Impact of unequal censoring on the comparison of survival curves

Oregon Convention Center 

This study investigates the impact of unequal censoring on comparisons of survival distributions. We evaluated the performance of five statistical tests: the log-rank (LR), Gehan-Breslow generalized Wilcoxon (GB), Tarone-Ware (TW), Peto-Peto (PP), and modified Peto-Peto (mPP) tests. Using 1,000 simulations comparing two survival curves, we assessed their size and power under four censoring patterns: overall, early, middle, and late censoring (total of 16 combinations). Additionally, we explored different scenarios with sample sizes of 20 and 50 per group and varying levels of censoring (10% and 25%). Regardless of sample size, censoring percentage, censoring patterns, LR test demonstrated the highest power overall while GB showed the least power. For a sample size 20 per group, only early-overall censoring reached more than 80% power for LR and TW tests. For a sample size 50 per group, early-overall, early-early, early-middle, early-late censoring appeared to minimally impact all five individual tests. We further investigated the effect of combining their p-values into a single value. The combined p values using the generalized Fisher method had higher power than LR test; however, the type I error rates were not well maintained compared to LR test.

Log-rank test

Gehan-Breslow generalized Wilcoxon test

Tarone-Ware test

Peto-Peto test

Modified Peto-Peto test

Survival analysis 

Biometrics Section