Imbalanced censoring as a bias in evaluation of randomized controlled trials: A simulation-based study.

e13606 Background: Censoring imbalances between experimental and control arms in clinical trials efficacy can compromise efficacy analysis especially if surrogate endpoints are used. We conducted a simulation-based study to assess the impact of imbalanced censoring in clinical trials on the probability of false-positive outcomes. Methods: With BayesCTDesign R package we created a data set of 500,000 simulated trials. The trials had initially negative results, varying sample sizes (100-2000 patients) and median survival times (1.5-9.5 mo.) with no baseline censoring. Trials with no significant baseline differences (p > 0.2) between experimental and control arms were selected. We randomly censored samples of patients from control (0-20% of sample) and experimental (0-40%) arms for analysis. A logistic regression model was created to assess probability of positive results due to censoring (HR<1 and p value after censoring <0.05). Sample sizes and difference in % of censored subjects were predictors as well as interaction term between these variables. Results: A data set consisting of 399,946 trials was included in the analysis. After random censoring, a statistically significant positive transformation was observed in 61797 trials with corresponding HR<1 (0.59-0.91; p < 0.05) in favor of experimental arm. A logistic regression model with above mentioned variables returned a following equation (p <0.05 for all the variables): y = -6.45005797 + 0.27661437*B1 + -0.000070584*B2 + 0.000304167*B3 where B0 is the intercept, B1 – difference in percentage of censored subjects between arms (per %), B2 – sample size (per subject) and B3 – an interaction term (B1*B2). Pseudo R2 coefficient was 0.715. Based on the model we estimated a predicted probability of false-positive results for the trials. Conclusions: Clinicians and researchers should pay special attention to censoring imbalances in the clinical trials as these may have a significant impact on results consistency, validity and lead to false-positive findings. [Table: see text]