The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series Estimating Deterministic Trends in the Presence of Serially Correlated Errors Eugene Canjels and Mark W. Watson Working Papers Series Macroeconomic Issues Research Department Federal Reserve Bank of Chicago September (WP-94-19) E s tim a tin g D ete rm in istic T re n d s in the Presence o f S erially C orrela ted E rro rs E u g e n e Canjels D e p a r t m e n t of Economics, Northwestern University M a r k W . Watson* D e p a r t m e n t of Economics, Northwestern University a n d Federal Reserve B a n k of Chicago A u g u s t 1994 Abstract This paper studies the problems of estimation and inference in the linear trend model: y t = a + fit + u t , where u t is follows an autoregressive process with largest root p, and /? is the parameter of interest. We contrast asymp totic results for the cases \p\ < 1 and p = 1 and argue that the most useful asymptotic approximations obtain from modeling p as local-to-unity. Asymp totic distributions are derived for the OLS, first-difference, infeasible GLS and three feasible GLS estimators. These distributions depend on the local-to-unity parameter and a parameter that governs the variance of the initial error term, ac. The feasible Cochrane-Orcutt estimator has poor properties, and the fea sible Prais-Winsten estimator is the prefered estimator unless the researcher has sharp a p r i o r i knowledge about p and k . The paper develops methods for constructing confidence intervals for (3 that account for uncertainty in p and k . We use these results to estimate growth rates for real per capita G DP in 128 countries. JEL: C22,040 •This is a substantially revised and expanded version on WA Note on Estimating Deterministic Trends in the Presence ofSerially Correlated Errors” by Watson. We thank Sergio Rebelo forposing a question that motivated this research. We have benefited from discussions and comments by two referees, Edwin Denson, Michael Horvath, Robert King, James MacKinnon, Gauri Prakash and James Stock. Financial support was provided by the National Science Foundation through grants SES-91-22463 and SBR-94-09629. 1 1 Introduction Many economic time series display clear trends well represented by deterministic linear or exponential functions of time. The slope of the trend function represents the average growth in the series (or rate of growth, if the series is in logarithms) and is often a parameter of primary interest. Serial correlation in the data complicates efficient estimation and statistical inference about the trend function, and this paper studies trend estimation and inference when this problem is severe. To be specific, assume that a series can be represented as. yt = cx + 0 t + u t (1 ~ p L)u t = vt (1) (2) where y t is the level or log-level of the series, and u t denotes the deviations of the series from trend. These deviations are serially correlated, with a largest autoregressive root of p . The error term v t is an 1(0) process. If the u'ts are jointly normally distributed, and the precise pattern of serial correlation is known, then efficient estimators of a and 0 can be constructed by GLS, and statistical inference can be conducted using standard regression procedures. In practise, the distribution of the errors and the pattern of serial correlation is unknown, so that GLS estimation and exact inference are infeasible. Applied researchers typically use one of three feasible estimators, motivated by the asymptotic equivalence of these estimators to the infeasible GLS estimator. If \p\ < 1,so that u t is 1 ( 0 ), then the feasible GLS estimator is asymptotically equivalent to the infeasible GLS estimator, under general conditions. Moreover, the classic result of Grenander and Rosenblatt (1957) implies that the OLS estimators of a and 0 are asymptotically equivalent to the GLS estimators. Thus, if u t is 7(0), OLS or feasible GLS applied to the level of y t is asymptotically efficient. On the other hand, when p = 1, so that ut is 7(1), a can no longer be consistently estimated by any method, and the OLS estimator of 0 is no longer asymptotically efficient. In this case, the data should be differenced and the Grenander and Rosenblatt result implies that the sample mean of A y t (the OLS estimator of 0 in the differenced regression) is asymptotically equivalent to the efficient, but infeasible, GLS estimator of 0 . In summary, if u t is 7(0) then OLS from the levels regression produces the asymptotically efficient estimator, while if tt* is 7(1) then the sample mean of A y t is the asymptotically efficient estimator. Inference is just as dependent on the 1 ( 0 ) / 1 ( 1 ) dichotomy. Ideally, in either sit uation, inference should be carried out using the t-statistic from the infeasible GLS regression. When u{ is 7(0), this t-statistic can approximated using the OLS estima tor together with a serial correlation robust standard error estimated from the OLS residuals. Alternatively, when p = 1 and the data are 7(1), this t-statistic can be approximated using the sample mean of Ay* together with a serial correlation robust 2 variance estimated from the first differences of the data. Of course, since most re searchers can’t know a p r i o r i whether their data are 7(0) or 7(1), these results are of limited value. In this paper we study inference problems and the behavior of OLS, first-difference and feasible GLS estimators when the data are either 7(0) or 7(1) and p is unknown. Our analysis builds on two literatures. The first is the literature on the linear regression model with AR(1) errors exemplified by Cochrane and Orcutt (1949) and Prais and Winsten (1954). The second is the literature on inference in regressions with 7(1) variables exemplified by Dickey and Fuller (1979), Durlauf and Phillips (1988) and Elliott, Rothenberg and Stock (1992). Much of the former literature focuses on efficient estimation of regression parameters when the errors follow a stationary AR(1) process, and is directly relevant for our analysis when \p\ < 1 and v t is i i d .1 There are few exact analytic results in this literature because of the dependence of results on the regressors and the nonlinearity introduced by feasible GLS estimation.2 Moreover, the asymptotic results summarized above rely on |p| < 1 and are not refined enough to discriminate between OLS and feasible GLS estimators. Thus, the majority of work in this area has relied on Monte Carlo simulations. Equations (1) and (2) have also been extensible studied in the unit root literature, primarily with a focus on tests for the hypothesis that p = 1. In most of this literature, the regression coefficients a and 0 are nuisance parameters and p is the parameter interest.3 One of the purposes of this paper is to highlight what this analysis says about the feasible estimators of 0 and statistical inference. W e begin our analysis in Section 2 by presenting results on the asymptotic dis tributions of estimators of 0 . These include the OLS, first-difference, infeasible GLS and three different, but commonly used, feasible GLS estimators. W e avoid the sharp |/?| < 1 and p = 1 dichotomy in the asymptotic distributions by using local-to-unity asymptotics, with the hope that these provide better finite sample approximations. The asymptotic results for |p| < 1 and p = 1 are not new: they are reported here for completeness and because, particularly w’hen p = 1, the results may not be widely appreciated by applied researchers. In any event, the local-to-unity results are the most relevant, since in most econometric applications the errors are highly serially correlated, although perhaps not characterized by an exact unit root. These results 1There islarge literatureon this topic, including Beach and MacKinnon (1978), Chipman (1979), Kadiyala (1968), Maeshiro (1976 and 1979), Magee (1987), Park and Mitchell (1980), Rao and Griliches (1969), Spitzer (1979), and Thornton (1987). 2Two exceptions directly relevant for our analysis are Prais and Winsten (1954) and Chipman (1979). The first paper studies equations (1) and (2) when a = 0 and t>» is iid , and calculates the relative efficiency of the OLS and first-difference estimators as a function ofp and the sample size, T. Chipman (1979) relaxes the assumption on a and calculates the greatest lower bound of the efficiency of the OLS estimator for all T and p < 1. W’e discuss the Chipman (1979) analysis in more detail in Section 2.2.1 3A notable exception is Durlauf and Phillips (1988), which isdiscussed in more detail in Section 2.2.1. 3 show sharp differences in the relative efficiencies of the estimators and four conclu sions emerge from the analysis. First, the Cochrane-Orcutt estimator performs very poorly when p is large. Second, the OLS estimator is more robust to variations in p than the first-difference estimator. Third, the variance of the initial error term has an important effect on the relative efficiencies of the estimators. Finally, the asymptotic results suggest that the feasible Prais-Winsten estimator isthe best estimator in most applied situations. Section 2 concludes with a small finite-sample experiment that indicates that the asymptotics provide reasonable approximations to the finite-sample relative efficiencies. Section 3 studies the problem of statistical inference about /?. Existing Monte Carlo evidence suggests that methods relying on 7(0) asymptotic approximations greatly understate the uncertainty in /? when \p\ < 1 but large. This leads to confi dence intervals that are much too small and hypothesis tests with sizes that are too large. Asymptotic approximations that rely on p = 1 have analogous problems. This section uses the local-to-unity asymptotic approximations from Section 2 to construct bounds tests and conservative confidence intervals building on methods developed in Dufour (1990) and Cavanagh, Elliott and Stock (1993). In Section 4 we apply the methods to estimate and construct confidence intervals for real per-capita G D P growth rates for one hundred and twenty-eight countries using post-war data. Consistent with the analysis in Section 2, we find large differences between the Cochrane-Orcutt and other estimators for many of the countries. There are smaller, but economically important differences in the other estimators, and this highlights the importance of estimator choice. Finally, for most countries, the high degree of serial correlation and short sample leads to wide confidence intervals for /?. Finally, we offer a summary and some conclusions in Section 5, and the appendix contains proofs and other detailed calculations. 2 E stim ators 2.1 T h e M od el The statistical model for the observations {yt}JC_x is conveniently summarized in the following assumptions: 1. The data y t are generated by yt = a + 0t + ut 2 . The error term is generated by (1 — 3. ui 4. v t = d ( L ) e t,with d ( L ) = ut for t = 1,... ,T. p rL )u t = vu for t = 2,..., T Pi’Ui-*-. = £ £ 0 d,Z,', and £ £ 0i I 4 l< °°- (3) . 5. The error term t t isa martingale difference sequence with E ( t ] |et- i , 2,...) = 1 and with sup, E t 4 < 0 0 . Assumption (1) says that the data are generated as a linear trend plus noise; the parameter /? is the average trend growth in the series and is the parameter of interest. Assumptions (2) and (3) are written to include both 7(0) and 7(1) processes. When p T = p, with |p| < 1, then u, is 7(0); while when p j = 1, then u t is 7(1). More generally, when />r = (1 + j ) , then u t follows a “local-to-unity” 7(1) process, with c = 0 corresponding to an exact unit root and values of c 0 generating data that are less (c < 0) or more (c > 0) persistent then the exact unit root process.4 Assumption (3) incorporates a range of assumptions about the initial condition ui, depending on the value of k and p j . For example, when k = 0,then u\ = uj ,so that the initial value is assumed to be an O p(l) random variable. When k > 0, then ui is O p ( T */2) when u t is 7(1), but is O p(l) when tt, is 7(0). When p r = p , with \p\ < 1 and k T —► 00,then u i is drawn from the unconditional distribution of u t, and the process is covariance stationary.5 Assumption (5) implies that the functional central limit applies to the partial sums of Ct, i.e., T ~ * tt W(s), where W(s) is a standard Wiener process.6 Assumption (4) insures that the functional central limit theorem also applies the partial sums of v t , specifically T ~ 2 vt = *>• d(l)W(s). 2.2 A sy m p to tic P roperties o f E stim ators 2 .2.1 OLS, First-Difference and G L S Estimators Let P o l s denote the OLS estimator of ft in (1), let P f q = ( T — l)-1 Y lJ =2 & y t denote the first-difference estimator, and let P g l s denote the infeasible GLS estimator that corrects for non-zero p j - Specifically, P g l s is the OLS estimator in the transformed regression yt - p T V t - i = (1 ~ P t 1)] + ut - )q + P [ t - p r{ t - p r u t - 1, < = 2,3,...,r. (4) together with a ~ Xy\ = <r~xot 4- <r~l P + (T~l u i , (5) where a 2 = (1 — P j ^ +1^)/(1 — p \ ) f°r />t / 1 and c 7 = [/cT] + 1 for p r = 1 For simplicity, the GLS estimator ignores the 7(0) serial correlation associated with d ( L ) . 4These “local-to-unity” processes have been used extensively to study local power properties of unit root tests, construct confidence intervals for autoregressive parameters for highly persistent processes, and more generally, to study the behavior ofstatistics whose distribution depends on the persistence properties of the data. Some notable examples are Bobkoski (1983), Cavanagh (1985), Cavanagh, Elliot and Stock (1993), Chan and Wei (1987), Chan (1988), Phillips (1987), and Stock (1991). 5See Elliott (1993) for related discussion ofthe initialerror in the 1(1) model. 6A range of alternative assumptions will also suffice; see Phillips and Solo (1992) for discussion. 5 This allows us to focus on the major source of serial correlation, p r ^ 0, and leads to no loss of asymptotic efficiency for the models considered here (Grenander and Rosenblatt (1957)). ' ' A A A In large samples, the behavior of P o l S i P f d and P g l s is summarized in Theorems 1 and 2: Theorem 1 ( B e h a v io r o f P o l s >P f d U n d e r a s s u m p t io n s ( l ) - ( 5 ) w ith p r Tl( (a ) (h) T ( P ance T (P (c) Po - ls — fd + uar(ui), ~ P) Vj = 12(1 - />)-*<£(l)2. w h e re /,• = ^ ‘_0 p ^ ~ ^ d j. T h e l i m i t i n g d is t r ib u t io n o f d e p e n d s o n th e d is t r ib u t io n o f th e e 's , a n d s o in g e n e r a l is n o n - n o r m a l, T*( Pq Proof. w ith 1 ( 0 ) E r r o r s ) : |/>| < 1: c o n v e r g e s in d is t r ib u t io n to a r a n d o m v a r ia b le w ith z e r o m e a n , v a r i P ) Vi = fd g ls p, a n d N ( 0,Vi), w h e re P ) P and = ls — P ) — + N(0,Vi), w h ere V i is s p e c ifie d in ( a ) . P a r t (a ) a n d (c ) f o llo w f r o m l i m i t th e o re m . a s t r a ig h t f o r w a r d a p p lic a t io n o f the c e n t r a l T o s h o w p a r t (b ), n o te th a t T ( P fd — P) = u j — fro m w h ic h the re s u lt f o llo w s im m e d ia t e ly .□ Theorem 2 ( B e h a v i o r o f P o l s > P f d L e t S c ( t ) = (— 2c)-1(l — e2rc). T h e n (a) T 3 (P o ls - P) and P q ls w ith 7(1) E rro rs): u n d e r a s s u m p t io n s ( l ) - ( 5 ) , w ith p r N ( 0,7?0, w h e re = d(l)V5[18(c - 2 ) V C + 72c(c - 2)ec + 12c3 + 54c2 + 72c - 72] R i cec+ c — 2(ec — 1) |2 +d(l)2144Sc(«)[ 2c2 ^ (b)THPF D ~ P ) N ( 0 , R 2), w h e re 7?2 = d(l)2[5c(l) + (l-ec)25c(/c)]. ( c) T H P gls - P ) - ^ Rz Proof. N ( 0 , R 3), w h e re 5c(/c)c2 + 1 = d(l)2[ (Sc(/c)c2 + 1)(1 - c + §c2) - 5e(/c)(ic2 - c)2 S e e A p p e n d ix . Corollary 3 ( B e h a v io r o f P o ls >P f d U n d e r a s s u m p t io n s ( l ) - ( 5 ) , w ith p r — (<•) T l 0 o i . s - 0 ) and 1.' N ( 0 ,id ( in (b) T i ( P F D - l ) ) ± N ( 0 , 4 i n (c) Tl( f a s - 0 ) -i-N(0,J(in = (1 + y); 6 P o ls w hen p — l ) : W e highlight five features of these results. First, P o l s > P f d and P g l s converge to ft faster in the 7(0) model than the 7(1) model. This results obtains because the variance of the errors is bounded in the 7(0) model and increases linearly with t in the 7(1) model. Sampson (1991) discusses the implication of this result for long-run forecast confidence intervals. Second, the averaging in P o l s in the 7(0) and 7(1) cases and in P f d in the 7(1) case leads to asymptotically normal estimators. In contrast, since T ( P f d — P ) = [ T / ( T — l)](ur — ui), no such averaging occurs for P f d in the 7(0) case, so that P f d is not asymptotically normally distributed in general. (See Quah and Wooldridge (1988) and Schmidt (1993) for related discussion.) 1 Third, P g l s is the asymptotically efficient estimator regardless of the value of p and it corresponds to the B L U E estimator when d ( L ) = d , a constant. The efficiency of the FD and the OLS estimator relative to the GLS estimator differs dramatically in the 7(0) and 7(1) c a s e s . When the errors are 7(0), then P f d converges to P more slowly than does P g l s > and thus has an asymptotic relative efficiency of 0. In this case, P o l s is asymptotically efficient, the familiar result from Grenander and Rosenblatt (1957). When the errors are 7(1), P o l s i P f d and P g l s converge at the same rate and the relative efficiency depends on the parameters c and k . Figure 1 plots the asymptotic relative efficiencies (defined as the ratio of the asymptotic variances of P o l s and P f d to the asymptotic variance of P g l s ) in the 7(1) model for a range of values of c and k . When c = 0, both P o l s and P f d are invariant to u i and so their variances and the relative efficiency do not depend on k . In this case P f d is asymptotically efficient and P o l s has an efficiency of 5/6. This result is derived in Durlauf and Phillips (1988), who study the properties of trend estimators in the model with p = 1 (equivalently, c = 0). When c is sufficiently negative, P o l s dominates P f d f°r all values of «. The intersection point of the P o l s and P f d relative efficiency curves depends on k . For example, when k = 0, P f d is efficient relative to P o l s for values of — 18.6 < c < 1.2,and P o l s dominates P f d for c outside this range. When k = 1.0, the range narrows to — 7.6 < c < 0.9. Fourth, when k = 0, so that u\ is O p ( 1), the relative efficiency of both P o l s and P f d increases monotonically with c. The relatively poor performance of these estimators when uj is O p(l) has been noted elsewhere, notably by Elliott, Rothenberg and Stock (1992) in the context of unit root tests. On the other hand, when k > 0,so that is O p ( T * ) , the relative efficiency of P o l s is U-shaped, with a minimum that depends on the specific value of k . For example, when k = 1, the minimum relative efficiency of P o l s occurs at c = — 3.006 where it takes on the value of 0.7535. As k — ► oo, the minimum relative efficiency of P o l s i s .7538 and occurs at c = — 3.076, a result that was also derived by Chipman (1979) using methods different from those employed here.7 7Chipman (1979) also shows that, when d (L ) = d, this asymptotic relative efficiency value isthe greatest lower bound for the relative efficiency of P o l s for all n > 2. Because of a slight numerical error in Chipman’s paper, his reported numerical results are different from those reported here. 7 Finally, when the errors are 7(1), the variances of 0 o l s i P f d , and 0 g l s depend on c and k in important ways. For example, Figure 2 plots the variance of 0 g i s as a function of cand k . A s c increases, the persistence of the errors increases and so does the associated variance of P g l s - Similarly, as k increases, the variance of uj increases, leading to an increased variance in P g l s 2 .2.2 Feasible G L S Estimators The efficient GLS estimator relies on two parameters, p and /c, whose values are typi cally unknown. In this section we analyze feasible analogues of P g l s - The parameter p is easily estimated from the data, and as we show below, replacing p with an es timate has little effect on 0 g l S ‘ On the other hand, it is impossible to construct accurate estimates of k , since this parameter only affects the data through the vari ance of the single observation uj. W e therefore analyze three feasible GLS estimators that differ in their treatment of the initial observation. W e find large differences in the relative performance of these estimators across different values of k . To focus attention on the parameter k , we begin by analyzing the estimators assuming that p is known; a simple modification of these results yields the results for unknown p . As above, the GLS estimators ignore the serial correlation associated with the 7(0) dynamics in d ( L ) , since the Grenander-Rosenblatt (1957) results imply that OLS or GLS treatment of d ( L ) has no asymptotic effect on the estimators of /?that we consider. Let f l c o denote the Cochrane-Orcutt (1949) GLS estimator that ignores the levels information in the first observation; that is, (3co denotes the OLS estimator of f i in equation (4). Let f i c c denote the GLS estimator constructed under the assumption that uo = 0. This assumption is often made in the unit root literature (see, e.g., Elliott, Rothenberg and Stock (1992)) and is referred to as the “conditional case.” Thus, j 3 c c is the OLS estimator of (3 from (4) together with: j/i = a + 13 + ux. (6) Finally, let f j p w denote the Prais-Winsten (1954) estimator; that is, the OLS estima tor of (3 from (4) together with: (1 - p r ) xl2y i = (1 - P2 t Y /2o + (1 - P t ) 1,2P + (1 “ Pt )1/2u»• (7) The Prais-Winsten estimator is defined for p j < 1, and we limit our discussion to this situation. In the notation introduced in the last section, f r e e corresponds to the GLS estimator constructed using k = 0, and /3pw is the limiting value of the GLS estimator as k —► oo. When p r = p , with |/>| < 1 (i.e., u t is 7(0)), each of the GLS estimators is asymptotically efficient and the large sample distribution is given in Theorem 1. (Specifically, the value of c that we report (c = —3.07558) is a more accurate estimate of the root to his polynomial (3.3) than the value reported in his paper (c= —3.09485).) 8 Thus, we need only consider the behavior of the estimators in the 7(1) model, and this is done in the following lemma: Lemma 4 ( B e h a v io r o f G L S E s t i m a t o r s w ith />r = (1 + U n d e r a s s u m p t io n s ( l ) - ( 5 ) , w ith (a ) T ' l \ 0 c o - W(0,Gi), 0)-^ „ G\ 7(1) E rro rs): f): w h ere 12d(l)2 . = -- -— ,f o r c / 0, a n d c1 G\ — (b) T W tfc c - 0 ) ~ ^ N ( 0 , G 2), w h ere J(1)2 fl I Sfrl (c “ 2°2)2 1 - 1 _ C + I C2[1 + c— & ]- C (c ) d(l)2,f o r c = 0 . T l t* (0 p w - 0 ) - ^ N ( 0 , G 3 ), w h ere G3 — <i(1)I ■[cI5,(«) + 1 + i c !]. Proof. See Appendix. Part (a) of the lemma implies that G \ , the limiting variance of T * { 0 c o — 0 ) , is discontinuous at c = 0. This occurs because the regression constant term, a , becomes unidentified as c —+ 0. For values of c close to zero, a is very poorly estimated, and the collinearity between the two regressors (1,<) in equation (4) means that f i c o is also a poor estimate of 0 . When c = 0, a disappears from equation (4) and so this source of variance disappears from 0 c o - Figure 3 shows the efficiency of each of the estimators relative to 0 g l s • The Cochrane-Orcutt estimator, 0 c o i performs very poorly for small values of c regardless of the value of k . This result is consistent with a large literature on the poor performance of the Cochrane-Orcutt estimator with trending regressors and p close to unity.8 The relative performance of the other two estimators depends on the values of k and c. When k — 0, fa c e is the asymptotically efficient estimator; while f ip w is the efficient estimator as k — ► 00. From Figure 3, 0 p w is approximately efficient even when k is very small. For example, for k = .01 the relative efficiency of is larger than 0.73 for all values of c; for k = .05 the relative efficiency is larger than .92; and for all values of k > .10, 0 p w is essentially efficient. While f i c c is efficient when k = 0 ,this efficiency gain disappears quickly for moderate values of c as k increases. 8See Prais and Winsten (1954), Maeshiro (1976,1978), Beach and MacKinnon (1978), Park and Mitchell (1980), Thornton (1987) and Davidson and MacKinnon (1993, Section 10.6). 9 W e are now ready to discuss the feasible GLS estimators with p r unknown. These estimators are calculated like their infeasible counterparts, using an estimator of p r in equations (4) and (7). These estimators will be denoted as 0 f c o ,0 f c c , and 0 f p w Analysis of these estimators is complicated by the fact that they implicitly depend on the estimator for pr, and a variety of estimators of p j have been suggested. For 0 f c o the non-linear least squares estimator is often employed, and this estimator is studied by Nagaraj and Fuller (1991) for the model with general regressors. Their analysis can be simplified here because of the special structure of the regressors: equation (4) together with assumption (2) can be combined as: y t = a + bt + p r y t - i + ft, for t = 2,3,..., T, (8) where a = a(l — p i ) 4-0 p r and 6 = /?(1 — pj).^Thus, 0 f c o can be formed from the OLS estimators from equation (8) as 0 f c o = 6/(1 — p r ) for p T ^ 0 and 0 f c o = 2 for p r = 1, where 2, 6,and p j are the OLS estimators of the coefficients in equation (8). Equivalently, 0 f c o can be constructed as the OLS estimator of 0 in (4) using p x in place of pr* Since the asymptotic distribution T(l— p r ) is readily deduced when p T = (1 + f ) , (see Stock (1991), for example), the asymptotic distribution of o — 0 ) can also be readily deduced. The problem is more complicated when analyzing 0 f c c and 0 f p w , since these estimators are generally based on iterative schemes for estimating pr,or, and 0 . Iter ative schemes are often used to construct 0 f c o as well. Since the limiting distribution of p r depends in important ways on the precise way the data are “detrended” (for example, see Schmidt and Phillips (1992) and Elliott, Rothenberg and Stock (1992)), the limiting distribution of 0 f c c > and 0 f p w will depend on the precise specification of the iterations. Rather than present results for specific versions of these estima tors, we present limiting representations of 0 f c C i and 0 f p w written as functions of c = plimT(l — pr). Different estimators of pr will lead to different limiting random variables c and different asymptotic distributions for the estimator of 0 . A specific example is contained in Durlauf and Phillips (1988, Theorem 4.1), who derive the lim iting distribution of 0 f c o when c = 0 and c is constructed from the Durbin-Watson statistic calculated from the levels OLS regression. Before presenting the limiting distributions for the feasible GLS estimators, it is useful to introduce some additional notation. The error term in the feasible GLS version of (4) is v t = V t — p r u t - u and the limiting values of the feasible GLS estimators can be written in terms of initial condition Ui and partial sums of v t . In the appendix we show that T ~ * u i =» IV^/c) ~ A^O, 5c(k )), where 5C(«) is defined in Theorem 2; we also show that T ~ j =» W ( s ) where IV’(s) is a functional of IF(s) and W e{,c). With this notation established, we now present the limiting distribution of the feasible GLS estimators: T i (/3f c 10 Theorem 5 ( B e h a v io r o f F e a s ib le G L S E s t i m a t o r s ) : S u p p o s e th a t a s s u m p t io n s ( l ) - ( 5 ) a re s a tis fie d , p p = (1+ f), and p U m ( p T - l) = 8^0. Then: (a ) T L > 0 FCO ~ 0 ) = > c-42Jftl - s)d W (s), => [1 - 3 + i c V ^ - j c 2) ^ * ) - f 0l (c s - l ) d W ( s ) ) , -/?)=► [1 - § 3 + i c 2] - 1[3W'e(« ) - f o i l + \ c - t s ) d W ( s ) ] . ( b) T H f o c c ~ 0 ) (c ) T \ 0 fpw Proof. __ and S e e A p p e n d ix . This theorem allows us to offer practical advice about choice of estimators. First, notice that 8 appears in the denominator of the limiting representation of (0 p c o ~ 0 ) . For most commonly used estimators of p , p F can take on values arbitrarily close to 1 with positive probability, so that 8 can be very close to zero. This means that 0 f c o can be very badly behaved, since realizations of 8 close to zero will often lead to extreme realizations of 0 F c o • On the other hand, 0 p c c and 0 F p w are better behaved, since [1— 8 + 58s] > 0 and [1 — ^8 + j^c*] > 0 for all values of 8. This can be seen in Figure 4 which plots the limiting probability densities of T * { 0 p c o — /?),T’a[ 0 F c c — 0 ) and T i { 0 p p w — 0 ) , for the case with c = 0, k = 1,and d(l) = l.9 Also plotted is the probability density of the exact (infeasible) GLS estimator (which in this case is the standard normal). The estimators 0 F c c > and 0 F p w have probability distributions very close to the infeasible efficient estimator. On the other hand, the distribution of 0 p c o is much more disperse, with thicker tails than the other distributions. For example, the limiting probability that \ T * ( 0 p c o — 0)\ exceeds 2 isapproximately 20%; while the corresponding values for 0 F c c and 0 F p w are approximately 5%. Figure 4 suggests that littleislost using in using either 0 F c c and 0 F p w in place of the infeasible efficient estimator, at least for this value of cand k , and that 0 F c o performs poorly. Additional calculations (not shown) indicate that the relative efficiencies of 0 p c c and 0 p p w are close to their infeasible analogues for a wide range of values of c and k . Table 1 summarizes many of the results in this section by presenting the average mean squared error for the different feasible estimators and different values of k ,av eraged over different ranges of c.10 As a benchmark, the first row of the table shows results for the efficient, but infeasible, GLS estimator. The next two rows are the OLS and first-difference estimators, followed by two of the feasible GLS estimators. (Since the asymptotic mean squared error of 0 F c o does not exist, this estimator is not included in the table.) The last row of the table shows results for a “pre-test” 9The densities for the feasible GLS estimators are estimates based on 5000 draws from approxi mations to the asymptotic distributions (constructed usingT=500). The estimators0 f c o and 0 p c c were constructed using p r constructed as the OLS estimator of (8). The Prais-Winsten estimator used min(l,/>r). l0These MSE’s were estimated using the simulations described in footnote 9. 11 estimator ( P p t ) constructed from the OLS and F D estimator. Figure 1 provides the motivation for this estimator. Since the OLS estimator dominates the first-difference estimator for large negative values of c and is dominated by the first-difference esti mator for small values of c, the pre-test estimator corresponds to the OLS estimator when 8 is large and negative and corresponds to the FD estimators when 8 is close to zero. Specifically, P p t = P o l s when 8 < c and P p t = P f d when 8 > c, where c is pre-specified threshold. The results shown in the table are for c = — 15, a value that produced good results over the range of values of k and c that we considered. Table 1 and the figures shown above suggest five conclusions: (i) The infeasible GLS estimator P c o performs very poorly for values of c close to 0. This poor performance is inherited by the feasible GLS estimator. For all values of c ^ 0 and for all values of /e, this estimator is dominated by P o l s • Thus, this estimator should not be used and is ignored in the remaining discussion. (ii) For very small values of c (say, — 2 < c < 0), f i p p isthe preferred estimator with a mean squared error approximately 5% lower than P f c c and P f p w • For this range of values of c, the OLS estimator, P o l s i has a relative efficiency of approximately .75. The pre-test estimator performs well, and is l%-2.5 % less efficient than P f D i depending on the value of k . (iii) For values of c in the range — 10 < c < — 2 , the relative performance of the estimators depends critically on the value of the initial error, parameterized by k . When k = 0,P f c c dominates the other estimators; P f p w is the preferred estimator when k > .10. When k = .05 the feasible GLS estimators and P f d are comparable. (iv) For values of — 30 < c < —10 and when k = 0,P f c c isthe preferred estimator. When k > 0.05, the variance of P f c c is more than twice as large as the variance of the best estimator, P f p w • The first difference estimator also performs poorly relative to P f p w when k > .05. (v) Items (ii)-(iv) show clearly that the best estimator depends on the values of c and k . Neither of these parameters can be consistently estimated from the data, and so a good choice must depend on either prior knowledge or robustness considerations. Our reading of the results suggests that P f p w is the most robust estimator, with a M S E close to the optimum for all values of the parameters considered. The pretest estimator is a reasonable alternative to P f p w iit has slightly better performance w’hen c close to 0 but somewhat worse performance for large negative c. 2.3 Sm all Sam ple P rop erties o f E stim ators The asymptotic results summarized in Theorems 1, 2 and 5 are potentially useful for two reasons. First, the asymptotic relative efficiencies can provide a criterion for choosing among the estimators even in finite samples. Second, the asymptotic distributions provide a basis for constructing confidence intervals and carrying out hypothesis tests. In this section we evaluate the first of these uses, and ask whether the 7(0) and 7(1) asymptotic variances provide a useful guide for choosing among the 12 estimators in small samples. In the following section, we discuss confidence intervals and statistical inference. Table 2 shows the exact relative efficiencies of 0 o l s i 0FD i 0FCC, 0 f p w , and 0 P t for the model with d(L) = d, et ~ NIID( 0,1), for various values of T , p, and for k = 0 (panel A) and k = 1.0 (panel B).n Also shown in the table are the relative efficiencies implied by the 7(1) asymptotics, calculated using c = T(p — 1). The 7(0) asymptotic relative efficiencies are not shown because they do not vary with T, £ or «; from Theorem 1 they are 1.00 for P o l s , 0 f c c > 0 f p w , and 0 p j and 0.00 for 0 f d In all cases, the 7(0) asymptotic relative efficiency suggests indifference between the four estimators 0 o l s > 0 f c c > 0 f p w and 0 p t , and suggests that these estimators are preferred to 0 f d W h e n p = 0.5, the finite sample results in Table 2 suggest that 0 olsi 0FCCi and 0 f p w are essentially efficient for all of the sample sizes considered. These estimators are significantly better than 0 f d • The pre-test estimator has a relative efficiency intermediate between 0 o l s and 0pp when T = 30, and very close to 0 ol s for larger values of T. Thus the 7(0) relative efficiency predictions are quite accurate when p = 0.5. The predictions based on the 7(1) asymptotic relative efficiencies are off the mark. The 7(1) asymptotics suggests that 0 f c c strongly dominates the other estimators when k = 0 and is strongly dominated by both 0ol s an<l 0 f p w when k = 1. O n the other hand, the estimator with the largest 7(1) asymptotic relative efficiency coincides with the largest finite sample relative efficiency, even when p = 0.5. For all of the other values of p that are considered (0.8, 0.9, 0.95, 1.0), the rank ings implied by the 7(1) asymptotic relative efficiencies are more accurate the 7(0) rankings. Indeed in all cases studied in the tables, the estimator with the largest 7(1) asymptotic relative efficiency has the largest finite sample relative efficiency as well. Thus, this experiment suggests that the 7(1) asymptotic relative efficiencies provide a useful criterion for ranking estimators in typical econometric settings. 3 3.1 Confidence intervals C onstruction o f confidence intervals. In this section we discuss methods for constructing confidence intervals for 0. W h e n p < 1 (so that the errors are 7(0)) confidence intervals can be constructed in the usual way by inverting the “t-statistic” constructed from any of the asymptotically equiv alent estimators 0 o l s » 0 f c o , 0 f c c , 0 f p w , or 0 p t • These t-statistics can be formed using an estimator for the variance Vi in Theorem 1, constructed by replacing p and d( 1) with consistent estimators. While these confidence intervals are asymptoticallyl llThe mean squared errors for 0pcc> 0FPW< and 0 p r , were estimated using 10,000 Monte Carlo draws, using p = £lT=2 2iUi-i/£^=2'2?, where u, are the OLS residuals from the regression of y, onto (1,t). This estimator ofp issuggested by the simulation results in Park and Mitchell (1980). 13 valid, they can greatly understate the uncertainty about 0 when p is large and the sample size is small. (See Park and Mitchell (19S0) for simulation evidence.) Thus, in most situations of practical interest, confidence intervals based on 7(0) approxima tions are not satisfactory. An alternative method pursued here is to construct confidence intervals using approximations based on 7(1) asymptotics. As we show below, this method yields confidence intervals with coverage rates closer to the nominal size than the 7(0) ap proximations. Unfortunately, the method isalso more complicated. The complication arises because in the 7(1) model, the asymptotic distribution pf the various estimators of 0 depends on the nuisance parameters c and k , and these parameters cannot be consistently estimated from the data. Thus, the variances of the estimators cannot be consistently estimated, so that t-statistics will not have the appropriate limiting standard normal distribution. While this problem cannot be circumvented entirely, it is possible to construct asymptotically conservative confidence intervals following the procedures developed by Dufour (1990) and Cavanagh, Elliott and Stock (1993).12 Specifically, let B K(c) denote a 100(1 — c*i)% confidence interval for 0 constructed conditional on a specific value of c and k . Similarly, let C K denote a 100(1 — o i l )% confidence interval for c conditional on k . Assume that 0< k < 7c, where 7c is pre-specified constant. Then the Bonferoni confidence interval, Uo<*<*Uc€C. B K( c ) , is a conservative 100(1 — ct\ — 0 2 )% confidence interval for 0 . This confidence interval requires the conditional confidence interval for 0 , B K( c ) , and the marginal confidence interval for c, denoted C K. Since B K(c) conditions on the nuisance parameters c and «, an asymptotically valid approximation can be con structed using any of the estimators 0 o i s > 0 F D , 0 c c , or 0 p w , and their asymptotic variances given in Theorem 1 and Lemma 4. (These variances require d ( 1), which can be consistently estimated using standard spectral estimators.) The marginal confidence intervals for c, C K, can be constructed using the methods developed in Stock(1991).13 12Dufour (1990) considers the problem ofstatistical inference in the regression model with Gaus sian AR(1) disturbances. He develops “bounds” tests and associated confidence intervals based on exact distributions. Cavanagh, Elliott and Stock (1993) consider testing for Granger-Causality in a regression with a highly serially correlated regressor modeled as a local-to-unity process. They develop bounds tests and associated confidence intervals based on asymptotic distributions. 13Stock (1991) considers the case with k = 0 and, using our notation, develops methods for con structing confidence sets Co- However, it iseasy to modify his analysis for k > 0. Specifically, fol lowing Stock, we construct confidence intervals by inverting the Dickey-Fuller t-statistic, rT. Under the assumption that k = 0, Stock shows fT => ( f * VFer(s)Jds)J[c+ f g W J ( s ) d l V ( s ) / ( f g WT (*)*</*)], where W J (s ) isthe “detrended" diffusion: W J (s ) = W e(s) — a i( r ) W e( r)d r — s fg aj(r)lVe(r)dr, where the diffusion W e(s) is defined in the appendix, ai = 4 — 6r, and aj = —6 + 12r. These results rely on the fact that => d(l)lVe(s) when k = 0. As shown in the appendix, when k ^ 0,T~£ u[ ,t ] => d(l)[We(s) + e,e W e(K)], where W e(/c) ~ JV(0,Se(«)) and is independent of W e($). Using this, itisstraightforward to show that allof Stock’s analysis continues to hold, with lVc(s) + e, c \Vc(K) replacing W e(s) in the above limiting representation for rT. 14 In general, this procedure is quite demanding. For each 0 < k < 7c, C K must be formed, then B K(c) must be constructed for all c € C«, and the union taken over all of these confidence sets. There are three special features of the linear trend model that simplify this procedure. First, from Theorem 2,the asymptotic variances of P o l s and P f d are monotonically increasing in c. Thus, when B K(c) are formed using t-statistics constructed from P o l s or then Uc€C„ B * i c ) = #«(c), where c = supc{c € C K ) . While this simplification does not necessarily hold for the GLS estimators P e c and P p w , experiments that we have performed suggest that Ucgc* B * ( c ) v B K(c) appears to be a good approximation for confidence sets constructed from these estimators as well. The second simplifying feature is that the distributions of the statistics used to form C K change little as k changes, so that C o as C K for all k . u Finally, for all of the estimators, the asymptotic variance is increasing in k and the limit exists as k — ► oo, so that B k ( c ) C B oo( c ) for all k . Putting these three results together implies that U o<*<k Ucec* B K(c) as .^(c), where c = supc{c € Co}. Thus approximate 100(1 — g-i — 0:2)% confidence intervals can be formed by (i) choosing the largest value of c in the 100(1 — a^)% confidence interval constructed using the procedure from Stock (1992), and (ii) constructing a^l00(l — confidence interval for P using this value of c together with P o l s i P f d , P c c , or P p w and an associated variance from Theorem 2 or Lemma 4 evaluated at k = oo. W e make two final points before evaluating the small sample properties of this procedure. First, since the variance of all of the estimators is increasing in c, smaller confidence intervals for P can be obtained by constructing 1-sided confidence intervals for c. Second, when the B K(c) confidence intervals are constructed by inverting the tstatistics for the estimators, the widths of the intervals will be non-random conditional on c and «. This implies that the narrowest of the confidence intervals (across all estimators) will also have coverage rate exceeding 100(1 — Qi — aj)%. Thus, for example, since P o l s is efficient relative to P f d when c < — 7.6 and k is large, the confidence interval can be constructed using P o l s when c < — 7.6 and using P f d when c > — 7.6. 3.2 Sm all sam ple perform ance o f confidence intervals Table 3 shows estimated coverage rates for confidence intervals for different values of T and c, calculated as described above. In panel A, the confidence intervals are calculated as the narrowest of the OLS and FD confidence intervals. Panel B shows results for confidence intervals constructed from the Prais-Winsten estimator. The design was much the same as in Section 2.3, i.e., d ( L ) = d and et ~ iV(0,1). Results u YVhen c = 0 the distribution of ?T isinvariant to k . This isnot strictly true for other values of but the distribution changes very little. For example, when c = —1.0 the 97.5 percentiles for r*1 are -3.72, -3.70, -3.70 and -3.70 when k = 0.0, 0.5, 1.0, and 10.0, respectively. The corresponding percentiles are -3.89, -3.84, -3.84, -3.84 for c = -5.0; -4.20, 4.20, 4.20, -4.20 for c = —10.0; and 4.52, -4.54, -4.54, -4.54 for c = —20.0. These percentiles are based on 5,000 simulations with T = 500. c, 15 are reported for conservative 90%, 95% and 99% confidence intervals constructed with Qi = aj. Results for non-symmetric ou and or? are similar and are not reported. The confidence interval for p was constructed from the ?T statistic constructed from the regression of y t onto A y t_i and (l,f) using the sample t = 2 , The sample residual variance from this regression was used as the estimator of <f(l)2 in the con struction of the confidence intervals for 0 . Finally, since the Prais-Winsten estimator is defined for \p\ < 1 we restricted the upper limit of the confidence interval to p = 1. For comparability, this restriction was also used in the 0 o l s and 0 f d confidence intervals. The coverage rates are close to their nominal level for c = 0. When c < 0, the confidence intervals are conservative, with coverage rates exceeding the nominal level. This occurs because of the sharp increase in the variance of estimators for small c. So for example, when the true value of c = — 5, then c = 0 is often in the confidence set Co, the variance of the estimators is much larger when c = 0 than when c = — 5 (see Figure 2) and this leads to a wide confidence interval for 0 . 4 E conom ic G row th R ates for th e Postw ar P e riod Table 4 shows estimated annual growth rates of real G D P per capita for 128 coun tries over the postwar period. The data are annual observations from the Penn World Table (version 5.5) described in Summers and Heston (1991) (series RGDPCH). The data set contains 150 countries, and we limited our analysis to those 128 countries with 20 or more annual observations. The first column of the table shows the country identification number from the Penn World tables, and the next column shows the country name. Columns 3-6 present four estimates of average trend growth (0 o l s , 0 f d »0 f c o »and 0 F P W i respectively); column 5 shows the estimate of c used to con struct the feasible GLS estimates (c); column 6 shows the Dickey-Fuller unit-root test statistic (?T) used to construct a confidence interval for c, and columns 7 and 8 present lower and upper limits of the approximated 95% confidence interval for 0 con structed from the 0 p w (0min and 0 max, respectively). The estimate c was calculated as explained in footnote 11. The fT statistic was calculated from the regression of A y t onto y t -\,Ayi_i and (1,t) using data from t = 3,.. .T, and the point estimates from this regression were used to estimate d(l). W e highlight five features of the results. First, for the majority of the countries, the different estimators give similar results. For example, for the Congo (country 12) the estimates range from 2.8% ( 0 f d ) to 3.4% (0 f c o )• Second, while the 0 f c o estimates are usually similar to the other estimates, they occasionally deviate substantially. For example, the estimates for Suriname (country 81) constructed from 0 o i S i 0FD > and 0 f p w range from 0.4% to 1.4%, while the estimate constructed from 0 f c o is — 212%. Indeed for 31 of the 128 countries, 0 f c o differs from 0 o i s by more than 5 percentage points. Third, while the differences 16 in the other three estimators are much smaller, these differences can be quantitatively important. For example, 0o l s > 0FD differ by more than 1% in 5 cases and by more than |% in 35 cases. Fourth, the confidence intervals are often wide and include negative values for /?. This results from three factors: a small sample size, a large error variance and a high degree of persistence in the annual growth rates. For example, the approximate 95% confidence interval for Algeria (country 1) is — 1.18 < /? < 4.10. For Algeria, the Dickey-Fuller t-statistic is — 1.46 which implies that c = 0 (i.e., p = 1) is contained in the 97.5% confidence interval for c. Thus, for this value of c, f ip w corresponds to the first-difference estimator. The mean growth rate for Algeria over the sample period is 1.45% (= 0 f d ) and this is the center of the confidence interval. The standard devia tion of the annual growth rates is 7.3%; thus, ifthe annual growth rates were serially uncorrelated, the standard deviation of the sample mean (= 0 f d — P p w ) would be 1.33% (= 7.3%/>/30). For Algeria, the growth rates are slightly negatively corre lated and the estimated standard deviation of f ip w used to construct the confidence interval was 1.18%. Finally, a few of the confidence intervals are quite narrower. For example, the estimated confidence interval for the U K (series 140) is 2.07 < 0 < 2.44. This series is less persistent than most of the others, and the Dickey-Fuller t-statistic is -4.51. This leads to a confidence interval for c with an upper limit of c = — 14.1 (corresponding to p = 0.66). From Figure 2,estimates of /? are much more precise when c = — 14.1 than when c = 0 . Indeed the ratio of the asymptotic standard deviation for f ip w for c = — 14 and c = 0 is 0.2, which approximately corresponds to the difference between the widths of the confidence intervals for 0 for the U K and the US (country 71). 5 C oncluding Rem arks In this paper we study the problems of estimation and inference in the deterministic model. While the structure of the model is very simple, serial correlation in the errors can make efficient estimation and inference difficult. Asymptotic results are presented for 7(0) and local-to-unity 7(1) error processes, with the latter being the most rele vant for econometric applications. The asymptotic distribution of the estimators is shown to depend on two important parameters: (i) the local-to-unity parameter that measures the persistence in the errors and (ii) a parameter that governs the variance of the initial error term. Three conclusions emerge from our analysis. First, the Cochrane-Orcutt estimator is dominated by the other feasible estimators and should not be used. When the data are highly serially correlated (i.e., the local-to-unity parameter is close to zero), the distribution of the Cochrane-Orcutt estimator has very thick tails, and large outliers are common. Second, the feasible Prais-Winsten estimator is the most robust across the parameters governing persistence and initial variance. This is the preferred es timator unless the researcher has sharp a p r i o r i knowledge about these parameters. 17 Finally, inference that ignores uncertainty about p or the variance in the initial er ror term can be seriously flawed and lead to large biases in confidence intervals for trend growth rates. It is not clear how to optimally account for uncertainty in these parameters, but conservative confidence intervals and tests are easily constructed. 18 A A.l Appendix: Theorem Proofs Preliminaries: From assumption (5), T ~ i *t =► W(s); in addition, this result, together with as sumption (4) implies T ~ » Vt =► d(l)VF(s), where W(s) is a standard Weiner pro cess. Analogously, accumulating the errors backwards from time 0, T ~ a et =*► W(s) and T -1 d(l)W(s), where W(s) is a standard Weiner process, independent of W(s). Let it, = E ‘=o P T v t - i p j - (1+f). Then =*• d(l)Wc(a), where W c(s) denotes the diffusion process generated by d W e( s ) = c W e(s)<fs + d l V ( s ) . Similarly, T ~ 2U\ — T~% p'jV\-i_= > < f ( l w h e r e W c(k ) denotes the diffusion process generated by d W c(.s) = cVVr(.(s)ds -I-dW(s). Note that W c(k ) ~ N ( Q , S c ( k )), where S c ( k ) = (— 2c)-1(l — e 2cK). Finally, write u t = u t 4-p { j l u i , so that =» d(l)[Wc(s) + e“ W c(/c)]. A.2 P r o o f o f T h e o r e m 2: A.2.1 Proof of (a): By direct calculation: t _1 —( T - ‘ EL, j- ) ( T - i EL, T -i Z L , H P o l s -0) T -' E Z U t )1 ~ (T-‘T .L ,( W Thus, T Ti0OLS - 0) = =* <f(l)12j where i?i = A i + \ s 1 2 T-i’ E u . d - i) + 0 ,(1 ) t= 1 1 ^ - ^)[Wc(s) + e “ W e( K ))d s ~ N ( 0, R x), A i ,with A\ = var (<1(1)12 f j , - i)ny»)<M and A 2 = rar{Wc(/c)<f(l)12 J (s — ^)e*cds}. To calculate Aj, note that: jf (* - = j \ s - 19 j) J ‘ ^ - TU \ V ( r ) i s = j \ j \ * - \ y is ) '- " d W ( r ) with 6 (r) = { / (s — i)d“ ik}e“eT. = / Jo 2 Jt Thus, A, = 144d(l)2 [ l b ( s ) 2d s , Jo and A 2 = 144d(l)25c(/c)[jf (s - |)e'cds]2. The first term in i?i is Aj after simplification, and the second term is Aj. A .2.2 P ro o f of (b): t H P fd - P ) = T - > ut - T - i u , = T - i u T - T - i u ^ l - => <f(l)[^(l) - (1 - e‘)VTt(«)) ~ JV(0,i(I)*[S«(l) + (1 - e‘)!S,M]. A .2.3 P ro o f of (c): This GLS estimator is constructed by OLS applied to an equation of the form y t = x \8 + e t, where 8 = (a /?)', X\ = (a-1 a ~*)', x t = [(1 — pr) t — p j { t — 1)]' for = 2,... , T . Let Q = J 2 x tx ti an<^ r = with elements q ij and r; for i , j = 1,2. Then 0 g l s — ft) = (?n922 — 922)-1(9n r2 — 9i2ri)- The various parts of the theorem will be proved by evaluating the relevant expressions for and r,-. Specifically, t T 9n = a Zi + (T — 1)(1 — pr)2; 912 — 922 = <*Zi + ( T — 1)pj + 2pr(l — r ri = ^ ’ “ 1 + (1 ~ P t ) XZ t=2 W e consider the cases with « = 0 and k — 0 : <?Zi + { T — k l)pr(l — Pt ) ^Z * + U — <=2 r r2 = + 5Z <=2 > 0 in turn. p t ) . By direct calculation: T-l(9n 922 - 922) - * (1 - T~ *qn r 2 = JZ^C1 <=2 20 c + -c2) c ^r) + ° p ( 1)' 1 + (1 — Pr)2^Z <=2 P t )2 5Z (=2 “ P t ) + P t ]- T~>ql2ri 0 . So that, T* (PcLS ~ 0) - + op(l) (l-o+W) (9) The result follows by noting that (1 —c + jc2 ) - 1 = k R3 evaluated at k = 0. > 0: By direct calculation: (?n<?22 ~ qu) ( S c ( k )~1 + c2)(l - c + ^c2) - (^c2 - c)2 T i q n r 2 = ( S c{ k )~1 + c2)(T"J £ v t{l - c ~ )) + op(l); t=2 1 T i q i 2r l = (^c2 - c)(5e(K)-1 T-*ti, - cT ’ ]>%<) <=2 Thus, T h -(5 /9 c 2 - c)&(«) 1 tPG/,S P j d(l)[-(lc2 - -irr r - U 1 + r-^Er=2^[(i-cjt)(5e(Ac) - 1 + c2) + (^c3 - c 2)] /O (•^(/c) " 1 J./-2V1 + c2)(l -- n c ++ k 2) _ - U ( rk*2 --A c*) 2 + c2)(l ~ c ) W e( K ) S e( K ) - ' + cs) + (1c2 - c ) c] < W ( j )] N ( Q , R 3 ), (c2 + 5c(/c)~1)(l - c + ic 2) - ^(ic 2 - c)2 3 ^ 2 ' where i?3 = d(l)2[ A.3 c2 + (5 c(k )c2 + 1 ) ( 1 - c+ 3 1 C2) - 5 c(k )(|c2 - c)>] ‘ P r o o f o f L e m m a 4: As in the proof the part (c) of Theorem 2, each of the estimators can be written as the OLS estimator from an equation y t = x't 6 + et, where 6 = ( a /?)', and the estimators differ in their definition of xj and e 2. As above, let Q = £ x<x«i and r = £ x te{, with elements ^ and r,- for i ,j = 1,2. Then, for each estimator ( 0 - / 3 ) = (q u q 22 — q i 2)~ l { q u r 2 — q\2r i ) and the for the proof we evaluate these expressions for each estimator. 21 A.3.1 When Proof of (a): c = 0, T i 0 Co — 0) = T~ j v t, and the result follows directly. For c ^ 0, r 9n = (T - 1)(1 - 9i2 = P t )2\ (T - l)/>r(l - Pt ) + (1 - P t )2 JE2 ^ f= 2 922 = ( T — l)/>r + 2px(l - p t )£ 1 + (1 - P t )2 £ i= 2 T r\ = (1 - Pt ) t= 2 T r2 = 51 v*[^(l - Pt ) + Pr]- Thus, 9 i W22 - 9 n c2( ! ~ c + ^ 2) “ ( | c2 ~ c ) 2 = ^ c 4; T*qn r2 = -c2T~J r 29i2n = c2(l - Y i v *(c f ~ !) + OpU): \c)T ~ i Y , vt + 0, ( 1 ). So that, T h h c o - 0 ) = -(^)[r-i Z m 4 = (~)[r "’ !C ^ - 1) + (1 - ;c)T-i £ + °?(1 ) ^C-f)d(i)Jo c - - S)dvns). The result follows by noting that (T W ) jf(5 " ' W * ) ~ ^(».G >). where „ ,12,,.,,,, /',1 12d(l)3 G\ = (—rd(l) c Jq/ (0 z - *) * = —3 c*— ■ 22 »,] + 0 ,(1 ) (1 0 ) (1 1 ) A.3.2 Proof of (b): For f r e e . T 9u = 1)(1 - />t )2; 1 + {T - 912 = 1+ - (T — Pt ) + (1 - Pt )2 $3 E t=2 = 922 - / > t ) 5 Z< + 1 + { T - \)p \ + 2pr{l (1 ~/> t )2 $3<2; (= 2 T ri = ux + (l - T r2 = ui + ]Tt'([*(l-/?r) + />r]- P t )'5 2 ,v * 1 = 2 1 = 2 Thus, *(911922 - 9n) -♦ (1 - C+ ^c2); T T~±qn r2 f - , - 1 - 1 = T-Jtix - rn T-* ] [ > ( 4 - 1) + op(1); (= 2 T 39l2n = (1 - C + ^C2)T JU! + Op(l). So that, t i ,a m T»(fec c(l - 5c)r ‘*ui - r ^ ^ ( c f - 1) | - m ---------- (i - c + jc»)-------- + M 1 ) =*• 1 — C+ 5 C* - =c)ive(ic) - / (cs - l)<iU,(s)]. L J0 The result follows by noting that <1(1)(1 - c + I c V M i - \ c )W M - £ (c s - i)<nr(.)] ~ a '(o ,g ,), where ^ - (1 _ + j y - [i + 5 (K)i£__ifi2 L] l - C + i C2 ll + ^ ( 1 — C + r c 2 * 23 (1 2 ) (13) A .3.3 P r o o f o f (c): For $ p w i qn = (l-p2 T) + (T-l)(l-pT?\ t= 2 922 = (1 - Pt ) + (T - 1)Pt + 2pr(l ~ pT)52t + (l - p T f ^ t 2] 1 = 2 t= 2 T r i = (1 — />r)«i + (1 “ PT) T S r 2 = (1 - P r)« i + 1 = 2 “ t l Pt ) + P t ]- 1 = 2 Thus, 911922 - 9?2 -» (c2 - 2 = (c2 - 2 r * 9 nr 2 = -(c2 - 2 c)(l - c + ^c2) - (^c2 - c) 2 c)(l - | c + ^ c 2); c)T“i X > t= 2 ( 4 - 1 )+ T ^ r , = - i ( c J - 2 c ) ( 2 c T - i u , + c T ' l £ > , ) + o„(l). 1 = 2 So that, rr\,o a\ T*(ppw — p) cT-tlH-T-iZvticji-ic-l) J..— ^----- + Op(l) 1 jC-t- 12C -------- =*• 1 — :C + TZC* !<*%(■») - J0 A « 2 1 12 - b i - W 4 The result follows by noting that rf(l)(l - jC + i c 3)[cWc(«) - j \ c s - i e - 1)<W M] ~ N ( 0 , G 3), where * = (14 + W - C^ M + /°’(cs “ 5C- 1)V,> J+ ^(l) 2 n _ (1 ~ 2 C + U C2)2' 24 (14) (15) A.4 P r o o f o f T h e o r e m 5: It is straightforward to verify that the analogues of (6), (8 ), and (10) continue to hold for the feasible GLS estimators, with 8 replacing c and v t = u t — P T u t - u replacing v t. The theorem then follows from (7), (9) and (11) using T ~ » v t => W ( c ) . To see this, and to derive an expression for VF(c), write Vt = Ut- PTVt-i = v t- (pT - Pt )vt—\ t-2 = Vt - (CT - c)r_,( 2 / 4 w«-i-i + />r’1“i]i j=0 where 87 = T (1 — p j). Thus [sT] H E ? , * t= 1 T - i E «=1 t- 2 (jT] [»T] «• - (w - c)T -' [,T] E r-i(E t=l - (Sr - c ) ( T - iu ,) T - ' j =0 E P t '' 1 (=1 ^ d ( l) W ( s ) , where iy(s) = W'(a) - (8 - c) f Jo and the last line follows from c t W c( T ) d r c, P?Tv [TT\-j —cJC —c □ 1 - (8 - c)We(/c)— 25 — —C =* Wc(^)? and T ""1 Pr"1 References [1] Beach, C.M. and J.G.MacKinnon (1978), “A Maximum Likelihood Procedure for Regression with Autocorrelated Errors,” E c o n o m e t r ic a , 46, no. 1,pp. 51-58. [2] Bobkosky, M.J. (1983), H y p o t h e s is T e s t in g in N o n s t a t i o n a r y thesis, Department of Statistics, University of Wisconsin. T im e S e r i e s ,Ph.D. [3] Cavanagh, C.L. (1985), “Roots Local to Unity,” manuscript, Department of Economics, Harvard University. [4] Cavanagh, C., G. Elliott and J. Stock (1993), “Inference in Models with Nearly Nonstationary Regressors,” manuscript, Kennedy School of Government, Har vard University. [5] Chan, N.H. (1988), “On Parameter Inference for Nearly Nonstationary Time Series,” J o u r n a l o f the A m e r i c a n S t a t is t ic a l A s s o c i a t i o n , 83, pp. 857-862. [6] Chan, N.H. and Wei, C.Z. (1987), “Asymptotic Inference for Nearly Nonstation ary AR(1) Processes,” T h e A n n a l s o f S t a t is t ic s , 15, pp. 1050-63. [7] Chipman, John S. (1979), “Efficiency of Least-Squares Estimation of Linear Trend When Residuals are Autocorrelated,” E c o n o m e t r ic a , 47, pp. 115-128. [8] Cochrane, D. and G.H. Orcutt (1949), “Applications of Least Squares Regres sion to Relationships Containing Autocorrelated Error Terms,” J o u r n a l o f the A m e r i c a n S t a t is t ic a l A s s o c i a t io n , 44, pp. 32-61. [9] Davidson, R. and J.G. MacKinnon (1993), E s t i m a t i o n m e t r ic s , Oxford University Press: New York. a n d I n fe r e n c e in E c o n o [10] Dickey, D.A. and W.A. Fuller (1979), "Distribution of the Estimators for Au toregressive Time Series with a Unit Root,” J o u r n a l o f th e A m e r i c a n S t a t is t ic a l A s s o c i a t io n , 74, no. 366, pp. 427-31. [11] Dufour, Jean-Marie (1990), “Exact Test and Confidence Sets in Linear Regres sions With Autocorrelated Errors,” E c o n o m e t r ic a , 58, no. 2, pp. 475-494. [12] Durlauf, S.N. and P.C.B. Phillips (1988), “Trends Versus Random Walks in Time Series Analysis,” E c o n o m e t r ic a , 56, pp. 1333-54.3 1 [13] Elliott, Graham (1993),“Efficient Tests for a Unit Root when the Initial Ob servation is Drawn from its Unconditional Distribution,” manuscript, Harvard University. 26 [14] Elliott, Graham, Thomas J. Rothenberg and James H. Stock (1992), “Efficient Tests of an Autoregressive Unit Root,” N B E R Technical Working Paper 130. [15] Grenander, U. and M. Rosenblatt (1957), S t a t is t ic a l S e r ie s ,John Wiley and Sons: New York. A n a l y s i s o f S t a t io n a r y T im e [16] Kadiyala, K.R. (1968), “A Transformation Used to Circumvent the Problem of Autocorrelated Errors,” E c o n o m e t r ic a , 36, pp. 93-96. [17] Maeshiro, Asatoshi (1976), “Autoregressive Transformation, Trended Indepen dent Variables and Autocorrelated Disturbances,” R e v ie w o f E c o n o m i c s a n d S t a t is t ic s , 58, pp. 497-500. [18] Maeshiro, Asatoshi (1979), “On the Retention of the First Observations in Serial Correlation Adjustment of Regression Models,” I n t e r n a t i o n a l E c o n o m i c R e v ie w , 20, no. 1, pp. 259-265. [19] Magee, L. (1987), “A Note Cochrane-Orcutt Estimation,” J o u r n a l r ic s , 35, pp. 211-18. o f E c o n o m e t [20] Nagaraj, N.K. and W.A. Fuller (1991), “Estimation of the Parameters of Linear Time Series Models Subject to Nonlinear Restrictions,” A n n a l s o f S t a t is t ic s , 19, no. 3, pp. 1143-54.. [21] Park, R.E. and Mitchell, B.M. (1980), “Estimating the Autocorrelated Error Model with Trended Data,” J o u r n a l o f E c o n o m e t r ic s , 13, pp. 185-201. [22] Phillips, P.C.B. (1987), “Toward a Unified Asymptotic Theory for Autoregres sion,” B io m e t r i k a , 74, pp. 535-47. [23] Phillips P.C.B. and V. Solo (1992), “Asymptotics for Linear Processes,” o f S t a t is t ic s , 20, pp. 971-1001. A n n a ls [24] Prais, S.J. and C.B. Winsten (1954), “Trend Estimators and Serial Correlation,” Cowles Foundation Discussion Paper 383. [25] Quah, D. and J. Wooldridge, (1988), “A Common Error in the Treatment of Trending Time Series,” Manuscript, Department of Economics, M.I.T.. [26] Rao, P. and Z. Griliches (1969), "Small-Sample Properties of Several Two-Stage Regression Methods in the Context of Auto-Correlated Errors,” J o u r n a l o f the A m e r i c a n S t a t is t ic a l A s s o c i a t io n , 64, pp. 253-72.7 2 [27] Sampson M. (1991), “The Effect of Parameter Uncertainty on Forecast Vari ances and Confidence Intervals for Unit Root and Trend Stationary Time-series Models,” J o u r n a l o f A p p li e d E c o n o m e t r ic s , 6,no 1,pp. 67-76. 27 [2S] Schmidt, Peter (1993), "Some Results on Testing for Stationarity Using Data Detrended in Differences,” Economics Letters, 41. pp. 1-6. [29] Schmidt, P. and P.C.B. Phillips (1992), “L M Tests for a Unit Root in the Pres ence of Deterministic Trends,” O x f o r d B u ll e t i n o f E c o n o m i c s a n d S t a t is t ic s , 54, pp. 257-87. [30] Spitzer, J.J. (1979), “Small Sample Properties of Nonlinear Least Squares and Maximum Likelihood Estimators in the Context of Autocorrelated Errors,” J o u r n a l o f th e A m e r i c a n S t a t is t ic a l A s s o c i a t io n , 74, pp. 41-47. [31] Stock, James H. (1991), “Confidence Intervals of the Largest Autoregressive Root in U.S. Macroeconomic Time Series,” J o u r n a l o f M o n e t a r y E c o n o m i c s , 28, pp. 435-60. [32] Summers, Robert and Alan Heston (1991), “The Penn World Table (Mark 5): An Expanded Set of International Comparisons, 1950-1988,” Q u a r t e r ly J o u r n a l o f E c o n o m i c s , 106, pp. 327-368. [33] Thornton, D.L. (1987), “A Note the Effect of the Cochrane-Orcutt Estimator in the AR(1) Regression Model,” J o u r n a l o f E c o n o m e t r ic s , 36, pp. 369-76. 28 k — 0.00 ** *= 0.01 *- 0 . 1 0 Figure 1 A Asymptotic Relative Efficiencies of Pq ^ A and 1.00 0 .2 0 .3 0.4 o Figure 2 Asymptotic Variance of $ q l s k — 0.00 0.01 **■ 0 . 1 0 Figure 3 A A A Asymptotic Relative Efficiencies of Pq q , Pqq anc* Ppy 1.00 0 .4 0 .3 0 .2 0.1 0 .0 Figure 4 Densities of Feasible GLS Estimators Table 1 Average Mean Square Error of Estimators A. Average MSE for -30<c<0 Estimator A £g l s £o l s ^FD ' a FCC £f p w ^PT 0 . 0 0 0 0 , 0 1 0 0.057 0.108 0.077 0.065 0.077 0.082 0.067 0 . 1 1 0 0.085 0.083 0.081 0.085 0.050 0.081 0.116 0 . 1 2 0 0.250 0.097 0.126 0 . 1 0 2 0 . 1 1 1 0 . 1 2 1 0.113 0.088 0.094 0.128 0.094 0.142 0 . 1 0 2 0 . 1 1 0 0 . 1 0 0 0.089 0 . 1 0 1 1 . 0 0 0 0.105 0.133 0.129 0.152 0.108 0.118 £g l s £o l s £ fd £fcc £f p w ^PT B . Average MSE for -2<c<0 K ------0 . 0 0 0 0.050 0 . 1 0 0 0 - 0 1 0 0.493 0.497 0.512 0.529 0.678 0.682 0.695 0.710 0.498 0.502 0.516 0.532 0.598 0.641 0.595 0.615 0.544 0.529 0.531 0.559 0.509 0.512 0.526 0.541 Estimator A £g l s £o l s ^FD £fcc £fpw ^PT C. Average MSE for -10<c<i-2 tc ------. 0 . 0 0 0 0 . 0 1 0 0.050 0 , 1 0 0 0.069 0.083 0.115 0.136 0.168 0.172 0.186 0.197 0.134 0.154 0.097 0.107 0.075 0.093 0.135 0.168 0.114 0.106 0.131 0.146 0 . 1 1 0 0.141 0.118 0.160 0.179 0.206 0.165 0.182 0.173 0.192 D. Average MSE for -■30<c<-10 K ------0 . 0 1 0 0 . 0 0 0 0.050 0 . 1 0 0 0.008 0.018 0.025 0.026 0.028 0.028 0.030 0.030 0.049 0.027 0.035 0.052 0.054 0.008 0.028 0.060 0.026 0 . 0 2 0 0 . 0 2 2 0.025 0.028 0.029 0.035 0.033 0,250 0.026 0.030 0.053 0.063 0.027 0.035 0.026 0.030 0.053 0.063 0.027 0.035 Estimator Estimator £g l s £o l s £ fd £fcc £f p w ^PT 0-250 0.566 0.743 0.567 0.676 0.590 0.574 0-250 0.158 0 . 2 1 1 1 . 0 0 0 0.634 0.808 0.635 0.766 0.664 0.649 l.ooo 0.168 0.219 0.190 0 . 2 2 0 1 . 0 0 0 Notes: The entries in the table are the mean squared error averaged over the indicated range of c. A. k - 0.00 (i) p-0.50 1 . 0 0 0 0.301 0.258 0.273 0.259 0.997 0.355 0.260 0.488 0.859 0.917 0.786 0.815 (iii) p-0.90 0.511 0.386 0.803 0.698 0.944 0.940 0.632 0.772 0.778 0.631 0.498 0.616 0.925 0.695 0.560 0.305 0.451 0.962 0.442 0.343 0.678 0.978 0.889 0.924 0.978 0.635 0.971 0.797 0.874 0.905 (iv) p-0.95 0.572 0.529 0.923 0.902 0.899 0.875 0.840 0.830 0.906 0.860 0.452 0.760 0.939 0.697 0.720 0.386 0.698 0.942 0.619 0.622 0.860 0.833 0.842 0.833 1.000 1.000 1.000 1.000 0.907 0.944 0.969 0.830 0.965 0.986 0.892 0.977 0.821 0.974 0.996 0.358 0.632 0.950 0.560 0.529 0.579 0.912 0.887 0.809 0.882 A &>LS ^FD a FCC ^FPW ^PT 0.941 1 . 0 0 0 0.350 0.272 m ) 0.258 0.113 0.714 0.466 0.947 0.839 0.675 0.607 0.810 0.922 0.838 0.805 A £o l s ^FD ^FCC £fpw ^PT 1 . 0 0 0 • • hi 1 . 0 0 0 ... TExact 0.958 0.185 0.974 0.975 0.953 (ii) p-0.80 0.305 0.451 0.968 0.454 0.363 A £o l s ^FD ^FCC ^FPW --- T-50 --Exact ■1 (1 ) 0.924 0.268 0.335 0.213 1 £o l s £ fd £fcc £f p w ^PT --- T-30 --Exact .I,( 1 1 0.891 0.283 0.492 0.330 0.990 0.958 0.390 0.957 0.705 0.276 o o r-t Table 2 Relative Efficiencies of Estimators Exact and 1(1) Approximation 0.631 0.670 0.937 0.824 0.664 (v) ^OLS ^FD ^FCC ^FPW ^PT p-1.00 0.833 1.000 1.000 0.859 0.903 0.964 0.966 0.993 0.992 0.850 1.000 Table 2 (Continued) Relative Efficiencies of Estimators Exact and 1(1) Approximation B. * - £o l s /?FD £fcc £fpw ^PT ... t -30 --Exact i m 0.856 0.950 0.550 0.463 0.470 0.981 1 . 0 0 0 0.979 0.698 0.735 1 . 0 0 (i) p-0.50 --- T-50 --i m Exact 0.966 0.902 0.308 0.381 0.960 0.328 0.979 0.974 0.932 0.859 (ii) P-0.80 0.867 0.817 0.667 0.698 0.870 0.591 0.974 0.958 0.699 0.721 £o l s £ fd a FCC £fpw ^PT 0.839 0.841 0.932 0.975 0.834 0.774 0.859 0.756 0.990 £o l s £ fd £f c c £fpw ^PT 0.801 0.970 0.916 0.950 0.942 0.753 0.971 0.863 0.989 0.971 (Hi) p-0.90 0.800 0.764 0.895 0.902 0.870 0.786 0.942 0.960 0.858 0.869 0.764 0.997 0.861 0.959 0.975 (iv) p-0.95 0.781 0.755 0.983 0.983 0.876 0.829 0.948 0.957 0.970 0.962 £o l s ^FD ^FCC ^FPW ^PT ^OLS ^FD ^FCC ^FPW hi 0.803 0.997 0 . 8 8 6 0.951 0.994 0 . 8 6 6 ... T-100 ... Exact i m 0.982 0.946 0.166 0.213 1 . 0 0 0 0.174 1 . 0 0 0 0.994 1 . 0 0 0 0.949 0.915 0.420 0.844 0.883 0.451 0.393 1 . 0 0 0 1 . 0 0 0 0.836 0.809 0.842 0.683 0.805 0.980 0.720 0.817 0.698 0.621 0.994 0.746 0.782 0.898 0.863 0.983 0 . 8 8 8 0.764 0.902 0.786 0.979 0.898 0.842 0.833 0.860 0.833 (v) p-1.00 0.850 0.833 1 . 0 0 0 1 . 0 0 0 1 . 0 0 0 1 . 0 0 0 1 . 0 0 0 1 . 0 0 0 0.936 0.989 0.895 0.991 0.927 1 . 0 0 0 1 . 0 0 0 1 . 0 0 0 0.879 0.965 0.996 0.863 0.958 0.980 0.838 0.959 0.981 1 . 0 0 0 Notes: The relative efficiency is the ratio of the variance of the infeasible GLS estimator to the variance of the estimator given in column 1. The columns labeled 1(1) are the asymptotic relative efficiencies using c-T(p-l). The corresponding 1 (0 ) relative efficiencies are 1 , 0 , 1 , 1 , 1 , 1 , respectively for the estimators in column 1 and for all T and |p|<1. Table 3 Confidence Interval Coverage Rates (%) A A A. Smallest of anc* (i) Level 90.0 90.0 90.0 95.0 95.0 95.0 99.0 99.0 99.0 K 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 95.0 95.0 95.0 99.0 99.0 99.0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 95.0 95.0 95.0 99.0 99.0 99.0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 .1 z!0 _ z2Q_ 96.0 95.6 94.0 97.8 97.4 97.1 97.3 96.6 96.8 94.5 94.5 94.3 93.0 92.9 93.6 97.9 97.9 96.7 98.8 98.7 98.5 98.6 98.3 98.5 96.9 97.1 96.9 97.9 98.0 97.9 99.4 99.5 99.2 99.7 99.6 99.6 99.5 99.6 99.7 99.0 99.1 99.2 T - 50 90.7 90.8 90.5 97.3 96.3 95.2 98.3 97.9 97.9 98.2 97.9 98.1 96. 96. 96. 94.9 94.8 94.9 98.7 98.3 97.7 99.2 99.0 99.0 99.3 99.1 99.1 98. 98. 98. 98.6 98.7 98.8 99.7 99.6 99.6 99.8 99.8 99.8 99.9 99.8 99.9 99. 99. 99. r - 100 97.8 97.5 96.4 98.8 98.4 98.3 98.6 98.8 98.8 98.5 98.3 98.3 (ill) 90.0 90.0 90.0 T - 30 88.7 88.7 89.2 0 ( W 90.0 90.0 90.0 Confidence Intervals 91.4 92.3 91.9 95.6 95.8 95.9 99.0 99.0 98.5 99.5 99.3 99.3 99.5 99.5 99.6 99.4 99.4 99.4 99.1 98.8 99.0 99.8 99.9 99.8 99.9 99.9 99.9 100.0 99.9 100.0 99.9 99.9 99.9 Table 3 (Continued) Confidence Interval Coverage Rates (%) B. Confidence Intervals Constructed from 1 . 0 89.0 93.0 92.9 93.4 97.8 97.8 96.6 98.7 98.5 98.5 98.5 | 96.5 98.2 96.5 98.4 96.5 97.8 97.9 97.9 99.3 99.5 99.2 99.7 99.6 99.7 99.5 99.5 99.6 98.9 99.0 99.0 (ii) T - 50 90.5 97.1 96.2 90.7 90.3 95.1 98.4 97.8 97.8 98.4 97.9 98.0 96.5 96.4 96.4 0 . 1 90.0 90.0 90.0 95.0 95.0 95.0 99.0 99.0 99.0 l ro O 88.5 8 8 . 6 99.0 99.0 99.0 99.0 99.0 99.0 97.2 96.5 96.5 0 . 0 0 . 1 0 . 1 95.0 95.0 95.0 -5 97.7 97.3 97.1 -l 96.0 95.4 93.7 95.0 95.0 95.0 90.0 90.0 90.0 0 K 1 Level 90.0 90.0 90.0 K. (i) 0 . 0 1 . 0 0 . 0 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 . 0 0 . 1 1 . 0 0 - 1 0 93.9 93.6 93.6 94.8 94.7 94.9 98.6 98.2 97.6 99.1 99.0 99.0 99.3 99.1 99.1 98.5 98.2 98.3 98.6 98.7 98.8 99.7 99.5 99.6 99.8 99.8 99.7 99.9 99.8 99.9 99.7 99.6 99.7 98.9 98.4 98.3 98.9 98.8 98.8 98.6 98.2 98.2 99.6 99.5 99.6 99.4 99.3 99.3 (Hi) T - 100 91.3 97.7 92.1 97.5 91.8 96.3 95.6 95.8 95.8 99.0 98.9 98.5 99.5 99.3 99.3 99.1 98.8 99.0 99.8 99.8 99.8 99.9 99.9 99.9 1 0 0 . 0 99.9 99.9 99.9 99.8 99.9 Notes: The table shows the exact coverage rates (in percent) for conservative confidence intervals constructed with an asymptotic level given in the first column. The confidence intervals in panel A were constructed as the narrowest of the intervals constructed from the OLS and first-difference estimators. The confidence intervals from panel B were constructed from the Prais-Winsten estimator. Table 4 Annual Real Per-Capita Growth Rates - 1 2 C ou ntry Sn irl P e r i o d c t l ALGERIA 1960 1990 ^OLS2.736 ^FD— 1.459 ^FCO2.788 ^FFW2.098 -6 .0 7 0 2 ANGOLA 1960 1989 -2 .0 3 8 -1 .0 0 4 -2 .7 5 9 -1 .3 4 5 -3 .7 5 3 -2 .2 9 8 3 B EN IN 1959 1989 -0 .4 0 3 -0 .3 7 0 -0 .5 8 1 -0 .3 9 7 -1 5 .1 8 0 -3 .2 3 0 4 BOTSWANA 1960 1989 5.839 6.079 6 .1 7 3 5 .9 0 6 -1 0 .9 5 5 -3 .2 6 2 5 BURKINA FASO 1959 1990 0.859 0.006 -2 0 .3 7 8 0.006 0 .2 9 0 -3 .4 9 8 6 BURUNDI 1960 1990 0.567 -0 .4 2 5 2.043 0 .0 4 1 -5 .5 8 6 -4 .2 4 7 7 CAMEROON 1960 1990 2.698 1.888 1.455 2.052 -2 .4 3 9 -1 .8 1 6 8 CAFE VERDE I S . 1960 1989 3.880 3 .3 3 8 4.693 3.545 -4 .3 6 9 -2 .5 4 8 9 CENTRAL A F R . R . 1960 1990 -0 .4 8 6 -0 .5 8 9 -0 .8 7 9 -0 .5 4 2 -5 .3 3 1 j- l . 164 10 CHAD 1960 1990 -2 .5 8 4 -2 .0 1 0 -2 .5 8 0 -2 .4 4 2 -1 2 .2 0 0 -2 .6 2 8 -1 .4 8 4 1 1 COMOROS 1960 1987 -0 .0 4 4 0.520 -0 .7 8 0 0.219 -6 .4 9 0 -3 .0 9 6 12 CONGO 1960 1990 3.314 2.788 3 .3 6 0 3.082 -7 .0 9 1 -2 .4 5 5 14 EG YP T 1950 1990 3.002 2.385 3.696 2.481 -2 .0 1 8 -3 .7 2 5 15 ETHIOPIA 1951 1986 0.831 0.669 0.688 0.761 -7 .4 5 2 -1 .6 1 4 16 GABON 1960 1990 2.298 2 .6 2 0 - 3 7 . 3 4 2 2.616 -0 .3 5 9 -1 .4 1 3 17 GAMBIA 1960 1990 1.1 5 0 0 .9 3 5 -0 .2 2 2 1.015 -4 .2 4 3 -1 .1 1 2 18 GHANA 1955 1989 -0 .2 6 6 -0 .0 8 9 -0 .4 2 5 -0 .2 0 5 -9 .5 0 1 -2 .4 1 8 19 G U INEA 1959 1989 -0 .3 0 4 -0 .2 1 5 -0 .5 9 0 -0 .2 6 0 -6 .1 3 0 -2 .1 0 8 1.036 -5 .6 5 4 -2 .0 8 9 20 G U IN E A -B IS S 1960 1990 0 .3 7 7 -0 .2 1 2 0.724 21 IV O R Y COAST 1960 1990 1 .0 7 3 0.633 5.807 0 .6 3 3 2.305 0.611 22 KENYA 1950 1990 1 .1 7 9 1 .1 6 6 1.191 1.177 -1 7 .9 1 9 -2 .8 7 8 23 L ESOTHO 1960 1990 4.402 4 .0 5 3 2.0 9 9 4.129 -2 .5 8 5 -1 .6 8 1 1986 0.682 0 . 3 1 0 449 .3 8 3 0.310 0 .0 2 3 0.194 24 LIB E R IA 1960 25 MADAGASCAR 1960 1990 -1 .9 6 2 -1 .8 1 8 -2 .4 4 1 -1 .8 8 0 -5 .0 8 2 -2 .0 6 4 26 MALAWI 1954 1990 1 .1 7 1 1 .2 2 6 0.684 1 .197 -6 .7 7 4 -2 .0 1 6 27 MAL I 1960 1990 0.877 0 .1 5 0 3.057 0 .2 7 4 -2 .1 1 7 -2 .7 5 8 28 M AUR IT ANIA 1960 1990 -0 .1 6 4 -0 .2 0 7 -0 .6 4 1 -0 .1 8 2 -7 .6 5 2 -1 .6 6 9 29 M AUR IT IUS 1950 1990 1.340 1.399 3.893 1.385 -2 .7 2 9 -1 .6 7 7 30 MOROCCO 1950 1990 2.814 2.355 2 .8 0 4 2.600 -7 .0 1 7 -2 .6 8 9 31 MOZAMBIQUE 1960 1990 -2 .3 0 9 -1 .4 2 6 -7 .8 5 6 -1 .4 9 3 -1 .2 0 7 -2 .2 4 0 32 N AM IB IA 1960 1989 0 .3 8 4 0.509 -5 .5 9 1 0 .4 9 3 -1 .7 3 4 -1 .4 0 9 33 N IG E R 1960 1989 -0 .4 1 5 -0 .2 5 6 -4 .6 1 2 -0 .2 8 8 -2 .3 8 7 -1 .9 2 4 34 NIG ER IA 1950 1990 1.989 1 .3 3 7 -0 .2 5 5 1.475 -2 .5 7 8 -1 .8 0 8 35 REUNION 1960 1988 3.764 3.799 2.695 3.784 -4 .6 4 1 -2 .2 5 2 36 RWANDA 1960 1990 1 .974 0.791 1 .9 1 6 1.366 -5 .8 4 2 -2 .5 1 6 37 SENEGAL 1960 1990 0.138 0.204 0.101 0.139 -2 8 .5 8 2 -3 .9 6 2 38 SE YC HE LLE S 1960 1989 3.896 3.449 4.072 3 .7 8 3 -1 1 .8 6 7 -2 .2 8 9 39 S IE R R A LEONE 1961 1990 0.049 0.593 -4 .6 2 7 0.519 -1 .7 8 9 -2 .2 3 9 40 SOMALIA 1960 1989 -0 .4 4 8 -0 .5 5 1 -0 .3 8 7 -0 .4 6 0 -1 8 .5 2 9 -2 .4 8 6 41 SOUTH A F R IC A 1950 1990 1.792 1.3 4 3 - 1 6 . 3 0 4 1.3 44 -0 .2 0 8 -0 .1 8 6 42 SUDAN 1971 1990 -0 .2 8 6 -0 .8 8 6 -0 .3 7 3 -0 .5 0 9 -7 .4 1 6 -2 .5 1 8 43 SWAZILAND 1960 1989 1.626 1.985 -4 .7 0 4 1 .9 3 2 -1 .9 1 0 -1 .2 6 1 44 T A N Z A N IA 1960 1988 1.5 3 2 1.686 0 .7 1 0 1.626 -4 .4 3 1 -1 .9 1 9 45 TOGO 1960 1990 1. 777 1.7 9 7 -1 .4 8 5 1.7 9 3 -2 .1 3 0 -1 .4 0 8 46 TU N ISIA 1960 1990 3.761 3.222 2.980 3.298 -1 .8 5 5 -1 .1 2 3 47 UGANDA 1950 1989 -0 .1 8 8 0.946 -0 .0 3 1 0 .1 7 7 -1 0 .5 5 8 -2 .5 7 5 48 ZAIRE 1950 1989 0.339 0.648 19.1 57 0.648 0.527 -1 .2 5 0 49 ZAMBIA 1955 1990 -0 .6 1 3 -0 .5 9 7 -2 .6 3 3 -0 .6 0 2 -3 .1 9 5 -1 .1 2 0 50 ZIMBABWE 1954 1990 0.904 1.018 0.795 0 .9 5 0 -8 .2 9 2 -2 .7 6 8 Table 4 (Continued) Annual Real Per-Capita Growth Rates ID Country Sm nl P e r i o d c r 52 BARBADOS 1960 1989 ^OLS3.4 7 1 2 .2 5 7 ^FP W 3 .5 8 5 -2 .7 2 2 - 1 . 572 54 CANADA 1950 1990 2 .7 6 2 2.503 2.849 2.626 -5 .9 7 4 - 2 . 255 55 COSTA R I C A 1950 1990 2.310 2.363 0.049 2 .3 5 5 -1 .8 6 7 - 2 . 151 57 DOMINICAN REP. 1950 1990 2.425 1.9 7 9 2 .0 5 0 2.2 5 7 -9 .0 6 8 - 1 . 127 58 EL SALVADOR 1950 1990 1.0 63 1.0 0 9 -4 .3 2 0 1.0 1 1 -0 .8 1 3 - 1 . 965 60 GUATEMALA 1950 1990 1 .239 0.790 -6 .3 9 2 0. 794 -0 .3 1 8 - 0 . 985 61 H A ITI 1960 1989 0.1 4 7 -0 .3 3 1 0.083 -0 .0 8 1 -6 .4 0 5 - 1 . 177 - 1 . 664 ^FD— 3.619 & TCCT 62 HONDURAS 1950 1990 1 .101 0.788 0 .6 6 0 0 .8 8 6 -3 .6 9 2 63 JAMAICA 1953 1989 1.4 20 2 .0 2 0 13 .2 76 2.020 0.732 - 1 . 810 64 MEXICO 1950 1990 2 .5 3 6 2.259 1 .738 2 .3 3 5 -3 .2 4 8 - 1 . 718 65 NICARAGUA 1950 1987 1 .0 2 1 0.943 -1 1 .8 2 8 0.946 -0 .8 6 9 - 1 . 186 66 PANAMA 1950 1990 2 .8 2 1 2.181 1 .626 2.331 -2 .8 1 3 -1 . 202 67 PUERTO R IC O 1955 1989 3 .6 4 9 3 .9 3 0 0 .1 9 3 3.911 -1 .1 2 3 - 1 . 902 70 T R I N I D AD&TOB AG 1950 1990 2.870 2 .5 9 6 -4 .6 1 3 2 .6 1 5 -1 .1 5 4 - 0 . 750 71 U .S .A . 1950 1990 1.9 40 1.8 94 1 .959 1 .926 -1 1 .4 8 8 - 2 . 776 72 ARG ENTIN A 1950 1990 0 .9 2 2 0.366 -1 .1 2 3 0 .453 -2 .0 3 4 - 0 . 307 73 B OLIVIA 1950 1990 1.3 17 0 .6 3 2 0 .4 9 8 0 .7 1 1 -1 .6 2 3 - 1 . 319 74 B R A Z IL 1950 1990 3.469 2 .8 5 8 -1 .3 2 2 2.881 -0 .7 7 9 -o.613 75 C HILE 1950 1990 0.925 1.234 0.801 1.0 38 -9 .3 9 1 - 3 . 067 76 COLOMBIA 1950 1990 2.146 1.92 7 2 .2 0 2 2.017 -4 .9 6 3 - 1 . 607 77 ECUADOR 1950 1990 2 .7 5 1 2 .1 6 5 0.848 2 .2 1 7 -1 .3 6 4 - 1 . 174 78 GUYANA 1950 1990 -0 .2 1 8 -0 .9 9 8 -0 .7 6 3 -0 .7 0 2 -4 .5 2 1 - 1 . 592 79 PARAGUAY 1950 1990 2.068 1.4 07 2.636 1 .596 -3 .3 7 6 - 2 . 082 80 PERU 1950 1990 1 .4 0 6 0.886 6.271 0 .8 8 6 1. 957 - 0 . 197 81 SURINAME 1960 1989 1.398 0.418 -2 1 .1 9 5 0 .4 3 7 -0 .5 4 0 - 0 . 232 82 URUGUAY 1950 1990 0 .3 7 2 0 .5 7 9 0.251 0 .4 3 7 -1 0 .9 1 0 - 2 . 907 83 VENEZUELA 1950 1990 0 .4 3 9 0 .5 4 9 6.766 0 .5 4 9 1. 237 - 1 . 338 85 BANGLADESH 1959 1990 1 .2 0 8 1.392 1 .183 1.2 61 -1 0 .8 9 5 - 2 . 813 87 C HI NA 1968 1990 5 .7 5 2 5 .984 5 .5 5 6 5.8 84 -4 .7 2 8 - 2 . 310 - 3 . 896 88 HONG KONG 1960 1990 6 .264 6.250 6 .0 5 1 6.261 -1 1 .7 0 5 89 IN D IA 1950 1990 1.43 7 1.79 4 1.6 5 5 1.6 0 8 -6 .7 9 4 - 1 . 473 90 IN D O N E S IA 1960 1990 4.471 3 .7 7 9 - 4 8 . 7 2 3 3 .7 7 9 0.0 9 7 - 3 . 505 91 IRAN 1955 1989 1.91 3 1 .528 - 5 5 . 1 9 4 1 .529 -0 .2 3 3 - 1 . 186 92 IRAQ 1953 1987 1.7 67 0 .4 7 6 -2 .7 6 7 0.8 0 1 -2 .9 5 6 - 0 . 791 93 ISRAEL 1953 1990 3 .6 2 0 3.6 3 7 10 .6 01 3 .6 3 7 0 .9 2 6 - 1 . 004 94 JAPAN 1950 1990 5.781 5.742 14 .0 77 5.742 0 .7 2 9 - 0 . 743 95 JORDAN 1954 1990 3.589 3.110 1 .6 9 8 3 .2 7 8 -4 .1 0 3 - 1 . 845 96 KOREA, REP. 1953 1989 5 .944 5 .6 9 2 -3 6 7 . 4 7 0 99 M ALA YSIA 1955 1990 4.2 64 3.871 101 MYANMAR 1950 1989 2.411 102 NEPAL 1960 1986 1.92 4 104 PAKIS TAN 1950 1990 2.355 105 P H IL IP PIN E S 1950 1990 2.001 108 SINGAPORE 1960 1990 6 .724 109 S R I LANKA 1950 1989 1. 854 110 SY R IA 1960 1990 3 .7 0 2 5.69 0.017 - 1 . 906 4.5 4 7 4.0 54 -5 .6 8 5 - 2 . 818 2.563 2 .3 1 8 2 .4 4 8 -1 3 .4 6 8 1.5 47 2 .1 2 7 1.8 00 -9 .2 3 3 2 .1 2 6 2 .6 1 4 2 .2 3 4 -5 .8 9 9 - 2 . 323 2 .0 7 3 -7 .9 9 6 2 .0 7 2 -0 .4 5 9 - 2 . 569 6.190 6 .3 4 7 6 .2 8 7 -2 .2 2 4 - 1 . 638 1. 838 2.454 1.844 -4 .1 0 2 - 1 . 325 3 .2 2 2 2 .9 5 8 3.489 -7 .0 4 8 - 0 . 608 -2 . 110 - 2 . 342 Table A (Continued) Annual Real Per-Capita Growth Rates ID C ountry S to o l P a rlo d -r ^FD — 5 .6 0 3 ^FCO6.659 ^F£W5 .6 1 3 c _ r 111 TAIWAN 1951 1990 A dls5 .6 5 3 -2 .3 7 1 -2 .5 0 2 112 TH AILAN D 1950 1990 3.922 3 .5 7 0 3 7 .9 8 1 3.571 -0 .1 8 7 -3 .6 1 4 1.9 8 9 5.151 114 YEMEN 1969 1989 4.727 5.676 3.739 5 .3 2 9 -3 .9 2 5 -2 .0 8 3 2.950 8.402 115 AUSTRIA 1950 1990 3 .6 4 0 3.664 25.762 3 .6 6 4 0.176 -0 .9 6 7 2.578 4.750 115 BELGIUM 1950 1990 2.908 2 .7 6 7 2.788 2.803 -3 .0 4 3 -1 .3 6 6 1 .7 9 5 3.739 113 CYPRUS 1950 1990 3.970 4 .0 9 8 3.994 3.994 -1 6.3 47 -3 .7 3 3 2.363 5.788 119 CZECHOSL OVAKIA 1960 1990 3.315 3 .0 4 1 -0 .3 9 4 3.061 -1 .1 7 9 -0 .6 8 5 1 .1 2 1 4 .9 6 1 4.288 6 .9 1 8 120 DENMARK 1950 1990 2.644 2.412 2.378 2.487 -3 .8 4 3 -1 .2 8 5 1 .3 2 2 3 .5 0 1 121 F IN LA N D 1950 1990 3 .4 3 4 3.452 3.249 3.441 -9 .0 6 3 -2 .6 6 2 2.075 4.828 122 FRANCE 1950 1990 3.080 3.008 11.8 23 3.008 0.351 -0 .5 5 5 2.014 4.002 123 GERM ANY, WEST 1950 1990 3.199 3.576 5.273 3.576 2.718 -2 .8 8 1 2.174 4.978 124 GREECE 1950 1990 4.328 3.887 1 9.217 3 .8 8 7 0.318 0.130 2.369 5.4 0 4 125 HUNGARY 1970 1990 2 .2 3 4 2.322 8.202 2.322 1 .4 0 2 -0 .8 3 6 0.4 7 5 4.169 126 IC ELA N D 1950 1990 3.422 2.969 3.322 3 .2 8 0 -1 0 .9 1 3 -3 .1 4 3 1 .0 0 4 4.933 127 IR ELA N D 1950 1990 3.207 3.102 3.445 3 .1 5 1 -5 .8 0 1 -2 .7 0 2 1.716 4 .4 8 9 128 ITA LY 1950 1990 3.752 3 .7 4 9 14.7 7 9 3.749 0.359 -0 .8 4 6 2.7 0 5 4.793 129 LUXEMBOURG 1950 1990 2.185 2.246 2.297 2.199 -1 4 .2 7 8 -3 .1 1 0 0.843 3.649 130 M ALTA 1954 1989 5 .4 9 6 5 .024 6.840 5.1 0 4 -2 .1 4 8 -1 .8 4 8 3.072 6.976 131 NETHERLANDS 1950 1990 2.763 2.588 1.6 04 2.611 -1 .7 9 5 -1 .4 6 7 1.222 3 .9 5 4 132 NORWAY 1950 1990 3.346 3.051 3.125 3.138 -3 .4 9 0 -2 .0 7 0 2.114 3 .9 8 8 133 POLAND 1970 1990 0 .6 9 4 1.242 -1 8 .3 5 4 1 .224 -0 .7 2 6 -2 .7 8 5 -1 1 .7 9 5 1 4.280 134 PORTUGAL 1950 1990 4 .3 2 0 4.213 2.890 4.228 -1 .8 8 5 -1 .2 5 0 2.450 5.975 136 S P A IN 1950 1990 3.786 3.998 -9 .2 5 2 3.995 -0 .3 8 6 -0 .9 8 8 2.297 5.699 137 SWEDEN 1950 1990 2.375 2.312 0.048 2 .3 1 4 -0 .8 1 7 -1 .1 7 0 1 .4 7 8 3.146 138 SW ITZER LAND 1950 1990 2.083 2.219 1 .174 2.1 8 9 -2 .6 7 0 -1 .6 9 0 0.918 3.519 139 TURKEY 1950 1990 2.7 4 6 3 .1 4 4 2 .4 6 0 2.875 -1 0 .5 8 9 -2 .5 3 4 1.511 4.778 140 U .K . 1950 1990 2.241 2.306 2.249 2.253 -1 6 .9 1 5 -4 .5 0 5 2.072 2.438 141 U .S .S .R . 1970 1989 3.272 3.377 -9 .9 2 7 3.376 -0 .2 1 5 -1 .9 8 1 2.679 4.074 142 YU G O S LA VIA 1960 1990 3.630 2.812 13.4 77 2.812 1.129 0.919 0.236 5 .3 8 8 143 A U S TR A LIA 1950 1990 2 .1 8 4 1.8 7 0 2.158 2.086 -1 0 .9 7 3 -2 .0 6 5 0.752 2.987 144 F IJ I 1960 1990 2.043 2.0 0 6 1 .853 2.021 -4 .6 2 9 -1 .5 1 6 -0 .3 9 2 4.405 145 NEW ZEALAND 1950 1990 1 .674 1 .3 8 8 1 .5 9 9 1.559 -8 .3 5 3 -2 .0 2 1 0.035 2 .7 4 2 146 PAPUA N .G U IN E A 1960 1990 0.215 0.643 3.196 0 .6 4 3 3 .4 2 0 -2 .6 8 9 -1 .2 2 9 2 .5 1 5 Notes: The column labeled ID shows the country ID from the Penn World Tables. The estimators ^q l s * ^FD’ ^FCO’ ^PW are described in the text; c *s an estimate of the local-to-unity paramater, constructed as T(p-l); rr is the augmented Dickey-Fuller t-statistic; and /J are the endpoints of the 95% confidence interval for /} constructed using the Prais-Winsten estimator, as described in the text. Working Paper Series A series of research studies on regional economic issues relating to the Seventh Federal Reserve District, and on financial and economic topics. REGIONAL ECONOMIC ISSUES Estimating Monthly Regional Value Added by Combining Regional Input With National Production Data WP-92-8 P h ilip R . Isra ile v ic h a n d K en n eth N . K u ttn er Local Impact of Foreign Trade Zone WP-92-9 D a v id D . W eiss Trends and Prospects for Rural Manufacturing WP-92-12 W illiam A . T esta State and Local Government Spending-The Balance Between Investment and Consumption WP-92-14 R ic h a rd H . M a tto o n Forecasting with Regional Input-Output Tables WP-92-20 P J i. Isra ile v ic h , R . M a h id h a ra , a n d G J D . H e w in g s A Primer on Global Auto Markets WP-93-1 P a u l D . B a lle w a n d R o b e rt H . S ch n orbu s Industry Approaches to Environmental Policy in the Great Lakes Region WP-93-8 D a v id R . A lla rd ic e , R ic h a rd H . M a tto o n a n d W illiam A. T esta The Midwest Stock Price Index-Leading Indicator of Regional Economic Activity WP-93-9 W illiam A . S tra u ss Lean Manufacturing and the Decision to Vertically Integrate Some Empirical Evidence From the U.S. Automobile Industry WP-94-1 T hom as H . K lie r Domestic Consumption Patterns and the Midwest Economy W P-94-4 R o b e rt S ch n orbu s a n d P a u l B a lle w 1 Working paperseriescontinued To Trade or Not to Trade: Who Participates in RECLAIM? WP-94-11 Thom as H . K lie r a n d R ic h a r d M attoon Restructuring & Worker Displacement in the Midwest WP-94-18 P a u l D . B a lle w a n d R o b e rt H . S chn o rb u s ISSUES IN FINANCIAL REGULATION Incentive Conflict in Deposit-Institution Regulation: Evidence from Australia WP-92-5 E d w a rd J . K a n e a n d G eo rg e G . K aufm an Capital Adequacy and the Growth of U.S. Banks WP-92-11 H e rb e rt B a e r an d Jo h n M c E lra v e y Bank Contagion: Theory and Evidence WP-92-13 G e o rg e G . Kaufm an Trading Activity, Progarm Trading and the Volatility of Stock Returns WP-92-16 Ja m e s T . M o s e r Preferred Sources of Market Discipline: Depositors vs. Subordinated Debt Holders WP-92-21 D o u g la s D . E v a n o ff An Investigation of Returns Conditional on Trading Performance WP-92-24 Ja m e s T . M o se r a n d J a c k y C . So The Effect of Capital on Portfolio Risk at Life Insurance Companies WP-92-29 E lija h B r e w e r I I I , Thom as H . M o n d sch ea n , an d P h ilip E . Strahan A Framework for Estimating the Value and Interest Rate Risk of Retail Bank Deposits WP-92-30 D a v id E . H u tch iso n , G eo rg e G . P en n a cch i Capital Shocks and Bank Growth-1973 to 1991 WP-92-31 H e rb e rt L . B a e r a n d Jo h n N . M c E lra v e y The Impact of S&L Failures and Regulatory Changes on the CD Market 1987-1991 WP-92-33 E lija h B r e w e r a n d Thom as H . M on dschean 2 Working paperseriescontinued Junk Bond Holdings, Premium Tax Offsets, and Risk Exposure at Life Insurance Companies WP-93-3 E lija h B r e w e r III a n d T h om as H . M on dsch ean Stock Margins and the Conditional Probability of Price Reversals WP-93-5 P a u l K ofm an a n d J a m e s T. M o se r Is There Lif(f)e After DTB? Competitive Aspects of Cross Listed Futures Contracts on Synchronous Markets P a u l K o fm a n , T ony B ou w m an a n d J a m e s T. M o se r W P-93-11 Opportunity Cost and Prudentiality: A RepresentativeAgent Model of Futures Clearinghouse Behavior H e rb e rt L . B a e r , V irgin ia G. F ra n c e a n d J a m e s T. M o se r WP-93-18 The Ownership Structure of Japanese Financial Institutions WP-93-19 H esn a G en a y Origins of the Modem Exchange Clearinghouse: A History of Early Clearing and Settlement Methods at Futures Exchanges WP-94-3 J a m e s T. M o se r The Effect of Bank-Held Derivatives on Credit Accessibility A. M in ton a n d J a m e s T. M o se r WP-94-5 E lija h B r e w e r III , B e rn a d ette Small Business Investment Companies: Financial Characteristics and Investments WP-94-10 E lija h B r e w e r III a n d H e sn a G en a y MACROECONOMIC ISSUES An Examination of Change in Energy Dependence and Efficiency in the Six Largest Energy Using Countries-1970-1988 WP-92-2 J a c k L . H e rv e y Does the Federal Reserve Affect Asset Prices? WP-92-3 Vefa T arhan Investment and Market Imperfections in the U.S. Manufacturing Sector WP-92-4 P a u la R. W orthin gton 3 Working paperseriescontinued Business Cycle Durations and Postwar Stabilization of the U.S. Economy W P-92-6 M a r k W. W atson A Procedure for Predicting Recessions with Leading Indicators: Econometric Issues and Recent Performance WP-92-7 Ja n ie s H . S to c k a n d M a r k W. W atson Production and Inventory Control at the General Motors Coxporation During the 1920s and 1930s WP-92-10 A n il K . K a sh ya p an d D a v id W . W ilco x Liquidity Effects, Monetary Policy and the Business Cycle WP-92-15 L a w re n c e J . C h ristia n o an d M a rtin Eich en bau m Monetary Policy and External Finance: Interpreting the Behavior of Financial Flows and Interest Rate Spreads WP-92-17 K en n eth N . K u ttn e r Testing Long Run Neutrality WP-92-18 R o b e rt G . K in g a n d M a rk W. W atson A Policymaker's Guide to Indicators of Economic Activity C h a rle s E v a n s , Steven Stro n g in , and F ra n c e s c a E u g e n i WP-92-19 Barriers to Trade and Union Wage Dynamics WP-92-22 E lle n R . R issm a n Wage Growth and Sectoral Shifts: Phillips Curve Redux WP-92-23 E lle n R . R issm a n Excess Volatility and The Smoothing of Interest Rates: An Application Using Money Announcements WP-92-25 Steven S tro n g in Market Structure, Technology and the Cyclicality of Output WP-92-26 B r u c e P e te rse n a n d Steven Stron gin The Identification of Monetary Policy Disturbances: Explaining the Liquidity Puzzle WP-92-27 Steven S trongin 4 Working paperseriescontinued Earnings Losses and Displaced Workers WP-92-28 L o u is S . J a c o b s o n , R o b e r t J. L a L o n d e, a n d D a n ie l G . Sullivan Some Empirical Evidence of the Effects on Monetary Policy Shocks on Exchange Rates WP-92-32 M a rtin E ich en bau m a n d C h a rle s E v a n s An Unobserved-Components Model of Constant-Inflation Potential Output WP-93-2 K en n eth N . K u ttn er Investment, Cash Flow, and Sunk Costs W P-93-4 P a u la R . W orthin gton Lessons from the Japanese Main Bank System for Financial System Reform in Poland WP-93-6 T akeo H o sh i, A n il K a sh y a p , a n d G a ry L ovem an Credit Conditions and the Cyclical Behavior of Inventories C. S tein WP-93-7 A n il K . K a s h y a p , O w en A . L a m o n t a n d J erem y Labor Productivity During the Great Depression WP-93-10 M ic h a e l D . B o rd o a n d C h a rle s L . E va n s Monetary Policy Shocks and Productivity Measures in the G-7 Countries WP-93-12 C h a rle s L . E v a n s a n d F ern a n d o S a n to s Consumer Confidence and Economic Fluctuations WP-93-13 Joh n G . M a tsu sa k a a n d A rg ia M . S b o rd o n e Vector Autoregressions and Cointegration WP-93-14 M a rk W. W atson Testing for Cointegration When Some of the Cointegrating Vectors Are Known WP-93-15 M ic h a e l T. K . H o rv a th a n d M a rk W. W atson Technical Change, Diffusion, and Productivity WP-93-16 J effrey R . C a m p b e ll 5 Working paperseriescontinued Economic Activity and the Short-Term Credit Markets: An Analysis of Prices and Quantities B en ja m in M , F ried m a n a n d K en n eth N . K u ttn er WP-93-17 Cyclical Productivity in a Model of Labor Hoarding WP-93-20 A r g ia M . S b o rd o n e The Effects of Monetary Policy Shocks: Evidence from the Flow of Funds WP-94-2 L a w re n c e J. C h ristia n o , M a rtin E ich en b a u m a n d C h a rle s E v a n s Algorithms for Solving Dynamic Models with Occasionally Binding Constraints /. C h ristia n o a n d J o n a s D M , F ish e r WP-94-6 Identification and the Effects of Monetary Policy Shocks WP-94-7 L a w re n c e L a w re n c e J . C h ristia n o , M a rtin E ich en b a u m a n d C h a rle s L . E v a n s Small Sample Bias in GMM Estimation of Covariance Structures WP-94-8 J o se p h G . A lto n ji a n d L e w is M . S e g a l Interpreting the Procyclical Productivity of Manufacturing Sectors: External Effects of Labor Hoarding? WP-94-9 A r g ia M . S b o rd o n e Evidence on Structural Instability in Macroeconomic Time Series Relations WP-94-13 J a m e s H . S to ck a n d M a rk W. W atson The Post-War U.S. Phillips Curve: A Revisionist Econometric History WP-94-14 R o b e r t G . K in g a n d M a rk W. W atson The Post-War U.S. Phillips Curve: A Comment WP-94-15 C h a rle s L . E v a n s Identification of Inflation-Unemployment WP-94-16 B e n n e tt T. M c C a llu m The Post-War U.S. Phillips Curve: A Revisionist Econometric History Response to Evans and McCallum R o b e r t G . K in g a n d M a rk W. W atson W P-94-17 6 Working paperseriescontinued Estimating Deterministic Trends in the Presence of Serially Correlated Errors WP-94-19 E u g en e C a n je ls a n d M a rk W. W atson 7