OBJECTIVES: Many modern oncology trials are designed to assess the benefit of combining a novel therapy with chemotherapy or other standard of care. It is insightful to estimate the proportion of patients who attain markedly improved survival outcomes arising from the novel agent.

METHODS: Bayesian parametric mixture models (B-PMMs) were applied to progression-free survival (PFS) data from the CLEOPATRA trial comparing a trastuzumab and docetaxel regimen with and without the targeted therapy pertuzumab (PER+TRA+DOC versus TRA+DOC) in metastatic HER2-positive breast cancer. Patients in the PER+TRA+DOC arm who do not respond to PER were assumed to exhibit a survival pattern similar to that of the TRA+DOC arm; a condition imposed via informative priors on the corresponding latency distribution. Scenario analyses explored alternative prior distributions for the proportion of responders to PER: uniform (vague), beta (optimistic [mean 30%], moderately informative) and logit-normal (optimistic [mean 30%], strongly informative).

RESULTS: The proportion of responders to PER estimated from models where prior expectation for this fraction was vague or moderately informative were highly similar (15.9% [95% CrI (credible interval): 6.6-25.5%] uniform vs 18.0% [95% CrI: 10.3-26.9%] beta). Predicted subpopulation and overall population survival patterns were likewise in close agreement (e.g., 5-year PER+TRA+DOC non-responder PFS: 9.8% [95% CrI: 6.1-14.2%] uniform vs 9.1% [95% CrI: 5.7-12.9%] beta; responder PFS: 80.2% [95% CrI: 65.2-93.6%] uniform vs 79.5% [95% CrI: 65.4-92.0%] beta). In comparison, the responder fraction estimate from the strongly informed, and therefore misled, model was somewhat greater (26.4% [95% CrI: 21.9-31.1%]). The modest uncertainty in the more weakly informed models suggests that they are sufficient in this application.

CONCLUSIONS: B-PMMs can be used to infer the proportion of patients who achieve durable response attributable to the addition of a targeted or other novel oncology therapy to a treatment regimen. Furthermore, B-PMMs appear generally robust to moderate perturbations in component prior distributions.