that test statistics will be inflated and the false positive rate will increase. This problem will be dramatically exacerbated if the discovery set is not very large, for example, coming from a single GWAS cohort, with there being a higher chance of observing a substantial correlation due to random sampling and, further, if the selected SNPs fitted in the model explain a large proportion of variance. In our method implementation, we keep the residual variance constant at the same level of the phenotypic variance, even after fitting SNPs that cumulatively explain a substantial proportion of phenotypic variation in the model (Online Methods). Although this approach is conservative, because we know that fitting the 180 known height-associated SNPs in the model reduces the residual variance by ~10% and therefore increases power to detect additional variants, it has the benefit of keeping the type-I error rate at the same level as that in the meta-analysis and thus avoids over-fitting.
