Significance levels in multiple comparison tests.

This article is concerned with the ubiquitous problem of determining the appropriate significance level when a test is multiple in nature, or the appropriate confidence region when more than one parameter is under consideration. A simple but little-known rule is given for correctly determining the level or region; its application is illustrated with some problems common in health services research, including cost comparisons among several health plans sampled over several cities, comparisons of the proportional frequencies of different types of surgery among several hospital types, and investigation of all possible pair correlations within a group of patient attributes.