Inclusion bias affects common variant discovery and replication in a health-system linked biobank

Electronic health record (EHR)-linked biobanks hold promise for precision medicine by enabling association studies between genetic variants and clinical phenotypes for individual risk assessment. However, most biobanks use opt-in consent protocols to recruit individuals interacting with healthcare systems. This strategy may lead to both participation and recruitment bias, the effects of which on genetic analyses remain understudied. We leverage the UCLA ATLAS Community Health Initiative as a use case to determine possible sources of bias and evaluate their impact on genetic analyses. We find that a wide array of factors are associated with participation, such as receiving primary care at UCLA (odds ratio [OR] = 8.44, p < 1e-300), frequency of healthcare utilization (OR = 1.04, p < 1e-300), and various sociodemographic factors. Together, features recorded in EHRs differentiate biobank participants from the broader healthcare system population (area under the receiver operating characteristic curve [AUROC] = 0.85, area under the precision-recall curve [AUPRC] = 0.82). By weighting the sample using inverse probability weights derived from probabilities of enrollment, we replicate 54% more known genome-wide association study (GWAS) variants than models not accounting for bias (e.g., associations between variants in PPARG and type 2 diabetes). Potential effects of bias were also present in polygenic score phenome-wide association studies (PGS-PheWAS), where across a panel of five PGS with varying genetic architectures, association patterns were affected by the reweighting strategy. Our results highlight that genetic analyses within EHR-linked biobanks may be affected by participation and recruitment bias and that ad hoc analyses within each healthcare system can identify possible sources of confounding.