A Tutorial on Improving RCT Power Using Prognostic Score Adjustment for Linear Models

The use of historical data to increase power in clinical trials has been a topic of interest for many years. A recent approach adjusts linearly for a prognostic score. This is supported by asymptotic optimality results using influence functions for asymptotically linear estimators as well as finite sample optimality results. We review plug-in and linear estimators of average treatment effect in randomized clinical trials, sample size determination, and linear adjustment for a prognostic score. Guidelines and recommendations for the implementation of linear adjustment for a prognostic score are given including curation of historical data and construction of a prognostic score based on the historical data. A simulation study is conducted to investigate the performance in finite samples, comparing it to standard procedures such as propensity score matching for RCTs (PSM-RCT) and ANCOVA using simple baseline adjustment. Unlike PSM-RCT, linear adjustment for a prognostic score avoids biased treatment effect estimates and maintains control of type I error probability. The simulation study shows that the method is robust against deviations from method assumptions and poor performance of the prognostic model. A case study demonstrates an increase in prospective power using linear adjustment with a prognostic score in a phase IIIb clinical trial for type 2 diabetes. A final discussion considers limitations of the method for example in regard to subgroup analysis and the existence of already known prognostic baseline covariates.