Abstract

Surrogate markers are widely used in clinical trials across many areas, including diabetes, cardiovascular disease, cancer, and HIV, when primary outcomes are costly, invasive, or slow to observe. In diabetes, for example, glycemic markers such as HbA1c routinely stand in for long-term outcomes like cardiovascular events and mortality. Yet a surrogate that has worked well in one trial population does not necessarily work in another, and diabetes research offers cautionary examples in which treatments improved glycemic control while raising concerns about harm to the outcomes patients care most about. This is the "surrogate paradox," in which a treatment appears beneficial based on the surrogate but is actually harmful with respect to the primary outcome. In this talk, we ask not only whether a surrogate is valid, but for whom it can be trusted. We consider a setting with an existing study (Study A) in which both the surrogate and the primary outcome were measured, and a new study (Study B), possibly in a different patient population, in which only the surrogate is measured. Rather than assuming that the relationship between surrogate and outcome transports directly from Study A to Study B, we consider a class of conditional mean functions for Study B that deviate from those estimated in Study A in controlled ways. Using this class, we estimate the distribution of potential treatment effects on the unmeasured primary outcome in Study B. From that distribution we define three resilience measures. The resilience probability is the probability that the treatment effect on the primary outcome is in the same direction as the effect on the surrogate. The resilience bound is the degree of deviation from Study A that can be tolerated before the surrogate paradox becomes plausible. The resilience set is the collection of deviations under which conclusions based on the surrogate remain valid. Together, these measures complement traditional surrogate validation by quantifying how robust surrogate-based conclusions are when moving to new populations, making them a practical tool for precision health, where treatment decisions increasingly target specific patient groups. We evaluate the proposed measures through simulation and application to clinical trial data, and discuss implications for use in clinical trial research.