A two stage procedure in which the samples from the original bootstrap population are themselves bootstrapped. The technique can give confidence intervals of more accurate coverage than simple bootstrapping.