On the finite-sample performance of CCE and FE estimators in dynamic heterogeneous panels with mixed integration orders of the latent factors

Recent advances in the panel literature show that CCE estimators can be valid under more general time-series properties of the latent common factors than previously assumed. Asymptotic normality can be derived by normalizing the second moment matrix of the factors, thereby ensuring that its limit is full rank even when the factors are nonstationary. This study provides Monte Carlo evidence on the finite-sample performance of CCE estimators under the empirically relevant case characterized by the coexistence of dynamic specification, heterogeneous slopes, and mixed integration orders of the factors. In addition, we compare CCE and FE estimators across scenarios in which the regressors are either uncorrelated with or correlated with the factor loadings. Furthermore, we distinguish between weak and strong factors, according to the asymptotic behavior of the factor loadings. We impose full rank on the mean factor loading matrix across all experiments. Under uncorrelated loadings, the CCE estimators perform well and are only negligibly affected by the presence of nonstationary factors. Efficiency improves markedly as N increases. When the DGP is dynamic, the CCE-MG estimator outperforms the other techniques considered. However, under static specifications, CCE-MG and the mean-group version of the two-way FE estimator perform very similarly in terms of bias, with the former still being the most efficient. Under correlated loadings, both CCE and FE estimators can be substantially biased and typically lose efficiency. The pooled OLS estimator remains severely biased and inefficient across all experiments, except under static specification with uncorrelated loadings and weak factors.