Improved inference for interactive fixed effects model under cross-sectional dependence

Abstract

This paper proposes an inference procedure for the interactive fixed effects model that is valid in the presence of cross-sectional dependence. When the error terms are cross-sectionally dependent, the least square (LS) estimator of this model is asymptotically biased and therefore the associated confidence interval tends to have a large coverage error. To address this, we propose a bias correction of the LS estimator and a cross-sectional dependence robust variance estimator to construct associated test statistics. The paper also discusses practical issues in implementing the proposed method, including the construction of distance that reflects the decaying pattern of cross-sectional dependence and the selection of the bandwidth parameters. Monte Carlo simulations show our procedure works well in finite samples. As empirical illustrations, we apply our procedure to study the effect of divorce law reforms on divorce rates and the impact of clean water and sewerage interventions on child mortality.