The Wild Bootstrap, Tamed At Last

Various versions of the wild bootstrap are studied as applied to regression models with heteroskedastic errors. We develop formal Edgeworth expansions for the error in the rejection probability (ERP) of wild bootstrap tests based on asymptotic t statistics computed with a heteroskedasticity consistent covariance matrix estimator. Particular interest centers on the choice of the auxiliary distribution used by the wild bootstrap in order to generate bootstrap error terms. We find that the Rademacher distribution usually gives smaller ERPs, in small samples, than the version of the wild bootstrap that seems most popular in the literature, even though it does not benefit from the latter's skewness correction. This conclusion, based on Edgeworth expansions, is confirmed by a series of simulation experiments. We conclude that a particular version of the wild bootstrap is to be preferred in almost all practical situations, and we show analytically that it, and no other version, gives perfect inference in a special case.