To account for heteroskedasticity, we weight by 1(1+Nhg2ℓj/M)2, which is the reciprocal of the conditional variance function Var[χj2|ℓj] under our model if we make the additional assumption that per-normalized genotype effect sizes are normally distributed (note that violation of this assumption does not bias the regression, it only increases the standard error. A derivation is provided in the Supplementary Note).
