Heteroskedasticity-Robust Standard Errors for Dynamic Panel Data Models with Fixed Effects

For linear dynamic panel data models with fixed effects, practitioners often use clustered covariance estimators for inference in the presence of cross-sectional or temporal heteroskedasticity in idiosyncratic errors. The performance of a clustered estimator heavily depends on the magnitude of the cross-sectional dimension(n). When n is small, inferences using clustered estimators are compromised. A paper by Stock and Watson (2008) provides a solution under strict exogeneity if the idiosyncratic errors are possibly heteroskedastic but serially uncorrelated. Their method, however, is not generalizable to dynamic panel data models, although heteroskedasticity-robust inferences have natural relevance to dynamic models due to the requirement of serial uncorrelatedness for model identification. In the present paper, we provide a solution for instrumental variables and generalized method of moments estimators using predetermined instruments, including popular estimators for dynamic panel models. Asymptotics are established, and the findings are verified by simulations.