1 PERFORMANCE COMPARISON OF ESTIMATORS OF DYNAMIC PANEL DATA MODELS WITH CROSS-SECTIONAL HETEROSCEDASTICITY: MONTE CARLO EVIDENCES Muhammad Abdullah 1 Ph.D. Research Scholar, Department of Statistics, B.Z. University, Multan, Pakistan E-mail: mambkbzu@yahoo.com G. R. Pasha National College of Business Administration & Economics, Multan Campus, Multan, Pakistan E-mail: drpasha@bzu.edu.pk ABSTRACT: The correlation between lagged endogenous regressor and the error term, in case of dynamic panel data models, causes the least squares dummy variable (LSDV) estimator to be biased and inconsistent. This problem persists even in case of heteroscedastic errors. In 2006, Bun and Carree, firstly addressed this situation and proposed a bias- corrected LSDV estimator. Earlier other authors have considered the case of simple (static) panel data models allowing heteroscedasticity, and have proposed several versions of estimated generalized least squares (EGLS) estimators using different ways, through which the variance components are estimated. In this study, we have customized them for the dynamic panel data models with cross-sectional heteroscedastic remainder errors and have analyzed their performance as compared to the bias-corrected LSDV estimator using Monte Carlo experiments. The experimental evidences showed that the proposed estimators, particularly extended Heteroscedastic Consistent Covariance Matrix (HCCM)-based EGLS estimators HGLS1s, are attractive choices in the sense of bias and mean squared error (MSE). Key words: Bias corrected estimator; Estimated GLS; Heteroscedastic Error Components; Variance Components, HCCM. JEL Classification: C13; C23 1. INTRODUCTION A series of studies shows that the correlation between lagged endogenous regressor and the error term, in case of Dynamic Panel Data Model (DPDM), causes the least squares dummy variable (LSDV) estimator to be biased and inconsistent; see Trognon (1978), Nickell(1981), Sevestre and Trognon (1985) etc. among others. This inconsistency led researchers not only to consider the other estimating procedures (like Maximum Likelihood-ML estimation, Generalized Method of Moments-GMM estimation, Generalized Least Squares-GLS estimation etc.) but also development of new estimators or some new modifications. Whereas Bun and Carree (2006) not only showed that this problem persists even in case of heteroscedastic errors, but also suggested a bias-corrected LSDV (referred as BCLSDV hereafter) estimator for these particular DPDMs. They considered the case of heteroscedasticity coming into model through the unit-time varying (remainder) error component. Earlier Li and Stengos (1994), following the Baltagi (1988) and Wansbeek (1989), had treated the case of this type of heteroscedasticity for Simple (Static) Panel Data Models suggesting an adaptive estimator that relies on nonparametric methods. Aslam (2006) suggested an out performing adaptive GLS estimator for the simple panel data models with heteroscedastic errors extending the considerations of Mazodier and Trognon (1978), Baltagi and Griffon (1988) and Roy ( 2002). Although they all considered the case of heteroscedasticity coming into model through the unit specific error component but it is also mentionable that the suggestions of Aslam (2006), Baltagi and Griffin (1988) are parametric, whereas the suggestion of Roy (2002) is nonparametric. Lejeune (2004) was concerned with the estimation and specification testing of the one-way error components model allowing for heteroscedasticity in both the individual-specific and the remainder error term as well as for the unbalanced panel data models. His considerations were also limited to the Simple (Static) Panel Data Models. Baltagi, Bresson and Pirotte (2005) checked the sensitivity of two adaptive heteroscedastic estimators suggested by Li and Stengos (1994) and Roy (2002) for an error component regression model to misspecification of the form of heteroscedasticity. And found that Li and Stengos’s (1994) estimator that considered the case of heteroscedasticity on the remainder error term, performs better under this type of misspecification than the corresponding estimator of Roy (2002). Similarly they concluded from the Monte Carlo results that ignoring the presence of heteroscedasticity on the remainder error term has a much more dramatic effect than ignoring the presence of heteroscedasticity on the individual specific error. This is why, in this study, we concentrate on the case 1 Corresponding Author