Bootstrapping subset test statistics in IV regression

The finite-sample performance of bootstrap procedures is studied by simulation in a linear regression model containing 2 endogenous regressors. Besides several residual-based bootstrap procedures, we also consider the GMM bootstrap. The test statistics include Wald type t-statistics based on k-estimators and the robust subset (quasi) LR statistic. In the simulations, the restricted
fully efficient (RFE) bootstrap based on Fuller estimates and the LIML t-statistic performs best of the Wald type statistics. Unfortunately, the bootstrap only marginally reduces the conservativeness of the subset QLR statistic. Finally, the GMM bootstrap does not seem to improve upon the asymptotic approximation. An empirical example illustrates the use of these procedures.