The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
A SIMPLE ESTIMATOR OF COINTEGRATING VECTORS IN HIGHER ORDER INTEGRATED SYSTEMS James H. Stock and Mark W. Watson Working Paper Series Macro Economic Issues Research Department Federal Reserve Bank of Chicago February, 1991 (WP-91-3) A Simple Estimator of Cointegrating Vectors in Higher Order Integrated Systems James H. Stock Department of Economics University of California, Berkeley Berkeley, CA 94720 and Mark W. Watson Department of Economics Northwestern University Evanston, IL 60208 and the Federal Reserve Bank of Chicago First Draft: This revision: September 1989 January 1991 Abstract Efficient estimators of cointegrating vectors are presented for systems involving deterministic components and variables of differing, higher orders of integration. The estimators are computed using GLS^or OLS, and Wald statistics constructed from these estimators have asymptotic x distributions. These and previously proposed estimators of cointegrating vectors are used to study long-run U.S. money (Ml) demand. Ml demand is found to be stable over 1900-1989; the 95% confidence intervals for the income elasticity and interest rate semielasticity are (.90, 1.03) and (-.124, -.088), respectively. Estimates based on the postwar data alone, however, are unstable, with variances that suggest substantial sampling uncertainty. Keywords: Error correction models, unit roots, money demand JEL classification number: 210 The authors thank Manfred Deistler, Robert Engle, Peter Phillips, Danny Quah, Kenneth West, and the participants in the NBER/FMME Summer Institute Workshop on New Econometric Methods in Financial Time Series, July 17-20, 1989 for helpful comments on earlier versions of this paper. Helpful suggestions by Lars Hansen and two anonymous referees are gratefully acknowledged. The authors also thank Robert Lucas for kindly providing the annual data analyzed in Section 7. This research was supported in part by the Sloan Foundation and National Science Foundation grants SES-86-18984 and SES-89-10601. 1. Introduction Parameters describing the long-run relation between economic time series, such as the long-run income and interest elasticities of money demand, often play an important role in empirical macroeconomics. If these variables are cointegrated as defined by Engle and Granger (1987), then the task of describing these long-run relations reduces to the problem of estimating cointegrating vectors. Recent research on the estimation of cointegrating vectors has focused on the case that each series is individually integrated of order 1 (is 1(1)), typically with no drift term. Johansen (1988a) and Ahn and Reinsel (1990) independently derived the asymptotic distribution of the Gaussian MLE when the cointegrated system is parameterized as a vector error correction model (VECM), and Johansen (1989) extended this result to the case of nonzero drifts. In a series of papers, Phillips and coauthors have considered efficient estimation based on a different model for cointegrated systems, the triangular representation. Phillips (1988a) studied estimation in a cointegrated model with general 1(0) errors; Phillips and Hansen (1989) considered a two-step zero frequency seemingly unrelated regression estimator; and Phillips (1988b) used spectral methods to compute efficient estimators in the frequency domain. This paper proposes two alternative, computationally simple estimators of cointegrating vectors, which readily extend to systems with arbitrary deterministic components and with higher orders of integration and cointegration. The estimators are motivated as Gaussian MLE's for a particular parameterization of the triangular representation. However, under more general conditions they are shown to be asymptotically efficient in Phillips' (1988a) sense, having an asymptotic distribution that is a random mixture of normals and producing Wald test statistics with asymptotic chi-squared null distributions. 1 In the 1(1) case with a single cointegrating vector, one simply regresses one of the variables onto contemporaneous levels of the remaining variables, leads and lags of their first differences, and a constant, using either ordinary or generalized least squares. It is argued that the resulting "dynamic OLS" (respectively GLS) estimators are asymptotically equivalent to the Johansen/Ahn-Reinsel MLE. The proposed estimators treat the parameters describing the short-run dynamics of the process as nuisance parameters; the object is to obtain efficient estimates of the cointegrating vector, which are typically of independent interest. If desired, the estimates subsequently can be used to study the short-run dynamics, say by estimating a triangular system (Campbell (1987), Campbell and Shiller (1987, 1989)) or a constrained VECM (King, Plosser, Stock and Watson (1987)). These estimators are used here to investigate the long-run demand for money (Ml) in the U.S. from 1900 to 1989. Other researchers (recently including Hoffman and Rasche [1989] and Baba, Hendry and Starr [1990]) have argued either explicitly or implicitly that long-run money demand can be thought of as a cointegrating relation among real balances, real income and the interest rate in postwar data. We find this characterization empirically plausible for the longer annual data as well, and therefore use these estimators of cointegrating vectors to examine Lucas's (1988) suggestion that there is a stable long-run Ml money demand relation spanning the twentieth century. The paper is organized as follows. The model and estimators are introduced in Section 2 for 1(1) variables, and they are extended to 1(d) variables in Section 3. The large-sample properties of the estimators and test statistics are summarized in Section 4. In Section 5, the proposed estimators are related to Johansen's (1988a) MLE, and the 1(2) case is examined in detail. results are presented in Section 6. given in Section 7. Monte Carlo The application to long-run Ml demand is Section 8 concludes. 2 Readers primarily interested in the empirical results can skip Sections 3-6 with little loss of continuity. 2. 1 Representation and Estimation in 1(1) Systems Suppose that each element of the n-dimensional time series y t is 1(1), that EAyt»0, and that the nxr matrix of r cointegrating vectors a is a - (-$ where B is the Ir)', rx(n-r) submatrix of unknown parameters to be estimated and Ir is the rxr identity matrix. The triangular representation for yt is, (2.1a) (2.1b) 1 2 1 2 where yt is partitioned as (yt , yt), where yt is (n-r)xl and yt is rxl and 1 2 where ut — (ut ' u^')' is a stationary stochastic process with full rank spectral density matrix. This representation has been used extensively in theoretical work by Phillips (1988a,1988b), typically without parametric structure on the 1(0) process u^, and in applications by Campbell (1987) and Campbell and Shiller (1987, 1989) (also see Bewley [1979]). For the moment, u^ is assumed to be Gaussian to permit the development of the Gaussian MLE for B . The parameterization that forms the basis for the proposed estimators is obtained by making the error in (2.1b) independent of (u^). Because ut is 2 1 2 1 1 Gaussian and stationary, E[ut |{Ayt )]«E[ut |{u^}]*d(L)Ayt , where d(L) is in general two-sided and {Ay^} denotes {Ay^, t-l,...,T). written, ( 2 . 2) y\ - 0y\ + d(L)Ay^ + 3 Thus (2.1b) can be -2 2 2 1 where ut*ut"E[ut |{ut J] . independent. 1 -2 By construction, (Ayt ) and {ut J are 1 -2 In addition, ut and ut have Wold representations 1 1 2 2 1 2 ut*c^^(L)€t and uta*c 2 2 ^'^€t ’ where {c^} and {ct } are independent. Thus (2.1a) and (2.2) can be written, (2.3a) Ay* - cn (L)eJ: (2.3b) y2 - 0y* + d(L)Ay* + c22(L)e2 2 ’ so ^ct^ where ct is NIID(0,2€), E €~diag(2^, 1 ^dependent °f (yt) . The two-sided triangular representation (2.3) provides a nonstandard asymptotic factorization of the Gaussian likelihood. and Let A^ denote the parameters of c-^(L) let A 2 denote the parameters of d(L) , C2 (y^,...,yj), i—1,2. (2.4) 2 (L), and ^ 2 * anc^ ^et ^ denote Then (2.3) implies that the likelihood can be factored as f(Y1 ,Y2 |9,A1 ,A2) - f(Y2 |Y1 ,0,X 2 )f (Y1 | . This differs from the usual prediction-error factorization because the conditional 2 1 mean of yt involves future as well as past value of y^. The representations (2.3) and (2.4) provide concrete guides to estimation and inference in these Gaussian systems. [0,Ag], then If there are no restrictions between A^ and is ancillary (in Engle, Hendry and Richard's [1983] terminology, weakly exogenous, extended to permit conditioning on both leads and lags of Ay^) for 0, so that inference can be carried out conditional on Y^. In this case, the 2 1 MLE of 0 can be obtained by maximizing f(Y |Y ,0,A2 ) . This reduces to estimating the parameters of the regression equation (2.3b) by GLS. Because the regressor y^ is 1(1), as is shown in Section 4 an asymptotically equivalent estimator of 0 can be obtained by estimating 0 in (2.3b) by OLS; this will be referred to as the 4 dynamic OLS (DOLS) estimator, to distinguish it from the static OLS (SOLS) 2 1 estimator obtained by regressing yt against only a constant and y^. Similarly, the feasible GLS estimator of 6 in (2.3b) will be referred to as the dynamic GLS (DGLS) estimator. The representation (2.3) warrants three remarks. First, Sims' (1972) Theorem 2 implies that the projection d(L)Ay^ will involve only current and lagged values 1 2 1 of Ayt if (and only if) u^ does not Granger-cause Ayt . C2 2 If so, and if (l) ^ has finite order p, then (2.3b) can be rewritten as an r-dimensional 2 1 error correction model, i.e. as a regression of Ayt onto (Ayt , Ayt Ay t - 2 , . . . , Ayt , yt_i~0'yt-i)• this case, the nonlinear least squares estimator (with Ayt included as a regressor; Stock [1987]) is the Gaussian MLE. Second, the large-sample properties of the OLS and GLS estimators of 0 are 2 readily deduced from the representation (2.3). Because €t is uncorrelated with the regressors at all leads and lags, conditional on the GLS estimator has a normal distribution and the Wald statistic testing the hypothesis that 0«0q (where rank(0)=r is maintained) has a x 2 1 distribution. Because yt is 1(1), the conditional covariance matrix of the GLS estimator differs across realizations of y\ even in large samples; thus unconditionally the GLS estimator of 6 has a large-sample distribution that is a random mixture of normals and the Wald statistic has a x 2 distribution. Phillips (1988a) provides an insightful discussion of the asymptotic mixed normal property of the MLE of 0, the local asymptotic mixed normal (LAMN) behavior of test statistics, and the efficiency of 2 the Gaussian maximum likelihood estimator of 0 . Third, although the interpretation of (2.3) as a factorization of the likelihood (2.4) assumes Gaussianity, the two-sided triangular representation with 2 1 ' EetAyt-j*0 for all j can be constructed under the weaker condition that ut is linearly regular and covariance stationary. 5 This result, summarized in the following lemma (proven in the Appendix) , provides an alternative to the Wold representation theorem. 3 Let ^ i 1 and denote the linear manifolds spanned by {u^}, -®<s<t and {u^}, -»<s<«> (respectively), closed with respect to convergence in mean square, and let ^(u^l^) denote the projection of onto . Lemma 2.1. 1 Let 2 — (u^' ut ')' he a mean zero linearly regular weakly stationary stochastic process with Eu^u^ < ®, where u^ is 2 (n-r)xl and ut is rxl. --- 1 O (L) II 1-1 ut 2 Then u^ has the representation, _c21 KJ c22 r li £t 2 LetJ or ut~c(L)€t (where c(L) and ct are partitioned conformably with ut) , where Ect^O, Ect€^.=diag(S^^, ^ 1 2 ) and Ec^Cg-0 f°r * and where {c^} are 2 innovations of {ut ) and {et ) are innovations of {ut*ut2 1 P(ut 1^oq) } • Also, cn (L) and c 2 2 -2 2 ^ ) are one-sided, in general ^^(L) is two-sided, and c(L) is square summable. In the representation (2.5), C2 ^(z) is in general two-sided because {e^} are constructed to be the innovations of {u^}. Note however that {et ) are not innovations for {ut).^ 3. Representation in 1(d) Systems This section extends the triangular representation (2.3) to systems in which variables may be integrated and cointegrated of different orders and in which there are polynomial time trends. The 1(d) generalization of (2.1), derived in 6 the Appendix from the Wold representation of an 1(d) system with drifts and multiple cointegrating vectors, is (3.1) .d 1 A yt Ad-1 2 A yt Ad-2 3 A yt **1,0 + ut ■ i /,d-l/?d-l„l, »d "l/i d “1 1 v . ^2,0 + 1*2,lz + *2,1(A yt) + u *. 0 **3,0 + + ^ 3 , 2 t2 T Ad-1 /Ad - 1 1 N ^ -d-2 /Ad - 2 1 . ^ .d-2 ,Ad - 2 ^3 ,1 (A yt> + ^ 3 ,1 (A y t > + * 3 , 2 (A 5 d+1 yt d vd .d-i j-l/-i-j*d+lfj for t-l,...,T, where the 3 + ut d+1 + ut are kjXl vectors which form a partition of yt , i.e. yt=(y^' y^' ... y^+^')', and A^=(l-L)(1-L), etc. By assumption, ut=(u^' 2 d+1 u£' ... u£ ')' is weakly stationary with mean zero. It is assumed that the highest order of integration of the elements of yt is 1(d). However, not all elements of yt need to be 1(d) for (3.1) to apply (see the examples in Section 5(B) and the empirical application to long-run money demand in Section 7). Moreover, some blocks of (3.1) might not appear. For example, with d~2 and n=2, yt could be Cl(2,1) in Engle and Granger's (1987) terminology, in which case k^=l, k 2 =l, and kg=0; if yt is Cl(2,2), then k^-1, and kg**l. This representation partitions yt into components corresponding to stochastic trends of different orders. Abstracting from the deterministic components, y^ is a k-^-vector corresponding to the k-^ distinct 1(d) stochastic trends in the system. In the second block of k 2 equations, 2 d -1 1 l^t corresPon<^s to the k 2 distinct I(d-l) elements in the system; for rows of 2 zero, yt is I(d-l), while for nonzero rows of 1 2 (yt , yt) are CI(d,l) . 2 yt is 1(d) and The kg equations in the third block describe the distinct I(d-2) components, and so forth. d-1 that equal 7 It is straightforward to generalize the representation (3.1) to include higher order polynomials in t, or to specialize it to the common case in which higher-order polynomials are suppressed. By assumption, (3.1) requires contemporaneous linear combinations of levels of yt , or levels of yt plus differences, to be integrated of at least order zero so that ut has a finite nonsingular spectral density matrix. This assumption is made without loss of generality, and serves to fix the maximum order of integration d. In practice, this is achieved by replacing yt by A ^yt or Ayt as needed for u^ to be 1(0) and not cointegrated. As in the 1(1) case, the errors are orthogonalized by projecting onto leads and lags of the errors in the preceding equations. 2.1, u^' ... u£+1')' has the representation u t - C(L)ct (3.2) where By repeated application of Lemma A. d+lfv , 2 , ... «t «t' ')' and E£t€4-S£-diag(211,S22.... sd+l,d+l>’ and where C(L) is a block lower triangular matrix partitioned conformably with ut , with diagonal blocks c^(L) that are one-sided lag polynomials and with lower offdiagonal blocks c^j(L) that in general are two-sided. We focus our attention on the first i blocks of equations, and make the additional assumption that {c^^(L)}, i—l,...,i-l are invertible, so that the first 1 blocks of (3.2) can be written, 1 (3.3) ■^2 i(L) • • U.ua> 0 • • • 0 1 0 • • • • 'di,i-l(L) * * • 1 ■ c-na) . 0 • ut 0 . cJei(L) where in general d^j (L) are two-sided (for example, d 2 ^(L)~C2 ^(L)c^(L) Substitution of the i-th equation in (3.3) into (3.1) then yields, 8 ) i t where {/xp ., j»l,...,jB-1} are functions of [u x t J lag polynomials projection of NIID(0,Se J ., j-l,...,m-l, m«l,...,i}. The which generalize d(L) in (2.3), arise from the onto {u1^} for m*=l,...,i-1. When {y^} is Gaussian, e t is ) . 5 i 1 iThe subspaces that cointegrate y t with (yt , . . . , yt 1 ) and their differences are determined by the matrices (0^ j} appearing in the second >J term on the right hand side of (3.4). Note that the i-th block of equations contains all of the cointegrating vectors for m<i, which appear in the higher order error correction terms making up the third term on the right hand side of (3.4). For example, in a system with d-2 the equations describing cointegration in the levels contain any cointegrating relations between the first differences. As in the 1(1) case, one motivation for considering (3.4) is as the conditional mean of a nonstandard factorization of the Gaussian likelihood. Let A^ denote the parameters of c^(L) and <^m (L), m*l,...i-l, let 0^ denote {0^ j, j«l,...,i1, i«j ,...,i-1) , let denote {/i^ j,j*0,...,i-l), and let A, 0, and /x represent the collection of A^, 0^, and /x^. (3.5) The Gaussian likelihood can be written as: f(Y,0 ,M»A) - f where Y=(y£, y£, ... , y^,)' and Y1 »(yj' , y \ ' ...... y^')' for 9 i-1.... d+1. If the parameters ( 0 t^le higher-order cointegrating vectors in (3.1) are known and if there are no restrictions between (A^,0^,/i^) and 1 1-1 {(Aj ,0j ,/ij ) ,j<i} , then (Y ,...,Y ) are ancillary for 0^. of 0^ is obtained by estimating the system (3.4) by GLS. In this case the MLE If some of (0g ,...,0^_]_) 1 1-1 are unknown then (Y ,...,Y ) are no longer ancillary (weakly exogenous) for 0^, and the GLS estimator of {0^’j} in (3.4) is not the exact Gaussian MLE. However, as is made precise in Section 5 for the 1(2) case, the DGLS and DOLS estimators of {0^ j} in (3.4) still have desirable properties. We therefore consider regression estimators of (3.4). Because the regressors in (3.4) can have stochastic or deterministic trends in common, it is convenient to transform the regressors to isolate these different trends. regressors in each equation in (3.4). Let denote the It is assumed that Xt is a known nonrandom 1 1-1 (not data-dependent) linear combination of Y ,...,Y . Define zt - DXt , where D is an invertible matrix of constants (possibly unknown) , chosen so that zt are the canonical regressors in the sense of Sims, Stock and Watson (1990). The choice of transformation matrix D depends on the specific application (see tioi 5(B) for examples). Section 1 2 In general, partition z^ as (z z ^ ' 21 9 \ 9 z~')', where by construction z^ is an 1(0) vector with mean zero (z^ contains the required leads and lags of {u^, mC0} , dictated by the polynomials 2 3 4 5 (d^m (L)}), zt~l, zt is dominated by a martingale, zt=t, zt is dominated by 6 2 integrated martingales, z^=t , and so forth. 0 (T*~^) for i>2. P T i i In general Xt*lztzt From Sims, Stock, and Watson (1990, Section 2), z. can be u written as zt * G(L)ut , where G(L) is a block lower triangular matrix and 0 V « t * i 1 «t and where ^ C «t— * i-lw )', where 0■= /(1£ , 1 £ 2,' mmm£ i-l,w ' l ’ ') is defined recursively by ?t“^ s - l ^ s ^ or Also> let i 21 denote the dimension of z£, and let g=$\_^g^ be the dimension of zt . With these definitions, the system (3.4) equivalently can be written as, 10 (3.6) A d-i+i z yt - <x t ® Tki>^ + et or d->e+i i yt - <zt ® A (3.7) where ej.-c^^L) ct> tj€t €t + et coe^ficients P are ana K L€t the coefficients on the regressors appearing in (3.4); these are related to the parameters of interest (the cointegrating parameters) are the coefficients on the integrated elements of z^, it is convenient to partition the gk^-vector 8 as 6*=(6^ 8*2 ... $2 ^)' > where 8 ^ is the g^k^-vector of coefficients on z^. 4. Estimation and Testing This section examines the least squares estimation of the parameters 8 in (3.7). It is assumed that yt has the triangular representation (3.1) with ut given by (3.3). We consider the case that zt and 8 are finite dimensional, i.e. in which {d^m (L), m<i} have fixed finite orders. It is assumed that {e^) in (3.2) is a martingale difference sequence with E[et€^.| ct et_2 >**•]“2 €- d i a g ( 2 ^ , /| S22,..., 2d+l,d+l) nonsingular and maxisuptE[(c.t) I61-1 *£t-2 '***^<0°* There are two natural estimators of the parameters in (3.6): the feasible GLS estimator based on an estimator of c^^(L), suitably parameterized, and the dynamic OLS estimator (respectively, the DGLS and DOLS estimators). (4.1) (4.2) 11 These are where zt-[zt ® 4>(L) '] and yt - $>(L)yt , where $(L) is an estimator of ^ ( D - s ^ c ^ a r 1. Associated with the DGLS estimator is the Wald statistic testing the h restrictions R£»r (where R and r have dimensions hxgk^ and hxl, respectively), (4.3) WGLS * ^R^GLS’r^ [R (X^tz£) R ^ (R^GLS"r^ ‘ Because the disturbance in (3.6) is serially correlated, the Wald statistic for $Ols must constructed using a modified covariance matrix. When the hypotheses of interest do not involve the coefficients on the mean-zero stationary regressors in (3.6), this is the spectral density matrix of e*. at frequency zero, Qi>e*cii(l)2iicii(l) 9 , estimated by (4.4) That is, W q LS - [R^OLS’r l#tR [(Iztzt ^ 0 Define the gxg scaling matrix R> * to be a block diagonal matrix partitioned conformably with zt , with diagonal blocks i>2. [R^OLS“r l * and for Also, for w t weakly stationary, define rw (j) - E[wt-E(wt) ] [wt_j-E(wt) ]' . The next four theorems, proven in the Appendix, summarize the asymptotic distributions of these statistics. Theorem 4.1 Suppose that y^ satisfies (3.4) and (3.6) where c..(L) is d+l-j c JJ summable, j-l,...,d+l, that c^(L)~^ has known order q<«, and ^^(L) has a known finite order. Then (T^ ® * k ^ ^ G L S ’^ Q partitioning Q and </> conformably with S : Qn - E[(zJ: ® $(L)')(z*' ® $(L))], Qj_j - 0, j>2, and Q jj - [V4j ® $(!)'$(!)] for i,j >2, where V22 - 1, 12 where after % > - Gmm<1> l J ' M “ ‘1>/2<s>H lP ‘l)/2<s>'dsl°pp<1>'' m-3,5,7.... 2i-l, p-3,5,7.... 2i-l vmp - G»n<1 >lJ'is<m'2 >/M P '1 )/'2 <8 >'<,s)GpP<1)' ' Vpm - m-2,4,6.... 2i, p-3,5,6,...,2i-l Vmp “ 2 /<P+m'2 >Gmm(1 )GPP (1)'’ m " 2 ’ 4 ' 6 .... 2i> P" 2 ’ 4 ’ 6 .... 2i> 4>x - N<0,E[(z£ ® •(L)')Sii(zJ» ® *(L))]), K “ / i (Gnun( K “ / ^ Gmm(1 )Wim '1)/2(s) ® *(1)')<W2 C), m-3,5,7.... 2i-l, where 1 ) s < m ' 2 ) / 2 ® *(l)')<W2 (s), m-2,4,6.... 2i, and W 2 are independent standard Wiener processes of dimension an<* respectively, where (t)«/^w£m i~l,2 and m=2,3,. ..,g, and where <f> i s Theorem 4.2. (s)ds for independent of <f>m , m>l. Under the assumptions of Theorem 4.1, (a) (T,p ® 1 ^ ) (^oLS"^ => where after partitioning V and w conformably with 8: «1 ~ N ( ° - V ’ Where wm “ /o(Gmm ( 1 ) s ( m " 2 ) / 2 trzl<J) ® ® « ^ ) d W 2(s), m-2,4,6.... 2i, «m = / o (Gmm(1 )Wim '1)/2(s) ® 0 ii)dW2(s)’ m-3,5,7.... 2i-l, where w ^ is independent of wm , m>l, and where V * [V-^j ] , i,j * 1, 2,...,2i, where 4.1. - Ez^zJ;', V^. * 0, j>2, and V^j , i,j > 2 are given in Theorem This holds even if c^(L) ^ has infinite order as long as c^(l) is 1- summable and c^^Cl) is nonsingular. (b) Partition 8 « ( 5 6 * ' ) ' so t^iat denotes the g e l e m e n t s of 8 corresponding to z^ and 8 * corresponds to the remaining (g-g^)k^ elements of S. Similarly partition ^q l S ’ ^GLS’ zt“ ^zt T^=diag(TiT,T^T) • z t ^ ' ’ an<* Then (T*T ® 1 ^ ) (^*q l s ”^*GLS^ ^ G ' 13 Theorem 4.3. Under the Assumptions of Theorem 4.1, W ^ g —> Theorem 4.4. Suppose that the first ^ o columns of R equal zero and that T* 1 6 1 1 under the assumptions of Theorem 4.1, W q l S ** VGLS ^ ® and WOLS _ > x h ‘ Note that c^^(L) ^ must be be finitely parameterized to implement the DGLS estimator. Although this is not strictly needed for the DOLS estimator, and therefore c^(l) must be consistently estimated to construct W q l s » which in practice entails estimating a parametric approximation. The asymptotic equivalence of the DOLS and DGLS estimators of S* (Theorem 4.2(b)) is a consequence of the trends in z^: for m>2 the GLS-transformed regressors are asymptotically collinear with their untransformed counterparts. This result extends the familiar result for the case of a constant and polynomial time trend (Grenander and Rosenblatt [1957]) and the results of Phillips and Park (1986) for 1(1) regressors to the general integrated regression model with regressors of various orders of integration. Although the theorems are stated in terms of 5, typically the results are translated into results on the coefficients of interest, f). j&Gls *-s obtained from $GLS“ ^ ^OLS* P«R(D' ^ k ^ ^ G L S an<* The distribution of and similarly for Moreover, the Wald statistic testing R 6 =r equivalently tests P£=r, where -1 ®I^g). Theorem 4.3 implies that WGLg is asymptotically x the Wald statistic testing P£=r is asymptotically x 2 for all P. 2 for all R, so When P/J=r places no restrictions on coefficients that can be written as coefficients on mean-zero stationary regresssors, Theorem 4.4 implies that the Wald test of P/3=r based on 2 j&OLs (with a serial correlation-robust covariance matrix) is asymptotically x • Importantly, the result concerning the asymptotic x 14 2 distribution of the Wald statistic testing restrictions on cointegrating vectors applies whether or not the integrated regressors have components that are polynomials in time. However, the limiting distribution of the estimator itself will differ depending on whether time (say) is included as a regressor and whether some of the regressors have a time trend component; for specific examples in the 1(1) case, see West (1988) and Hansen (1989). These theorems apply to the case that there are a fixed number of regressors. Conceptually, one could view this estimator as semiparametric by embedding this parametric regression in a sequence of regressions where the number of regressors increase as a function of the sample size. A formal treatment of this extension would entail generalizing the univariate 1(0) results of Berk (1974) and the univariate 1(1) results of Said and Dickey (1984) to the 1(d), vector-valued case, an extension not undertaken here.^ 5. Examples This section examines two examples in detail. The first compares the model of Section 2, and the associated dynamic OLS and GLS estimators, with the 1(1) VECM formulation studied in Engle and Granger (1987) and Johansen (1988a, 1988b). second example examines various cases of the A. 1 The (2 ) specialization of (3.1). Comparison with the 1(1) vector error correction model. One representation of a purely stochastic 1(1) cointegrated system is the VECM, (5.1) Ayt - 7<*'yt - l + A(L)Ayt - 1 + ft , ft NIID(0,Sf), t-1, ...,T where yt is nxl, A(L) has finite order and is unrestricted, a is a nxr matrix of 15 cointegrating vectors and 2^. is unrestricted. Johansen (1988a) and Ahn and Reinsel (1990) derived the limiting distribution of the Gaussian MLE unknown parameters of a in (5.1). for the Here, the MLE for (5.1) is related to the DOLS estimator. Because the asymptotic information matrix for (uMTF, AMTR(z)) in (5.1) is block diagonal, let A(L)*0. vector, partition 7 -(7 ^' 7 2 1 2 Partition yt-(yt ' y t)9 into a (n-r)- and r- )9 conformably, where 7 ^ is (n-r)xr and 7 2 is rxr and normalize a as a«(-0 Ir)9, where 6 is rx(n-r). Without loss of generality assume that 7 g is nonsingular. /co\ (5.2a) The block triangular form of (5.1) is, A 1 - ut 1, Ayt 1 2 1 +.^ft y\ “ *j\ + ut> (5.2b) > 2 Clearly ut depends on 6 and, if 7 ^ P-1r+a''t ’ 1 0 , so does ut , so the factorization of the likelihood (2.4) results in restrictions between 0 and Ag and between A^ and (0 ,A2 ). In this model, the exact MLE (9MT F) is the system estimator studied by Johansen (1988a) and Ahn and Reinsel (1990), not the single equation estimator examined in Section 4. To study the behavior of ^LE* ^ **s convenfent to reparameterize the VECM (5.1) (with A(L)-O) as: (5.3a) Ay 1 - IIAy^ + i; 1 (5.3b) Ayj: - where n - 1^1^, + ^ t - l + *t P2~12 > and rhT^\' 1 1 1 1 8 parameterization is convenient because (5.3) is an unconstrained linear triangular simultaneous equation model so that MLE's correspond to iterated SUR estimators 16 (see Lahiri and Schmidt [1978]). The MLE's of the cointegrating vectors in (5.1) can be recovered from the MLE's of (5.3) as ^ L E “ ’^2^MLE^1 MLE* Isolating the regressors of different orders of integration, equation (5.3b) can be written as: (5.3c) Ay l - Sxz\ + «3zJ + r,\ 1 2 1 3 1 where zt«yt ^-0 yt_^=a'yt ^ * 3 parameterization zt are canonical * an<i ^3 ~^l+^ 2 ^ • In t^1^s 1 (1 ) regressors, z^ are mean zero 1 (0 ) 1 1 ~^2 ~7 2 regressors, the true value of S^ is zero, and ^m LE“^“ "^1^MLE^3 MLE* (5.a) and (5.c), ^gg and 2 ^gg can be written as OLS estimators from the 1 3 *1 l 2 regression of Ayt onto z£, zt and ^-Ay^-ft^ggAy^. 11 $ 1 From 2 Because ft^gg and 1 MTF are consistent, Eytyt' is Op(T ), and 2yt_^Ayt ' is Op(T), by direct calculation, (5.4) T(0m l e -0) - '^1 ^MLET^3,MLE “ ^t^t^t-l^ 1 2 Et-2yt-lyt-l) 1 + °p^1J 2 1 lv 7 where at-=f?^-E(»y^| »/^) . The single equation estimators are obtained from the regression, (5.5) y\ - 8y\ + d(L)AyJ: + et For finite order VECM's, d(L) will typically be infinite order, so the finite approximations used in Sections 2-4 will result in misspecified regression equations. If, however, the order of d(L) (say q) is such that q-*x> as T-*® and 3 q /T-*0, then the results of Berk (1974) and Said and Dickey (1984) suggest that this misspecification will vanish asymptotically. With this interpretation, the 17 single-equation dynamic OLS estimator can be written, (5.6) where T<»0LS-U> - "^1 + op<l) t^ie l° 1 ^ 1 2 n 6 run component of e^, where 0*c(l)Sj.c(l)' is (2n times) the spectral density matrix of at frequency zero. A straightforward calculation demonstrates that et*"^2^at+l’ so t^iat T< W Jq l s ^ ^ 0. Thus, even though the VECM likelihood cannot be factored as in (2.4), the single equation estimator is asymptotically equivalent to the MLE for this model. An interpretation is that, even though there are constraints across equations and across parameters that appear when the triangular system (5.2) is derived from a VECM, these constraints only involve coefficients that obey conventional JT asymptotics. Thus, asymptotically they convey no information about 6 , beyond that contained in the unrestricted second equation. Indeed this is a general property of regressions involving integrated regressors: the asymptotic distribution of estimators of coefficients on canonical integrated regressors remains unaltered when efficient estimators of coefficients on zero mean 1(0) regressors are replaced by consistent esimators. One example of this is given by the MLE of 6 for (5.1); asymptotically equivalent estimators can be constructed as where 2 1 regression of Ayt onto yt 2 and are t^ie OLS estimators from the $ 2 - 1 1 ~ 2 Y^-l an<* *7 t“Ayt-nAyt , w^ere ® any consistent estimator of II in equation (5.3a). Their asymptotic equivalence notwithstanding, it is useful to think of DOLS and #Ml e as applying to two different models. For finite order VECM's, Johansen's (1988a) estimator is the MLE, while for models that support the factorization (2.4) with d(L) finite order, the single equation estimators are the MLE. 18 B. Examples of 1(2) Systems. The following examples concern specification and inference in general 1(2) systems. To simplify exposition, all deterministic terms are omitted and their coefficients are taken to be zero. general From (3.1), the (2 ) model is, 1 (5.7a) (5.7b) (5.7c) Some of the 0's can have rows of zero, or be zero, and the second block of equations might not be present at all (i.e. k 2 ~ 0 ). These possibilities are examined here by considering a series of special cases with k 2 ~ 0 general cases can be analyzed by combining these special cases. or k 2 ~l; more It is assumed that {d^ m (L)} in (3.3) have known finite order. 2 . Then (5.7b) does not appear in the system and yt is omitted from Case 1: (5.7c). The dynamic OLS and GLS estimators of (0^ p 0^ ^) are asymptotically 2 efficient and inference is x • Case 2: k^^l, 0^ ^ ^nown and nonzero. f Then the estimation equation (3.4) becomes, _ 3 1 ~3 2 where Eutus ' and Eutug ' are zero f°r all t9s. 2 1 2 the regressors A yt , Ay t 1 - 0 2 1 anc* 1 Because 0g ^ is known, leads and lags are 1(0) with 19 1 1 mean zero, so these comprise zt . 2 Because yt and yt are CI(2,1), we can set 3 1 2 1 1 5 1 zt«(Ayt , yt-0 2 iYt)' and zt«yt (other assignments of z^ are possible 3 5 The coefficients on z^ and z^ are respectively but yield the same results). f3 /f3 r3N /al .0 8 **(5-^, ^ 2 ™ ^ 3 i » ^ 2 N , c5 n0 'and o ^ 1 ^3 2*^2 an<* ^3 1**^“^2 1^2* (T,T2), 0^ i> A ”^ ^ 2 ,0 1^3 2 * *1 *3 ^3 i"*®i> Because (i^, $"*) converge at rates 2 an<* ^3 1 individually converge at the rate T. ^ 3 Jointly, (0® \+&\ 1^3 2*^3 2 ’^3 1^ converge at rates (T2 ,T,T). The 2 estimators are efficient and inference is x • Case 3: Jc2 -Z, ^\ \ known to be zero. 1 2 02 1 The estimation equation is (5.8) with 2 Leads and lags of A yt and Ayt are 1(0) with mean zero, and these 1 3 1 2 2 5 1 comprise zt . Also set zt«(Ayt ' yt)' (yt is 1(1)) and zt“yt . Thus 0 0 1 2 2 (ij i » ^ 3 2*^3 1^ converge at (T ,T ,T) and inference is x • 1 Case 4: ( ^ 3 k<£*l, 82 ^ unknown. 1 2 Although (Y ,Y ) are not weakly exogenous for i » ® 3 2*^3 1^ *-n ^ i s case» the dynamic OLS and GLS estimators nevertheless have desirable properties. With 02 ^ unknown, the estimation equation (5.8) becomes, y\ (5.9) -d3,2<1 >*2,l>Ayt + d3,2<1>Ayt + *3 ,lyt + «3,2yt (d3(l(L)-d3j2(L)^2,l^A2yt + d3,2^L^A2yt + where either 2 1 (L)“ (1 *L) (0 ) (if 0 2 ( b ) 2 ^^' ^ j/0 ) or I(-l) (if 0 Because A^yJ: *-s 1 (0 ) an<* A^y^ ^ ]_” 0 ) > and because both have mean zero, their presence does not affect the asymptotic distribution of the other estimators and they will be ignored in this discussion. Whether or not 02 -j~0, a valid 1 2 1 1 3 1 2 1 1 assignment of zfc is zt-(Ayt - 0 2 j A y fc)', zt-(Ayt , yt-0 2 >iyt)'» and 20 zj^-y^. Evidently 03 ^ is not identified from (5.9) alone; using a 1 consistent estimator of 02 1 from (5.7b) would result in loss of x (although the resulting estimator would be consistent). 0^ 2 2 inference However, 03 ^ and are s®Parately identified in (5.9) and individually converge at rate T. 3 5 Together, the coefficients on (zt , z^) have an asymptotic mixed normal distribution. Moreover, the distribution of (0® case 2, when the true value of 0^ 1 *-s known. ^3 2^ Thus (0^ t*ie same as *-n ^3 ^ are 2 asymptotically efficient even if 0^ ^ is unknown, for general 0^ exception to this is the special case of 0 ^ ^ known to be zero, in which case Ay^ would not enter as a regressor in (5.8) were 0^ ^ known. inference on (0 ^ 2 The Even here, • ) Is . Monte Carlo Results 6 This section summarizes a study of the sampling properties of seven estimators of cointegrating vectors in three bivariate probability models. The data were generated by the model: (6 .1 a) Ay* - u* (6 .1 b) - 0 y* + NIID(0,2j.), where ut-(u* u^) '. with $(L)ut—ft , $(L)«l2 -$L, in the series is zero. Because ut follows a VAR(l), yt follows a VAR(2). Under (6 .1 ), T (5 -0 ) is invariant to loss of generality 0 The true drift 0 for all the estimators considered, so without is set to zero. The six estimators considered are the static OLS estimator (SOLS; Engle and Granger [1987], Stock [1987]), the dynamic OLS estimator 21 (DOLS) , the dynamic GLS estimator (DGLS), the zero frequency band spectrum estimator of Phillips (1988b) (PBSR), the fully modified estimator of Phillips and Hansen (1989) (PHFM), and Johansen's (1988a) VECM maximum likelihood estimator (JOH). JOH were calculated as described in Section 5(A). t-statistics for Two serial correlation-robust estimators of the covariance matrix of the DOLS estimator were considered, one based on a weighted sum of the autocovariances of the errors (D0LS1), the second using an autoregressive spectral estimator (D0LS2). all estimation procedures. A constant was included in The details of the construction of the estimators are given in the notes to Table 1. The design (6.1) parsimoniously nests several important special cases. (case A), when all elements of $ except First equal zero and Ej. is diagonal, Ay^ is strictly exogenous in (6.1b) and SOLS is the MLE. In this case, all the efficient estimators are asymptotically equivalent to SOLS, although they estimate nuisance parameters that in fact are zero. of $ is zero and $ 2 1 ^® or Second (case B) , if the second column is not diagonal or both, then SOLS is no longer the MLE and does not have an asymptotic mixed normal distribution, but the DOLS, DGLS, and JOH estimators are correctly specified and are asymptotically MLE's (the difference again being the unnecessary estimation of some nuisance parameters). In this case, PBSR and PHFM are efficient if interpreted semiparametrically. Third (case C) , for general $ and 2^., JOH with one lag is the exact MLE and DOLS, DGLS, PBSR, and PHFM are asymptotically efficient when interpreted semiparametrically. Results for cases A, B and C are reported in the respective panels of Table 1 for T**100 and 300. Panel A verifies that the estimation of the nuisance parameters per se in the efficient estimators does not substantially reduce performance in the special case that OLS is the MLE. Panel B explores the performance of the estimators in 22 models in which DOLS, DGLS, and JOH are 22 correctly specified. Even when ^21~^’ SOLS can have substantial bias; for example, for T-100 and ^^--.90, the 5%, 50%, and 95% points of the SOLS distribution are -.001, .076, .196. The DOLS, DGLS, and JOH estimators eliminate this bias, although when the regressor exhibits strong serial correlation, the DOLS t-statistics tend to have heavier tails than predicted by the asymptotic distribution theory. to SOLS. The PBSR and PHFM estimators tend to have biases comparable When this bias is small (for example when 21*®^ ’ their t-statistics have approximately normal distributions. The final case ($, 2^. unrestricted) introduces two additional parameters, and it is beyond the scope of this investigation to explore this case in detail. Rather, case C is examined by generating data from a model relevant to the empirical analysis in Section 7, namely a bivariate model of log Ml velocity (v) and the commercial paper rate (r), estimated using annual data from 1904-1989 (earlier observations were used for initial lags) imposing a long-run interest g semielasticity of -.10. The estimated VAR(l) for the triangular system (vt , Vt+.10rt) is reported in panel C of Table 1. The results for this system indicate large bias in SOLS and, to a lesser extent, in DGLS, PBSR, and PHFM. DOLS exhibits less bias and, not surprisingly since it is the exact MLE in this system, JOH is essentially unbiased. The dispersion of the distributions are comparable, except for the JOH estimator which has some large outliers for T-100. The x 2 approximation to the Wald statistic (testing 0«-.lO) works best for JOH, next best for D0LS2 and DGLS, less well for the remaining efficient estimators. To interpret DOLS and DGLS results, it is useful to write (6.1) in the triangular form given in (2.3). Write the VAR(l) for u^ as tf(L)ut-at , where -h -H tf(L)=E£. $(L) and at~2j. et , so that E(ata£)*I. 1 Then Ayt has a univariate ARMA(2,1) representation and c^^(L) in (2.3a) is given by c^(L)~/c(L) |tf(L) | where k (L) is the first degree polynomial with roots outside the unit circle that 23 solves /c(L)/c(L ^)-^2 2 ^L ^ 2 2 ^L # T^ie Pr°ject^on °f y^-fly^ onto {Ay^} is d(L)Ay^; for this design d(L)— [ ^ ^ ( L ) ^ ^ ^ " ^ ) + ^ll(L)^i2 (L’^) ] [#c(L)/c(L ^)] C2 2 Finally, the residual from this regression, 2 (L)€t in (2.3b), follows an AR(1) with C2 2 (L) -1 -*(L ). Thus *c(L) dictates both how quickly the coefficients on leads and lags of Ay^ in the DOLS/DGLS regressions die out and the degree of serial correlation in the regression error. In cases A and B, /c(L)»l, and the DOLS/DGLS regressions have no omitted variables. In case C, *(L)*1-.66L so the true d(L) is infinite order and the DOLS/DGLS regressions omit leads and lags of Ay^. The results from the experiments can be summarized as follows. First, SOLS is biased in almost all trials, with nonstandard distributions for the estimator and test statistics. Second, DOLS and DGLS are unbiased for cases A and B, but exhibit bias in Case C. The relatively large root of /c(L) suggests that the bias is attributable to the truncation of d(L) in the DOLS/DGLS regressions. o Third, in results not shown in the table, doubling the number of leads and lags for DOLS and DGLS and the order of the AR correction for DGLS has little effect in cases A and B and reduces the bias in case C . ^ Fourth, the PBSR and PHFM bias has the same sign as, but is somewhat less than, the SOLS bias. A possible explanation is that both PBSR and PHFM rely on initial biased SOLS estimates of 0 , which result in inaccurate spectral density estimates subsquently used to compute PBSR and PHFM. Fifth, for case C (where the error is highly serially correlated) the autoregressive spectral estimator used in D0LS2 produces a more normallydistributed t-statistic than does the kernel estimator used in D0LS1. Sixth, tripling the sample size noticeably improves the quality of the asymptotic approximations. This modest Monte Carlo experiment suggests three conclusions. First, all the estimators (except the correctly-specified JOH) exhibit bias in some of the 24 simulations, although the bias is in each case less than for SOLS: no single estimator is a panacea. Second, the distributions of the t-ratios tend to be spread out relative to the normal distribution, suggesting that the usual confidence intervals will overstate precision. has shortcomings: Third, in case C each estimator the DGLS, PBSR, and PHFM estimators are substantially biased, and the JOH estimator, while unbiased, has an empirical distribution with a much greater dispersion than the other efficient estimators; DOLS has the lowest RMSE. Fourth, of the two procedures for computing the covariance matrix, the autoregressive estimator produces t-statistics that are more normally-distributed than does the kernel estimator. For this reason, the DOLS standard errors reported in the empirical analysis in Section 7 are based on the autoregressive covariance estimator. 7. Application to the Long-Run Demand for Money in the U.S. This section addresses two questions. First, is there a stable long-run Ml demand equation spanning 1900-1989 in the United States? Second, what are the income elasticity and interest semielasticity, and how precisely are these estimated? The long-run demand for money plays an important role in the quantitative analysis of the effects of monetary policy. Unfortunately, estimates of long-run income and interest elasticities obtained using postwar data have been sensitive to the sample period and specification (see the reviews by Laidler [1977], Judd and Scadding [1982], and Goldfeld and Sichel [1990]). In his review of this research and of early work by Meltzer (1963), Lucas (1988) presented informal but highly suggestive evidence that this apparent sensitivity resulted not from a breakdown of the prewar long-run Ml demand relation, but from a paucity of low frequency information in the postwar data. 25 This section examines Lucas's interpretation using the formal econometric techniques for the analysis of cointegrating relations developed in this paper and elsewhere. Our analysis focuses on the annual data studied by Lucas (1988), extended to cover 1900-1989, although for comparison with other studies selected results using postwar monthly data are presented as well.^ A . Results for annual data. The annual time series are Ml (in logarithms, m ) , real net national product (in logarithms, y ) , the net national product price deflator (in logarithms, p ) , and the commercial paper rate (in percent at an 12 annual rate, r) . Real Ml balances (m-p, plotted with y in Figure la) grew strongly over the first half of the century, but experienced almost no net growth over most of the postwar period. Over the entire period, velocity (y-m+p) and r (plotted in Figure lb) exhibit strikingly similar long-run trends, dropping from the 1920's to the 1930's, growing from 1950 to 1980, then declining after 1981. Inspection of these figures suggest that real balances, output, velocity and interest rates might be well-characterized as being individually integrated, and formal tests support this view. Specifically, the following characterizations appear consistent with the observed series: m-p is both halves of the sample), with drift; r is 1 drift; and (m-p), y and r are cointegrated. Whether p and m are individually 1(1) or 1 1 (1 ) for the full sample (and (1 ) with no drift; y is 1 (1 ) with (2 ) is unclear: the inference depends on the subsample and the test specification. The evidence suggests, but is not conclusive, that r-Ap is 1(0). Because rt is nonnegative, characterizing rt as 1(1) raises conceptual difficulty. This decision is driven by the empirical evidence that r^ exhibits considerable persistence, and is consistent with interest rate specifications used by other researchers (e.g., Campbell and Shiller [1987] and Hoffman and Rasche [1989]). The applicability of the DOLS and DGLS estimators to 1(1) and 1(2) systems 26 13 makes it possible to estimate 6 to test whether 0^-1. in the cointegrating relation, m~0pP’0yy*0rr , and consider three specifications. First, if m and p are 1(1), then (m,p,y,r) constitute the 1(1) system analyzed in Section 2 with one 2 cointegrating vector, modified for nonzero drifts, and inference is x • Second, if m and p are 1(2) and (r,Ap) are not cointegrated, then this is an 1(2) system with y*-pt , yt~<rt yt^’ and yt“mt ’ where *2 ,1 “° ’ *3 ,1 ~0 ’ fl3 ,l“ (*y and 0® l"^p- ’ This is case 3 in Section 5(B), and inference on (0y , 0r , 0^) • Third, if p is 1(2) and r-Ap is 1(0), this is a using DOLS or DGLS is 1 2 - combination of cases 2 and 3 in Section 5(B), with yt~Pt > yt=(A *2,1-(1’0)'’ *3,l“ (*r’0)'’ *3 ,l“V B\ iAyJ-(rt-Apt,Ayt)'. and 0°>2 «(O,0y )'. 1 rt ,yt), Then Ay*- Here, the cointegrating vector (1,-1) is imposed on (rt ,Apt) as implied by the elementary economic hypothesis that the real interest rate is stationary. 2 Again, inference using DOLS or DGLS is x • Estimates for the four-variable system are reported in Table 2 for these three specifications. 14 The estimates of 0^ do not differ from one at the 10% (two- sided) level in any of the specifications. In all but two cases, 0y is statistically indistinguishable from 1 at the 5% level, and in the two exceptions 0y is estimated imprecisely. To be consistent with theory and with the rest of the money demand literature, we henceforth impose the estimation of 0 0 p“l and study in more detail y and 6r . Estimates of Ml demand cointegrating vectors in the system (m-p,y,r) are presented in panel A of Table 3. The estimators are those studied in the Monte Carlo experiment, plus the single-equation nonlinear least squares estimator (NLLS), which is used by Baba, Hendry and Starr (1990) to estimate their long-run Ml demand equation. The full-sample estimates are similar across estimators and none of the efficient estimators reject the hypothesis that 6 ^ 1 at the sided level. 1 0 % 2 - Using only the first half of the sample, the efficient estimators 27 provide smaller income elasticities and larger interest elasticities, but this difference is modest. In sharp contrast to the first-half estimates, the postwar estimates in Table 3 differ greatly across estimators. The SOLS estimate is close to zero, the NLLS elasticities have the "wrong" sign, and the JOH estimator is highly sensitive to the number of lagged first differences used.^ The final set of estimates refer to the system (m-p,y,r ), where r smoothed commercial paper rate. reasons. 16 is the A smoothed interest rate is used for two First, the empirical money demand literature is indecisive on whether a long- or short-term interest rate is most appropriate. Because there is no consistent risk-free long-term rate with constant tax treatment over the full sample, r can be interpreted as a proxy for a long-term rate which, under the risk-neutral theory of the term structure, is an average of current and expected future short-rates. * Second, r can be viewed (indeed is constructed as) an estimate of the permanent component in interest rates. The cointegrating vector relates the permanent components of m-p, y, and r; to the extent that r is a particularly noisy measure of its permanent component, the cointegrating regressions will suffer from a small-sample version of errors-in-variables bias, and using r could reduce this bias. The results in Table 3 are for a two-sided smoother, but they are typical of results for other smoothed rates. sample estimates change only slightly using r . "ic and standard errors are larger with r than r. The full- The postwar income elasticities The JOH and NLLS estimates are* * 17 quite sensitive to using r , and the differences across point estimates remain. The differences between the prewar and postwar estimates raise the possibility that there has been a shift in the long-run money demand relation. To evaluate this (and to explore the source of the instability in the postwar estimates), we examine four related pieces of evidence. The first consists of formal tests of the null hypothesis of a constant cointegrating relation, against the alternative 28 of different cointegrating vectors over 1900-1945 and 1946-1989, under the maintained hypothesis that the parameters describing the short-run relations are constant, using the DOLS estimator. not conclusive. The results, given in panel B of Table 3, are Although two of the four specifications reject constancy at the 10% level, only one rejection is at the 5% level and the shift parameters Sy and $r are imprecisely estimated. Second, 95% confidence regions for (0 , 0r) implied by the point estimates in panel A of Table 3 generally overlap, or nearly overlap, near 0^,-1.00 and 0r«-.ll. These regions are plotted in Figure 2a-2d for, respectively, the DOLS, DGLS, PBSR, and PHAN estimators. 18 For each estimator, the only nonoverlapping region is for the postwar estimator based on r; the major axes of the prewar and postwar ellipses are approximately orthogonal; and the confidence region for the full sample is much smaller than for either half. 19 The third piece of evidence concerns the properties of the cointegrating residuals, zt« mt-0yyt-0rrt . These residuals exhibit quite different properties for the different point estimates: residuals constructed using either the full-sample or first-half point estimates are consistent with cointegration, while the residuals based on the postwar estimates are not. based on the full-sample point estimates, appears stationary. cointegrating vector is estimated over the first half and r second half, asymptotically r h a s distribution: As shown in Table 4, When the is computed over the the standard univariate Dickey-Fuller (1979) using f , non-cointegration is rejected in the postwar data at the 5% one-sided level for the DOLS prewar cointegrating vector. In contrast, the residuals from the postwar money demand equations exhibit more serial correlation but lower variance than the residuals constructed using the prewar or full-sample estimates. Fourth, the postwar VECM likelihood, concentrated to be a function of (0^, 0r), 29 is bimodal for both the r and r data sets. Moreover the JOH point estimates are quite sensitive to the number of nuisance parameters estimated (number of lagged first differences included). Inspection of the concentrated likelihood, plotted in Figure 3 for 3 lags (J0H(3) in Table 3A), indicates two conclusions: that the JOH MLE's for 2 and 3 lags lie on a ridge that corresponds to the major axis of the postwar confidence ellipses in Figure 2, and that the likelihood is not well approximated as a quadratic. This explains, in a mechanical sense, the instability of the JOH estimates with respect to the lag length, and suggests that the JOH estimator might be poorly approximated as normally distributed. These four pieces of evidence lead us to conclude that, despite the apparently large differences in the prewar and postwar point estimates, the evidence against Lucas's (1988) interpretation of a stable long-run money demand relation is weak, and indeed that the best summary of the evidence is that long-run Ml demand has been stable over 1900-1989. imprecisely estimated. Using the postwar data alone, the elasticities are The postwar data is dominated by the 1950-1980 trends in velocity and interest rates; as Lucas (1988) pointed out, this requires the estimates to lie on the "trend line" given by A(m-p) -0yAy-0rAr (where Ay is the average annual growth rate of yt , etc.). This line constitutes the major axis of the postwar confidence ellipses in Figure 2 and the ridge in the postwar VECM likelihood in Figure 3. Several such trend lines (or low frequency movements) can be drawn from the prewar sample, resulting in tighter confidence regions when only the 1900-1945 sample is used. When the 1900-1945 and 1946-1989 subsamples are combined, the 1900-1945 and 1946-1989 trend lines solve for point estimates 0y~l.OOO and 0r«-.145. Because the efficient estimators of cointegrating vectors exploit this same low-frequency information, albeit in a more sophisticated way, the sampling uncertainty of the full-sample estimates is much smaller than that based on the prewar and especially the postwar data. 30 B. R e s u l t s f o r p o s t w a r m o n th ly d a t a . Cointegrating vectors estimated using postwar monthly data on Ml, real personal income, the personal income price deflator, and a variety of interest rates are reported in Table 5, panel A. Compared to the postwar annual results, the income elasticities estimated over 1949:1-1988:6 are higher and there is somewhat less disagreement across the efficient estimators, with income elasticities ranging from .30 to .89 based on the commercial paper rate. The estimates are stable across the choice interest rate (the exception is the DGLS estimates, for which GLS effectively firstdifferences the data, as in the postwar annual estimates). The point estimates agree closely with Baba, Hendry and Starr's (1990) NLLS estimate of .5 obtained over 1960-1988, strikingly so since they used GNP rather than personal income, quarterly rather than monthly data, and several additional regressors designed to account for shifts in short-run money demand relation. 20 Although the point estimates are not sensitive to the start date of the regression, they are quite sensitive to the final regression date. For example, JOH estimates of the income elasticity, estimated over 60:1 to the last month in each quarter from March 1984 through June 1988 using the commercial paper rate ( 8 lags), range from -3.00 to 3.54; for the NLLS estimator, this range is .29 to 1.08. When computed over 60:1-78:12, the JOH, NLLS and DOLS income elasticities are -.27, -.13, and .11. Comparable instability is present for each of the interest rates studied in Table 5, whether estimated in logarithms or in levels. Because we do not provide uniform critical values for tests based on these "recursive" estimates, this observed instability does not provide formal evidence on the stability of the cointegrating vector estimated with the postwar data. This sensitivity to terminal regression dates is, however, consistent with our interpretation of the annual data. Specifically, the data from 1950 to 1982 are 31 dominated by the single upward trend in real balances, income and interest rates, which results in estimated income and interest elasticities that are strongly negatively correlated and are imprecisely estimated, except that they must fall on the trend line. Only with the most recent data, which reflect the second trend (increasing income, declining velocity and interest rates), are the estimates more precise with values that are comparable across estimators. C. Discussion and Summary. This analysis is restricted in several regards. Only one monetary aggregate has been considered, Ml. Much of the money demand literature has focused on the search for a stable short-run demand function, an issue avoided here altogether. The analysis has relied heavily on asymptotic distribution theory to construct formal confidence intervals and tests, and the estimation procedures typically entail the estimation of many nuisance parameters relative to the sample size. Although we have only limited evidence, this leads us to suspect that the precision of the foregoing results is overstated. Even with these caveats, these results suggest three conclusions. First, when viewed over 1900-1989, there appears to be a stable long-run Ml demand function. Estimated over the entire sample, 95% confidence intervals based on the DOLS estimator are, for the income elasticity, (.90, 1.03), and for the interest semielasticity, (-.124, -.088). Similar intervals are obtained using the other efficient estimators over the full sample. Second, our results are consistent with Lucas's (1988) suggestion that there is a stable long-run money demand relation over the pre- and postwar periods. A key piece of evidence for this is the apparent stationarity of the postwar residuals computed using the first-half estimates of the cointegrating vector. Third, in isolation the postwar evidence says little about the parameters of the cointegrating vector: the estimates have large standard errors and moreover 32 are sensitive to the subsample and estimator used. The main reason for this is that the postwar data are dominated by steadily rising income and interest rates and effectively no growth in real balances; only after 1982 is there a decline in interest rates that reduces multicollinearity sufficiently to estimate the money demand relation. We suspect that the postwar standard errors understate the sampling variability, particularly for the monthly results, because of this sensitivity to terminal dates, some evidence that the large-sample mixed normal distribution provides a poor approximation to the postwar sampling distributions, and the presence of this problem in the Monte Carlo analysis in Section 8 6 . . Conclusions The empirical investigation suggests some observations that might apply more generally beyond this particular application. First, the precise estimation of long-run money demand appears to require a long span of data: estimates over the full span are more precise than over the first half of the century alone, and the data since 1946 contain quite limited information about long-run money demand when viewed in isolation. Second, the use of several efficient estimators is a valuable check of the sensitivity of the estimates to changes that should be asymptotically negligible. In the case of postwar money demand, the sensitivity of the postwar estimates to the choice of estimator and to the estimation period drew attention to the low-frequency multicollinearity between postwar income, velocity and interest rates in the postwar data. 33 Appendix Proof of Lemma 2.1. The proof is a modification of Anderson's (1971, Theorem 7.6.7) and Rozanov's (1967, Chapter 2.3) proofs of the Wold decomposition. i—1,2, form Hilbert spaces. 2 Note that Then u^“cn(L)€^ is the Wold representation of 1 1 1 and by construction ct is the innovation process for ut . ^t ® ^t-1* so t^iat and construction Let forms a basis for D^. p<utl*i>“p <utlUs~«Ds)“5 — ' »C2 1 j £t-j“c2 1 Then (L>€t> where c2 1 j“ Now at_ut-p<uti®i)“ut-c 2 i(L)et is ~2 2 stationary, has E(ut) <», and is linearly regular, and so possesses the Wold ~2 decomposition utsssC2 2 2 2 2 2 ^F^ €t ’ where €t~ut-P(ut 1 ^ 1 2 2 <D ^t- 1 ^* - 2 1 2 construction, ct is the innovation process for ut , E€t« 0 , E€tes all t,s, and Ec^c^'-E.^ for t-s a**d equals 0 otherwise. ' = 0 for Finally, c(L) is square summable because Eu^u^'^ by assumption. □ Derivation of (3.1). Assume that the nxl vector yt has the Wold representation A<*yt — A* + F^(L)at , where (i) at is a martingale difference sequence with E(atat '|at ^ ,at_2 »•••) and max^suptE(a^t)<®, (ii) as~0 for s<0, (iii) Zj«oJd |Fjl<00' F ^ L ^ ^ J ^ q Fj L^ , with F^(e’^w ) is nonsingular for t^O, and (v) rank[F^(l) ]=k^<n. The triangular representation (3.1) is constructed by repeated application of the following Lemma: Lemma A.l. Axt - Assume that the nxl vector xt is generated by + F(L)at , where at satisfies (i) and (ii), F(L) is Z- summable and satisfies (iv), and rank[F(l)]=k<n. Without loss of generality arrange xt so that the upper kxn block of F(l) has full row rank. - 34 - Then xt can be represented as: Axt - + D 1 (L)at “ 4 + *xt + D 2 (L)at 1 2 1 2 where xt«(xt ' xt ')', where xt is kxl, xt is (n-k)xl, and D(L)*[D^(L)' D 2 (L)']' is (i-1) summable. Proof. When /im lies in the column space of The result holds trivially for k-n, so consider k<n. Order xt so that F(L) can be partitioned as F(L)=[F^(L)' F2 (L)']' where F^(L) is kxn, F 2 (L) is (nk)xn, and F^(l) has full row rank. for some kxr matrix 6. (A. 1) Because F^(l) has full row rank, F2 (l)-0 'F-^(l) Now partition as (/*£ ^ so that Ax* - 0'AxJ; - S _ o O * 2 ,i-*>l,i>ti + [F2 <L>-,9 'Fl<L)]af Accumulating (A.l) yields x^ - 0'x^ - + ^ 2 ^ ^ at ’ w^ere D 2 (L) - F^CL) - 6 'F*(L) , where F*(L) - (1-L)'1 (Fi(L)-F^l) ) , i-1,2. "fc is H summable, F^(L) is (i-1) summable. then 1*2 m*® so ^ 1 ^ 2 m+ 1 *®* Because Fi(L) If /im lies in the column space of F(l) , T^e Lemma f ° H ° ws by setting / i ^ » (i=0,...,m) and D 1 (L)=F1 (L). □ To construct the triangular representation (3.1), apply Lemma A.l to xt=A^"^yt to yield the decomposition: ad ~ 1 ~ A yt .d-l- A ^d■1 “ n , o + Fi 2 - - /. \ (L)at „ ^ .d-l/Ad-l-lN y t " ^2,0 + 1* 2 , 1 * + ° 2 ,1 ( A + Ff‘1 (L)at y t> ~1 -2 where yt has been partitioned into k^xl and (n-k^)xl components yt and yt . Now assume that F^’^(1)-[F^ ^(1)' F^ ^(1)']' has rank kj^^^n, and apply the - 35 - lemma to xt - [Ad "^y^f (Ad’2 y 2 -0d pAd ~^y\) 1 • Continuing this process yields the triangular representation (3.1), with u^.«Dj(L)at , j- 1 , ...,d+1, where rank [Dj(l)]-kj for j-l,...,d. While £>-[0^(1)' D 2 (l)' ... Dd(l)']' has full row rank by construction, nothing so far ensures that D^+^(l) is linearly independent of the rows of D. completed. If it is, then the construction is If it is not, then redefine the variables yt to be A ^yt and d to be d+1 , and repeat the construction until [6 ' D^+^(l)]' has full rank, so that ut is 1(0) with a full rank spectral density matrix. This yields (3.1) for variables of arbitrary finite orders of integration and cointegration. Proof of Theorem 4.1. First consider the infeasible GLS estimator than $(L) . constructed using $(L) rather -1 , Note that (Tt ®I)(6 q LS-£) - <T ^T “•p, > where Q, (T^®I)Xtztz£(T^®I) and <f>^ - (T^®I)£tztet , with zt - [zt®$(L)'] and e^ - $(L)et . (Unless otherwise stated, the remaining identity matrices have dimension k^ so this subscript is suppressed.) The convergence of Qppp- to follows from a standard application of the weak law of large numbers. with i or j > q1jt - 2 For Qpjp : aii®i)Et[ZSU(4-m ®v ) [ ig .0(4-h ® - < T i^ t[2 -o a -o < 4 -» 4 :h ® - < T ^ i> x t [ 2 ^ I g . 0( 4 4 ’ ® »;% )i(T ji® D + op <i) -> (Vtj ® 0 ^ ) where the last two lines follow from Lemma 1 of Sims, Stock and Watson (1990) (SSW) and $(1) h i For <f>iT, i > 2: 8 - - (T‘^®I)Xt(z^ ® $( 1 )')]Z’^ + Opd) 36 J /i<Gmm ( 1 ) s ( m ' 2 ) / 2 ® *<l),)‘W 2 <s). ">-2,4,6.... 2 i u i (Gmm(1 )Wim "1)/2(s) ® $(l)')dW2 (s), m-3,5,7, ...,2i-l where the last line follows from Lemma 1 of SSW. The joint convergence and distribution of ^ follows from SSW Lemmas 1 and 2. To prove that the feasible GLS estimator has the same limit, let qt - (Tx1®i)lti l % < Zt.m ® *;>n2-o<*t-h ® ^)i'(Ti1®i> so (Tj1®!) (SG h s-S G h s) j=l,...,q. - 0T (^T -^T ) + (0t -Qt)*t - Assume that B for Because Qrj, - Q ^ 0, Q^, - Q 5 0, Q is a.s. invertible, and 0, GLS and feasible GLS are asymptotically equivalent. ^ □ Proof of Theorem 4.2. (a) By assumption, Cj j (L) is d+l-j summable for j-1,2,...,d+l. This implies that the diagonal entries Gjj(L) of G(L) corresponding to the stochastic elements, in from equation (3.7) are j summable. The theorem then follows from Lemma 1 of SSW. □ 1 1 ^ 1 — Theorems 4.1 and 4.2 imply that TlT^tztzt'T*T * 0 • (b) consider the infeasible GLS estimator 6 First GLS, defined in the proof of Theorem 4.1. Theorem 4.1 implies that (T*t ® I)(«*GLS-S*) - B;£[(T;£®I)Xt(zt®S(L)')E‘^ ] where B*x - (T^®I) [£t(z*®$(l) ') (z*'®$(l) ) ]( T ^ I ) . + op (l) Now B*T “ (T;i®IH I(g-gl)ki®$ <1 >'$ <1 > > ^ t (ztzt'®I>l(T*T®I> SO K t - [(T*x®I)Et(z*®I)(z*'®I)(T^^sl)]*1 tl®$(l)'$(1 )]_1. Also, ( t * t®d ( S * o l s - « * ) - [(T ;^® i)X t ( z *® i) ( z *'® i) ( T ;^ ® i) ] * 1 37 x (T;J»I)Xt(z*®I)Cii(L)«^ + op (l) - B*£(I®$(1) '$(1) ) (T;^®I)][t;(z*®I)cii(L)e^ + op (l) - B;^(T;^®i)j;t(z*®$(i)'$(i))cii(L)£^ + opd) . Thus (T*t ®I) (^*OLS”^*GLS^ “ ®*T^ *x®^) '^(1 ) - It<z*®S(L)')2 - ^ } €fc + op (l) - B;J(T;^®i)5;t{(z*®$(i)')$(i)[cXi(i)£^ + c ^ ( l )a c ^] - ( z * ® * d ) ' ) S ^ £t + [z*®($(l)-$(L))']S^} + Opd) “ A£j) + O p d ) -H where the final equality follows from $d)cjjjj(l)-£^, and where ^(D-d-D'^KD-Kl)), c^LHl-D'^c^aj-c^d)), and A 1T - (T;^®I)Xt(z*®$(l)'$(l))c*i(L)A£^ A2t - (T;^®I)Xt[z*®d(l)-$(L))']S^e^ “ -It [(T;^Az*)®$*(L)')S^e^ • Because -> Q* (the (*,*) block of Q given in Theorem 4.1), the result follows if A^t ^ 0 and A 2 ^ ^ 0. Because $*(L) has a finite order by assumption and E(€^| {z*} )*=0, standard telescoping arguments imply that A ^ 5 0. In addition, ^2T ^ 0 as a consequence of the results for □ in Theorem 4.1. Proof of Theorem 4.3. The result follows from Theorem 4.1 above and Theorem 4 of Johansen (1988a) or alternatively from section 4 of Phillips (1988a). □ Proof of Theorem 4.4. This follows directly from Theorem 4.3 and the proof of Theorem 4.2. 38 □ Footnotes 1. Since submitting this paper it has come to our attention that the estimator proposed here was independently developed by Phillips (1988a) (also see Phillips and Loretan [1989]) and Saikkonen (1989). The earliest reference of which we have become aware is Hansen (1988). The current paper extends previous results to higher, differing orders of integration, handles deterministic time trends, and applies the results to the estimation of long-run money demand. 2. Similar results hold in the Gaussian model with explosive roots, see Domowitz and Muus (1988). 3. This lemma has antecedents (but to our knowledge no previous formal statement and proof) in the literature on optimal filtering, for example Whittle (1983) Chapter 5 or Brillinger (1980) Section 8.3, or Sims' (1972) discussion of Granger causality. 4. Although the construction leading to (2.5) makes (2.5) unique, alternative triangular representations exist. For example, it is possible to^construct a one sided triangular^representation analogous to (2.5), except that will not be innovations of u£ and in general c-^(z) will not be invertible. Such a representation was derived by Hansen, Roberds and Sargent (1990) to study restrictions on the consumption and labor income process implied by the balanced budget constraint.5 7 6 5. Johansen (1988b, 1990) studied the restrictions on the coefficients of vector autoregressions implied by the existence of cointegration in higher order systems. Johansen (1988b) examined systems with, in Granger and Engle's (1987) terminology, cointegration of the form CI(d,b), where d^b. As Johansen (1990) points out, this excludes cointegration of the general form (3.4), which generalizes what Granger and Lee [1988] term "multicointegration". Johansen (1990) complements our derivation, since it explicitly handles multicointegration; it relates multicointegration to restrictions on the parameters of the levels VAR, whereas the current derivation refers to the moving average representation of the d-th difference. 6 . 7. This result has recently been provided by Saikkonen (1989) in the d=l case. Equations (5.3) and (5.4) provide an easy way to construct standard errors for ^MLE’ namely from (5.4) or alternatively from the usual NLLS formula from the 2 regression of Ayt onto 1 2 * 1 1 2 yt_^, ^t*"^t‘®MLE^t^ * extends directly to the case of general finite-order A(L) by including lags of Ayt in the regression. As discussed below, an asymptotically equivalent estimator of 6 and its standard error can be constructed by replacing of by any consistent estimator of II. 39 usec* *-n construction 8 . See Section 7 for a description of the data. 9. This^interpretation is supporte^ by jn additional Monte Carlo experiment in which Ay. was replaced by [k (L)k (L’ ;]Ayt . (Of course in an empirical application /c(L) would be unknown.) This eliminates nearly all of the bias: for T-100, the bias falls from .026 to -.006 for DOLS and from .045 to -.008 for DGLS. 10. For Model 3 and T-100, the DOLS bias was reduced to .007. The main effect of doubling the number of leads and lags and order of the autoregression for Models 1 and 2 was to increase the dispersion of the t-statistic; for example, for T-100 t 9 5 -t qc for D0LS1 increased from 3.86 to 4.13. A similar increase in dispersion of the’distribution of JOH t-statistics occurred when the number of lags in the VECM was doubled. 11. Most empirical analyses of money demand predate the literature on cointegration. Exceptions are Hoffman and Rasche (1989), who apply Johansen's (1988a) estimator to monthly U.S. Ml data from 1953 to 1987, and Johansen and Jesulius (1990), who apply Johansen's (1988b) procedure to the long-run demand for money in Denmark and Finland. Baba, Hendry, and Starr (1990) focus on short-run U.S. Ml demand (1960-1988, quarterly), but a preliminary step is their estimation of long-run Ml demand using a single equation error correction model (the "NLLS" estimator). With the same purpose and methodology, Hendry and Ericsson (1990) present results for the U.K. as well as the U.S. 12. Data sources and construction: Ml: 1947-1989: The monthly Citibase Ml series (FM1) was used for 1959-1989; the earlier Ml data was formed by splicing the Ml series reported in Banking and Monetary Statistics, 1941-1970, Board of Governors of the Federal Reserve System to the Citibase data in January 1959. The monthly data were averaged to obtain the quarterly or annual observations. Data prior to 1947 are those used by Lucas; from 1900-1914 the data are from Historical Statistics, series X267 and from 1915-1946 they are from Friedman and Schwartz (1970), pp. 704-718, column 3. Y: U.S. Net National Product 1947-1989, Citibase GNNP. Prior to 1947, Lucas's (1988) data (Friedman and Schwartz real net national product (1982 dollars), Table 4.8). For the monthly data 1959-1989, we used personal income (GMPY). P: Price deflator for NNP. 1947-1989, Citibase GDNNP. Prior to 1947, Lucas's (1988) data -- same source as NNP. For monthly data we used the price deflator for GMPY. Interest rates: Commercial Paper rate. 19471989, 6 -month commercial paper, Federal Reserve Board (FYCP), prior to 1947, Lucas's (1988) data (Friedman and Schwartz (1982), Table 4.8, column 6 ).1 3 13. Univariate Dickey-Fuller (1979) r and ?r statistics, computed with 2 and 4 lags on the full data set, fail to reject a single unit root in each of m, p, y, r, m-p, and log velocity at the 1 0 % level; the unit root hypothesis is not rejected for y with 4 lags, but is rejected at the 10% (but not 5%) level with 2 lags. A unit root in Ay, Ar, and A(m-p) are each rejected at the 1% level. Similar inferences obtain when the sample is split 1900-1945, 1946-1989. Whether m and p have two unit roots is less clear: for m, two unit roots are rejected in favor of one at the 5% level for both subsamples, but not the full sample, while the reverse is true for p. For r-Ap (Ap in percentages), one unit root is rejected (vs. zero) for the full sample at the 1 0 % level, but not in either subsample using the rT statistic (f rejects at 1 0 % in both subsamples). 40 The Stock-Watson (1988) qr(3,l) statistic, applied to the system (m-p,y,r) over the full sample, rejects the hypothesis of three unit roots in favor of one unit root at the 5% level over the full sample ^ith 1-4 lags. The evidence on three vs. two unit roots is less strong: the q*(3,2) statistic (2 lags) has a p-value of .33. However, the Engle-Granger (1987) augmented Dickey Fuller test based on the residual from regressing m-p on y and r (with a constant and time trend in the regression and using the appropriate critical values for a trivariate detrended system, two lags) rejects non-cointegration at the 5% level over the full sample. The details are available from the authors on request. 14. The dates in Table 2 and henceforth refer to the dates over which regressions are run; earlier and later observations are used as initial and terminal conditions as needed. 15. Likelihood ratio tests of 2 vs. 3 lags in the VECM are, respectively, 12.08, 11.12, and 3^.88 over the periods 1904-1986, 1904-1945, and 1946-1986. With asymptotic Xg distributions, these suggest specifying p-3 over 1946-1986. 16. The smoothed interesj rate was constructed to be the two-sided estimate of the germanent^component r. calculated using the Kalman smoother for the model r^»rt+/i^t:, Ar^/i^, ^lt*^2t^ independent and var(/i^t)/var(|t2 t)“ 3 . Other filters that yield similar results are a one-sided exponentially weighted moving average filter with coefficient .95 and the Hodrick-Prescott filter. 17. The results in Table 3 are robust to changes in the details of the computation of each of the estimators, in particular: using a Bartlett kernel with 7 lags for PBSR and PHFM, using 3 rather than 2 leads/lags for DOLS and DGLS, using 1 or 2 rather than 3 lags for JOH. The only exception is the postwar instability of the JOH estimates, discussed in more detail below. 18. Because Wald statistics testing hypotheses about (0 , 0r) using the efficient estimators have large-sample x distributions, the usual*7approach can be used to construct confidence regions for (0 , 0r) . Asymptotically the estimators in the two subsamples are independent, but*7for the small samples considered here, the short-run dependence in the data, the presence of initial and terminal leads and lags, and possible deviations from the large-sample mixed normal distribution will result in a lack of independence. 19. The anomalous region is the postwar unsmoothed interest rate region for DGLS. This is best understood by noting that the estimated GLS transformation for DGLS approximately differenced the data (the estimated AR(2) filter is 1-1.39L+.41L ;, so that the DGLS point estimates are in effect determined by covariances between first differences of the data, not their levels. 20. Using the JOH estimator with monthly data on log real personal income, log real Ml, and the logarithm of the 90-day Treasury bill rate, 1953-1988 (3 lags), Hoffman and Rasche [1989] estimate the income elasticity to be .78; the difference between their estimate and the corresponding value from Table 5 (.462) arises from our use of levels, not logarithms, of interest rates. 41 References Ahn, S.K. and G.C. Reinsel (1990), "Estimation for Partially Nonstationary Autoregressive Models," Journal of the American Statistical Association, 85, 813-823. Anderson, T.W. (1971), The Statistical Analysis of Time Series. Wiley: New York. Baba, Y. , D.F. Hendry, and R.M. Starr (1990), "The Demand for Ml in the USA, 19601988," manuscript, Department of Economics, University of California, San Diego (revision of University of California, San Diego Working Paper # 8 8 -8 , "U.S. Money Demand, 1960-84," December 1987). Berk, K.N. (1974), "Consistent Autoregressive Spectral Estimates," Annals of Statistics, 2, 489-502. Bewley, R.A. (1979), "The Direct Estimation of the Equilibrium Response in a Linear Model," Economic Letters, 3, 357-361. Brillinger, D.R. (1980), Time Series, Data Analysis and Theory, Expanded Edition, Holden-Day, San Francisco. Campbell, J.Y. (1987), "Does Savings Anticipate Labor Income? An Alternative Test of the Permanent Income Hypothesis," Econometrica, 55, 1249-1274. Campbell, J.Y. and R.J. Shiller (1987), "Cointegration and Tests of Present Value Models," Journal of Political Economy, 95, 1062-1088. Campbell, J.Y. and R.J. Shiller (1989), "The Dividend-Price Ratio and Expectations of Future Dividends and Discount Factors," The Review of Financial Studies, 1, 195-228. Dickey, D.A. and W.A. Fuller (1979), "Distribution of the Estimators for Autoregressive Time Series with a Unit Root," Journal of the American Statistical Association, 74, 427-431. Domowitz, I.H. and L. Muus (1988), "Likelihood Inference in the Nonlinear Regression Model with Explosive Linear Dynamics," in W. Barnett, E. Berndt and H. White (eds), Dynamic Econometric Modelling, Cambridge University Press, Cambridge. Engle, R.F. and C.W.J. Granger (1987), "Cointegration and Error Correction: Representation, Estimation, and Testing," Econometrica, 55, 251-276. Engle, R.F., D.F. Hendry, and J.F. Richard (1983), "Exogeneity," Econometrica 51, no. 2, 277-304. Friedman, M. and A.J. Schwartz (1970), Monetary Statistics of the United States. New York: Columbia University Press for the National Bureau of Economic Research. 42 Friedman, M. and A.J. Schwartz (1982), Monetary Trends in the United States and the United Kingdom. Chicago: Unversity of Chicago Press for the National Bureau of Economic Research. Goldfeld, S.M. and D.E. Sichel (1990), " The Demand for Money," Chapter 8 in B.M. Friedman and F.H. Hahn (eds.), Handbook of Monetary Economics, vol. 1 , Amsterdam: North-Holland, 299-356. Granger, C.W.J. and T-H. Lee (1988), "Multicointegration," Discussion Paper #24, Department of Economics, University of California, San Diego. Grenander, U. and M. Rosenblatt (1957), Statistical Analysis of Stationary Time Series, John Wiley and Sons: New York. Hansen, B.E. (1988), "Robust Inference in General Models of Cointegration," manuscript, Yale University. Hansen, B.E. (1989), "Efficient Estimation of Cointegrating Vectors in the Presence of Deterministic Trends," manuscript, University of Rochester. Hansen, B.E. and P.C.B. Phillips (1988), "Estimation and Inference in Models of Cointegration: A Simulation Study," Cowles Foundation Discussion Paper No. 881. Hansen, L.P., W. Roberds, and T.J. Sargent (1987), "Time Series Implications of Present Value Budget Balance and of Martingale Models of Consumption and Taxes," manuscript, Department of Economics, University of Chicago. Hendry, D.F. and N.R. Ericsson (1990), "Modeling the Demand for Narrow Money in the United Kingdom and the United States," International Finance Discussion Paper no. 383, Board of Governors of the Federal Reserve System, Washington, D.C. Hoffman, D. and R.H. Rasche (1989), "Long-Run Income and Interest Elasticities of Money Demand in the United States," NBER Working Paper no. 2949. Johansen, S. (1988a), "Statistical Analysis of Cointegration Vectors," Journal of Economic Dynamics and Control, 12, 231-255. 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. (1989), "Estimation and Hypothesis Testing of Cointegrating Vectors in Gaussian Vector Autoregression Models," manuscript, Institute for Mathematical Statistics, University of Copenhagen. Johansen, S. (1990), "A Representation of Vector Autoregressive Processes Integrated of Order 2," Preprint 1990 no. 3, Institute of Mathematical Statistics, University of Copenhagen. 43 Johansen, S. and K. Juselius (1990), "Maximum Likelihood Estimation and Inference on Cointegration -- with Applications to the Demand for Money," Oxford Bulletin of Economics and Statistics, 52, no. 2, 169-210. Judd, John P. and J.L. Scadding (1982), "The Search for a Stable Demand Function: A Survey of the Post-1973 Literature," Journal of Economic Literature, Vol. XX, pp. 993-1023. King, R . , C. Plosser, J.H. Stock, and M.W. Watson (1987), "Stochastic Trends and Economic Fluctuations," NBER Discussion Paper No. 2229; forthcoming, American Economic Review. Lahiri, K. and P. Schmidt (1978), "On the Estimation of Triangular Structural Systems," Econometrica 46, 1217-1222. Laidler, D.E.W. (1977), The Demand for Money: Theories and Empirical Evidence. New York: Dun-Donnelly. Lucas, R.E. (1988), "Money Demand in the United States: A Quantitative Review," Camegie-Rochester Conference Series on Public Policy, 29, 137-168. Meltzer, A.H., "The Demand for Money: The Evidence from the Time Series," Journal of Political Economy, 71, 219-246. Phillips, P.C.B. (1988a), "Optimal Inference in Cointegrated Systems," Cowles Foundation Discussion Paper No. 8 6 6 (revised August 1989). Phillips, P.C.B. (1988b), "Spectral Regression for Cointegrated Time Series," Cowles Foundation Discussion Paper No. 872. Phillips, P.C.B. and B.E. Hansen (1989), "Statistical Inference in Instrumental Variables Regression with 1(1) Processes," forthcoming, Review of Economic Studies. 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 J.Y. Park (1986), "Asymptotic Equivalence of OLS and GLS in Regression with Integrated Regressors," Cowles Foundation Discussion Paper No. 802. Phillips, P.C.B and P. Perron (1988) " Testing for a Unit Roots in a Time Series Regression," Biometrika, 75, 335-346. Rozanov, Y. (1967), Stationary Random Processes. San Francisco: Holden Day. Said, S.E. and D.A. Dickey (1984), "Testing for Unit Roots in AutoregressiveMoving Average Models of Unknown Order," Biometrika, 71, 599-608. Saikkonen, P. (1989), "Asymptotically Efficient Estimation of Cointegrating 44 Regressions,” manuscript, Department of Statistics, University of Helsinki. Sims, C.A. (1972), "Money, Income and Causality," American Economic Review. 62, 540-552. Sims, C.A., J.H. Stock, and M.W. Watson (1990), "Inference in Linear Time Series Models with Some Unit Roots," Econometrics, Vol. 58, No. 1.. Stock, J.H. (1987), "Asymptotic Properties of Least Squares Estimators of Cointegrating Vectors," Econometrics, 55, 1035-1056. Stock, J.H. and M.W. Watson (1988), "Testing for Common Trends," Joumsl of the American Ststisticsl Associstion, 83, 1097-1107. West, K.D. (1988), "Asymptotic Normality when Regressors Have a Unit Root," Econometrics, 56, 1397-1418. Whittle, P. (1983), "Prediction and Regulation by Least Squares," 2nd Edition, revised, University of Minnesota Press, Minneapolis. 45 Table 1 Monte Carlo Results t: :]• vC a A. T=100 Estimator Bias(0) a(0) T-300 t t .05 P(W>3 84) 1 .95 Bias(0) P(W>3..84) .000 .021 -1.67 1.68 054 .000 .007 -1.63 1.70 .052 .000 .023 -1.86 1.87 083 .000 .007 -1.69 1.79 .062 DOLS2 .000 .023 -1.87 1.86 087 .000 .007 -1.66 1.78 .061 DGLS .000 .024 -1.80 1.76 073 .000 .007 -1.64 1.75 .056 PBSR .000 .021 -1.78 1.81 073 .000 .007 -1.67 1.76 .060 PHFM .000 .022 -1.88 1.88 086 .000 .007 -1.71 2.81 .065 JOH .000 .025 -1.98 1.96 077 .000 .007 -1.84 1.67 .057 SOLS 0.8 '.95 SOLS T=100, *u * * *21 ::]■ 0. 0 \05 DOLS1 B. 21 <r(0) ft 11 DOLS1 bias(0) t DOLS2 t t .95 .05 1 :* ■ € L*5 5' 1 DGLS t .95 '.05 PBSR *.95 '.05 PHFM t .95 '.05 JOH t .95 '.05 *. 95 -.90 .084 -1.80 1.84 -1.84 1.84 -1.77 1.77 -1.46 1.69 -1.06 2.98 -1.95 1.83 -.80 .092 -1.81 1.84 -1.85 1.86 -1.77 1.76 -1.49 1.74 -1.17 2.74 -1.95 1.83 -.70 .089 -1.82 1.84 -1.84 1.85 -1.77 1.76 -1.52 1.77 -1.25 2.55 -1.95 1.83 -.60 .081 -1.83 1.84 -1.85 1.84 -1.77 1.76 -1.56 1.78 -1.31 2.40 -1.94 1.84 -.50 .071 -1.83 1.84 -1.84 1.84 -1.77 1.76 -1.58 1.79 -1.38 2.32 -1.95 1.86 .00 .026 -1.85 1.83 -1.86 1.83 -1.77 1.77 -1.76 1.72 -1.67 2.05 -1.97 1.90 .50 .000 -1.87 1.89 -1.90 1.87 -1.80 1.77 -2.05 1.46 -2.01 1.65 -1.99 2.01 .60 -.002 -1.88 1.89 -1.89 1.89 -1.81 1.80 -2.15 1.34 -2.11 1.55 -1.97 2.01 .70 -.003 -1.88 1.91 -1.91 1.91 -1.82 1.82 -2.25 1.25 -2.21 1.42 -1.96 2.05 .80 -.003 -1.92 1.94 -1.91 1.94 -1.81 1.82 -2.36 1.13 -2.33 1.28 -1.97 2.04 .90 -.002 -1.90 1.93 -1.93 1.94 -1.85 1.83 -2.45 1.04 -2.42 1.15 -1.99 2.02 -.90 -.283 -1.80 1.84 -1.84 1.84 -1.77 1.77 -0.58 1.23 -3.80 0.69 -1.90 1.79 -.80 -.078 -1.81 1.84 -1.85 1.85 -1.77 1.76 -0.77 1.46 -1.80 1.34 -1.90 1.79 -.70 .007 -1.82 1.84 -1.84 1.85 -1.77 1.75 -0.86 1.62 -1.31 1.75 -1.91 1.79 -.60 .048 -1.83 1.84 -1.85 1.84 -1.77 1.76 -0.94 1.73 -1.15 2.00 -1.91 1.79 -.50 .068 -1.83 1.84 -1.84 1.84 -1.77 1.76 -0.98 1.82 -1.09 2.13 -1.92 1.79 .00 .065 -1.85 1.83 -1.86 1.83 -1.77 1.77 -1.08 2.08 -1.09 2.28 -1.97 1.80 .50 .028 -1.87 1.89 -1.89 1.87 -1.80 1.77 -1.14 2.18 -1.16 2.30 -2.00 1.84 1.85 .60 .021 -1.88 1.89 -1.89 1.89 -1.81 1.80 -1.15 2.20 -1.18 2.30 -2.01 .70 .015 -1.88 1.91 -1.91 1.91 -1.82 1.82 -1.17 2.19 -1.19 2.31 -2.00 1.87 .80 .010 -1.92 1.94 -1.91 1.94 -1.81 1.82 -1.24 2.16 -1.23 2.30 -2.03 1.84 .90 .005 -1.90 1.93 -1.93 1.94 -1.85 1.83 -1.45 2.18 -1.38 2.36 -1.98 1.87 .951 -.039 [-.062 , .643 Z * € .499 .499' 1.374 T=100 T-300 P(W>3. Bias(0) P(W>3.84) Estimator Bias(0) SOLS .085 .120 -1.95 5.16 .466 .033 .045 -1.90 5.29 .483 DOLS1 .026 .125 -2.10 2.71 .188 .007 .041 -1.79 2.32 .118 DOLS2 .026 .125 -1.72 2.25 .111 .007 .040 -1.55 1.97 .071 DGLS .045 .131 -1.52 2.35 .111 .012 .042 -1.43 2.08 .076 PBSR .039 .123 -1.83 2.79 .180 .012 .041 -1.64 2.41 .122 PHFM .041 .122 -1.91 3.01 .206 .011 .041 -1.69 2.46 .131 JOH .003 .330 -2.40 2.07 .095 -.001 .044 -1.97 1.75 .064 < r(6 ) '.05 '.95 <r(9) '.05 '.95 Notes to Table 1: Bias(0) and a(d) are the Monte Carlo bias and standard deviation of 0, respectively. t qc and t ^ 5 are the empirical 5% and 95% critical values of the t-ratios, and P(ft>3.84) is the percent rejections at the asymptotic 5% level of the test statistic testing 0-0q which, for all but JOH, is the square of the tstatistic, and for JOH is the likelihood ratio statistic. 5000 Monte Carlo replications were used. The number of observations (100 and 300) refer to the span of the regressions; additional observations were used for initial and terminal conditions. All regression include a constant term. The estimators are: SOLS 1 2 Static OLS regression of yt on yt . D0LS1 Dynamic OLS regression of yt on (yt ,Ayt ,Ayt+^,...,Ayt+k ) , 1 2 where k-2 for T-100, k—3 for T-300. 2 2 2 The covariance matrix is estimated by averaging the first k error autocovariances using the Bartlett kernel, where k-5 for T-100, k-8 for T-300. D0LS2 1 2 2 2 2 Dynamic OLS regression of yt on (yt ,Ayt ,Ayt±1,...,Ayt±k), where k-2 for T-100, k-3 for T-300. The covariance matrix is estimated by an autoregressive spectral estimator with 2 lags for T-100, 3 lags for T-300. DGLS 1 2 2 2 2 Dynamic GLS regression of yt on (yt ,Ayt ,Ayt+^,...,Ayt+^ ) , where k-2 for T-100, k-3 for T-300. The errors were modeled as an AR(2) for T-100 and AR(3) for T-300. PBSR Phillips (1988b) band spectral regression, where the spectral density at frequency zero was estimated using the Bartlett kernel with 5 lead/lags for T-100 and 8 lead/lags for T-300. PHFM Phillips-Hansen (1989) fully modified estimator using the Bartlett kernel with 5 lead/lags for T-100 and 8 lead/lags for T-300. JOH Johansen (1988a) VECM MLE based on the estimated model yt £k^A ^ A y t ^ + at , where k-4 for T-100 and k-6 for T-300. 7 a'yt ^ + Standard errors were computed using the formulas given in Section 5(A). The DOLS and DGLS standard errors were computed using a degrees-of-freedom adjustment, specifically df-number of periods in the regression - number of regressors in the DGLS or DOLS regression - number of autoregressive lags in the GLS transform (DGLS) or AR spectral estimator (DOLS). The JOH standard errors were computed as described in Section 5(A) with a degrees-of-freedom adjustment (df-number of periods in the regression - number of regressors in a single equation of the VECM). The degrees-of-freedom corrections are motivated by analogy to the classical linear regression model. No such adjustments were made for PBSR or PHFM. Table 2 Estimated Cointegrating Relations: mt - a + 0ppt + 0yyt + ^rrt; Specifications: I. II. For p t 1(1): mt - /i+0ppt+0yyt+0rrt+dp (L)Apt+dy (L)Ayt+dr (L)Art+et For pt 1(2) and r, Apt not cointegrated: mt " ^+% pt+*yyt+*rrt+dp(L)A2 Pt+dy (L>Ayt+dr (L>Art+et III. For pt 1(2) and r-Apt 1(0): mt - 'i+V t +V t +*rrt+dp (L)A2 pt+dy (L)Ayt+dr (L)(rfc-Apt)+e t Estimates (Standard Errors) Specification Estimator I II III Period no . leads/lags °y DOLS 1903-87 2 DOLS 1904-86 3 DGLS 1903-87 2 1 . 0 0 0 DGLS 1904-86 3 DOLS 1904-87 2 DOLS 1905-86 3 DGLS 1904-87 2 DGLS 1905-86 3 DOLS 1904-87 2 1.143 (.185) 1.205 •(-177) (.213) 1.219 (.152) .838 (.154) .794 (.145) .322 (.289) .798 (.125) - .119 (.016) - .128 (.014) - .042 (.019) -.133 (.013) 1.183 (.190) 1.304 (.2 0 0 ) 1.041 (.166) 1.292 (.180) .820 (.159) .732 (.165) .932 (.143) .763 (.147) -.119 (.016) -.132 (.016) -. 1 0 0 (.016) - .134 (.016) .949 (.138) .887 (.118) .355 (.289) .842 (.103) -.096 (.015) -.106 (.013) -.024 (.017) -.115 (.0 1 1 ) 1 . 0 1 1 (.165) Notes: DOLS 1905-86 3 DGLS 1904-87 2 DGLS 1905-86 3 1 . 1 0 0 (.145) .982 (.209) 1.180 (.128) di(L)“Zj— kdijL^ , where k is the number of leads/lags listed in the third column. Standard errors are in parentheses . An AR(2) < error process was used to implement the GLS transformation for the ]DGLS estimator and to estimate the DOLS covariance matrix when k=2, and an AR(3) was used for k«3. The shorter regression periods for k-3 in panel B relative to k~2 in panel B, and for k=2 in panel A relative to k=2 in Panel B, allow for necessary initial and terminal conditions (leads and lags). Table 3 Honey Demand Cointegrating Vectors: Estimates and Tests, Annual Data Dynamic OLS/GLS estimation equation: A. mt-pt * A* + ^y^t + ^rrt + dy(L)Ayt + dr (L)Art + et Point Estimates (standard errors) 1903 -1945 1903 -1987 Estimator 1946-1987 e #y #y e y r #y 1946 -1987 0 r* SOLS .929 -.083 .916 -.089 .193 -.016 .412 -.046 NLLS .898 -.093 1.104 -.093 -.445 .084 .298 -.023 DOLS .965 (.031) -.106 (.009) .859 (.142) -.117 (.028) .270 (.213) -.027 (.025) .413 (.320) - .047 (.042) DGLS .960 (.037) -. 1 0 0 (.0 1 0 ) .972 (.170) .945 (.308) - . 0 2 0 (.030) (.009) 1.171 (.132) - .091 (.013) .903 (.103) -.103 (.018) .216 (.091) -. 0 2 0 (.0 1 1 ) .367 (.147) -.042 (.019) (.008) .911 (.082) -. 1 0 2 (.015) .205 (.054) - .018 (.006) .393 (.106) - .045 (.014) .971 (.031) -.109 (.009) .878 (.094) 35.588 (1787.6) -5.088 (266.0) -2.344 (4.581) .340 (.645) .976 (.030) -.115 (.009) .940 (.119) -.473 (.390) .075 (.055) - .131 (.274) .033 (.039) .960 (.033) PBSR .956 (.032) PHFM J0H(2) J0H(3) B. Interest rate - . 1 0 1 (.009) - . 1 0 0 - . 1 0 0 - . 1 1 1 (.018) - . 1 2 1 (.0 2 2 ) Tests for Breaks in the Cointegrating Vector Based on DOLS , break date - 1946 no. leads/ lags r 2 3 ]r* 2 3 Wald statistic (p-value) Point estimates (standard errors) ey 6.05 (.05) .969 (.107) 5.03 (.08) .983 (.103) 2.90 (.24) 3.82 (.15) •t S ) -.452 (.269) .061 (.029) ) -.446 (.280) .056 (.027) .862 (.108) -.141 (.025) -.197 (.352) .068 (.048) .862 (.099) -.144 (.023) -.285 (.320) .076 (.043) - . 1 1 1 (.0 2 2 - . 1 1 1 (.0 2 2 Notes to Table 3: Panel A: NLLS is the nonlinear least squares estimator; the other estimators are defined in the notes to Table 1 (DOLS here and in subsequent tables is D0LS2 in Table 1). lagged first differences. JOH(k) is the JOH estimator evaluated using k J0H(3) was computed over regression dates 1904- 1986, 1904-1945, and 1946-1986. on (m-p)t l , yt.1 , rt_1( and 2 For the NLLS estimator, A(m-p)t is regressed lags each of A(m-p)t_1 , Ayt_1( and A r ^ ; and $r are estimated from the coefficients on the lagged levels. 9 DOLS and DGLS used 2 leads and lags of the first differences in the regressions and an AR(2) process for the error. The frequency zero spectral estimators required for PBSR and PHFM were computed using a Bartlett kernel with 5 lags. All regressions included a constant. Panel B: The statistics are based on the regression, (m-p)t«fi+0yyt.+0rrt+ iy(yt-yr)l(t>r)+6 r (rt-rr)l(t>r)+dy(L)Ayt+dr (L)Art , where 1(*) is the indicator function and dy(L) and dr (L) have the number of leads and lags stated in the second column. 1904-1986. * 2 Regressions with k*=2 were run over 1903-1987, with k=3, over The Wald distribution. The covariance matrix was computed using and AR(2) spectral estimator. statistic tests the hypothesis that 8 -$r-0 and has a Table 4 Properties of Error Corrections Terms Esti -- 1904-86 — Estimation — zt “ mt 1904-45 — rrt — 1946-86 — mator Period DOLS 1903-87 0.965 -0.106 -4.646 0.397 0.156 -3.618 0.334 0.136 -3.496 0.360 0.172 DGLS 1903-87 0.960 -0.100 -4.542 0.411 0.151 -3.685 0.301 0.133 -3.314 0.407 0.166 JOH 1903-87 0.971 -0.109 -4.673 0.396 0.159 -3.549 0.361 0.139 -3.553 0.347 0.176 DOLS 1903-45 0.859 -0.116 -3.289 0.674 0.195 -3.654 0.282 0.137 -3.211 0.456 0.192 DGLS 1903-45 0.972 -0.100 -4.531 0.409 0.151 -3.667 0.312 0.134 -3.215 0.432 0.167 JOH 1903-45 0.878 -0.111 -3.592 0.617 0.179 -3.688 0.271 0.134 -3.438 0.389 0.180 DOLS 1946-87 0.270 -0.027 -1.512 0.972 0.468 1.136 1.065 0.347 -3.134 0.463 0.050 DGLS 1946-87 0.945 -0.020 -1.047 0.951 0.240 -1.114 0.848 0.187 -1.526 0.970 0.279 JOH(2) 1946-87 35.588 -5.088 -1.541 0.958 23.401 0.222 1.009 19.997 -3.377 0.444 9.014 JOH(3) 1946-86 -0.473 0.075 -1.327 0.978 1.017 -1.453 0.964 1.017 -0.078 0.998 1.017 DOLS* 1946-87 0.413 -0.047 -1.331 0.968 0.379 0.656 1.052 0.279 -2.969 0.664 0.042 DGLS* 1946-87 1.170 -0.091 -1.740 0.897 0.207 -2.898 0.520 0.157 -1.641 0.938 0.173 JOH(2)* 1946-87 ★ JOH(3) 1946-86 Notes: 8 y 8 r f p » t P & f p a -2.344 0.340 -1.583 0.986 2.152 1.607 1.038 1.761 -2.542 0.811 0.257 -0.131 0.033 -1.447 0.980 0.773 -1.446 0.964 0.773 0.201 1.001 0.773 The point estimates for the indicated estimator and sample period are taken from Table 3. DOLS*, DGLS*, and JOH* refer to these estimators evaluated using the smoothed interest rate rt . A The summary statistics f^, p, and a are respectively the Dickey-Fuller t-statistic testing p—1 with a constant and 3 lags in the autoregression, the sum of the autoregressive coefficients in the regression of zt on a constant and 3 lags, and the standard devation of z^.. The reported entries are these statistics, computed for z^ constructed using the the point estimates in the first columns for each row, with regressions run (and statistics computed) over the subsample given in the column heading. Table 5 Honey Demand Cointegrating Vectors: Period: 49:1 - 88:6 Interest rate: Coon. Paper Estimator e e y 49:1 - 88:6 Conm. Paper e r e y r Estimates, Monthly Data 60:1 - 88:6 60:1 - 88:6 Coon. Paper 90-day T-bill e y e r e y e r 60:1 - 88:6 10-yr T-bond 9 9 y r SOLS .272 -.016 .398 -.035 .339 -.017 .362 -.021 .480 -.031 NLLS .539 -.044 .259 .034 .570 -.030 -.483 -.026 .353 .012 -.044 .398 -.027 .415 -.030 .529 -.037 DOLS .326 -.025 .457 (.187) (.026) (.136) (.019) (.208) (.025) (.165) (.020) (.210) (.023) .889 -.008 DGLS (.203) (.003) .302 -.021 PBSR (.037) (.005) .302 -.021 PHFM (.033) (.004) .561 -.068 JOH (.199) Notes: (.032) NLLS and JOH DGLS used 8 .525 -.026 1.139 -.009 1.195 -.011 1.046 -.019 (.109) (.007) (.289) (.003) (.268) (.003) (.173) (.003) .404 -.036 .367 -.022 .389 (.045) (.006) (.053) (.005) (.049) (.005) .370 -.022 .393 (.045) -.037 .412 (.042) (.006) .629 (.129) used 8 (.048) (.004) -.076 .520 -.075 (.020) (.202) (.039) .462 (.137) -.025 .500 -.034 (.052) (.005) -.025 .511 -.035 (.004) (.047) (.005) -.060 (.024) .631 (.144) -.067 (.021) lagged differences of the variables; DOLS and leads and lags of the first differences in the regressions. An AR( 6 ) error was assumed for DGLS and for the calculation of the standard errors for DOLS. The frequency zero spectral estimators required for PBSR and PHFM were computed using a Bartlett kernel with 18 (monthly) lags. regressions included a constant. All Figure 1 U.S. real net national product (solid line) and real Ml, 1900-1989 A. 1 300 B. 13 10 1320 1330 1 3 H0 1350 1360 1370 1380 U.S. short-term commercial paper rate (solid line; left scale) and the logarithm of Ml velocity, 1900-1989 2.0 1 . 8 1 . 6 1 .9 1 .2 1 .0 .8 . 6 1900 1908 1916 192H 1932 19HQ 1998 1956 1969 1972 1980 1988 Figure 2. 95% confidence regions for the income elasticity 9 and the interest semielasticity 9r , estimated over 1903-1987 (solid line), 1903-1945 (dashes), 1946-1987 (short dashes), and, using the smoothed interest rate r , 1946-1987 (dash-dots), based on the D0LS, DGLS, PBSR, and PHFM estimators. DOLS B. DGLS 0 .0 0 -0 .0 4 -0 .0 8 -0 .1 2 -0 .1 6 -0 .2 0 Interest s e m ie la stic ity 0 .0 4 A. 0 .0 0 -0.0 4 -0 .0 8 -0.1 2 -0 .1 6 Interest s e m ie la sticity 0 .0 4 C. PBSR D. PHFM Figure 3. Concentrated vector error correction model (VECM(3)) likelihood surface in (0 , 0r) space, 1946-1986