Correction for participation bias in the UK Biobank reveals non-negligible impact on genetic associations and downstream analyses

Abstract

While large-scale volunteer-based studies such as the UK Biobank (UKBB) have become the cornerstone of genetic epidemiology, the study participants are rarely representative of their target population. Here, we aim to evaluate the impact of non-random participation in the UKBB, and to pin down areas of research that are particularly susceptible to biases when using non-representative samples for genome-wide discovery. By comparing 14 harmonized characteristics of the UKBB participants to that of a representative sample, we derived a model for participation probability. We then conducted inverse probability weighted genome-wide association analyses (wGWA) on 19 UKBB traits. Comparing the output obtained from wGWA (Neffective=94,643 - 102,215) to standard GWA analyses (N=263,464 - 283,749), we assessed the impact of participation bias on three estimated quantities, namely 1) genotype-phenotype associations, 2) heritability and genetic correlation estimates and 3) exposure-outcome causal effect estimates obtained from Mendelian Randomization. Participation bias can lead to both overestimation (e.g., cancer, education) and underestimation (e.g., coffee intake, depression/anxiety) of SNP effects. Novel SNPs were identified in wGWA for 12 of the included traits, highlighting SNPs missed as a result of participation bias. While the impact of participation bias on heritability estimates was small (average change in h2: 1.5%, maximum: 5%), substantial distortions were present for genetic correlations (average absolute change in rg: 0.07, maximum: 0.31) and Mendelian Randomization estimates (average absolute change in standardized estimates: 0.04, maximum: 0.15), most markedly for socio-behavioural traits including education, smoking and BMI. Overall, the bias mainly affected the magnitude of effects, rather than direction. In contrast, genome-wide findings for more molecular/physical traits (e.g., LDL, SBP) exhibited less bias as a result of selective participation. Our results highlight that participation bias can distort genomic findings obtained in non-representative samples, and we propose a viable solution to reduce such bias. Moving forward, more efforts ensuring either sample representativeness or correcting for participation bias are paramount, especially when investigating the genetic underpinnings of behaviour, lifestyles and social outcomes.