Generalized pairwise comparisons for censored data: An overview.

Area of Science:

• Biostatistics
• Survival Analysis
• Nonparametric Statistics

Background:

• The Wilcoxon-Mann-Whitney test is a standard for comparing two groups.
• Censored data presents challenges in statistical analysis due to missing information.
• Various generalized pairwise comparison (GPC) methods exist to address censored data.

Purpose of the Study:

• To evaluate generalized pairwise comparison (GPC) methods for censored data.
• To compare GPC methods in hypothesis testing and estimation of treatment effects.
• To provide recommendations for selecting appropriate GPC methods based on censoring levels.

Main Methods:

• Review and comparison of GPC methods, including those ignoring noninformative pairs (Gehan, Harrell, Buyse), imputation-based methods (Efron, Peron, Latta), and inverse probability of censoring weighting (IPCW, Datta, Dong).
• Evaluation of statistical power and estimation properties under varying censoring proportions.
• Assessment of the 'net benefit' measure as a treatment effect estimator.

Main Results:

• Methods ignoring uninformative pairs demonstrate comparable power to complex methods in low censoring scenarios.
• These simpler methods show superior performance with high censoring proportions (>40%).
• Harrell's c-index is an unbiased estimator for net benefit under proportional hazards; imputation or IPCW methods are unbiased up to 60% censoring.

Conclusions:

• Simpler GPC methods (ignoring uninformative pairs) are efficient and powerful for censored survival data, particularly with high censoring rates.
• The choice of GPC method depends on the specific research question (hypothesis testing vs. estimation) and the extent of censoring.
• Imputation or IPCW methods are recommended for unbiased net benefit estimation when proportional hazards do not hold and censoring is substantial.