Many simulation experiments have shown that, in a variety of circumstances, bootstrap tests perform better than current asymptotic theory predicts. Specifically, the discrepancy between the actual rejection probability of a bootstrap test under the null and the nominal level of the test appears to be smaller than suggested by theory, which in any case often yields only a rate of convergence of this discrepancy to zero. Here it is argued that the Edgeworth expansions on which much theory is based provide a quite inaccurate account of the finite-sample distributions of even quite basic statistics. Other methods are investigated in the hope that they may give better agreement with simulation evidence. They also suggest ways in which bootstrap procedures can be improved so as to yield more accurate inference.