Performance Improvement of Parameter Estimation by Two-Step Analysis Incorporating Generalized Log-Rank Statistics

• Activity Number: 363
• Type: Contributed
• Date/Time: Tuesday, August 2, 2016 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Junji Moriya* and June Li
• Companies: Kyowa Kirin Pharmaceutical Development and Kyowa Kirin Pharmaceutical Development
• Keywords: Time To Event Data ; Stratified Cox Model ; Small Sample ; Parameter Estimation ; Simulation and Application

Abstract

In a randomized trial involving a Time-To-Event endpoint, the Cox model has been utilized to estimate the hazard ratio (HR). Also, taking into consideration about a stratum, the stratified Cox model often has been utilized. However, for small trials, the estimated HR from the Cox model could be unstable. The stratified Cox model needs the assumption that the HRs are constant across all the strata. If the assumption is not correct, the estimated HR could be biased. To address those problems, Mehrotra and Roth (2001, 2011) provided an efficient method based on a generalized log-rank (GLR) statistic to estimate HR for small trials. Additionally, Mehrotra, et al. (2012) provided the two-step analysis method which is more applicable than the usual stratified Cox model. They suggested that the two-step analysis method may be enhanced more by incorporating the GLR statistic. However, it has never shown how far the method can actually be improved. To clarify the performance improvement, we will show the simulation results among the usual stratified Cox model and the two-step analysis methods with/without GLR statistic under some realistic situations.