Software Application Profile: McBias: a user-friendly R package and RShiny application for modeling bias in observational studies

Genome-wide association studies (GWAS) have identified hundreds of thousands of single nucleotide polymorphisms (SNPs) associated with complex traits, yet many SNPs do not fully explain trait heritability.1 Thus, in GWAS analyses, biobanks, electronic health records, and other large datasets of convenience are increasingly being used to identify SNPs with smaller effect sizes on traits.2 However, potential biases in these datasets may create spurious effect sizes.3 These spurious findings can hinder the transferability of GWAS results4 and incorrectly identify trait-associated SNPs.5 GWAS summary statistics quantify statistical precision through confidence intervals and standard errors, but they do not quantify bias. Bias remains a consistent challenge, as it can impact the reported results.6 For example, higher educational attainment and a non-smoking status are predictive of UK Biobank (UKB) participation.7,8 Thus, within UKB, non-smokers are reported to have lower educational attainment compared with smokers.7 Another study reported a hazard ratio of either 1.13 or 0.96 for the association between increased hemoglobin A1c and myocardial infarction, depending on whether an outpatient or emergency room population was used, respectively.9 While many studies may be prone to selection bias, scientific reports have infrequently quantified this potential selection bias.6 Additional biases and their implications are visualized by using directed acyclic graphs (DAGs) in Fig. 1.