Applying Bias Correction Methods to Build Hybrid Controls Using Real-World Patients for a Phase IIb Randomized Controlled Trial of Baricitinib for Rheumatoid Arthritis

Abstract

Studies incorporating randomized controlled trial (RCT) controls with real-world (RW) patients face concerns of selection bias and unmeasured confounding. To evaluate two methods (Match, Test-then-Pool [MTP] and Matching and Bias Adjustment [MBA]) to construct a hybrid control arm using concurrent RCT control patients and trial-eligible RW patients from the CorEvitas Rheumatoid Arthritis (RA) Registry to replicate the treatment effect obtained from the original RCT. The measures considered estimate the effect of baricitinib on the proportion achieving 20% improvement in American College of Rheumatology response (ACR20) and change in disease activity (Clinical Disease Activity Index, CDAI) for patients with moderately to severely active rheumatoid arthritis. Patients included RCT treated (n = 102), a 50% random sample of the RCT control (n = 49), and trial-eligible RW (n = 359). MTP used optimal propensity score (PS) matching to replace half of the RCT control group with a comparable RW control group (n = 49) and tested for differences between RCT and RW control outcomes before combining into the hybrid control arm for analysis. MBA leveraged 3 sets of optimal PS matches to identify a RW control group and adjusted for residual differences in outcomes between RCT and RW controls during analysis when creating the hybrid control arm. Optimal matching produced a matched RW cohort with satisfactory balance on the majority of key covariates compared to RCT participants. Using MTP, the odds ratio (OR) for ACR20 from covariate adjusted modeling was 4.71 (95% confidence interval [CI] 2.51, 9.15). Using MBA, ACR20 treatment effect from bias and covariate adjusted modeling was 4.76 (2.10, 10.81). For reference, the OR for the treatment difference of ACR20 for the trial-eligible population was 3.81 (2.13, 6.95) and was 4.74 (2.32, 10.00) for the RCT population with 50% sample of controls. With MTP with covariate adjustment, the treatment effect of mean CDAI change was -8.96 (-12.58, -5.35). For MBA with bias and covariate adjustment, the treatment effect of mean CDAI change was -9.85 (-15.16, -5.02). In the original RCT, the mean CDAI change was -8.88 (-12.58, -5.18) and for the RCT population with 50% sample of controls, the mean CDAI change was -10.10 (-14.46, -5.74). With both hybrid control methods using the combination of well-matched RWD and RCT control patients, we replicated the Phase IIb trial results. Trial Registration: Data were used from Eli Lilly and Company's clinical trial (NCT01185353) and the CorEvitas Rheumatoid Arthritis Registry (NCT01402661), both registered at clinicaltrials.gov.

Keywords: external control arm; hybrid control arm; randomized controlled trial; real‐world data; registry data.