ABSTRACT
The standard tool for analyzing time-to-event data from a randomized two-group trial is the Cox proportional hazards model [1]. The Cox model is an attractive semiparametric option when the trial size is reasonably large. However, for small trials, such as many cancer clinical trials or pilot=early phase trials in other areas, point and interval estimates of the relative risk (hazard ratio) based on the Cox model can be inefficient, thereby adversely impacting the design of potential follow-up trials. Mehrotra and Roth [2] provided a heuristic explanation for the small sample inefficiency of the Cox model and developed a more efficient method for relative risk estimation and inference based on a generalized logrank (GLR) statistic. They showed that the GLR method and the Cox model yield similar results for large trials (>100 subjects per treatment group), but the GLR-based estimates of the relative risk have notably smaller mean squared errors (MSEs) for small trials.
