Consideration of stratification in confirmatory trials with time-to-event endpoint

Stratification in randomization and analysis are widely employed to balance treatment groups in clinical trials. However, the potential power loss due to under-stratification or over-stratification has not been thoroughly evaluated in the typical setting of confirmatory clinical trials. In cases where there are too many strata and some have small sample sizes or a small number of events, it is common practice to combine these small strata during analysis. However, there is a lack of guidance on how those small strata should be combined. This paper presents extensive simulation studies to evaluate the impact of under-stratification or over-stratification on the power of survival analysis and the estimate of hazard ratio using stratified log-rank test and Cox PH model, respectively. The difference in power between stratified and unstratified log-rank tests is also investigated under different scenarios. Our results suggest that failing to consider prognostic stratification factors with strong effects, and/or accounting for non-prognostic factors such as noise and predictive factors, may reduce the power of the stratified log-rank test. Additionally, methods of combining small strata are explored and compared.