Unmasking Bias: A Framework for Evaluating Treatment Benefit Predictors Using Observational Studies

Treatment benefit predictors (TBPs) map patient characteristics into an estimate of the treatment benefit tailored to individual patients, which can support optimizing treatment decisions. However, the assessment of their performance might be challenging with the non-random treatment assignment. This study conducts a conceptual analysis, which can be applied to finite-sample studies. We present a framework for evaluating TBPs using observational data from a target population of interest. We then explore the impact of confounding bias on TBP evaluation using measures of discrimination and calibration, which are the moderate calibration and the concentration of the benefit index ($C_b$), respectively. We illustrate that failure to control for confounding can lead to misleading values of performance metrics and establish how the confounding bias propagates to an evaluation bias to quantify the explicit bias for the performance metrics. These findings underscore the necessity of accounting for confounding factors when evaluating TBPs, ensuring more reliable and contextually appropriate treatment decisions.