Avoiding Loss of Power in Conditional Tests with 2x2xK Contingency Tables

The commonly-used statistic for testing conditional independence in multiple 2×2 tables has a conditional distribution with relatively few possible values. This unnecessary discreteness leads to conservative tests. Kim and Agresti (1995) remedy this conservativeness by partitioning the sample space based on a secondary static-the null hypothesis probability of various tables with the same sufficient statistic. However, the Kim and Agresti test requires calculation of the exact product hypergeometric probabilities. We propose an alternative test statistic that may assume as many distinct values as there are elements of the conditional sample space. For this statistic, tests may be based on either exact computation of the p-value or a chi-square distribution approximation.

کلیدواژه ها:

exact test ، hypergeometric ، Mantel-Haenszel