OBJECTIVES: Basket trials are increasingly being used to investigate novel therapies targeting rare cancer mutations common to multiple histologies. Analyses of basket trials for histology-independent therapies (HIT) often employ complete pooling of information across histologies to improve power (assuming that outcomes are homogenous across histologies), or no pooling whatsoever. However, a third increasingly used option is application of Bayesian hierarchical models (BHM) to allow for partial pooling of information across histologies, with the amount of pooling dependent on the degree of between-histology heterogeneity. We extend these BHM approaches to indirect treatment comparisons (ITC) for survival endpoints.

METHODS: We modify the BHM approach of Murphy et al. (2020) to allow for unanchored ITCs between the arms of two basket trials (or artificially-constructed baskets) for survival endpoints. We allow for prognosis to differ by histology via a histology-specific random effect to mitigate confounding due to imbalances in histology. We simulate exponentially-distributed survival data for two single-arm basket trials with imbalances in the distribution of prognostically important histologies. We demonstrate the impact of partial pooling on survival curve estimates under a comparative BHM and assess the model’s ability to reduce bias in the treatment effect estimate versus a complete pooling approach on 20 simulated datasets.

RESULTS: 95% credible intervals (CrI) for the hazard ratio under the BHM captured the true effect in 19/20 simulated datasets in contrast to the complete pooling approach which only captured the true effect in 8/20. Estimated survival curves and 95% CrIs by histology will be presented.

CONCLUSIONS: We propose a BHM approach for performing ITCs between HITs for survival endpoints. The approach is implementable using individual patient data (IPD) or pseudo-IPD by histology for two basket trials or artificially-constructed baskets. We demonstrate that the approach has potential to reduce bias in effect estimation relative to a complete pooling approach.