OBJECTIVES: Real-world databases may lack information on important confounders for synthetic control arm analysis (SCA) when evaluating the effectiveness of immunotherapy. We used external adjustment for unmeasured confounders to bias-correct effect estimates comparing the immune checkpoint inhibitor atezolizumab versus docetaxel and compared these to non-bias-adjusted and randomized effect estimates to quantify the impact of unmeasured confounding on SCA results.

METHODS: Initiators of atezolizumab were identified from the OAK and POPLAR trials. Eligible initiators of docetaxel were selected from the Flatiron Health electronic health records database, which lacks complete information on confounders PD-L1 expression and TMB. Inverse probability weighting was used to adjust for a set of measured confounders including demographics, tumour characteristics, and baseline laboratory values. Overall survival (OS) hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using an inverse probability weighted hazards model. A combination of clinical input, published literature, and external datasets were used to estimate the confounding strength for PD-L1 expression (>1%) and TMB (assumed weakly prognostic), which were then used to produce bias-corrected HRs and 95% CIs using expectation maximization-based simulations of individual-level values for these two unmeasured confounders.

RESULTS: Comparing atezolizumab arm from the OAK trial (n=613) to docetaxel (n=164) after adjustment for measured confounders only, the SCA OS HR was 0.71 [95% CI: 0.54-0.93]. After bias-adjustment for PD-L1 positivity and TMB status, the SCA OS HR was 0.76 [95% CI: 0.55-1.08], which was closer to the randomized trial’s OS HR of 0.78 [95% CI: 0.68-0.89] reported by the OAK study. Findings were consistent with the POPLAR trial.

CONCLUSIONS: PD-L1 expression and TMB should be considered in SCA analyses estimating the effectiveness of anti-PD-L1 immunotherapies where they are deemed to be confounders. Bias-adjustment using external information may be useful when important variables are unmeasured or poorly measured in real-world data.