In the application of statistical methods to clinical practice and research, the difficulty of controlling type I error is well recognized. In general, the investigator has two choices - either let the actual type I error exceed the desired nominal level by an unknown, and sometimes substantial, amount or impose a restriction for keeping the type I error within the target nominal value. The former choice will rely on methods based on large sample approximations, so- called approximate methods, and the latter choice will involve one of the exact inference procedures. Approximate...