Evaluation of methods for building a hybrid real-world (RW) control arm for baricitinib Phase IIB randomized controlled trial (RCT) in patients with rheumatoid arthritis (RA)

Background: Studies augmenting RCT controls with data from RW patients face concerns of potential selection bias and unmeasured confounding.

Objectives: To evaluate performance of two methods that utilize hybrid control arms to estimate effectiveness of baricitinib on change from baseline in disease activity as measured by Clinical Disease Activity Index (CDAI).

Methods: A random sample of 50% RCT controls was optimal propensity score matched to trial-eligible RW patients from the CorEvitas RA registry as an initial step for both a Match-Test-and-Pool (MTP) and Matching and Bias Adjustment (MBA) analysis. The MTP limits bias from RW patients by testing for outcome differences between the matched RCT and RW control groups and pooling if no significance

difference is found. The MBA method additionally matches 50% RCT control sample to RCT treated patients, and remaining RCT treated to RW controls. The MBA method then estimates a bias adjustment parameter using a Bayesian regression model with covariates that control for remaining differences between control groups. 500 random samples were drawn from posterior distribution of the bias parameter and used for bias correction to RW controls in regression models comparing the RCT treated to the hybrid controls. Rubin’s rules were used to pool treatment effect estimates.

Results: Matching of control groups considerably improved on covariate imbalance. According to the MTP method, the mean change in CDAI in RCT treatment group was -21.6 (95% CI -24.1, -19.0) and -12.6 (95% CI -15.2, -10.0) in the hybrid control group. For MBA, the estimated bias between control groups remained consistent across models; posterior mean of the bias estimate ranged from -5.4 to -4.0. Covariate-adjusted bias-corrected models had marginal mean in RCT treatment group of -21.7 (95% CI -25.9, -17.5) and -11.4 (95% CI -15.5, -7.2) in hybrid control group. The estimates from both methods were similar to original trial estimates ‒ mean change in CDAI in treated group -21.4 (95% CI -23.9, -18.9) vs. -12.5 (95% CI -15.3, -9.7) in the control group.

Conclusions: Using data from RA registry patients, we obtained well-matched controls and replicated the inference of the baricitinib Phase IIb trial using both hybrid control methods. MBA adjusts for additional bias in the outcome estimate. The estimated bias was consistent across all models. The residual bias not accounted by matching on observables is important to consider for the estimate of treatment effect. This study demonstrated that it is feasible to have accurate estimation of treatment effect using a hybrid control group, when the data collected on patients is comparable to that collected in the trial.