INTEGRATE pooled phase 2/3 results are robust to postprogression switching and the winner's curse

Abstract

Background: The INTEGRATE phase 3 trial in advanced gastric and esophagogastric junction cancer involved pooling overall survival data with its preceding phase 2 trial, raising concerns about misalignment due to treatment switching in phase 2, or the "winner's curse." We evaluated phase 2 results, adjusted for these opposing effects, against phase 3 according to the prespecified statistical analysis plan.

Methods: Overall survival estimates were adjusted for treatment switching using the rank-preserving structural failure time model (RPSFTM) and inverse probability of censoring weights (IPCW) method. A novel shrinkage approach mitigated overestimation from the winner's curse, and Bayesian prediction methods predicted phase 3 outcomes from phase 2 estimates. A simulation study modeled 10 000 seamless phase 2/3 trials to quantify bias in the pooled estimate.

Results: The observed phase 3 hazard ratio (HR = 0.71, 95% CI = 0.54 to 0.93) for overall survival was more conservative than the adjusted phase 2 estimates (RPSFTM and novel shrinkage approach: HR = 0.61, 95% CI = 0.29 to 1.29; RPSFTM and Bayesian prediction: HR = 0.59, 95% CI = 0.48 to 0.73; IPCW and novel shrinkage approach: HR = 0.55, 95% CI = 0.31 to 0.99; IPCW and Bayesian prediction: HR = 0.58, 95% CI = 0.46 to 0.72). Simulations indicated negligible bias in the pooled log hazard ratio of ‒0.011 and 0.005 under the null and alternative hypotheses, respectively.

Conclusion: Adjusting phase 2 estimates for both treatment switching and the winner's curse produced point estimates similar to the unadjusted phase 3 results. A prospective plan to pool trial data under a closed testing procedure may be a reasonable strategy when a recruitment shortfall in phase 3 is anticipated, provided that potential sources of misalignment are thoroughly assessed.

Clinical trial information: ACTRN12612000239864 (INTEGRATE I)NCT02773524 (INTEGRATE IIA).

Figure 1.

Hazard ratios for estimands of interest. Abbreviations: HR = hazard ratio; IPCW = inverse probability of censoring weights; RPSFTM = rank-preserving structural failure time model.

Figure 2.

Impact of violating the common treatment effect assumption on the hazard ratio in the rank-preserving structural failure time model (log rank, no recensoring). A larger ratio of treatment effect after progression to random assignment means that the effect of regorafenib is stronger after disease progression than at random assignment.