Even when the initial association is unlikely to be a stochastic artifact due to multiple testing, it may still be an artifact due to bias. For common variants, the anticipated effects are modest—for binary traits, odds ratios smaller than 1.5; for continuous traits, percent variance explained less than 0.5%—and very similar in magnitude to the subtle biases that may affect genetic association studies—most notably population stratification bias. For this reason, it is important to see the association in other studies conducted using a similar (but not identical) study base. In principle, careful design and analysis should eliminate or greatly reduce bias due to population stratification in association studies using unrelated individuals [19–21]—and in practice these methods have effectively removed some worrying systematic inflation in association statistics. [22] Family-based designs can provide additional evidence that an observed association is not due to population stratification bias, but these designs are not cost-efficient, and have their own unique sources of bias. For example, non-differential genotyping error can inflate Type I error rates in some family-based analyses, although it does not change the Type I error rate. [23]
