Bootstrap for Panel Data

Abstract

This paper considers bootstrap methods for panel data. Theoretical results are provided for the sample mean. It is shown that the resampling only in the cross section dimension is not valid in the presence of temporal heterogeneity. The block resampling only in the time series dimension is not valid in the presence of cross section heterogeneity. The double resampling that combines the two previous resampling methods, is valid for panel data models with cross section and/or temporal heterogeneity, with or without spatial dependence. This approach also avoids multiple asymptotics that may arise in large panel models. Simulations confirm these theoretical results.

Keywords: Bootstrap, Panel Data Models

JEL Classification: C15, C23