of causal variants84; a limitation of this approach is that SNPs are assumed to be uncorrelated, which can be approximately achieved by LD-pruning but precludes analyses of dense marker panels. The heritability explained by SNPs can alternatively be estimated from the slope of LD score regression85, in which χ2 statistics for each SNP are regressed against LD scores (sum of squared correlations with all SNPs), leveraging the fact that SNPs with higher LD scores are expected to contain more polygenic signal86. This approach explicitly allows for LD between SNPs and can distinguish between polygenicity and confounding, although it assumes a linear model that may not hold in practice and makes strong assumptions about effect sizes of rare variants that only enable robust estimates for common variants. Another recent method models LD while treating SNP effects as fixed rather than random (similar to ref.84), enabling estimation of heritability explained by common SNPs in local regions as well as genome-wide10. Overall, summary statistic based methods provide a useful alternative to methods for estimating heritability explained by SNPs from individual-level data using restricted maximum likelihood (REML) and its extensions87,88.
