vector, the realizations are such that researchers willing to
use the l0 percent signihcance level would incorrectly conclude that there
-0.8
are two cointegrating vectors. Failing to reject an incorrect null hlpothesis is
-1 always an inherent danger of using wide confidence intervals.
3 8 13 18 23 28 33 38 43 48 53 58 63 68 73 78 83 88 93 98
STEP 3: Analyze the normalized cointegrating vector(s) and speed of adjustment coef-
Long-run error ----Residual ficients. If we select r : I , the estimated cointegrating vector (Bs h 0z 0) is
FIGURE 6.4 Long-Run and Short-Run Errors B : (0.005s3 0.4rs32 0.42988 -0.42207)
366 cHAprER 6 coTNTEGRATToN AND ERRoR-coRRECTroN MoDELS GENERAL.TO-SPECIFICMODELING 367

If we normalize with respect to Pbthe normalized cointegrating vector The theoretical framework underling the approach begins with an autoregressive
and the speed of adjustment parameters are distributed lag (ADL) model. In the two-variable first-order case, the ADL model is
p: (-0.01331 -1.0000 -1.0350 1.0162) lt-- attlrt1, at2zt-t at3zt 1* t
ar: 0.54627
As will be demonstrated shortly, it is useful to write the ADL model in the
a,:0.16578 equivalent form
d*:0.21895
. a(ya- ]rt_)+ pLz,* e,
Ly,-- (6.63)
Recall that the data were constructed imposing the long-run relationship
wt : lt * zi hence, the estimated coefficients of the normalized B vector are where o : (at - l), p: ap,and 0: @n+ ap)l(l - a;n).
close to their theoretical values of (0 -l -1 1).16 Consider the following tests: In a more general setting, with n variables, the ADL model has the form
1. The test that Bg:0 entails one restriction on one cointegrating xy: A{L)x1,_, + A2(L)x2r * ... * A,(L)xr, + e,
vector; hence, the likelihood ratio test has a y2 distribution with
The point is that the ADL is extremely general; as such, it is a useful way to
one degree of freedom. The calculated value of X2 -- 0.011234 is
begin modeling the variable x1r. For our purposes, the essential point is that an
not signif,rcant at conventional levels. Hence, we cannot reject the
error-correction model is embedded within the ADL; as such, appropriately restrict-
null hypothesis that Bs: 0. Thus, it is possible to use the form of
ing the ADL yields an error-correction model. To return to the two-variable case,
the model in which there is neither a drift nor an intercept in the
suppose thaty, and z, are cointegrated of order (1, 1) and that the error-correcting
cointegrating vector. Thus, to clarifu the issue conceming the
model is represented by
number of cointegrating vectors, it would be wise to reestimate
the model excluding the constant from the cointegrating vector. Ly,: a{yt- 0r;) + et (6.64)
) To restrict the normalized cointegrating vector such that 0z: -l
and B3: 1 entails two restrictions on one cointegrating vector;
A,z,: a2$t,_1 - pz,s) + e2, (6.6s)

hence, the likelihood ratio test has a y2 distribution with two Notice that (6.64) and (6.65) are in reduced form and not in structural form. As
degrees of freedom. The calculated value of y2 :0.55350 is not in Chapter 5, there is no reason to suppose that the error terms are uncorrelated with
significant at conventional levels. Hence, we cannot reject the each other. Given that we do not want to impose that restriction Ee11e21: 0 we can
null hypothesis that B2: -1 and fu: l. we can let the variance/covariance matrix of the errors be given by
3. To test the joint restriction P : (0, -l , -1, I ) entails the three l_ _ I
ut2l
restrictions 0o : 0, 0z: -7 , and B3: 1 . The calculated value of X:1"'t
y2 withthree degrees of freedom is 1.8128 so that the signih- lotz ozz)
cance level is 0.612. Hence, we cannot reject the null hypothesis
where o;i: Ee,ei.
that the cointegrating vector is (0, -1, -1, l).
As long as o12 = 0, e11 and e2y are correlated. It is possible to write the contem-
STEP 4: Finally, innovation accounting and causality tests on the error-correction model poraneous relationship between the two errors as
of (6.62) could help to identiff a structural model and determine whether the e11: Pe2,+ , (6.66)
estimated model appears to be reasonable. Since the simulated data have no
where p is o9/o22 and e, is a white-noise error term.
economic meaning, innovation accounting is not performed here.
If we substitute (6.65) and (6.66) into (6.64), we obtain
1 O. GENERAL.TO.SPECIFIC MODELING Ly,: a{y,-t - 0z;) + pe2t+ t e
: at(tt_t _ 0r,_) + plNt _ dz(tt _ Bz,_)) + e,
In some sense, the general-to-specific approach is a compromise between a pure error- : (CIl - pa) (l; - ]ra) + pLz,+ e,
correction model and a traditional regression-based approach. An error,-correction
Now, if we let o : al - pab we can write
model can be overparameterized while a pure regression can seem ad hoc unless you
are sure of the correct specification. The general-to-specific modeling approach is Lyt: a(y;- 0r,-)* pLz,* e, (6.67)
more of a philosophy than a specific statistical test. As the name suggests, you begin Hence, we can write the error-correction equation for y, given by (6.6$ and (6.65)
with an extremely general model and pare it down by testing various coefficient in the exact fonn as the reparameterizedADl given by (6.63). Of course, if we start from
restrictions. In this way, the final model is always nested within the initial model. a higher-order process, additional lags of Ayr-; and A,z,_ishould be added to the equation.
368 cHAprER 6 colNTEGRATToN AND EBRoR-coRBECTroN MoDELS GENERAL-TO.SPECIFICMODELING 369

Weak Exogeneity should have a proportional effect on the CPI. As such, a testable hypothesis is p, +

In a cointegrated system, if a variable does not respond to the discrepancy from the
0z+ h: 1. Of course, the theory also requires that the {er} sequence be stationary.
The paper used quarterly data for each ofthe four variables over the 1976l.3 to
long-run equilibrium relationship, it is weakly exogenous.20 Hence, if the speed of 1993:3 period for a total of 69 observations. The time paths of the CPI, unit labor
adjustment parameter aiis zero, the variable in question is weakly exogenous. The costs, and import prices are shown in Figure 6.5.22 All three series have a decidedly
practical importance is that a weakly exogeneous variable does not experience the upward movement and seem to track each other reasonably well. The first step is to
type of feedback that necessitates the use of a VAR. analyze each for possible unit roots versus deterministic time trends. Since the
To explain (see Appendix 6.1 for more details), suppose that you estimate an orders of integration are unclear, tests are also conducted for multiple unit roots.
ADL using OLS. You know that this equivalent to estimating the equation in the form The augmented Dickey-Fuller tests (using four lags of each variable) have the fol-
of (6.67). Hence, when z, is zol weakly exogenous (i.e., when d.2 = 0), you estimate lowing form:
the unrestricted equation
43
Ay,:01,-t+ ]zzrt+ BrL,z,+ e, (6.68) Lyt : ao t.y!t_r+ a2t +Dg,Lr,_, +ldiD;t * et (6.70)
where from (6.67) the estimated coefficients are such that B1: a1 pa2, 0z: (ar -
pa)B and fu: p. The only variables needing explanation are the centered seasonal dummy
Notice that the contemporaneous correlation coefficient between e11 and e2, (i.e., p) variables {Du}, {Dzt}, and {D3,}. Each seasonal dummy Di, has a value of 314 in
appears in the estimate for each coefficient of (6.68). However, if z, is weakly exoge- quarter i and -l/4 in the other three quarters. In this way, each dummy has a mean
nous (i.e., when a2 : 0), your coefficient estimates are such that B1 : at 0z: at/, value of zero (i.e., it is centered around zero) and so cannot contribute to the magni-
and B3: p. Hence, you can identi$ ap B, and p from 0y 02, and 83. More impor- tude of the drift. The estimated values of 7 and the associated t-statistics for the null
tantly, the fact that az: 0 means that the contemporaneous correlation between e1, hypothesis 7: 0 are
and e2, does not affect the estimate of the cointegrating parameter B: By'ayThe fact
thatz,is correlated with the innovations in e1, is not important for the estimates of o1
pt ulc, ih pett
and B when az : 0. Hence, if a2: 0, it is appropriate to estimate (6.68) by OLS and 'y 0.02 -{.0s -0.06 -0.12
to construct confidence intervals for p andlor a1 using the normal distribution.2l r-statistic (0.91) (-0.93) (-r.2s) (-2..20)
At the 5 percent significance level, the Dickey-Fuller critical value is -3.50;
lnflation in Australia hence, the null hypothesis of a unit root cannot be rejected. Moreover, there is sup-
In a very instructive paper, de Brouwer and Ericsson (1998) show the proper way to port for the claim that {p,} ard {ip,} are both (2) variables since the null of a unit
model a single variable in a cointegrated system using the general-to-specific root cannot be rejected for the {Lpr} and {Lip,\ series. At this point, there is a crit-
approach. Two other excellent examples of the approach include Hendry and Ericsson ical choice to make. One is to allow for the possibility of multicointegration.
(1991) and Davidson, Hendry Srba, and Yeo (1987).
The goal is to develop a model of the Australian inflation rate as measured by the
140
"underlying" CPI. This measure of inflation excludes certain volatile items such as
fruit and vegetables, automotive fuel, and mortgage interest rates. Without investigat- 120
ing the underlying theoretical model, for our purposes, it is sufficient to posit that the
price index may be represented by 100
o
o
pt: 0o+ Brulc,+ ]zipt+ B3pet,+ e, (6.6e) -80
il

where: pr: the underlying CPI

Hoo
ulcr: an index of unit labor costs o,
ip,: index of tariff-adjusted import prices converted to Austra.lian dollars -40
petr: index of gasoline prices measured in Australian dollars
20
e,: the deviation from the long-run equilibrium
All variables are expressed in natural logs. 0r
In a cost-push model of inflation, increases in unit labor costs, import prices, and
Jun 76 Jun 79 Jun 82 Jun 85 Jun 88 Jun 91

gasoline prices should all increase the Australian CPI. Hence, gy gz, and B3 should ---ip ---ulc
FIGURE 6.5 -cpi
Prices and Costs in Australia
all be positive. Moreover, proportional increases in each of the subcomponent prices
37O cHAprER 6 colNTEGRATtoN AND ERRoR-coRRECTIoN MoDELS
However, with only 69 total observations, the Dickey-Fuller tests have very low
power to detect a truly stationary process. Moreover, if an equation in the form of
(i.e., L2p, and
iA.ZO; Ir estimated using the second differences of these two variables
L2ip),thevalues of lare foundto be-0.44 (/:-3.01) and-0.68 (t:-2.55), respec-
tively. Given that the point estimates of the characteristic roots are far from unity, the
authors decided to treat all four variables as (1) processes.
The next step was to obtain the appropriate lag length for the VAR' Given that the
data set is quarterly and that the sample size is small, they estimated an equation in the
form of (6.54) using four lags of {Ax,}, a constant, and three centered seasonal dummy
variables. The multivariate SBC for the rrodels with four through one lags are -29.05,
-29.99,-30.49, and -31.15, respectively.23 Since the SBC continually decreases as the lag
length is reduced, the authors chose to use the first-order VAR to test for cointegration.
The values of the ),ru." and .\*u* statistics are such that the null hypothesis of no
cointegration (i.e., r:0) is soundly rejected. Also notice that the )ou"" statistic is such
that the null hypothesis ofr ( 1 cannot be rejected and the )-u, statistic is such that
the nuil of r: 1 againstthe alternative of r:2 cannotbe rejected. As such, it is evi-
dent that there is only one cointegrating vector among the four variables.
r:0 r1l r<2 r<3
)r.u". 101.7 22.2 12.6 3.9
)-^, 79.4 9.7 8.7 3.9
Given that there exists a single cointegrating vector, the issue is to see if it is pos-
sible to impose the homogeneity restriction gt + 0z+ gZ: 1. The sample value of 12
is 0.61, and the critical value at the 5 percent significance level is 3.8. Hence, the restric-
tion is not binding; the normalized cointegrating vector with the restriction imposed is
0.426u1c,+ 0.481ip,+ 0.093pet, (6.71)
Pt: 0o+
Equation (6.71) indicates that the long-run elasticity of gasoline prices for the
CPI is approximately 9 percent. The elasticities of unit labor costs and import prices
are about 42 atd48 percent, respectively. Next, a y2 test is used to determine whether
each of the coefficients is statisticatly different from zero. The sample values of 12
forthe exclusion of ulc1,ip,,andpet,ate7.3,29.2,and 2.8, respectively. If you com-
pare these values to 3.8 (i.e., the critical value at the 5 percent significance level), you
can see that only the coefficientfot pet, can be set equal to zero.
Weak Exogeneity Again
If we ignore the deterministic regressors and the stationary dynamics, each variable
has the error-correction form
L,x i, : n - Bs - O.426ulc,-1 - 0.481ip,'1' 0.093pet,-1) +'''
ai@
where the notation x;, (l : 1 to 4) refers to the variable p,, ulc,, ipr, and petr, respectively'
s
Each speed of adjustment coefficient (i.e., each value of a,.) measures the degree
to which the variable in question responds to the deviation from the long-run equilib-
rium relationship. It is interesting to note that the speed of adjustment coefflcient of
the L,prequation is -1.0. As such, the Australian CPI experiences sharp increases with
GENERAL-TO.SPECIFIC MODELING 371
increases in unit labor costs, imporl prices and gasoline prices. On the other hand, the
values of the a; for the unit labor cost, import price and gasoline price equations are
only -0.061, -0.075, and -0.096, respectively. Not only are these coefficients small in
absolute value (implying a slow adjustment), they are of the "wrong" sign. As already
noted, this is not especially troublesome since the Johansen test indicates that the char-
acteristic root of the system implies convergence to the long-run equilibrium.
Next, de Brouwer and Ericsson (1998) go on to formally test whether the four
values of o, are equal to zero. The sample values of 12 for the restrictions a1 : O, a2
:0,cr3:0,andaq:0are25.4, l.89,l.34,and0.43,respectively.Onlythespeedof
adjustment coefficient for the p, equation has a yz value exceeding 3.8. Hence, ulc,,
ip,, andpetl are all weakly exogenous. As such, the CPI equation can be estimated and
tested in a single-equation framework.
Now consider an equation in the form of (6.69) that is estimated using OLS. As
discussed in Appendix 6.1, it is appropriate to conduct inference on the cointegrating
vector if the errors in the equations that determine ulc,, ipr, and pet, are uncorrelated
with e, and if {er} is not serially correlated. The weak exogeneity tests suggest that
the errors in the equations that determine ulcu ipu and pett are not correlated with er.
One way to eliminate any serial correlation in {er} is to include lagged changes of the
endogenous variable and the regressors. Thus, a reasonable specification is
:
Pt 0o + BPlc, + /ziPt + BPet, +A{L)LP t
+ A2(L)L,ulcA + 4(L)Lip;. + Aa(L)Lpet,_1 + e,
or the more general autoregressive distributed lag model
pt -- A{L)prt + Az(L)ulct + Ar(L)ip, + Ao(L)pet,
+ deterministic regressors * e, (6.72)
where variables up to lag four are included in the estimation and the deterministic
regressors consist of a dummy variable to account for changes in tax rates, an inter-
cept, and three centered seasonal dummy variables.
Note that the ADL represented be (6.72) can be estimated by OLS. The individ-
ual coefficients of (6.72) are not especially important. However, the long-run solution
of the estimated equation is such that
pt: -0.094 + 0.5l6ulc, + 0.463ip, + 0.55pet, (6.73)
where the coefFrcient I is obtained asAlQ)lll -lr(l)1.
There are two interesting features of the long-run solution. First, (6.73) looks
remarkably like the long-run equilibrium relationship (6.71) obtained from the
Johansen methodology. Second, (6.72) is quite general in its structure. In particular,
the error-correction model is clearly a special case of (6.72). Towards this end, the
authors used standard /-tests and F-tests to determine whether it is possible to pare
down the ADL model to an error-correcting form. Not only did they test exclusion
restrictions on the various lags, but they were not able to reject these hypotheses:
1. The coefficients of lags two through four on all variables can be set equal
to zero.
2. The sum of the coefficients on ulc.-1, ipr1, and pet,-l minus the coefficient
onprl equals zero.
372 cHAprER 6 colNTEGRATToN AND EBRoR-coRBECTroN MoDELS ouESTIoNS AND EXEBCIsES 373

3. The coeffrcients on the contemporaneous values of Lyu A,ulc,, and L,ip, The Johansen and Juselius tables are extended to ailow for more than five variables in
equal zero. osterwald-Lenum (1992). Also, there is a growing body of work that examines hypothesis
testing in a cointegrated system. Park (1992) develops a nonparametric method for the estima-
The final model (with standard errors in parenthesis) is tion and testing of cointegrating vectors. Johansen and Juselius (1992) and Horvath and Watson
Lp, : 0.0 1 41 L.p et, + 0.07 63 gap,-1 - 0. 089 1 (p/-t - 0.465 ulc,-1 (1993) discuss the testing of structural hypotheses within a cointegration framework. A useful
(0.006) (0.011) (0.007) (0.056) review ofthe hypothesis tests are found in Johansen (1991). Appendix 6.1 and Section 10 dis-
cuss the Phillips-Hansen (1990) method.
- 0.443ip,-1 - 0.092pet,-y) + deterministic regressors * e, (6.74)
If all but one variable is weakly exogenous, that variable can be modeled in a single
(0.0s0) (0.026) equation framework. In an example of the Australian inflation rate, it was shown how to use
where gap, was added a measure of the Australian output gap obtained as the resid- a general-to-specific modeling strategy to obtain a parsimonious model.
ual from a regression of output on a linear time trend. The literature is proceeding in several interesting directions. Friedman and Kuttner (1992)
As it stands, (6.74) is extremely parsimonious as compared with an equation in use cointegration tests to show that significant relationships between money, income, and inter-
an unrestricted VAR or error-correction model. As such, it is important to put the est rates break down in the 1980s. The paper makes an excellent companion piece with this
regression equation through a battery ofdiagnostic tests. Obviously, the residuals of chapter since it also uses innovation accounting techniques. Another interesting money demand
study using the techniques in this chapter is Baba, Hendry and Starr (1992). Gregory and
the final model should appear to be white noise. Moreover, goodness-of-fit measures
Hansen (1992) consider the possibility of a structural break in a cointegrated system. The inter-
and the out-of-sample forecasts should be superiorto those of rival models. There is
cept and/or slope coefficients of the cointegrating vector are allowed to experience a regime
little point in reporting the details ofsuch tests here. Instead, there is a general point shift at an unknown date. Baltagi (2000) contains a collection ofpapers that try to extend the
to be made about cointegrated variables. Suppose lhat xy through x,r, are cointegrated results of panel unit toot tests to cointegrated systems. King, et al. (1991) combine cointegra-
of order C( I , 1) and that the following regression has been estimated by OLS : tion tests with the type of structural decompositions considered in Chapter 5.
xu: 0o+ ]zxzt + '.' + p,{rt+ e,
If x2, through xnr are all weakly exogeneous, and {e,} is not serially correlated, it OUESTIONS AND EXERCISES
is possible to perform inference onthe Bi using standard l-tests or F-tests. If e, is seri
ally correlated, the regression can be augmented by lagged changes (or levels) ofthe
l. Let equations (6.14) and (6.15) contain intercept terms such that

endogenous and exogenous variables. The general specification of a distributed-lag lt= ato* attlrl-| ay2zp1* eyt and zt: a2o* aztlrt:_ a22zr4t er,
model allows for single-equation estimation and testing. As in any OLS equation using a. Show that the solution fory, can be written as
(1) variables, you perform lag length (but not Granger causality or exogeneity) tests.
.yr
: [(1 - a22l)eyt + (7 - a22)arc'r a12Le,1 + onazo) / [(1 - a1{)Q - qzzL) - apa21L2)
11. SUMMARY AND CONCLUSIONS b. Find the solution for 2,.
Many economic theories imply that a linear combination of certain nonstationary variables must c. Suppose thaty,andz,are CI(|,1). Use the conditions in (6.19), (6.20),and(6.21)
be stationary. For example, if the variables {xv}, {rz,}, and !r3,} are (1) and the linear combi- to write the error-correcting model. Compare your answer to (6.22) and (6.23).
natiol e,: ps+ 0f u+ p*zt+ B3x31is stationary, the variables are said to be cointegrated of order Show that the error-correction model contains an intercept term.
(1, l). The vector (pe, 0v 02, 0) is called the cointegrating vector. Cointegrated variables share d. Show that {y,} and {2,} have the same deterministic time trend (i,e., show that the
the same stochastic trends and so carmot drift too far apart. Cointegrated variables have an error- slope coefficients of the time trends are identical).
correction representation such that each responds to the deviation from "long-run equilibrium." e. What is the condition such that the slope ofthe trend is zero? Show that this condi-
One way to check for cointegration is to examine the residuals from the long-run equilib- tion is such that the constant can be included in the cointegrating vector.
rium relationship. Ifthese residuals have a unit root, the variables cannot be cointegtated of order 2. The data file COINT6.PRN contains the three simulated series used in Sections 5 and 9.
(1, 1). Another way to check for cointegration among (1) variables is to estimate a VAR in first You should find the properties of the data are such that
differences and include the lagged level of the variables. If we use a multivariate generalization
ofthe Dickey-Fuller test, the vector can be checked for the presence ofunit roots. In an /? equa- Series Obs Mean Std Error Minimum Maximum
tion system, n minus the number of unit roots equals the number of cointegrating vectors. Y 100 4.2810736793 t.4t48612773 -6.3307043375 -t.2512548288
The ),ru"" and )-u* test statistics can be used to help determine the number of cointegrat- Z 100 1.79sr179043 -s.7040632238 0.6257029853
-2.1437335637
ing vectors. These tests are sensitive to the presence ofthe deterministic regressors included in w 100 -6.3677952867 2.3914380011 -9.6848404427 -1.4460513399
the cointegrating vector(s). Restrictions on the cointegrating vector(s) and/or the speed of
adjustment parameters can be tested using 12 statistics. You should be aware of the role of the a. Use the data to reproduce the results in Section 5

deterministic regressors in a cointegration framework. Johansen (1994) shows how to test to b. Use the data to reproduce the results in Section 9.
determine whether there is a deterministic trend, drift terms outside of the cointegrating vector 3. The file COINT_PPP.XLS contains quarterly values of German, Japanese, and Canadian
or constants appearing in the cointegrating vector. wholesale prices and bilateral exchange rates with the United States. The file also contains
374 cHAprEH 6 cotNTEGRATtoN AND ERRoB-coRREcloN MoDELS OUESTIONS AND EXERCISES 375

the U.S. wholesale price level. The names on the individual series should be self-evident. Using Rl0 as the dependent variable, are the three interest rates cointegrated?
For example, p_zs is the U.S. price level and ex j
is the German exchange rate with the d. Estimate an error-correcting model using 12 lags of each variable. Use the residu-
United States. Al1 variables except the mark/dollar exchange rates run from 1973:Q4 to als from part b as the error-correction term and do not include a separate intercept.
2001:Q4 and all have been normalized to equal 100 in i973:Q4. You should find that the error-corrections are such that
a, Form the log of each variable. Estimate the long-run relationship between Canada
and the United States as
ATBILLT: 4.040e,-, + ... /-statistic for the error-correction term: -0.983
AR3,: 0.088e,_1 + ... ,-statistic for the error-correction term: 2.638
lo g(ex _c a) : 4. 12 + 0.93 7 1o gQt _c a) - 0. 8 3 0log(p_rzs) ARl0,: 0.096e,_1 +... ,-statistic for the error-correction term: 3.741
Do the point estimates of the slope coefficients seem to be consistent with where er_1 is the lagged residual from your estimate in part b.
long-run PPP? i. Perform the appropriate diagnostic tests on the system. In particular, determine
b. Let p, denote the residuals from the long-run relationship. Use these residuals to whether the three residual series appear to be white noise. Are the lags lengths
perform the Engle-Granger test for cointegration. If you use three lagged changes, unnecessarily long?
you should find ii. Discuss the nature of the adjustment. Are any of the rates weakly exogenous?
Ap1: In response to a deviation from the long-run relationship, how are the tfuee
-0.08911.y-1 - 0.063Aprt + 0.048A/-r ra+ 0.236Ap,t2 + vt
rates predicted to change?
The f-statistic on the coeffici ent for p,_1 is -2.17 . Explain why the test indi-
cates that long-run PPP fails. Estimate the model using the Johansen procedure. Use 12 lags and include an inter-
cept in the cointegrating vector. You should find the following:
c. Repeat parts i and ii using Germany and Japan.
List ofcharacteristic roots (i.e., eigenvalues) in descending order:
The second, fourth, and fifth columns of the file labeled INT_RAIES.XLS contain the inter-
est rates paid on U.S. 3-month, 3-year, and l0-year U.S. govemment securities. The data run
0.055 0.025 0.007
from 1954:7 Io 2002:72. These columns are labeled TBILL, R3, and Rl0, respectively. Trace Tests Maximum Eigenvalue Tests
a. Pretest the variables to show that the rates all act as unit root processes.
Null Altemative )ou"" Null Alternative )*u* S%oYalne
SYoYa\rc
Specifically, perform augmented Dickey-Fuller tests using the lag length selected
by the AIC. If you include an intercept (but no time trend) you should obtain
r:0 r)7 50.78 34.91 r: O r: | 32.41 22.00
r <7 r)2 18.38 19.96 r=l r=2 14.45 15.67
Series Lags Estimated a1 /-statistic prob-valte r 12 r:3 3.93 9.24 r=2 r=3 3.93 9.24
TBILL t2 -0.0140 -2.08 0.259 i. Explain why the )trace test strongly suggests there is exactly one cointegrat-
R3 t2 -0.0106 -1.95 0.308 ing vector.
R10 t2 -0.0078 -1.85 0.355 ii. To what extent is this result reinforced by the )*u* test?
b, Estimate the cointegrating relationships using the Engle-Granger procedure. f. Verify that the normalized cointegrating vector is
Perform augmented Dickey-Fuller tests on the residuals. Using TBILL as the TBILL/ :
-0.379 + 1.603R3/ - 0.665R10/
"dependent" variable, you should find
i. Compare this result to your answer in part b.
TBILLT: -0.101 + 1.899R31- 0.997 R10' ii. Veri$, that the three speeds of adjustment parameters (with t-statistics in
(4.147)(4t.7 62) (-2 1. I 83) parentheses) are
where l-statistics are in parentheses.
Perform the Engle-Granger test on the residuals from the equation above.
TBILL: -0.037 R3: 0.082 R10: 0.101
Why is it appropriate to use nine lags in the augmented form of the test? If you
(-0.e30) (2.4s) (3.e8)

use nine lags, you should find that the coefFrcient on the lagged residual (i.e., e,_1) 5. Suppose you estimate zr to be
is -0.1347 with a /-statistic of -5.18.
The 5 percent critical value is about -3.76. Based on this data, do you con-
[o.o -0.5 o.2l
clude that the variables are cointegrated? ,.:10., -o.2s o.1l
Repeat part b using R1 0 as the dependent variable. If you use 4 and 12 lags in the lr., -1.0 o4]
augmented form of the Engle-Granger test, (i.e., estimate Aer: apr_1+ ....) you
should find
a. Show that the determinant ofz-is zero.
b. Show that two ofthe characteristic roots are zero and that the third is 0.75.
lags estimated o, /-statistic c. Let B' :
Q, -2.5, 1) be the single cointegrating vector normalized with respect to
4 -u.u t4 -4.01 x3,. Find the (3 . 1) vector o
such that ir: dB' . How would a change if you nor-
t2 -0.072 -3.35 malized B with respect to x1r?
 376 cHAprER 6 cotNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS ENDNOTES 377

d. Describe how you could test the restriction h + Az: O. a. Use the data set MONEYDEM.XLS to estimate the error-correction model. Use
Now suppose you estimate z.to be: hve lags of each variable.
b. As shown in the manual, you can use the effor-correction model to obtain the
Io.s
tl 0.4 o.ol impulse response functions with a Choleski decomposition. Your responses should
7r: | 0.1 0.1 0.0 I look like:
[o.zs o.2s o5] lmpulse Response Functions
e. Show that the three characteristic roots are 0.0, 0.5, and 0.9.
f. Select B such that: Shocks to the OneYear Rate Shocks to theThree Month Rate
1.00 0.30
lo.s 0.751 0.95 0.25
s:lo.q o.2sl 0.90 o.20

[oo o5l 0.85

0.15
0.10
Find the (3 . 2) matrix a such that r: a0' .
0.80
0.05
6. Suppose thatx1,andx2, are integrated oforders 1 andZ,respectively. You are to sketch the o.75
0.00
proof that any linear combination of x1randx2, is integrated of order 2. Toward this end: 0.70
-0.05
a. Allow xr, and x2rto be the random walk processes 0.65 -0.10
x17: xy-1* y and xt,: xt, r 4 t. 0.60 -0.1 5
05101520 05101520
Given the initial conditions xrg and x29, show that the solutions for x1, and x2,
----3month ----3month
have the form x1, : 110 * Ee r,_, and x2, : xzo * E zt_t
-1year -1year
lr. Show that the linear combination /firt + Azxzr will generally contain a sto- c, In the equation for A.tblyr,, the coefficient on the error-correction term is -{.098 with
chastic trend. a ,-statistic equal to 4.427. Why is it possible to argue that tblyr,is weakly exoge-
iii. What assumption is necessary to ensure that x1, and, x2, are CI(l, l)? nous? How can you model the long-term rate using the general-to-specific approach?
b. Now let x2,be integrated of order 2. Specifically, let Lx2, : Ax2,_1+ zr. Given
initial condition for x2s and x21, find the solution for x2,. (you may allow e1, and ENDNOTES
e 2, to be perfectly correlated.)

i. Is there any linear combination ofxl, and x2rthat contains only a stochas- 1. To include an intercept term, simply set all realizations of one {x,r} sequence equal to unity. In the
tic trend? text, the long-run relationship with an intercept will be denoted by Bo + Blrl / + . . . + 8,x,,: O. Also
note that the definition rules out the trivial case in which all elements of B equal zero. Obviously if
ii. Is there any linear combination ofx1, and x2rthat does not contain a stochas-
all the B,: 0, Bxi : 0.
tic trend? 2. As a technical point, note that if all elements ofx, are (0), it is possible for e, to be integrated of
c. Provide an intuitive explanation for the statement: If x1, ar,d x2, arc integrated of order -1. However, this case is of little interest for economic analysis. Also note that if {xr} is sta-
orders d1 and d2 where dz, dt any linear combination of x1, and x2, iS integrated tionary, Ad.r, is stationary for all d> 0.
of order d2. 3. The issue is trivial ifboth trends are deterministic. Simply detrend each ofthe variables using a
7. The Programming Manual that accompanies this text contains a discussion of nonlinear deterministic polynomial time trend of the form do + cr i + art2 + ....
4. The error-correction term could have been written in the form - 1zrs,_). Normalization
least squares and maximum liketihood estimation. If you have not already done so, "!t(p{r,_t
withrespecttothelong-termrateyields(6.9),wherear:ojp, andB:B2/Bl.Here,thecointegrat-
download the manual and programs from the Wiley Web site.
ing vector is (1, -@.
a. Section 5.5 in chapter 4 contains discussion ofthe problem ofconducting infer- 5. Note that (6.1 1 ) and (6.12) represent a system of first-order difference equations. The stability con-
ence on the parameters of a cointegrating vector. Execute Program 4.10. why is it ditions place restrictions on the magnitudes ofa5, ar, and the various values of ay(k).
a problem that only 16.8 percent of the true values of B1 lie within a 95 percent 6. Equation (6.18) can be written as A2: a1,\-r azwhere ar =(an+ arr) ar,da2= (ana2r- attazz).
confidence interval? Now refer all the way back to Figure 1.5 in Chapter 1. For )1 : l, the coefficients of(6.18) must
lie along line segment BC. Hence, at + a2= 1, or all + a22+ ar2azt - a,az2: 1. Solving for all
b. How would you modify the program so as to generate the Engle-Granger criti-
yields (6.19). For l)21 < 1, the coefficients must lie inside region A0BC. Given (6.19), the condi-
cal values?
tion a2- a1 = 1 is equivalent to that in (6.21).
8. Section 3 of chapter 2 of the Programming Manual that comes with this text estimates 7. Arother interesting way to obtain this result is to refer back to (6.i4). If ap: 0, yt= arflFt + eyr
the relationship between the long-term and short-term interest rate as Imposing the condition that [yr] is a unit root process is equivalent to setting all = 1 so that Ay, : er1.
8. The stability condition is thatl < a1 < 0. Hence, ifar is found to be sufficiently negative, we need
tb Iyr, : 0.698 + 0.976tb3mo, to be able to reject the null hypothesis: ar:2.
378 CHAPTEH 6 COINTEGRATION AND ERROR-CORRECTION MODELS APPENDIX 6.1: INFERENCE ON A COINTEGRATING VECTOR 379

As shown in Section 3, the values ofo, and o, are directly related to the characteristic roots ofthe standard erors. For example, the cointegrating relationship between yb ztu arrd wt it
difference equation system. Direct convergence necessitates that o/ be negative and a, be positive.
Section 5 was reported as (with l-statistics in parentheses)
10. EngleandGranger(1987)doprovideastatistictotestthejointhypothesis dt:ar= 0.However,
their simulations suggest this statistic is not very powerful, and they recornmend against its use. h: -0.4843 - 0.92732,+ 0.97687w, + er,
Ofcourse, x,, and xr, can be CI(2,2) so that the linear combination xr,- B2x2,is stationary. In this (-0.s7s 1x-38.095) (s3.462)
circumstance, 71 = 0 since zt cailL'tot be cointegrated with x1, - 02xy. However, this is a rather
unusual circumstance, so that the literature focuses on the case in which x11 dnd x2, are CI(2,1). However, {err} mal be serially correlated and zr and wt may not be exogenous
t2. Wholesale prices and period average exchange rates were used in the study. Each series was converted variables. As in a traditional regression with stationary variables, you need to correct
into an index number such that each series was equal to unity at the beginning of its respective period
for serial correlation and the problem ofendogenous regressors. To illustrate rhefully
(either 1 960 or 1973). In the frxed exchange rate period, all values of {e,} were set equal to unity.
13. :
A second set of regressions of the form p, 0o + gt| + p, was also estimated. The results using this
modified least squares procedure developed by Phillips and Hansen (1990), consider
altemative normalization are very similar to those reported here. the simple two-variable example
14. The numbers are slightly different from those reported by Johansen and Juselius (1990) due to
rounding. Brz,* e1,
15. As discussed in Appendix 6.2, the Johansen procedure consists of the matrix of vectors of the ^1,-B:+
Azr- e2t
squared canonical correlations between the residuals of.r, and A.rr_, regressed on lagged values of
Ar,. The cointegrating vectors are the rows ofthe normalized eigenvectors. The first equation is the cointegrating relationship and the second indicates
16. You can normalize with respect to any variable with only one caveat. If the true value of p, : 0 (so that {zr\ is the stochastic trend. The notation elt and e2, is designed to illustrate the
that x, does not enter the cointegrating relationship), normalizing with respect to x,, will wildly point that the residuals from both equations are stationary. However, they may be
inflate the other coefficients and their significance levels. In essence, you will have divided all other serially correlated and may be correlated with each other. As such, the second
magnitudes by a coefficient that is estimated to be near zero.
equation is actually quite general since L.z, can be correlated with its own lags and
In the general case of an n-variable system, let rank(r) = r and define the n . r matrices a and B
such tlrat a0 : r. Also, let aland Bl denote the orthogonal complements of a and B. Johansen with values of ;:,.
(1995) obtains the value ofs fiom aL'lB1 once the value ofr has been chosen. Clearly, the relationship between the lwo errors is crucial. We begin with the sim-
i8. Johansen shos,s that this two-step procedure has the following properties: (i) ifthe rank ofn is r ple case wherein:
and there are no (2) components, the procedure picks out the true value of r with a high probabil-
ity, (ii) a value of r that is too low is selected with a limiting probability of zero, and (iii) if there
are (2) components, the procedure will accept no ,I(2) components with a small probability. ddl[:]
19.
Jorgensen, Kongsted, and Rahbek ( 1 996) show how to simultaneously select the values of r and s.
Hence, the characteristic roots are those of a r'lp,
l::,:,,):N [T ;]l
20. Engle, Hendry, and Richard (1983) provide a comprehensive analysis ofvarious types ofexogene-
Case 7: In this circumstance, the errors are serially uncorrelated and the cross-
ity. In general, a variable;r,, is weakly exogenous for the parameter set P if the marginal distribu-
tion of x;, contains no useful information for conducting inference on P. Hence, xi, can be
correlations are zero. Hence, the OLS regression of y, on zt and a con-
exogenous in one econometric model but not another. stant is such that the explanatory variable (i.e., z,) is independent of the
21" However, since {y,} and, {2,} are (1), the test statistics of the null hypothesis B1 :0 and B2- 0 er.ror tefin e1r. As indicated in the text, the OLS estimates of Bg and B1
are nonstandard. can be tested using the nonnal distribution. Hence, /-tests and F-tests are
22. The data shown in Figure 6.5 were provided by David Gruen.
appropriate. If the disturbances are not normally distributed, the asymp-
23. The paper also reports the results ofthe likelihood ratio statistics for testing the reduced lag lengths
These tests reinforce the results of the SBC. totic results are such that f-tests and F-tests are appropriate.
Case 2: In general, eu and e2twlll be correlated with each other so that Ee1,e2,
APPENDIX 6.1: INFERENCE ON A = 0. In order to conduct inference on the parameters ofthe cointegrat-
COINTEGRATING VECTOR ing vector, it is necessary to correct for the endogeneity ofzr. You do this
by including leads and lags of {Az,} in the cointegrating relationship.
The Johansen procedure allows you to test restrictions on one or more cointegrating Hence, you estimate the equation
vectors. However, it is very tempting to use the l-statistics on a coefflcient of a cointe-
grating vector estimated by OLS in the Engle-Cranger methodology. Nevertheless, you yt:00+ pp,+ ...+ 1-1L,z,ar+ loMt+ltNrt+ ... + elt
must avoid this temptation since the coefficients do not have asy,rnptotic l-distributions In essence, you are controlling for innovations in z/ since the equa-
except in one special circumstance. The problem is that the coefficients are super-con- tion is equivalent to
sistent but the standard errors are not. Nevertheless, it is typical for a published study
to report the coefficients ofthe cointegrating vector and the associated l-statistics or lt: 0o + Bgt+ ... * 'l1e2t+t -t 'yoe2t * 1f2t-r + - .. + ett
380 cHAprER 6 cotNTEGRATtoN AND ERRoR-coRRECTtoN MoDELS APPENDIX 6.2: CHARACTERISTIC HOOTS, STABILITY AND RANK 381

Letvar(e1) be denoted by o12.If {e1,} is serially uncorrelated, you APPENDIX 6.2: CHARACTERISTIC ROOTS,
can form a l-statistic to determine whether the estimated value of p1 STABILITY AND RANK
(i e, B,) equals the hypothesized value B1 using the t-statistic
Gharacteristic Roots Defined
t:Gr-A)to" '
Let A be an (n ' n) square matrix with elements aU and let x be an (n 1) vector. The
Case 3: In the most general case, Eeye2,
= 0 and the residuals from the coin- scalar ) is called a characteristic root of A if
tegrating vector (i.e., the estimated values of e1,) are likely to be seri-
Ax: )x (46.1)
ally correlated. Hence, you also need to modify the f-statistic so that
you use the appropriate estimate of the variance of e1,. If the {e1r} Let 1be an (n ' n) identity matrix so that we can rewrite (A6.1) as

series is serially correlated, you adjust the /-statistic using the follow- (A - ),1)x:0 (A6.2)
ing procedure:
Since x is a vector containing values not identically equal to zero, (A6.2) requires that
STEP 1: Estirnate the equation fory, and obtain the estimated {e1r} series. Denote the rows of (A - A1) be linearly dependent. Equivalently, {A6.2) requires that the
the l-statistic for the null hypothesis p1 : B1 as ts. determinant I A - N | : 0. Thus, we can find the characteristic root(s) of (A6.1) by
STEP 2: Estimate the {e1r} series as an AR(p) process to correct for autocorrelation. hnding the values of .\ that satis$r
In particular, use the residuals from Step I to estimate the equation
lA- xl:0 (A'6.3)
e1y: d.1ey-1 + ... + aOey-O* , Example I
Let o2 denote the estimated variance of e, so that o is the standard devia- LetAbe the matrix:
tion. Construct the value ,\ as
a _l
I o.s -o.2
\:o(1 -a1-...-op) ^-l-.o.z o.s
STEP 3: Multiply t6by o"/ ). The resulting value is the appropriate /-statistic for the
so that
null hypothesis gr: gt Compare the corrected l-statistic to that in a t-table.
As you can see, the corrected /-statistic uses a more appropriate estimator
lA-.\1 l=l
Io.s-.r -o.z I

for var(e1r). I

Little is altered if we allow z, to be a vector of variables. However, a word of cau-

| -0.2 o.s-^l
tion is in order. There are many possible sources of error in the three-step methodol- Solving for the value of ) such thatl A : 0 yields the quadratic equation:
- )1 |
ogy outlined above. You could use too few or too many lags in Step 1. A similar .\2-)+0.21 :0
problem arises in Step 2 becausep is unknown. The Johansen procedure circumvents
The two values of ,\ which solve the equation are .\: 0.7 and : 0.3. Hence, 0.7 and
many of these problems in that all variables are treated as jointly endogenous and the
0.3 are the two characteristic roots. ^
VAR residuals are not serially correlated. Hence, you can conduct inference on the
cointegrating vector(s) directly. Example 2
The procedure is not always as difficult as it sounds. As a practical matter, many Now change I such that each element in column 2 is twice the corresponding value
researchers correct for serial correlation by adding lagged changes of Lyrto the esti- in column 1. Specifically,
mated equation. If the augmented equation eliminated the serial correlation, Steps 2
and 3 are unnecessary. The estimated equation has the form A
lo.s 1

^-l_.0.,
-l
-0.4
lt: 0o+ Brzt+ AtQ)Ly._1+... + 1aN,*1 + 1sN,+ 11L,2,_1 + ... + t
where A{L) is a polynomiat in the lag operator L Now,
and {e ,} is serially uncorrelated.
Section l0 contains an interesting example applied to the Australian price
level. The key feature of the example is that the regressors are shown to be
weakly exogenous. -o.4 -^l
Again, there are two values of .\ which solve I A - N | : 0. Solving the quadratic
equation )2 -0.1.\:0 yields the two characteristic roots .\1 :0 and )z = 0.1.
382 cHAprER 6 cotNTEGRATIoN AND ERRoR-coRRECTtoN MoDELS APPENDIX 6.2: CHARACTERISTIC ROOTS, STABILIry AND RANK 383

Gharacteristic Equations The proof of this important proposition is straightforward since the values \y \2,
..., .\, solve (A6.4). Yet, from the algebra of polynomials, the product of the factors
Equation (A6.3) is called the characteristic equation of the square mafrx A. Notice that
of (,4.6.4) is equal to (-l)"b,:
the characteristic equation will be an nth-order polynomial in ). The reason is that the
determinantlA-Nl:0containsthenthdegreeterm)nresultingfromtheexpression:
(a11- \)(a22- ))(a:s - )) ... (onn- \)
Iil=1 ^, = (-r)n bn
As such, the characteristic equation will be an nth-order polynomial of the form From the discussion, we also know that (-l)nbn: I A l. Hence (A6.5) must hold
+ bl^n-r + [r\n-2 + [r)n-3 + ... + bn_l\+ bn: o (A6.4) in that the product ()r)(A) ... (,\,) : (-l)nbn: I A l.
^n
From (46.4) it immediately follows that an (n . n) square matrix will necessarily Examples I to 3 Continued
have n characteristic roots. As we saw in chapter I, some of the roots may be repeat- In Examples I and2,the characteristic equation is quadratic of the form A2 + bi+ b2
ing and some may be complex. In practice, it is not necessary to actually calculate the : 0. To find the roots of this quadratic equation, we seek the factors .\1 and )2 such that
values of the roots solving (A6.4). The necessary and sufficient conditions for all char- ()-.\1x)-)z):o
acteristic roots to lie within the unit circle are given in the appendix to Chapter l.
Notice that the term b, is of particular relevance because b, : (_l)"1 I l. After
all, b, is the only expression resulting from I A - x I that is not multiplied by ). In 1z - (.\.\1 + .\.\j + )1)2 : 0
terms of (A6.4), the expressions )u and bnwlrl have the same sign if n is even and or
opposite signs if n is odd. InExample 1, the characteristic equation is ,\2
-
: 0 so that b2: 0.21. Since i I l: 0.21, it follows that b2: (-l)2(0.21). Similarty, )2-(Ar+)2)^+)1)2:o
^+0.21in
Example 2, the characteristic equation is )2 - 0. I : 0, so that b2: 0. Since it is also Clearly, the value of )1.\2 must equal b2. To check the formulas in Example I , recall
the case thar I A I : 0, it also follows that b2: (-l)21
^ I l. In Example 3 below, we con- that the characteristic equation is )2 - ,\ + 0.21 : 0. In this problem, the value of b2
sider the case in which n : 3. is 0.21, the product of the characteristic roots is }r)z: (0.7X0.3) :0.21, and the
Example 3 determinant of I is (0. S;z - (0.212 : 0.21 . In Example 2, the chatacteristic equation
Let A be such that is ,\2 - 0.1.\:0 so that bz: O. The product of the characteristic roots is )r)z:
(0.0X0.1) : 0.0, and the determinarlt of A is (0.5)(0.a) - (0.2) : 0.
o.2 0.2 In Example 3, the characteristic equation is cubic: )3 - 1.5)2 + 0.63) - 0.081 :
0.5 -.\ 0.2 0. The value of b3 is -0.081, product of the characteristic roots is (0.9X0.3)(0'3):
0.081 , and the determinant of A is 0.081 .
0.2 0.5 -)
The characteristic equation is
Gharacteristic Roots and Rank
.ls - t.s.\z + 0.63.\ - 0.081 : o.'' The rank of a square (n . n) matrix /4 is the number of linearly independent rows
and the characteristic roots are (columns) in the matrix. The notation rank(l) : r means that the rartk of A is equal to
r. The matrix I is said to be of full rank if rank(l) : m.
)1 : 0.9, )2 : 0.3, and .\3 : 0.3 From the discussion above, it follows that the rank of A is equal to the number
The determinant of A is 0.081 so that 63 : -0.08 I: (-l)3 | I l. of its nonzero characteristic roots. Certainly, if all rows of A are linearly independ-
ent, the determinant ofl is not equal to zero. From (46.5) it follows that none of the
Determinants and Characteristic Roots characteristic roots can equal zero if I A | = 0. At the other extreme, if rank(l) : 0,
each element of ,4 must equal zero. When rank(l):0, the characteristic equation
The determinant of an (n ' n) matrix is equal to the product of its characteristic roots, degenerates into ,\u : 0 with the solutions )r : Iz ), : 0. Consider the inter-
that is mediate cases wherein 0 < rank(l): r <n. Since interchanging the various rows of
n a matrix does not alter the absolute value of its determinant, we can always rewrite
lAl=fl), (4.6.s) IA - N I
: 0 such that the first r rows comprise the r linearly independent rows of
i=1
A.The determinant of these first r rows will contain r characteristic roots. The other
where .\1, \2, ...,.\, are the n characteristic roots of the (n . z) matrix A. (n-r) roots will be zeroes.
384 cHAprER 6 corNrEGRATIoN AND ERRoR-coRREcloN MoDELS APPENDIX 6.2: CHARACTERISTIC ROOTS, STABILITY, AND RANK 385

In Example 2, rank(A): I since each element in row I equals -2.5 times the cor- so that the following system of equations must be satisfied:
responding element in row 2. For this case, I A | : 0 and exactly one characteristic
root is equal to zero.In the other two examples, I is of full rank and all characteris- lor- \ al2 at3 aln ll.,l 0 I

tic roots differ from zero. Io^ azz - \ az3 a2n ll.,ll 0
Example 4
Now consider a (3 ' 3) matrix I such that rank(l) : l. Let I

I anl An2 An3 a,r, 0

^l[.;]:l
-.1 0.2 0.2
I
o.s For a nontrivial solution to the system of equations, the following determinant must
le-.lr 1=l I 0.4-.\ 0.4 equal zero:

l-o.zs -o.l -0.1-) - Il aD at3 al,

l(arr
The rank ofl is unity since row 2 is twice row 1 and row 3 is -0.5 times row 1. The I azt Gzz- \) az3 42,,
determinant of I equals zero and the characteristic equation is given by I

)3_0.8)2:0
The three characteristic roots are )1 :0.8, Az:0, and .\3:0.
l,^ an2 an3

The determinant will be an zzth-order polynomial that is satisfied by n values of

).Denote these n characteristics roots by .\1, .\2, . . . ,\r. Since each is a solution to the
Stability of a First-order VAR homogeneous equation, we know that the following linear combination of the homo-
geneous solutions is also a homogeneous solution:
Let x, be the (n ' I ) vector (xtt xzn .. ., *n)' and consider the first-order VAR
xt: AO* Aflrt* t (4.6.6)
xi,: d1)./ + d2\2t + ... + dn^nt

where: Ao: (n . 1) vector with Note that each {x;r} sequence will have the same roots. The necessary and suffi-
elements a,6
At: (n . n) square matrix cient condition for stability is that all characteristic roots lie within the unit circle.
with elements ay
t: (n ' l) vector of white-noise disturbances (e1p e27, ..., nt)'
To check the stability of the system, we need only examine the homogeneous Gointegration and Rank
equation:
The relationship between the rank of a matrix and its characteristic roots is critical in
x1: AP7-1 (A6.7) the Johansen procedure. Using the notation from Section 7, let
We can use the method of undetermined coefficients and for each x;rposit a solu- x1:A1xr-1 *e,
tion of the form
so that
xi1: ciX (A6.8)
N,: (At - I)x,-t + e,
where c; is an arbitrary constant. : 1fxF1 + et
If (46.8) isto be a solution, it must satisfu each of the n equations represented
If the rank of z- is unity, all rows of zr can be written as a scalar multiple of the
by (,4.6.7). Substituting r it= ci\t and xi,_1 -- c.\t-1 for each of the -x,, in (,4.6.7), we get
first. Thus, each of the {&;r} sequences can be written as
c1\t: a1p1)t-1 + at2c2^t'1 + ... + agcn\t-l
Lx i, : s i * r t2x2t-t +
c2)t:
(tr tx t'._t .. .+ T bl nt-t) + it
a21c1\t'1 + a22c2^t-l + ... + a2ncrAt'|
ca\t: a3p1\t-1 + a32c2yt-l + ... + a3ncr\t-l where s1 : I and si: riilrli.
Hence, the linear combination: Tnxl.1 * T12xzt-1 + ... + Tlrlnrt -- (Lxt -
cnX: arpl\,-\ + onz"z\r-l + ... + anncr\t-l e;r)/s; is stationary since Ax;, and ei, are both stationary.
The rank of n- equals the number of cointegrating vectors. If rank(zi.) -- r, there
Now, divide each equation by 1t-t and collect terms to form
are r linearly independent combinations of the {x,r} sequences that are stationary. If
c{ay-\) * c2a12* caaB ...* cnatn:0 rank(n) : n, all variables are stationary.
cp21* c2@22-\) + caa23 ... * cna2r: 0 The rank of z- is equal to the number of its characteristic roots that differ from

c1ar1 * C2dn2t caan3 ... + cn(an -\):0 you to determine the number of roots that are statistically different from zero. The
 386 cHAprER 6 cotNTEGRATIoN AND ERRoH-conHECTroN MoDELS

relationship between A1 and, n- is such that if all characteristic

unit circle, zr is of full rank.
roots of lr are itp the

calculating the characteristic Roots in Johansen,s Method

Although commercially available software packages can obtain the characteristic
roots
of zr, you might be interested in prograrnming the method yourself (or
at least under-
standing the method). First select the most appropriate Laglength p
in for the vAR
xr: Apr_1+ ... + A1X;o+ e,
STEP 1: Estimate the VAR in first differences, that is, estimate
Lx,= BlLxr_l + ...+ BnlAxrp+t+ ett
STEP 2: Regress x7-1 o the lagged changes; that is, estimate a vAR of the form
xr-;,: C1N,_1 + ...+ CnlNt_o+t+ ezt
STEP 3: compute the squares of the canonical correrations between e1,
and e2r. rn an
n -equation vAR, the r canonical correrations
are the , vatues of ),ltne ),
are obtained as the solutions to

l),szz-srzsrrtsirl:o
T7
where S,; :T-rDri,(ei)', Sp :T-tD@2,"t),
and e2, are the column vectors of residuals obtained in Steps
ld
STEP 4: The "r,
maximum likelihood estimates of the cointegrating vectors are the n
7 ard 2.
columns that are nontrivial solutions for

)iS zztr i : 512Sn1Sj 2a.,

ID
4fds*J ]SECOND EDITION
t{ o b"tn \tLvl*

APP..LIED ECONOMETNC
TIME SERTES

Walter Enders
University of Alabama

@wtrEv
www. wiley. com/college/enders