Augmenting external control arms using Bayesian borrowing: a case study in first-line non-small cell lung cancer

Aim: This study aimed to improve comparative effectiveness estimates and discuss challenges encountered through the application of Bayesian borrowing (BB) methods to augment an external control arm (ECA) constructed from real-world data (RWD) using historical clinical trial data in first-line non-small-cell lung cancer (NSCLC). Materials & methods: An ECA for a randomized controlled trial (RCT) in first-line NSCLC was constructed using ConcertAI Patient360™ to assess chemotherapy with or without cetuximab, in the bevacizumab-inappropriate subpopulation. Cardinality matching was used to match patient characteristics between the treatment arm (cetuximab + chemotherapy) and ECA. Overall survival (OS) was assessed as the primary outcome using Cox proportional hazards (PH). BB was conducted using a static power prior under a Weibull PH parameterization with borrowing weights from 0.0 to 1.0 and augmentation of the ECA from a historical control trial. Results: The constructed ECA yielded a higher overall survival (OS) hazard ratio (HR) (HR = 1.53; 95% CI: 1.21–1.93) than observed in the matched population of the RCT (HR = 0.91; 95% CI: 0.73–1.13). The OS HR decreased through the incorporation of BB (HR = 1.30; 95% CI: 1.08–1.54, borrowing weight = 1.0). BB was applied to augment the RCT control arm via a historical control which improved the precision of the observed HR estimate (1.03; 95% CI: 0.86–1.22, borrowing weight = 1.0), in comparison to the matched population of the RCT alone. Conclusion: In this study, the RWD ECA was unable to successfully replicate the OS estimates from the matched population of the selected RCT. The inability to replicate could be due to unmeasured confounding and variations in time-periods, follow-up and subsequent therapy. Despite these findings, we demonstrate how BB can improve precision of comparative effectiveness estimates, potentially aid as a bias assessment tool and mitigate challenges of traditional methods when appropriate external data sources are available.

Researchers and health agencies need accurate and reliable estimates of the relative effectiveness of treatments. In some cases, an ‘external control group’ can be created from pre-existing healthcare records, eliminating the need for a concurrent control and the associated patient burden. However, it can be challenging to find comparable patients with similar characteristics, and comparable information on health outcomes for those patients. Often, there are just not enough patients available in existing records to make a meaningful comparison. Bayesian borrowing can be used to include control arm information from historical studies, even if these studies differed somewhat from the present study.

In this article, we investigate the ability of external controls and Bayesian borrowing to replicate findings from a randomized controlled trial (RCT – the gold standard). We begin with an RCT and then remove the control group before constructing an external control group using real-world data from electronic health record data. We then use Bayesian borrowing to add information from another historical trial in the same indication to see whether the original RCT results can be replicated, and under what conditions. Our example illustrates common challenges of working with real-world data and provides practical insights for the incorporation of additional data sources when comparing the effectiveness of treatments that have not been compared directly in an RCT.