Generalized pairwise comparisons using pseudo-observations for time-to-event censored data in a randomized controlled...

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Survival Analysis

Background:

• Generalized pairwise comparison (GPC) methods extend Mann-Whitney for prioritized outcome ranking in randomized controlled trials (RCTs).
• GPC methods can analyze censored time-to-event data, but bias can arise from indeterminate pairs due to censoring.
• The win ratio (WR) is a GPC measure interpretable as the inverse hazard ratio, but sensitive to censoring.

Purpose of the Study:

• To propose a novel GPC method using pseudo-observations to address uninformative pairs caused by censoring in time-to-event outcomes.
• To evaluate the performance of the proposed method against existing GPC approaches via simulations under diverse censoring scenarios.

Main Methods:

• Development of a novel GPC approach utilizing pseudo-observations to handle censored data.
• Comparative simulation studies assessing performance under various censoring proportions and distributions (equal vs. unequal dropout).
• Evaluation of bias, root mean squared error, and statistical power against established GPC methods (Gehan and Latta).

Main Results:

• The proposed method shows comparable performance to existing GPC methods under equal dropout and administrative censoring.
• For unequal dropout, performance is contingent on censoring proportion and distribution, with the new method reducing bias and RMSE.
• Statistical power improvements were not observed with the proposed method, despite reductions in bias and error.

Conclusions:

• The novel pseudo-observation GPC method effectively mitigates bias and error associated with censoring in time-to-event data, particularly in scenarios with unequal dropout.
• While not increasing statistical power, the method offers a valuable alternative for analyzing RCT data with censored outcomes.
• The approach was successfully illustrated using reconstructed randomized controlled trial datasets.