GMM Bootstrapping and Testing in Dynamic Panels

Two different bootstrap approaches for GMM estimation have recently been suggested for use in dynamic panel data models (Brown & Newey (1995) and Hall & Horowitz (1996)) In this paper we compare the small sample properties of these estimators, suggest how sequential testing can be conducted within the GMM bootstrapping framework, and investigate the performance in a sequence of tests where we seek to find the correct lag length of a dynamic model. This comparison is carried out by means of Monte Carlo experiments. Our findings are that i) the Brown and Newey method has superior size properties, but cannot be used in a sequence of tests without modifications, and that ii) the Hall and Horowitz method works better than the usual asymptotic tests in a sequence of tests.