Abstract:
|
Motivated by a randomized clinical trial, where some prognostic factor was discovered during the conduct of the study but not pre-specified in the protocol as a stratification factor, this presentation focuses on the evaluation of the estimation bias, coverage and power of analysis of different types of endpoints with or without adjusting for the prognostic factor in the analysis. Simulation studies show that for time to event data, when the analysis model doesn't adjust for such factor, bias and loss of power exist even if the balance of treatment groups with respect to the prognostic factor is maintained. For binary data, when the cell size decided by the prognostic factor is not balanced across the treatment groups, the estimation can be either biased up or down, mainly driven by the responses contributed from the cell with larger cell size. By contrast, stratified analysis or regression analysis adjusting for such prognostic factor in general gives unbiased estimate with proper coverage and higher power. Thus stratified or regression analysis adjusting for the prognostic factor is suggested even though it is not pre-specified as a factor in stratified randomization.
|