|
Abstract:
|
Methods to identify valid surrogate endpoints are increasingly sought after, however strong correlation is not sufficient to ensure a seemingly useful surrogate endpoint will predict a beneficial treatment effect. Principal surrogacy has been proposed as a solution to incorporate potential outcomes under two treatments. We consider previous work that models the joint distribution of four potential outcomes. In this work, we incorporate covariates in the statistical validation process of a surrogate endpoint by allowing the mean structure to depend on covariates. By adjusting for baseline patient characteristics, we determine for which individuals the surrogate may be valid and assess the plausibility of conditional independence assumptions using Bayesian methods. We explore to what extent patient subgroups affect marginal metrics such as the causal effect predictiveness curve.
|
