Primary Submission Category: Generalizability/Transportability
Multi-Source Conformal Inference Under Distribution Shift
Authors: Larry Han, Yi Liu, Alexander Levis,
Presenting Author: Larry Han*
Conformal inference is a set of methods used to construct distribution-free, nonparametric prediction intervals with finite-sample marginal coverage guarantees. These methods have generally focused on covariate shift while assuming that conditional outcome distributions are invariant across environments. However, conditional outcome invariance is often violated in the real world. In this paper, we consider the problem of obtaining distribution-free prediction intervals for a target population leveraging multiple potentially biased data sources. Our approach is based on the efficient influence functions for the quantiles of unobserved outcomes in the target and source populations, combined with machine learning prediction algorithms to estimate nuisance functions, and a data-adaptive strategy to upweight informative data sources for efficiency gain and downweight non-informative data sources for bias reduction. We highlight the robustness and efficiency of our proposals for a variety of conformal scores and data-generating mechanisms via extensive synthetic experiments. We showcase the benefits of our method for obtaining more informative prediction intervals for pediatric patients undergoing high-risk cardiac surgery using data from 100 congenital heart centers in the United States.