Controlling FDR in selecting group-level simultaneous signals from multiple data sources with application to the National Covid Collaborative Cohort data

TL;DR

This paper introduces a knockoff-based method for controlling the false discovery rate when identifying group-level signals across multiple data sources, demonstrated with COVID cohort data.

Contribution

It develops a novel FDR-controlling algorithm for union of group-level tests applicable to heterogeneous multi-source data, with theoretical guarantees.

Findings

Exact FDR control achieved in simulations

Effective identification of true signals in N3C data

Method robust to heterogeneity across sources

Abstract

One challenge in exploratory association studies using observational data is that the associations between the predictors and the outcome are potentially weak and rare, and the candidate predictors have complex correlation structures. False discovery rate (FDR) controlling procedures can provide important statistical guarantees for replicability in predictor identification in exploratory research. In the recently established National COVID Collaborative Cohort (N3C), electronic health record (EHR) data on the same set of candidate predictors are independently collected in multiple different sites, offering opportunities to identify true associations by combining information from different sources. This paper presents a general knockoff-based variable selection algorithm to identify associations from unions of group-level conditional independence tests (simultaneous signals) with exact FDR control guarantees under finite sample settings. This algorithm can work with general regression settings, allowing heterogeneity of both the predictors and the outcomes across multiple data sources. We demonstrate the performance of this method with extensive numerical studies and an application to the N3C data.