OBJECTIVES: To evaluate treatment effect in a randomized control trial (RCT) using external control arm (ECA) from real-world data (RWD) in the presence of poor balance of baseline characteristics between RCT and ECA.

METHODS: Participants were randomized to one of two experimental arms or an internal control arm (ICA) of a phase II RCT and were evaluated for bleeding events within three months. The Finnish healthcare system was used to develop an anonymized cohort, fulfilling eligibility for the RCT, to use as an ECA. Propensity-score (PS) methods based on 28 known confounders were used to create statistically comparable groups between RCT participants and those eligible from RWD. After PS trimming, inverse probability of treatment weighting (IPTW), G-computation, augmented IPTW (AIPTW), targeted maximum likelihood estimation (TMLE), and overlapping weights (OW) methodologies were applied for comparing the ICA to ECA, and then the treatment effect of the combined ICA+ECA to the RCT experimental arm was evaluated. In addition, a comprehensive Bayesian Dynamic Borrowing approach and the propensity score composite likelihood (PSCL) method were applied.

RESULTS: The RCT had n=505 participants randomized to the experimental arms and n=250 to the ICA. A pool of eligible Finnish RWD (n=3327) was established, of which n=2847 were included after trimming. Although the PS distributions showed poor overlap between the RCT and Finnish ECA subjects, all methods evaluated showed similar results for the proportion of participants experiencing an event. Additionally, from RCT data with augmented control arm (ICA n=250 + ECA n=2847 = 3097 control subjects), the same conclusions can be drawn as the original RCT study.

CONCLUSIONS: Several causal inference methods are available for analyzing RCT data combined with RWD. Many of these methods were applied and showed similar results arriving to the same conclusion in the presence of poor PS distribution overlap.