Evaluating prognostic score adjustment and information borrowing in randomized trials with binary endpoints

Randomized controlled trials (RCTs) with binary endpoints often face limitations in statistical power due to small sample sizes. Two strategies to improve efficiency are covariate adjustment using prognostic scores and borrowing information from historical control data. This simulation study compares the performance of these approaches under varying trial conditions.

We conducted simulations calibrated to parameters from three ulcerative colitis trials (QUASAR, UNIFI, VEGA). We assessed eight statistical methods: Cochran-Mantel-Haenszel (CMH); covariate-adjusted logistic regression model (Logit); prognostic score-adjusted linear probability model (PSA-LPM); prognostic score-adjusted G-computation (PSA-Gcomp); Two-Step frequentist dynamic borrowing; and three Bayesian borrowing approaches-normalized power prior (Bayes-NPP), empirical power prior (Bayes-EP), and robust meta-analytic predictive prior (Bayes-MAP). We varied treatment effect size; control response rate; prognostic score strength; and between-trial heterogeneity. The estimand was the marginal risk difference under a treatment-policy strategy.

Logit, PSA-Gcomp, and PSA-LPM consistently improved power relative to CMH while generally maintaining appropriate type I error. Their performance was broadly comparable across the evaluated scenarios. Bayesian borrowing methods provided substantial power gains under low heterogeneity but showed increased bias and inflated type I error as heterogeneity rose. Bayes-NPP and Two-Step were highly sensitive to discordance between current and historical data, reducing reliability under moderate-to-high heterogeneity. Bayes-MAP demonstrated greater robustness to these changes.

Prognostic score adjustment provides a reliable and efficient alternative to historical borrowing, particularly when between-trial heterogeneity is present, that performs as well as other covariate adjustment methods. Bayesian borrowing methods can deliver large power gains when historical and current data are well-aligned, but require careful evaluation of prior compatibility and heterogeneity to preserve validity.