Super Learner G-Computation as a Model-Specification Sensitivity Analysis for Population-Adjusted Indirect...

Parametric G-computation can align population-adjusted indirect comparisons (PAICs) with marginal estimands, but estimates may depend on unverifiable outcome-model specification. We evaluated Super Learner G-computation as a flexible outcome-modeling extension for anchored PAICs.

We simulated 18 anchored A vs C and B vs C trial scenarios with index-trial individual patient data and comparator-trial aggregate data. Scenarios varied sample size, population overlap, and correct vs misspecified parametric outcome models. The target was the comparator-trial population; the estimand was the marginal A vs B log-odds ratio. We compared parametric G-computation, Super Learner G-computation, matching-adjusted indirect comparison, conventional mean-plug-in simulated treatment comparison, split-averaged Super Learner G-computation, and targeted maximum likelihood estimation. We used root mean squared error (RMSE) as the primary performance measure and empirical coverage to assess interval calibration.

In the correctly specified large-sample good-overlap scenario, parametric G-computation had the lowest RMSE (0.263). In the nonlinear small-sample poor-overlap scenario, Super Learner G-computation had RMSE 0.513 vs 0.811 for parametric G-computation, 0.901 for conventional simulated treatment comparison, and 1.153 for matching-adjusted indirect comparison. The split-averaged version reduced RMSE to 0.475. Across 9 nonlinear scenarios, flexible plug-in estimators had lower RMSE than parametric G-computation in 7, mainly under poor or moderate overlap, but coverage was lower in all 9. Targeted maximum likelihood estimation improved coverage in some settings but did not consistently reduce RMSE.

Super Learner G-computation may be useful as an outcome-model specification sensitivity analysis for marginal PAIC estimands. Lower RMSE came with lower coverage; therefore, uncertainty intervals require caution.