Several other studies have reported pseudo-R 2 from the regression of disease on the polygenic score [3], [6], [9]. Although prediction was not emphasised by those studies, they may still be evaluated for that purpose. Recently, Lee et al have argued that, for genetic predictors, R 2 on the liability scale is a more interpretable measure of accuracy for binary traits [25]. In Table 3, liability R 2 derived from those reports are compared to analytic values assuming different levels of heritability explained by the markers. The choice of has little bearing on these results so it is set to 0 throughout. The reported values are consistent with the markers explaining around half the heritability, with variation above and below. This is in line with the estimates of explained variance that were reported by those studies, and those estimates also agree well with those obtained using the method proposed here (equation 6). The low reported values of R 2 do not directly reflect the degree of missing heritability; rather they reflect the effect of sampling variation on the variance explained
