Distributionally Robust Trial-Level Surrogacy Models

Develop models with weaker distributional assumptions than bivariate normality that estimate conditional sign risk, correct for measurement error, detect drift, and provide asymptotic guarantees without relying on bivariate normality.

Background

The paper’s shipping gate is derived under a bivariate normal working model for intervention-level eval and online effects, together with additive measurement error. Under that model, the conditional sign risk needed to decide whether to ship or kill can be computed from estimated means and covariances. The authors explicitly identify the development of less restrictive models as the most important unresolved problem.

A successful extension would need to handle several issues simultaneously: estimating the conditional probability of a harmful online effect, correcting measurement error in historical intervention effects, detecting post-deployment drift, and establishing asymptotic guarantees that remain valid without assuming bivariate normality. This problem is included because the paper states that the current moment-based gate does not generally determine the relevant conditional tails outside the working model.