Cross-section dependence in nonstationary panel models: a novel estimator

This paper uses Monte Carlo simulations to investigate the impact of nonstationarity, parameter heterogeneity and cross-section dependence on estimation and inference in macro panel data. We compare the performance of standard panel estimators with that of our own two-step method (the AMG) and the Pesaran (2006) Common Correlated Effects (CCE) estimators in time-series panels with arguably similar characteristics to those encountered in empirical applications using cross-country macro data. The empirical model adopted leads to an identification problem in standard estimation approaches in the case where the same unobserved common factors drive the evolution of both dependent and independent variables. We replicate the design of two recent Monte Carlo studies on the topic (Coakley et al, 2006; Kapetanios et al, 2009), with results confirming that the Pesaran (2006) CCE approach as well as our own AMG estimator solve this identification problem by accounting for the unobserved common factors in the regression equation. Our investigation however also indicates that simple augmentation with year dummies can do away with most of the bias in standard pooled estimators reported --- a finding which is in stark contrast to the results from earlier empirical work we carried out using cross-country panel data for agriculture and manufacturing (Eberhardt & Teal, 2008; Eberhardt & Teal, 2009). We therefore introduce a number of additional Monte Carlo setups which lead to greater discrepancy in the results between standard (micro-)panel estimators and the novel approaches incorporating cross-section dependence. We further highlight the performance of the pooled OLS estimator with variables in first differences and speculate about the reasons for its favourable results