Extended Data Fig. 9: Consistency of GENESIS predictions at different sample sizes.
%h2GWAS and MGWAS were predicted for 19 traits (Supplementary Table 3) based on N=145k vs. N=460k summary statistics, with target sample size equal to 460k, 2M or 10M. At N=460k, predictions of large-N %h2GWAS were slightly smaller (panels b-c), while predictions of \(M_{{\mathrm{GWAS}}}\) were slightly larger (panel f). This difference could result from a less severe form of the power-dependent bias that is known to affect the point-normal (2-component) model when it is misspecified: as sample size increases, SNPs with smaller effect sizes become detectable, and estimates shift toward a larger number of causal SNPs with smaller effect sizes. (This only occurs when the model is misspecified, with a larger-than-expected number of small-effect SNPs). The 3-component model ameliorates this bias by including a small-effect heritability component even at small sample sizes. However, if this model too is misspecified (for example when there is a mixture of small-, medium- and large-effect SNPs), then it would be affected in the same way as the point-normal model, to a lesser degree. Numerical results are presented in Supplementary Table 3. The same analysis using FMR is presented in Extended Data Fig. 7.
