Randomized Phase III Oncology Trials: A Survey and Empirical Bayes Inference

Abstract

Mandatory registration of eligible clinical trials at clinicaltrials.gov provides an unprecedented opportunity to survey existing trials that is immune to potential publication bias. Phase III oncology clinical trials represent the gold-standard evaluation on new therapeutic interventions to fight one of the most deadly diseases. Yet, the collective performance of these trials is not well understood. Through clinicaltrials.gov, we identified 130 eligible randomized phase III oncology trials in 2008–2012 and extracted results from 122 using clinicaltrials.gov and other sources. We estimated the distribution of the effect size of randomized phase III oncology trials through a new Bayesian deconvolution method, which allows the calculation of several performance measures. We found that about 22% of the interventions for cancer treatment that reach phase III have null or negative efficacy, and another 30% with strong positive efficacy. For the rest interventions, the majority have modest efficacy, which leads to insufficient power. The false positive rate is low at 0.13%, whereas as high as 33.5% of the trials could be false negatives. The distribution of the effect sizes also provide a tangible prior for Bayesian inference of future trials.