Identifying and Quantifying Biases in Cohorts through Directional Acyclic Graph Theory

Abstract

Genome wide association studies (GWAS) often use large, repurposed datasets of convenience, such as electronic health records or biobanks. Understanding how bias influences associations within these datasets is important for discerning meaningful results from spurious ones. Bias quantification methods exist but are rarely implemented due to a lack of user-friendly software. To address this need, we developed an, R package, McBias, and an accompanying RShiny web browser app to further improve accessibility. McBias models and quantifies bias using Bayesian networks (BNs) and Monte Carlo simulations, based on user supplied variable distributions and effect sizes. Our rShiny app aids users in writing equations for probabilistic model creation, but has some additional limitations. From the models and simulations, McBias calculates and visualizes total bias that may arise in userdefined analyses. Our first use case demonstrates how genetic variant rs7903146-T affects body mass index and waist circumference differently under different adjustment sets. Our second use case replicates estimates from a published study which quantified the selection bias in the UK Biobank. We discuss potential oversights and data misinterpretations when bias and spurious results are unaccounted for.