Controlling FDR in selecting group-level simultaneous signals from multiple data sources with application to the...

Area of Science:

• * Biostatistics and Bioinformatics
• * Computational Epidemiology
• * Health Data Science

Background:

• * Exploratory association studies with observational data face challenges from weak/rare associations and complex predictor correlations.
• * False Discovery Rate (FDR) control is crucial for replicability in predictor identification.
• * The National COVID Collaborative Cohort (N3C) offers multi-site electronic health record (EHR) data, presenting opportunities and challenges due to data heterogeneity.

Purpose of the Study:

• * To address the challenge of heterogeneous data types for clinical endpoints across multiple data sources.
• * To present a general knockoff-based variable selection algorithm for identifying associations from unions of group-level conditional independence tests.
• * To provide exact False Discovery Rate (FDR) control guarantees in finite sample settings.

Main Methods:

• * Developed a general knockoff-based variable selection algorithm.
• * Utilized unions of group-level conditional independence tests to identify simultaneous signals.
• * Algorithm accommodates general regression settings with heterogeneous predictors and outcomes across sites.

Main Results:

• * Demonstrated the algorithm's performance through extensive numerical studies.
• * Successfully applied the method to analyze data from the National COVID Collaborative Cohort (N3C).
• * Achieved exact FDR control guarantees under finite sample settings.

Conclusions:

• * The proposed algorithm effectively handles heterogeneous data types in multi-site EHR data.
• * It provides a robust framework for identifying true associations in complex observational studies.
• * The method enhances the reliability and replicability of findings from large-scale health data initiatives like N3C.