WPS7045 Policy Research Working Paper 7045 Testing Weak Exogeneity in Cointegrated Panels Enrique Moral-Benito Luis Serven Development Research Group Macroeconomics and Growth Team September 2014 Policy Research Working Paper 7045 Abstract For reason of empirical tractability, analysis of cointe- is obtained by first letting the time dimension of the panel grated economic time series is often developed in a partial go to infinity and then letting its cross-sectional dimen- setting, in which a subset of variables is explicitly mod- sion go to infinity. The paper evaluates the accuracy of eled conditional on the rest. This approach yields valid the asymptotic approximation in finite samples via sim- inference only if the conditioning variables are weakly ulation experiments. Finally, as an empirical illustration, exogenous for the parameters of interest. This paper pro- the paper reports tests of weak exogeneity of disposable poses a new test of weak exogeneity in panel cointegration income and wealth in an aggregate consumption equation. models. The test has a limiting Gumbel distribution that This paper is a product of the Macroeconomics and Growth Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at lserven@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Testing Weak Exogeneity in Cointegrated Panels* Enrique Moral-Benito Luis Serven Banco de España The World Bank JEL Codes: C23, C32. Keywords: Panel data, cointegration, weak exogeneity, Mote Carlo methods. * We thank A. Ludwig for kindly providing us with the data. 1 Introduction Analysis of economic time series under cointegration is often developed using their autoregressive– error correction model (VAR-ECM) representation. In many cases, however, the number of variables involved makes impractical the full-system approach in which all of them are jointly modeled (e.g. Johansen, 1988). Thus, researchers often resort to conditional (or partial) VAR-ECM analysis, which models a smaller set of variables conditional on the rest. Conditional modeling relates naturally to economic theory, which typically speci…es relations involving endogenous variables along with a (potentially large) number of exogenous variables. Thus, the conditional approach may be a logical choice if theory makes clear predictions regarding the relations among a subset of variables, but o¤ers little guidance on how the remaining variables are determined. Alternatively, the choice may be dictated by empirical tractability, when the large dimensionality of the full system is likely to make its empirical analysis overly cumbersome. However, the conditional approach yields e¢ cient inference only if the variables in the condi- tioning set are weakly exogenous with respect to the parameters of interest. If this is not the case, e¢ cient inference generally requires a full-system approach. In the context of cointegrated time se- ries, the parameters of interest typically are those of the cointegrating vector(s), in which case the conditioning variables are said to be weakly exogenous if they do not display error-correcting behav- ior. This topic has been discussed at length by Johansen (1992) and Boswijk (1995) among others.1 This literature has developed variable-addition tests to verify the validity of the weak exogeneity asumptions underlying the conditional approach; see Urbain (1995) and Boswijk and Urbain (1997) for further details and empirical applications. To date, the focus of this literature has been con…ned to the time-series context. But application of the VAR-ECM approach to panel settings has become increasingly popular (see, e.g., Breitung and Pesaran, 2008). This is partly due to the power de…ciencies of pure time-series approaches to cointegration (especially with relatively short time samples, as is commonly the case), which under appropriate conditions can be mitigated by combining time-series information with cross-sectional information. It is also motivated by the increasing availability of both micro and macroeconomic panel data sets. Modelling panel data within the VAR-ECM framework is arguably more demanding than in the case of pure time series data. For instance, the researcher must elicit certain assumptions regarding parameter heterogeneity and interactions among units in the panel. The laxer the assumptions, the larger the number of parameters to be estimated. Conditional analysis can help reduce that number drastically, and this makes it very appealing for the practitioner in the panel setting. 1 See also Dolado (1992), Ericsson (1995), and Hendry (1995). 1 Validity of the inference obtained from the conditional ECM in a panel context depends on weak exogeneity requirements similar to those that apply in the time-series case, and this naturally poses the need for weak exogeneity testing in panel implementations of the conditional approach. However, although the tests from the time-series literature cited above can shed light on the assumption of weak exogeneity for each cross-sectional unit in the panel, it is not obvious how they should be combined in order to jointly test weak exogeneity for the panel as a whole. Possible options range from testing whether weak exogeneity holds on average across all units,2 to simply joining all individual-unit tests into a panel-wide Wald test.3 In this paper we present a new panel weak exogeneity test. Like most of the time-series literature, we focus on the case in which the parameters of interest are those of the cointegrating vector(s). Weak exogeneity then amounts to the requirement that the cointegrating vector(s) not enter the marginal model. Following Westerlund and Hess (2011), we propose a test of this hypothesis based on the maximum of the individual Wald test statistics over all the cross-sectional units. We show that this maximal test has a limiting Gumbel distribution as both T (the time series sample size) and N (the number of units) go to in…nity. We also discuss how parameter heterogeneity and cross- sectional dependence of the type considered by Bai and Kao (2006) can be accommodated within our testing strategy. Simulation experiments investigate the size and power properties of the test in …nite samples. Overall, our simulation results suggest that the test performs well under sample sizes commonly encountered in applied research, although when T and N are roughly similar it exhibits a small size distortion that disappears with larger T .4 Also, the test proposed here has power against alternatives under which the test of “average weak exogeneity”has zero power. Finally, we illustrate the usefulness of the approach in practice by testing weak exogeneity of disposable income and wealth in an aggregate consumption function. In a related contribution, Trapani (2014) develops a Hausman-inspired test for the null hypothesis of endogeneity (i.e. existence of non-zero long-run correlation between the regressors and the error 2 See, e.g., Gemmell, Kneller and Sanz (2012) and Acosta-Hormaechea and Yoo (2012) for recent examples of this approach. 3 Canning and Pedroni (2008) examine alternative panel testing strategies in a related setting, in which the concern long-run causality’rather than the validity of conditional inference. is what they call ’ 4 The conventional joint Wald test can be expected to exhibit much worse behavior in terms of both power and size for the types of samples considered here. For example, in a related panel setting Phillips and Sul (2003) …nd the Wald test for homogeneity highly unreliable owing to its unacceptable size distortions, which remain severe even large N’ setting, the Wald test is not consistent because the number of with fairly large values of T. Further, in a ‘ restrictions to be tested increases with the sample size. However, it is also true that when the time dimension of the panel greatly exceeds its cross-sectional dimension, the testing strategy based on the joint Wald test could allow better accommodating ‡exible forms of cross-unit dependence, for example through SURE estimation of the parameters of the conditional model allowing for an unrestricted covariance matrix. 2 term). The setting is somewhat di¤erent, however. His testing approach is based on the reduced-form panel cointegrating equation, instead of the full VAR-ECM representation we use; moreover, his main objective is to help choose between estimation techniques rather than to test the appropriateness of the econometric speci…cation. The rest of the paper is organized as follows. Section 2 lays out the analytical framework and presents the weak exogeneity test. Section 3 reports the simulation experiments. Section 4 presents an empirical application of the test. Finally, section 5 o¤ers some concluding observations. 2 The Model Consider the m 1 vector Xit , which corresponds to the observation for unit i (i = 1; :::; N ) in period t (t = 1; :::; T ), and de…ne the panel VAR(pi ) model pi X Xit = ij Xit j + it (1) j =1 where it are independent identically distributed it Nm (0; i ). Note that this assumption rules out dependence across units in the panel, but below we shall relax this restriction and discuss how cross-sectional dependence can be accommodated in our testing procedure. The model admits an error-correction representation (see Engle and Granger, 1987):5 pi 1 X Xit = Ai Xit 1+ Aij Xit j + it (2) j =1 where we assume cointegration by imposing rank (Ai ) = ri with 0 < ri < m 8i. We can therefore 0 decompose Ai as i i, where i and i are m r matrices of full column rank for all i. While i contains the adjustment parameters, the columns of i represent the cointegrating vectors. In what follows we assume that the latter are the parametes of interest. Note that all parameters are allowed to be individual-speci…c. Therefore, our testing procedure can handle situations in which both short-run dynamics and long-run relationships di¤er across the units in the panel. However, we do need to assume that the number of cointegration relations is the same for all units, i.e., ri = r 8i with 0 < r < m; otherwise, the null hypothesis of weak exogeneity would di¤er across units hampering the construction of a panel-wide test, our main objective in this paper.6,7 5 To avoid notational clutter, the model omits unit-speci…c constants and time-trends. 6 Larsson and Lyhagen (2000) discuss in detail how to test for a common cointegrating rank across units in the panel. 7 Note that the formulation in (2) rules out cointegration across units in the panel, and excludes also the possibility that deviations from the long-run equilibrium in a given unit could impact the behavior of other units. 3 2.1 Conditional Analysis and Weak Exogeneity Based on economic theory, or just to reduce the dimensionality of the system for computational tractability, a researcher might be interested in modelling the equations for a subset of variables in the above formulation, taking the rest as given. For this purpuse, we can partition the vector Xit 0 0 0 as (yit ; zit ) , where yit represents a g -vector containing the variables of interest and the k -vector zit contains the set of conditioning variables (m = g + k ). Against this background, the model in (2) can be decomposed into two components. First, the conditional model for yit given zit : pi 1 X 0 yit = 0i zit + i;yz i Xit 1 + Aij;yz Xit j + it;yz (3) j =1 1 where i, Aij , and i are partitioned so that 0i = i;yz i;zz , i;yz = i;y 0i i;z , Aij;yz = Aij;y 0i Aij;z for j = 1; :::; pi 1, and it;yz = it;y 0i it;z . And, second, the marginal model of zit ; which consists of the last k equations of (2): pi 1 X 0 zit = i;z i Xit 1 + Aij;z Xit j + it;z (4) j =1 Since the parameters of the marginal and conditional models are interrelated, analysis of the full system in (2) is generally required for drawing e¢ cient inference about i. However, when the vector zit is weakly exogenous for i in the sense of Engle et al. (1983), analysis of the conditional system is e¢ cient and equivalent to full-model analysis. Formally, the conditioning variables zit are said to be weakly exogenous for the parameters of interest if, (i) the parameters in the conditional model and the parameters in the marginal model are variation free (i.e. they are not subject to any joint restrictions), and (ii) the parameters of interest are functions of the parameters in the conditional model only. If the parameters of interest are those corresponding to the cointegrating vectors i, a necessary and su¢ cient condition for zit to be weakly exogenous is that i;z = 0 (see Johansen, 1992 Theorem 1). In this case, e¢ cient inference on i can be conducted by analyzing (3) alone. As suggested by Johansen (1992), if g r one can easily test weak exogeneity using variable addition tests. In particular, we can estimate the cointegrating vectors from the conditional system (3), and insert the superconsistent estimates ^ into the marginal equations (4). Hence, for …xed i i = ^i we can test the hypothesis that the coe¢ cients i;z are jointly zero by, for instance, a Wald test. This testing procedure is straightforward in the time-series framework discussed in Johansen (1992). However, in the panel setting considered here, the choice of testing procedure is less obvious. 4 One might be tempted to construct a Wald test for the null that i;z = 0 for all i simultaneously, but such test will be poorly behaved when N and T are roughly similar. Alternatively, one might consider a test of the null that the average of i;z across i equals zero, but this would be informative on average’among the about a somewhat di¤erent hypothesis, namely that weak exogeneity holds ’ cross-sectional units. There are situations in which such hypothesis would hold even though unit- speci…c weak exogeneity does not; for instance, if the i;z are distributed symmetrically around zero, so that some units exhibit i;z > 0 and others i;z < 0. In such setting, a test of weak exogeneity based on whether the average i;z equals zero would have zero power. We next suggest an alternative test based on the maximum of the individual Wald test statistics. 2.2 A Panel Test of Weak Exogeneity Stacking the observations for all units in the panel, we can write the marginal system in (4) as follows: zi = [Ik ^ (5) i( 1) ] i;z + [Ik Xi( j ) ]Ai;z + i;z where zi = ( zi11 ; :::; ziT1 2 ; zi21 ; :::; ziT k 0 ; :::; zik1 ; :::; ziT ) is a kT 1 vector, ^i( 1) is a T 1 0 vector given by (^i0 ; :::; ^iT 1 )0 with ^it 1 = ^ i Xit 1 . Finally, Xi( j ) is the T m(pi 1) matrix of data corresponding to all lags of Xit included in the model with Ai;z = (A0i1;z ; A0i2;z ; :::; A0ip 1;z ) 0 the corresponding km(pi 1) 1 vector of coe¢ cients. We can estimate the coe¢ cients in (5) by ordinary least squares (OLS) and obtain:8 ! ! 1 ! ^ i;z Ik ^0 ^ Ik ^0 Xi( (Ik ^0 i( 1) i( 1) i( 1) j) i( 1) ) zi = (6) ^i;z A Ik Xi0( j ) ^i( 1) Ik Xi0( j ) Xi( j) (Ik Xi0( j ) ) zi with covariance matrix estimated by: ! ! 1 ^ i;z Ik ^0 ^i( Ik ^0 Xi( j) ^ V = ^2 i( 1) 1) i( 1) (7) i ^i;z A Ik 0 ^ Xi( j ) i( 1) Ik Xi0( j ) Xi( j) where ^ 2 i is the least squares variance estimator. For a given unit in the panel we can consider the following Wald test of the null of weak exogeneity: h i 1 ^ i = ^ 0i;z V W ^ (^ i;z ) ^ i;z (8) 2 which is asymptotically distributed as a k as T ! 1. 8 In this particular case, these estimates coincide with equation-by-equation estimates because right-hand-side variables are the same in all equations and there are no cross-equation restrictions. 5 In the panel setting, we formulate the null hypothesis to be tested as H0 : i;z =0 for all i (9) against the alternative Ha : i;z 6= 0 for at least one i (10) In parallel with the poolability test of Westerlund and Hess (2011),9 consider using the maximum of the individual Wald test statistics (8) for testing H0 versus Ha : ^i ^ max = max W W (11) 1 i N To derive its asymptotic distribution, we de…ne the normalized test statistic: ^ Zmax = 1 W W ^ max dN (12) cN 1 1 where cN = 2 and dN = F 1 N with F (x) being the chi-squared distribution function with k degrees of freedom evaluated at x. Theorem 1 Under H0 , as T ! 1, N ! 1: ^ Zmax x) ! exp( e x ) P (W Proof See Appendix. Theorem 1 indicates that the WZmax test has a limiting Gumbel distribution as T ! 1, N ! 1. Because the asymptotic distribution of the test is obtained by …rst letting T ! 1 and then letting N ! 1, this implies that the test is appropriate in cases where N is moderate and T is large (see Phillips and Moon, 1999); this type of data con…guration can be expected for instance in multy- country macroeconomic data. In the Monte Carlo experiments we investigate the accuracy of the asymptotic approximation under di¤erent (N; T ) con…gurations. 2.3 Cross-Sectional Dependence The recent panel time-series literature has paid considerable atention to the possibility that the individual units in a panel dataset may be interdependent. We next discuss how cross-sectional dependence can be accommodated in our testing procedure. Speci…cally, we consider a setting along 9 The Westerlund and Hess (2011) test is based on the maximum across panel units of their individual Hausman test statistics for the null that their respective cointegrating vector parameters equal those of all the other cross seciton units. 6 the lines of Bai and Kao (2006), in which the long-run disequilibrium error it has a stationary common component (ft ) and a stationary idiosyncratic component ( it ): 0 0 i Xit = it = i ft + it (13) where ft is a q 1 vector of latent common factors, i is a q r matrix of factor loadings and it is a vector of idiosyncratic errors. Importantly, by assuming that the factors ft are stationary we rule out the presence of common stochastic trends (see, e.g., Bai et al., 2009), and thus the possibility of cointegration between the variables and the factors (see, e.g., Gengenbach et al., 2012) and, in particular, cointegration across units.10 From the practical perspective, whether stationarity of ft represents a restrictive assumption depends on the application at hand. In the international business cycle literature, for example, global technology shocks – widely assumed the primary force driving world business cycles – are typically viewed as persistent but stationary (see, e.g., Kehoe and Perri, 2002); the same applies to global demand (spending) shocks (e.g., Boileau et al., 2010). In other situations, it may be more di¢ cult to rule out common stochastic trends; for example, panel tests of purchasing power parity usually involve country-speci…c I(1) variables de…ned relative to a reference country, and this tends to introduce common stochastic trends in the analysis (e.g., Urbain and Westerlund, 2011). Our testing procedure can accommodate cross-sectional dependence of the assumed form with- out modi…cation. In practice, the test statistic can be computed as described above but replacing the original variables yit and zit with the residuals of a regression of these variables on their cross- sectional averages, i.e., the common correlated e¤ects (CCE) approach proposed in Pesaran (2006). Furthermore, it would be straightforward to allow also for stationary common factors in the condi- tioning variables (as noted in Remark 1.1 in Bai and Kao, 2006), for example by also employing in the variable-addition tests the CCE approach. A particular (and more restrictive) form of cross-sectional dependence worth mentioning arises from further imposing that i = 8i together with q = 1. In this case, one can remove the common e¤ect prior to the test just by cross-sectional de-meaning of the data. In the simulations below, we investigate the impact of these types of cross-sectional dependence on the …nite-sample behavior of the test. 10 Strictly speaking, if the common factors are I(1), the CCE approach in Pesaran (2006), whose use is suggested below in the text, would still remove the common components (Kapetianos et al., 2011) and allow performing the panel weak exogeneity test. However, such procedure would not take proper account of the possible presence of cross-unit cointegration. 7 3 Monte Carlo Evidence To evaluate the …nite-sample behavior of the proposed testing procedure, we build upon the simu- lation designs in Phillips and Hansen (1990), Phillips and Loretan (1991), and Boswijk (1995). For each unit in the panel (i = 1; :::; N ) we consider the case of two cointegrated variables under the following data generating process (DGP):11 yit = i zit + uy;it (14) zit = zit 1 + uz;it (15) uit = it + i it 1 where uit = (uy;it ; uz;it )0 and it =( 0 1;it ; 2;it ) . Moreover, it N (0; ) with ! ! 1 11 12 = and i = (16) 1 i 22 The i parameter is the error correction coe¢ cient in the marginal model for zit and thus deter- mines whether or not zit is weakly exogenous for i. In our simulations, we consider a fraction of units not satisfying the weak exogeneity of zit ; for these units, i = 0:8; as in Phillips and Loretan (1991). The remaining (1 )N units do satisfy the weak exogeneity condition (i.e. i = 0:0). With respect the remaining parameter values to be chosen, in our baseline design we also follow Phillips and Loretan (1991) and …x = 0:5, 11 = 0:3, 12 = 0:4, and 22 = 0:6 for i = 1; :::; N . The cointegrating vector is also common to all units in the baseline design with i = 2 8i. Given the simulated samples for each of the N units over T periods, our simulation exercises consist of 2 steps. First, for …xed i we compute the individual Wald statistics of the form (8) for the null H0 : i = 0. Second, from the N individual Wald statistics thus obtained we compute ^ Zmax statistic and evaluate its size and power computing the rejection frequencies the normalized W under di¤erent values of , the fraction of units in which zit is not weakly exogenous. This …rst exercise assumes a known i, common across units, i.e., i = 8i. In real applications, however, needs to be estimated. Hence we perform a second set of simulations involving an additional step estimating at the beginning of each replication. In particular, we consider two di¤erent estimation strategies, one based on pooling the data for all the units, and another based 11 In a recent paper, Gengenbach et al. (2013) derive a Granger-type representation theorem given the triangular representation of a panel cointegration model as in (14)-(15) but including common factors in both the long-run relationship and the conditioning variables. In particular, they show that such a panel cointegration model also admits an error-correction representation. It can thus be shown that a marginal model for zit as in equation (4) can be derived from the error-correction representation with common factors entering in both the long-run relationship and the short-run dynamics (see Remark 2 in Gengenbach et al., 2013). 8 on unit-by-unit estimation. Under homogeneity of the cointegrating vector, the pooled approach is more e¢ cient than the unit-by-unit approach. However, we also consider the unit-by-unit strategy because the speed of adjustment in (3) might vary among cross-sectional units even in the case of homogeneous .12 For both the pooled and the unit-by-unit approaches we follow Boswijk (1995) and employ OLS including three lags of zit and two lags of yit as suggested by Phillips and Hansen (1990). Holding the resulting ^ …xed, we repeat the two steps above. Panel A of Table 1 presents the simulation results based on 50; 000 replications. For sample sizes commonly found in practice, we see that the test is appropriately sized given the nominal 5% size considered. However, in cases when N is not substantially smaller than T , the test appears slightly oversized, which is not surprising in view of the sequential limit theory underlying its asymptotic distribution. Power rises steadily with both N and T , as well as with , the number of units violating weak exogeneity in the simulated samples. Indeed, as sample sizes grow, the power of the test becomes considerable even when only 20% of units violate weak exogeneity. In Panel B of Table 1, we explore the performance of the test when using the estimated ^ based on the pooled approach instead of the true .13 While power is uniformly lower than when using , the di¤erence is almost negligible.14 Next, in Panel C of Table 1 we present the results based on the unit-by-unit estimates of the common cointegrating vector. Power is always slightly smaller than in the pooled approach reported in Panel B, as expected from the superior e¢ ciency of the pooled estimator under homogeneity. However, the di¤erences are in all cases fairly modest. In turn, in Table 2 we evaluate the performance of the test in situations in which cross-sectional units are not independent from each other. For this purpose, we consider the DGP in (14) but allowing cross-section dependence through the error term uit . In particular, we allow for cross-section dependence in the long-run relationship by re-specifying equation (14) as 0 yit = i zit + i ft + uy;it (17) 12 An alternative approach could be the pooled mean group — PMG— estimator discussed in Pesaran et al. (1999). The PMG estimator can accommodate homogeneous long-run coe¢ cients together with heterogeneous speed of ad- justment and short-run coe¢ cients. 13 In practical terms, this means that the subsequent estimation of i and the construction of the Wald statistics (8) employ a generated regressor. In the time-series context (i.e., N = 1), this should be of little consequence for the behavior of the variable-addition test, owing to the super-consistency of ^i with respect to T ; see, e.g., Boswijk and Urbain (1997). In the panel context, however, the situation is somewhat less clearcut, because ^ based on the pooled p approach converges at just the standard rate N with respect to the cross-sectional sample size. This might raise concerns about the performance of our testing procedure unless T is very large. Nevertheless, the results shown in the table reveal that performance is little a¤ected by the use of the pooled estimator of the long-run relationship. 14 Further, closer comparison of panels A and B of Table 1 reveals that the decline in power from using ^ instead of is inversely related to T =N , as should be expected in light of the theoretical argument mentioned in the previous footnote. 9 Table 1: Size and Power under Cross-section Homogeneity Parameters N T = 0:0 = 0:2 = 0:4 = 0:6 = 0:8 = 1:0 Panel A: Using i i = 2:0 8i 10 50 0.073 0.323 0.504 0.636 0.736 0.806 = 0:5 10 100 0.059 0.639 0.862 0.947 0.981 0.993 11 = 0:3 10 150 0.056 0.863 0.980 0.997 1.000 1.000 12 = 0:4 20 50 0.090 0.431 0.639 0.775 0.858 0.912 22 = 0:6 20 100 0.069 0.796 0.955 0.990 0.998 0.999 20 150 0.065 0.961 0.998 1.000 1.000 1.000 Panel B: Using pooled ^ 10 50 0.074 0.321 0.501 0.632 0.731 0.801 10 100 0.059 0.636 0.860 0.946 0.980 0.992 10 150 0.056 0.861 0.980 0.997 1.000 1.000 20 50 0.091 0.429 0.636 0.771 0.855 0.909 20 100 0.069 0.795 0.955 0.989 0.998 0.999 20 150 0.065 0.960 0.998 1.000 1.000 1.000 Panel C: Using unit-by-unit ^i 10 50 0.074 0.286 0.445 0.572 0.671 0.744 10 100 0.058 0.612 0.838 0.934 0.974 0.989 10 150 0.055 0.850 0.977 0.996 0.999 1.000 20 50 0.090 0.380 0.573 0.709 0.800 0.861 20 100 0.068 0.766 0.944 0.985 0.997 0.999 20 150 0.065 0.954 0.998 1.000 1.000 1.000 refers to the fraction of units for which weak exogeneity does not hold. For the case Notes: 50; 000 replications; = 0:0 the numbers represent the size of the W^ Zmax test. For the remaining cases with > 0:0, the numbers in the table refer to power. The nominal level test is …xed at 5%. We …x i = 0:8 in N units and i = 0:0 in the remaining N (1 ) units. In Panel A we consider the simplest case of a common time e¤ect by setting ft N (0; 1) and i = 1 8i. As explained above, in this case we deal with dependence across units in our panel by employing in the testing procedure the cross-sectionally demeaned variables. In Panel B we consider a more general form of cross-section dependence as discussed in Bai and Kao (2006). Speci…cally, the common factors are allowed to have an heterogeneous impact on the individual units by setting ft N (0; 1) and i N (1; 1) as in Westerlund and Hess (2011). The Monte Carlo simulations reported in Table 2 show that cross-sectional dependence of these 10 Table 2: Size and Power under Cross-section Dependence Parameters N T = 0:0 = 0:2 = 0:4 = 0:6 = 0:8 = 1:0 Panel A: Cross-Section Dependence with a Common Time E¤ect i = 1:0 8i 10 50 0.076 0.240 0.400 0.557 0.689 0.795 i = 2:0 8i 10 100 0.058 0.471 0.749 0.898 0.965 0.991 = 0:5 10 150 0.055 0.701 0.931 0.988 0.999 1.000 11 = 0:3 20 50 0.093 0.372 0.587 0.741 0.843 0.908 12 = 0:4 20 100 0.068 0.716 0.924 0.983 0.997 0.999 22 = 0:6 20 150 0.065 0.921 0.996 1.000 1.000 1.000 Panel B: Cross-Section Dependence as in Bai and Kao (2006) i N (1; 1) 10 50 0.077 0.240 0.403 0.557 0.689 0.797 i = 2:0 8i 10 100 0.056 0.471 0.749 0.897 0.966 0.990 = 0:5 10 150 0.057 0.706 0.932 0.988 0.998 1.000 11 = 0:3 20 50 0.093 0.370 0.580 0.736 0.843 0.908 12 = 0:4 20 100 0.068 0.714 0.923 0.983 0.997 0.999 22 = 0:6 20 150 0.064 0.921 0.995 1.000 1.000 1.000 Notes: 50; 000 replications; refers to the fraction of units for which weak exogeneity does not hold. For the case = 0:0 the numbers represent the size of the W^ Zmax test. For the remaining cases with > 0:0, the numbers in the table refer to power. The nominal level test is …xed at 5%. We …x i = 0:8 in N units and i = 0:0 in the remaining N (1 ) units. forms has a fairly modest e¤ect on the size and power properties of our testing procedure. While power is generally smaller than in the case of cross-sectional independence, it is still considerable in all cases. In Table 3, we explore the e¤ects of introducing heterogeneity in the long-run relationship and considering an alternative type of serial correlation in the idiosyncratic errors. In particular, in Panel A of Table 3 we allow the long-run parameter i to vary across units. Overall, the size and power properties of the test for di¤erent values of are very similar to those found in Table 1 under homogeneity. In a majority of cases power is slightly smaller than in the homogeneous case, but the di¤erences are just in the third decimal. Therefore, we conclude that cross-section heterogeneity can be appropriately handled by our testing procedure. In Panel B of Table 3 we consider autoregressive idiosyncratic errors instead of moving average errors as in the simulations above. For this purpose, we de…ne: uit = i uit 1 + it (18) 11 Table 3: Size and Power under Cross-section Heterogeneity and AR(1) Errors Parameters N T = 0:0 = 0:2 = 0:4 = 0:6 = 0:8 = 1:0 Panel A: Cross-Section Heterogeneity i U (1; 3) 10 50 0.074 0.324 0.506 0.638 0.734 0.803 = 0:5 10 100 0.058 0.641 0.866 0.948 0.979 0.993 11 = 0:3 10 150 0.057 0.862 0.981 0.998 0.999 1.000 12 = 0:4 20 50 0.092 0.431 0.638 0.773 0.858 0.910 22 = 0:6 20 100 0.067 0.795 0.956 0.990 0.998 1.000 20 150 0.066 0.962 0.998 1.000 1.000 1.000 Panel B: AR(1) Errors i = 2:0 8i 10 50 0.072 0.346 0.545 0.683 0.781 0.841 = 0:5 10 100 0.047 0.695 0.900 0.968 0.990 0.996 11 = 0:3 10 150 0.041 0.901 0.989 0.999 1.000 1.000 12 = 0:2 20 50 0.087 0.466 0.686 0.817 0.893 0.937 22 = 0:6 20 100 0.056 0.843 0.975 0.996 0.999 1.000 20 150 0.048 0.978 0.999 1.000 1.000 1.000 Notes: 50; 000 replications; refers to the fraction of units for which weak exogeneity does not hold. For the case ^ test. For the remaining cases with > 0:0, the numbers in the = 0:0 the numbers represent the size of the W Z table refer to power. The nominal level test is …xed at 5%. We …x i = 0:8 in N units and i = 0:0 in the remaining N (1 ) units. where it is de…ned above and ! 11 12 i = (19) 21;i 22 In order to ensure comparability with the DGP based on moving average errors and guarantee stationarity of the resulting autoregressive errors, we set 11 = 0:3, 12 = 0:2, 22 = 0:6; moreover, we choose a value of 21;i that produces i;z = i = 0:8 as in the simulations above. For this purpose, we make use of the results in Cappuccio and Lubian (1996) and Gengenbach et al. (2013) who derive the conditional and marginal ECM representation for a cointegration model as in (14)- (15). In particular, when the idiosyncratic errors are AR(1) it can be shown that the error correction coe¢ cient in the marginal model has the form: 21;i i;z = (20) 1 11 22 21;i 12 + 11 22 Given the values above for 11 , 12 and 22 , we set 21;i so that i;z = 0:8 = i. By doing this we ensure that the error correction behavior of the conditioning variables is the same under both the 12 autoregressive and the moving average errors for the N units not satisfying the weak exogeneity condition. For the remaining (1 )N units we set 21;i = 0. When considering autoregressive errors in Panel B of Table 3, power is slightly larger and size slightly smaller than in the case of moving average errors in Panel A of Table 1. We thus conclude that this alternative form of serial correlation in the errors can be appropriately handled by our testing procedure. Finally, following Westerlund and Hess (2011), we also investigated the …nite sample properties of an alternative panel weak exogeneity test based on the normalized sum (instead of the maximum) of the individual Wald test statistics: 1 X ^ N p ^ WZ = p Wi N (21) N i=1 which is asymptotically distributed as a N (0; 2) as T ! 1, N ! 1 (see Westerlund and Hess, 2011). Simulation results15 indicate that, in …nite samples, this test has similar power to that of the W ^ Zmax test appears preferable. ^ Zmax test, but exhibits larger size distortions; hence the W 4 Empirical Illustration In this section we apply the panel test of weak exogeneity to the estimation of a consumption function along the lines of Ludwig and Sløk (2004). They analyze the impact of changes in household wealth on consumption distinguishing between two di¤erent components of wealth, namely, housing wealth and stock market wealth, using quarterly data for a sample of 16 OECD countries over the period 1960-2000. Our purpose here is to illustrate our testing procedure in practice; we thus follow the baseline speci…cation of Ludwig and Sløk (2004) as closely as possible.16 In analogy with equation (3) above, Ludwig and Sløk (2004) consider the following conditional error correction model:17 d sw hw d sw hw cit = i cit 1 0 1 yit 2 wit 3 wit + a1 yit + a2 wit + a3 wit + it (22) d where cit refers to log private per capita consumption in country i at quarter t, yit is log per capita sw hw disposable income, wit is the log stock market wealth proxied by a stock market price index, and wit refers to log housing wealth proxied by a house price index (see Ludwig and Sløk (2004) for full details 15 These results are not reported to save space, but they are available upon request. 16 Hence we do not test other aspects of their speci…cation, such as the fact that cross-sectional dependence is not taken into account. 17 To avoid notational clutter we present here an ARDL(1,1,1) version of the model. However, in practice we allow for country-speci…c dynamics, with lag length determined by the Schwarz criterion. 13 d sw hw 0 on the data used). In the notation of Section 2, we can de…ne the vector Xit = (cit ; yit ; wit ; wit ). Analysis of the conditional model in (22) provides e¢ cient inference as long as the conditioning vari- ables y d , wsw , and whw are weakly exogenous for the long-run parameter vector =( 0; 1; 2; 3 ). Otherwise, analysis of the full model would be required. We next use our panel test to assess the validity of the weak exogeneity assumption implicitly made by Ludwig and Sløk (2004) when estimating the conditional model in (22). Table 4 presents the country-speci…c Wald tests of weak exogeneity, together with the panel tests for each of the three conditioning variables –separately as well as jointly. d According to the individual tests in Panel A of Table 4, disposable income (yit ) and housing hw wealth (wit ) appear to be weakly exogenous with respect to the long-run relationship in equation (22) in each of the 16 countries in the sample. The panel tests reported in Panel B strongly support the assumption of weak exogeneity of disposable income and housing wealth with respect to the long-run consumption function, thus con…rming the verdict of the individual tests. Regarding stock sw market wealth (wit ), the individual tests in column 2 reject the null of weak exogeneity at the 5% level in 3 out of 16 countries. Overall, however, the panel test at the bottom of the table does not provide much evidence against the weak exogeneity of stock market wealth. Based on these …ndings, we conclude that none of the three conditioning variables –disposable income, stock market wealth and housing wealth – reacts signi…cantly to deviations of aggregate consumption from its long-run equilibrium trajectory. The last column of Panel A in Table 4 reports joint tests of weak exogeneity of all three regressors for each individual country. The null cannot be rejected at the 5% level for any country, and only in 2 countries (the Netherlands and Sweden) out of 16 is it is rejected at the 10% level. In turn, the panel joint test at the bottom of the table, which considers all countries simultaneously, yields a p-value of 0:746, so that weak exogeneity of all three regressors for the panel as a whole cannot be rejected. All in all, these results indicate the the conditional analysis based on estimation of equation (22) alone, as done by Ludwig and Sløk (2004), is equivalent to full-model analysis in this context. 5 Concluding Remarks In cointegrated panel settings, it is common for reasons of empirical tractability to model only a subset of variables conditional on another subset whose marginal processes are not modelled. This approach is as e¢ cient as the full-model approach only when the conditioning variables are weakly exogenous for the parameters of interest. The time series literature has proposed variable-addition tests of the null of weak exogeneity for the case of a single cross-section unit (e.g., Johansen 1992, Urbain 1995). In this paper, we extend 14 Table 4: Weak Exogeneity Tests Panel A: Country-by-country tests d sw hw Country yit wit wit Joint Australia 0.214 (0.644) 0.666 (0.414) 0.000 (1.000) 0.880 (0.830) Belgium 0.019 (0.890) 1.242 (0.265) 0.000 (1.000) 1.261 (0.738) Canada 0.157 (0.692) 3.005 (0.083) 0.012 (0.913) 3.174 (0.366) Denmark 0.087 (0.768) 1.452 (0.228) 0.001 (0.975) 1.541 (0.673) Finland 0.082 (0.775) 0.008 (0.929) 0.014 (0.906) 0.104 (0.991) France 0.050 (0.823) 1.799 (0.180) 0.001 (0.975) 1.850 (0.604) Germany 0.020 (0.888) 2.668 (0.102) 0.023 (0.879) 2.710 (0.439) Ireland 0.019 (0.890) 1.126 (0.289) 0.001 (0.975) 1.146 (0.766) Italy 0.045 (0.832) 0.837 (0.360) 0.088 (0.767) 0.969 (0.809) Japan 0.226 (0.635) 0.400 (0.527) 0.004 (0.950) 0.630 (0.890) Netherlands 0.005 (0.944) 6.572 (0.010) 0.011 (0.916) 6.588 (0.086) Norway 0.041 (0.840) 5.020 (0.025) 0.004 (0.950) 5.065 (0.167) Spain 0.061 (0.805) 0.056 (0.813) 0.011 (0.916) 0.128 (0.988) Sweden 0.082 (0.775) 6.602 (0.010) 0.000 (1.000) 6.684 (0.083) United Kingdom 0.221 (0.638) 1.170 (0.279) 0.077 (0.781) 1.468 (0.690) United States 0.054 (0.816) 5.395 (0.020) 0.007 (0.933) 5.456 (0.141) Panel B: Panel tests d sw hw yit wit wit Joint ^ Zmax statistic W -1.622 1.566 -1.691 -0.316 p-value 0.994 0.189 0.996 0.746 d is log per capita disposable income, w sw is the log stock market wealth, and w hw refers to log housing Notes: yit it it wealth. Sample period is 1960:Q1-2000:Q4. In panel A, p-values are in parentheses. All p-values in the Table refer to the null hypothesis of weak exogeneity of the corresponding regressor(s). this approach to panel settings. In particular, for the case in which the parameters of interest are those of the cointegrating vector(s), we propose a panel test based on the maximum of the individual Wald statistics for testing weak exogeneity, and we obtain the asymptotic Gumbel distribution of the resulting test statistic as both T — the number of time periods— and N — the number of cross- sectional units— tend to in…nity. Monte Carlo simulations indicate that the proposed panel test performs quite well in sample 15 sizes commonly encountered in applied research. Moreover, the simulations also suggest that the test is robust to cross-sectional heterogeneity of the long-run parameters, as well as cross-sectional dependence of the type discussed in Bai and Kao (2006). Thus, the test should have wide applicability. Finally, we illustrate its use by assessing the weak exogeneity of disposable income and wealth for the estimation of an aggregate consumption function. 16 A Appendix A.1 Proofs ^ Zmax test, we consider Proof of Theorem 1 In order to derive the asymptotic distribution of the W a sequential limit in which T ! 1 followed by N ! 1 as suggested by Phillips and Moon (1999). This sequential approach is based on …xing one of the indexes, say N , and allowing the other (i.e. T ) to pass to in…nity. By then letting N pass to in…nity, a sequential limit is obtained. We …rst use time series limit theory for any given unit i in the panel, i.e., we pass T to in…nity for …xed N . In this case, it is a well-known result that: ^i ! W 2 as T ! 1 (23) k for each unit in the panel i = 1; :::; N . Given this result, having passed T to in…nity in equation (11) we obtain: Wmax = max Wi (24) 1 i N 2 where Wi k for all i. We next allow N ! 1 and apply limit theory to the statistic in (24). In particular we make use of a well-known result in extreme value theory which provides us with a set of limit laws for maxima. Let MN be the maximum of a sequence of N iid chi-squared variables, Mn = max(X1 ; :::; XN ). Then, the Fisher-Tippett theorem states that: cN1 (MN dN ) ! (25) as N ! 1. Where, for x 2 R, (x) = exp( e x ) refers to the Gumbel distribution function, and the norming constants cN > 0 and dN 2 R are the mean excess function and the empirical version of the (1 n 1 )-quantile of the underlying distribution function respectively (see Embrechts et al., 1997). Combining this result with the large T , …xed N asymptotic test in (24) we have that: cN1 (Wmax dN ) ! as N ! 1 (26) where, as discussed in Westerlund and Hess (2011), the norming constants are given by dN = 1 1 1 1 F (1 N ) and cN = F (1 Ne ) dN , or cN = 2 as N ! 1. Note that F (x) here refers to the chi-squared distribution function with k degrees of freedom. Combining the results in (23) and (26): ^ Zmax ! W as T ! 1; N ! 1 (27) ^ Zmax has a limiting Gumbel distribution as T ! 1,N ! 1 which indicates that W 17 References [1] Acosta-Ormaechea, S. and J. Yoo (2012) "Tax Composition and Growth: A Broad Cross- Country Perspective", IMF Working Paper 12/257. [2] Bai, J. and C. Kao (2006) "On the estimation inference of a panel cointegration model with cross- sectional dependence", In: Baltagi, Badi (Ed.), Contributions to Economic Analysis. Elsevier, pp. 3-30. [3] Bai, J., C. Kao and S. Ng (2009) "Panel cointegration with global stochastic trends", Journal of Econometrics, vol. 149, pp. 82–99. [4] Boileau, M., M. Normandin and B. Fosso (2010) "Global versus country-speci…c shocks and international business cycles", Journal of Macroeconomics 32, 1-16. [5] Boswijk, P. (1995) "E¢ cient inference on cointegration parameters in structural error correction models," Journal of Econometrics, vol. 69, pp. 133-158. [6] Boswijk, P. and J. Urbain (1997) "Lagrange-multiplier tests for weak exogeneity: a synthesis", Econometric Reviews vol. 16, pp. 21-38. [7] Breitung, J. and H. Pesaran (2008) "Unit Roots and Cointegration in Panels", in L. Matyas and P. Sevestre (eds), The Econometrics of Panel Data, Ch. 9, pp. 279-322. [8] Canning, D. and P. Pedroni (2008) "Infrastructure, long-run economic growth and causality tests for cointegrated panels", The Manchester School vol. 76, pp. 504-527. [9] Cappuccio, N. and D. Lubian (1996) "Triangular representation and error correction mechanism in cointegrated systems", Oxford Bulletin of Economics and Statistics vol. 58, pp. 409–415. [10] Dolado, J. (1992) "A Note on Weak Exogeneity in VAR Cointegrated Systems," Economic Letters, vol. 38, pp. 139-43. [11] Embrechts, P., C. Klüppelberg, and T. Mikosch (1997) "Modelling Extremal Events for Insur- ance and Finance," Springer: Berlin. [12] Engle, R. and C. Granger (1987) "Cointegration and error correction: Representation, estima- tion, and testing," Econometrica, vol. 55, pp. 251-276. [13] Engle, R., D. Hendry, and J. Richard (1983) "Exogeneity," Econometrica, vol. 51, pp. 277-304. 18 [14] Ericsson, N. (1995) "Conditional and structural error correction models," Journal of Economet- rics, vol. 69, pp. 159-171. [15] Gemmell, N., R. Kneller and I. Sanz (2012) "Does the composition of government expenditure matter for economic growth?", manuscript, University of Nottingham. [16] Gengenbach, C., J.-P. Urbain and J. Westerlund (2012) "Panel error correction testing with global stochastic trends", University Maastricht, Mimeo. [17] Gengenbach, C., J.-P. Urbain and J. Westerlund (2013) "Alternative representations for cointe- grated panels with global stochastic trends", Economics Letters, vol. 118, pp. 485-488. [18] Hendry, D. (1995) "On the interactions of unit roots and exogeneity," Econometric Reviews, vol. 14, pp. 383-419. [19] Johansen, S. (1988) "Statistical analysis of cointegration vectors," Journal of Economic Dynam- ics and Control, vol. 12, pp. 231-254. [20] Johansen, S. (1992) "Cointegration in partial systems and the e¢ ciency of single-equation analy- sis," Journal of Econometrics, vol. 52, pp. 389-402. [21] Kapetanios, G., H. Pesaran, and T. Yamagata (2011) "Panels with non-stationary multifactor error structures," Journal of Econometrics, vol. 160, pp. 326-348. [22] Kehoe, P. and F. Perri (2002) "International business cycles with endogenous incomplete mar- kets," Econometrica 70, 907-928. [23] Larsson, R. and J. Lyhagen (2000) "Testing for common cointegrating rank in dynamic panels," Working Paper Series in Economics and Finance 378, Stockholm School of Economics. [24] Ludwig, A. and T. Sløk (2004) "The Relationship between Stock Prices, House Prices and Consumption in OECD Countries," The B.E. Journal of Macroeconomics (Topics), vol. 4(1), pp. 1-26. [25] Pesaran, H. (2006) "Estimation and inference in large heterogeneous panels with a multifactor error structure", Econometrica, vol. 74, pp. 967-1012. [26] Pesaran, H., Y. Shin, and R. Smith (1999) "Pooled mean group estimation of dynamic hetero- geneous panels," Journal of the American Statistical Association, vol. 94, pp. 621-634. [27] Phillips, P. and B. Hansen (1990) "Statistical inference in instrumental variable regression with I(l) processes," Review of Economic Studies, vol. 57, pp. 99-125. 19 [28] Phillips, P. and D. Sul (2003) "Dynamic panel estimation and homogeneity testing under cross section dependence," Econometrics Journal, vol. 6, pp. 217-259. [29] Phillips, P. and M. Loretan (1991) "Estimating long-run economic equilibria," Review of Eco- nomic Studies, vol. 58, pp. 407-436. [30] Phillips, P. and H. Moon (1999) "Linear Regression Limit Theory for Nonstationary Panel Data," Econometrica, vol. 67, pp. 1057-1112. [31] Trapani L. (2014) "Testing for Exogeneity in Cointegrated Panels," Oxford Bulletin of Economics and Statistics, forthcoming. [32] Westerlund, J. and W. Hess (2011) "A new poolability test for cointegrated panels," Journal of Applied Econometrics, vol. 26, pp. 56-88. [33] Urbain, J. (1995) "Partial versus full system modelling of cointegrated systems: an empirical illustration," Journal of Econometrics, vol. 69, pp. 177-210. [34] Urbain, J. and J. Westerlund (2011) "Least Squares Asymptotics in Spurious and Cointegrated Panel Regressions with Common and Idiosyncratic Stochastic Trends," Oxford Bulletin of Eco- nomics and Statistics, vol. 73, pp. 119-139. 20