To quantify the magnitude of the bias that MAF-dependent genetic architectures could introduce, we simulated a frequency-dependent genetic architecture where effect size was uncorrelated with MAF (Online Methods). For most phenotypes, this model should represent a reasonable bound of the genetic architecture. We observed minimal bias: in these simulations, the mean LD Score regression intercept was 0.994 (Supplementary Figure 6, Supplementary Table 7). Nevertheless, there exist extreme genetic architectures where LD Score regression is not effective: for instance if all causal variants are rare (MAF < 1%, which may be an appropriate model for a phenotype under extreme negative selection), then LD Score regression will often generate a negative slope, and the intercept will be exceed the mean χ2 (Supplementary Figure 7).
