Comparison of Four Tests for Equality of Survival Curves in the Presence of Stratification and Censoring

Abstract

Monte Carlo experiments investigated the small sample behavior of 4 tests used to analyze stratified censored human survival data. The size of the generalized Gehan and Mantel-Haenszel statistics were near nominal level .alpha. = 0.05. The observed size of the Cox partial likelihood ratio statistic was typically 0.065 for fewer than 32 observations per stratum and fell to near nominal levels for larger sample sizes. The continuity corrected Mantel-Haenszel statistic was consistently conservative, with observed size often lower than 0.035, until total observations exceeded 32 per stratum. The generalized Gehan statistic was only slightly less powerful than the other statistics against the proportional hazards alternatives studied. The power of these statistics depended mainly on the total number of deaths observed and the otherwise little affected by the degree of censoring or the number of strata.