A framework for evaluating markers used to select patient treatment

There is growing interest in markers that can be used to identify which patients are most likely to benefit from a treatment. For example, the Gail breast cancer risk prediction model may be useful for identifying a subset of older women for whom the benefit of tamoxifen for breast cancer prevention is likely to outweigh the harm. Two general classes of approaches to evaluating treatment selection markers have been developed. The first uses data on a cohort of untreated subjects to develop a risk prediction model, such as the Gail model, which is used to identify a high-risk subset of subjects. This model is paired with a measure of treatment effect to assess the impact of identifying and treating the high-risk subset. The second approach uses data from a randomized trial to model the treatment effect on a composite outcome that includes all effects of treatment (positive and negative). The treatment effect model is used to identify a subset of subjects with positive treatment effects and to assess the impact of identifying and treating this subset. We describe a framework that includes both existing approaches as special cases. In doing so, we review the existing approaches, clarify their underlying assumptions, and facilitate the evaluation of markers under less restrictive assumptions.