variable. An example of the situation in Figure 3 (panel B) is when the phenotype increases mortality32–35–for example, in studies of smoking as a phenotype, where smoking is associated with premature mortality. In a cohort study which examines smoking, and then follows participants up for Alzheimer’s disease, those who die early (perhaps because of smoking-related illness) will never have the chance to be diagnosed with Alzheimer’s disease, and therefore smoking will appear to be a protective factor. Figure 3 (panels C to E) also shows examples where selection will bias the estimation of the causal effects of SNPs on the outcome. In these examples, SNPs that do cause the outcome directly via the phenotype will show either increased or decreased association in the selected sample, depending on the underlying genetic and environmental aetiology of both traits. In the situations depicted in Figure 3A, C and E, the association between phenotype and outcome (e.g. in an observational study) would also be biased. In contrast, Figure 3F shows a situation where selection will bias the association of the phenotype with the outcome, but the association of the SNP with the outcome will be unbiased. Other, more complex, situations can also lead to
