Quantifying the extent to which index event biases influence large genetic association studies

Abstract

As genetic association studies increase in size to 100,000s of individuals, subtle biases may influence conclusions. One possible bias is "index event bias" (IEB), also called "collider bias", caused by the stratification by, or enrichment for, disease status when testing associations between gene variants and a disease-associated trait. We first provided a statistical framework for quantifying IEB then identified real examples of IEB in a range of study and analytical designs. We observed evidence of biased associations for some disease alleles and genetic risk scores, even in population-based studies. For example, a genetic risk score consisting of type 2 diabetes variants was associated with lower BMI in 113,203 type 2 diabetes controls from the population based UK Biobank study (-0.010 SDs BMI per allele, P=5E-4), entirely driven by IEB. Three of 11 individual type 2 diabetes risk alleles, and 10 of 25 hypertension alleles were associated with lower BMI at p<0.05 in UK Biobank when analyzing disease free individuals only, of which six hypertension alleles remained associated at p<0.05 after correction for IEB. Our analysis suggested that the associations between CCND2 and TCF7L2 diabetes risk alleles and BMI could (at least partially) be explained by IEB. Variants remaining associated after correction may be pleiotropic and include those in CYP17A1 (allele associated with hypertension risk and lower BMI). In conclusion, IEB may result in false positive or negative associations in very large studies stratified or strongly enriched for/against disease cases.

Adjustment for index event bias in genome-wide association studies of subsequent events

“…Existing guidance for dealing with index event bias includes evaluating the associations of SNPs with the trait under selection and with potential confounders 2 ; associations with confounders may be adjusted for directly or via propensity scores 14 . In some circumstances, modelling the selection process may mitigate the bias 17 . However, these approaches cannot entirely account for bias when there are unknown confounders of incidence and prognosis.…”

“…We have not considered the case in which the trait of interest is a precursor of the trait under selection ( figure 2). Selection bias also occurs in this case 10,17,20 but cannot immediately be corrected by our approach because it would require knowledge of the effects of all confounders (Supplementary Text 2). Methods exist to adjust for selection in this situation 18,19,21,22 , although they do not allow for unmeasured confounders.…”

“…Previously, expressions for index event bias have been derived when selecting on a binary disease trait 13,16,17 and when adjusting for a heritable covariate 11 . These studies, however, have not identified methods for correcting this bias.…”

“…On the other hand, some authors have considered the bias arising when analysing a risk factor for the trait under selection [17][18][19][20] . For example, accepting BMI as a cause of type-2 diabetes, a SNP with a direct effect on type-2 diabetes risk may have a biased association with BMI when studied within a case/control sample of type-2 diabetes (figure 2).…”

“…This situation is different from the one considered here as the trait of interest is a precursor of, and not subsequent to, the index event. This situation has also been described as an index event bias 17 , but here we will reserve the term strictly for the situation in which the trait of interest is subsequent to the selection criterion (figure 1). Several methods are available to adjust analyses of risk factors for selection into case/control studies 18,19,21,22 , but they do not apply to the situation considered here, in which selection bias acts entirely through unobserved confounders.…”