On Testing Overidentifying Restrictions in Dynamic Panel Data Models

The conventional Sargan (1958)/Hansen (1982) test of overidentifying restrictions and the Tilting Parameter test of Imbens, Spady and Johnson (1998) are compared in the context of the AR(1) dynamic panel data model using Monte Carlo experiments. Interestingly, the size properties of the former are found to be superior in this setting. Nevertheless, the Sargan/Hansen test is found to have no power in panels of dimensions that are commonly encountered in empirical work. A simple procedure for reducing the number of moment conditions tested is shown to improve both the size and power properties of the conventional test. A useful diagnostic procedure is also suggested.