This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

2 (1990)

0305-9049 $3.00

**MAXIMUM LIKELIHOOD ESTIMATION AND INFERENCE ON COINTEGRATION - WITH APPLICATIONS TO THE DEMAND FOR MONEY
**

Seren Johansen, Katarina Iuselius

I.

INTRODUCTION

1.1. Background

Many papers have over the last few years been devoted to the estimation and testing of long-run relations under the heading of cointegration, Granger (1981), Granger and Weiss (1983), Engle and Granger (1987), Stock (1987), Phillips and Ouliaris (1986), (1987), Johansen (1988b), (1989), Johansen and Juselius (1988), canonical analysis, Box and Tiao (1981), Velu, Wichern and Reinsel (1987), Pena and Box (1987), reduced rank regression, Velu, Reinsel and Wichern (1986), and Ahn and Reinsel (1987), common trends, Stock and Watson (1987), regression with integrated regressors, Phillips (1987), Phillips and Park (1986a), (1988b), (1989), as well as under the heading testingfor unit roots, see for instance Sims, Stock, and Watson (1986). There is a special issue of this BULLETIN (1986) dealing mainly with cointegration and a special issue of the Journal of Economic Dynamics and Control (1988) dealing with the same problems. We start with a vector autoregressive model (cf. (1.1) below) and formulate the hypothesis of cointegration as the hypothesis of reduced rank of the longrun impact matrix n = api, The main purpose of this paper is to demonstrate the method of maximum likelihood in connection with two examples. The results concern the calculation of the maximum likelihood estimators and likelihood ratio tests in the model for cointegration under linear restrictions on the cointegration vectors p and weights a. These results are modifications of the procedure given in Johansen (1988b) and apply the multivariate technique of partial canonical correlations, see Anderson (1984 )or Tso (1981 ). For inference we apply the results of Johansen (1989) on the asymptotic distribution of the likelihood ratio test. These distributions are given in terms of a multivariate Brownian motion process and are tabulated in the Appendix. Inferences on a and P under linear restrictions can be conducted using the usual X2 distribution as an approximation to the distribution of likelihood ratio test. We also apply the limiting distribution of the maximum likelihood estimator to a Wald test for hypotheses about a and p.

**170 Consider the model
**

HI :XI=UIXr_1

BULLETIN

1.2. The Statistical Model

+ ... + UkXr-k + P. + 4JDr+ e.; (t = 1, ... , T),

(1.1)

where E1'''',£r are IINp(O, A) and X-k+I, ... ,Xo are fixed. Here the variables Dr are centered seasonal dummies which sum to zero over a full year. We assume that we have quarterly data, such that we include three dummies and a constant term. The unrestricted parameters (p., 4J, UI, ...,Uk, A ) are estimated on the basis of T observations from a vector autoregressive process. For a p-dimensional process with quarterly data this gives Tp observations and p + 3p + kp? + p(p + 1)/2 parameters. In general, economic time series are non-stationary processes, and VARsystems like (1.1) have usually been expressed in first differenced form. Unless the difference operator is also applied to the error process and explicitly taken account of, differencing implies loss of information in the data. Using !'l = 1 - L, where L is the lag operator, it is convenient to rewrite the model (1.1) as !'lXr = rlt'lXr-! where (i = 1,... , k - 1), and

( 1.3)

+ ... + I', -'I!'lXr-HI

+ UXr-k + P. + «I»Dr + e;

(1.2)

Notice that model (1.2) is expressed as a traditional first difference VA Rmodel except for the term nx k. It is the main purpose of this paper to investigate whether the coefficient matrix U contains information about long-run relationships between the variables in the data vector. There are three possible cases:

l_

(i) Rank(U) = p, i.e. the matrix U has full rank, indicating that the vector process XI is stationary. (ii) Rank(U) = 0, i.e, the matrix n is the null matrix and (1.2) corresponds to a traditional differenced vector time series model. (iii) 0 < rank(n ) = r < p implying that there are p x r matrices a and {J such that n = a{Jl. The cointegration vectors (J have the property that p'X! is stationary even though XI itself is non-stationary. In this case (1.2) can be interpreted as an error correction model, see Engle and Granger (1987), Davidson (1986) or Johansen (1988a). Thus the main hypothesis we shall consider here is the hypothesis of r cointegration vectors

H2

:n= api,

(1.4)

where a and pare p x r matrices.

INFERENCE

ON COINTEGRATION

171

We further investigate linear hypotheses expressed in terms of the coefficients p., a and P, and in particular the relation between the constant term and the reduced rank matrix n. If n is restricted as in H2, see (1.4) and Il t- 0 the non-stationary process XI has linear trends with coefficients which are functions of Il only through a'l u; where a 1 is a p x (p - r) matrix of vectors chosen orthogonal to a. Thus the hypothesis", = ap;), or alternatively «,» = 0, is the hypothesis about the absence of a linear trend in the process. Note that when Il = afJ;) we can write

the calculations. Since the asymptotic distributions of the test statistics and estimators depend on which assumption is maintained, it is important to choose the appropriate model formulation. This has been pointed out for instance by West (1989), Dolado and Jenkinson (1988). The mathematical results for the multivariate model (1.2) are given in Johansen (1989). J.3. The Data We have chosen to illustrate the procedures by data from the Danish and Finnish economy on the demand for money.' The relation m =f(y,p, c) expresses money demand m as a function of real income y, price level p and the cost of holding money c. Price homogeneity was first tested and since it was clearly accepted by data the empirical analysis here will be for real money, real income and some proxies measuring the cost of holding money. Money, income and prices were measured in logarithms, since multiplicative effects are assumed. The two data sets differ both as to which variables are included and the length of the sample. More interestingly, however, the institutional relations in the two economies have been quite different in the sample period. In Denmark, financial markets have been much less restricted than in Finland, where both interest rates and prices have been subject to regulation for most of the sample period. One would expect this to show up in the empirical results and it does. For the Danish data the sample is 1974.1-1987.3. As a proxy for money demand m2 was chosen because the data available on a quarterly basis are based on more homogeneous definitions for m 2 than for m 1. The cost of holding money was assumed to be approximately measured by the difference between the bank deposit rate, id, for interest bearing deposits (which are the main part of m2) and the bond rate, i'', which plays an important role in the Danish economy. The two interest rates were included unrestrictedly in the analysis, but subsequently tested for equal coefficients with opposite signs. The inflation rate, !1p, was also included as a possible proxy for the cost of

I For a general review of theoretical and empirical results on the demand for money, see for instance Laidler ( I 985).

**aP'X i=« + = aP'X I-k + afJ'0 = ap""X* I-k' where fJ* = (fJ', fJ~), and Xi _k = (X; - k, 1)'. This is useful for
**

II f'"

_..----------------~--------~---------- ., ---..

.

172

BULLETIN

holding money, but since it did not enter significantly into the cointegration relation for money demand it was omitted from the present analysis. For the Finnish data the sample is 1958.1-1984.3. In this case m1 was chosen since the m 1 cointegration relation was found to enter the demand for money equation more significantly and hence illustrated the methodology better. Since interest rates have been regulated, a good proxy for the actual costs of holding money is difficult to find. The inflation rate, Ap, is a natural candidate and therefore is included in the data set. Moreover, the marginal rate of interest, i'", of the Bank of Finland is included in spite of the fact that the marginal rate measures restrictedness of money rather than the cost of holding money. It has. however, been chosen as a determinant of Finnish money demand in other studies and therefore is also included here. All series are graphed in Figure 3 and Figure 4 in Section IV. The data are available from the authors on request. The paper is structured as follows: Section II discusses the various hypotheses we shall investigate and in Section III the notation is introduced for the maximum likelihood procedure. The next section derives the estimates of a and p under-the assumption of cointegration and the last two sections investigate estimates and tests for P and a under linear restrictions. Throughout, the two examples are used to motivate the statistical analysis and to illustrate the mathematically derived concepts,

II.

A CLASSIFICATION

OF THE VARIOUS HYPOTHESES

The hypotheses we consider consists of the hypothesis H2 on the existence of cointegrating relations combined with linear restrictions on either the cointegrating relations or their weights: H2:D=aP', H3: D= aplH'

H4: (or P=Htp), a = A 1/1),

n = A ",P' (or

**Hs: n = A "pin' (or P = Hp and a = Atp),
**

and Hj is Hj augmented by II- = ap ~for j = 2, ... , 5. Note that the hypothesis HI' where n is unrestricted, can be written as Hz with r =p. Hence, in this case the restriction II- = ap~ is the same as having IIunrestricted. When we estimate model (1.2) under the hypothesis D = ap' the choice of hypothesis about II- becomes important. For the Danish data there does not seem to be any linear trend in the non-stationary processes (cf, Figure 3) and we wiJI estimate models of the form Hj. For the Finnish data, however, there seems to be a linear trend in the non-stationary processes (cf. Figure 4) and models of the form Hj will be estimated. The matrices A( p x m) and H( p x s) are known and define linear restrictions on the parameters a (p X r) and p (p X r). The restrictions red uce the

p = (I{3~ n = A !/I{3'. and the space spanned by a. starting with the most general VAR model (HI) and introducing the restriction of cointegration (H2) as well as linear restrictions on the cointegration vectors (fJ) and the weights (a) in 11. . In fact all hypotheses arc special cases of H. The important distinction between the H and the H* hypotheses is that H* define a restriction on u. All these hypotheses are restrictions of the matrix D which under H) contains p2 parameters. the choice a~ and fJ(~')-1 will give the same matrix D. Note also that for each value of r(O:S r:5 p) there is a corresponding hypothesis li2(r) of r or fewer cointegrating relations. usually m 1 and m2 respectively. P = a{3~ Fig.) is indicated by a". The relations between the various hypotheses are illustrated in Figure 1. = a{3~ II=A'/Iq/H' II = A '/Irp'H'. the cointegration space. I. Finally mr+(s -r) r parameters remain under H5• Note also that the parameters a and fJ are not identified in the sense that given any choice of the matrix g(r X r). Under the hypothesis H2 there are pr+(p .INFERENCE ON COINTEGRATION 173 parameters to rp(S Xr) and 1/J(m xr). The analysis makes it H* 3 II =A H* 4 !/I{3' II = arp'H'. In general we present the results normalized by the coefficient of some of the variables. namely that it lies in the space spanned by a or that a ~Ii = 0. where r~s:Sp and r~m~p. if we choose either A or H as the identity matrix. but can easily be extended to include the case of H*. and H4• The assumption of no trend f.r) r under H4. hence that no trend is present. and that H3 c H2 and H4 eli]. One way of expressing this is to say that what the data can determine is the space spanned by the columns in fl.r) r parameters which are further restricted to sr +(p . In the above scheme note that H5 = H4 n H. In the following the discussion will be concerned with the H hypotheses.l = ap. and hence determine the same probability distribution for the variables. f1. The relation between the various hypothesis studied.

..1 p = O. T).. -1~Xt-k+1 + nXI_k + Jl + CJ)DI+ £1' (3. (3. III....••~ possible to conduct inferences about the value of r by testing H2(r) in HI or by testing H2(r) in Hk + 1).. Similarly. + I'.and n. but mention how results are modified when a'1 P = O. We first introduce the notation ZUt = ~XI' ZII denotes the stacked variables ~XI_I.~-----------------------174 BULLETIN •. We shall give details in the case of hypotheses about a and P without restricting #-. We now consider maximum likelihood estimation of the parameters in the unrestricted model: ~Xt= fJ~Xt_1 + . maximum likelihood estimation consists of a regression of ZOt . 1. since all the models we are interested in are expressed as restrictions on #. The model with p = 0 and CJ)= 0 was analysed in Johansen (1988b) and Johansen and Juselius (1988).. THE MAXIMUM LIKELIHOOD PROCEDURE In the following we will use the parameterization (1. and Zkl=X k. The reason for this is that the parameters (f.. I'k -I' CJ) and p....2) For a fixed value of n..5) +nMkl...9) are well-known but reproduced to establish the notation. . .2)-(3.1) The results (3. . .nzh. --- --. n. it is possible to maximize over all the other parameters once and for an.. . Generally we use a superscript * to indicate that we are analysing a model where a '. (3.4) Then (3. I' is the matrix of parameters corresponding to Z II' i.2).D" and 1. on Zit giving the normal equations L: t= ITT ZOtZ"I=f L I ZIIZ'II+n L ZkIZ'II' (3.. .3) I The product moment matrices are denoted: Mil = T- I L t= ZitZ." I (i. The model expressed in these variables becomes I- (/=1..3) can be written as: MOl =fM11 or (3.. . . #-. the matrix consisting of I'1. . k)..j=O. ..fk-l. This will be useful for discussing the estimators and tests later. <ll.e. A) are variation independent and. Thus Zll is a vector of dimension p( k -1) + 3 + 1 and I' is a matrix of dimensionp x(p(k -1) +3 + 1).~Xt_k+l.

MOl M11IZI Rk.MklMlllZli' i' i.~k=(X.:. f k _ I .::.. k). . Model (3.::.-k' Further.10) -I and A ..--- ---.:.. ""' ..6) (3. Note in particular that Stk is (p + 1) x (p + 1).-k' 1).... the residuals we would obtain by regressing tJ.l + q..~. :ssas:ss&:: ••••.SOkSkk SkO' L.. ---.e.Soo .•. the test statistic for the residuals in the . and formulate these well-known THEOREM 3_1: In the model: k -I HI: ~XI the parameters = I f. -----------. For k = 2. Similarly we define Mij and Sij.. INFERENCE ON COINTEGRATION 175 (3. -.3. Ror= ZOr.).7) ~Xr and XI .-..5) gives the estimate of f.l = aJ3.:.=Mr)-M'IMI-IIMI" results in (i. q..D1 + EI' I~.1. For the Finnish data. Ihe Empirical Analysis a/the Unrestricted Model H.}F(HI)=IAI.. (3. The concentrated likelihood function becomes: (3_8) We express the estimates under the model HI by introducing I the notation (3.DI' whereas zt:=X.• .•-.~~ ..k on This leads to the definition of the residuals 11 l~ -...j=o. it is convenient to define Z~( = ZOI= ~XI and let Zfl be the stacked variables ~XJ-l"".. ~ •. . Thus we have moved the constant from the regressors into the vector X.1) including a constant term and seasonal dummies is fitted to the Danish and Finnish money demand data described in Section 1. .. the residuals for the Danish data passed the test for being uncorrelated (see Table 1 below).~XI-k+I.. we define I'" as the matrix of the relevant parameters I'I .9) Sij= r: I . Under the hypothesis Ht: n = aJ31 and /. The estimate of •- (3.~~XI_j+ nXI_k + /.= Zk. •.~XI_k+l' D" and 1. 3. which will be investigated for the Danish data.XI_I> ..I are estimated by ordinary least squares and we have: U=SOkSl:kl.-I RirR.11 ) (3_12) n inserted into (3. •.... ""' "---'.

019 T. The normality assumption is tested by the Jarque & Bera test (Jarque and Bera.8). The robustness of the ML cointegration procedure for deviations from normality has not been investigated so far.019 19. (4. . ~.So"P(P'SkkP)-1 p'SU) = SOO d(P) (P'SkkP) d(P)'. For the Finnish data the residuals from the !1i'" and /1p equations do not pass the test.034 TIrf(i= T'<m ( I ..30 10. 10) . model (3.X'(2). For fixed p. it is easy to estimate a and A by regressing ~IJ on P'R'-k to obtain: a(p) = SOkP(prSkkPt 1.1:16 0... The estimation of I'1"". .007 Ai'/ 7. which . error of regression estimate.06 0. and reported below. They are. After conditioning on the first two data realizations.99 28. but since the seasonal autocorrelation is rather small we have chosen to ignore this.02 0. Accordingly.l is the same as before leading to (3.88 0. EK') SK-+--"4 m . T2=-(. where n = apr.x~(I0). TABLE I Some Test Statistics for the niid Assumption for the Residuals in the Model (1. p.48 1.--. they are not reported.045 Ay Ai'" Ay Ap 6.61 0.. 1980)..1) . I'k _[.1) with k = 2 was fitted to both data sets. = 11. cJ» and J. The estimates of n are reported in Table 7 Section VI. Since the parameter estimates of r I.005 Ami 11. The deviations from normality are mainly due to too many large residuals. DERIVATION OF THE ESTIMATES OF (l AND f3 UNDER THE HYPOTHESIS n = (lfJ' AND THE UKELIHOOD RATIO TEST FOR THIS HYPOTHESIS Consider the model H2.~-----------------------.-176 BULLETIN equation for ~y is almost significant The autocorrelogram suggests that there is some seasonality left in the residuals. SK is skewness and EK is excess kurtosis.011 ur where 7.029 4.93 0. approximately symmetrically distributed around zero.2) With k=2 The Danish data Am2 TI T. the number of observations left for estimation was 53 in the Danish and 104 in the Finnish data. and A are not of particular interest in this paper. - (4.30 1. a.2) ":::: A(P)= Soo. The Finnish data Ail> 10.34 1.57 1. and the standard error of regression estimates in Table 1 below. . however. probably is less serious than a skewed distribution.15 2_12 0.61 0. is the standard m is the number of regressors..- . IV.-~-----------..21 1.

.1: Under the hypothesis H2:D=aP'. II =U. one proceeds to ISoo= SOkf3(f3'Skkf3) . The likelihood ratio test statistic for the hypothesis HI is a special case of H2 for the choice r = p.M2).INFERENCE ON COINTEGRATION 177 (4.V Skk V = I.2In(Q. where MI =Skk' M2 = SkOSi)(11SOk' The results will be summarized in THEOREM 4. where A is symmetric..B I/IP'MIP I can be minimized by solving the equation 1 AMI.B'Sko I = I s. Now (4. '" i > ip and eigenvectors V= (v 1"'" VI') normalized P (4.. is: .8) This is based on the general result where in this application A ~Sllil' B = SlIkP and C = P'SkkP. the maximum likelihood First solve the equation estimator of fJ is found by the following procedure: (4.. (4.B'Skk..). 'Many computer packages contain procedures for solving the eigenvalue problem I AI .5) is equivalent to IAI -C' ISkOS"OSlIkC' which has the same eigenvalues are then found as v.1.B I I 1. 'I 'he eigenvectors of (4..5) giving the eigenvalues I>'" such that.AI = 0.5) .B'SkOSi)ol SOk.SkiJSiuiSOk) I· (4. for some non-singular p x p matrix C.I . i > . The choice of is now \ A . ••• .3) and As shown in Johansen (1 988b) estimate f3 by applying the identity.= C' 'e.6) which gives (4.5) to this problem by first decomposing SH = CC'.. One can easily reduce (4.' (see also Tso (1981 )).4) This is minimized by noting that an expression like IP'(MI . H21 HI)= 2 /J into the above H2 if! HI' since -T L p In( 1-· .. > i I l• hut eigenvectors ci el. II (fSkkf3 f31/1 P'SHP .M21 =0.7) The estimates of the other parameters are found by inserting equations.B1 I s.. II f3'(Su .

j=I.(t) = Vi(t) -J L Vi(s)ds.. V.10) IJ where U(l) =l V1(t).r)-dimensional Brownian motion and the (p .8) and (4.. P . r\r+1)= ~ T In(I-Ar+I). and are not given by the usual X2 distribu- tions.] is defined as the matrix of stochastic integrals" F. (4. (i = 1. . ... sec the Appendix.11 ) (J Further f L }'F'd t is a (p .J-.(t)l is defined by F. A.13) II .. and can be tabulated by simulation.. = ap. Consider first the case a 'LJ. P .l = O.9) Under the hypothesis H~:n= aP' and p. ..r).9) are found in Johansen (1989)... . . By sujtably normalizing the equation (4.r) X (p .. . .r) matrix of random variables defined by (I the ordinary Riemann integrals (i.r).. These distributions are conveniently described by certain stochastic integrals.dVj (i. . .p-r). (4. Fp . but as multivariate versions of the Dickey-Fuller distribution.5) and letting T'" one can show that T(Ar+I.) converge to the roots of the equation 0:) IpJI FF'dt-JI F(dU')JI (dU)F' [=0. i IJ I) (4. j = 1. (4.(t)f is a (p .12) and L J L F(dU') = [1: (dU)F.r)-dimcnsional stochastic process F(t) ={FL(t)..178 UULLETIN Similarly the likelihood ratio test statistic for testing H2( r) in 112( r + 1) is given by ·-21n(Q.. (4. . ..) the same results hold but derived from Sij rather than Sij' The asymptotic distributions of the likelihood ratio test statistics (4..

8) is therefore called the trace statistic (trace). I ) "" n .9) is called the maximal eigenvalue statistic (Amc. the number of non-stationary "The definition of a stochastic integral is analogous to the definition of a Riemann integral. i. r + I.r. sup(tdIk( U(/.8) satisfies ~2In(Q.c< 00. The statistic (4.IS infinite variation but finite quadratic van allan.e. Then we consider a partion of the unit interval and the Riemann sum R=LJ(t. ~! r: f. since the trace equals the (maximal) eigenvalue.. 373. where triMf denotes the trace of the matrix M.) .~I (dU)F'[II 11 Il FF'dt]-IJI I) F(dU')j. such that E(R ··fFdU)2 converges to zero.A r+ converges weakly to . We let U and F be two continuous stochastic processes on the unit interval like the Brownian motions.10) depend only on the dimension p .=tr!JI .. This statistic is the square of the statistic ill tabulated in Fuller ( 1976) p.)' If r = l' . The statistic (4. f. there exists a random variable.1. The distribution of the trace and the maximal eigenvalue of the roots of (4. then both U and F arc one dimensional.T In(1 . which we shall call fFdU. where AmaJMj denotes the marginal eigenvalue of the matrix M.U(/k-I))' <:.I)(U(/k)-U(t.e. Similarly ~ 2 In( Q. distribution of the statistic can be expressed as and the asymptotic where O=Jl II U(u)du. T hen the test statistics are equal.INFERENCE ON COINTEGRAI'lON 179 With this notation it follows that the statistic (4. . This can be used to prove the existence of the limit of R in /'2' i. i.H21/Jl)=~T which converges weakly to L j:-r+1 P In(l~Ai)= L i=r-i-I I' ' i'p.e. I)) The function U(·) h. r I r + 1) = .

r). the trend is present under the null hypothesis.9) can be shown to be distributed as above but with F defined by: FJU)=Uj(U) Fi(u)=] (i = 1.» = 0 has broader tails.r-------. .¥. n: 1fFdU'j.. .r ... 2. Tables Al and A2 in the Appendix. j (i=p-r). Next consider the case where a'j_#. These distributions are tabulated by simulation in Tahle A3. i._------------------------- -. p . The relation between the applications of the three distributions is illustrated in Figure 2 below: H1~ T2 = trlf dUF'[fFF' duJ 1JFd U'}. Under Hi (i.15) F (t)=t-l/2. '" = ap.14) (4. (i =p . The distribution of the trace and maximal eigenvalue of the equation (4. The relation between the hypotheses HI' H 2 and Nt and the test statistics used to test them. n . (4.r+ 1).e.. Since the distribution with a '. It is instructive to consider again the case of p .e.r = 1.0. F=U-O Tf = trif dUF'[JF'F' dul F=jU'.1).10) with this choice of F is tabulated by simulation in the Appendix and given in Table AI. p . We can express the results in this case by choosing a different definition of F: F((ll=Uj(l)-f' iI Uj(s)ds. the p-value should be calculated from the latter distribution. . Note that = Tz + U(J )' U( 1)... This is the well-known result {West (1989)) that if the linear trend is present under the hypothesis of non-stationarity then the usual asymptotics hold for the likelihood ratio test.l) the asymptotic distribution of the test statistics (4. The distributions are tabulated by simulation and are given in the Appendix in Table A2.~ BULLETIN 180 components under the hypothesis.8) and (4. where the statistic reduces to which is distributed as x2(1). (i = 1. U(l)'U(') I------_Hi Fig. ct.

and weights W = SOk V are reported. In Table 3 the likelihood ratio test statistics are calculated and compared to the 95 percent quantiles of the appropriate limiting distribution. not even in the limit.14. see Theorem 4. 35.99.This hypothesis will be maintained below. VAR Econo- . Two versions of the test procedure are reported in Table 3. Thus all the tests performed in Table 3 are highly dependent on one another. If we test the hypothesis that r= 0 in HI we get a test value of 49.5 percent We conclude that there is only one cointegration vector in the Danish data. 1) we find . Thus a constant 1 was appended to the vector X/-2. The calculations have been performed metrics.Ai)l .07. First we consider the number of cointegration vectors.TI. 1) (see Figure 2) was performed: .1. 1) in H2(r:5. -2In(Q)= . f. The Empirical Analysis 5 In Table 2 the estimated eigenvalues the normalized eigenvectors V. = . The first is based on the trace and the second on the maximum eigenvalue. giving the matrices The results of the eigenvalue and eigenvector calculations are given in Table 2. Hence there is no evidence in the Danish data for more than one cointegration relation. which is found to have a p-value of appro 10 percent. for the two data sets. :: . and the calculations were performed as described in Section 1. . Note that the eigenvectors vi for the Danish data are of dimension 5.21n( Q. r = 0 I r:5.08 which is in the upper tail of the distribution of Am"x for r= 0 with a p-value of 2. 1) = 30. see Table A3.8) and (4. Using the trace test procedure gives st. is not significant.T L In{(I- i ii)l = 1.9). For the Finnish data It is assumed to contain effects both from the intercept and from the linear trend (sec discussion in Section J). The graphs corresponding to the eigenvectors and the original data are presented in Figures 3 and 4. Hi(l) 4 I H2( 1))= T In] IS60 10t)J(I- An/ ISOli I( 1 . The Danish Data On the basis of the plots of the series (see Figure 3) a model without a linear trend in the non-stationary part of the process was assumed. It must be noted that since the TAi are ordered they cannot be independent. If instead we apply the maximum eigenvalue test.21n( Q.INFERENCE ON CO INTEGRATION 181 4. in the computer package RATS. where the last coefficient is the estimated intercept. (4. as a check that the maintained assumption about the absence of trend is data consistent.' i=2 . i. and test H 2( r = 0) in H2( r:5.06._21n(I-Aj)= 19.: . Inc/Doan Associates . in the asymptotic distribution. Finally. beginning with the hypothesis r$ 1 versus the general alternative H. the test for Hi(r:5. The 95 percent quantile.

.--182 BULLETIN o o ::. <') o I c :xi cc o ::.~.3 r') I N N . o I o c~ o 0'1 o I o C ~ o c ~ -< c ..f I N o o o N .. C ~O 00 00 o ~ ae- o ~o 00 '''': r.f I C .~-------~---------~--~-------------------..

(5) 2.35 2.1.95) Amax Amzt'( (0.42 3.28 27.08 14. which indicate that at least 2 but possibly 3 cointegration vectors are present. and the hypothesis that some subset of the adjustment coefficients is zero will be formally tested in Section VI. and it is natural to give them an economic meaning in terms of the average speed of adjustment towards the estimated equilibrium state.06 r=O 49. The Finnish Data As discussed earlier a model that allows for linear trends is fitted to the Finnish data. In the first equation.(17 53.69 r<:.213.115. vectors and weights are given in Table 2 and the test statistics in Table 3. which usually would be considered too high. But since the power of .14 ----. The estimated eigenvalues. (5) H)_ r~3 trace A max A nl<!. In particular the last two adjustment coefficients are low.65 76. 1 r=O trace 3. Similarly ci is found as the first column in the matrix =(- W* = S02 V*: 0.029).75 21. since I SIlIlI nt( 1 .) and versus the general alternative HI (trace) for the Danish and Finnish data. 1 19.09 20.{ trace Ji2 r::::3 r::::2 r s.INFERENCE ON COINTEGRATION TABLE 3 183 Test statistics tor the hypothesis fij and f-Ltor various values of r versus r+ I (Am".34 10.11 11.A.17 8.14 (0. . .5.34 relation holds. The acceptance of the third relies on a p-value of approximately 20 percent.0.60 21.09 15. 0.21 it> + 4. The interpretation of the cointegration vector as an error correction mechanism measuring the excess demand for money IS straightforward. with the estimate of the equilibrium relation given by This m2 ci' = 1.64 38.22 d + 6. (U)23. whereas in the remaining three equations the adjustment coefficients are lower though of the 'correct' sign.89 28. which measures the changes in money balances. The coefficient estimates of the cointegrating relation are found in Table 2 as the first column in V*.213.17 35.03 Y . such that a low coefficient indicates slow adjustment and a high coefficient indicates rapid adjustment.26 48.t) = I SIlIISlIkSkklSUJ The asymptotic distribution of the test statistic is X2(3) and thus not significant. the average speed of adjustment is approximately 0.35 6.49 8.08 9.37 30.35 8..) = I S~III nT( 1 . The 95% quantiles are taken from Table A 2 (H 2) and A 3 (fi~) ---------~--The Finnish data the Danish data trace (0.95) (0.01 37.--~ -«: 9. The coefficients of ci can be interpreted as the weights with which excess demand for money enters the four equations of our system..0R 17$4 31.11 7.90 26.06.

40 1976 1978 1980 (a) 1982 1984 1986 1988 REAL INCOME IN DENMARK .184 BULLETIN REAL MONEY STOCK IN DENMARK 7.50r--------------------------------------------------. 7.

.130 1976 1978 1980 (c] 1982 1984 DEPOSIT RATE IN DENMARK 0. Continued . 0.130 .-----------------------------. 1976 1978 1980 (d) 1982 1984 19B6 1988 Fig.170 0.INFERENCE ON COINTEGRATION 185 BOND RATE IN DENMARK 0..----------------------------. 3.230.

. Continued .050 0.050 0. .000 -0. 186 BULLETIN COINTEGRATION RELATION 1 .•..150 1976 1978 1980 (e) 1982 1984 1986 1988 COINTEGRATION RELATION 2 0.•..100 -0..100 -0.100 0.. 0..050 -0..050 .100 0..150 1976 1978 1980 (f) 1982 1984 1986 19B8 Fig.0. -0.000 ~ ~ ...150 0.150 0... 3.

. 3.200 1982 (9) 1984 1986 1988 COINTEGRATION RELATION 4 1.----------------------------.200 0.INFERENCE ON COINTEGRATION 187 COINTEGRATION RELATION 3 0..8a 1976 1978 1980 1982 (h) 1984 1986 1988 Fig..000 -0.100 -0. Continued .100 0. .00~-----------------------------.500. 0.

80 3.4 and the original Finnish data.00r------------------------------------------------------. 4.20 4.XI for i= 1. The sample is: 19S5.40 3.00 1966 1970 (al 1974 1978 1982 1986 REAL INCOME IN FINLAND 4.60 3. The graphs of the cointcgration relation V. 3.60 4.3.188 BULLETIN REAL MONEY STOCK IN FINLAND 4.20 3.1 ~19R4. .00 1966 1970 {b) 1974 1978 1982 1986 Fig.

260 0.160 0.INFERENCE ON COINTEGRATION 189 MARGINAL INTEREST RATE IN FINLAND 0.110 IV 1962 (e) INFLATION RATE IN FINLAND 1974 1978 1982 1986 0.0200 rN " r-r-' L..310 r--~--------------------------' 0.0300 f- 0.210 0..0500 0.0600 0.J ~ IV ~ N \ 1\ 0.. Continued . 4.0100 1962 1966 1970 (d) ~ If 1974 1978 1982 1986 Fig.0400 ~ 0.

4.0.50 -3. .00 1962 1966 1970 (e) 1974 1978 1982 1986 .-------------.070 -0. -2..00. COJNTEGRATION RELATION 2 .190 BULLETIN COINTEGRATION RELATION 1 -1.060 . 1962 1966 1970 (f) 1974 1978 1982 1986 Fig..00 -2.-------------------------------------.0.-------------------------------------------------. Continued .080III .

Continued .5~ r------------------------------------------------------. -0.80 i: .INFERENCE ON COINTEGRATION 191 CO INTEGRATION RELATION 3 -.0 1962 1966 1970 (h) 1974 1978 1982 1986 Fig. ~.. 4.90 -0. r f -0.95 ~. 1962 1966 1970 (9) 1974 1978 1982 1986 COINTEGRATION RELATION 4 10.

0. The sign to be expected for 'excess demand for money' should be negative. f3il = . but au dominates II ' so that the 'excess demand for money' enters with a negative sign in the first equation. q Since the asymptotic distribution of this statistic is X 1( 1). see Section V. (0. =.y.0). For the case r > 1.1. i'" and 8p are stationary.A. )/( 1. ". which is most correlated with the stationary part of the model. . whereas PI and P3 seem to contain information about m 1 . they should do so in all cointegration vectors.192 BULLETIN the tests are likely to be low for cointegration vectors with roots close to but outside the unit circle. . The estimates reported in Table 2 indicate that {J2 is approximately measuring the inflation rate. one would expect that the linear combination. i. the interpretation of and a is not straightforward. A heuristic interpretation is however possible by considering the estimates in Table 2.e. it is significant. .. Note that Pi. and (0. it seems reasonable in certain cases to follow a test procedure which rejects for higher p-valucs than the usual 5 percent.1.'J \:: .f3i2 seems consistent with the data for the 3 eigenvectors. However. -c-. .0. a linear combination between these three vectors might be more stable (in terms of the roots of the characteristic polynomial) than the individual vectors themselves and this linear combination could in fact be the economically interesting relation.0.1). i.>~ J.0.j)}= 4.0). Note also that a II and alJ have opposite sign.I' i =). implying that m I .Pi. namely the first eigenvector.. Next we test the hypothesis that the linear trend is absent. i(~(I-iil) tnl( 1 .1). the P a It seems reasonable to denote the first coordinate of the cointcgration vector In ordinary matrix notation we then have f~il = {3. by {Ji]' . One reason why we have kept r= 3 in this case is that the hypothesis of proportionality between money and income. Hj_(3) in H1(3).. ff~(3)IH1(3))= T . i' j Pi' say... is of special interest.1. The value of a'2 can be interpreted as the weight with which the inflation rate enters equation i.T In!IS~III.0. and . In particular..?.that (J2 is approximately proportional to (0. 3.. We find {J as the first three columns of V from Table 2 and a as the corresponding columns of the weights W.0. In Table 7 Section VI. and the hypothesis H 2( 3) is maintained. PI' fJ2 and P3 can be approximately represented as linear combinations of the vectors (.~ (l-in/ISool . Note that given the fuJI matrices V and Wane can estimate a and fJ for any value of r. We found that -2In(Q. -c-.y. This means that the only interesting cointegration relation found is between m 1 and y. -c-. the ordering of the eigenvectors provided by the estimation procedure is likely to be useful. Thus.2"" .e. Although there is some arbitrariness in the case r > 1. -c-.' If In 1 and y appear in a cointegration vector with equal coefficients of opposite sign.78.

i 3.H'SkOSI)(/SOkHI=0. (4. (5.1). and rp has to be chosen to minimize (5.. a and A is given as for fixed {J = Hrp.1. the estimator without subscript will be the estimator under H2 or H!.2) I AH'SkkH for . '''..1: Under the hypothesis HJ.. Thus it will in general be possible to find some relation which has.• . from r (4.1 ) over the set of all s x r matrices tp. and Ht in H2.1 below. The hypothesis specifies the same restriction on all the cointegration vectors. equal coefficients with opposite sign to m and y. Throughout. the estimate of the combined effects of all three cointegration vectors. is it meaningful to say that we have found a unit elasticity. and only if the proportionality restriction is present in all f3 vectors. I'k _ I' CfJ. I n {~1 (1 - i d.r) and and find the estimates of a.3) Choose tjJ = (V3.f3=Hrp.1. but that means that the estimation of I'1 ' . ESTIMATION AND TESTING UNDER LINEAR RESTRICTIONS ON fJ Model H:> : = H rp is a formulation of a linear restriction on the cointcgration vectors. Likelihood Ratio Tests Under fl3 we have the restriction f3 = Hrp where H is (p x s). V3. P 5. (5. we find the maximum likelihood estimator of {J as follows: First solve (5. corresponding to a long-run unit elasticity. enter then any linear combination of these relations will also be a cointegrating relation. This completes the investigation of the model He and H! in HI and we turn now to the models H. IJ. A subscript indicates which hypothesis we are currently working with..e. say. i.···. THEOREM 5. This is clearly not interesting. A and maximized likelihood becomes P3 = HcjJ. The L L~}jT in 3) = I s.2). is reported. and V3 = (V3.5). This problem has the same kind of solution as above and we formulate the results in Theorem 5. The reason for this is the following: If we have two cointegration vectors in which m and y. and (3. V.1 v 3\·) normalized by V~(H'SkkH) V3 = 1. > i3..4) . say...1 > .INFERENCE ON COINTEGRATION 193 estimate of n = ap'. It is striking how well the proportionality hypothesis between money and income is maintained in all equations of the system.

1'1. The Empirical The Finnish Data Analysis ". The Danish Data In the Danish data we found r= 1./ j .. both of which are economically meaningful: H*·(1 3.. Thus the conclusion about the Finnish data is that the last two variables t= and !!./ . the same results hold... The corresponding restricted p-estimates hardly change at all compared to the unrestricted estimates of Table 2 and they are therefore not reported here..5) degrees of freedom.: .". i. y and m 1."J ".p arc already stationary. and the first two.21n( Q)= 0. _. Based on the unrestricted estimates in the previous section it seems natural to formulate two linear hypotheses in this case.l =_R..y.. .1' f'l.194 BULLETIN which gives the likelihood ratio test of the hypothesis H3 in H. are cointegrated. Thus the hypothesis of equal coefficients with opposite sign for In 1 and y.. = af3:). In this space we can choose to present the results in any basis we want and it seems natural to consider the three variables In 1 .. as ~ 2In( Q. .. )~ -:-:..:-~ .• J ".29 = 3.:f3 = HIP and p.e. . _. . Under the hypothesis Ii"."_' We consider the hypothesis that there is proportionality between money and income. Thus the hypothesis H 3 is really the hypothesis of a complete specification of sp({3).•• J .~ f3 = ( I o 0 1 where ffJ is a 3 x 3 matrix.82 which is compared to xk95(r(p-s))=X2(3(4-3))=7...2' 1 if: Ir :i: "..3) In matrix notation the hypothesis can be formulated as: ~ 1 0 0) 100 0 1 0 tp.."J 5. Solving (5. i'" and tsp. is clearly accepted.. so that the coefficients of money and income are equal with opposite sign.1..'J .. (i= 1.." ::t: I I . These are compared to the eigenvalues of the unrestricted model H2• The test statistic is calculated as .2) gives the eigenvalues in Table 4..1.2.51 + 0. H1ili2)= T L 1=1 In{(l ~ id/(1 ~ Ai))' T he asymptotic distribution of this statistic is shown in Johansen (1989) to be X2 with r(p .. _. . With the imposed proportionality restriction we now have three cointegration vectors restricted to a three dimensional space defined by the restriction that m 1 and y have equal coefficients with opposite sign. . ./ :... "...02 + 3..81.

88.5. o 0] 000 001 I () o 1 where cp is a 4 x 1 vector.S2) = 1(4.-. . and we can thus accept the hypothesis that for the Danish data the coefficients of m2 and yare equal with opposite sign.3) = 1 degree of freedom.-.095. The test for the hypothesis is given by .iI..032).o~~~!o cp. Solving (5.) -In( 1- i ill = 0.023.21n( Q) = 53(ln(1- Jet.0.4' In matrix formulation the first hypothesis is expressed as /3= [ . .~r""""":~~"''~' ' and ---- .88. Since Ilt.2) we get the eigenvalues reported in Table 4.88 .05 . r -~ o o where cp is a 3 x 1 vector. we will test Nt2 within Hj'.177.0. The second hypothesis that the coefficients for the bond rate and the deposit rate are equal with opposite sign is now tested. This will now be formulated in matrix notation as 1 .3) = 1. These are compared to the eigenvalues of the unrestricted Hi modeL The test of Ht. -.00.5. The test statistic is clearly not significant.0.21 ). . {3* = -.2 In( Q. HI. The asymptotic distribution of this quantity is given by the X2 distribution with degrees of freedom rip .0.. It is not significant and we conclude the analysis of the cointegration vectors for the Danish demand for money by the restricted estimate /3* = (1. The corresponding estimate of a is given by d * = ( ..s) = 1(4 .6. . This hypothesis implies that the cost of holding money can be measured as the difference between the bond yield and the yield from holding money in bank deposits. in Hf consists of comparing Aj.-. . INFERENCE ON COINTEGRATION 195 HI2: P'J = - f31..00.1. Solving (5. which should be compared with the X 2 quantiles with r( s.2)-ln( 1- AL)l = 0. and Af by the test . . was strongly supported by the data. IHi) = T{ln{1 .2) gives the eigenvalues in Table 4. ..

433 0_172 0. --.1_ 1) (.-.l..-.073 0. the quantity w=T 1!2K*'Pf 1[(£ f .178 0.i. .070 7.2..433 0.. 0.---------.e.49 38..1.. The above test statistics require the normalization of fJ and v as in (4. and since r s: s s: p = 4 and.. then the Wald statistic is given by w2 = T0 tr!(K'/J(O-1 .1 = 3.-.04 29.: 0.044 0. the contrast involves only the coefficients of the variables. 0)'.60 3.5). £2' £3) (see ::~ the Finnish data in Table 5).. .199 Eigenvalues Ar 0.. . The asymptotic distribution of this statistic is X2C1i with r(p . corresponding to a completely specified f3. An alternative expression for this statistic which can be applied for any normalization is d ...) 26.s) degrees of freedom. This statistic is easily calculated from Tahle 5.309 0.89 7.34 2.T In(1-in 10. as in the Finnish data..-----------"""!'IJIII 196 BULLETIN TABLE 4 The Eigenvalues and the Corresponding [est Staustics for Testing Restrictions on f3 Eigenvalues A. we can .. .32 2.L (K*IVfJ2)) 1/2 is asymptotically Gaussian with mean 0 and variance 1.-... . such that K*'/J* = K'/J... -------. In this case r = p .36 0. (see the Danish data in Table 5) then.045 The Finnish data -jln(l-A..35 0. . . since s = p is no restriction.16 2.44 112: H.113 0. i.. The idea is to express the restrictions on P asJ K'P = 0 and then normalize Kip by its 'standard deviation'.It I P'K) (K'vv'Kt I L «·1 where v is the eigenvector corresponding to £4' and D =diag(£l. only test a hypothesis with s = r =p -1 = 3.13 The Danish data . not the constant term. where K is p x (p .-~----. ~~ _ ..64 0..423 0..3{) 0_043 30Jl9 0 () . the Wald Test Instead of the likelihood ratio tests which require estimation under the model}: :::::: H 2 and H 3' one can directly apply the results of model H 2 given in Table 2 to::~: calculate some Wald tests.()O() io.32 0 5. in case r= 1...s). -. .11 In: II~I : 1I~2: 0..006 6.309 0. Hence K*=(K'.or 30. If more than one cointegration vector is present.030 38.--.." It is shown in Johansen (1989) that if v* denotes the eigenvectors corresponding to .226 0.----.~..47 23...

24)~ +--..4 is tested in a similar way.• ': "-r ":": ~ " :": --~--"_---'--"""-"'-. H*).24. In Table 5 the eigenvalues and eigenvectors for this normalization are reported. The Empirical Analysis Since the calculations are numerically simpler for the normalization V'Sa" = I.4332 . First we find from Table 5 that K'vv'K = (1.0.. and I (K*'vff i=2 =( 14.96 and K 'p"(A_1 )-lp"' (-2.•. The Danish Data We start by the hypothesis H*3.2' (i= 1.22)2 = 12.2-2-60--~1 -1 (-11.31/( 1/0.1·"'-U .66 .83/12.13 + 10.•."'--" •••• •• •.R "'-1.66.O.:Pil =- Pi.0)P=0.64)2 +(1.1.•.20.+ '--O-. This can be formulated as K'P = (1. Note however.2 is now tested within H! and not within Hjl' The test statistics becomes 1.309r 1_ 1 _. Tl/2K'/J= 531/2( . Both these statistics are asymptotically normalized Gaussian and the values found are hence not significant..1) x 360.31._~_-_ -~~=0.96 = 5.R =.86)2 (4. This just .25.83.2.1.58-6. INFERENCE ON COINTEGRATION 197 5..02-1.3).~ . The Finnish Data For the Finnish data we only test the hypothesis: H. Then the test statistic becomes w = 5.21.2.-1.05)2 + (7.0731 1_1 The test statistic becomes w2 = 104 x l'. Notice that the Wald test in all cases gives a value of the test statistic which is larger than the value for the likelihood ratio test statistic.97 + 22.2. it will be used to illustrate the Wald tests.0) P= 0..R The second hypothesis --f3 - 1.38 + 2. which is not significant in the X 2 distribution with 3 degrees of freedom.93)2 + (11. The Wald statistic is then calculated as follows: First.3 .13.13)1/2 = 0.70) = 5.32. 0.36 -7.95 .06)2 D -J K = 0.20)2 = 360.93+2. that l1r.1.2 expressed as K'P=(1.

I r- 0 N N N ": -c:t -.. N ~ 0- N E .j. I"'l r0.-..j- I q o N Vi \0 N I :5 _. _.198 BULLETIN N N N 00 ('j o N r'l N 0\ I \0 00 N 00 Vi N o o 00 N N 0- I o v) I o N tx: 00 -c:t -.

4) {6.· In the following.I(B'R{]')/2)' (6. Then the hypothesis H4 can be expressed as B a = O.ap'Rk1} = A'RoI -.5} and the maximized likelihood function from the marginal distribution L'. Sak. = A'SOk - A'SooB(B'SooBt The factor corresponding to the marginal distribution of B'Ro.. etc.b1Sbk IB'SOk.2) B'(Ro! .) BiRo.1 ) (6. (6. ESTIMATION AND TESTING UNDER RESTRICTIONS Let us now turn to the hypothesis H4 where a is restricted by a = A 1JI in the model Hz. = (6. we define: Aua = A'AA.INFERENCE ON COINTEGRATION 199 emphasizes the fact that we are relying on asymptotic results and a careful study of the small sample properties is needed.. Here A is a (p x m) matrix.-A'AtpP'Rk' -AahAj. and Rk' conditional on B'Rr. It is convenient to introduce 8(px(p-m))=Al' such that B'A=O. The concentrated likelihood function (3.I is found by regression for fixed tp and .h are variation independent and hence that the estimate of AahAI~I. Aub =A'AB.h1B'RoJ (6.A'A tpP'Rkr. ON a VI.bl-Tf2exr!- r*i {A'Ro.ap'Rk. is given by IBIBITI2IAhhl"TI2exp!- r~i (BI~)YAh/.b = SUk -SahSj.8) can be expressed in the variables given by I A'(Ro.! and is given by IAIAITI2IAaa.3) and gives the estimate Abb = Sbh = B'SooB.' Al/bAh"i and AUI/.6) It is a well-known result from the theory of the multivariate normal distribution that the parameters Ali. we factor out that part of the likelihood function which depends on B'~)" since it does not contain the parameters tp and p.a2 F= ISbhI/IB'BI· The other factor corresponds to the conditional distribution of A'Ro. To save notation.

...12). > A4. . I . (6. The same result holds for testing H!: a = A tp in Hi. (6. . and estimated from (3.b ..p) normalized (6.AIAt}t'P'A'A = S"u.. ..A' a4d~A.14) i=1 The asymptotic distribution of this test statistic is given by a X 2 distribution with ri p ..1 > . I ASKb - Sk(J.8) which means that the estimation of fJ follows as before.8) .7) and (6.hSu~. P is found as follows: First solve the (6. = A4p . the maximum likelihood estimator of equation giving A4.SahSi.1 ol.. . such that V~SkU = I.m > A4. =0 and V4 =(v4.- SkbShb1 B'Ro/· In terms of R(I/ and Rkr the concentrated likelihood function has the form (3.200 BULLETIN P giving A "bAh.4).AJAtpfJ'Skb) Sbb1.9) which gives the estimates (6.R.h = Sad and the maximized likelihood function L'.lhS'lkbl= 0.m) degrees of freedom. = TIL 1= Ri... " It is convenient to calculate the relevant product moment matrices as J Si/J.. fJ)=(Sub and new residuals defined by R'lI = A'Ro.1 v4.)' i~1 i~l (6.13) The estimate of A can be found from (6..."...11) (6. B'R .)=ISool n (1-'£4. Now take . The likelihood ratio test statisic of H4 in Hz is r is (6.1 ( tp.10) and a4 = At}t= A(A'A t IA'(SOk -SooB(B'S(I(lB)-IB'SOk) P4..5)..7) Rk/ = Rk.1: Under the hypothesis H4:a=Atp." THEOREM 6. (6.12) Au"..... "4 +1 = .£4..a~r(H4)=IB'BI-IIA'AI'IISbbIISa"bl n (l-. see Johansen (1989).

then ordinary least squares analysis of the first equation will give the maximum likelihood estimation for the cointegration vector. The Empirical Analysis In Section V it was shown that the hypothesis about proportionality between money and income. namely by solving the eigenvalue problem (6. D and the constant. {3i.15) This gives the final solution to the estimation problem of Hi.e. in particular. then only one eigenvalue is non-zero. a is proportional to (1. = H4 n H3. If.8). . and that the hypothesis {3I. and that a linear restriction on a implies a conditioning. then we can correct for these before solving the eigenvalue problem.1. Finally. It is now clear how one should solve the model H.8) that since the matrix SkahS. Notice how (6. In this case we note that PIRkt=tp'H'Rkr which leads to solving (5. .PIA was accepted for the Danish data.2. It is seen from (6. and the corresponding eigenvector is proportional to SZk1. was accepted both for the Danish and the Finnish data. note that a linear restriction on fJ implies a transformation of the process.. 8'a = 0.0). i. where we have conditioned on B'Rol' In other words if we assume that the equations for B'Ror do not contain the parameter a. Thus all the calculations can easily be performed starting with the product moment matrices Sij and applying the usual operations of finding marginaJ (transformed) and conditional variances followed by an eigenvalue routine.INFERENCE ON COINTEGRATION 201 The following very simple Corollary is useful for explaining the role of singleequation analysis: COROLLARY 6. .h' which is exactly the regression coefficient of Rkt obtained by regressing A'Rol on B'Ror and Rkt• This can of course be seen directly from (6. and the Ht .6) since A'A tp is 1 x 1 and can be absorbed into fJ.{3i. 6..u1hSakb is singular and in fact of rank 1. however. which shows that P is given by the regression as described. chosen to present the analysis of restrictions on p and a separately in order to simplify the notation.hSko. Finally we just state briefly how one solves the estimation and testing of the model Hs:fJ=Htp and a=Atp. . JJ PROOF: It suffices to notice that when m = r= 1.3 = . We have. Thus it seems natural to move directly to the H. L\X H1. Since a = A 1/J we solve (6. where restrictions have been imposed on fJ as well as on a.2: If m = r = 1 then the maximum likelihood estimate of Pis found as the coefficients of Xt-k in the regression of A'AXt on Xt-k.15) contains the previous problems by choosing either H = I or A = lor both.. B'AXp and AX'_I .0.2) where Rkr has been replaced by H'Rkr• Thus restricting P to lie in sp(H) implies that the levels of the process should be transformed by H '. then only one cointegration vector has to estimated.1 = . An empirical illustration of this will be presented below..

4. normalized at A 1(0)= 1. The single equation estimation corresponds to the calculation of the static long-run solution of general autoregressive model Al(L) m. and A..6..23.698). We conclude the empirical analysis by a comparison of the estimated "matrices under the f<I1I unrestricted HI-model and the final version" = apl with data consistent restrictions on a and P (see Table 7). ..57.l)=( -0.4.25.~5(3). .67 = 30. + A3(L) i.16) evaluated at L = 1 gives: (A I( I)..254.b' S~(/. as .57i<l + 6.06) i.357 and consequently one eigenvector fl.A3(1). For the Danish data . as well as one estimate a. 0). Solving the eigenvalue problem (6. " " The test statistic for this hypothesis about a is then given by .0. see Corollary 6.211. .0.2. 0.1.83) (1.A4(1).16) where e.' + as +A~D.. .81 = X6.204 BULLETIN The appropriate A and B matrices are now on the basis of which the matrices stu. .19) (0.+ e. as! = (0.00.4. . Although the test statistic is not significant at the 5 percent level it would be so at a slightly higher level.76.09 < 7.2 In( Q'. are inderv rdent Gaussian variables with mean zero and variance a.A 4. =A2(L) y.8=AI(1)-I[ -AIO).e. Normalized by the coefficient of m 2 the estimates are . i= 1. 0. A2(1). a J and Q4 to zero. A4(1).8) gives one eigenvalue 0. lI*4.76il> + 2. from which the static long-run solution can be calculated as m = 0. a. .b and S~kh can be calculated. 1. The static long-run solution is obtained by evaluating (6. On the basis of this we conclude that there is no strong support for restricting a2. Sku.16) at L= 1.A *)j * I .244. (0.8(HJI) = (1.96.58. is a lag polynomial of order 2.l a=AI(l)· The OLS estimation of (6.46) (2.1 I lIt)_ = T{ln(I .58) a(H!. exactly the same estimates as in the restricted maximum likelihood procedure.96y.' + A4(L) i.654.2..(L). which gives the estimate of P and a normalized by the coefficient of m.0.42 =7. .1 ) -In( 1 . (6.

INFERENCE ON CO INTEGRATION 205 .

The hypothesis of the existence of cointegration vectors is formulated as the hypothesis of reduced rank of the long-run impact matrix. Inference concerning linear restrictions on the cointegration vectors and their weights can be performed using the usual X 2 methods. VII. The order of cointegration was one for the Danish version. The test procedures are in general likelihood ratio tests. which simplified the interpretation of the cointegration vectors as a long-run relation in the levels of the process.constant term and seasonal dummies. The proposed methods are illustrated by money demand data from the Danish and the Finnish economy. The applications were chosen to illustrate various aspects of the cointegration method. For the Finnish data there were three cointegration vectors which served to illustrate the interpretational problems when there are several cointegration vectors in the data. . whereas the Finnish model allows for linear trends in the non-stationary part of the model. and these are tabulated by simulation. and it is demonstrated that the estimation procedure as well as the distribution of the test statistics of the reduced rank model is strongly affected by the assumption of how the constant term is related to the stationary and the nonstationary part of the model. The solution we propose is to start with a relatively simple model specifying a vector valued autoregressive process (VAR) including a . The model for the Danish demand for money is specified without assuming a linear trend in the data. as well as estimates and tests for linear hypotheses about the cointegration vectors and their weights. and with independent Gaussian errors. The asymptotic inferences concerning the number of cointegrating vectors involve non-standard distributions. The role of the constant term is closely related to the question of whether there are linear trends or not in the levels of the data.I····-~-------··-----·------·-··-----·--------···--··- 206 BULLETIN the number of parameters (excluding the constant) has been reduced from 16 to 4. see Johansen (1989). In this way we can derive estimates and test statistics for the hypothesis of a given number of cointegration vectors. It is shown that the inclusion of the constant term in the general VARmodel has significant effects on the statistical properties of the described tests for the reduced rank model. whereas for the Finnish data the reduction is from 16 parameters to 6. Institute of Mathematical Statistics and Institute University a/Copenhagen Date of Receipt of Final Manuscript: December 1489 0/ Economics. SUMMARY In this paper we have addressed the estimation and testing problem of longrun relations in economic modelling. but in the case of linear restrictions on f3 a Wald test procedure is suggested as an alternative to the likelihood ratio test procedure. This is given a simple parametric form which allows the application of the method of maximum likelihood and likelihood ratio tests.

d.11) the stochastic matrix fFF'dt and [F dU' are approximated by T-2 and T-1 I J~! r (X'_I-X~I)(X'_I-X_IY I r~l 1 (X.r. Gaussian variables (t= 1.000 simulations the quantiles are found as the appropriate order statistics.. i = 1.p -r) and calculate X.-I I T X.. . .i.INFERENCE APPENDIX.r) array of i.0.14) and (4._I-X_l}E.= 0.T /2._ I . .16) and (4.. . i = 1. i = 1. .-At)} is approximated by a random walk with T= 400 steps.17) then X _ I is dropped and XI _ I is extended by an extra component 1. The (p ... . from X'i= . On the basis of 6.r ). SIMULATION ON COINTEGRATION OF THE LIMITING DISTRIBUTIONS 207 The limit distributions are expressed as functions of the stochastic matrix see Section IV. In case the process F is given by U . . . and if F is given by (4. where X_ I= T- I . U".15). -1' From these expressions we calculate From this matrix the trace and the maximum eigenvalue are calculated.. . with Xo. respectively..\~l I f'i? ( t = 1. we replace in the above calculation the last component of X. p ..X _ I by t . If instead F is given by (4..r)-dimensional Brownian motion U(t) =( VI(t). . see (4. T.... Thus we generate a Tx(p . 1.. ..p .

810 23.192 14.332 17. 2.024 12.030 7.551 77.691 12.576 18.977 Simulations arc performed by replacing (he Brownian motion by a Gaussian random walk with 400 steps and the process is simulated 6.700 9.962 14J)36 20.518 24.616 38.658 19.951 18.778 27.691 15. 20.164 23.521 31.184 52.249 82.279 27.181 68.436 45.030 13.338 26.330 2.024 18.154 32.140 6. 52.038 3.373 61.759 4.215 90'%.106 3.599 35.002 29.956 8.000 7. 0.816 12.809 35. 9.178 Trace 5. 17.335 35. 53. 23.132 18.5'%.566 -------~~- 8.017 34. 2.103 55.844 31.782 26.953 6.917 30.197 29.419 69.699 10.342 34.816 13. 95% 97.030 8.291 55. dim 50% ---~-1.576 21.911 3.962 15. 23.774 2.083 17.936 19.568 18.852 3.672 5.699 11. 95% 97.031 11.000 times.521 22.065 32.000 times. 7.905 Simulations are performed replacing the Brownian motion by a Gaussian random walk with 400 steps and the process is stimulated 0.415 2. 18.707 4.313 50.341 33.801 73..783 18. 17.818 6. 33.955 1.397 53.455 12. 34.208 BULLETIN TABLE Al Distribution of the Maximal Eigenvalue and Trace of the Stochastic Matrix J(dU) F'lfFF'dut IJF(dU') where U is an m-dimensional Brownian motion and F is U component is replaced by t-1/2.474 3.275 23.099 18.611 34.475 53.030 8.998 2.125 16.905 10.062 51.299 32.415 4.248 65.278 18.092 1.403 23.905 2.583 28.936 17.262 Trace 9.638 3.936 25. 5 'X. 6.nl 43.623 5.546 5.030 9.132 SO'X. mean Maximal eigenvalue 4.192 12. 9O'X.163 29. 12.332 15.362 29.159 56.959 24.310 35.1 mean that the last dim 50'%. see Theorem D.837 28.899 1. .880 84.447 2.680 3.658 1.256 48.875 5. sr« 1. 7.211 1. TABLE A2 Distribtuion of the Maximal Eigenvalue and Trace of the Stochastic Matrix J(dlJ) F'[fFF'duJ-'JF(dU') where U is an m-dimensional Brownian motion and F = U 99% D va.447 2.113 27.879 20.451 23.341 27.949 7.341 6.658 16.964 65J)63 3. 99% var Maximal eigenvalue I.719 5.509 47.250 60.424 72.588 2.962 37.712 30.858 11.169 33.324 22.943 38. except 4.792 76.666 16.250 19.335 13.868 40. 12.188 25.445 4. 0.083 14.549 23.873 41.595 21.381 4.607 24.

868 39. 'A Canonical Analysis of Multiple Time Series with Applications'.337 3. in Karlin.741 24.397 Trace 10..607 28. mean var 95% 97. Vol. L.172 24.094 20.877 15.969 4.381 1 23. Wiley.622 23. C. 55.068 12.431 58.---'-'-".. C.954 6. pp.938 28. Time Series and Multivariate Statistics. J. S. (1983)..449 78. Mimeo.-"'-~.752 21. W.068 8. University of Wisconsin. 3. (1981). ----"-'- . (1987). INFERENCE ON COINTEGRATION TABLEA3 209 Distribution a/the Maximal Eigenvalue and Trace of the Stochastic Matrix J(dU) F'lJFF'duJ. Academic Press.021 18. and Weiss. (1987).474 2. SO'X. Granger. and Goodman. ~=" .698 23. H.611 31.IJF(dU') where U is an m-dimensional Brownian motion and F is an (m + 1i-dimensional process equal to U extended by I. T.925 71. 11.628 17. (1976). and Reinsel.635 66.(194 15. 'Cointegration: A Survey of Recent Developments'. 64. Biometrika. C..). J.121 39. G.. W. Amemiya.~~- . in Studies in Econometrics.328 10. J.050 19. dim 50'). J.857 12. Anderson.672 4.093 49.563 13.017 23. W. 38.262 36. G.709 22.. (1988). Oxford University. W.796 25. pp.347 75.494 4. 16. Econometrica. (1981).. (1986).192 37. 23. T.359 28.H77 11. A.. . 'Some Properties of Time Series Data and their Use in Econometric Model Specification'.068 53.. 'Co-integration and Error Correction: Representation. Institute of Economics and Statistics. journal of Econometrics. R. A. YO'::.529 59.738 13.. 'Time Series Analysis of Error Correction Models'... 'Co integration in Linear Dynamic Systems'. P.. Discussion paper.054 82. A. Dolado.768 45. Davidson.000 times.. pp.817 >: ~~ t~ : 5..894 28. S. Fuller. G. 3. LSE.361 5. J. and Jenkinson.167 34. K. W.854 89.472 9.. 121-30.643 7. A. -"'-.625 12. 355-65. (1984).. see Theorem 4. 13. F. Wiley.988 40.198 60.954 I. 251-76... REFERENCES Ahn. New York...5% 99% Maximal eigenvalue I.. 18. and Tiao. Introduction to Statistical Time Series. Engle.603 56. C.255-78.957 32. E.917 14.'-. C.1 ------_.738 19.474 22..844 5.592 9. 58.624 7. New York.409 33.. Estimation and Testing'. and Granger.781 19. J.592 5.243 4.474 2.202 37.072 Simulations are performed by replacing the Brownian motion by a Gaussian random walk with 4()Osteps and the process is simulated 6. 'Estimation for Partially Non-stationary Multi- variate Autoregressive Models'. T.168 35. New York. Granger. (cds.563 17.834 26. Box. Vol. Vol.709 17. pp. . An Introduction to Multivariate Statistical Analysis.836 30.433 6.740 19.

E.. (1988). 836-43. (1986). 8.-S. E. Vol. P. . pp. Econometric Theory. Contemporary Mathematics. P. G. Phillips. R.. 53. 'Inference in Linear Time Series Models with some Unit Routs'. 3. H. C. C. A. Vol. 'Reduced Rank Modelsfor Multiple Time Series'. 'Reduced-rank Regression and Canonical Analysis'. S. N. pp. (1981). Vol. . D. 468-97. C. 'Testing for Common Trends'. Part 2'. Johansen. C. Technical Report The University of Michigan. 'Testing for Cointegration'. y. (1987). 1. (1986a). Stock. 'A Note on Non-stationary and Canonical Analysis of Multiple Time Series Models'.-. 83. P. M. Stock. pp. Johansen. Johansen.. 852Phillips. 183~89. 4. Vol. 255-59. and Reinsel. . D. 'Statistical Analysis of Cointegration Vectors'.-~. B. G. (1988b). B. 1. (1985). P. '"--. The Mathematical Structure of Error Correction Models'.-. W. (1989). (1980). 95-131. (1987). 'Statistical Inference in Regressions with Integrated Processes. 'Hypothesis Testing for Cointcgration VectorsWith an Application to the Demand for Money in Denmark and Finland'. (1989)_ 'Asymptotic Normality when Regressors Have a Unit Root'. Phillips. Phillips. pp. Journal of the American Statistical Association. Biometrika. Pena. Oxford. B. 43. S. (1986). B. 55. 80. 'Some Aspects of Asymptotic Theory with Applications to Time Series Models'. (1987)_ 'Identifying a Simplifying Structure in Time Series'. Laidler.. Y. 'Asymptotic Properties of Least Squares Estimates of Co integration Vectors'. Johansen. Phillips. 1097 ~1107. Preprint University of Copenhagen. P.~. W. (1988)_ 'Statistical Inference in Regressions with Integrated Processes. Vol.. 1035~56. A..-. D. Cowles Foundation discussion paper No. Phillips. Homoscedasticity and Serial Independence of Regression Residuals'.. and Juselius. D. Y. (1987).. journal of the Royal Statisical Society B. 'Multiple Time Series Regression with Integrated Processes'.-.. C. 231-54. D. Cowles Foundation discussion paper No. S. pp. Cowles Foundation discussion paper No. R. K. M. Po. 802. M. Journal of Time Series Analysis. 82. S. Vol. pp.. B. S. 866. and Park. C. 12. P. 105~ 18. and Box.forthcoming Econometrica. Economic Letters. pp.~-"- . P. and Watson. American Mathematical Society. 73. C. Vol. pp. S. W. Vol. K. (1988). Econometrica. K. pp. (1986). Review of Economic Studies. 847. and Ouliaris. Velu. C. and Bera. (1988). 'Asymptotic Properties of Residual Based Tests for Cointegration'. Vol. Phillips. P. 479~87. W. 6. 'Optimal Inference in Co integrated Systems'. (1988a). 'The Demand for Money'. West. and Wichern. pp. Tso. B. Vol. and Ouliaris. Econometrica. and Park 1. 473-95. 'Multiple Regression with Integrated Time Series'. 'Efficient Tests for Normality. D. R (1987). journal of Economic Dynamics and Control.. P. (1988).. C. Cowles Foundation discussion paper No. Stock. Phillips. C. J. forthcoming. H. Cowles Foundation discussion paper No. journal of the American Statistical Association. (1986). B. J. and Watson.. Preprint. P. 359-86. Part 1'. Reinsel. and Park. Vol. Sims. G. S. and Durlauf. 'Asymptotic Equivalence of OLS and GLS in Regression with Integrated Regressors'. Velu. 1. 'Estimation and Hypothesis Testing of Cointegration Vectors in Gaussian Vector Autoregressive Models'. Econometric Theory. M. 809. pp. (1989). Vol. Jeganathan. H.---~~---~------BULLETIN 210 Jarque. Philip Allan. -. P. C. pp. Wichern. W..

- On Writing
- The Shell Collector
- I Am Having So Much Fun Here Without You
- The Rosie Project
- The Blazing World
- Hyperbole and a Half
- Quiet Dell
- After Visiting Friends
- Missing
- Birds in Fall
- The Great Bridge
- Who Owns the Future?
- The Wife
- On Looking
- Catch-22
- A Farewell to Arms
- Goat Song
- Birth of the Cool
- Girl Reading
- Galveston
- Telex from Cuba
- Rin Tin Tin
- The White Tiger
- The Serialist
- No One Belongs Here More Than You
- Love Today
- Sail of Stone
- All My Puny Sorrows
- Fourth of July Creek
- The Walls Around Us

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd