The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series V e c to r A u to r e g r e s s io n s a n d C o in te g r a tio n Mark W. Watson Working Papers Series Macroeconomic Issues Research Department Federal Reserve Bank of Chicago December 1993 (W P-93-14) FEDERAL RESERVE BANK OF CHICAGO V ecto r A u to reg ressio n s a n d C o in teg ratio n Mark W. Watson Northwestern University Evanston, Illinois 60208 and Federal Reserve Bank of Chicago First Draft: September 1992 This Draft: August 3, 1993 This paper was prepared for the Handbook o f Econometrics, Vol. 4 (edited by R.F. Engle and D. McFadden). The paper has benefited from comments by Edwin Denson, Rob Engle, Neil Ericsson, Michael Horvath, Soren Johansen, Peter Phillips, Greg Reinsel, James Stock and students at Northwestern University and Studienzentrum Gerzensee. Support was provided by the National Science Foundation through grants SES-89-10601 and SES-91-22463. 1. In tro d u c tio n Multivariate time series methods are widely used by empirical economists, and econometricians have focused a great deal o f attention at refining and extending these techniques so that they are well suited for answering economic questions. This paper surveys two o f the most important recent developments in this area: vector autoregressions and cointegration. Vector autoregressions (VARs) were introduced into empirical economics by Sims (1980), who demonstrated that VARS provide a flexible and tractable framework for analyzing economic time series. Cointegration was introduced in a series o f papers by Granger (1983), Granger and W eiss (1983) and Engle and Granger (1987). These paper developed a very useful probability structure for analyzing both long-run and short-run economic relations. Empirical researchers immediately began experimenting with these new models, and econometricians began studying the unique problems that they raise for econometric identification, estimation and statistical inference. Identification problems had to be confronted immediately in VARs. Since these models don’t dichotomize variables into "endogenous" and "exogenous," the exclusion restrictions used to identify traditional simultaneous equations models make little sense. Alternative sets o f restrictions, typically involving the covariance matrix o f the errors, have been used instead. Problems in statistical inference immediately confronted researchers using cointegrated models. At the heart o f cointegrated models are "integrated" variables, and statistics constructed from integrated variables often behave in nonstandard ways. "Unit root" problems are present, and a large research effort has attempted to understand and deal with these problems. This paper is a survey o f some o f the developments in VARs and cointegration that have occurred since the early 1980s. Because o f space and time constraints, certain topics have been omitted. For example, there is no discussion o f forecasting or data analysis; the paper focuses entirely on structural inference. Empirical questions are used to motivate econometric issues, -1 - but the paper does not include a systematic survey o f em pirical work. Several other papers have surveyed som e o f the m aterial covered here. In particular, the reader is referred to the survey on V A R ’s by C anova (1991), and to surveys on statistical issues in integrated and cointegrated systems by C am pbell and Perron (1991), Engle and Yoo (1991), P hillips(1988), and Phillips and Loretan (1991). Before proceeding, it is useful to digress for a m om ent and introduce som e notation. Throughout this paper, 1(d) w ill denote a variable that is integrated o f ord er d , w here d is an integer. F o r our purposes an 1(d) process can be defined as follows. Suppose that 4>(L)x q t = 0(L )et, w here the the roots o f the polynom ial <f>(z) and 6(z) are outside the unit circle and et is a m artingale difference sequence with variance <P". In other w ords, Xq t follows a covariance stationary and invertible A RM A process . Let x<j t be defined recursively by xd , t = E s = l xd - l , s ’ f ° r d = l * ••• • Then x^ is defined as 1(d). This definition says that an 1(d) process can be interpreted as a d-fold partial sum o f stationary and invertible ARM A process. M any o f the statistical techniques surveyed in this chapter w ere developed to answ er questions concerning the dynam ic relationship between m acroeconom ic tim e series. W ith this in m ind, it is useful to focus the discussion o f econom etric techniques on a set o f concrete econom ic questions. The questions concern a m acroeconom ic system com posed o f eight tim e series: the logarithm s o f output (y), consum ption (c), investm ent (i), em ploym ent (n), nominal wages (w), money (m ), prices (p), and the level o f nom inal interest rates (r). Econom ic hypotheses often restrict the G ranger (1969) causal structure o f the system . A classic exam ple is H all’s (1978) interpretation o f the perm anent incom e/life-cycle m odel o f consum ption. In H all’s m odel, consum ption follows a m artingale, so that ct. j is an optim al forecast o f ct. Thus, the model predicts that no variables in the system w ill G ranger-cause consum ption. W hen the data are integrated, some im portant and subtle statistical issues arise w hen this proposition is tested. F o r exam ple, M ankiw and Shapiro (1985) dem onstrate that unit root problem s plague the regression o f Act onto yt. j : standard critical values for G ranger- -2- causality test statistics lead to rejection of the null hypothesis far too frequently when the null is true. On the other hand, Stock and West (1988) show that these unit root problems disappear when Granger causality is tested using the regression of ct onto ct. j and yt. j , but then reappear in the regression of ct onto ct.j and mt_j. The Mankiw-Shapiro/Stock-West results are explained in Section 2 which focuses on the general problem of inference in regression models with integrated regressors. Economic theories often restrict long-run relationships between economic variables. For example, the proposition that money is neutral in the long run implies that exogenous permanent changes in the level of mt have no long-run effect on the level of yt. When the money-output process is stationary, Lucas (1972) and Sargent (1972) show that statistical tests of long-run neutrality require a complete specification of the structural economic model generating the data. However, when money and output are integrated, Fisher and Seater (1993), show that the neutrality proposition is testable without a complete specification of the structural model. The basic idea is that when money and output are integrated, the historical data contain permanent shocks. Long-run neutrality can be investigated by examining the relationship between the permanent changes in money and output. This raises two important econometric questions. First, how can the permanent changes in the variables be extracted from the historical time series? Second, the neutrality proposition involves "exogenous" components of changes in money; can these components be econo metrically identified? The first question is addressed in Section 3, where, among other topics, trend extraction in integrated processes is discussed. The second question concerns structural identification, and is discussed in Section 4. One important restriction of economic theory is that certain "Great Ratios" are stable. In the eight variable system, five of these restrictions are noteworthy. The first four are suggested by the standard neoclassical growth model. In response to exogenous growth in productivity and population, the neoclassical growth model predicts that output, consumption and investment will grow in a balanced way. That is, even though yt, ct, and ^ increase permanently in -3- response to increases in productivity and population, there are no perm anent shifts in ct-yt and it-yt . T he m odel also predicts that the m arginal product o f capital w ill be stable in the long run, suggesting that a sim ilar long-run stability should be present in ex-post real interest rates, r-Ap. A bsent long-run frictions in com petitive labor m arkets, real wages should equal the m arginal product o f labor. T hus, w hen the production function is Cobb-D ouglas (so that m arginal and average products are proportional), (w -p)-(y-n) should b e stable in the long run. Finally, many m acroeconom ic m odels o f money (e .g ., Lucas (1988)) im ply a stable long-run relation between real balances (m -p), output (y) and nom inal interest rates (r), such as m -p= i3yy+/3rr; that is, these m odels im ply a stable long-run "money dem and" equation. Kosobud and Klein (1961) contains one o f the first systematic investigations o f these stability propositions. They tested w hether the determ inistic grow th rates in the series w ere consistent with the propositions. H ow ever, in m odels w ith stochastic grow th, the stability propositions also restrict the stochastic trends in the variables. These restrictions can be described succinctly: Let xt denote the 8 x 1 vector (yt, c t, it , nt , w t , mt , pt , rt). A ssum e that the forcing processes o f the system (productivity, population, outside m oney, etc.) are such that the elem ents o f xt are potentially 1(1). The five stability propositions im ply that z ^ a ’Xf is 1(0), w here: a — i 1 -1 ’ f i y - i 0 0 oy 0 -1 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 -1 -1 0 0 0 0 0 0 0 0 0 0 1 T he first tw o colum ns o f a are the balanced grow th restrictions, the third colum n is the real w age - average labor productivity restriction, the fourth colum n is stable long-run m oney dem and restriction, and the last colum n restricts nom inal interest rates to be 1(0). I f m oney and prices are 1(1), Ap is 1(0) so that stationary real rates im ply stationary nom inal rates.* -4- These restrictions raise two econometric questions. First, how should the stability hypotheses be tested? This is answered in Section 3.c which discusses tests for cointegration. Second how should the coefficients /Jy and /3f be estimated from the data, and how should inference about their values be carried out? This the subject of Section 3.d on estimating cointegrating vectors. In addition to these narrow questions, there are two broad and arguably more important questions about the business cycle behavior of the system. First, how do the variables respond dynamically to exogenous shocks? Do prices respond sluggishly to exogenous changes in money? Does output respond at all? And if so, for how long? Second, what are the important sources of fluctuations in the variables. Are business cycles largely the result of supply shocks, like shocks to productivity? Or do aggregate demand shocks, associated with monetary and fiscal policy, play the dominant role in the business cycle? If the exogenous shocks of econometric interest -- supply shocks, monetary shocks, etc. -can be related to one-step-ahead forecast errors, then VAR models can be used to answer these questions. The VAR, together with a function relating the one-step-ahead forecast errors to exogenous structural shocks is called a "structural" VAR. The first question —what is the dynamic response of the variables to exogenous shocks -- is answered by the moving average representation of the structural VAR model, and its associated impulse response functions. The second question —what are the important sources of economic fluctuations —is answered by the structural VAR’s variance decompositions. Section 4 shows how the impulse responses and variance decompositions can be computed from the VAR. Their calculation and interpretation is straightforward. The more interesting econometric questions involve issues of identification and efficient estimation in structural VAR models, and the bulk of Section of 4 chapter is devoted to these topics. Before the proceeding to the body of the survey, three organizational comments are useful. First, the sections of this survey are largely self contained. This means that the reader -5- interested in structural VAR’s can skip Sections 2 and 3 and proceed directly to Section 4. The only exception to this is that certain results on inference in cointegrated systemsm, discussed in Section 3, rely on asymptotic results from Section 2. If the reader is willing to take these results on faith, Section 3 can be read without the benefit o f Section 2. The second comment is that Sections 2 and 3 are written at a somewhat higher level than Section 4. Sections 2 and 3 are based on lecture notes developed for a second year graduate econometrics course, and assumes that students have completed a traditional first year econometrics sequence. Section 4, on structural VAR’s, is based on lecture notes from a first year graduate course in macroeconomics, and assumes only that students have a basic understanding o f econometrics at the level o f simultaneous equations. Finally, this survey focuses only on the classical statistical analysis o f 1(1) and 1(0) systems. Many o f the results presented here have been extended to higher order integrated systems, and these extensions will be mentioned where appropriate. 2. Inference in V A R’s w ith In teg rated Regressors 2 .a Introductory Comments: Time series regressions that include integrated variables can behave very differently than standard regression models. The simplest example o f this is the AR(1) regression: yt = p y t_ i+ e t, where p = 1 and et is iid(0,o^). As Stock shows in his chapter o f the Handbook, p, the OLS estimator o f p, has a non-normal asymptotic distribution, is asymptotically biased, and yet is "super consistent," converging to its true value at rate T. Estimated coefficients in VAR’s with integrated components, can also behave differently than estimators in covariance stationary VAR’s. In particular, some o f the estimated coefficients behave like p, with non-normal asymptotic distributions, while other estimated coefficients behave in the standard way, with asymptotic normal large sample distributions. This has profound consequences for carrying out statistical inference, since in some instances, the usual test statistics will not have asymptotic x distributions, while in other circumstances -6- they will. For example, Granger causality test statistics will often have nonstandard asymptotic distributions, so that conducting inference using critical values from the table is incorrect. On the other hand, test statistics for lag length in the VAR will usually be distributed x* in large samples. This section investigates these subtleties, with the objective o f developing a set of simple guidelines that can be used for conducting inference in VAR’s with integrated components. We do this by studying a model composed o f 1(0) and 1(1) variables. Although results are available for higher order integrated systems (see Park and Phillips (1988)(1989), Sims, Stock and Watson (1990) and Tsay and Tiao (1990)), limiting attention to 1(1) processes greatly simplifies the notation with little loss o f insight. 2.b An Example: Many o f the complications in statistical inference that arise in VAR’s with unit roots can be 3 analyzed in a simple univariate AR(2) model: (2 . 1) yt = < f > \ y t- \ + ^ i - 2 Assume that <t>i+<t>2 = 1 and | ^ simple, assume that -*■ T \ I < 1. so that process contains one unit root. To keep things is NIID(0,1). Let xt = (y t_j y ^ ) ’ and _i estimator is 0 = ( E xtxp denote E t = + -*■ ( E xtyt) and t y ' , so that the OLS .i = ( E xtx[) ( E xt7 p* (Unless noted otherwise, £ will 1 throughout this paper.) a In the covariance stationary model, the large sample distribution o f 4> is deduced by writing T*^ ($-<£)= (T ’ ^ E x ^ ’)*1^ T ' 1 E xt*t ^ E x^ijj), and then using a law of large numbers to show that ■ V, and a central limit theorem to show that T ''^ E x^ t ^ N ( 0 ,V). These results, together with Slutsky’s theorem, imply that ^ N (0 ,V ‘ ^). When the process contains a unit root, this argument fails. The most obvious reason is that, when p = 1 , E(Xjxp is not constant, but rather grows with t. Because o f this, T"* E -7- ^ T ' ^ £ xti7t no longer converge: convergence requires that £ xtx’ be divided by T 2 instead o f T, and that £ xtr?t be divided by T instead o f T ^ . Moreover, even with these new scale factors, T £ xtxj converges to random matrix rather than a constant, and T £ xti7t converges to a non-normal random vector. However, even this argument is too simple. The standard approach can be applied to a specific linear combination o f the regressors. To see this, rearrange the regressors in (2.1) so that (2 . 2 ) yt = 7 i Ayt- i + 72yt-l + *ft where 7 j =-<t>2 and 72=^1 +$2 • Regression (2.2) is equivalent to regression (2.1) in the sense that the OLS estimates o f </>j and <f>2 are linear transformations o f the OLS estimators o f 72 * and In terms of the transformed regressors: - -1 (2.3) .72-72. 7 ^ and 72 (and hence <t>) can be analyzed by studying the large sample behavior o f the cross products £ A y 2_j, £ A y ^ y ^ j , £ y 2_^, £ Ayt_jijt, and and the asymptotic behavior o f Eyt-i’lf To begin, consider the terms £A y^_j and £ Ayt_jjjt. Since (2.4) Ayt = 2 = 7 2 = 1> + VV Since 14>21 < 1. Ayt (and hence Ayt_j) is covariance stationary with mean zero. Thus, standard asymptotic arguments imply that T"* £ Aj^ . j o^y and T’ *^ £ Ayt. j 7jt ^ N (0 ,o ^ y ). This means that the first regressor in (2.2) behaves in the usual way. "Unit root" complications arise -8- only because o f the second regressor, yt_j. To analyze the behavior o f this regressor, solve (2.4) backwards for the level o f yt: (2.5) yt = ( l + ^ f 1^ + yo + ^ where f t = E ^ l ^ s 311(1 st = * d + ^ 2)"1 £ i = o ( -^ 2)*+ ^t-i» 311(1 *?i= 0 for i^ O has been assumed for simplicity. Equation (2.5) is the Beveridge-Nelson (1981) decomposition o f yt. It decomposes yt into the sum o f a martingale or "stochastic trend" ( ( H - ^ ) " 1^ ) , a constant (yq) and an 1(0) component (st). The martingale component has a variance that grows with t, and (as is shown below) it is this component that leads to the nonstandard behavior o f the cross products ry?_i, S y t-i^yt-i»311(1 lyt-i^t- Other types o f trending regressors also arise naturally in time series models, and their presence affects the sampling distribution of coefficient estimators. For example, suppose that the AR(2) model includes a constant, so that: (2 . 6) yt = oc + 7 iAyM + 72yt-l + V This constant introduces two additional complications. First, a column of l 's must be added to the list o f regressors. Second, solving for the level of yt as above: (2.7) yt = (i+<f>2)"l a t + ( 1+<f>2)" 1f t + yo + ^ The key difference between (2.5) and (2.7) is that now yt contains the linear trend ( H - ^ ' ^ t This means that terms involving yt. j now contain cross products that involve linear time trends. Estimators of the coefficients in equation (2.6) can be studied systematically by investigating the behavior of cross products o f (i) zero mean stationary components (like ijt and Ayt_j), (ii) -9- constant terms, (iii) martingales and (iv) time trends. We digress to present a useful lemma that shows the limiting behavior o f these cross products. This lemma is the key to deriving the asymptotic distribution for coefficient estimators and test statistics for linear regressions involving 1(0) and 1( 1) variables, for tests for cointegration and for estimators o f cointegrating vectors. While the AR(2) example involves a scalar process, most o f the models considered in this survey are multivariate, and so the lemma is stated for vector processes. 2.c A Useful Lemma Three key results are used in the lemma. The first is the functional central limit theorem. Letting 7jt denote an n x 1 martingale difference sequence, this theorem expresses the limiting behavior o f the sequence o f partial sums £t = £ jijs, t= 1 ,...,T , in terms o f the behavior o f a n x 1 standardized W iener or Brownian motion process B(s) for 0 < s ^ 1.^ That is, the limiting behavior o f the discrete time random walk is expressed in terms o f the continuous time random walk B(s). The result implies, for example, that = > B(s)~N(0,s), for 0 :£ s < 1, where [sT] denotes the first integer less than or equal to sT. The second result used in the lemma is the continuous mapping theorem. Loosely, this theorem says that the limit o f a continuous function is equal to the function evaluated at the limit o f its arguments. The nonstochastic version o f this theorem implies that T*^ jt= T ‘* i(t/T)-* J Q sds=xh . The stochastic version implies that T ' ^ ^ _ ^ t =T"^ ^ _ j ( T ’ 1/^ | t) = > J QB(s)ds. The final result is the convergence o f T"* £ y ^ jj^ to the stochastic integral J QB(s)dB(s)J which is one o f the moments directly under study. These key results are discussed in W ooldridge’s chapter of the Handbook. For our purposes they are important because lead to the following lemma. -10- Lemma 2.c Let ?jt be an n x 1 vector o f random variables with E f y | E ^ t^ t I V l ’ — » ’7 i)= I n ’ ^ , i7j ) = 0 , bounded fourth moments. Let F (L )= J ^ L qF jL^ and G(L) denote two matrix polynomials in the lag operator with J ^ - Q i | Fj | < » and S f - Q i | Gj | < oo . Let | t = S s = i^s* ^ 1 dimensional Brownian motion process. *et E(s) denote an n x Then the following converge jointly: = > F (l) f B(s)ds, (a) r ' A £F(L>T)t = > (c) r 1 e {,[F(L)i,t]' jB (s)d B (s)\ = > F ( l) ’ + j B(s)dB(s) (d) r 1 E [F(L)i,t][G(L)ijt] ’ 2 L T °?i = l,Fr iV G! Ji 3/ 2 E t[F (L )» ,+ 1 r ( f ) T ' 3/ 2 E ? t (g) r 2 E ? t?; (h) r 5/2 e t?t = > (e) r where, to simplify notation J = > j B(s)ds, = > j B(s)B(s)’ds = > q f sdB(s)’F ( l ) \ J sB(s)ds. is denoted by j . The lemma follows from results in Chan and Wei (1988) together with standard versions o f the law of large numbers and the central limit theorem for martingale difference sequences (see White (1984)). Many versions o f this lemma (often under assumptions slightly different from those stated here) have appeared in the literature. For example, univariate versions can be found in Phillips (1986, 1987a), Phillips and Perron (1988) and Solo (1984), while multivariate versions (in most cases covering higher order integrated processes) can be found in Park and Phillips (1988) (1989), Phillips and Durlauf (1986), Phillips and Solo (1992), Sims, Stock and Watson (1990), and Tsay and Tiao (1990). The specific regressions that are studied below fall into two categories: (i) regressions that -11 - include a constant and a martingale as regressors or, (ii) regressions that include a constant, a time trend and a martingale as regressors. In either case, the coefficient on the martingale is the parameter o f interest. The estimated value o f this coefficient can be calculated by including a constant or a constant and time in the regression, or alternatively by first demeaning or detrended the data. It is convenient to introduce some notation for the demeaned and detrended martingale and their limiting Brownian motion representations. Thus, let £ ^ = £ t-T"* E j = i£ s denote the demeaned martingale, and let ? [ = ? t- ^ 1-3 2t denote the A. detrended martingale, where regression o f a and and /32 are the OLS estimators obtained from the onto ( I t ) . Then, from the lemma, a straightforward calculation yields: T ' ^ l f s T ] = > B (s)- f o S W d r ^ B ^ s ) and T ' ^ l [ sT ]= > B (s)- f ^ a 1(r)B(r>drs j Qa2 (r) 13(r)d r= /3T(s), where a^(r)= 4 -6 r and a 2(r)= - 6 + 12r. 2.d Continuing with the Example W e are now in a position to complete the analysis o f the AR(2) example. Consider a scaled version o f (2.3): -1 'T-15>y?.i T'3/25>yt.iyt.i A From (2.5) and result (g) o f the lemma, T*^ £ y ^ .j = > (1 +<£2)'^ J B^(s)ds and from (b) T ' 1 £ y t.iTJt= > ( 1 + 4 >2)_1 $ B(s)dB(s). Finally, noting from (2.4) that A y ^ l + ^ L ) * 1^ , (c) implies that T * ^ £ Ayt. i y t. i $ 0. This result is particularly important because it implies that the limiting scaled "X’X" matrix for the regression is block diagonal. Thus, Tl/4(7i-7i) = (T'1EAy2.1)-1r ' AEAy[. 1tlt + op(l) ? N(0,^y), and -12- T (7 2 ^ 2 ) = c r ^ l y t - p ’ ^ ' ^ y i - i ^ t ) + V 1) = > ( 1 + ^ t J B2(s)ds]_1[ j B(s)dB(s)]. Two features of these results are important. First, 7j and 72 converge at different rates. These rates are determined by the variability o f their respective regressors: y j is the coefficient 72 is the coefficient on a regressor with a variance A that increases at rate t. The second important feature is that 7 j has an asymptotic normal A distribution, while the asymptotic distribution o f 72 is non-normal. Unit root complications will complicate statistical inference about 72 but not y j A A Now consider the estimated regression coefficients <f>j and <#>2 in the untransformed A A 1/ A __ A A A A regression. Since <f>2 = - 71 , T * N(0,o^y). Furthermore, since </>j = y j + 72 * T I/^ ( ^ j - ^ p = T 1/^( 7 j- 7 j ) + T ,/^( 72 -72 ) = T 1'^( 7 j- 7 j)+ O p (l). That is, even though depends on A A A both 7 ^ and 72 , the "super consistency" o f 72 implies that its sampling error can be ignored on a regressor with bounded variance, while in large samples. Thus, so that both </>j and $2 converge at rate T ^ and have asymptotic normal distributions. Their joint distribution is more complicated. Since <t>i+<l>2 = 72> T 1/^ ( ^ j- 0 p + T 1/^(<^2- 02 ) “ T ,/^( 72 -72 ) ^ » the joint asymptotic distribution of T \L -*■ \L +• A A (4>\-4>i) and T singular. The linear combination <t>^+4>2 converges at rate T to a non-normal distribution: > f B(s)2ds]*^[ j B(s)dB(s)]. There are two important practical consequences o f these results. First, inference about <f>j or about <f>2 can be conducted in the usual way. Second, inference about the sum o f coefficients 4>\+4>2 must be carried out using nonstandard asymptotic distributions. Under the null hypothesis, the t-statistic for testing the null HQ: <f>^=c converges to a standard normal random variable, while the t-statistic for testing the null hypothesis HQ: 0 i +^2 = l converges to [ \ B(s)2ds]~'^[ j B(s)dB(s)], which is the distribution o f the Dickey-Fuller r statistic (See Stock’s chapter o f the Handbook). As we will see, many o f the results developed for the AR(2) carry over to more general -13- settings. First, estimates o f linear combinations o f regression coefficients converge at different rates. Estimators that correspond to coefficients on stationary regressors, or that can be written as coefficients on stationary regressors in a transformed regression (y j in this example), converge at rate and have the usual asymptotic normal distribution. Estimators that correspond to coefficients on 1( 1) regressors, and that cannot be written as coefficients on 1(0) regressors in a transformed regression (72 in this example), converge at rate T and have a nonstandard asymptotic distribution. The asymptotic distribution o f test statistics are also affected by these results. Wald statistics for restrictions on coefficients corresponding to 1(0) 7 regressors have the usual asymptotic normal or x distributions. In general, Wald statistics for restrictions on coefficients that cannot be written as coefficients on 1(0) regressors have nonstandard limiting distributions. W e now demonstrate these results for the general VAR model with 1( 1) variables. 2.e A General Framework: Consider the VAR model: (2.8) Yt = o + E ? = 1W i + «, where Yt is an n x 1 vector and et is a martingale difference sequence with constant conditional variance (abbreviated mds(Ef)) with finite fourth moments. Assume that the determinant o f the autoregressive polynomial | I-'& jZ -^z^- ... - | has all of its roots outside the unit circle or at z = 1, and continue to maintain the simplifying assumption that all elements o f Yt are individually 1(0) or 1(1).^ For simplicity, assume that there are no cross equation restrictions, so that the efficient linear estimators correspond to the equation-by-equation OLS estimators. We now study the distribution o f these estimators and commonly used test statistics.** -14- 2 .e .l Distribution o f Estimated Regression Coefficients To begin, write the i’th equation o f the model as: (2.9) yi>t = + «ift, where yj t is the i’th element o f Yt, Xt = ( l Y [.j YJ_2 ... Y|_p)’is the ( n p + 1) vector o f regressors, 0 is the corresponding vector o f regression coefficients, and ej t is the i’th element o f et. (For notational convenience the dependence o f 0 on i has been suppressed.) The OLS estimator o f 0 is P = ( l X ^ t) ' l ( l X ty i t ), so that 0 -0 = ( L X tx p - 1( £ X t€i>t). A*. As in the univariate AR(2) model, the asymptotic behavior of 0 is facilitated by transforming the regressors in a way that isolates the various stochastic and deterministic trends. In particular, the regressors are transformed as Zt = D X t, where D is nonsingular and Zt = ( z i t Z2 t ... z4 t) ’, where the Zj t will be referred to as "canonical" regressors. These regressors are related to the deterministic and stochastic trends given in Lemma 2.c by the transformation: ' FU (L) zl,t Z2 ,t F3l(L) rt Z3,t fN o - F41<L> 0 0 O' 0 0 1 F3 2 F33 0 ^t - F42 F43 F44- _t f2 2 \-i 1 or Zt = F(L) i/j. j where = (t?| 1 | t’ t)! The advantage o f this transformation is that it isolates the terms o f different orders of probability. For example, Zj t are zero mean 1(0) regressors, z ^ is a - 15- constant, the asymptotic behavior o f the regressor Z3 t is dominated by the martingale component F 336t . i 1 ^ d z ^ is dominated by the time trend F ^ t . The canonical regressors Z2 >t and z^ t are scalars, while z ^ t and z ^ are vectors. In the AR(2) example, Z l,t= Ay t - l = ( l + ^ 2 L)‘ \ - l « 80 ^ f h ( L ) = ( 1+ ^ 2L )'^; Z2 >t is absent, since the model did not 23^ —y t-1 —( i + ^ 2)”1€t- i + Yo + h - l >80 ^ F 33 = ( 1+ ^ 2)"1. F3 2 = yo 311(1 F 3 i( L ) = 02 (l + ^ 2)”^(1 and z4 tt 15 ab^n* since yt contains no deterministic drift. contain a constant; Sims, Stock and Watson (1990) provide a general procedure for transforming regressors from an integrated VAR into canonical form. They show that can always be formed so that the diagonal blocks, F ^, i > 2 have full row rank, although some blocks may be absent. They also show that F j 2 = 0 , as shown above, whenever the VAR includes a constant. The details o f their construction need not concern us since, in practice, there is no need to construct the canonical regressors. The transformation from the Xt to the Zj regressors is merely an analytic device. It is useful for two reasons. First, X |D ’(D ’) ' *18= 2 ^7 , with 7 = (D ')’ 1j3. A Thus the OLS A estimators o f the original and transformed models are related by D ’? = 0 . Second, the A asymptotic properties o f 7 are easy to analyze because of the special structure o f the A regressors. Together these imply that we can study the asymptotic properties o f 0 by first A studying the asymptotic properties of A 7 and then transforming these coefficients into the /S’s. The transformation from Xt to Zt is not unique. All that is required is some transformation that yields a lower triangular F(L) matrix. Thus, in the AR(2) example we set z ^ t =A yt_i and Z3 t = y t_i, but an alternative transformation would have set z j t = A yt_j and Z3 t = y t_2 - Since we always transform results for the canonical regressors Zt back into results for the "natural" regressors Xt, this non-uniqueness is o f no consequence. A W e now derive the asymptotic properties o f y- t = Z | 7 +ej t. Writing sequence from Lemma 2.c, then 7-7 =*( £ Z jZ p 7 constructed from the regression where ijt is the standardized n x 1 martingale difference t = u'ijt = tjJw, where w’ is the i’th row o f 2^, and ( £ ZtJ?t’«). Lemma 2.c can be used deduce the asymptotic behavior o f £ ZtZ[ - 16- and E Z S o m e care must be taken, however, since each o f the Zj t elements of Zt are growing at different rates. Assume that z j t contains k j elements, Z3 t contains k^ elements, and partition 7 conformably with Zj as 7 = ( 7 j 72 73 74)* where 7j are the regression coefficients corresponding to z: t. Let r H T Jk X 0 0 0 0 T*1 0 0 0 0 0 TIk 3 0 l Cs| N CO H 0 1 0 and consider ^ j ( 7 - 7 ) = 0k j * E ZtZ^kj E Z ^ w ). The matrix 'k j multiplies the various blocks o f (7^-7^), E ZtZJ, and E Z ^ by the scaling factors appropriate from the lemma. The first block of coefficients, 7 j, are coefficients on zero mean stationary components and are scaled up by the usual factor o f T ^ ; the same scaling factor is appropriate for 72 , the 73 are coefficients on regressors dominated by martingales, and these need to be scaled by T; finally, 74 is a coefficient on a regressor constant term; the parameters making up dominated by a time trend and is scaled by T ^ . Applying the lemma, we have ¥ E ZtZ j * = > V, where, partitioning V conformably with Zt: E jF H j F h j - / F§2 = v 22 = > F33[ T'j/ 2 S z i >tZj<t S F44/3 o P f 22 f 44/2 r ” = ^ ^33 I sB(s)dsF^ l V 4 ,l - 17- > II (N > II T'2 ^ z2 ,tz 4 1t II CO <N > II = > F22 i B(s)$d sF jj II ^ B(s)B(s)’ds]F33 > II T ' 3 E (z 4jt)2 \ II < U) u> T 1^ = vn II < £ 5 II T where the notation reflects the fact that F 22 and F 44 are scalars. The limiting value o f this scaled moment matrix shares two important characteristics with its analogue in the univariate AR(2) model. First, V is block diagonal with V jj = 0 for j £ 1. (Recall that in the AR(2) model 7 *3/2 ^ Ayt_jyt.jB o). AR(2) model T -2 Second, many o f the blocks o f V contain random variables. (In the 2 £ y |_ j converged to a random variable.) Now, applying the lemma to ¥ j 1 £ ZjijJw yields ' k j 1 £ = > A, where, partitioning A conformably with Zt: F 22 J dB(s)’w= = A = > F 33 \ B(s)dB(s)’w = > F 44 J sdB(s)’« Putting the results together, ^ ( 7A -7 ) A1 Tf < II r l r z3 ,tit’" T-3/2 r* * T E z 4 (tijtU = > = > N [0,(w\o)V n ] II - -1 = > V A, and three important results follow. First, the A. individual coefficients converge to their values at different rates: 7j A and 72 converge to their values at rate T ^ , while all of the other coefficients converge more quickly. Second, the block diagonality of V implies that T ^ ( 7 j -7 j ) ^ N (0 ,a ? V ‘ J j), where o?= w ’w=var(ej). Moreover, A j is independent o f Aj for j > 1 (Chan and Wei (1988), Theorem 2.2), so that \L * T (7 j *7 j ) is asymptotically independent of the other estimated coefficients. Third, all o f the other coefficients will have non-normal limiting distributions, in general. This follows because Vj 3 =£ 0 for j > 1, and A 3 is non-normal. A notable exception to this general result is when the canonical regressors do not contain any stochastic trends, so that Z3 t is absent from the model. In this case V is a constant and A is normally distributed, so that the estimated coefficients n have a joint asymptotic normal distribution. The leading example o f this is polynomial regression, when the set o f regressors contains covariance stationary regressors and polynomials -18- in time. Another important example is contained in West (1988), who considers the y a i? r unit root AR(1) model with drift. The asymptotic distribution o f the coefficients & that correspond to the "natural" regressors Xt can now be deduced. It is useful to begin with a special case o f the general model: (2 . 10) yi>t = fix + x 2 ,t’02 + x3 , t h + ei,t where x j t = 1 for all t, %2 t is a h x 1 vector of zero mean 1(0) variables, and X3 t contains the other regressors. It is particularly easy to transform this model into canonical form. First, since x j t = 1, we can set Z2 t =x^ t; thus in terms of the transformed regression, /3j = 72- Second, since the elements o f X2 t are zero mean 1(0) variables, we can set the first h elements o f Zj t equal to X2 t; thus is equal to the first h elements of The remaining elements of zt are some linear combination o f the regressors that need not concern us here. In this example, since #2 is a subset o f the elements o f 7 j , T (^ -Z ^ ) is asymptotically normal and independent of the coefficients corresponding to trend and unit root regressors. This results is very useful since it provides a constructive sufficient condition for estimated coefficients to have an asymptotic normal limiting distribution: whenever the block o f coefficients can be written as coefficients on zero mean 1(0) regressors in a model that includes a constant term they will have a joint asymptotic normal distribution. A A Now consider the general model. Recall that fi=D’y. Let dj denote the j ’th column o f D, and partition this conformably with 7 , so that dj = (djj A A d^ A d^j)’, where djj and A 7 - are the same dimension. Then the j ’th element o f 0 is /3j = £ jd*j7 |. Since the A A A components o f 7 converge at different rates, /3j will converge at the slowest rate o f the 7 ^ -*■ |L included in the sum. Thus, when d jj^ O , /9j will converge at rate T , the rate o f convergence A of 7i- - 19- 2.e.2 Distribution o f Wald Test Statistics Consider Wald test statistics for linear hypotheses o f the form R £ = r, where R is a q x k matrix with full row rank: W = (R0-r) ’[R( l X jX p ' 1R ’] '1(Rj3-r)/^. (Recall that /3 corresponds to the coefficients in the i’th equation, so that W tests withinequation restrictions.) Letting Q = R (D ’), an equivalent way o f writing the Wald statistic is in terms of the canonical regressors and their estimated coefficients 7: W = ( Q r O ’f Q f E Z t Z j r ^ T ^ Q Y - r ) / ^ . Care must be taken when analyzing the large sample behavior o f W because the individual A. coefficients in 7 converge at different rates. To isolate the different components, it is useful Q to assume (without loss o f generality) that Q is upper triangular. Now, partition Q, A 7 and the canonical regressors making up Zj, so that Q = [q-j ] where qy is matrix representing qj constraints on the kj elements in 7j. Since Q is assumed to be conformably with q -x k j upper triangular and o f full row rank, the matrices q ^ have full row rank for all i. Partition r = ( r i r^ r j r^)’ conformably with Q. A A Now consider the first q j elements o f Q 7 : q^ ^ j >2 A converges more quickly than A A A + cli2 7 2 +cll3 T 3 + (ll4 7 4 ' Since A t j, for A 7 ^ and 72 , the sampling error in this vector will be A A dominated asymptotically by the sampling error in Q n 7 i+ < li 272 * Similarly, the sampling A A error in the next group o f q 2 elements o f Q 7 is dominated by q2272> “ A ^ 33^ 3 ’ etc* ^ us» A appropriate scaling matrix for Q 7 *r is: -20- next <13 t y T4T „ <U 0 0 0 th i 0 0 0 TI 0 0 0 <12 0 1 0 \ q3 0 T3/2I 1 j Now, write the Wald statistic as: W = (Q 7 -r)’1irT ,[1irTQ (i: ZtZp* 1Q ,,irT] ' 1,i rT(Q 7 -r)/oj But, under the null, |/ T A. A. A A J/ (£lll7 i+ < li2 'y 2 +<ll373+ < li474 " r l) = T T° ' 1)/ 2 <<'jj^j + - A, A, (91171+91272 ' r l) + °p O ). ^ + <>44^4 ' r4> = T0 ' 072^ ^ * rj) + V ' ) . Thus, if we let qn q12 0 0 0 q22 0 0 0 0 q^j 0 .0 0 0 Q q4J , then ^ x (Q7-D = Q * t <7-7) + op(l) under the null. 9 Similarly, it is straightforward to show th a t: -21 - + Op(l). Finally, since ¥ 7 (7 -7 ) = > V "1A and ¥ T** E Z j Z ^ j 1= > V, then W = > ( Q V U j ’C Q V ^ Q J '^ Q V U ). The limiting distribution o f W is particularly simple when q ^ = 0 for i ^ 2 . In this case, all of the hypotheses o f interest concern linear combinations o f zero mean 1(0) regressors, together with the other regression coefficients. When q ^ - O , so that the constant term is unrestricted, we have: W - [qn(7i-7i)]’[qii(Eziftzl,t’) M i l *fall(7i*7i)] + Op(l) so that W -* Xqj. When the constraints mvolve other linear combinations o f the regression 2 coefficients, the asymptotic x distribution o f the regression coefficients will not generally obtain. This analysis has only considered tests o f restrictions on coefficients from the same equation. Results for cross equation restrictions are contained in Sims, Stock and Watson (1990). The same general results carry over to cross equation restrictions. Namely, restrictions that involve subsets o f coefficients that can be written as coefficients on zero mean stationary regressors in regressions that include constant terms can be tested using standard asymptotic distribution theory. Otherwise, in general, the statistics will have nonstandard limiting distributions. -22- 2. f Examples: 2 .f 1 Testing Lag Length Restrictions Consider the V A R (p+s) model: *t - “ + r pi l i ^ c - i + «r and the null hypothesis HQ:$ p + \ = ^ p + 2 ~ " ' = ^ p + s = ®* says that the true model is a VAR(p). When p 2: 1, the usual Wald (and LR and LM) test statistic for HQ has an asymptotic \ 9 distribution under the null. This can be demonstrated by rewriting the regression so that the restrictions in HQ concern coefficients on zero mean stationary regressors. Assume that AYt is 1(0) with mean n, and then rewrite the model as: Y, = a + A Y ,.! + E jP_+1S' 10|(A Y t.!-n) + e,, where A = E ^ ! ^ . 0 , ^ P j a n d a=a+ V The restrictions $ p + l = ,l»p_l_2 = . . . = $ p + s = 0 , *n the original model are equivalent to 9 p = 0 p + i = .. . = 9 p + s .^ in the transformed model. Since these are coefficients are zero mean 1(0) regressors in regression 9 equations that contain a constant term the test statistics will have the usual large sample x distribution. 2.f.2 Testing fo r Granger Causality: Consider the bivariate VAR model: yi.t - “1 + E ? = l * l l fiyi,t-i + S ? = 1 ^ 12 ,iy2 ,t-i + el,t y2,t = a2 + 1*21^1,t-i + E?=i*22,iy2,t-i + c2,f -23- The restriction that y 2 >t does not Granger cause y j t corresponds to the null hypothesis H0 : ^ i 2 , l =<^12,2= - '- = ^ 1 2 ,p = ®* ^ en ( y ^ t y i,\) ^ covariance stationary, the resulting Wald, LR or LM test statistic for this hypothesis will have a large sample Xp distribution. When (yi t y 2 ,t) ^ integrated» the distribution of the test statistic depends on the location o f unit roots in the system. For example, suppose that y j t is 1(1), but that y2 t is 1(0). Then, by writing the model in terms o f deviations o f y 2 t from its mean, the restrictions involve only coefficients on zero mean 1(0) regressors. Consequently, the test statistic has a limiting Xp distribution. When y 2 t is 1(1), then the distribution o f the statistic will be asymptotically \ when y j t and y 2 t are cointegrated; when y j t and y 2 t are not cointegrated, the Granger causality test statistic will not be asymptotically x , in general. Again, the first result is easily demonstrated by writing the model so the coefficients o f interest appear as coefficients on zero mean stationary regressors. In particular, when y j t and y2 t are cointegrated, there is an 1(0) linear combination o f the variables, say wt = y 2 t-Xyj t, and the model can be rewritten as: yi,t = “ i + E?=i^n,iyi,t-i+ S?=i^>i2 ,i(w t-r^w)+ «i,t where nw is the mean o f wt, a j = a + E ? = i ^ i 2 ,iMw and i^» i = l . —»p. In the transformed regression, the Granger causality restriction corresponds to the restriction that the terms wt_ ^ w do not enter the regression. But these are zero mean 1(0) regressors in a regression that includes a constant, so that the resulting test statistics will have a limiting Xp distribution. When y j t and y 2 t are not cointegrated, the regression cannot be transformed in 2 this way, and the resulting test statistic will not, in general, have a limiting x distribution. 10 The Mankiw-Shapiro (1985)/Stock-West (1988) results concerning H all’s test o f the lifecycle/permanent income model can now be explained quite simply. Mankiw and Shapiro considered tests o f H all’s model based on the regression o f Act (the logarithm o f consumption) -24- onto y t_i (the lagged value o f the logarithm o f income). Since y t_j is (arguably) integrated, its regression coefficient a n d t-statistic will h a v e a non-standard limiting distribution. Stock a n d West , following Hall’s (1978) original regressions, considered regressions o f c t onto ct.j a n d y t_ j. Since, according the life-cycle/permanent i n c o m e model, ct_j a n d y t.j are cointegrated, the coefficient o n y t_j will b e asymptotically n o r m a l a n d its t-statistic will h a v e a limiting standard n o r m a l distribution. H o w e v e r , w h e n y t.j is replaced in the regression with m ^ (the lagged value o f the logarithm o f m o n e y ) , the statistic will not b e asymptotically normal, since ct.j a n d m t_i are not cointegrated. A m o r e detailed discussion o f this e x a m p l e is contained in Stock a n d W e s t (1988). 2.f.3 Spurious Regressions In a very influential paper in the 1 9 7 0 ’s, G r a n g e r a n d N e w b o l d (1974) presented M o n t e Carlo evidence re minding economists o f Y u l e ’s (1926) spurious correlation results. Specifically, 2 G r a n g e r a n d N e w b o l d s h o w e d that a large R a n d a large t-statistic w e r e not unusual w h e n o n e r a n d o m w a l k w a s regressed o n another, statistically independent, r a n d o m walk. Their results w a r n e d researchers that standard meas ur es o f fit can b e very misleading in "spurious" regressions. Phillips (1986) s h o w e d h o w these results could b e interpreted quite simply, a n d his analysis is s u m m a r i z e d here. Let yj t a n d y 2 t b e t w o independent r a n d o m walks: yi,t = yi,t-i + ci,t y2,t= y2,t-i+ c2,t w h e r e €t= ( e j 1 62 t)’ is a m d s ( E f) with finite fourth m o m e n t s , a n d {tj are mutually independent. F o r simplicity, set yj Q = y 2 y2 ,tont°yi,t: -25- q a n d {«2 ,tJt=l = 0- Cons id er the linear regression of (2 .U ) y2>t = ^ y 1>t + ut, where ut is the regression error. Since y ^ t and y 2 t are statistically independent 0 = 0 and ut=y2,f Now consider three statistics, the OLS regression estimator, the regression R2 and the usual t-statistic for testing the null that 0 = 0 : 2 = [£y2,tyi,tHI(y,it)2] r2 = t l y 2 ,tyi,,l2/[r(yi,,)2 m / l r = 0 /S J where (S£) 2 = T ’ *[ £ (y2 t )2-0 £ y 2 , t y i , t ^ E (yi >t)2] ‘s toe usual formula for the variance o f 0. When y j t and y 2 t are mutually independent and HD, standard results imply that 0 -fl}, R 2 ^ 0 <e and r -* N (0 ,1). When y j t and y 2 t are mutually independent random walks things are quite different. Let B(s) denote a 2 x 1 Brownian motion process, V = J B(s)B(s)’ds and vy denote the ij’th element o f V. Then, utilizing the Lemma 2.c: (2.12) 0 = > (<72/alXv21^v ll) (2.13) R2 = > (2.14) T l/zr = > v 21/(v n v 22-v ^1) ,/4, (v 2 i)/(v n v22) where o |= v a r(€ j), i = 1,2. Thus, both 0 and R 2 converge to non-degenerate random variables, while r diverges. This shows that large absolute values o f the t-statistic should be expected in "spurious" regressions. When the estimated regression (2.11) contains a constant or a constant and time trend, -26- similar results obtain with the demeaned and detrended Brownian motion processes B^(s) and Bt( s) replacing B(s). When the regression contains a constant, the results are invariant to the initial conditions for y ^ t and y 2 t; when the regression contains a constant and a time trend the results are invariant to initial conditions and drift terms in y j t and y2 t. See Phillips (1986) for a more detailed discussion. 2.f.4 Estimating Cointegrating Vectors by Ordinary Least Squares Now suppose that y j t and y 2 t are generated by: (2.15) Ay l t = (2.16) » 1>t y2 t = /3y,_t + u2 t where ut = (u j t U2 t) ’ = D et, where et is a mds(l2) with finite fourth moments. Like the spurious regression model, both y j t and y2 t are individually 1( 1): y j t is a random walk, while y 2 t follows a univariate ARIM A(0,1,1) process. Unlike the spurious regression model, one linear combination of the variables y2 t-jSyj t =U2 t is 1(0), and so the variables are cointegrated. Stock (1987) derives the asymptotic distribution of the OLS estimator o f cointegrating vectors. In this example, the limiting distribution is quite simple. W rite (2.17) 5-iS = [ E y ljtu2 t] / [ S ( y i , t)2l. let dy denote the ij’th element o f D, and D j= (d ^ d ^ ) denote the i’th row o f D. Then the limiting behavior or the denominator o f 0-/3 follows directly from the lemma: (2.18) r 2 E ( y j t)2 = D [[T ‘2 E { t{;]D i = > D ,t i B(s)B(s)'ds]Di -27- where is the bivariate random walk, A |t = e t and B(s) is a 2 x 1 Brownian motion process. The numerator is only slightly more difficult: (2.19) T^IyjtU^ = T ^ E y j ^ U j t + T ^ E A y j^ t = D j l T 11 *t.i41D i + 1 1 et€|]D^ » > D j[ J B(s)dB(s)’]D£ + D ^ . Putting these two results together: (2.20) TOS-0) = > [ D ^ J B(s)dB(s)’]D ^ + D 1D ^ ] P 1[ f B(s)B(s)’ds]D i]*1. There are three interesting features of the limiting representation (2.20). First, j3 is "super A consistent," converging to its true value at rate T. Second, while super consistent, 0 is asymptotically biased, in the sense that the mean o f the asymptotic distribution in not centered at zero. The constant term D j D ^ d ^ ^ + ^ n ^ l ^ at i*1 the numerator o f (2.20) is primarily responsible for this bias. To see the source o f this bias, notice that the regressor y j t is correlated with the error term U2 t. In standard situations, this "simultaneous equation bias" A. is reflected in large samples as an inconsistency in /3. With cointegration, the regressor is 1(1) and the error term is 1(0), so no inconsistency results; "the simultaneous equations bias" shows A up as bias in the asymptotic distribution of 0 . In realistic examples this bias can be quite large. For example, Stock (1988) calculates the asymptotic bias that would obtain in the OLS estimator o f the marginal propensity to consume obtained from a regression o f consumption onto income using annual observations with a process for ut similar to that found in U .S. data. He finds that the bias is still -. 10 even when 53 years o f data are used. * * Thus, even though the OLS estimators are "super" consistent, they can be quite poor. The third feature o f the asymptotic distribution in (2.20) involves the special case in which -28- d 12= d 2 1 =0 50 ^ u l , t 311(1 u2 , t 816 statistically independent. In this case the OLS estimator corresponds to the Gaussian MLE. When d j 2 =cl2 1 = ®* (2-20) simplifies to: (2.21) T<0-0) = > (d22/d 1j)[ J B ^ d B ^ s ) ] ! J B ^ d s ] * 1. where B(s) is partitioned as B (s)= (B j(s) B2 (s))’. This result is derived in Phillips and Park (1988) where the distribution is given a particularly simple and useful interpretation. To develop the interpretation, suppose for the moment that u2 t = d 22c2 t was HD and normal. (In large samples the normality assumption is not important; it is made here to derive simple and A exact small sample results.) Now, consider the distribution o f 0 conditional on the regressors {yi t}t=s j. Since u2 t is NUD, the restriction d j 2 = d 2 j = 0 implies that u 2 t is independent o f {y i >t} [ - 1• This means that 0 - 0 1 { y i>t}^_ \ - N(0,d22[ £ ( y ^ ) 2]"1), so that the A. unconditional distribution >3-/3 is normal with mean zero and random covariance matrix, d22[ E (y i >t) V 1• In large samples, T '2 E (y i ^ = > d 11 f B2(s)ds, so that T( 0 -0 ) converges to a normal random variable with a mean of zero and random covariance matrix, (d22/d j j) [ 5 B |(s)ds] . Thus, T(/3-/3) has an asymptotic distribution that is a random mixture of normals. Since the normal distributions in the mixture have a mean o f zero, the A asymptotic distribution is distributed symmetrically about zero, and thus 0 is asymptotically median unbiased. A The distribution is useful, not so much for what it implies about the distribution o f 0 , but A for what it implies about the t-statistic for 0. When d j 2 or d2 j are not equal to zero, the tstatistic for testing the null /3=/3q has a non-standard limiting distribution, analogous to the distribution o f the Dickey-Fuller t-statistic for testing the null o f a unit AR coefficient in a univariate regression. However, when ^ \ 2 ~ ^ 2 \ ^ t_statistic has a limiting standard normal distribution. To see why this is true, again consider the situation in which u2 t is HD and normal. When d j 2 = d 2 j = 0 , the distribution o f the t-statistic for testing /3=/Sq conditional on -29- T {y1>t}{= j has an exact Student’s t distribution with T -l degrees o f freedom. Since this distribution does not depend on this is the unconditional distribution as well. This means that in large samples, the t-statistic has a standard normal distribution. As we will see in this next section, the Phillips and Park (1988) result carries over to a much more general setting. In the example developed here, ut = D et is serially uncorrelated. This simplifies the analysis, but all o f the results hold more generally. For example, Stock (1987) assumes that ut =D (L)et, where D(L) = E ° j° = 0° ^ , I D (l) | * 0 and (2.22) ji | Dj | < « . In this case, T(/3-0) = > [D 1(l)[ J B(s)dB(s)’]D 2 ( l ) ’+ r T = 0D i , iDi,i3CD i( l) [ f B(s)B(s)’ds]D 1( l ))-1 where D j(l) is the j ’th row o f D (l) and Dj j is the j ’th row o f Dj. Under the additional A. assumption that d i 2 ( l ) =<* 2 l(l)= 0 ’ T03-/3) is distributed as a mixed normal (asymptotically) and the the t-statistic has an asymptotic normal distribution when d j 2 ( l ) = d 2 i ( l )= 0 (see Phillips and Park (1988) and Phillips (1991a)). 2.g Implications for Econometric Practice The asymptotic results presented above are important because they determine the appropriate critical values for the tests o f coefficient restrictions in VAR models. The results lead to three lessons that are useful for applied practice: (1) Coefficients that can be written as coefficients on zero mean 1(0) regressors in regressions that include a constant term are asymptotically normal. Test statistics for restrictions on these coefficients have the usual asymptotic x distributions. For example, in the model -30- (2.23) yt = 7 i z i #t + 72 + 73 z3 ,t + 7^ + where z j t is a mean zero 1(0) scalar regressor and z^ t is a scalar martingale regressor, this result implies that Wald statistics for testing HQ: 7 j= c is asymptotically x^. (2) Linear combinations o f coefficients that include coefficients on zero mean 1(0) regressors together with coefficients on stochastic or deterministic trends will have asymptotic normal distributions. Wald statistics for testing restrictions on these linear combinations will have large 9 sample x distributions. Thus in (2.23), Wald statistics for testing HQ: R j 7 j + ^ 2 have an asymptotic x distribution if R j £ 0 . 73 + ^ 7^ = r, will (3) Coefficients that cannot be written as coefficients on zero mean 1(0) regressors (e.g. constants, time trends, and martingales) will, in general, have nonstandard asymptotic distributions. Test statistics that involve restrictions on these coefficients that are not a function of coefficients on zero mean 1(0) regressors will, in general, have nonstandard asymptotic distributions. Thus in (2.23), Walds statistics for testing: HQ: R (72 73 2 non-x asymptotic distributions, as do tests for composite hypotheses of the form Ho : R<72 73 74) ’ = r and 74) ’ = r have =c. When test statistics have a nonstandard distribution, critical values can be determined by Monte Carlo methods by simulating approximations to the various functionals o f B(s) appearing in Lemma 2.c . As an example, consider using Monte Carlo methods to calculate the asymptotic distribution o f sum o f coefficients < f> ^ + < f> 2=72 model (2.1). Section 2.d showed that T(72-72) = > in univariate AR(2) regression +<h ) i J B^(s)ds]'*[ f B(s)dB(s)], where B(s) scalar Brownian motion process. If xt is generated as a univariate Gaussian random walk, then 9 one draw o f the random variable [ J B (s)ds] -31 - 1 [ J B(s)dB(s)] is well approximated by (T 2 £ x t) ^ E xtAxt+ j) with T large. (A value of T =500 provides an adequate approximation for most purposes.) The distribution o f taking repeated draws o f (T '2 £ x2)*1^ ' 1£ x ^ H 6311 *hen be approximated by - l ) multiplied by (1 + 4 ^ ). An example o f this approach in a more complicated multivariate model is provided in Stock and Watson (1988). Application o f these rules in practice requires that the researcher know about the presence and location o f unit roots in the VAR. For example, in determining the asymptotic distribution o f Granger causality test statistics, the researcher has to know whether the candidate causal variable is integrated and, if it is integrated, whether it is cointegrated with any other variable in the regression. If it is cointegrated with the other regressors, then the test statistic has a asymptotic distribution. Otherwise the test statistic is asymptotically non-x , in general. In practice such prior information is often unavailable, and an important question is what is to be done in this case? 12 The general problem can be described as follows. Let W denote the Wald test statistic for a hypothesis of interest. Then the asymptotic distribution o f the Wald statistic when a unit root is present, say F(W | U), is not equal to the distribution o f the statistic when no unit root is present, say F(W | N). Let Cy and Cj^ denote the "unit root" and "no unit root" critical values for a test with size a . That is, Cj j and c ^ satisfy: P (W > C y | U )= P (W > C j^ | N ) = a . The problen that Cj j ^C j^, and the researcher does not know whether U or N is the correct specification. In one sense, this not an unusual situation. Usually, the distribution o f statistics depends on characteristics o f the probability distribution o f the data that are unknown to the researcher, even under the null hypothesis. Typically, there is uncertainty over certain "nuisance parameters," that affect the distribution o f the statistic o f interest. Yet, typically the distribution depends on the nuisance parameters in a continuous fashion, in the sense that critical values are continuous functions o f the nuisance parameters. This means that asymptotically valid inference can be carried out by replacing the unknown parameters with consistent estimates. -32- This is not possible in the present situation. While it is possible to represent the uncertainty in the distribution o f test statistics as a function o f nuisance parameters that can be consistently estimated, the critical values are not continuous functions o f these parameters. Small changes in the nuisance parameters — associated with sampling error in estimates - may lead to large changes in critical values. Thus, inference cannot be carried out by replacing unknown nuisance parameters with consistent estimates. Alternative procedures are required. ^ Development o f these alternative procedures is currently an active area o f research, and it is too early to speculate on which procedures will prove to be the most useful. It is possible to mention a few possibilities and highlight the key issues. The simplest procedure is to carry out conservative inference. That is, to use the largest of the "unit root" and "no unit root" critical values, rejecting the null when W >m ax(Cu,Cj^). By construction, the size of the test is less than or equal to a . Whenever W > m ax (cjj,c^), so that the null is rejected using either distribution or W < m in (c u ,c ^ ), so that the null is not rejected using either distribution, one need not proceed further. However a problem remains when m in(C u,c^) < W <m ax(Cy,Cj^). In this case, an intuitively appealing procedure is to look at the data to see which hypothesis — unit root or no unit root -- seems more plausible. This approach is widely used in applications. Formally, it can be described as follows. Let 7 denote a statistic helpful in classifying the stochastic process as a unit root or no unit root process. (For example 7 might denote a Dickey-Fuller "t-statistic" or one of the test statistics for cointegration discussed in the next section.) The procedure is then to define a region for 7 , say Ry, and when 7 6 ! ^ , the critical value Cjj is used; otherwise the critical value Cj^ is used. (For example, the unit root critical value might be used if the Dickey-Fuller "t-statistic" was greater than -2, and the no-unit root critical value used when the DF statistic was less than -2.). In this case, the probability of type 1 error is: P(Type 1 error) = P(W > Cy 17 £ R ,j)P (7 € R y )+ P (W > cN 17 € R (j)P (7 « % ) • -33- The procedure will work well, in the sense o f having correct size, and power close to the power that would obtain when the correct unit root or no unit root specification were known, if two conditions are met: First, P f y E R y ) and P fy E R jj) should be near 1 when the unit root and no-unit root specification are true, respectively. Second P (W > C y | y E R ^ ) and P(W > Cjsj 17 € R j j ) should be near P(W > Cy | U) and P(W > C j^ | N) respectively. U nfortunately,: practice neither o f these conditions may be true. The first requires statistics that perfectly discriminate between the unit root and non-unit root hypothesis. W hile significant progress has been made in developing powerful inference procedures (e.g., Dickey-Fuller (1979), Elliot, Rothenberg and Stock (1992), Phillips and Ploberger (1991), Stock (1992)), high probability of classification errors are unavoidable in moderate sample sizes. In addition, the second condition may not be satisfied. An example presented in Elliot and Stock (1992) makes this point quite forcefully. (Also see Cavanagh and Stock (1985).) They consider the problem o f testing whether the price-dividend ratio helps to predict future changes in stock prices. ^ A stylized version o f the model is: (2.24) pt-dt = <£(pt_j-dt_j) + (2.25) Apj = £(Pt-i*dt_i) e2,t where pt and dt are the logs o f prices and dividends respectively and (cj t ^ t) ’ is a mds(Ee). The hypothesis o f interest is HQ: 0 = 0 . Under the null, and when | <f>| < 1, the t-statistic for this null will have an asymptotic standard normal distribution; when the hypothesis ^ = 1, the tstatistic will have a unit root distribution. (The particular form o f the distribution could be deduced using Lemma 2.c, and critical values could be constructed using numerical methods.) The pretest procedure involves carrying out a test o f <f>= 1 in (2.24), and using the unit root critical value for the t-statistic for 0 = 0 in (2.25) when <t> = 1 is not rejected. If ^ = 1 is rejected, -34- the critical value from the standard normal distribution is used. Elliot and Stock show that the properties of this procedures depend critically on the correlation between and 62 f To see why, consider an extreme example. In the data, dividends are much smoother than prices, so that most o f the variance in the price-dividend ratio comes from movements in prices and not from dividends. Thus, ej t and ^ ta r e likely to A be highly correlated. In the extreme when they are perfectly correlated, 03-/3) is proportional a to (<£-<£), and the "t-statistic" for testing /3=0 is exactly equal to the "t-statistic" for testing <f>=l. A In this case F(W 17 ) is degenerate and does not depend on the null hypothesis. All o f the information in the data about the hypothesis j8=0 is contained in the pre-test. While this example is extreme, it does point out the potential danger o f relying on unit root pretests to choose critical values for subsequent tests. 3. C ointegrated Systems 3.a Introductory comments An important special case of the model analyzed in Section 4 is the cointegrated VAR. This model provides a framework for studying the long-run economic relationships discussed in the Introduction. There are three important econometric questions that arise in the analysis o f cointegrated systems. First, how can the common stochastic trends present in cointegrated systems be extracted from the data? Second, how can the hypothesis o f cointegration be tested? And finally, how should unknown parameters in cointegrating vectors be estimated, and how should inference about their values be conducted? These questions are answered in this section. We begin, in Section 3.b, by studying different representations for cointegrated systems. In addition to highlighting important characteristics o f cointegrated systems, this section provides an answer to the first question by presenting a general trend extraction procedure for cointegrated systems. Section 3.c discusses the problem o f testing for the order o f cointegration, and Section 3.d discusses the problem of estimation and inference for unknown parameters in -35- cointegrating vectors. To keep the notation simple, the analysis in Sections 3.b-3.d abstracts from deterministic components (constants and trends) in the data. The complications in estimation and testing that arise when the model contains constants and trends is the subject o f Section 3.e. In addition, only 1(1) systems are considered. Using Engle and Granger’s (1987) terminology, the section discusses only CI(1,1) systems; that is, for systems in which linear combinations o f 1(1) and 1(0) variables are 1(0). Extensions for CI(d,b) systems with d and b different from 1 are presented in Johansen (1988b)(1992c), Granger and Lee (1990), and Stock and Watson (1993). 3.b Representations for the I fll Cointegrated Model Consider the VAR (3.1) xt = + et where xt is an n x 1 vector composed o f o f the variables in the system are n ( z )= I - 1(0) 1(0) or and 1(1), 1(1) variables, and is a mds(E€). Since each the determinantal polynomial | II(z) | , with —jlljz 1, contains at most n unit roots. When there are fewer than n unit roots, then the variables are cointegrated, in the sense that certain linear combinations o f the xt’s are 1(0). In this subsection we derive four useful representations for cointegrated VARs: (1) the vector error correction VAR model, (2) the moving average representation o f the first differences o f the data, (3) the common trends representation o f the levels o f the data, and (4) the triangular representation o f the cointegrated model. All of these representations are readily derived using a particular Smith-McMillan factorization o f the autoregressive polynomial II(L). The specific factorization used here was originally developed by Yoo (1987) and was subsequently used to derive alternative representations o f cointegrated systems by Engle and Yoo (1991). Some o f the discussion -36- presented here parallels the discussion in this latter reference. Yoo’s factorization o f II(z) isolates the units roots in the system in a particularly convenient fashion. Suppose that the polynomial II(z) has all o f its roots on or outside the unit circle, then the polynomial can be factored as II(z)=U (z)M (z)V (z), where U(z) and V(z) are n x n matrix polynomials with all of their roots outside the unit circle, and M(z) is an n x n diagonal matrix polynomial with roots on or outside the unit circle. In the case o f the 1(1) cointegrated VAR, M(L) can be written as: where Ajc=(1-L)I j£ and k + r = n . This factorization is useful because it isolates all o f the VAR’s nonstationarities in the upper block of M(L). W e now derive alternative representations for the cointegrated system: 3 .b .l The Vector Error Correction VAR Model (VECM): To derive the VECM, subtract xt_j from both sides of (3.1) and rearrange the equation as: (3.2) Axt = I I x ^ + E P l j ^ A x ^ j + et where n = - I n + E f = 1n i = - n ( l) , and * i = -E P a s i+ 1IIj, i= * l,...,p -l. Since II(1)=U(1)M (1)V(1), and M (l) has rank r, n = - I I ( l ) also has rank r. Let a denote an n X r matrix whose columns form a basis for the row space o f II, so that every row of II can be written as a linear combination of the rows o f a ’. Thus, we can write n = $ a ’, where 5 is an n X r matrix with full column rank. Equation (3.2) then becomes: (3.3) Axt = S a’Xj.j + ^ P l ^ A x ^ + €t, -37- or (3.4) Axt = 5wt_j + + et, where wt = a ’xt. Solving (3.4) for wt_j shows that wt. j = (5 ’5)‘ 1 6’[Axt- j ^ A xj. j-Cj], so that wt is 1(0). Thus, the linear combinations o f the potentially 1(1) elements o f xt formed by the columns o f a are 1 (0 ), and the columns o f a are cointegrating vectors. The VECM imposes k < n unit roots in the VAR by including first differences o f all o f the variables and r= n -k linear combinations o f levels o f the variables. The levels o f xt are introduced in a special way — as wt = a ’xt — so that all o f the variables in the regression are 1(0). Equations o f this form appeared in Sargan (1964), and the term "error correction model" was introduced in Davidson, Hendry, Srba and Yeo (1 9 7 8 ).^ As explained there and in Hendry and von Ungem Sternberg (1981), a ’xt = 0 can be interpreted as the "equilibrium" o f the dynamical system, wt as the vector o f "equilibrium errors" and equation (3.4) describes the self correcting mechanism o f the system. 3 .b .2 The M o vin g A v e ra g e R ep resen ta tio n To derive the moving average representation for Axt , let so that M (L )M (L )= (l-L )In . Then: M(L)M(L)V(L)xt = M(L)U(L)_1€t, so that V(L)Axt = MfLlUCL)'1^ , -38- and (3.5) Axt = C(L)et where C (L )= V (L )' 1 M (L )U (L )'1. There are two special characteristics o f the moving average representation. First, C (1)=V (1)’ *M(1)U(1)'^ has rank k and is singular when k < n . This implies that the spectral density matrix of Axt evaluated a frequency zero, C(1)E€C(1)’, is singular in a cointegrated system. Second, there is a close relationship between C (l) and the matrix o f cointegrating vectors a . In particular, a ’C ( l ) = 0 . ^ Since wt = a ’xt is 1(0), Awt = a ’Axt is I(-l) so that it’s spectrum at frequency zero, a ’C (l)E eC (l)’a , vanishes. The equivalence o f vector error correction models and cointegrated variables with moving average representations o f the form (3.5) is provided in Granger (1983) and forms the basis of the Granger Representation Theorem (see Engle and Granger (1987)). 3 .b .3 The C om m on T ren ds R ep resen ta tio n : The common trends representation follows direcdy from (3.5). Adding and subtracting C (l)et from the right hand side o f (3.5) yields: (3.6) Axt = C (l)et + [C(L)-C(l)]et. Solving backwards for the level o f xt, (3.7) xt = C (l)£ t + C*(L)et + xq where = E ‘ = 1 es and C *(L)=(1-L )‘ 1[C (L )-C (1 )]= I^ _ 0C ’ , w h e r e c ’ = - I ^ = i+ 1 Cj and €j= 0 for i ^ 0 is assumed. Equation (3.7) is the multivariate Beveridge-Nelson (1981) -39- decomposition o f xt; it decomposes xt into its "permanent component," C ( l ) |t +XQ, and its "transitory component," C (L)et. Since C (l) has rank k, we can find a nonsingular matrix G, such that C (1)G =[A °n x rl> where A is an n x k matrix with full column rank. Thus C a ^ C W G G ' 1^ , so that (3.8) Xj = A rj + C (L)ej + Xq where r t denotes the first k components o f G '*£t. Equation (3.8) is the common trends representation o f the cointegrated system. It decomposes the n x 1 vector xt into k "permanent components" r t and n "transitory components" * C (L)et. These permanent components often have natural interpretations. For example, in the eight variable (y,c,i,n,w ,m ,p,r) system introduced in Section 2, five cointegrating vectors were suggested. In an eight variable system with five cointegrating vectors there are three common trends. In the (y,c,i,n,m ,p,r) systems these trends can be interpreted as population growth, technological progress, and trend growth in money. The common trends representation (3.8) is used in King, Plosser, Stock and Watson (1991) as a device to "extract" the single common trend in a three variable system consisting o f y, c, and i. The derivation o f (3.8) shows exactly how to do this: (i) estimate the VECM (3.3) imposing the cointegration restrictions; (ii) invert the VECM to find the moving average representation (3.5); (iii) find the matrix G introduced below equation (3.7); and finally (iv) construct recursively from *t= r t - l + e t» where et is the first element o f G'*et, and where et denotes the vector o f residuals from the VECM. Other interesting applications o f trend extraction in cointegrated systems are contained in Cochrane and Sbordone (1988) and Cochrane (1990). 3 . b . 4 The T ria n g u la r R e p re sen ta tio n The triangular representation also represents xt in terms o f a set o f k non-cointegrated 1(1) -40- variables. Rather than construct these stochastic trends as the latent variables r t in the common trends representation, a subset o f the xt variables are used. In particular, the triangular representation is of the form: (3.9) A x1>t = uu (3.10) x2 i,-0 x u - u2>t where xt = (x j t x£ t) \ x j t is k x 1 and X2 >t is r x 1. The transitory components are ut = (u i t U2 t) ’ =D (L )et, where (as we show below) D (l) has full rank. In this representation, the first k elements o f xt are the common trends and X2 t-0 x j t are the 1 (0 ) linear combinations o f the data. To derive this representation from the VAR (3.2), use II(L)=U (L)M (L)V (L) to write: (3.11) U(L)M(L)V(L)xt = €t, so that (3.12) M(L)V(L)xt = U O ^ e j . Now, partition V(L) as v u ( L) v 12<L> v2i(L) v 22(l>. V(L) where v j j(L) is k x k , v ^ f L ) is k x r , V2 j (L) is r x k and v 2 2 (L) is r x r . Assume that the data have been ordered so that v 2 2 (L) has all o f its roots outside the unit circle. (Since V(L) has all of its roots outside the unit circle, this assumption is made with no loss o f generality.) Now, let -41 - C(L) - Xk 1>3(L) 0 Ir. 1. w here/ 3 (L)=-V 2 2 (L) v 2 l(^>- ^ en (3.13) MCIOVtLJCCLJCCL)'^ = U(L)_1€t or, rearranging and simplifying 'vli(L)+vi2(L))9 (l-L)v12(L)' ’Axl,t (3.14) hJ N / CM CM > -v22(L)/J (L) .1 - U(L) ict -x2 , t ^ xl,t- where /3*(L)=(1-L)'^[/3(L)-/5(1)] and £ = /3 (l). Letting G(L) denote the matrix polynomial on the left hand side o f (3.14), the triangular representation is obtained by multiplying equation (3.14) by G (L)'*. Thus, in equations (3.9) and (3.10), ut =D (L)et, with D (L )= G (L )‘ *U(L)"*. When derived from the VAR (3.2), D(L) is seen to have a special structure that was inherited from the assumption that the data were generated by a finite order VAR. But o f course, there is nothing inherently special or natural about the finite order VAR; it is just one flexible parameterization for the xt process. When the triangular representation is used, an alternative approach is to parameterize the matrix polynomial D(L) directly. An early empirical study using this formulation is contained in Campbell and Shiller (1987). They estimate a bivariate model o f the term structure that includes long term and short term interest rates. Both interest rates are assumed to be 1(1), but the "spread" or difference between the variables is assumed to be 1(0). Thus, in terms o f (3.9)-(3.10), x ^ t is the short term interest rate, X2 t is the long rate, and /3=1. In their empirical work, Campbell and Shiller modeled the process ut in (3.10) as a finite order VAR. In empirical work, the triangular representation is no more or less convenient than the -42- VECM. However in theoretical econometric work concerned with estimating cointegrating vectors, the triangular representation is arguably more convenient than the VECM. The reason is that coefficients making up the the cointegrating vectors appear only in (3 . 1 0 ), and the system (3.9)-(3.10) "looks like" a standard triangular simultaneous equation system; estimators developed for that model, suitably modified, can be used to estimate the cointegrating vectors. Phillips (1991a) gave the triangular representation its name and demonstrated its usefulness for developing and analyzing the properties o f estimators of cointegrating vectors. ^ The representation has subsequently been used by many other researchers who have developed a large number of asymptotically efficient estimators. Regardless o f the representation used, model building for cointegrated systems involves two steps. In the first step, the degree of cointegration (or equivalently the number of unit roots in the model) is determined. In the second step, the unknown parameters o f the model are estimated. Statistical procedures for carrying out these steps are the subject o f the next two sections. 3.c Testing for Cointegration in 1(1) Systems It is convenient to cast our discussion in terms o f the VAR in equation (3.2). W e are interested in tests concerning r=rank(II) for this equation. The null and alternative hypotheses are: H0 : ra n k (II)= r= r 0 Ha : ra n k (II)= r= r 0 + r a where ra > 0 . The hypotheses are written so that ra denotes the additional cointegrating vectors that are present under the alternative. For example, when rQ= 0 , the null specifies that there are no cointegrating vectors, while the alternative implies that there are r a -43- > 0 cointegrating vectors. Specifying the null as "no cointegration" and the alternative as "cointegration" is natural, since when r = 0 , then 11=0 in equation (3.2), while when r £ 0 , then II =£0; the null and alternative are then Hq : n = 0 and Ha : n £ 0 (but restricted to have rank ra). As might be expected, the distribution o f test statistics for cointegration are complicated by the presence o f unit roots. Using the results developed in Section 2, these complications are easily overcome. To keep things as simple as possible, this section ignores constant terms and deterministic growth in the model. In terms o f the analysis in Section 2, this eliminates the canonical regressors corresponding to the constant (Z2 t) and the deterministic time trends (z4 t). Hypothesis testing when deterministic components are present is discussed in Section 3.e. There are a many tests for cointegration: some are based on likelihood methods, using a Gaussian likelihood and the VECM representation for the model, while others are based on more a d h o c methods. Section 3.c. 1 presents likelihood based (Wald and Likelihood Ratio) tests for cointegration constructed from the VECM. The non-likelihood based methods o f Engle and Granger (1987) and Stock and Watson (1988) are the subject o f Section 3.C.2, and the various tests are compared in Section 3.C.3. 3.C .1 L ik e lih o o d B a se d T ests f o r C o in teg ra tio n : 20 In Section 3 .b .l the general VECM was written as: (3.3) Axt = 5a’xt.j + i f ^ l ^ ^ t - i + «t. To develop the restrictions on the parameters in (3.3) implicit in the null hypothesis, first partition the matrix o f cointegrating vectors as a = [ a Q a a] where a Q is an n x r Q matrix whose columns are the cointegrating vectors present under the null and a a is the n x r a matrix o f additional cointegrating vectors present under the alternative. Partition 5 conformably as 5 = [5 0 6 a], let T = ( $ 1 #2 ••• $ p - i ) 211(1 let zt =(Ax^_i AxJ_ 2 ... AXt-p+lT* The VECM can -44- then be written as: (3.15) Axt = + 5aofaxt-l + Fzt + ft’ where, under the null hypothesis, the term 5aa£xt_j is absent. This suggests writing the null and alternative hypotheses as HQ: $a = 0 vs. Ha : 5a £ 0 . Written in this way, the null is seen as a linear restriction on the regression coefficients in (3.15). An important complication is that the regressor a axt_j depends on parameters in a a that are potentially unknown. Moreover, when 6 a = 0 , a axt. j does not enter the regression, and so the data provide no information about any unknown parameters in <*a . This means that these parameters are econometrically identified only under the alternative hypothesis, and this complicates the testing problem in ways discussed by Davies (1977)(1987), and (in the cointegration context) by Engle and Granger (1987). In many applications, this may not be a problem of practical consequence, since the coefficients in a are determined by the economic theory under consideration. For example in the (y,c,i,w ,n,r,m ,p) system, candidate error correction terms with no unknown parameters are y-c, y-i, (w-p)-(y-n) and r. Only one error correction term, m-p-/Syy-/S^R, contains potentially unknown parameters. Yet, when testing for cointegration, a researcher may not want to impose specific values o f potential cointegrating vectors, particularly during the preliminary data analytic stages o f the empirical investigation. For example, in their investigation o f long-run purchasing power parity, Johansen and Juselius (1992) suggest a twostep testing procedure. In the first step cointegration is tested without imposing any information about the cointegrating vector. If the null hypothesis o f no cointegration is rejected, a second stage test is conducted to see if the cointegating vector takes on the value predicted by economic theory. The advantage o f this two-step approach is that it can uncover cointegrating relations not predicted by the specific economic theory under study. The -45- disadvantage is that the first stage test for cointegration will have low power relative to a test that imposes the correct cointegrating vector. It is useful to have testing procedures that can be used when cointegrating vectors are known and when they are unknown. With these two possibilities in mind, we write r = r k + r u, where r^ denotes the number o f cointegrating vectors with known coefficients, and ru denotes the number of cointegrating vectors with unknown coefficients. Similarly, write r 0 = r 0 k+ r 0u and ra = rak+ rau» where the subscripts "k" and "u" denote known and unknown respectively. O f course, the r&k subset o f "known cointegrating vectors" are present only under the alternative, and a axt is 1 ( 1 ) under the null. Likelihood ratio tests for cointegration with unknown cointegrating vectors (i.e., HQ: r = r Qu vs. H a : r = r 0 u+ r au) are developed in Johansen (1988a), and these tests are modified to incorporate known cointegrating vectors (nonzero values o f r ^ and r ^ ) in Horvath and Watson (1993). The tests statistics and their asymptotic null distributions are developed below. For expositional purposes it is convenient to consider three special cases. In the first, ra = rak, so that all of the additional cointegrating vectors present under the alternative are assumed to be known. In the second, ra = r 3u, so that they are all unknown. The third case allows nonzero values of both r ^ and r ^ . To keep the notation simple, the tests are derived for the rQ= 0 null. In one sense, this is without loss o f generality, since the LR statistic for H 0 :r = r Q vs. Ha : r = r Q+ r a can be always be calculated as the difference between the LR statistics for [Hq : r = 0 vs. Ha : r = r Q+ r a] and [HQ: r = 0 vs. Ha : r = r Q]. However, the asymptotic null distribution o f the test statistic does depend on r ^ and r ^ , and this will be discussed at the end o f this section. Testing H Q:r = 0 vs. Ha : r = r ^ : When rQ= 0 , equation (3.15) simplifies to: (3.16) Axt = S^a^Xj.j) + Tzj + et. -46- Since a ^ t - l IS known» (3.16) is a multivariate linear regression, so that the LR, Wald and LM statistics have their standard regression form. Letting X = [x j X2 ... x ^ ]’, X_1 = [ x q Xj ... Xy_j]’, A X = X -X .j, Z = [ z j Z2 ... z ^ ]’, 62 ••• <uid Mz = tI-Z (Z ,Z)‘ 1 Z ’], the OLS estimator o f 5a is 5a =(A X ’Mz X . 1 a a)(a a ’X . 1 ’Mz X . 1 a a ) * 1 which is the Gaussian MLE. The corresponding Wald test statistic for HQ vs. Ha is: (3.17) W = [vec(3a)]’[(aa ‘X .j 'M z X . ^ j ) '1 ® E / W * , ) ] = [vecfAX'MzX.^jM’K a j'X.fMzX^aj)'1 ® E ^ K v ^ A X ’MzX.,^)]. A A where L £ is the usual estimator value o f E6 (Ef = T 1 c’«, where e is the matrix o f OLS residuals from (3.16)), "vec" is the operator that stacks the column o f a matrix, and the second line uses the result that vec(A B C )=(C ’ xA)vec(B) for conformable matrices A,B and C. The corresponding LR and LM statistics are asymptotically equivalent to W under the null and local alternatives. The asymptotic null distribution o f W is derived in Horvath and Watson (1993), where it is shown that (3.18) W = > Trace[( f B 1 (s)dB(s)’) ’( f B 1 (s)B 1 (s)’ds)_1( J B 1 (s)dB(s)’)], where B(s) is a n x 1 W iener process partitioned into ra and n-ra components B j(s) and B2 (s) respectively. A proof o f this result will not be offered here, but the form o f the limiting distribution can be understood by considering a special case with T = 0 (so that there are no lags o f Axt in the regression), Ef = In and « a = P rt 0]. In this case, xt is a random walk with NIID(0,In) innovations, and (3.16) is the regression o f Axt onto the first ra elements o f xt. j , say Xj t_j. Using the true value o f E€, the Wald statistic in (3.17) simplifies to: -47- W =» [v ec(£ A x ,x j t. , ) ] ’[ ( r x i . m M . m )" 1 ® W CvecCEAx^i j.j)). = Trace[(EAxtxiit.I)(Sxljt.1xiit.i)-1( I x lt.1Ax;)] = Trace[(T'1i;4 x txiit. 1)(T'2 E x lit. 1xijt. 1)"1(T‘ 1i;x lj(. 1Ax;)l = > Trace[( i B,(s)dB(s)T( J B,(s)B1(s)T 1( J B^sJdBfs)’)] where the second line uses the result that for square matrices, Trace(AB) = Trace(BA), and for conformable matrices Trace(A BCD )=(vec(D ))’(A x C ’)vec(B’) (Magnus and Neudecker (1988, pag 30)), and the last line follows from Lemma 2.c. This verifies (3.18) for the example. Testing HQ:r = 0 vs. H ^ r ^ r a ^ When a a is unknown, the Wald test in (3.17) cannot be calculated because the regressor <*a ’Xt_i depends on unknown parameters. However, the LR statistic can be calculated, and useful formulae for the LR statistic are developed in Anderson (1951) and Johansen (1988a). In the context o f the VECM (3.3), Johansen (1988a) shows that the LR statistics can be written as: (3.19) LR = - T E j ^ l n a - y j) , where 7 ^ are the ordered squared canonical correlations between Axt and xt. j , after controlling for Axt_j ,..., Axt_p + j . These canonical correlations can be calculated as the eigenvalues o f T - 1 S, where S = E ; 1^ (A X ,M ZX . 1 )(X : 1 MZX . 1 )*1 (X : 1 M ZAX)E;,/4’, and where Ee= T ‘ *(AX’M z AX) is the estimated covariance matrix o f et, computed under the null (see Anderson (1984, Chapter 12) or Brillinger (1980, Chapter 10)). Letting Xj(S) denote the eigenvalues o f S ordered as Xj (S )^ X 2 (S )£ :... ^ X ^ S ) , then 7 ^ from (3.19) is 7 j =T"^X j (S). Since elements o f S are O p(l) from Lemma 2.C., a Taylor series expansion o f ln (l-y p shows that LR statistic can be written as: -48- (3 .2 0 ) L R = + op( l) . Equation (3.20) shows why the LR statistic is sometimes called the "Maximal eigenvalue statistic" when 1^ = 1 and the "Trace-statistic" when ra^= n (Johansen and Juselius (1 9 9 0 )).^ One way to motivate the formula for the LR statistic given in (3.20), is by manipulating W (a). 23 To see the relationship between LR and W in this case, let L($a ,a a) denote the log likelihood written as a function of $a and a a , and let 5a(a a) denote the M LE o f $a for fixed a a. When is known, then the well known relation between the Wald and LR statistic in the linear regression model (Engle [1984]) implies that the Wald statistic can be written as: (3.21) W (aa) = 2[L(5a(a a) ,a a)-L (0,aa)] = 2[L(5a(a a) ,a a)-L(0,0)] where the last line follows since a & does not enter the likelihood when $a = 0 , and where W (aa) is written to show the dependence o f W on a a . From (3.21), with E€ known, (3.22) Supa t W (aa) = Supa t 2[L(5a(a a) ,a a)-L(0,0)] = 2[L(5a ,a a) - L(0,0)] = LR where the Sup is taken over all n x r a matrices a a . When E€ is unknown, this equivalence is asymptotic, i.e ., supa# W (aa) = LR + Op(l). To calculate Supa< W (aa), rewrite (3.17) as: (3.23) W (aa) = [vec(AX’MZX . { a a)]’[(a^X I { MZX .ja a ) ‘ 1 ® ^ ^ [ v e ^ A X ’M z X ^ o ^ ] -49- = TR[E;‘A (A X ’M z X_1ca) (^X.1' M z X.iaa)-1(a'X:1M z AX)E;'/l’J = T R I^'^C A X ’M z X .!) D D '(X : 1 M z AX)E;‘A ’], where D - « a(a ’X : 1 M z X . 1 a a)''A = T R P ’(X : 1 MZA X )E ' 1 (AX’MZX . 1 )D] = TR [F’C C ’F], where F = ( X : 1 M z X _ p ^ a a( a ;X ’_1 M z X_1 a a) ' V\ and C = p c : 1 M z X . 1 ) '^ ( X : 1 M z AX)E;V4’. Since F ’F = I fiu, (3.24) Supa# W («a) = SupF »F = I TR [F’(CC’)F] = E ^ X ^ C C ’) = = LR + o p(l) where Xj(CC’) denote the ordered eigenvalues o f (CC’), and the final two equalities follow from the standard principal components argument (for example, see Theil (1971,page 46)) and X^(CC’)= X j(C ’C). Equation (3.24) shows that the likelihood ratio statistic can then be calculated (up to an op(l) term) as the largest ra eigenvalues o f C ,C = E ' 1A(AX,M ZX_1 )(X : 1 MZX_1 ) ' 1 (X : 1 MZA X ) E ^ \ T o see the relationship between the formulae for the LR statistics in (3.24) and (3.20), notice that C ’C in (3.24) and S in (3.20) differ only in the estimator o f C ’C uses an estimator constructed from residuals calculated under the alternative, while S uses an estimator constructed from residuals calculated under the null. In general settings, it is not possible to derive a simple representation for the asymptotic distribution o f the Likelihood Ratio statistic when some parameters are present only under the alternative. However, the special structure o f the VECM makes such a simple representation possible. Johansen (1988a) shows that the LR statistic has the limiting null asymptotic distribution given by -50- (3 .2 5 ) L R = > E jt jX jd i) where H = [ \ B(s)dB(s)’] ’[ J B(s)B(s)’ds] f B(s)dB(s)’], and B(s) in an n x 1 W iener process. To understand Johansen’s result, again consider the special case with T = 0 and 1 ^ = ^ . In this case, C ’C becomes (3.26) C ’C = (AX’X . j K X I j X . p '^ ^ A X ) = [ £ A x jx j.jlt E xt- lxt- ll xt_i Ax{] = [T"1E AXjX !]’[T‘2 E *t.ix|.j]’1[T"1E xt.jAx J] = > [ S B(s)dB(s)’] ’[ f B(s)B(s)’ds]"1[ f B(s)dB(s)’] from Lemma 2.c. This verifies 3.25 for the example. Testing H p: r=Q vs. Ha: The model is now: (3.27) AX, = a ^ X M ) + a ^ X , . , ) + 0 Z , + <t. where <*a has been partitioned so that contains the r ^ known cointegrating vectors, contains the r ^ unknown cointegrating vectors, and Sa has been partitioned conformably as <5=(5afc 8 ^ ) . As above, the LR statistic can be approximated up to an Op(l) term by maximizing the Wald statistic over the unknown parameters in a ^ . Let Mzk= M z 'Mzx - l 0 fa)C(c^kX- l Mzx - l a ak^"^°^fcMzx - l Mz denote the matrix that partials both Z and out o f the regression (3.27). The Wald statistic (as a function o f a - ) can then be written as «u 24 : (3.28) W ^ . a ^ ) = [vK tA X ’M ^ a ^ n ^ ' X . j ' M ^ . i a and -5 1 - ^ ) '1 ® E‘ 1 ][vec(AX'M 2 X_1 a at)] + [vec(4X ,Ml k X . 1 o aii)],[(oiw,X .,'M 2 kX .I a a ii) - 1 ® The first term is identical to equation (3.17) above, and the second term has the same form except that both Z and X . j a ^ have been partialled out o f the regression. W e can derive the LR statistic as above with one modification: when maximizing W f a ^ a ^ ) over the unknown cointegrating vectors in o ^ , attention can be restricted to cointegrating vectors that are linearly independent o f a ^ . Thus, the LR statistic is obtained by maximizing (3.28) over all n x r ^ matrices satisfying 0 ^ 0 ^ = 0 . Let G denote an (arbitrary) n x (n -r^ ) matrix whose columns span the null space o f the columns o f a ^ . Then the columns of G, so that a ^ ^ G a ^ , where a zu' “ a t = can be written as a linear combination o f is an ( n - r ^ x r ^ matrix, so that “ a t = ^ ^or ^ “ a„- Substituting G a ^ into (3.28) and carrying out the maximization yields: (3.29) Supa _o W (aa i,aao) = [ v e c W X ' M . X . ^ ’K a ^ X . f ® E ;‘ ][vec(AX'MzX. + E jr “ i XjfH’H) = LR + op(l), where H 'H = E ;l' 4 (AX’MzkX . 1 G )(G ’X : 1 MzkX . 1 G )' 1 (G 'X : 1 M zkAX)E;1'4 : The statistic is calculated as follows. Regress AX onto a ^ X .j and Z and form the usual Wald statistic. This is the first term on the right hand side o f (3.29). Let G be an arbitrary matrix whose columns span the null space o f the columns o f a ^ . (The columns o f G can be constructed as the eigenvectors o f c r ^ a ^ corresponding to zero eigenvalues.) The second term on the right hand side o f (3.29) is the sum o f the r ^ largest eigenvalues o f E‘ 1/^(AX’MzjcX_1 G)(G’X ’ 1 MzkX_1 G)*1 (G’X_1 MzkAX)Eg,/^ \ Ee can be replaced by any consistent estimator o f Ee without affecting the large sample behavior o f the statistic. Two particularly simple estimators are E 6 = T '* A X ’MZAX and the residual covariance matrix from -52- the regression of Xt onto p lagged levels o f Xt. The asymptotic null distribution o f the LR statistic in (3.29) is derived in Horvath and Watson (1993). They show (3.30) LR = > Trace[( f B 1 (s)dB(s)’) ’( \ B j W B j f l ’ds)’ ^ J B 1 (s)dB(s)’)] + ^ ( C ’C) where B(s) is an (n x 1 ) W iener process partitioned into r ^ and n -r ^ components B j(s) and B2 (s) respectively, C ’C = [ J B2 (s)dB(s)’] ’[ J B2 (s)B2 (s)’]'* [ J B2 (s)dB(s)’], and B2 (s) is the residual from the regression of B2 (s) onto B j(s), i.e ., I ^ s^ I ^ sJ-t-B ^ s), where 7=[ J B ^B jfe)’]’ 1 J B2(s)B l ( s y ? 5 As pointed out above, when the null hypothesis is HQ:r = r 0 k+ r 0u, the LR test statistic can be calculated as the difference between the LR statistics for [HQ: r = 0 vs. Ha : r = r Q+ r a] and [HQ: r= 0 vs. Ha : r = r Q]. So for example when testing HQ: r = r Qu vs. Ha : r = r 0 u+ r a^, the LR statistic is: (3.31) LR = - T E j S ^ I n d - r , ) = + Op(l), where yj are the canonical correlations defined below equation (3.19) (see Anderson (1951) and Johansen (1988a)). Critical values for the case ^ = ^ = 0 and n - ^ ^ 5 are given in Johansen (1988a) for the trace-statistic (so that the alternative is rau= n-r0u); these are extended for nr ^ ^ 11 in Osterwald-Lenum (1992), who also tabulates asymptotic critical values for the maximal eigenvalue statistic (so that ^ = ^ = 0 and ^ = 1 ) . Finally, asymptotic critical values for all combinations o f r ^ , r ^ , r ^ and r ^ with n - r ^ ^ 9 are tabulated in Horvath and Watson (1992). 3 .C .3 N on L ik e lih o o d b a s e d a p p ro a c h e s: In addition to the likelihood based tests discussed in the last section, standard univariate unit -53- root tests and their multivariate generalizations have also been used as tests for cointegration. To see why these tests are useful, consider the hypotheses H0 : r = 0 vs. Ha : r = l , and suppose that a is known under the alternative. Since the data are not cointegrated under the null, wt = a ’xt is 1(1), while under the alternative it is 1(0). Thus, cointegration can be tested by applying a standard unit root test to the univariate series wt. To be useful in more general cointegrated models, standard unit root tests have been modified in two ways. First, modifications have been proposed so that the tests can be applied when a is unknown. Second multivariate unit root tests have been developed for the general testing problem HQ: r = r Q vs. Ha : r = r Q+ r a . We discuss these two modifications in turn. Engle and Granger (1987) develop a test for the hypotheses HQ: r = 0 vs. Ha : r = l when a is unknown. They suggest using OLS to estimate the single cointegrating vector and applying a standard unit root test (they suggest an augmented Dickey-Fuller t-test) to the OLS residuals, A A A A wt = a x t. Under the alternative, a is a consistent estimator of a , so that wt will behave like A wt. However, under the null, a is obtained from a "spurious" regression (see Section 2.f.3) A and the residuals from a spurious regression (wt) behave differently than non-stochastic linear combinations o f 1(1) variables (wt). This affects the null distribution o f unit root statistics A A calculated using wt. For example, the Dickey-Fuller t-statistic constructed using wt has a different distribution than the statistic calculated using wt, so that the usual critical values given in Fuller (1976) cannot be used for the Engle-Granger test. The correct asymptotic null distribution o f the statistic is derived in Phillips and Ouliaris (1990), and is tabulated in Engle and Yoo (1987) and MacKinnon (1991). Hansen (1990a) proposes a modification o f the EngleGranger test that is based on an iterated Cochrane-Orcutt estimator which eliminates the "spurious regression" problem and results in test statistics with standard Dickey-Fuller asymptotic distributions under the null. Stock and Watson (1988), building on work by Founds and Dickey (1986), propose a multivariate unit root test. Their procedure is most easily described by considering the V A R(l) -54- model, xt = $ x t_j + e t, together with the hypotheses HQ: r = 0 vs. Ha : r = r a . Under the null the data are not cointegrated, so that $ = I n . Under the alternative there are ra covariance stationary linear combinations o f the data, so that $ has ra eigenvalues that are less one in modulus. The A Stock-Watson test is based on the ordered eigenvalues o f $>, the OLS estimator of $ . Writing these eigenvalues as | Xj | ^ [ X2 1 ^ ..., the test is based on eigenvalue. Under the null, the ra ’th smallest = 1, while under the alternative, | | < 1 . The asymptotic null distribution o f T (#-I) and T( | Xj | -1) are derived in Stock and Watson (1988), and critical values for T( | Xn.ra. j | -1) are tabulated. This paper also develops the required modifications for testing in a general VAR(p) model with rQ^=0. 3 .d .4 C o m p a riso n o f th e T ests The tests discussed above differ from one another in two important respects. First, some of the tests are constructed using the true value of the cointegrating vectors under the alternative, while others estimate the cointegrating vectors. Second, the likelihood based tests focus their attention on 5 in (3.3), while the non-likelihood based tests focus on the serial correlation properties of certain linear combinations of the data. O f course, knowledge o f the cointegrating vectors, if available, will increase the power o f the tests. The relative power o f tests that focus on 6 compares and tests that focus on the serial correlation properties o f wt = a ’xt is less clear. Some insight into this is obtained by considering a special case of the VECM (3.3): (3.32) Axt = 5a( a ’xM ) + «t Suppose that a a is known and that the competing hypotheses are H0 : r = 0 vs. Ha : r = l . Multiplying both sides o f (3.31) by o£ yields: ( 3 .3 3 ) A w t = 0 w t_i + et -55- where w t = a£ x t, 0=o£5a and et =o^et. Unit root tests constructed from wt test the hypotheses HQ: 0 = (a a$a)=O vs. H&: 0=(o£Sa)< O , while the VECM-based LR and Wald statistics test HQ: 5a = 0 vs. Ha : 5a £ 0 . Thus, unit root tests constructed from wt focus on departures from the 5a = 0 null in the direction o f the cointegrating vector a a . In contrast, the VECM likelihood based tests are invariant to transformations o f the form P a ^ . j when a a is known and Pxt. j when a a is unknown, where P is an arbitrary non-singular matrix. Thus, the likelihood based tests aren’t focused in a single direction like the univariate unit root test. This suggests that tests based on wt should perform relatively well for departures in the direction o f a , but relatively poorly in other directions. As an extreme case, when a£5a = 0 , the elements o f xt are 1(2) and wt is 1(1). (The system is 0 ( 2 ,1 ) in Engle and Granger’s (1987) notation.) The elements are still cointegrated, at least in the sense that a particular linear combination o f the variables is less persistent than the individual elements o f xt, and this form o f cointegration can be detected by a non-zero value o f 5 in equation (3.32) even though 0 = 0 in (3.33). A systematic comparison of the power properties o f the various tests will not be carried out here, but one simple Monte-Carlo experiment, taken from a set o f experiments in Horvath and Watson (1993), highlights the power tradeoffs. Consider a bivariate model o f the form given in (3.32) with et-NIID( 0 ,l 2 ), a a = ( l -1)’ and $a = (5ai $a2) ’. This design implies that 0 = 5 ai-Sa2in (3.33) , so that the unit root tests should perform reasonably well when 15a(-5a 2 1 is large and reasonably poorly when | $aj-5a21 is small. Changes in 5a have two effects on the power o f the VECM likelihood based tests. In the classical multivariate regression model, the power o f the likelihood based tests increase with f = 5 a ( + 5 a2< However, in the VECM, changes in 5aj and 5a 2 also affect the serial correlation properties o f the regressor, wt = a ’xt_j, as well as f. Indeed, for this design, wt-A R (l) with AR coefficient 0 = 5 a - 8 a (see equation (3.33)). Increases in 0 lead to increases in the variability o f the regressor and increases in the power o f the test. Table 1 shows size and power for four different values o f 0a when T = 100 in this bivariate -56- system. Four tests are considered: (1) the Dickey-Fuller (DF) t-test using the true value o f a ; (2) the Engle-Granger test (EG-DF, the Dickey-Fuller t-test using a value o f a estimated by OLS); (3) the Wald statistic for HQ: $a = 0 using the true value o f o ; and (4) the LR statistic for Hq : 5 = 0 for unknown a . The table contains several noteworthy results. First, for this simple design, the size of the tests is close to the size predicted by asymptotic theory. Second, as expected, the DF and EGDF tests perform quite poorly when | $ai*$a 2 1 is small. Third, increasing the serial correlation 2 2 in wt =o;axt, while holding 5 ^ + 5 ^ constant, increases the power o f the likelihood based tests. (This can be seen by comparing the 5a = (.05,.055) and 5a = (-.05,.055) columns.) Fourth, increasing $a j + 5 a2, while holding the serial correlation in wt constant, increases the power of the likelihood based tests. (This can be seen by comparing the 5a = (-.05,.055) and 5a = (.105,.00) columns.) Fifth, when the DF and EG-DF are focused on the correct direction, their power exceeds the likelihood based tests. (This can be seen from the 5a = (-.05,.055) column.). Finally, there is gain in power from incorporating the true value o f the cointegrating vector. (This can be seen by comparing the DF test to the EG-DF test and the Wald test to the LR test.) A more thorough comparison o f the tests is contained in Horvath and Watson (1993). 3.d Estimating Cointegrating Vectors 3 . d . l G a u ssian M axim um L ik e lih o o d E stim ation b a s e d on th e T ria n g u la r R ep resen ta tio n In Section 3.b.4 the triangular representation of the cointegrated system was written as: (3.9) A xl t = uM (3.10) x2 , f ^ l , t = “2,t where ut =D (L)et. In this section we discuss the MLE estimator o f j3 under the assumption that et-NIID(0,I). The NIID assumption is used to only motivate the Gaussian MLE. The -57- asymptotic distribution o f estimators that are derived below follow under the weaker distributional assumptions listed in Lemma 2.c. In Section 2 .f.4 we considered the OLS estimator o f 0 in a bivariate model, and paid particular attention to the distribution o f the estimator when d i 2 = ^21 = 0* *n dus case» x l , t 1S wea^ y exogenous for 0 and the MLE estimator corresponds to the OLS estimator. Recall (see Section 2.f.4) that when = ®» OLS estimator o f 0 has an asymptotic distribution that can be represented as a variance mixture o f normals and that the t-statistic for 0 has an asymptotic null distribution that is standard normal. This means that tests concerning the value o f 0 and confidence intervals for 0 can be constructed in the usual way; complications from the unit roots in the system can be ignored. These results carry over immediately to the vector case where Xj t is k x 1 and X2 t is r x 1 when D i2 = 0 k x r ^ d d 21 = ®rxk* Somewhat surprisingly, they also carry over to the M LE o f 0 in the general model with ut =D (L)et, with D ^C L) and D 2 j(L) nonzero, so that the errors are both serially and cross correlated. Intuition for this result can be developed by considering the static model with ut =D et and D 2 1 and D 2 1 nonzero. Since Uj t and U2 t are correlated, the MLE o f 0 corresponds to the SUR estimator from (3.9)-(3.10). But, since there are no unknown regression coefficients in (3.9), the SUR estimator can be calculated by OLS in the regression: (3.34) where x2>t = 0 x lft + 7 7 Ax1>t + e 2 >t is the coefficient from the regression o f U2 t onto Uj t, and e 2 t =U 2 t_Et u 2 1 1 U1 tl *s residual from this regression. By construction, e 2 t is independent o f {xj T}^._ j for all t. M oreover, since 7 is a coefficient on a zero mean stationary regressor and 0 is a coefficient on a martingale, the limiting scaled "X ’X" matrix from the regression is block diagonal (Section 2.e. 1). Thus from Lemma 2.c, -58- (3.35) T (j}-» - (T - 1 r e 2 .tx I ,t')(T ’2 E x 1 ,tx l , t ' ) ‘ 1 + V 1) = > ( E „ , i B jfs X U ty s J T ^ K E * j B 1 (s)B 1 (s)-dsE ^ where EUi= v ar(u j t), E g ^ v a r ^ t) ^ ) -1 B(s) is an n x l Brownian motion process, partitioned as B (s)= (B j(s)’ B2 ( s) ’) ’, where B j(s) is k x 1 and B2 (s) is r x 1 . Except for the change in scale factors and dimensions, equation (3.35) has the same form as (2.20), the asymptotic distribution A. A of 0 in the case d 1 2 = d 2 1 = 0- Thus, the asymptotic distribution o f 0 can be represented as a variance mixture of normals. Moreover, the same conditioning argument used when d i 2 = d 2 j implies that the asymptotic distribution of Wald test statistics concerning 0 have their usual large-sample x distribution. Thus, inference about 0 can be carried out using standard procedures and standard distributions. Now suppose that ut =D (L)et. The dynamic analogue o f (3.34) is x2>t = 0 x 1>t + 7 (L)Ax1>t + e ^ H H (3.36) where 7 (L)Axl t = E [u2>t| {Ax1>T} ; = 1 ]= E [u 2>t | {u1>T}{. = 1], and e 2 >t= u 2 >fE[u2>t| {ul r } From classical projection formulae (e.g. Whittle (1983), chapter 5), Y(L) = D 2 j (L)[D j (L)D j (L"V]"^> where D j(L ) denotes the first k rows o f D ( L ) .^ Equation (3.36) differs from (3.34) in two ways. First, there is potential serial correlation in the error term of (3.36), and second, 7 (L) in (3.36) is a two-sided polynomial. These differences complicate the estimation process. To focus on the first o f complication, assume that 7 (L )= 0 . In this case, (3.36) is a regression model with a serially correlated error, so that (asymptotically) the M LE o f 0 is just the feasible GLS estimator in (3.36). But, as shown in Phillips and Park (1988), the GLS correction has no effect on the asymptotic distribution o f the estimator: the OLS estimator and GLS estimators of 0 in (3.17) are asymptotically equivalent. -59- 28 Since the regression error e 2 t and the regressors {xl r}^ = 1 are independent, by analogy with the serially uncorrelated case, A T03-j8) will have an asymptotic distribution that can be represented as a variance mixture o f normals. Indeed, the distribution will be exactly o f the form (3.35), where now EUi and represent "long-run" covariance matrices. 29 Using conditioning arguments like those used in Section 2 .f.4, it is straightforward to show that the Wald test statistics constructed from the GLS estimators o f j8 have large sample distributions. However, since the errors in (3.36) are serially correlated, the usual estimator of the covariance matrix for the OLS estimators o f /? is inappropriate, and a serial correlation robust covariance matrix is required. 30 Wald test statistics constructed from OLS estimators o f £ together with serial correlation robust estimators o f covariance matrices will be asymptotically 2 X and will be asymptotically equivalent to the statistics calculated using the GLS estimators o f /3 (Phillips and Park (1988)). In summary, the serial correlation in (3.36) poses no major obstacles. The two-sided polynomial >(L) poses more of a problem and three different approaches have developed. In the first approach, polynomial. 31 7 (L) is approximated by a finite order (two-sided) In this case equation (3.36) can be estimated by GLS, yielding what Stock and Watson (1993) call the "Dynamic GLS" estimator o f 0. Alternatively, utilizing the Phillips and Park (1988) result, an asymptotically equivalent "Dynamic OLS" estimator can be constructed by applying OLS to (3.36). To motivate the second approach, assume for a moment that 7 (L) were known. The OLS estimator o f j3 would then be formed by regressing x2 t*7 (L)Axj t onto X j t. But, since T ’ 1 E [7 (L)Axj t] x \ t = T ‘ * £ [7 ( 1 )Ax j t]x[ t +O p(l) (by (c) o f Lemma 2.c), an asymptotically equivalent estimator can be constructed by the regression o f X2 t*7 (l)A x j l onto x ^ t. Estimators o f this form are proposed in Park (1993) and Phillips and Hansen (1990), where 7 ( 1 ) is replaced with a consistent estimator. The final approach is motivated by the observation that the low frequency movements in the -60- data asymptotically dominate the estimator of /3. Phillips (1991b) demonstrates that an efficient band spectrum regression, concentrating on frequency zero, can be used to calculate an estimator asymptotically equivalent to the MLE estimator in ( 3 .1 7 ) .^ All of these suggestions lead to asymptotically equivalent estimators. The estimators have asymptotic representations o f the form (3.35) (where EUj and Ee2 represent long-run covariance matrices), and thus their distributions can be represented as variance mixtures o f normals. Wald test statistics computed using the estimators (and serial correlation robust matrices) have the usual large sample x distributions under the null. 3 .d .2 G a u ssian M axim um L ik e lih o o d E stim ation b a se d on th e V E C M Most of the empirical work with cointegrated systems has utilized parameterizations based on the finite order VECM representation shown in equation (3.3). Exact MLEs calculated from the finite order VECM representation o f the model are different from the exact MLEs calculated from the triangular representations that were developed in the last section. The reason is that the VECM imposes constraints on the coefficients in ?(L) and the serial correlation properties o f e2 t in (3.17). These restrictions were not exploited in the estimators discussed in Section 3.d. 1. While these restrictions are asymptotically uninformative about 0, they nevertheless impact the estimator in small samples. Gaussian M LE’s o f jS constructed from the finite order VECM (3.2) are analyzed in Johansen (1988a)(1991) and Ahn and Reinsel (1990) using the reduced rank regression framework originally studied by Anderson (1951). Both papers discuss computational approaches for computing the M LE’s, and more importantly, derive the asymptotic distribution o f the Gaussian MLE. There are two minor differences between the Johansen (1988a)(1991) and Ahn and Reinsel (1990) approaches. First, different normalizations are employed. Since n=Sa’ = $ F F ’ ^a for any nonsingular r x r matrix F, the parameters in 5 and a are not econometrically identified without further restrictions. Ahn and Reinsel (1990) use the same -61 - identifying restriction imposed in the triangular model, i.e ., <*’ = [-£ If]; Johansen (1991) uses the normalization a ’R a = I r , where R is the sample moment matrix o f residuals from the regression o f xt_j onto Axt_j, i = l , p-1. Both sets o f restrictions are normalizations in the sense that they "just" identify the model, and lead to identical values o f the maximized likelihood. Partitioning Johansen’s M LE as a= (& j* &2 ’) ’* where a j is k x r and A <*2 is r x r , 1 then the MLE of /3 using Ahn and Reinsel’s normalization is /S= -(a jc *2 ) ’. The approaches also differ in the computational algorithm used to maximize the likelihood function. Johansen (1988a), following Anderson (1951), suggests an algorithm based on partial canonical correlation analysis between Axt and xt. j given Axt_j, i= l ,. . . , p - l . This framework is useful because likelihood ratio tests for cointegration are computed as a byproduct (see equation 3.19). Ahn and Reinsel (1990) suggests an algorithm based on iterative least squares calculations. M odem computers quickly find the M LE’s for typical economic systems using either algorithm. Some key results derived in both Johansen (1988a) and Ahn and Reinsel (1990) are transparent from the latter’s regression formulae. As in Section 3.c, write the VECM as (3.37) Axt = 6a ’xt_j + Tzt + et = a[x2>t.i-f?*iit.i] + rz, + «t where zt includes the relevant lags o f Ax^ and the second line imposes the Ahn-Reinsel normalization o f a . Let w ^.j=X 2 t-l'/3x l denote the error correction term, and let 0=[vec(5)’ vec(T)’ vec(/3)’] ’ denote the vector o f unknown parameters. Using the well known relations between the vec operator and Kronecker products, vec(Tzt) = (z|® In)vec(T), vec(5wt. 1 ) = ( w |. 1 0 I n)vec(T) and vec(6/3x1>t. 1)= (x J t. 1 ®5)vec(/3). Using these expressions, and defining Qt = [(z t ® In) (wt_ j 0 I n) ( x i>t_ i® $ ) ]\ then the Gauss-Newton iterations for estimating 6 are: -62- 0.38) Si+1 = «* + (EQ;s;1Q t]'1[i;Q|E;1«t] where <r denotes the estimator o f 6 at the i’th iteration, E€= T evaluated at (r. 33 £ ete|, and Qt and ct are Thus, the Gauss-Newton regression corresponds to the GLS regression o f et onto (z’® In), (w [.j® In), and (xj ® 5). Since the zt and wt are 1(0) with zero mean and x j t is 1(1), the analysis is Section 4 suggests that the limiting regression "X’X" matrix will be block diagonal, and the M LE’s o f 5 and T will be asymptotically independent o f the MLE o f /3. Johansen (1988a) and Ahn and Reinsel (1990) show that this is indeed the case. In addition they demonstrate that the MLE of /S has a limiting distribution o f the same form as shown in A, equation (3.35) above, so that T(/3-/3) can be represented as variance mixture o f normals. Finally, paralleling the result for M LE’s from triangular representation, Johansen (1988a) and Ahn and Reinsel (1990) demonstrate that - N(0,I) [ Z Q f e l Q tf \ o - 8 ) so that hypothesis tests and confidence intervals for all o f the parameters in the VECM can be 2 constructed using the Normal and x distributions. 3 .d .3 C o m p a riso n a n d E fficien cy o f th e E stim a to rs The estimated cointegrating vectors constructed from the VECM (3.3) or the triangular representation (3.9)-(3.10) differ only in the way that the 1(0) dynamics o f the system are parameterized. The VECM models these dynamics using a VAR involving the first differences Axj t, AX2 t and the error correction terms, X2 t-/3xj t; the triangular representation uses only Axi t and the error correction terms. Section 3 .d .l showed that the exact parameterization of the 1(0) dynamics - 7 (L) and the serial correlation o f the error term in (3.36) — mattered -63- little for the asymptotic behavior o f the estimator from the triangular representation. In particular, estimators o f 0 that ignore residual serial correlation and replace y(L) with 7 ( 1 ) are asymptotically equivalent to the exact MLE in (3.36). Saikkonen (1991) shows that this asymptotic equivalence extends to Gaussian M LE’s constructed from the VECM. Estimators o f /3 constructed from (3.36) with appropriate nonparametric estimators o f 7 ( 1 ) are asymptotically equivalent to Gaussian M LE’s constructed from the VECM (3.3). Similarly, tests statistics for H 0 :R[vec(/3)]=r constructed from estimators based on the triangular representation and the VECM are also asymptotically equivalent. Since estimators o f cointegrating vectors do not have asymptotic normal distributions, the standard analysis o f asymptotic efficiency -- based on comparing estimator’s asymptotic covariance matrices -- cannot be used. Phillips (1991a) and Saikkonen (1991) discuss efficiency o f cointegrating vectors using generalizations o f the standard efficiency definitions.”^ Loosely speaking, these generalizations compare two estimators in terms of the relative probability that the estimators are contained in certain convex regions that are symmetric about the true value of the parameter vector. Phillips (1991a) shows that when ut in the triangular representation (3.9)-(3.10) is generated by a Gaussian ARMA process, then the M LE is asymptotically efficient. Saikkonen (1991) considers estimators whose asymptotic distributions can be represented by a certain class o f functionals o f Brownian motion. This class contains the OLS and NLLS estimators analyzed in Stock (1987), the instrumental variable estimators analyzed in Phillips and Hansen (1990), all o f the estimators discussed in Section 3 .d .l and 3 .d .2 ., and virtually every other estimator that has been suggested. Saikkonen shows that the Gaussian MLE or (any o f the estimators that are asymptotically equivalent to the Gaussian MLE) are asymptotically efficient members o f this class. Several studies have used Monte Carlo methods to examine the small sample behavior o f the various estimators o f cointegrating vectors. A partial list o f the Monte Carlo studies is Ahn and Reinsel (1990), Baneijee, Dolado, Hendry and Smith (1986), Gonzalo (1989), Park and Ogaki -64- (1991), Phillips and Hansen (1990), Phillips and Loretan, and Stock and Watson (1993). A survey o f these studies suggests three general conclusions. First, the static OLS estimator can be very badly biased even when the sample size is reasonably large. This finding is consistent with the bias in the asymptotic distribution o f the OLS estimator (see equation 2.22) that was noted by Stock (1987). The second general conclusion concerns the small sample behavior o f the Gaussian MLE based on the finite order VECM. The Monte Carlo studies discovered that, when the sample size is small, the estimator has a very large mean squared error, caused by a few extreme outliers. Gaussian MLEs based on the triangular representation do not share this characteristic. Some insight into this phenomenon is provided in Phillips (1991c) which derives the exact (small sample) distribution o f the estimators in a model in which the variables follow independent Gaussian random walks. The MLE constructed from the VECM is shown to have a Cauchy distribution and so has no integer moments, while the estimator based on the triangular representation has integer moments up to order T -n + r. While Phillips’ results concern a model in which the variables are not cointegrated, it is useful because it suggests that when the data are "weakly" cointegrated - as might be the case in small samples -- the estimated cointegrating vector will (approximately) have these characteristics. The third general conclusion concerns the approximate Gaussian MLEs based on the triangular representation. The small sample properties o f these estimators and test statistics depend in an important way on the estimator used for long-run covariance matrix o f the data (spectrum at frequency zero), which is used to construct an estimator o f 7 ( 1 ) and the long-run residual variance in (3.36). Experiments in Park and Ogaki (1991), Stock and Watson (1993), and (in a different context) Andrews and Moynihan (1990), suggests that autoregressive estimators or estimators that rely on autoregressive pre-whitening outperform estimators based on simple weighted averages of sample covariances. -65- 3.e The Role of Constants and Trends 3 . e . l The M o d e l o f D e te rm in istic C om pon en ts Thus far, deterministic components in the time series (constants and trends) have been ignored. These components are important for three reasons. First, they represent the average growth or non-zero level present in virtually all economic time series; second, they affect the efficiency o f estimated cointegrating vectors and power of tests for cointegration; finally, they affect the distribution o f estimated cointegrating vectors and cointegration test statistics. Accordingly, suppose that the observed n x 1 time series yt can be represented as: (3.39) yt = /x0 + H \t + xt where xt is generated by the VAR (3.1). In (3.39), HQ+ H \ t represents the deterministic component o f yt, and xt represents the stochastic component. In this section we discuss how the deterministic components affect the estimation and testing procedures that we have already surveyed. There is a one simple way to modify the procedures so that they can applied to yt. The deterministic components can be eliminated by regressing yt onto a constant and time trend. Letting y [ denote the detrended series, the estimation and testing procedures developed above can then be used by replacing xt with y [. This changes the asymptotic distribution o f the statistics in a simple way: since the detrended values o f yt and xt are identical, all statistics have the same limiting representation with the Brownian motion process B(s) replaced by BT(s), the detrended Brownian motion introduced in Section 2.c. W hile this approach is simple, it is often statistically inefficient because it discards all o f the deterministic trend information in the data, and the relationship between these trends is often the most useful information about cointegration. To see this, let a denote a cointegrating vector and consider the "stable" linear combination: -66- (3.40) where a ’yt = Xq + X jt + wt Xq = ck> q , Xj = a > j , and wt = a ’xt. In most (if not all) applications, the cointegrating vector will annihilate both the stochastic trend and deterministic trend in a ’yt. That is, wt will be 1(0) or and Xi = 0 . As shown below, this means that one linear combination o f the coefficients in the cointegrating vector can be consistently estimated at rate T ^ . In contrast, when detrended data are used, the cointegrating vectors are consistently estimated at rate T. Thus, the data’s deterministic trends are the dominant source of information about the cointegrating vector, and detrending the data throws this information away. The remainder of this section discusses estimation and testing procedures that utilize the data’s deterministic trends. Most of these procedures are simple modifications o f the procedures that we developed above. 3 . e . 2 E stim a tin g C o in teg ra tin g V ectors We begin with a discussion of MLE o f cointegrating vectors based on the triangular representation. Partitioning yt into (n-r) x 1 and r x 1 components, y j t and y2 t, the triangular representation for y t is: (3.41) Ay1>t = 7 + u 1>t ( 3 . 4 2 ) y 2 ft-0 y i,t = Xq + X it + u2>t. This is identical to the triangular representation for xt given in (3.9)-(3.10) except for the constant and trend terms. The constant term in (3.41) represents the average growth in y j t. In most situations Xj = 0 in (3.42) since the cointegrating vector annihilates the deterministic trend in the variables. In this case, Xq denotes the mean o f the error correction terms, which is -67- unrestricted in most economic applications. Assuming that Xj = 0 and Xq and 7 are unrestricted, efficient estimation o f 0 in (3.42) proceeds as in Section 3 .c .l. The only difference is that the equations now include a constant term. As in Section 3 .c .l, Wald, LR or LM test statistics for testing HQ: R[vec(/3)]=r will have limiting x distributions, and confidence intervals for the elements o f 0 can be constructed in the usual way. The only result from Section 3.c. 1 that needs to be modified is the asymptotic A. distribution o f 0 . This estimator is calculated from the regression o f y 2 t onto y j t, leads and lags o f Ayj t and a constant term. When the y ^ t data contain a trend (7 ^ 0 in (3.41)), one of the canonical regressors is a time trend (z^ t from Section 2 .e .l), and the estimated coefficient on the time trend converges at rate T ^ . This means that one linear combination o f the estimated coefficients in the cointegrating vector converges to its true value very quickly; when the model did not contain a linear trend the estimator converged at rate T. The results for MLEs based on the finite order VECM representation are analogous to those from the triangular representation. The VECM representation for yt is derived directly from (3.2) and (3.39): (3.43) Ayt = m + ^ (a ’X t.p + E f l ^ A x ^ + €t = Ml + 5(a’yt_r Xo-Xit) + E where /xj = (I- 1 Ayt_i + et Xq = c* > q and Xj = a V j . Again, in most applications Xj = 0 , and the VECM is (3.44) Ayt = 0 + 6 ( a ’yM ) + E j ^ V i ^ t - i + €t where 6 = j i y S \ Q . When the only restriction on is a ’/ti = 0 , the constant term 6 is unconstrained, and (3.44) has the same form as (3.2) except that a constant term has been -68- added. Thus, the Gaussian M LE from (3.44) is constructed exactly as in Section 3.C.2 with the addition o f a constant term in all equations. The distribution o f test statistics is unaffected, but . for the reasons discussed above, the asymptotic distribution o f the cointegrating vector changes because of the presence of the deterministic trend. In some situations the data are not trending in a deterministic fashion, so that p j = 0 . (For example, this is arguably the case when yt is a vector o f U .S. interest rates.) When p^ = 0 , then p i =0 in (3.43), and this imposes a constraint on 6 in (3.44). To impose this constraint, the model can be written as (3.45) Ay, = + «, and estimated using a modification o f the Gauss-Newton iterations in (3.38), or a modification o f Johansen’s canonical correlation approach (see Johansen and Juselius (1990)). 3 . e . 3 T estin g f o r C o in teg ra tio n Deterministic trends have important effects on tests for cointegration. As discussed in Johansen and Juselius (1990), Johansen (1991)(1992a), it is useful to consider two separate effects. First, as in (3.43)-(3.44) nonzero values of p q and p^ affect the form o f the VECM, and this in turn affects the form of the cointegration test statistic. Second, the deterministic components affect the properties o f the regressors, and this in turn affects the distribution o f cointegration test statistics. In the most general form o f the test considered in Section 3.C.1, a was partitioned into known and unknown cointegrating vectors under both the null and alternative; that is, a was written as a = ( a 0k a ^ ) . When non-zero values of p q and /*j are allowed, the precise form o f the statistic and resulting asymptotic null distribution depends on which o f these cointegrating vectors annihilate the trend or constant (see Horvath and Watson (1993)). Rather than catalogue all o f the possible cases, the major statistical issues will be -69- discussed in the context of two examples. The reader is referred to Johansen and Juselius (1990), Johansen (1992a) and Horvath and Watson (1993) for a more systematic treatment. In the first example, suppose that r= 0 under the null, that a is known under the alternative, that fiQ and m are nonzero, but that a ’/tj =0 is known. To be concrete, suppose that the data are aggregate time series on the logarithms of income, consumption and investment for the United States. The balanced growth hypothesis suggests two possible cointegrating relations with cointegrating vectors (1, -1, 0) and (1, 0, -1). The series exhibit deterministic growth, so that /ij=£0, and the sample growth rates are approximately equal, so that a ’/ij =0 is reasonable. In this example, (3.44) is the correct specification of the VECM with 6 unrestricted under both the null and alternative and 5=0 under the null. Comparing (3.44) and the specification with no deterministic components given in (3.3), the only difference is that xt in (3.3) becomes yt in (3.44) and the constant term 6 is added. Thus, the Wald test for HQ: 5=0 is constructed as in (3.17) with yt replacing xt and Z augmented by a column of l ’s. Since a > j =0, the regressor a ’yt_i = a ’xt_j + a ’/iQ, but since a constant is included in the regression, all of variables are deviated from their sample means. Since the demeaned values of a ’yt.j and a ’xt_j are the same, the asymptotic null distribution of the Wald statistic for testing HQ:5=0 in (3.44) is is given by (3.18) with /3^(s), the demeaned Wiener process defined below Lemma 2.c, replacing B(s). The second example is the same as the first, except that now a is unknown. Equation (3.44) is still the correct VECM with 6 unrestricted under the null and alternative. The LR test statistic is calculated as in (3.19), again with yt replacing x( and Z augmented by a vector of l ’s. Now however, the distribution of the test statistic changes in an important way. Since the regressor yt_j contains a nonzero trend, it behaves like a combination of time trend and martingale components. When the n x 1 vector yt. j is transformed into the canonical regressors of Section 2, this yields one regressor dominated by a time trend and n-1 regressors dominated by martingales. As shown in Johansen and Juselius (1990), the distribution of the resulting LR -70- statistic has a null distribution given by (3.25) where now H = [ f F(s)dB(s)’]’[ J F(s)F(s)’ds]‘*[ J F(s)dB(s)’], where F(s) is an n x l vector, with first n-1 elements given by the first n-1 elements of /3^(s) and the last element given by the demeaned time trend, s-‘A. (The components are demeaned because of the constant term in the regression.) Johansen and Juselius (1990) also derive the asymptotic null distribution of the LR test for cointegration with unknown cointegrating vectors when =0, so that (3.45) is the appropriate specification of the VECM. Tables of critical values are presented in Johansen and Juselius (1990) for n -r^ < 5 for the various deterministic trend models, and these are extended in Osterwald-Lenum for n -r^ <11. Horvath and Watson (1992) extend the tables to include non zero values of r ^ and r^ . The appropriate treatment of deterministic components in cointegration and unit root tests is still unsettled, and remains an active area of research. For example, Elliot, Rothenberg and Stock (1992) report that large gains in power for univariate unit root tests can be achieved by modifying standard Dickey-Fuller tests by an alternative method of detrending the data . They propose detrending the data using GLS estimators or hq and ^ from (3.39) together with specific assumptions about initial conditions for the xt process. Horvath and Watson (1993) apply an analogous procedure in likelihood based tests for cointegration. Johansen (1992b) develops a sequential testing procedure for cointegration in which the trend properties of the data and potential error corrections terms is unknown. 4. Structural V e c t o r Autoregressions 4.a Introductory Comments Following the work of Sims (1980), vector autoregression have been extensively used by economists for data description, forecasting, and structural inference. Canova (1991) surveys VAR’s as a tool for data description and forecasting; this survey focuses on structural -71 - inference. We begin the discussion in Section 4.b by introducing the structural moving average model, and shows that this model provides answers to the "impulse" and "propogation" questions often asked by macroeconomists. The relationship between the structural moving average model and structural VAR is the subject of Section 4.c. This section discusses the conditions under which the structural moving average polynomial can be inverted, so that the structural shocks can be recovered from a VAR. When this is possible, a structural VAR obtains. Section 4.d shows that the structural VAR can be interpreted as a dynamic simultaneous equations model, and discusses econometric identification of the model's parameters. Finally, Section 4.e discusses issues of estimation and statistical inference. 4.b The Structural Moving Average Model. Impulse Response Functions, and Variance Decompositions In this section we study the model: (4.1) yt «C (L )€t where yt is an ny x 1 vector of economic variables and is an n€x 1 vector of shocks. For now we allow ny£ne. Equation (4.1) is called the structural moving average model, since the elements of et are given a structural economic interpretation. For example, one element of ct might be interpreted as an exogenous shock to labor productivity, another as an exogenous shock to labor supply, another as an exogenous change in the quantity of money, and so forth. In the jargon developed for the analysis of dynamic simultaneous equations models, (4.4) is the final form of an economic model, in which the endogenous variables yt are expressed as a distributed lag of the exogenous variables, given here by the elements of et. It will be assumed that the endogenous variables yt are observed, but that the exogenous variables et are not directly observed. Rather, the elements of et are indirectly observed through their effect on -72- the elements of yt. This assumption is made without loss of generality, since any observed exogenous variables can always be added to the yt vector. In Section 2, a typical macroeconomic system was introduced and two broad questions were posed. The first question asked how the system’s endogenous variables responded dynamically to exogenous shocks. The second question asked which shocks were the primary causes of variability in the endogenous variables. Both of these questions are readily answered using the structural moving average model. First, the dynamic effects of the elements of on the elements of yt are determined by the elements of the matrix lag polynomial C(L). Letting C(L)=C q+ C j L +C 2L + ... , where nyXn£ matrix with typical element [cy ^], then (4.2) Cjj ^ = dyjydej^-k = ^ y i,t+ k ^ €j,t where y^ t is the i’th element of yt, t is the j ’th element of et, and the last equality follows from the time invariance of (4.1). Viewed as a function of k, C:: u is called the impulse response function of ej t for yj t. It shows how y^ changes in response to a one unit "impulse" in e; t. In the classic econometric literature on distributed lag models, the impulse responses are called dynamic multipliers. To answer the second question concerning the relative importance of the shocks, the probability structure of the model must be specified and the question must be refined. In most applications the probability structure is specified by assuming that the shocks are HD(0,Ee), so that any serial correlation in the exogenous variables is captured in the lag polynomial C(L). The assumption of zero mean is inconsequential, since deterministic components such as constants and trends can always be added to (4.4). Viewed in this way, et represents innovations or unanticipated shifts in the exogenous variables. The question concerning the relative importance of the shocks can be made more precise by casting it in terms of the h-step -73- is a ahead forecast errors of yt. Let ytyt.h =E(yt | {es} s=.oo) denote the h-step ahead forecast of yt made at time t-h, and let a ^ t- h ^ f ^ t/t- h 31 £ k = 0 ^ k €t-k denote the resulting forecast error. For small values of h, a ^ .^ can be interpreted as "short-run" movements in yt, while for large values of h, a ^ .^ can be interpreted as "long-run" movements. In the limit as h-»», at/t-h= y f importance of a specific shock can then be represented as the fraction of the variance in atyt_h that is explained by that shock; it can be calculated for short-run and longrun movements in yt by varying h. When the shocks are mutually correlated there is no unique way to do this, since their covariance must somehow be distributed. However, when the shocks are uncorrelated the calculation is straightforward. Assume Ec is diagonal with diagonal elements o j, then the variance of the i’th element of a ^ .jj is E jC lf r ’j E k= O c ij,k l' “ that (4.3) ^ ij,h = t1*} ^ k = O c ij,k( ^ { r r a = l ^ n S k —Oc im,kW shows the fraction of the h-step ahead forecast error variance in y^t attributed to €j t. The set 9 of ne values of R |j ^ are called the variance decomposition of yj t at horizon h. 4.c The Structural VAR Representation The structural VAR representation of (4.1) is obtained by inverting C(L) to yield: (4.4) A(L)yt = €t where A(L)= Aq- £ j AylJ* is a one-sided matrix lag polynomial. In (4.4), the exogenous shocks et are written as a distributed lag of current and lagged values of y^. The structural VAR representation is useful for two reasons. First, when the model parameters are known, it can be used to construct the unobserved exogenous shocks as a function of current and lagged -74- values of the observed variables yt. Second, it provides a convenient framework for estimating the model parameters: with A(L) approximated by a finite order polynomial, equation (4.4) is a dynamic simultaneous equations model, and standard simultaneous methods can be used to estimate the parameters. It is not always possible to invert C(L) and move from the structural moving average representation (4.1) to the VAR representation (4.4). One useful way to discuss the invertibility problem (see Granger and Anderson (1978)) is in terms of estimates of et constructed from (4.4) using truncated versions of A(L). Since a semi-infinite realization of the yt process, {ys} £ , is never available, estimates of €t must be constructed from (4.4) T t 1 using {ys} s = \- Consider the estimator ?t = I4 = ()Ajyt_j constructed from the truncated realization. If ?t converges to et in mean square as t-*oo, then the structural moving average process (4.1) is said to be invertible. Thus, when the process is invertible, the structural errors can be recovered as a one-sided moving average of the observed data, at least in large samples. This definition makes it clear that the structural moving average process cannot be inverted if n y < n £. Even in the static model yt =Cet, a necessary condition for obtaining a unique solution for et in terms of yt is that n y ^ n g. This requirement has a very important implication for structural analysis using VAR models: in general, small scale VAR’s can only be used for structural analysis when the endogenous variables can be explained by a small number of structural shocks. Thus, a bivariate VAR of macroeconomic variables is not useful for structural analysis if there are more than two important macroeconomic shocks affecting the variables. In what follows we assume that ny=n€. This rules out the simple cause of non- invertibility just discussed; it also assumes that any dynamic identities relating the elements of yt when ny > ng have been solved out of the model. With ny=nf s n, C(L) is square and the general requirement for invertibililty is that the determinantal polynomial | C(z) | have all of its roots outside the unit circle. Roots on the unit circle pose no special problems; they are evidence of overdifferencing and can be handled by -75- appropriately transforming the variables (e.g., accumulating the necessary linear combinations of the elements of yt). In any event, unit roots can be detected, at least in large samples, by appropriate statistical tests. Roots of | C(z) | that are inside the unit circle pose a much more difficult problem, since models with roots inside the unit circle have the same second moment properties as models with roots outside the unit circle. The simplist example of this is the univariate MA(1) model yt =(l-cL)et, where et is HD(0,o^). The same first and second moments of yt obtain for the model yt =(l-2L)?t, where c= c‘ * and is nD (0,a|) with = c ^ . Thus, the first two moments of yt cannot be used to discriminate between these two different models. This is important because it can lead to large specification errors in structural VAR models that cannot be detected from the data. For example, suppose that the true structural model is yt =(l-cL)et with | c | >1 so that the model is not invertible. A researcher using the invertible model would not recover the true structural shocks, but rather 1 1 oo i ?t =(l-cL) yt =(l-cL) (l-cL)ft =et-(£-c) £ j =Cr €t-l-r A general discussion of this subject is contained in Hannan (1970) and Rozanov (1967). Implications of these results for the interpretation of structural VAR’s is discussed in Hansen and Sargent (1991) and Lippi and Reichlin (1993). For related discussion see Quah (1986). Hansen and Sargent (1991) provides a specific economic model in which noninvertible structural moving average processes arise. In the model, one set of economic variables, say xt, are generated by an invertible moving average process. Another set of economic variables, say yt, are expectational variables, formed as discounted sums of expected future x^s. Hansen and Sargent then consider what would happen if only the yt data were available to the econometrician. They show that the implied moving average process of yt, written in terms of OO the structural shocks driving xt, is not invertible. The Hansen-Sargent example provides an important and constructive lesson for researchers using structural VAR’s: it is important to include variables that are directly related to the exogenous shocks under consideration (xt in the example above). If the only variables used in the model are indirect indicators with important -76- expectation^ elements (yt in the example above), severe misspecification may result. 4.d Identification of the Structural VAR Assuming that the lag polynomial of A(L) in (4.4) is of order p, then structural VAR can be written as: (4.5) AQyt = A ^ j. j + A2yt_2 + ... + Apyt.p + Since Aq is not restricted to be diagonal, equation (4.5) is a dynamic simultaneous equations model. It differs from standard representations of the simultaneous equations model (see Hausman (1983)) because observable exogenous variables are not included in the equations. However, since exogenous and predetermined variables —lagged values of yt. j —are treated identically for purposes of identification and estimation, equation (4.5) can be analyzed using techniques developed for simultaneous equations. The reduced form of (4.5) is: (4.6) yt = V m + *2*1-2 + - + *p*t-p + et where $^=A q *Aj, for i = l , ... , p, and et = A ^ e t. A natural first question concerns the identifiability of the structural parameters in (4.5), and this is the subject taken up in this section. The well known "order" condition for identification is readily deduced. Since yt is n x 1, 2 - 1 - 1 there are pn elements in ($ j, # 2, ••• » $p) and n(n+l)/2 elements in Ec =A q E£(Aq )’, the covariance matrix of the reduced form disturbances. When the structural shocks are __ 2 NIID(0,Ee), these [n p + n(n+l)/2] parameters completely characterize the probability distribution of the data. In the structural model (4.5) there are (p+ l)n elements in (Aq, A j , -77- 9 , Ap) and n(n+1)/2 elements in E£. Thus, there are n* more parameters in the structural model than are needed to characterize the likelihood function, so that t? restrictions are required for identification. As usual, setting the diagonal elements of Aq equal to 1, gives the first n restrictions. This leaves n(n-l) restrictions that must be deduced from economic considerations. The identifying restrictions must be dictated by the economic model under consideration. It makes little sense to discuss the restrictions without reference to a specific economic system. Here, some general remarks on identification are made in the context of a simple bivariate model explaining output and money; a more detailed discussion of identification in structural VAR models is presented in Giannini (1991). Let the first element of yt, say yj t, denote the rate growth of real output, and the second element of yt, say y2 t denote the rate of growth of 39 money. Writing the typical element of A^ as ajj j£, equation (4.5) becomes: (4.7a) y u = -a120y2it + E ? = 1a l u y lit.j + E ? = i » i 2ijy2,,.j + aljt (4.7b) y2(t = -a2 i,o y i,t + r ? = i a 2i ijy i>t.i + £ ? = l a22,i72,t-i + 62,r Equation (4.7a) is interpreted as an output or "aggregate supply" equation, with t interpreted as an aggregate supply or productivity shock. Equation (4.7b) is interpreted as a money supply "reaction function" showing how the money supply responds to contemporary output, lagged variables, and a money supply shock ^ f Identification requires n(n-l)=2 restrictions on the parameters of (4.7). In the standard analysis of simultaneous equation models, identification is achieved by imposing zero restrictions on the coefficients for the predetermined variables. For example, the order condition is satisfied if yj enters (4.7a) but not (4.7b), and y j ^ enters (4.7b) but not (4.7a); this imposes the two constraints Z2 \ j = a j 1 ^ = 0. *n this case» y l,t-l outPut equation but not the money equation, while yj ^ shifts the money equation but not the output -78- equation. Of course, this is a very odd restriction in the context of the output-money model, since the lags in the equations capture expectational effects, technological and institutional inertia arising production lags and sticky prices, information lags, etc.. There is little basis for identifying the model with the restriction &2\ \ = a i l 2 = ®* ^ e e d there is little basis for identifying the model with any zero restrictions on lag coefficients. Sims (1980) persuasively makes this argument in a more general context, and this has led structural VAR modelers to avoid imposing zero restrictions on lag coefficients. Instead, structural VAR’s have been identified using restrictions on the covariance matrix of structural shocks, Ee, the matrix of contemporaneous coefficients Aq and the matrix of long-run multipliers A (l)”*. Restrictions on have generally taken the form that is diagonal, so that the structural shocks are assumed to be uncorrelated. In the example above, this means that the underlying productivity shocks and money supply are uncorrelated, so that any contemporaneous cross equation impacts arise through nonzero values of a ^ q and *21 O’ Some researchers have found this a natural assumption to make, since it follows from a modeling strategy in which unobserved structural shocks are viewed as distinct phenomena which give rise to comovement in observed variables only through the specific economic interactions studied in the model. The restriction that is diagonal imposes n(n-l) restrictions on the model, leaving only n(n-l)/2 . . additional necessary restrictions. 40 These additional restrictions can come from a priori knowledge about the Aq matrix in (4.5). In the bivariate output-money model in (4.7), if is diagonal, then only n(n-l)/2=l restriction on Aq is required for identification. Thus, a priori knowledge of a ^ o or a21,0 serve to identify the model. For example, if it was assumed that the money shocks affect output only with a lag, so that dyj ^Jde2 t = 0, then a ^ q = 0, and this restriction identifies the model. The generalization of this restriction in the n-variable model produces the Wold causal chain (see Wold (1954) and Malinvaud (1980, pages 605-608)), in which d y jy d e j^ O for i< j. This restriction leads to a recursive model with Aq lower triangular, yielding the required -79- n(n-l)/2 identifying restrictions. This restriction was used in Sims (1980), and has become the "default" identifying restriction implemented automatically in commercial econometric software. Like any identifying restriction, it should never be used automatically. In the context of the output-money example, it is appropriate under the maintained assumption that exogenous money supply shocks, and the resulting change in interest rates, has no contemporaneous effect on output. This may be a reasonable assumption for data sampled at high frequencies, but loses its appeal as the sampling interval increases. Other restrictions on Aq can also be used to identify the model. Blanchard and Watson (1986), Bemanke (1986) and Sims (1986) present empirical models that are identified by zero restrictions on Aq that don’t yield a lower triangular matrix. Keating (1990) uses a related set of restrictions. Of course, nonzero equality restrictions can also be used; see Blanchard and Watson (1986) and King and Watson (1993) for examples. An alternative set of identifying restrictions relies on long-run relationships. In the context of structural VAR’s these restrictions were used in papers by Blanchard and Quah (1989) and King, Plosser, Stock and Watson (1991).^ These papers relied on restrictions on A (l)= A Q -£ f_ jA^ for identification. Since C(1)=A(1)"*, these can alternatively be viewed as restrictions on the sum of impulse responses. To motivate these restrictions, consider the output-money exam ple.^ Let Xj t denote the logarithm of the level of output and X2 t denote the logarithm of the level of money, so that y ^ t = A x ^ t and y2,t= ^ x2,f Then ^rom (*.l), (4.8) dxi,t+ k ^ ej,t = £m = 0^yi,t+ n/^€j,t ~ £m =0cij,m » f o r ij = l ,2, so that (4.9) lim ^Q,, dxj^+j^dej^ = E m =0cij,m -80- which is the ij’th element of C(l). Now, suppose that money is neutral in the long run, in the sense that shocks to money have no permanent effect on the level of output. This means that lim ^ o o ^xl ,t + k ^ €2 ,t= ®» 50 ^ at CO) is a lower triangular matrix. Since A(1)=C(1)"*, this means that A(l) is also lower triangular, and this yields the single extra identifying restriction that is required to identify the bivariate model. The analogous restriction in the general nvariable VAR, is the long-run Wold causal chain in which e: t has no long-run effect on y: t for j < i. This restriction implies that A(l) is lower triangular yielding the necessary n(n-l)/2 45 identifying restrictions. 4.e Estimating Structural VAR Models This section discusses methods for estimating the parameters of the structural VAR (4.5). The discussion is centered around generalized method of moment (GMM) of estimators. The relationship between these estimators and FIML estimators constructed from a Gaussian likelihood is discussed below. The simplist version of the GMM estimator is indirect least squares, which follows from the relationship between the reduced form parameters in (4.6) and the structural parameters in (4.5): (4.10) Aq'A j = i = l ........ p (4.11) A0EeA$ = Zt Indirect least squares estimators are formed by replacing the reduced form parameters in (4.10) and (4.11) with their OLS estimators and solving the resulting equations for the structural parameters. Assuming that the model is exactly identified, a solution will necessarily exist. A A A. A A Given estimators 4>j and Aq , equation (4.10) yields A|=A q4>j . The quadratic equation (4.11) is more difficult to solve. In general, iterative techniques are required, but simpler methods are presented below for specific models. -81 - To derive the large sample distribution of the estimators and to "solve" the indirect least squares equations when there are overidentifying restrictions, it is convenient to cast the problem in the standard GMM framework (see Hansen (1982)). Hausman, Newey and Taylor (1987) show how this framework can be used to construct efficient estimators for the simultaneous equations model with covariance restrictions on the error terms, thus providing a general procedure for forming efficient estimators in the structural VAR model. Some additional notation is useful. Let ^ ( y ^ j , Yt-2* •••» Yt-p)’ denote the vector of predetermined variables in the model, and let 6 denote the vector of unknown parameters in Aq, A j , ... , Ap and E€. The population moment conditions that implicitly define the structural parameters are: (4.12) E(€tz p = 0 (4.13) E(etep = Ee A. where et and are functions of the unknown 6. GMM estimators are formed by choosing 6 so that (4.12) and (4.13) are satisfied, or satisfied as closely as possible, with sample moments used in place of the population moments. The key ideas underlying the GMM estimator in the structural VAR model can be developed using the bivariate output-money example in (4.7). This avoids the cumbersome notation associated with the n-equation model and arbitrary covariance restrictions. (See Hausman, Newey and Taylor (1987) for discussion of the general case.) Assume that the model is identified by linear restrictions on the coefficients of Aq , A j , ... Ap and the restriction that E(cl t€2 ^ et W1 t denote the variables appearing on the right hand side of (4.7a) after the restrictions the structural coefficients have been solved out, and let coefficients. Thus, if a ^ q= 0 denote the corresponding is the only coefficient restriction in (4.7a), then only lags of yt appear in the equations and Wj t =(y|_j, y|_2» •••* Yt-p)’* ^ die long-run neutrality -82- assumption I ? = o a 12 >i = 0 is imposed in (4.7a), then w 1<t= ( y 1>t. 1, y l t _2 , .... y 1>t_p, Ay2>t, Ay2 t- i . ••• ^ . t - p + l ) ’ -46 Defining w2 t and ^ analogously for equation (4.7b), the model can be written as: (4.14a) yu = w JtJ j + « ,it (4.14b) y2,t ” w2,ts l + % f and the GMM moment equations are: (4.15a) E(zt€1>t) = 0 (4.15b) E(zte2 t ) = 0 (4.15c) E(el t e2 t ) = 0 (4.15d) E(e2iX <?€) = 0 , i= l,2 . The sample analogues of (4.15a)-(4.15c) determine the estimators and 52, while (4.15d) determines d2 and a2^ as sample averages of sums of squared residuals. Since the estimators of a2 { and a22 are standard, we focus on (4.15a)-(4.15c) and the resulting estimators of and 52. Let ut =(z|cj t, zje2 t, €j te2 t)’ and u=T'^ £ u t denote the sample values of the second moments in (4.15a)-(4.15c). Then the GMM estimators, Ak. and 62 , are values of (4.16) and ^ that minimize J = u 't n l u, .n 47 These estimators have a simple GLS or where Eu is a consistent estimator or E(utuJ). instrumental variable interpretation. To see this, let Z = (zj z2 ... z^-)’ denote the T x2p matrix of instruments; let W j = ( wj j wj 2 ... wj j ) ’ and W2 =(w2 j w 2 2 ... w2 j ) ’ denote the T x k j -83- and T x k 2 matrices of right hand side variables; finally, let Y j, Y2, ej and e2 denote the T x 1 vectors composed of yl t, y2>t, €1>t and «2 t respectively. Multiplying equations (4.14a) and (4.14b) by ^ and summing yields: (4.17a) Z’Yj = (Z’W p*! + Z’ej (4.17b) Z’Y2 = (Z’W ^ + Z’e2. Now, letting €j=Yj-Wj3j, for some3j (4.17c) e1’e2 + e 1,W232 +e2’W 131 = ( ^ ’W j ^ j + (ej’W j) ^ + * i* 2 + quadratic terms. Stacking equations (4.17a)-(4.17c) and ignoring the quadratic terms in (4.17c) yields: (4.18) Q = + P252 + V where Q =[(Z’Yt ) | (Z*Y2)| (71’72 + 71’W2^2 +72’W 13 1)], P r -[(Z *W j)|02pXki |( ^ W j) ] , p2 = [°2p x k 21(z ’w 2>I(«l ’W2)J» and V =[(Z ’e1)|(Z ’€2)|(€1’e2)], and where " | • denotes vertical concatenation ("stacking"). By inspection V=Tu from (4.16). Thus when Q, P j and P2 are A evaluated at A = 5j and ^2 = 62, the GMM estimators coincide with the GLS estimators from (4.18) . This means that the GMM estimators can be formed by iterative GLS estimation of “ equations (4.18), updating -I A A and $2 at each iteration and using T*1 £ utut as the GLS covariance matrix. Hausman, Newey and Taylor (1987) show that the resulting GMM estimators of 6j, $2* <r^i and o^2 are jointly asymptotically normally distributed when the vectors (zj e^)’ are independently distributed and standard regularity conditions hold. There results are readily -84- extended to the structural VAR when the roots of $(z) are outside the unit circle, so that the data are covariance stationary. Expressions for the asymptotic variance of the GMM estimators are given in their paper. When some of the variables in the model are integrated, the asymptotic distribution of the estimators changes in a way like that discussed in Section 2. This issue does not seem to have been studied explicitly in the structural VAR model, although such an analysis would seem to be reasonably straightforward.^® The paper by Hausman, Newey and Taylor (1987) also discuss the relationship between efficient GMM estimators and the FIML estimator constructed under the assumption that the errors are normally distributed. It shows that the FIML estimator can be written as the solution to (4.16), using a specific estimator of Eu appropriate under the normality assumption. In particular, FIML uses a block diagonal estimator of Eu, since E[[ej te2 t)(ej tzt)]=E[[Cl t€2,t^€2,tzt^ when the errors are normally distributed. When the errors are not normally distributed, this estimator of Eu may be inconsistent, leading to a loss of efficiency in the FIML estimator relative to the efficient GMM estimator. Estimation is simplified when there are no overidentifying restrictions. In this case, iteration is not required, and the GMM estimators can be constructed as instrumental variable estimators. When the model is just identified, only one restriction is imposed on the coefficient in equation (5.7). This implies that one of the vectors or 62 is 2px 1, while the other is (2 p + 1) x 1, and (4.18) is a set of 4 p + 1 linear equation in 4 p + 1 unknowns. Suppose, without loss of generality, that is 2 p x l. Then is determined from (4.17a) as 5 j =(Z ’W j )‘ 1(Z’Y j ), which is the usual IV estimator of equation (4.14a) using zt as instruments. Using this value for in (4.17c) and noting that Y 2 = ^ 2 ^2 + 1 2 ^ equation (4.17c) becomes (4.18) n 'Y 2 = A + .,-62 A where ej = Y^-W^Sj is the residual from the first equation. The GMM estimator of $2 is -85- formed by solving (4.17b) and (4.18) for 62. This can be recognized as the IV estimator of equation (4.14b) using zt and the residual from (4.14a) as an instrument. The residual is a valid instrument because of the covariance restriction (4.15c).^ In many structural VAR exercises, the impulse response functions and variance decompositions defined in Section 4.b are of more interest than the parameters of the structural VAR. Since C(L)=A(L)**, the moving average parameters/impulse responses and the variance decompositions are differentiable functions of the structural VAR parameters. The continuous mapping theorem directly yields the asymptotic distribution of these parameters from the distribution of the structural VAR parameters. Formulas for the resulting covariance matrix can be determined by delta method calculations. Convenient formulae for these covariance matrices can be found in Lutkepohl (1990) and Mittnik and Zadrozny (1993). Many applied researchers have instead relied on Monte Carlo methods for estimating standard errors of estimated impulse responses and variance decompositions. Runkle (1987) reports on experiments comparing the small sample accuracy of the estimators. He concludes that the delta method provides reasonably accurate estimates of the standard errors for the impulse responses, and the resulting confidence intervals have approximately the correct coverage. On the other hand, delta method confidence intervals for the variance decompositions are often unsatisfactory. This undoubtedly reflects the [0,1] bounded support of the variance decompositions and the unbounded support of the delta method normal approximation. -86- Footnotes 1. Since nominal rates are 1(0) from the last column of a, the long run interest semielasticity of money demand, 0r need not appear in the fourth column of a. 2. The values of (3y and 0T are important to macroeconomists because they determine (i) the relationship between the average growth rate of money, output and prices and (ii) the steadystate amount of seignorage associated with any given level of money growth. 3. Many of the insights developed by analyzing this example are discussed in Fuller (1976) and Sims(1978). 4. Throughout this paper B(s) will denote a multivariate standard Brownian motion process, i.e., an n x 1 process with independent increments B(r)-B(s) that are distributed I^O^r-s)^). 5. Higher order integrated processes can also be studied using the techniques discussed here, see Park and Phillips (1988) and Sims, Stock and Watson (1990). Seasonal unit roots (corresponding to zeroes elsewhere on the unit circle) can be also be studied using a modification of these procedures. See Tiao and Tsay (1989) for a careful analysis of this case. 6. The analysis in this section is based on a large body of work on estimation and inference in multivariate time series models with unit roots. A partial list of relevant references includes Chan and Wei (1988), Park and Phillips (1988)(1989), Phillips (1988), Phillips and Durlauf (1986), Sims, Stock and Watson (1991), Stock (1987), Tsay and Tiao (1991), and West (1988). Additional references are provided in the body of the text. -87- 7. A j, A2, and A4 are jointly normally distributed since j skdB(s)’« is a normally distributed 7k random variable with mean 0 and variance (w’w) j s ds. 8. This assumption is made without loss of generality since the constraint Q y= r (and resulting Wald statistic) is equivalent to CQ7=Cr for nonsingular C. For any matrix Q, C can chosen so that CQ is upper triangular. 9. q 12 is the only off-diagonal element appearing in Q. It appears because and ?2 both converge at rate T . 10. A detailed discussion of Granger-causality tests in integrated systems is contained in Sims, Stock and Watson (1990) and Toda and Phillips (1991)(1992). 11. Stock (1988), Table 4. These results are for durable plus nondurable consumption. When nondurable consumption is used, Stock estimates the bias to be -.15. 12. Toda and Phillips (1991)(1992) discuss testing for Granger causality in a situation in which the researcher knows that the number of unit roots in the model but doesn’t know the cointegrating vectors. They develop a sequence of asymptotic x tests for the problem. When the number of unit roots in the system in unknown, they suggest pretesting for the number of unit roots. While this will lead to sensible results in many empirical problems, examples such as the one presented at the end of this section show that large pretest biases are possible. 13. Alternatively, using "local-to-unity" asmptotics, the critical values can be represented as continuous functions of the local-to unity parameter, but this parameter cannot be consistently estimated from the data. See Bobkoski (1983), Cavanagh (1985), Chan and Wei (1987), Chan -88- (1988) and Phillips (1987b). 14. Hodrick (1992) contains an overview of the empirical literature on the predictability of stock prices using variables like the price-dividend ratio. Also see, Fama and French (1988) and Campbell (1990). 15. As Phillips and Loretan (1991) point out in their survey, continuous time formulations of error correction models were used extensively by A.W. Phillips in the 1950’s. I thank Peter Phillips for drawing this work to my attention. 16. To derive this result, note from (3.2) and (3.3) that n = -$ (l)= -U (l)M (l)V (l)= 5 a ’. Since M(l) has zeroes everywhere, except the lower diagonal block which is Ir, a ’ must be a nonsingular transformation of the last r rows of V(l). This implies that the first k columns of a ’V (l)'1 contain only zeros, so that a ’V(l)"1W (l)U (l)= a’C (l)=0. 17. The last component can be viewed as transitory because it has a finite spectrum at frequency zero. Since U(z) and V(z) are finite order with roots outside the unit circle, the Cj coefficients decline exponentially for large i, and thus £ ji | Cj | is finite. Thus the C* matrices * * are absolutely summable, and C (l)EeC (1)’ is finite. 18. The matrix G is not unique. One way to construct G is from the eigenvectors of A. The first k columns of G are the eigenvectors corresponding to the nonzero eigenvalues of A and the remaining eigenvectors are the last n-k columns of G. 19. While the usefulness of the triangular representation for analyzing estimators of cointegrating vectors was arguably demonstrated for the first time in Phillips (1991a), the -89- representation had been used in earlier work. For example, see Phillips and Durlauf (1986), Phillips (1988), and Park and Phillips (1988) (1989). 20. Much of the discussion in this section is based on material in Horvath and Watson (1993). 21. Formally, the restriction rank(5aa a)= ra should be added as a qualifier to Ha. Since, this constraint is satisfied almost surely by unconstrained estimators of (3.15) it can safely be ignored when constructing likelihood ratio test statistics. 22. In standard jargon, when r^ ^ O , the trace statistic corresponds to the the test for the alternative rau=n-rQu. 23. See Hansen (1990b) for a general discussion of the relationship between Wald, LR and LM tests in the presence of unidentified parameters. 24. The first term in (3.28) is the Wald statistic for testing 5 ^ = 0 imposing the constraint that 5 ^ = 0 . The second term is the Wald statistic for testing 6 ^ = 0 with a ^ X t_j and Zj partialled out of the regression. This form of the Wald statistic can be deduced from the partitioned inverse formula. 25. This compact way of writing the limiting distributions, using projections of Wiener processes, is taken from Park and Phillips (1988)(1989). 26. This example was pointed out to me by T. Rothenberg. 27. This is the formula for the projection onto the infinite sample, i.e. -90- 1 ^ 1 am 7 (L)Ax| =E[Uj | {Ax*} * —. cq]. In general, 7(L) is two-sided and infinite order, so that this is an approximation to E[Uj | {Ax*} * _ j]. The effect of this approximation error on estimators of j3 is discussed below. 28. This can be demonstrated as follows. When 7(L)=0, ^ t =D 2 2 ^ €2 t ^ X1 t = D l l ^ €l t* Let C (L )= [D 22(L )]and assume that matrix coefficients in C(L), D j j (L) and D22(L) are 1summable. Letting $=vec(/3), the GLS estimator and OLS estimators satisfy: TO q l s -S) = ( T ^ E q r t r ' f r - ' E q ^ t ) , T(5GLS-6) - (T'2 E q tq;)"1(T"1 E ^ 2 ,t ) . where qt = [xj t ®Ir], and defining the operator L so that ztL=Lzt =zt_j, qt = [xj t ®C(L)’]. Using the Lemma 2.c: t (5o l s -j ) = r r 2 E x 1,t*1,,’® irr , c r 1E (x 1y ® v D 22(i)«2,t] + °po>. = [ r '2 E * i,tx 1,t’® y ' 1(T '1E(X i,,’®D22(l)«2,t)] + ° P<1>, t («g l s -5) = [T'2 E { ca -)x litH xIit'c (L )';® irr 1iT '1E (x lit® c a ) > 2 , t ] + V » = i r ^ E x ^ t x ^ t ' s a D ’c f D r 'r r - 'E C x ^ t S c a ) ^ , , ) ] + op(i). Since C(l) ^ = ^ 2 2 0 )i "i ' (^OLS"^= ^ ' ^ G L S " ^ ^ O L S ^GLS^^’ 29. The long run covariance matrix for an n x 1 covariance stationary vector yt with absolutely summable autocovariances is ^ -.o o C o v fy ^y j.j), which is 2x times the spectral density matrix for yt at frequency zero. 30. See Wooldridge’s chapter of the Handbook for a thorough discussion of robust covariance matrix estimators. 31. This suggestion can be found in papers by Hansen (1988), Phillips (1991a), Phillips and -91 - Loretan (1989), Saikkonen (1991) and Stock and Watson (1993). Saikkonen (1991) contains a careful discussion of the approximation error that arises when 7(L) is approximated by a finite order polynomial. Using results of Berk (1974) and Said and Dickey (1984) he shows that a consistent estimator of 7(1) (which, as we show below is required for an asymptotically efficient estimator of /3) obtains if the order of the polynomial 7(L) increases at rate for 0 < 5 < 1/3. 32. See Hannan (1970) and Engle (1976) for a general discussion of band spectrum regression. 33. Consistent initial conditions for the iterations are easily constructed from the OLS a estimators of the parameters in the VAR (3.2). Let II denote the OLS estimator of II, A A A A A A A partitioned as 11=[IIj II2], where IIj is nx(n-r) and II2 is n x r; further partition II j =(TI|j E^j] ’ and n 2 =|TIi2 ^22l’» where II j j is (n-r)x(n-r), IL>i is rx(n-r), 11^ is (n-r)xr A A A jA and II22 is r x r. Then II2 serves an initial consistent estimator of 5 and -(1^22) “ 21 serves as an estimator of 0. Ahn and Reinsel (1990) and Saikkonen (1992) develop efficient two-step A estimators of /3 constructed from II, and Engle and Yoo (1991) develop an efficient three-step estimator of all the parameters in the model using iterations similar to those in (3.38). 34. See Basawa and Scott (1983) and Sweeting (1983). 35. We limit discussion to linear trends in yt for reasons of brevity and because this is the most important model for empirical applications. The results are readily extended to higher order trend polynomials. 36. Ogaki and Park (1990) define these two restrictions as "stochastic" and "deterministic" cointegration. Stochastic cointegration means that wt is 1(0), while deterministic cointegration means that Xj = 0. -92- 37. Blanchard and Quah (1989) and Canova, Faust and Leeper (1993) discuss special circumstances when some structural analysis is possible when n y< nfi. For example, suppose that yt is a scalar and the nf elements of et affect yt only through the scalar "index" et = D ’et, where D is n£x 1 vector. Then the impulse response functions can be recovered up to scale. 38. A simple version of their example is as follows: suppose that yt and xt are two scalar time series, with xt generated by the MA(1) process xt =€t-0et_j. Suppose that yt is related to xt by the expectational equation yt = Et E ? = ( A + i = xt + 7EtXt+1 = (1 -00)€t - m C(L)et where the second and third lines follow from the MA(1) process for xt. It is readily verified that the root of C(z) is (1-/86)/6, which may be less than 1 even when the root of (l-0z) is greater than 1. (For example, if 6=13=0.8, the root of (l-0z) is 1.25 and the root of C(z) is .8). 39. Much of this discussion concerning this example draws from King and Watson (1993). 40. Other restrictions on the covariance matrix are possible, but will not be discussed here. A more general discussion of identification with covariance restrictions can be found in Hausman and Taylor (1983), Fisher (1966), Rothenberg (1971) and the references cited there. 41. The appropriateness of the Wold causal chain was vigorously debated in the formative years of simultaneous equations. See Malinvaud (1980), pages 55-58 and the references cited there. -93- 42. Applied researchers sometimes estimate a variety of recursive models in the belief (or hope) that the set of recursive models somehow "brackets" the truth. There is no basis for this. Statements like "the ordering of the Wold causal chain didn’t matter for the results" say little about the robustness of the results to different identifying restrictions. 43. For other early applications of this approach, see Shapiro and Watson (1988) and Gali (1992). 44. The empirical model analyzed in Blanchard and Quah (1989) has the same structure as the output-money example with the unemployment rate used in the place of money growth. 45. Of course, restrictions on Aq and A(l) can be used in concert to identify the model. See Gali (1992) for an empirical example. 46. If a 12(L)= E?=(>ai2,iLi and a 12(l)= 0 , then a 12(L)y2>t=a*2(L)(l-L)y2>t=a*2(L)Ay2 t , where a*2(L) = iM> where a*2 j = - j y _ j + ja^2 y The discussion that follows assumes homogeneous (or zero) the linear restrictions on the structural coefficients. As usual, the only change required for nonhomogeneous (or non-zero) linear restrictions is a redefinition of the dependent variable. 47. When elements of and ur are correlated for t £ r , Eu is replaced by a consistent estimator of the limiting value of the variance of T*^u. 48. Instrumental variable estimators constructed from possibly integrated regressors and instruments is discussed in Phillips and Hansen (1990). -94- 49. While this instrumental variables scheme provides a simple way to compute the GMM estimator using standard computer software, the covariance matrix of the estimators constructed using the usual formula will not be correct. Using ej ^ as an instrument introduces "generated regressor" complications familiar from Pagan (1984). Corrections for the standard formula are provided in King and Watson (1993). An alternative approach is to carry out one GMM iteration using the IV estimators as starting values. The point estimates will remain unchanged, but standard GMM software will compute a consistent estimator of the correct covariance matrix. The usefulness of residuals as instruments is discussed in more detail in Hausman (1983), Hausman and Taylor (1983) and Hausman, Newey and Taylor (1987). -95- R e fe re nc es Ahn, S.K. and G.C. Reinsel (1990), "Estimation for Partially Nonstationary Autoregressive Models," Journal o f the American Statistical Association, 85, pp. 813-823. Anderson, T.W. (1951), "Estimating Linear Restrictions on Regression Coefficients for Multivariate Normal Distributions," Annals o f Mathematical Statistics, 22, pp. 327-51. Anderson, T.W. (1984), An Introduction to Multivariate Statistical Analysis, 2nd Edition, Wiley: New York. Andrews, D.W.K. and J.C. Moynihan (1990), "An Improved Heteroskedastic and Autocorrelation Consistent Covariance Matrix Estimator," Cowles Foundation Discussion Paper No. 942, Yale University. Baneijee, A., J J . Dolado, D.F. Hendry and G.W. Smith (1986), "Exploring Equilibrium Relationships in Econometrics through Static Models: Some Monte Carlo Evidence," Oxford Bulletin o f Economics and Statistics, Vol. 48, No. 3, pp. 253-70. Basawa, I. V. amd D J . Scott, (1983), Asymptotic Optimal Inference fo r Nonergodic Models, Springer Verlag: New York. Berk, K.N. (1974), "Consistent Autoregressive Spectral Estimates," Annals o f Statistics, 2, pp. 489-502. Bemanke, B. (1986), "Alternative Explanations of the Money-Income Correlation," CarnegieRochester Conference Series on Public Policy. Amsterdam: North Holland Publishing Company. Beveridge, Stephen and Charles R. Nelson (1981), "A New Approach to Decomposition of Time Series in Permanent and Transitory Components with Particular Attention to Measurement of the ’Business Cycle’," Journal o f Monetary Economics, 7, pp. 151-74 Blanchard, O.J. and D. Quah, (1989), "The Dynamic Effects of Aggregate Demand and Supply Disturbances," American Economic Review, 79, pp. 655-73. Blanchard, O.J. and Watson, M.W. (1986), "Are Business Cycles All Alike?," in R. Gordon (ed.), The American Business Cycle: Continuity and Change, Chicago: University of Chicago Press, pp. 123-179. Bobkosky, M.J. (1983), Hypothesis Testing in Nonstationary Time Series, Ph.D. thesis, Department of Statistics, University of Wisconsin. Brillinger, D.R. (1980), Time Series, Data Analysis and Theory, Expanded Edition, HoldenDay, San Francisco. Campbell, J.Y. (1990), "Measuring the Persistence of Expected Returns," American Economic Review, Vol. 80, No. 2, pp. 43-47. Campbell, J.Y. and P. Perron, (1991), "Pitfalls and Opportunities: What Macroeconomists Should Know about Unit Roots," Econometric Research Program Working Paper No. 360, Princeton University. -96- Campbell, J.Y. and R J. Shiller (1987), "Cointegration and Tests of Present Value Models," Journal o f Political Economy, 95, pp. 1062-1088. Reprinted in Long-Run Economic Relations: Readings in Cointegration, edited by R.F. Engle and C .W J. Granger, Oxford University Press, New York, 1991. Canova, Fabio (1991), "Vector Autoregressive Models: Specification Estimation and Testing," manuscript, Brown University Canova, F., J. Faust and E. M. Leeper (1993), "Do Long-Run Identifying Restrictions Identify Anything?", manuscript, Board of Governors of the Federal Reserve System. Cavanagh, C.L. (1985), "Roots Local to Unity," manuscript, Department of Economics, Harvard University. Cavanagh, C.L. and J.H. Stock (1985), "Inference in Econometric Models with Nearly Nonstationary Regressors," Manuscript, Kennedy School of Government, Harvard University. Chan, N.H. (1988), "On Parameter Inference for Nearly Nonstationary Time Series," Journal o f the American Statistical Association, 83, pp. 857-62. Chan, N.H. and Wei, C.Z. (1987), "Asymptotic Inference for Nearly Nonstationary AR(1) Processes," The Annals o f Statistics 15, pp. 1050-63. Chan, N.H. and Wei, C.Z. (1988), "Limiting Distributions of Least Squares Estimates of Unstable Autoregressive Processes," The Annals o f Statistics 16, no. 1, pp. 367-401. Cochrane, J.H. (1990), "Permanent and Transitory Components of GNP and Stock Prices," manuscript, University of Chicago. Cochrane, J.H. and A.M. Sbordone (1988), "Multivariate Estimates of the Permanent Components of GNP and Stock Prices," Journal o f Economic Dynamics and Control 12, pp. 255-296. Davidson, J.E., D.F Hendry, F. Srba and S. Yeo (1978), "Econometric Modelling of the Aggregate Time-Series Relationship Between Consumer’s Expenditures and Income in the United Kingdom," Economic Journal, 86. Davies, R. B. (1977), "Hypothesis Testing When a Parameter is Present Only Under the Alternative," Biometrika, Vol. 64, pp. 247-54. Davies, R. B. (1987), "Hypothesis Testing When a Parameter is Present Only Under the Alternative," Biometrika, Vol. 74, pp. 33-43. Dickey, D.A. and W.A. Fuller (1979), "Distribution of the Estimators for Autoregressive Time Series with a Unit Root," Journal o f the American Statistical Association 74, no. 366, 42731. Elliot, Graham, Thomas J. Rothenberg and James H. Stock (1992), "Efficient Tests of an Autoregressive Unit Root," NBER Technical Working Paper 130. -97- Elliot, Graham and James H. Stock (1992), "Inference in Time Series Regressions when there is Uncertainty about Whether a Regressor Contains a Unit Root," manuscript, Harvard University Engle, R.F. (1976), "Band Spectrum Regression," International Economic Review, Vol. 15, pp. Engle, R.F. (1984), "Wald, Likelihood Ratio, and Lagrange Multiplier Tests in Econometrics," in Z. Griliches and M. Intiligator (eds) Handbook o f Econometrics, Vol. 2, pp 775-826, North-Holland: New York. Engle, R.F. and C.W.J. Granger (1987), "Cointegration and Error Correction: Representation, Estimation, and Testing," Econometrica, 55, pp. 251-276. Reprinted in Long-Run Economic Relations: Readings in Cointegration, edited by R.F. Engle and C.W.J. Granger, Oxford University Press, New York, 1991. Engle, R.F., D.F. Hendry, and J.F. Richard (1983), "Exogeneity," Econometrica 51, no. 2, pp. 277-304. Engle, R.F. and B.S. Yoo (1987), "Forecasting and Testing in Cointegrated Systems," Journal o f Econometrics, 35, pp. 143-59. Reprinted in Long-Run Economic Relations: Readings in Cointegration, edited by R.F. Engle and C.W.J. Granger, Oxford University Press, New York, 1991. Engle, R.F. and B.S. Yoo (1991), "Cointegrated Economic Time Series: An Overview with New Results," in R.F. Engle and C.W.J. Granger (eds) Long-Run Economic Relations, Readings in Cointegration, Oxford University Press: New York. Fama E.F. and K.R. French (1988), "Permanent and Transitory Components of Stock Prices," Journal o f Political Economy, Vo, 96, No. 2, pp. 246-73. Fisher, F. (1966), The Identification Problem in Econometrics, Mew York: McGraw-Hill. Fisher, Mark E. and John J. Seater (1993), "Long-Run Neutrality and Supemeutrality in an ARIMA Framework," American Economic Review. Founds, N.G. and D.A. Dickey (1986), "Testing for a Unit Root Nonstationarity in Multivariate Time Series," manuscript, North Carolina State University. Fuller, W.A. (1976), Introduction to Statistical Time Series. New York: Wiley. Gali, Jordi (1992), "How Well does the IS-LM Model Fit Postwar U.S. Data," Quarterly Journal o f Economics, 107, pp. 709-738. Geweke, John (1986),"The Supemeutrality of Money in the United States: An Interpretation of the Evidence," Econometrica, 54, pp. 1-21. Gianini, Carlo (1991), "Topics in Structural VAR Econometrics," manuscript, Department of Economics, Universita Degli Studi Di Ancona. Gonzalo, J. (1989), "Comparison of Five Alternative Methods of Estimating Long Run Equilibrium Relationships," manuscript UCSD. -98- Granger, C.W.J. (1969), "Investigating Causal Relations by Econometric Methods and Cross Spectral Methods," Ecortometrica, Vol. 34, pp. 150-61. Granger, C.W.J. (1983), "Co-Integrated Variables and Error-Correcting Models," UCSD Discussion Paper 83-13. Granger, C.W.J. and A.P. Andersen (1978), "An Introduction to Bilinear Time Series Models," Vandenhoeck & Ruprecht: Gottingen. Granger, C. W.J. and P. Newbold (1974), "Spurious Regressions in Econometrics," Journal o f Econometrics, 2, pp. 111-20. Granger and Newbold (1976), Forecasting Economic Time Series, Academic Press. Granger, C.W.J and A.A. Weiss (1983), "Time Series Analysis of Error-Correcting Models," in Studies in Econometrics, Time Series, and Multivariate Statistics, pp. 255-78, Academic Press: New York. Granger, C.W.J. and T-H. Lee (1990), "Multicointegration," Advances in Econometrics, 8, pp. 71-84. Reprinted in Long-Run Economic Relations: Readings in Cointegration, edited by R.F. Engle and C.W.J. Granger, Oxford University Press, New York, 1991. Hall, Robert E. (1978), "Stochastic Implications of the Life Cycle - Permanent Income Hypothesis: Theory and Evidence," Journal o f Political Economy, Vol. 86, No. 6, pp. 97187. Hannan, E J. (1970), Multiple Time Series, Wiley: New York. Hansen, B.E. (1988), "Robust Inference in General Models of Cointegration," manuscript, Yale University. Hansen, B.E. (1990a), "A Powerful, Simple Test for Cointegration Using Cochrane-Orcutt," Working Paper No. 230, Rochester Center for Economic Research. Hansen, B.E. (1990b), "Inference When a Nuisance Parameter is Not Identified Under the Null Hypothesis," manuscript, University of Rochester. Hansen, B.E. and P.C.B. Phillips (1990), "Estimation and Inference in Models of Cointegration: A Simulation Study," Advances in Econometrics, 8, pp. 225-248. Hansen, Lars P. (1982), "Large Sample Properties of Generalized Method of Moments Estimators," Econometrica, 50, pp. 1029-54. Hansen, Lars P. and Thomas J. Sargent, (1991), "Two Problems in Interpreting Vector Autoregressions," in L. Hansen and T. Sargent (eds) Rational Expectations Econometrics, Westview: Boulder. Hausman, J.A. (1983), "Specification and Estimation of Simultaneous Equation Models," in Z. Griliches and M. Intiligator (eds) Handbook o f Econometrics, Vol. 1, pp. 391-448, NorthHolland: New York. -99- Hausman, Jerry A. and William E. Taylor (1983), "Identification in Linear Simultaneous Equations Models with Covariance Restrictions: An Instrumental Variables Interpretation," Econometrica, Vol. 51, No. 5, pp. 1527-50. Hausman, Jerry A., Whitney K. Newey and William E. Taylor (1987), "Efficient Estimation and Identification of Simultaneous Equation Models with Covariance Restrictions," Econometrica, Vol. 55, No. 4, pp. 849-874. Hendry, D.F. and T. von Ungem-Stemberg (1981), "Liquidity and Inflation Effects on Consumer’s Expenditure," in A.S. Deaton (ed.) Essays in the Theory and Measurement o f Consumer’s Behavior, Cambridge University Press: Cambridge. Hodrick, Robert J. (1992), "Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement," The Review o f Financial Studies, Vol. 5, No. 3, pp. 357-86. Horvath, Michael and Mark Watson (1992), "Critical Values for Likelihood Based Tests for Cointegration When Some Cointegrating May be Known," manuscript, Northwestern University. Horvath, Michael and Mark W. Watson (1993), "Testing for Cointegration When Some of the Cointegrating Vectors are Known," manuscript, Northwestern University. Johansen, S., (1988a), "Statistical Analysis of Cointegrating Vectors," Journal o f Economic Dynamics and Control, 12, pp. 231-54. Reprinted in Long-Run Economic Relations: Readings in Cointegration, edited by R.F. Engle and C.W.J. Granger, Oxford University Press, New York, 1991. Johansen, S. (1988b), "The Mathematical Structure of Error Correction Models," Contemporary Mathematics, vol. 80: Structural Inference from Stochastic Processes, N.U. Prabhu (ed.), American Mathematical Society: Providence, RI. Johansen, S. (1991), "Estimation and Hypothesis Testing of Cointegrating Vectors in Gaussian Vector Autoregression Models," Econometrica, 59, 1551-1580. Johansen, S. (1992a), "The Role of the Constant Term in Cointegration Analysis of Nonstationary Variables," Preprint No. 1, Institute of Mathematical Statistics, University of Copenhagen Johansen, S. (1992b), "Determination of Cointegration Rnak in the Presence of a Linear Trend," Oxford Bulletin o f Economics and Statistics, 54, pp. 383-397. Johansen, S. (1992c), "A Representation of Vector Autoregressive Processes Integrated of Order 2," Econometric Theory, Vol. 8, No. 2, pp. 188-202. Johansen, S. and K. Juselius (1990), "Maximum Likelihood Estimation and Inference on Cointegration —with Applications to the Demand for Money," Oxford Bulletin o f Economics and Statistics, 52, no. 2, pp. 169-210. Johansen, S. and K. Juselius (1992), "Testing Structural Hypotheses in a Multivariate Cointegration Analysis of the PPP and UIP of UK," Journal o f Econometrics, 53, pp. 21144. - 100- Keating, John (1990), "Identifying VAR Models Under Rational Expectations," Journal o f Monetary Economics, Vol. 25, No. 3, pp. 453-76. King, Robert G., Charles I. Plosser, James H. Stock and Mark W. Watson (1991), "Stochastic Trends and Economic Fluctuations," American Economic Review, 81, pp. 819-840. King, Robert G. and Mark W. Watson (1993), "Testing for Neutrality," manuscript, Northwestern University. Kosobud, Robert and Lawrence Klein (1961), "Some Econometrics of Growth: Great Ratios of Economics," Quarterly Journal o f Economics, 25, pp. 173-98. Lippi, M. and Reichlin, L. (1993), "A Note on Measuring the Dynamic Effects of Aggregate Demand and Supply Disturbances," American Economic Review, xx, pp. xx-xx. Lucas, Robert E. Jr. (1972), "Econometric Testing of the Natural Rate Hypothesis," in The Econometrics o f Price Determination, O. Eckstein, ed. Washington D.C.: Board of Governors of the Federal Reserve System. Lucas, R.E. (1988), "Money Demand in the United States: A Quantitative Review," CamegieRochester Conference Series on Public Policy, 29, pp. 137-68. Lutkepohl, H. (1990), "Asymptotic Distributions of Impulse Response Functions and Forecast Error Variance Decompositions of Vector Autoregressive Models," Review o f Economics and Statistics, 72, pp. 116-25. MacKinnon, James G. (1991), "Critical Values for Cointegration Tests," in R.F. Engle and C.W.J. Granger (eds) Long-Run Economic Relations, Readings in Cointegration, Oxford University Press: New York. Magnus, J.R. and H. Neudecker (1988), Matrix Differential Calculus, Wiley: New York. Malinvaud, E. (1980) Statistical Methods o f Econometrics, Amsterdam: North Holland. Mankiw, N.G. and M.D. Shapiro (1985), "Trends, Random Walks and the Permanent Income Hypothesis," Journal o f Monetary Economics, 16, pp. 165-74. Mittnik, S. and P.A. Zadrozny (1993), "Asymptotic Distributions of Impulse Responses, Step Responses and Variance Decompositions of Estimated Linear Dynamic Models," Econometrica, 61, pp. 857-70.. Ogaki, M. and J.Y. Park (1990), "A Cointegration Approach to Estimating Preference Parameters," manuscript University of Rochester. Osterwald-Lenum, Michael (1992), "A Note with Quantiles of the Asymptotic Distribution of the Maximum Likelihood Cointegration Rank Test Statistics," Oxford Bulletin o f Economics and Statistics, 54, pp. 461-71. Pagan, Adrian (1984), "Econometric Issues in the Analysis of Regressions with Generated Regressors," International Economic Review, 25, pp. 221-48. - 101 - Park, J.Y (1993), "Canonical Cointegrating Regression," E c o n o m e t r ic a , 61. Park, J.Y. and M. Ogaki (1991), "Inference in Cointegrated Models Using VAR Prewhitening to Estimate Shortrun Dynamics," Rochester Center for Economic Research Working Paper No. Park, J.Y. and P.C.B. Phillips (1988), "Statistical Inference in Regressions with Integrated Regressors I," E c o n o m e t r i c T h e o r y 4, pp. 468-97. Phillips, P.C.B. (1986), "Understanding Spurious Regression in Econometrics," J o u r n a l o E c o n o m e t r i c s , 33, pp. 311-40. Phillips, P.C.B. (1987a), "Time Series Regression with a Unit Root,: E c o n o m e t r ic a f , 55, pp. 277- Phillips, P.C.B. (1987b), "Toward a Unified Asymptotic Theory for Autoregression," B io m e t r i k a , 74, pp. 535-47. Phillips, P.C.B. (1988), "Multiple Regression with Integrated Regressors," M a t h e m a t i c s , 80, pp. 79-105. C o n te m p o ra ry Phillips, P.C.B., (1990), "To Criticize the Critics: An Objective Bayesian Analysis of Stochastic Trends," Cowles Foundation Discussion Paper no. 950; forthcoming, J o u r n a l o f A p p l i e d E c o n o m e t r ic s . Phillips, P.C.B. (1991a), "Optimal Inference in Cointegrated Systems," E c o n o m No. 2, pp. 283-306. e t r ic a , Vol. 59, Phillips, P.C.B. (1991b), "Spectral Regression for Cointegrated Time Series," in W. Barnett (ed.) N o n p a r a m e t r ic a n d S e m ip a r a m e t r ic M e t h o d s i n E c o n o m ic s a n d S t a t is t ic s , Cambridge University Press, pp. 413-436. Phillips, P.C.B. (1991c), "The Tail Behavior of Maximum Likelihood Estimators of Cointegrating Coefficients in Error Correction Models," manuscript, Yale University. Phillips, P.C.B. and S.N. Durlauf (1986), "Multiple Time Series Regression with Integrated Processes," R e v i e w o f E c o n o m i c S t u d ie s , 53, pp. 473-96. Phillips, P.C.B. and B.E. Hansen (1990), "Statistical Inference in Instrumental Variables Regression with 1(1) Processes," R e v ie w o f E c o n o m i c S t u d ie s , 57, pp. 99-125. Phillips, P.C.B. and M. Loretan (1989), "Estimating Long Run Economic Equilibria," Cowles Foundation Discussion Paper no. 928, Yale University. Phillips, P.C.B. and S. Ouliaris, (1990), "Asymptotic Properties of Residual Based Tests for Cointegration," E c o n o m e t r i c a , 58, pp. 165-94. Phillips, P.C.B. and J.Y. Park (1988), "Asymptotic Equivalence of OLS and GLS in Regression with Integrated Regressors," J o u r n a l o f t h e A m e r i c a n S t a t i s t i c a l A s s o c i a t i o n , 83, pp. 111115. Phillips, P.C.B. and P. Perron (1988), "Testing for Unit Root in Time Series Regression, " - 102- B io m e t r ik a , 75, pp. 335-46. Phillips, P.C.B. and W. Ploberger (1991), "Time Series Modeling with a Bayesian Frame of Reference: I. Concepts and Illustrations," manuscript, Yale University. Phillips P.C.B. and V. Solo (1992), "Asymptotics for Linear Processes," A r m pp. 971-1001. 20, a is o f S t a t is t ic s , Quah, D. (1986), "Estimation and Hypothesis Testing with Restricted Spectral Density Matrices: An Application to Uncovered Interest Parity," Chapter 4 of E s s a y s i n D y n a m ic M a r c o e c o n o m e t r ic s , Ph.D. Dissertation, Harvard University. Rothenberg, T.J. (1971), "Identification in Parametric Models," E c o n o m Rozanov, Y.A. (1967), S t a t io n a r y R a n d o m P ro c e s s e s e t r ic a , 39, pp. 577-92. , San Francisco: Holden Day. Runkle, D. (1987), "Vector Autoregressions and Reality," J o u r n a l o f B u s in e s s a n d E c o n o m ic S t a t is t ic s . Said, S.E. and D.A. Dickey (1984), "Testing for Unit Roots in Autoregressive-Moving Average Models of Unknown Order," B io m e t r ik a , 71, 599-608. Saikkonen, P. (1991), "Asymptotically Efficient Estimation of Cointegrating Regressions," E c o n o m e t r ic a T h e o r y , Vol. 7, No. 1, pp. 1-21. Saikkonen, P. (1992), "Estimation and Testing of Cointegrated Systems by an Autoregressive Approximation," E c o n o m e t r ic a T h e o r y , Vol. 8 , No. 1, pp. 1-27. Sargan, J.D. (1964), "Wages and Prices in the United Kingdom: A Study in Econometric Methodology," in P.E. Hart, G. Mills and J.N. Whittaker (eds), E c o n o m e t r i c A n a l y s is f o N a t i o n a l E c o n o m i c P l a n n i n g , London: Butterworths. Sargent, Thomas J. (1971), "A Note on the Accelerationist Controversy," J o u r n a l o B a n k i n g a n d C r e d i t , 3, pp. 50-60. r f M oney, Shapiro, Matthew and Mark W. Watson (1988), "Sources of Business Cycle Fluctuations," M a c r o e c o n o m i c s A n n u a l , Vol. 3, pp. 111-156. Sims, C.A. (1972), "Money, Income and Causality," A m 552. e r ic a n E c o n o m i c R e v ie w , 62, pp. 540- Sims, C.A. (1978), "Least Squares Estimation of Autoregressions with Some Unit Roots," University of Minnesota, Discussion Paper No. 78-95. Sims, C.A. (1980), "Macroeconomics and Reality," E c o n o m e t r ic a , 48, pp. 1-48. Sims, C. (1986), "Are Forecasting Models Usable for Policy Analysis?" Q u a r t e r l y Federal Reserve Bank of Minneapolis, Winter. Sims, C. (1989), "Models and Their Uses," A m 489-494. - 103- R e v ie w , e r ic a n J o u r n a l o f A g r i c u l t u r a l E c o n o m ic s , 71, pp. Sims, C.A., J.H. Stock, and M.W. Watson (1990), "Inference in Linear Time Series Models with Some Unit Roots," E c o n o m e t r i c a , Vol. 58, No. 1, pp. 113-44. Solo, Victor (1984), "The Order of Differencing in ARIMA Models," J o u r n a l o S t a t i s t i c a l A s s o c i a t i o n , 79, pp. 916-21. f t h e A m e r ic a n Stock, J.H. (1987), "Asymptotic Properties of Least Squares Estimators of Cointegrating Vectors." E c o n o m e t r i c a 55, pp. 1035-56. Stock, J.H. (1988), "A Reexamination of Friedman’s Consumption Puzzle," J o u r n a l o a n d E c o n o m i c S t a t i s t i c s 6 , no. 4, pp. 401-14. f B u s in e s s Stock, James H. (1991), "Confidence Intervals of the Largest Autoregressive Root in U.S. Macroeconomic Time Series," J o u r n a l o f M o n e t a r y E c o n o m ic s 28, pp. 435-60. Stock, James H. (1993), " xx ," forthcoming in R.F. Engle and D. McFadden (eds) H a n d b o o k E c o n o m e t r ic s , Vol. 4, North-Holland: New York. o f Stock, James H. (1992), "Deciding Between 1(0) and 1(1)," manuscript, Harvard University. Stock, James H. and Mark W. Watson (1988), "Interpreting the Evidence on Money-Income Causality," J o u r n a l o f E c o n o m e t r ic s , Vol. 40, Number 1, pp. 161-82. Stock, J.H. and M.W. Watson (1988), "Testing for Common Trends," J o u r n a l o f t h e A m e r i c a n S t a t i s t i c a l A s s o c i a t i o n , 83, pp. 1097-1107. Reprinted in L o n g - R u n E c o n o m i c R e l a t i o n s : R e a d in g s i n C o i n t e g r a t i o n , edited by R.F. Engle and C.W.J. Granger, Oxford University Press, New York, 1991. Stock, J.H. and M.W. Watson (1993), "A Simple Estimator of Cointegrating Vectors in Higher Order Integrated Systems," E c o n o m e t r ic a , 61, pp. 783-820. Stock, H.H. and K.D. West (1988),"Integrated Regressors and Tests of the Permanent Income Hypothesis," J o u r n a l o f M o n e t a r y E c o n o m ic s , Vol. 21, No. 1, pp. 85-95. Sweeting, T., (1983), "On Estimator Efficiency in Stochastic Processes," S t o c h a s t ic t h e i r A p p l i c a t i o n s , Vol. 15, pp. 93-98. Theil, Henri (1971), P r i n c i p l e s o f E c o n o m e t r ic s , P ro ce sse s a n d Wiley: New York. Toda, H.Y. and P.C.B. Phillips (1991), "Vector Autoregressions and Causality," Cowles Foundation Working paper no. 1001. Toda, H.Y. and P.C.B. Phillips (1992), "Vector Autoregressions and Causality: A Theoretical Overview and Simulation Study," E c o n o m e t r i c R e v ie w s . Tsay, R.S and G.C Tiao (1990), "Asymptotic Properties of Multivariate Nonstationary Processes with Applications to Autoregressions," A n n a l s o f S t a t is t ic s , 18. pp. 220-50. West, K.D. (1988), "Asymptotic Normality when Regressors Have a Unit Root," E c o n o m 56, pp. 1397-1418. White, Halbert (1984), A s y m p t o t ic T h e o r y f o r E c o n o m e t r ic ia n s , -104- e t r ic a , New York: Academic Press. Whittle, Peter (1983), P r e d i c t i o n a n d R e g u la t io n b y L i n e a r L e a s t - S q u a r e Edition, Revised, University of Minnesota Press: Minneapolis. Wold, H. (1954), "Causality and Econometrics," E c o n o m e t r ic a , M e th o d s , Second xx, pp xx-xx. Wooldridge, J. (1993), " xx ," forthcoming in R.F. Engle and D. McFadden (eds) H a n d b o o k E c o n o m e t r ic s , Vol. 4, North-Holland: New York. o f Yoo, B. Sam (1987), "Co-Integrated Time Series* Structure, Forecasting and Testing," Ph.D. Dissertation, UCSD. Yule, G.C. (1926), "Why Do We Sometimes Get Nonsense-Correlations Between Time-Series," J o u r n a l o f t h e R o y a l S t a t i s t i c a l S o c ie t y B , 89, pp. 1-64. -105- Table 1 Comparing Power of Tests for Cointegration D e s ig n : h r*ti 2 LxtJ £t [xj - x£] + 2 -52L£tJ '«i where «t-(€* S iz e ' ’NIID(0>I2) ' and t - 1 ,...,1 0 0 . fo r Po w er 5% fo r A s y m p t o t ic T e s ts C r it ic a l C a r r ie d O u t a t V a lu e s 58 and L e v e l S Test DF (a known) EG-DF (a unknown) Wald (a known) LR (a unknown) (0,0) 5.0 4.7 4.7 4.4 (.05,.055) 6.5 2.9 95.0 (-.05,.055) 81.5 31.9 54.2 86.1 20.8 (-.105,0) 81.9 32.5 91.5 60.7 Notes: These results are based on 10,000 replications. The fir s t column shows rejection frequencies using asymptotic c r itica l values. The other columns show rejection frequencies using 5% cr itic a l values calculated from the experiment in column 1. -106-