Making sense of clinical trial data: is inverse probability of censoring weighted analysis the answer to crossover bias?

Abstract

Ideally, therapeutic interventions are evaluated through randomized clinical trials. These trials are commonly analyzed with an intent-to-treat (ITT) approach, whereby patients are analyzed in their assigned treatment group regardless of actual treatment received. If an interim analysis of such trials demonstrates compelling evidence of a difference in benefit, ethical considerations often dictate that the trial be unblinded and participants be provided access to the more efficacious agent. Because interim analysis may not address longer-term outcomes of interest, important clinical questions such as overall survival benefit-the ultimate test of efficacy to many-may remain unanswered. The ensuing crossover disturbs randomization and may lead to biased longer-term analysis, compromising the utility of clinical data. This has been especially apparent in recent adjuvant and prevention breast cancer trials. We consider four such trials: HERA (Herceptin Adjuvant), NSABP P-1 (National Surgical Adjuvant Breast and Bowel Project Breast Cancer Prevention P-1), MA.17, and BIG 1-98 (Breast International Group 1-98), the long-term outcomes of which were complicated by unblinding and selective crossover. We also discuss the biases associated with ITT analysis and, alternatively, censoring of follow-up data (ie, dropping out) after selective crossover. Moreover, we discuss how the statistical procedure of inverse probability of censoring weighted (IPCW) analysis may be used to account for selective crossover as an alternative to ITT or censoring analysis, as was recently done for the BIG 1-98 trial. Notably, IPCW analysis may be particularly suited for detecting overall survival benefits that otherwise would not be detected with an ITT approach, as reported for the BIG 1-98 trial.