Professional Documents
Culture Documents
Hansen 1992
Hansen 1992
Hansen 1992
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Business &Economic Statistics.
http://www.jstor.org
KEYWORDS: Cointegration;
Fullymodifiedestimation;Quandtstatistic;Parameterconstancy;
Structuralchange.
One potential problem with time series regression sitions were given by Nabeya and Tanaka (1988), Ny-
models is that the estimated parametersmay change blom (1989), and Hansen (1990).
over time. A formof model misspecification,parameter The precedingworks did not consider models with
nonconstancy,may have severe consequences on in- integrated regressors.This article makes such an ex-
ference if left undetected. In consequence, many ap- tension. The test statistics mentioned previously are
plied econometriciansroutinelyapply tests for param- examined here in the context of cointegratingregres-
eter change. The most commontest is the sample split sions,makinguse of the fullymodifiedestimationmethod
or Chow test (Chow 1960). This test is simpleto apply, of Phillips and Hansen (1990). The asymptoticdistri-
and the distributiontheory is well developed. The test butionsof the test statisticsare found to dependon the
is crippled,however, by the need to specifya priorithe stochasticprocessdescribingthe regressors.It emerges
timing of the (one-time) structuralchange that occurs as an importantconclusionthat it is necessaryto know
under the alternative.It is hard to see how any non- the stochasticprocessof the regressorsbefore one can
arbitrarychoice can be madeindependentlyof the data. apply the tests consideredhere.
In practice, the selection of the breakpointis chosen An additionalfinding is that, since the alternative
either with historicalevents in mindor aftertime series hypothesisof a randomwalk in the interceptis identi-
plots have been examined.This impliesthat the break- cal to no cointegration, the test statistics are tests of
point is selected conditionalon the data and therefore the null of cointegrationagainst the alternativeof no
conventionalcriticalvalues are invalid. One can only cointegration.
conclude that inferencesmay be misleading. A relatedresearcheffortby ZivotandAndrews(1992)
An alternative testing procedure was proposed by and Banerjee, Lumsdaine,and Stock (1992) developed
Quandt(1960),who suggestedspecifyingthe alternative a distributionaltheory for the test of the unit-roothy-
hypothesisas a single structuralbreakof unknowntim- pothesis employed by Perron (1989). Perron specified
ing. The difficultywith Quandt'stest is that the distri- the alternativeto be a single structuralbreakof known
butional theory was unknown until quite recently. A timing,but the aforementionedarticlesspecifythe time
distributional theoryfor this test statisticvalidfor weakly of the break as unknown.These articlesaddressa dif-
dependent regressorswas presented independentlyby ferent question (testing the unit-root hypothesis), al-
Andrews (1990), Chu (1989), and Hansen (1990). Chu though using similarmethods.
consideredas well the case of a simplelineartime trend. Section 1 sets up the structureof the model, allowing
Another testing approachhas developed in the sta- for quite generalstochasticand deterministictrendsin
tistics literaturethat specifiesthe coefficientsunderthe the regressors.This model builds on and extends the
alternativehypothesisas randomwalks. Recent expo- setup used by Phillipsand Hansen (1990) and Hansen
321
(1992a). Section 2 describesthe method of fully mod- hand,if a time trendis requiredin the levels regression,
ified estimationfor the model. Section 3 describesthe then x1, = k1, = k, = (1, t)' and there is no k2t. Another
test statisticsfor parameterinstability.Section 4 gives commonspecificationis that there are no trendsin the
the asymptoticdistributiontheory for the test statistics system, so k1t = kt = 1 and there is no k2t.
under the null hypothesis. Section 5 discussesthe al- Some applied researchers have considered using
ternative of no cointegration.Section 6 reports three breakingtrendfunctionsin addition to the simple in-
simple applicationsof the tests. Section 7 concludes. tegerpowersof time consideredhere. Althoughin prin-
Proofs are in the Appendix. ciple it is straightforwardto define the estimatorsand
A GAUSS programthat implementsthe estimation test statisticswith these more general trend functions,
methods and test statisticsis availableon requestfrom this extension will not be consideredin this articlefor
the author. two reasons. First, the restrictionto powers of time
simplifiesthe asymptotictheory. Standardizedpowers
1. A COINTEGRATED
REGRESSIONMODEL of time converge uniformlyto limiting functions, but
We will be workingwith a fairly standardmultiple- this is not true of discontinuoustrend functions. See
regressionmodel under cointegration.The cointegrat- Zivot and Andrews(1992) for an econometricexample
ing equationsare of weak convergencewith discontinuousfunctions.Sec-
ond, breakingtrends may only make sense in a prob-
Yt = Ax, + ul,, t= 1, . . ., n,
abilitymodel if the timingand magnitudeof the break
where the process x, = (xt, x2t)' is determined by the is allowed to be random.This simplyreintroducesI(1)
equations or 1(2) components into the system that are already
capturedin system (1)-(2).
Xlt = klt The followingnuisanceparametersplay an important
H klt + Hk2t + X2t
role in the formulationof the statisticswe will be con-
x2t=
1 2 sidering.Define the matrices
n n
X2t = + (2) 1
Xt-1 U2t
l = lim - E E(uju')
Define the vectors n---,c t=l j=1
1 n t
u = ( ut U2 ); k = (kt k2t). A = lim- E E E(uju) (3)
mi m2 + P2 Pl P2 n->oc n t=l 1j=
timator of Saikkonen (1991) and Stock and Watson In most applications,it appears that the choice of
(1991). (See PhillipsandLoretan[1991]for an informa- kernel is much less importantthan the choice of band-
tive review.) width parameter.For currentconsistencyproofs (see
Hansen 1992b), it is required that M - ooat some rate
2.1 Estimation of Covariance Parameters slower than sample size. This, however, does not pro-
The semiparametricmethod of Phillipsand Hansen vide a useful guide for the selection of the bandwidth
(1990) is a two-step estimator in which the first step parameterin a particularapplication.In a recentarticle,
estimates the covarianceparametersQ1.2 and A1I de- Andrews(1991)providedsome usefulguidelines(based
fined in (4). Our suggestion is to use a prewhitened on the minimizationof asymptotic truncated mean
kernel estimator with the plug-in bandwidthrecom- squarederror) to their selection. He recommendeda
mended by Andrewsand Monahan(in press). We out- plug-inbandwidthestimator.For the Bartlett, Parzen,
line the procedurein this subsection. and QS kernels, the choices are
First, estimate (1) by ordinaryleast squares (OLS), Bartlett: M = 1.1147(c()n)l/3
yielding the parameterestimates A and the residuals
u1, = Yt - Axt. Second, estimate (2) by OLS in differ- Parzen: M = 2.6614(a(2)n)l/5
ences: Ax2, = lAk1,t + I2Ak2t + u2t, yielding the re- QS: M- = 1.3221(a(2)n)15,
siduals u2t.Set u = (uAl,u2t).
The covariancematricesfl and A could be estimated where a(1) and a(2) are obtainedfrom approximating
directlyfrom the residualsut via a kernel. In most ap- parametricmodels. A particularlyattractivechoice sug-
plications,the cointegratingresidualsu1,have a signif- gestedby Andrewsis univariate(autoregressive)AR(1)
icant degree of serial correlation. In this event, the models for each element, eat (a = 1, . . ., p) of e,. Let
kernel estimate will be highly biased, unless a large (pa,6-2) denote the autoregressive and innovation vari-
bandwidthparameteris used, whichincreasesthe vari- ance estimatesfor the ath element. Then
ance of the estimator.In suchcases, an estimatorbased P A P
W2 Ca
on prewhiteningis often preferablein moderatesample i(5(1)
=.4PaOa
--
tawi a1 (1 - a)6( + Pa)2 1 (1 - Pa)
sizes.We suggestusinga (vectorautoregressive) VAR(1),
althougha higher order VAR could also be used. We and
first fit a VAR to the residualsut: ut = ia,_- + et. A p 2 2
/
kernel estimatoris then appliedto the whitenedresid- a^ Pa4a
(2)
uals et. These take the form () = (1 a- p)8 a ( - Pa)
n
I
The use of a plug-inbandwidthparameterhas several
- e t
e tjt
w(jlM) E
n t=j+l
advantages.First, it removes the arbitrarinessassoci-
j=o
ated with the choice of bandwidth.Many applied re-
and searchershave been frustratedwith the semiparametric
n n branchof the unit-rootliteraturebecause the test out-
e = fie
~ w(jlM)-
w(j/M) et_jet , comes sometimesdepend on the choice of bandwidth.
j=-n n t=j+l e et'
Second, simulationresults of Park and Ogaki (1991)
demonstratedthat its use can dramaticallyimprovethe
where w(.) is a weight function (or kernel) that yields
mean squared error of semiparametricestimates of
positive semi-definiteestimates and M is a bandwidth
cointegratingrelationships.
parameter.The estimatorQe can be seen as a scaled
estimate of the spectraldensityof e, (when e, is covari-
ance stationary)and has its origin in the literatureon 2.2 Estimation of the Regression Parameters
spectraldensityestimation,whichdates backto Parzen
Partition A and ft as A and fi, set il1.2 =l -
(1957).
The covarianceparameterestimates of interest can i12i2'f1i21 and A2 = A21 - fikP2lh21 and define
be obtained by recoloring: Q = (I - <)-l'e(I - the transformed dependent variable y+ = y -
)- 1 and A = (I - )-Ae(I - <)-1 - (I
- li"2j2l12t. The FM estimatorof A is then given by
A)--1j , where t = (1/n) Z=1it^;.
These estimatesrequirea choice of kernel and band- A+ y;- .
(0 A1i'))) (xt;)
t=1
width parameter.Any kernel that yields positive semi- t=l
definiteestimatescanbe used.Theseincludethe Bartlett, Associated with these parameterestimates are the
Parzen, and quadraticspectral(QS) kernels. Andrews residualsau = y+ - A +x,. One interestingfeatureof
(1991) recommendedthe QS kernel, which takes the the FM estimates that will be importantfor our later
form
developmentsis that
n
1
w(x) = 122(sin(6(rx/5) - cos(6rx/5)) x
w()=127T2X26=rx5 n t=i w21
althoughin OLS regressionthe sum of the productsof martingaledifferencesis 0: Ho: 62 = 0. One alternative
the regressorsand residualsis identically0. Thus the hypothesis is H3 : 82 > 0, G, = (fi.2 0 V,,)-l; tln E
scores of the problemare the variables 9C,with test statistic
1
S, = (XtU -
(5)
MeanF =
nY tlne
F, where n* = >
tln 3
1.
(21))
which satisfy Sn=1 , = 0.
The final alternativeis H4: 62 > 0, Gt = (f1.2 ?
Mnn)-1, with test statistic
3. TESTS FOR PARAMETER INSTABILITY
Hansen (1990) outlined a general theory of testing Lc = tr{Mnn1 E St? S)}.
for parameterinstabilityin econometricmodels. The
test statisticscan be derivedas Lagrangemultiplier(LM)
The F,, test (fixed t) is computationallysimple, cor-
tests in correctlyspecifiedlikelihoodproblems.In this
section, we describethese test statisticsin the context respondingto the classical Chow test or sample split
test. The test statisticis computationallyequivalentto
of fully modifiedestimationof cointegratedregression
models. estimatingA1 andA2 on the two subsamplesandtesting
their equivalenceusing a Wald test, usingthe variance
We can modify(1) to incorporatepossibleparameter
estimatefor the full-sampleestimates.Thiscanbe easily
instabilityby allowingA to depend on time: seen if we consider the special case of least squares
Yt = A,xt + u,. (1)' estimation on a single equation (ml = 1). Then note
that
For all of the tests, the null hypothesisis that the coef-
ficientA, in (1)' is constant,althoughthe tests differin
the treatmentof alternativehypotheses. Mnt Snt = (E XiXi Xii
Ai = A1, i < t
that is, the score from the first part of the sample,
= A2, i>t. evaluatedat the estimate from the full sample, is pro-
The null hypothesisis Ho: A1 = A2. For the first test, portionalto the difference between the estimates ob-
the timing of the structuralbreak is known under the tainedfromjust the firstpartof the sampleand the full
alternativeH : A1 + A2, t known.A test for Hoagainst sample.It follows(witha little algebra)thatourstatistic
H1 is given by the statistic Fntis essentiallyequivalent to the Wald statistic that
tests the equivalenceof At and A. The only difference
Fn = vec(Snt)' (Q1.20 Vnt)-vec(Snt ) arises due to the choice of the varianceestimates.It is
= tr{SntVntl,SntL1},
well knownthatthisWaldstatisticis algebraicallyequiv-
alent to the classic Chow statistic, which is based on
where the differencebetweenthe estimatesobtainedfromthe
t two subsamples.For example,see SnowandIm (1991).
Sn = S i, (6) The distributionaltheory developed for this test
i=1
(asymptoticchi-squared)is only valid when t can be
chosenindependentlyof the sample.Thisis a restrictive
Vnt = Mnt- MntMnnl'Mn, (7)
assumptionin practiceand may be valid only when t is
and chosen in an arbitraryway, such as t = n/2. In this
t event, the test mighthave low power againstmany al-
Mnt = xiX'. (8) ternativesof interest.
i=l The SupF test dates back to Quandt(1960). Several
For the secondtest, the timingof the structuralbreak recentworkshave exploredthe distributionaltheoryin
is treatedas unknown:H2 : A1 4 A2, [tin] E -, where severalcontexts-those of Andrews(1990),Chu (1989),
5- is some compact subset of (0, 1), and [.] denotes and Hansen (1990). The only difficultyin implemen-
tation is the choice of the region S-. As pointedout by
"integerpart." This test statisticis simply
Anderson and Darling (1952) and emphasizedby An-
SupF = sup Fnt. drews (1990), the region 2- must not include the end-
tln E-
points 0 and 1; otherwisethe test statisticwill diverge
The thirdand fourthtests model the parameterA, as to infinityalmostsurely.The fix suggestedby Andrews
a martingale process: A, = A,_1 + E,; E(E,t-- 1) = 0, is to select T-= [.15, .85]. Although a reasonableap-
E(E,E') = 82G,. In this context, the null hypothesis can proach,this introducesan element of arbitrarinessthat
be written as the constraintthat the variance of the dilutes the appeal of the test.
The MeanF test statistic is derived from a different 7. k, = (1, t, t2, . . ., tP)', p = pi + P2 - 1; and
hypothesis structure but is seen to be simply the average 8. M51n = 0(1).
Fnt test. Although in principle the averaging can include
For the random sequence {ut, the assumptions im-
all values of r for which Fn,can be computed, in practice
some trimming will be required (since F,, will not be pose weak dependence through fairly mild conditions
on the strong mixing coefficients. The moment condi-
defined over all t). Thus the arbitrariness associated
tions are only slightly stronger than square integrability.
with the SupF test is not completely avoided. For the
- Assumption 5 implies that the elements of xt are not
remainder of the article, we set = [.15, .85] as for
the SupF test. mutually cointegrated, that x2tdoes not contain a lagged
The Lc test has a long history in statistical theory, dependent variable, and that xt and the error u1, are
not multicointegrated (see Granger and Lee 1990). As-
although it has not been fully understood until quite
sumption 6 says that the vector k2 plays a role in the
recently. It was first proposed by Gardner (1969) as a
asymptotic behavior of x2t. Assumption 7 restricts the
Bayes test for structural change. It was later indepen- vector kt to integer powers of time.
dently proposed by Pagan and Tanaka (1981), Nyblom Set Yt = jt=l uj. Our assumptions are sufficient for
and Makelainen (1983), and King (1987). These works
the following results:
all concerned tests on a single coefficient in a Gaussian
linear regression model. First attempts at a large-sample
=
distributional theory were made by Nyblom and Make- (1/Vn)Y,,[nr] > B(r) BM(Q); (9)
lainen (1983), Nabeya and Tanaka (1988), and Ley-
[nr] rf
bourne and McCabe (1989). A fairly complete theory
(l/n) YVtu+l > J0 Bdb' + rA; (10)
for maximum likelihood was given by Nyblom (1989) t=l
and was extended to general econometric estimators by
and
Hansen (1990). It has the advantage that it is much
easier to compute than the SupF and MeanF tests and A --- 1A'
, --> n. (11)
requires no decisions for trimming, hence excluding any Here and elsewhere, ">" denotes weak convergence
form of arbitrariness.
of the associated probability measures with respect to
The three proposed tests-SupF, MeanF, and Lc-
the uniform metric, and BM(fl) denotes a Brownian
are all tests of the same null hypothesis but differ in
motion with covariance matrix f. The invariance prin-
their choice of alternative hypothesis. In practice, all
of the tests will tend to have power in similar directions, ciple (9) was shown by Herrndorf (1984). Convergence
to the matrix stochastic integral (10) was shown by Han-
so the choice may be made on the computationalgrounds
sen (1992c). Consistent covariance parameter estima-
that Lc is much easier to calculate. But the appropriate
tion (11) was shown by Hansen (1992b).
test statistic for a particular application should also de-
We need to find a sequence of weight matrices that
pend on the purpose of the test. If the desire is to will appropriately standardize the regressors x, and the
discover whether there was a swift shift in regime, then
estimates A +. We adopt a method from Hansen (1992a).
the SupF test is appropriate. On the other hand, if one
Set n = diag(l, n- , n-2, .. n-P) and k(r) = (1,
is simply interested in testing whether or not the spec-
r, r2, . . ., rP)'. Thus
ified model is a good model that captures a stable re-
lationship, the notion of martingale parameters is more 8nk[nr]
-> k(r) as n - oc (12)
appropriate, since it captures the notion of an unstable
model that gradually shifts over time. If the likelihood uniformly in r. Partition 8, = diag(8n,, 82n) and k(r) =
of parameter variation is relatively constant throughout (k1(r), k2(r))' in conformity with k.
the sample, then the Lc statisticis the appropriatechoice. Equation (2) specifies that the stochastic regressors
x2t are driven by the processes k1,, k2t, and xt. Since k1,
is also in the levels regression, least squares will project
4. DISTRIBUTIONALTHEORY x2, orthogonal to k1,, leaving only k2tand x4. We would
like to isolate the effects of the stochastic trends from
The assumptions we require for the asymptotic dis-
the deterministic trends. Construct an (m2 + P2) x m2
tribution theory are summarized in the following. Let
matrix HI in the null space of I2. The matrix II will
{a,} denote the a-mixing (strong-mixing) coefficients
for {u,}. then annihilate the remaining deterministic component
k2t from x2t. Now define the weight matrix for x2t,
Assumptions. For some q > 3 > 5/2,
1. E(u,) = 0; =2n
2n
( 2J2(n
2 -1 1tf2
)112I)-1 2
\
1/2 )
\(1l/V)(n2*'22n2I)
2. am are of size -ql3/[2(q - /3)];
3. sup,-l EUtlq < o0; and the weight matrix for xt,
4. H as defined in (3) exists with finite elements;
5. H22 > 0 and H1.2 > 0; 8ln 2nO
r = -(
6. rank(I2) = P2; Fn F2nl r2J .
To see that this is a good choice for weightingmatrix, We are now in a position to analyze the asymptotic
note that distributionof these processes.
r =
-= 8nklt
+ X2t)
Theorem 2.
rnxt = Vr2(I2k2t
/ 1nklt \
rrnsn(T)SS*(T)11.2,
(a) ~nFrSn(r)
(a) 1r)2
32nk2t + 32n(I2II2) II2x2 1, (13)
-
(1I/v i)(rl 'f1
k(1/x/~)(II2 2
f22II2) 11/2fI* X
12 X2t] (b) Fn(r) > F(T)on rE
T ;
so by (9) and (12), where S*(r) = S(T) - M(r)M(1)-1S(1), S(T) =
/kl(r) fO5XdW,F(r) = tr(S*(T)'V(T)-1S*(r)),and W1 =
= k2(r) = X(r), say, (14)
B1/2B. BM(Iml), independent of X.
rnX[nr]
W2(r) The process S*(T) is a tied-downversionof the pro-
cess S(r), whichis a continuoustime martingale.Con-
where W2(r) = (11'122112)-/2I* 'B(r) - BM(Im2). ditional on 9; = 0 - r - 1), the sigma field
-r(X(r):
SincefoXX' > 0 for all T> 0 (PhillipsandHansen1990,
generatedby the processX(r), both are Gaussianpro-
lemma A.2), F, is an appropriateweightingmatrixfor cesses. Theirconditionalcovariancefunctionsare given,
the processx,. Equation (14) says that x, is asymptoti- for r1 < T2:
cally dominatedby the trend processeskl(r) and k2(r)
and an m2-dimensionalstochastictrend (W2). E(vec(S(Tr)vec(S(T2))'j!~) = Im2 M(r1)
The test statisticsof Section3 are functionsof partial = Im20 (M(rT)
E(vec(S*(rO))vec(S*(T2))'1;)
samplesums.It willbe convenientto expressthese sums
as functionsof the space [0, 1]. Specifically,define - M(r)M(1)-'M(T2)).
[ni]
The distributionaltheoryfor the limitprocessof S('r)
Mn(r) = Mnln = Z XXi is analogousto the distributionaltheory that arises in
modelswithouttrends;for example,see Nyblom(1989)
and V"(r) = Vn[n = Mn(r) - Mn(r)Mn(1)-lMn(r). or Hansen (1990). In models withouttrends, we find
We can now findthe function-spacedistributionallimits
of these randomfunctionals.
n Sn(T) -= B*(r) = B(r) - (M)M-1B(l)
Theorem1.
= B(T) - TB(1),
(a) lrnMn()r)' M(T)= XX';
a Brownianbridge. In this expression,M > 0 is a con-
stant matrix. The difference between this result and
(b) - rnVn(T)rn> V(T)= M(T) - M(T)M(1)-'M(T); Theorem 2 arises because, in models without trends,
samplecovariancematricesconvergeto constantmatri-
ces. In modelswithstochastictrends,samplecovariance
(c) irnx tu'j XdBl +r A2) matricesare randomvariablesthat change over time.
where ui = ul - ii1u2,;
Thus the expressionfor S*(r) depends on the matrix
processM(r), representingthe samplecovariancestruc-
ture of the regressors.
(d) Vnr 21P (A
.( We can now give expressionsfor the asymptoticdis-
tributionof the test statisticsfrom Section 3.
+- A)rI
(e) Vn-(A dB.2X'' xx , Theorem3.
where B1.2 BM(fi1.2) is independent of X(r), and (a) F-*d i, b = (1 + p + m2)mi;
A21 = (I'22n) -21/2I'2 A21-
(b) supF -d su F(T);
All of the test statisticsconsideredin Section 3 were
functionsof the stochasticprocessSnt.Threeof the tests
were also functionsof the process Fnt.It will be con- (c) MeanF-d f F(r)dr;
venient to write these also as functionals on [0, 1].
Define
Inr] (d) Lc -->
Jd tr(S*(T)'M()- lS*(Tr)).
= =
Sn(T) Sn[n] E Si
i=1 The asymptoticdistributionof the standardF test is
chi-squared.This test, however, as suggested earlier,
Fn(r) = Fn[nr.
has limitedapplicabilitydue to the restrictivenatureof
= vec(Sn(T)) '(V11.2 ) Vgn(T))
-vec(S(T)). the alternativehypothesisinvolved. The other test sta-
tisticsare nonstandardand dependon the natureof the a3), we can easily generate approximateasymptoticp
trends in X (i.e., p, ml, and m2). values automaticallyin the course of calculatingthe
Theorem 3 shows that it is importantto know the statistic.
trend propertiesof the regressorsbefore a parameter To calculate the parameters,I evaluated 38 upper
constancy test can be mounted. Since the asymptotic percentilesfrom the Monte Carlo distributions,from
distributionsonly depend on a few parameters,appro- .20 to .015 in steps of .005. Then I regressedthe per-
priate critical values can be tabulated. If X contains centiles on a third-orderpolynomial in the associate
only deterministictrends, then analytic methods are criticalvalues. For all cases, the fit was very good over
available. Nabeya and Tanaka (1988) derived the this region. Experimentationwith extendingup to the
asymptoticdistributionof Lc for x, = k, (in our nota- .010 percentile indicateda worsened fit, so it was not
tion) and iid errors u1. In this case, they have found done. The estimatedparametersare reportedin Table
expressionsfor the characteristicfunctionof the limiting 1. On their own, they are not interesting. But when
distribution.Theirmethoddoes not immediatelyextend incorporatedinto a computerprogram,they reducethe
to stochastic trends, so here I resort to simulation need to use tables. The polynomials should only be
methods. viewed as approximationsthat can produce p values
The asymptotic distributionsare approximatedby over the region[.20, .015].Thisis not a majorhandicap,
draws from samples of size 1,000 using iid normal since a p value below .20 is rarelytermed"significant."
pseudorandomnumbers. The calculationswere made The distributionaltheory of this section is asymp-
in GAUSS386 using its random-numbergenerator. totic. An investigationof the behaviorof the test sta-
Criticalvaluesfor the three tests are tabulatedin Tables tisticsin finite sampleswas undertakenby Gregoryand
1, 2, and 3 for the single equation setting (ml = 1). Nason (1991). These authorsassessed the testing pro-
The tables includep = 0, 1, 2 and m2 = 0, 1, 2, 3, 4. ceduresdescribedin this articleby applyingthe tests in
For m2 = 0, 1, 2, 25,000 replicationswere made. For the context of a linear-quadraticmodel. Their Monte
m2 = 3, 4, 10,000replicationswere made. The critical Carlodesigninvolvedsamplesizes of 100,200, and500.
values are noticeablydifferentfrom those for the case They found that the tests exhibitedvery little size dis-
of weakly dependent data. (For the supF statistic,see tortionin these samples.They also found that the tests
Andrews [1990, table 1]; for the MeanF and Lc statis- had good power againstsimple structuralbreaksat the
tics, see Hansen [1990, table 1]. first,second,andthirdquarterof the sample.The power
The criticalvalues of Table 1 are useful but require of the tests depended on a cost-of-adjustmentparam-
applied researchersto frequentlylook up tables when eter, which induces serial correlationinto the cointe-
making computations.It is more convenient to have gratingerror. As the degree of serialcorrelationin the
computer packages produce p values along with test errorincreases,the powerdecreases.Thisis not entirely
statistics.What we want is a functionp value = p(x), surprising,because a highly seriallycorrelatederroris
which maps an observedtest statisticx into the appro- close to a randomwalk, whichis equivalentto a random
priatevalue in the range[0, 1]. Thisis of specialinterest walk in the intercept.The abilityof the test to discrim-
when the p value falls into the range [0, .2]. Suppose inate between these two cases breaks down, and the
that we can well approximatethe functionp(x) by a power falls. Overall, Gregoryand Nason's study casts
low-order polynomial: p(x) - aO+ alx + a2x2 + a3x3. a favorable light on the finite-sampleperformanceof
Then if we can obtain the parametersa = (ao, a1, a2, the test statisticsadvocatedhere.
m2 p 1% 5% 10% ao a, a2 a3
0 1 16.4 12.9 11.2 1.954 - .373 .0245 -.00055
0 2 20.0 15.8 14.1 2.487 -.400 .0219 - .00041
1 0 16.2 12.4 10.6 1.960 -.350 .0213 -.00044
1 1 19.0 15.2 13.4 2.666 -.424 .0230 -.00043
1 2 22.0 17.8 15.9 3.480 -.505 .0250 -.00042
2 0 18.6 14.8 13.0 3.182 - .491 .0258 - .00046
2 1 21.4 17.3 15.3 3.652 -.511 .0243 -.00039
2 2 23.9 19.7 17.7 4.003 - .508 .0219 - .00032
3 0 21.0 17.2 15.3 2.882 -.403 .0193 -.00031
3 1 23.9 19.3 17.3 3.248 - .397 .0163 - .00023
3 2 26.0 21.4 19.4 4.488 -.523 .0206 -.00027
4 0 23.6 19.0 17.1 3.522 - .449 .0194 - .00028
4 1 25.2 21.2 19.1 4.030 -.472 .0187 -.00025
4 2 28.0 23.2 21.0 5.341 -.594 .0224 -.00029
NOTE: These tablesareforthe singleequationmodel(ml = 1). Criticalvaluesare calculatedby MonteCarlosimulationusingsamples
of size 1,000. 25,000 replicationswere made form2 = 0, 1, 2. 10,000 replicationswere made form2 = 3, 4.
m2 p 1% 5% 10% ao a, a2 a3
0 1 6.83 4.48 3.73 1.080 -.511 .0843 -.00478
0 2 8.85 6.22 5.11 1.595 -.613 .0818 -.00374
1 0 6.78 4.57 3.73 1.008 -.470 .0773 -.00438
1 1 8.61 6.22 5.20 1.386 -.501 .0629 -.00271
1 2 10.4 7.76 6.50 1.641 -.479 .0479 -.00163
2 0 8.50 6.17 5.18 1.477 -.557 .0729 -.00326
2 1 10.3 7.69 6.58 1.818 - .556 .0607 - .00223
2 2 11.9 9.12 7.88 2.121 -.550 .0489 -.00148
3 0 10.1 7.68 6.66 1.448 - .397 .0370 -.00117
3 1 12.0 9.21 7.89 2.22 - .580 .0520 - .00159
3 2 13.4 10.4 9.15 2.640 -.609 .0480 -.00128
4 0 11.7 9.08 7.87 2.162 -.563 .0505 -.00154
4 1 13.3 10.6 9.28 2.440 -.551 .0426 -.00113
4 2 15.0 11.9 10.4 3.287 -.702 .0512 -.00127
NOTE: See Note to Table 1.
5. A TEST OF COINTEGRATION
AGAINST ventionallydefined, fall into the set of alternativescon-
NO COINTEGRATION sidered by the specificationtests consideredhere?
For simplicity,rewriteModel (1) as
Many applied econometriciansbelieve that it is im-
portant that an econometricmodel be able to survive Yt = A1 + A2X2t + ut; (15)
statisticaltests of the assumptionsunderlyingthatmodel. that is, assume that k1t is simply a constant. Assume
In the case of estimatinga cointegratingrelationship, that y, and x2, are not cointegrated.This is equivalent
a naturalhypothesisto test is thatof cointegrationitself. to the statementthat the error u, is 1(1). Now we can
In contrast, most cointegrationtests, such as those of
decompose ut as ut = W, + v,, where W, is a random
Engle and Granger(1987), Stock and Watson (1988), walk (AWtis white noise) and vt is stationary.Thus
Johansen(1988), and Phillipsand Ouliaris(1990), take
Equation(15) can be writtenas
the null to be no cointegration.The one notable ex-
ception is the spuriousregressortest of Park, Ouliaris, Yt = Alt + A2x2t + vt, (16)
and Choi (1988). where A1t = A1 + Wt. Equation (16) is a special case
The specificationtests developed in Section 3 are of (1)', which is our model of cointegrationwith non-
clearlytests of the model of cointegrationproposedby stationarycoefficients.Specifically,we can see that no
Granger(1981) and developed by Engle and Granger cointegrationis equivalent to one coefficient, the in-
(1987). It is, of course, possible to generalizethe def- tercept, followinga randomwalk. This is a specialcase
inition of cointegrationto allow a nonstationarylinear of the alternativehypothesisfor which the Lc statistic
relationshipbetween the variables,but this would be a is an LM test statistic.We concludethat Lc is a test of
radicaldeparturefrom the idea Grangeroriginallyput the null of cointegrationagainst the alternativeof no
forward. But does the model of no cointegration, con- cointegration.
Table3. AsymptoticCriticalValuesfor Lc
m2 p 1% 5% 10% ao a, a2 a3
0 1 .723 .468 .361 .927 -3.536 4.771 -2.225
0 2 .758 .480 .382 1.120 -3.644 4.155 -1.628
1 0 .898 .575 .450 .769 -3.432 5.471 -3.041
1 1 .959 .623 .497 .996 -3.493 4.311 -1.834
1 2 .999 .654 .520 1.171 -3.421 3.507 -1.240
2 0 1.03 .690 .556 .855 -3.829 6.085 -3.342
2 1 1.13 .778 .625 1.074 -3.658 4.358 - 1.778
2 2 1.19 .814 .666 1.263 -3.511 3.404 -1.133
3 0 1.18 .834 .680 1.247 -3.393 3.235 -1.066
3 1 1.29 .901 .752 1.430 -3.623 3.185 -.959
3 2 1.33 .954 .793 1.496 -3.636 3.075 -.894
4 0 1.31 .934 .780 1.451 -3.515 2.942 -.841
4 1 1.45 1.03 .866 1.694 -3.835 2.992 -.795
4 2 1.51 1.10 .922 1.726 -3.729 2.792 -.716
NOTE: See Note to Table 1.
The SupF and MeanF statistics are not specifically nomics. In a recent article, Campbell (1987) showed
targetedfor the alternativeof random-walkcoefficients, that a strict infinite-horizonpermanent-incomemodel
but they will have asymptoticpower againstthis alter- yields a cointegratingrelationshipbetween aggregate
native as well. consumption and aggregate disposable income. We
Interestingly,there is a connectionbetweenthe SupF can now test the constancyof this cointegratingrela-
test, the MeanFtest, and the spuriousregressortest of tionship.
Park et al. (1988). For each t, F,, is the F test for the The data, from Blinder and Deaton (1985), are sea-
significanceof sample-splitslope dummies.In the ter- sonallyadjustedaggregatequarterlyU.S. consumption
minology of Park et al. (1988), these are "spurious and total disposable income (DI,) in real per capita
trends."By the same argumentused by these authors, units, for the period 1953:2-1984:4. Campbell esti-
each F,, statisticis consistentagainstthe alternativeof mated the equation for both total consumption(TC,)
no cointegration.Therefore, the SupF and MeanFsta- and nondurablesand servicesconsumption(NDS,). In
tistics will be also. the following regressions,a constantand a time trend
There is another intuitive concept linking no coin- were included:
tegrationand parameterinstability.Under the null of
cointegration,regressioncoefficientestimatesconverge TCt = -113 + -1.02t + .982DIt
uniformlyin differentparts of the sample space to the (194) (1.83) (.088)
cointegratingrelationship.Under the alternativeof no M = .89
cointegration,however, the regressionestimates con-
verge to randomvariables,whichwill take on different SupF = 12.3 (.15)
values in different samples. See Phillips (1986). Thus MeanF = 6.2 (.05)
sequential parameterestimates will display apparent
.51
parameterinstability.This simple observationimplies Lc= (.09);
an importantmessage. Rejectionof the null of constant
parametersdoes not implythe particularalternativethe NDS, = 518 + 2.96t + .526DIt
test was designed to detect. There are many possibili- (103) (.97) (.047)
ties. If the SupF test rejects, for example, it would be
AM= .97
quite inappropriateto conclude (on this piece of evi-
dence alone) that there were two cointegratingregimes, SupF = 8.7 (>.20)
which shifted at a particularpoint in the sample. The
MeanF = 3.0
only statisticallyjustifiedconclusionis that the standard (>.20)
model of cointegration,includingits implicit assump- Lc = .14 (>.20).
tion of long-runstabilityof the cointegratingrelation-
ship, is rejected by the data. First, examine the fully modified estimates. It ap-
6. APPLICATIONS pears that the correctionsare having an importantef-
fect. The OLS estimates of the first equation, for ex-
We now apply this testing method to three applica- ample, yield a coefficientfor disposableincome of .85
tions. In each example, the fully modified estimation ratherthan the economicallymore plausible .98. Note
method is used. The covarianceparametersare esti- that the estimatedbandwidthparameterfor both equa-
matedusing a QS kernelon residualsprewhitenedwith tions is less than 1, indicatingthat nearlyall of the serial
a VAR(1). (All of the reportedregressionswere also correlationin the residualswas capturedby the pre-
estimated using the Parzen and Bartlett kernels, and whiteningprocedure.
the results were nearly identical.) The bandwidthpa- The tests when applied to the first equation do not
rameterwas selectedaccordingto the recommendations yield clear results, with p values rangingfrom .05 to
of Andrews (1991), using univariateAR(1) approxi- .15. Although the evidence suggeststhat the relation-
mating models. In all regressionsreported, the esti- ship may indeed be unstable, the data are not suffi-
mates and standarderrors are the fully modified esti- ciently informativeto be able to reject the null of sta-
mates of Phillips and Hansen (1990). The estimated bility. On the other hand, the second equation (for
plug-in bandwidthparameter(M) is reported. All of nondurablesand services consumption)does not sug-
the SupF and MeanF statisticsare calculatedusing the gest instabilityat all, since none of the test statisticsare
trimmingregion [.15, .85]. The constancytest statistics significantat the 20% level. These results (for both
are reported along with their asymptoticp values (in equations)are robustto the choiceof kernelandwhether
parentheses), which are calculated according to the the equationsare estimatedafter taking logarithms.
method of Section 4. It is informativeto visuallyexaminethe sequence of
F statisticsfor structuralchange.Figures1 and2 display
6.1 Aggregate Consumption Function these sequences for each regression, along with 5%
The notion of an aggregaterelationshipbetweencon- criticalvalues for its largest value (SupF), its average
sumptionand income has a long history in macroeco- value (MeanF), and for a fixed known breakpoint.
I I I I I I I I I I I
4-1
0
U)
Ll-
C0
0 ? ? ? ? ?
1958 1962 1966 1970 1974 1978 1982
Figure 1. Total Consumption, 1953-1984: - , F Statistic Sequence; --, 5% Critical, SupF;...... 5% Critical, MeanF; ---, 5% Critical,
KnownBreak.
OD
T- I II I I _I I I I I
' I
C0
0
v--
0
4-J
U) 00
d-..,
0 I I I I I
Figure 2. NDS Consumption, 1953-1984: - , F Statistic Sequence; --, 5% Critical, SupF;...... 5% Critical, MeanF; ---, 5% Critical,
KnownBreak.
the stabilityof each specification: be unity, but the point estimate is significantlyabove
this value. The SupFand MeanFtests statisticssuggests
P, = -.15 + 32.1D,
that the relationshipis not stable. The plot of the se-
(.06) (3.5) quence of F statisticsis displayedin Figure4.
A = .78 Thisevidencesuggeststhatstockpricesanddividends
are indeed cointegrated,but the logarithmicapproxi-
SupF = 6.9 (>.20) mationused by Campbelland Shiller(1988a,b)may be
MeanF = 3.3 (.14) misspecified.
Lc = .30 (>.20);
6.3 Term Structure of Interest Rates
ln(P,) = 4.44 + 1.33 ln(D,) The theory of the term structureof interest rates
(.51) (.12) suggests that, if interest rates can be characterizedas
M = 1.07 1(1) processes,then they shouldbe cointegrated.Stock
and Watson (1988), for example, tested for cointegra-
SupF = 11.7 (.06) tion amongthree postwarU.S. interestratesand found
MeanF = 5.1 evidence of two cointegratingvectors (i.e., only one
(.03)
commontrend). They used monthlydata fromJanuary
Lc = .35 (.18). 1960to August 1979, presumablyto exclude a possible
regime shift in the term structuredue to the change in
The levels equation yields estimates very close to the FederalReserve'soperatingproceduresin 1979.We
those fromOLS (whichgivesa slope coefficientof 31.1). now test the hypothesisthat these relationshipsare sta-
Thiscorrespondsto a long-runreal-interestrateof 3.1%, ble over the entire period from January1960to March
which, as noted by Campbelland Shiller (1987), is be- 1990. We use the same series-the federal funds rate
low the sample mean returnof 8.2%. The relationship (FF), the 90-daytreasury-billrate (TB3), and the one-
appearsvery stable, however, with no significanttest year treasury-billrate (TB12)-and obtainedthe series
statistics. The plot of the sequence of F statistics is from the Citibasedata base.
displayedin Figure3. We report the results of two fully modified regres-
The logarithmicspecification does not perform as sions, TB3 on TB12 and TB3 on FF (the regressionof
well. The modelpredictsthatthe slope coefficientshould TB12 on FF yields resultsvery similarto the regression
00 -
0
I.s._
Figure 3. Stock Prices and Dividends, 1871-1986: , F Statistic Sequence; --, 5% Critical, SupF;...... 5% Critical MeanF; ---, 5%
Critical,KnownBreak.
in
1-
0
1880 1900 1920 1940 1960 1980
Figure 4. Logged Stock Prices and Dividends, 1871-1986: -- , F Statistic Sequence; --, 5% Critical SupF;...... 5% Critical, MeanF;
---, 5% Critical, Known Break.
00
4-
LL
0 -
0 I I I I I.
regression parameters in a cointegrating regression. It Adrian Pagan, Jim Stock, the associate editor, the ref-
should be possible to extend these results to tests on a eree, and seminar participants at the University of
subset of the parameters as well. This will be left to Rochester, Indiana University, and UCLA for helpful
future research. discussions and comments on earlier drafts. This research
was supported by the National Science Foundation.
ACKNOWLEDGMENTS APPENDIX: PROOFS OF THE THEOREMS
An earlier version of this article was entitled "Testing Proof of Theorem 1
for Structural Change of Unknown Form in Models
With Non-stationary Regressors." My thanks go to Don (a) The finite-dimensional result is immediate from
Andrews, Bill Brown, David Hendry, Peter Phillips, (14). Weak convergence follows from the continuous
0-
Co
Ic
U- r _
0
1960 1964 1968 1972 1976 1980 1984
and Federal Funds Rates, 1960-1990: --
Figure 6. Three-MonthTreasury-Bill , F StatisticSequence; --, 5% Critical,SupF; .....
5% Critical, MeanF; ---, 5% Critical, Known Break.
-
r2nA2+1 = | -1/2 2 (A21 - &22f221 21l) Proofof Theorem3
l(2*' 2\2 /2*)
Part(a) followsfromTheorem2 (b) and the fact that
0O conditional on 9x, for any r E ST,vec(S*(T)) = N(0,
II(n*' 1i22n 1/2
1.2 ( V(r)), and therefore
n1/
F(T)l -- X2b
Income?An AlternativeTest of the PermanentIncome Hypoth- Nyblom, J. (1989), "Testingfor the Constancyof ParametersOver
esis," Econometrica,55, 1249-1273. Time," Journalof the AmericanStatisticalAssociation,84, 223-
Campbell,J. Y., and Shiller,R. J. (1987), "CointegrationandTests 230.
of PresentValueModels,"Journalof PoliticalEconomy,95, 1062- Nyblom, J., and Makelainen,T. (1983), "Comparisonsof Tests for
1088. the Presence of Random Walk Coefficientsin a Simple Linear
(1988a), "Stock Prices, Earningand Expected Dividends," Model,"Journalof the AmericanStatisticalAssociation,84, 856-
TheJournalof Finance,43, 661-676. 864.
(1988b),"The Dividend-PriceRatio and Expectationsof Fu- Ogaki, M., and Park, J. Y. (1990), "A CointegrationApproachto
tureDividendsandDiscountFactors,"Reviewof FinancialStudies, EstimatingPreferenceParameters,"Paper236, RochesterCenter
1, 195-228. for EconomicResearch.
Chow, G. C. (1960), "Testsof EqualityBetweenSets of Coefficients Pagan,A. R., andTanaka,K. (1981), "A FurtherTest for Assessing
in Two LinearRegressions,"Econometrica,28, 591-605. the Stabilityof RegressionCoefficients,"unpublishedmanuscript.
Chu, C.-S. J. (1989), "New Tests for ParameterConstancyin Sta- Park,J. Y., andOgaki,M. (1991),"Inferencein CointegratedModels
tionaryandNonstationaryRegressionModels,"unpublishedman- UsingVAR Prewhiteningto EstimateShortrunDynamics,"Work-
uscript,Universityof Californiaat SanDiego, Dept. of Economics. ing Paper281, RochesterCenterfor EconomicResearch.
Engle, R. F., and Granger,C. W. J. (1987), "Co-integrationand Park,J. Y., Ouliaris,S., and Choi, B. (1988), "A New Approachto
Error-correction: Representation,EstimationandTesting,"Econ- Testing for a Unit Root," mimeo, Cornell University,Dept. of
ometrica,55, 251-276. Economics.
Gardner,L. A., Jr. (1969), "On DetectingChangesin the Mean of Park, J. Y., and Phillips,P. C. B. (1989), "StatisticalInferencein
NormalVariates,"TheAnnalsof Mathematical Statistics,40, 116- RegressionsWithIntegratedProcesses,Part1," EconometricThe-
126. ory, 4, 468-497.
Granger,C. W. J. (1981), "Some Propertiesof Time Series Data Parzen, E. (1957), "On ConsistentEstimatesof the Spectrumof a
and Their Use in EconometricModel Specification,"Journalof StationaryTime Series,"TheAnnalsof Mathematical Statistics,28,
Econometrics,16, 121-130. 329-348.
Granger,C. W. J., and Lee, T.-H. (1990), "Multicointegration," Perron,P. (1989), "The Great Crash,the Oil Price Shock, and the
Advancesin Econometrics,8, 71-84. Unit Root Hypothesis,"Econometrica,57, 1361-1401.
Gregory,A. W., and Nason, J. M. (1991), "Testingfor Structural Phillips, P. C. B. (1986), "UnderstandingSpuriousRegressionsin
Breaks in CointegratedRelationships,"unpublishedmanuscript, Econometrics,"Journalof Econometrics,33, 311-340.
Queen's University,Dept. of Economics. Phillips,P. C. B., and Hansen, B. E. (1990), "StatisticalInference
Hansen, B. E. (1990), "LagrangeMultiplierTests for Parameter in InstrumentalVariablesRegressionWithI(1) Processes,"Review
Instabilityin Non-linearModels," unpublishedmanuscript,Uni- of EconomicStudies,57, 99-125.
versityof Rochester,Dept. of Economics. Phillips, P. C. B., and Loretan, M. (1991), "EstimatingLong-run
(in press a), "EfficientEstimationand Testing of Cointe- EconomicEquilibria,"Reviewof EconomicStudies,58, 407-436.
gratingVectorsin the Presenceof DeterministicTrends,"Journal Phillips,P. C. B., and Ouliaris,S. (1990), "AsymptoticProperties
of Econometrics,52. of Residual Based Tests for Cointegration,"Econometrica,58,
(in press b), "ConsistentCovarianceMatrixEstimationfor 165-193.
DependentHeterogeneousProcesses,"Econometrica,60. Quandt,R. (1960), "Testsof the HypothesisThat a LinearRegres-
(in press c), "Convergenceto StochasticIntegralsfor De- sion SystemObeysTwo SeparateRegimes,"Journalof theAmer-
pendentHeterogeneousProcesses,"EconometricTheory,8. ican StatisticalAssociation,55, 324-30.
Herrndorf,N. (1984), "A FunctionalCentral Limit Theorem for Saikkonen,P. (1991), "AsymptoticallyEfficientEstimationof Coin-
WeaklyDependentSequencesof RandomVariables,"TheAnnals tegratingRegressions,"EconometricTheory,7, 1-21.
of Probability,12, 829-839. Snow, M. S., and Im, E. I. (1991), "The Equivalenceof Two Test
Johansen,S. (1988), "StatisticalAnalysisof CointegrationVectors," Statisticsfor Testing the Constancyof RegressionCoefficients,"
Journalof EconomicDynamicsand Control,12, 231-254. EconometricTheory,7, 419-420.
King, M. L. (1987), "An AlternativeTest for RegressionCoefficient Stock, J. H., and Watson, M. W. (1988), "Testingfor Common
Stability,"TheReviewof Economicsand Statistics,69, 379-381. Trends,"Journalof theAmericanStatisticalAssociation,83, 1097-
Leybourne,S. L., and McCabe, B. P. M. (1989), "On the Distri- 1107.
bution of Some Test Statisticsfor CoefficientConstancy,"Bio- (1991),"A SimpleEstimatorof Cointegrating Vectorsin Higher
metrika,76, 169-177. OrderIntegratedSystems,"unpublishedmanucript.
Nabeya, S., and Tanaka,K. (1988), "AsymptoticTheoryof a Test Zivot, E., andAndrews,D. W. K. (1992), "FurtherEvidenceon the
for the Constancyof RegressionCoefficientsAgainstthe Random GreatCrash,the Oil-PriceShock,andthe Unit-RootHypothesis,"
Walk Alternative,"TheAnnals of Statistics,16, 218-235. Journalof Business& EconomicStatistics,10, 251-270.