Statistical inference using generalized pairwise comparisons

Abstract

In this study, the power of the generalized pairwise comparisons was compared to the power of weighted log-rank tests using simulations for three scenarios of survival differences. These scenarios includes proportional hazards, delayed treatment effect and cure rate. The overall hazard ratio for all scenarios was kept the same for these three scenarios for comparability. Under proportional hazards scenario, the standard log-rank test is more powerful than all other tests. The log-rank test loses some power in the presence of censored observations. The net benefit is less powerful than the log-rank test and loses power with increase in threshold of clinical relevance. When there is a delay in treatment effect, the net benefit gains power with an increase in clinical relevance. It is equally powerful to the Fleming and Harrington weighted log-rank test which gives more power to late failure times, for large thresholds of clinical relevance. If a proportion of patients is cured, the net benefit is more powerful than any other test if large thresholds are used.