On Bootstrap Standard Errors in Dynamic Panel Data Models

It is a well-known property that standard GMM estimators for dynamic panel data might perform poorly in small samples. Several papers have noted this to be especially true for the estimated standard errors, which are normally biased downwards. The aim of the present paper is to compare how two recently suggested bootstrap procedures can assist inference in dynamic panel data models, when the mentioned small-sample bias is a potential problem. We do this by means of Monte Carlo experiments, forming tests using both standard errors estimated by asymptotic approximations, as well as by bootstrap procedures. The results give a fairly clear support for using bootstrap inference. Whereas the tests based on asymptotics have empirical levels that may deviate substantially from their nominal ones, the bootstrap procedures appear to perform quite well in the context of dynamic panel data estimation.