Credible set is sensitive to imputation quality and missing variants

Bayesian fine-mapping to obtain credible sets has been widely applied post GWAS to pinpoint causal variants. The calculation of credible sets generally assumes that all variants have been equally well genotyped, which is often not the case when a GWAS has been run on imputed data. In this work, we investigate the behavior of credible sets in imputed datasets utilizing 'held out' genotyped variants to measure accuracy. We show, via simulation, that: i) the coverage of credible sets decreases when using imputed variants in GWAS; ii) rare causal variants often fail to be tagged in credible sets when they are not present in the GWAS variant set. We develop a reweighting approach to take imputation quality into account during fine-mapping that only requires summary statistics, and demonstrate the approach with real data.