Abstract
This paper proposes two bootstrap-based tests that can be used to infer whether the individual slopes in a panel regression model are homogenous. The first test is suitable when wanting to infer the null of homogeneity versus the general alternative, while the second is suitable when wanting to infer the units of the panel that can be pooled. Both approaches are shown to be asymptotically valid, a property that is verified in small samples using Monte Carlo simulation.
Similar content being viewed by others
Notes
As pointed out by MacKinnon (2007), the advantage of bootstrapping \((y_{i,t}, x_{i,t}^\prime )\) rather than residuals often comes at the cost of poor small-sample performance, a finding that is supported by our preliminary Monte Carlo results.
An alternative bootstrap approach based on cross-sectional resampling has been suggested by Kapetanios (2008). Unlike the bootstrap approach used here, the cross-sectional resampling scheme provides asymptotically valid bootstrap procedures when \(N \rightarrow \infty \) but T remains fixed, but this is based on cross-sectional independence.
See Gonçalves (2011) for a set of assumptions that can be used in the case of arbitrary cross-sectional dependence and N asymptotics.
Kapetanios (2003) assumes the existence of a consistent pooled estimator of some common parameter \(\beta \), regardless of whether or not the units are poolable. Typically, this requires that a random coefficient assumption is satisfied, or alternatively, that the fraction of non-poolable units tends to zero as \(N \rightarrow \infty \).
We also considered a moving average model for \(f_{\varepsilon , t}\). The results were, however, very similar to the ones based the autoregressive model considered here and are therefore omitted.
The power results are not size corrected because such a correction is generally not available in practice. Hence, a test is useful for applied work only if it respects roughly the nominal significance level.
For a detailed discussion of these procedures, we refer to Lehmann and Romano (2005, Chapter 9).
References
Andrews DWK (1991) Heteroskedasticity and autocorrelation consistent covariance matrix estimation. Econometrica 59:817–858
Anselin L, Le Gallo J, Jayet H (2008) Spatial panel econometrics. In: Mátyás L, Sevestre P (eds) The econometrics of panel data. Springer, Berlin
Baltagi BH (2008) Econometric analysis of panel data. Wiley, Chichester
Baltagi BH, Bresson G, Pirotte A (2008) To pool or not to pool? In: Mátyás L, Sevestre P (eds) The econometrics of panel data. Springer, Berlin
Bun MJG (2004) Testing poolability in a system of dynamic regressions with nonspherical disturbances. Empir Econ 29:89–106
Chudik A, Pesaran MH, Tosetti E (2011) Weak and strong cross-section dependence and estimation of large panels. Econom J 14:C45–C90
Davidson J (1994) Stochastic limit theory. Oxford University Press, Oxford
Davidson R, MacKinnon JG (1999) The size distortion of bootstrap tests. Econom Theory 15:361–376
Fitzenberger B (1997) The moving blocks bootstrap and robust inference for linear least squares and quantile regressions. J Econom 82:235–287
Freedman DA (1981) Bootstrapping regression models. Ann Stat 9:1218–1228
Gonçalves S (2011) The moving blocks bootstrap for panel linear regression models with individual fixed effects. Econom Theory 27:1048–1082
Gonçalves S, White H (2005) Bootstrap standard error estimates for linear regression. J Am Stat Assoc 100:970–979
Hall P, Horowitz JL, Jing B-Y (1995) On blocking rules for the bootstrap with dependent data. Biometrika 82:561–574
Hidalgo J (2003) An alternative bootstrap to moving blocks for time series regression models. J Econom 117:369–399
Hsiao C (2003) Analysis of panel data. Cambridge University Press, Cambridge
Hsiao C, Pesaran MH (2008) Random coefficient panel data models. In: Mátyás L, Sevestre P (eds) The econometrics of panel data. Springer, Berlin
Kapetanios G (2003) Determining the poolability properties of individual series in panel datasets. Queen Mary, University of London Working Paper No. 499
Kapetanios G (2006) Cluster analysis of panel data sets using non-standard optimisation of information criteria. J Econ Dyn Control 30:1389–1408
Kapetanios G (2008) A bootstrap procedure for panel data sets with many cross-sectional units. Econom J 11:377–395
Lahiri SN (2002) On the jackknife-after-bootstrap method for dependent data and its consistency properties. Econom Theory 18:79–98
Lahiri SN, Furukawa K, Lee Y-D (2007) A nonparametric plug-in rule for selecting optimal block lengths for block bootstrap methods. Stat Methodol 4:292–321
Lehmann EL, Romano JP (2005) Testing statistical hypotheses. Springer, New York
Lin C-C, Ng S (2012) Estimation of panel data models with parameter heterogeneity when group membership is unknown. J Econom Methods 1:42–55
MacKinnon JG (2007) Bootstrap hypothesis testing. Queen’s economics department Working Paper No. 1127
Mathai AM, Provost SB (1992) Quadratic forms in random variables. Marcel Dekker, New York
Newey WK, West KD (1994) Automatic lag selection in covariance matrix estimation. Rev of Econ Stud 61:613–653
Pesaran MH, Tosetti E (2010) Large panels with common factors and spatial correlations. J Econom 161:182–202
Pesaran MH, Yamagata T (2008) Testing slope homogeneity in large panels. J Econom 142:50–93
Pesaran MH, Chudik A (2013) Econometric analysis of high dimensional VARs featuring a dominant unit. Econom Rev 32:592–649
Pesaran MH, Smith R, Im KS (1996) Dynamic linear models for heterogeneous panels. In: Mátyás L, Sevestre P (eds) The econometrics of panel data. Springer, Berlin
Phillips PCB, Sul D (2003) Dynamic panel estimation and homogeneity testing under cross section dependence. Econom J 6:217–259
Smeekes S (2015) Bootstrap sequential tests to determine the stationary units in a panel. J Time Ser Anal (forthcoming)
White H (2000) A reality check for data snooping. Econometrica 68:1097–1126
White H (2001) Asymptotic theory for econometricians. Academic Press, New York
Zhou Z, Shao X (2013) Inference for linear models with dependent errors. J R Stat Soc Ser B 75:323–343
Author information
Authors and Affiliations
Corresponding author
Additional information
The authors would like to thank Badi Baltagi (Editor), Edith Madsen, Hans Christian Kongsted, David Edgerton, and seminar participants at Lund University for many valuable comments and suggestions. Financial support from the Knut and Alice Wallenberg Foundation and the Jan Wallander and Tom Hedelius Foundation is gratefully acknowledged.
Appendix: Proofs
Appendix: Proofs
This appendix is concerned with the proofs of Lemma 1 and Theorems 1–3. The convergence is in probability, but we generally do not add this explicitly in order to simplify the notation. The sequence \(\{a_T\}\) is at most of order \(T^{\kappa }\) in probability, denoted \(a_T = O_p(T^{\kappa })\), if \(T^{-\kappa }a_T\) converges in distribution. The sequence is of order smaller than \(T^{\kappa }\) in probability, denoted by \(a_T = o_p(T^{\lambda })\), if \(T^{-\kappa } a_T \rightarrow _p 0\). The bootstrap stochastic order symbols, denoted \(O_{p^*}(\cdot )\) and \(o_{p^*}(\cdot )\), are defined in an analogous manner.
We begin by defining the following quantities which will be used throughout this Appendix:
where \(\tilde{\varepsilon }_{i,t}^{*} = y_{i,t}^* - \theta _i - x_{i,t}^\prime \beta _i\). Also, let \(z_t\) and \(\tilde{z}_t\) be the \(mN \times 1\) stacked vectors \(z_t = (z_{1,t}^\prime ,\ldots , z_{N,t}^\prime )^\prime \) and \(\tilde{z}_t = (\tilde{z}^{\prime }_{1,t},\ldots , \tilde{z}^{\prime }_{N,t})^\prime \), with similar definitions of \(z_{t}^{*}\) and \(\tilde{z}_{t}^{*}\).
Lemma 2
Under Assumption ERR, as \(T \rightarrow \infty \),
Proof of Lemma 2
Clearly
By Corollary 3.48 in White (2001), \(( \overline{x}_{i} - \mu _{i}) = T^{-1} \sum _{t=1}^T (x_{i,t} - \mu _{i}) = o_p(1)\), and by further use of his Theorem 5.20, \(T^{-1/2} \sum _{t=1}^T \varepsilon _{i,t} = O_p(1)\). It follows that
The required result now follows by applying to \(T^{-1/2} \sum _{t=1}^T \tilde{z}_t\) a central limit theorem for mixing processes (see, e.g., White 2001, Theorem 5.20). \(\square \)
Lemma 2*
Under Assumptions ERR and BL, as \(T \rightarrow \infty \),
Proof of Lemma 2*
We have \(\varepsilon _{i,t}^* = \tilde{\varepsilon }_{i,t}^{*} - (\hat{\theta }_i - \theta _i) - x_{i,t}^\prime ({\hat{\beta }}_i - \beta _i)\). Similarly, \(\hat{\varepsilon }_{i,t} = \varepsilon _{i,t} - (\hat{\theta }_i - \theta _i) - x_{i,t}^\prime ({\hat{\beta }}_i - \beta _i)\). Using these relationships, we can write
where we have used that fact that \(\sum _{t=1}^T (x_{i,t} - \overline{x}_i) \hat{\varepsilon }_{i,t} = 0\) by the first-order conditions for \({\hat{\beta }}_i\). By adding and subtracting appropriately, we have
where the last equality follows from using \((\overline{x}_i - \mu _i) = o_p(1)\) (see White 2001, Corollary 3.48), \(T^{-1/2}\sum _{t=1}^T \tilde{\varepsilon }_{i,t}^{*} = O_{p^*}(1)\) (see Fitzenberger 1997, Theorem 3.1), and \(T^{-1/2} \sum _{t=1}^T \varepsilon _{i,t} = O_p(1)\) (see White 2001, Theorem 5.20). The required result now follows from the same argument as in Fitzenberger 1997, giving
as \(T \rightarrow \infty \). \(\square \)
Proof of Theorem 1
We begin by proving the asymptotic distribution under \(H_0\), in which case
This implies
or, in stacked form,
where \(\xi _T = \big ( \xi _{1,T}^\prime ,\ldots ,\xi ^\prime _{N,T}\big )^\prime \), \({D} = {\tau }_N^\prime \otimes {I}_k\), \({Q}_{T} = \mathrm {diag}({Q}_{1,T} ,\ldots , {Q}_{N,T})\), and \(\hat{\Sigma }_T\) is a diagonal \(Nm\times Nm\) matrix whose diagonal is given by \((\hat{\sigma }_1^2,\ldots , \hat{\sigma }_N^2)^\prime \otimes {\tau }_k\).
By the properties of the LS residuals and Corollary 3.48 in White (2001), we have that \(\hat{\sigma }_i^2 - \sigma _i^2 = o_p(1)\), where \(\sigma _{i}^2 = E\big (\varepsilon _{i,t}^2\big )\). Note also that, by Lemma 2 and Assumption REGR, \(\xi _T \rightarrow _d \xi \) as \(T \rightarrow \infty \), where \(\xi \sim N({0}, \Sigma )\), and \(Q_{T} - Q =o_p(1)\), where \({Q} = \mathrm {diag}({Q}_{1} ,\ldots , {Q}_{N})\). It follows that
where \(=_d\) signifies equality in distribution, \(Z_1 =\Sigma _0^{-1/2} Q^{-1/2} \xi \), \(Z_1 \sim N(0, B)\) with \({B} = \Sigma _0^{-1/2} Q^{-1/2}\Sigma \Sigma _0^{-1/2} Q^{-1/2}\), \(\Sigma _0\) is \(\hat{\Sigma }_T\) with \(\hat{\sigma }_i^2\) replaced by \(\sigma _i^2\), \(H = {D} \Sigma _0^{-1/2} Q^{1/2}\) and \(A = {I}_{kN} -{H}^\prime ({H}{H}^\prime )^{-1}{H}\). Since A is symmetric,
where \(\lambda _1, \ldots , \lambda _{mN}\) are the eigenvalues of \({B}^{1/2} {A} {B}^{1/2}\) (see Mathai and Provost 1992). Noting that A and B are (non-stochastic) positive semidefinite matrices, we have that \(\lambda _j \ge 0\) for all j. This establish the required result under \(H_0\).
To show consistency, we consider an alternative hypothesis of the form \(H_{1} : \beta _i = \beta + \delta _i\), where \(\delta _i\) are \(m \times 1\) vectors of fixed constants such that \(\Vert \delta _i \Vert \le C < \infty \) for all i and \(\delta _i \ne \delta _h\) for some pair \(i \ne h\). Under this alternative,
with an obvious definition of \(c_T\). The last term on the right is \(O_p(1)\), suggesting that
Since \(\delta _i \ne \delta _h\) for some \(i \ne h\), it must hold that \(\delta _i - c_T \ne 0\) for at least one i. Since \({Q}_i/\sigma _i^2\) is positive definite, this means that \(T( {\hat{\beta }}_i - {\hat{\beta }}_\mathrm{{WFE}})^\prime Q_{i,T} ( {\hat{\beta }}_i - {\hat{\beta }}_\mathrm{{WFE}} )/\hat{\sigma }_i^2 \rightarrow \infty \) as \(T \rightarrow \infty \), and so the proof is complete. \(\square \)
Proof of Theorem 2
Define \(\hat{\Sigma }_T^*\) as the diagonal matrix whose diagonal is given by \((\hat{\sigma }_1^{*2},\dots , \hat{\sigma }_N^{*2})^\prime \otimes {\tau }_k\), and assume for the moment that there exists a diagonal matrix \(\Sigma _0^*\) satisfying \(\hat{\Sigma }_T^* - \Sigma _0^* = o_{p^*}(1)\). Under \(H_0\),
where \(\hat{\sigma }_i^{*2}\) is the usual LS estimate of the variance from the bootstrap procedure. This result, together with the same arguments used in the proof of Theorem 1, implies that
where \({H}^* = {D}(\hat{\Sigma }_T^*)^{-1/2} Q^{1/2}\) and \(Z_1^* \sim N({0}, {B}^*)\) with \({B}^* = (\hat{\Sigma }_T^*)^{-1/2} Q^{-1/2} \Sigma (\hat{\Sigma }_T^*)^{-1/2} Q^{-1/2}\). Hence, as in the proof of Theorem 1, if we can show that \(\hat{\Sigma }_T^* - \Sigma _0 =o_{p^*}(1)\), then
and the second result of the theorem follows. We now verify that \(\hat{\Sigma }_T^* - \Sigma _0 =o_{p^*}(1)\). Under \(H_0\),
where \({M}_{x_i} = {I}_T - x_i (x_i^\prime {M}_{\tau } x_i)^{-1}x_i^\prime \). The second equality follows from straightforward algebra under the null hypothesis. We can use an argument similar to Lemma 2 to show that \(T^{-1/2} x_i^\prime {M}_{\tau } \varepsilon ^*_i = O_{p^*}(1)\), and the third equality follows. Finally, by Theorem 3.1 in Fitzenberger (1997), since \(\varepsilon _{i,t}^* = \tilde{\varepsilon }_{i,t}^{*} + \hat{\varepsilon }_{i,t} - \varepsilon _{i,t}\) and \(T^{-1} \sum _{t=1}^T \hat{\varepsilon }_{i,t} = 0\), we have \(\overline{\varepsilon }_i^* = T^{-1} \sum _{t=1}^T (\tilde{\varepsilon }_{i,t}^{*} - \varepsilon _{i,t}) = O_{p^*}(T^{-1/2})\), and so the fourth equality follows. It remains to consider \(T^{-1} \sum _{t=1}^T (\varepsilon _{i,t}^{*})^2\). By MBB–Lemma A.3 in Fitzenberger (1997), \(T^{-1} \sum _{t=1}^T [(\tilde{\varepsilon }_{i,t}^{*})^2 - \varepsilon _{i,t}^2] = o_{p^*}(1)\). We now show that \(T^{-1} \sum _{t=1}^T [(\varepsilon _{i,t}^{*})^2 - (\tilde{\varepsilon }_{i,t}^{*})^2] = o_{p^*}(1)\). For this purpose, it is convenient to define \(v_{i,t} = w_{i,t} (w_i^\prime w_i)^{-1} \sum _{s=1}^T w_{i,s}^\prime \varepsilon _{i,s}\), where \(w_i = (\tau _T, x_i)\). Also, by the properties of LS residuals, \(\hat{\varepsilon }_{i,t} = \varepsilon _{i,t} - v_{i,t}\). Hence,
The second equality follows from the fact that \(T^{-1} \sum _{t=1}^T \hat{\varepsilon }_{i,t}^2\rightarrow _p\sigma _i^2\). As for the third equality, by Corollary 3.48 in White (2001), \(T^{-1} \sum _{s=1}^T w_{i,s}^\prime \varepsilon _{i,s} = o_p(1)\), by Assumption REGR, \((T^{-1}w_i^\prime w_i)^{-1}\) converges to a nonzero deterministic matrix, and by further use of MBB–Lemma A.3 in Fitzenberger (1997), \(T^{-1} \sum _{t=1}^T (\tilde{\varepsilon }_{i,t}^{*}w_{i,t} - \varepsilon _{i,t}w_{i,t}) =o_{p^*}(1)\). Finally, by Corollary 3.48 in White (2001), \(T^{-1} \sum _{t=1}^T \varepsilon _{i,t}w_{i,t} =o_p(1)\). This establishes that \(\hat{\Sigma }_T^* - \Sigma _0 =o_{p^*}(1)\), and so the proof of the first statement of the theorem is complete.
The second part of the theorem follows from the fact that in the bootstrap the data are generated under the null of equal slope coefficients. Consequently, the result in (6) also holds under the alternative hypothesis. \(\square \)
Proof of Lemma 1
Consider the following alternative hypothesis \(H_{1} : \beta _i = \beta _b + \delta _i\), where \(\delta _i\) is a \(m \times 1\) vector of fixed constants such that \(\Vert \delta _i \Vert \le C < \infty \). By the properties of the LS estimator, \({\hat{\beta }}_i - \beta _i = O_p(T^{-1/2})\). It follows that \(\hat{\beta }_i - \hat{\beta }_b - \delta _i = o_p(1)\). Also, \(( \hat{\sigma }_i^2 Q_{i,T}^{-1} + \hat{\sigma }_b^2 Q_{b,T}^{-1} )^{-1}\) converges to a positive definite matrix by Assumption REGR. Consequently, provided that \(i \notin \mathcal {Z}\), \(W_i \rightarrow \infty \) as \(T \rightarrow \infty \), which in turn implies
Consider the second result of the lemma. We only need to consider the order statistics for which \(i \in \mathcal {Z}\), as \(W_{(k+1)} \rightarrow \infty \) if \(k + 1 \le q\) by the first part. Hence,
The second result then follows from application of the continuous mapping theorem (see, e.g., White 2000, Proposition 2.2).
In order verify the third result, we proceed in two steps. First, we establish the exact expression of \(G_i\), the limiting distribution of \(W_i\) for \(i \in \mathcal {Z}\). Second, we show that the bootstrap analog, \(W_i^*\), converges to the same distribution. Define the \(m \times Nm\) matrix \({P}_i = (0 , \ldots , 0, {I}_{m} , 0, \ldots , 0, -{I}_{m}, 0, \ldots ,0)\), where \({I}_m\) and \(-{I}_m\) are in positions i and b, respectively. In addition, define \({\hat{\beta }} = ( {\hat{\beta }}_1, \ldots , {\hat{\beta }}_N )^\prime \). In this notation, since \(\beta _i = \beta _b\) for all \(i \in \mathcal {Z}\), \(W_i\) can be written as
where \(\tilde{{Z}}_2 = \sqrt{T} \, {P}_i {\hat{\beta }} = {P}_i Q_T^{-1} \xi _T\). By Lemma 2 and Assumption REGR,
as \(T \rightarrow \infty \), and \(\hat{\Sigma }_T Q_T^{-1} - \Sigma _0^* Q^{-1} =o_p(1)\).
To analyze the bootstrap test, \(W_i^*\), define \({\hat{\beta }}^* = ( {\hat{\beta }}_1^*, \ldots , {\hat{\beta }}_N^* )^\prime \). It follows that we can write \(W_i^*\) as
Note that \({\hat{\beta }}^*\) is estimated using \(y_{i,t}^*\), which is generated under the hypothesis that \({\beta }_i = {\beta }_b\) for all i. This implies that
where \(\tilde{{Z}}_2^{*} = \sqrt{T} {P}_i {\hat{\beta }}^* = {P}_i Q_T^{-1} \xi _T^*\). By Lemma 2 and Assumption REGR, we have that
In addition, from the proof of Theorem 1, \(\hat{\Sigma }_T^* - \Sigma _0 = o_{p^*}(1)\), which implies that \(\hat{\Sigma }_T^* Q_T^{-1} - \Sigma _0 Q^{-1} = o_{p^*}(1)\). This completes the proof of the third result.
The fourth result follows from the third and the continuous mapping theorem (White 2000, Proposition 2.2). \(\square \)
Proof of Theorem 3
The first result follows from the first and last results of Lemma 1. To show the second result, note that, by the first part of Lemma 1, \(W_i \rightarrow _p \infty \) if \(i \notin \mathcal {Z}\). Therefore, if \(k = q\), \(\mathcal {D}_k^c\) equals \(\mathcal {Z}\) in the limit with probability one. It follows from second and fourth parts of Lemma 1 that the probability of not rejecting the null, \(H_0(k)\), is equal to \(1 - \alpha \) (the confidence level in Algorithm SEQBOOT). Finally, as for the third result of Theorem 3, note that
From the first two parts of the theorem, \(\sum _{k=0}^{q} \lim _{T \rightarrow \infty } P(\hat{q} = g) = 1 - \alpha \). \(\square \)
Rights and permissions
About this article
Cite this article
Blomquist, J., Westerlund, J. Panel bootstrap tests of slope homogeneity. Empir Econ 50, 1359–1381 (2016). https://doi.org/10.1007/s00181-015-0978-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00181-015-0978-z