Exposing flaws in S-LDSC; reply to Gazal et al.

Abstract

In our recent publication,1 we examined the two heritability models most widely used when estimating SNP heritability: the GCTA Model, which is used by the software GCTA2 and upon which LD Score regression (LDSC) is based,3 and the LDAK Model, which is used by our software LDAK.4 First we demonstrated the importance of choosing an appropriate heritability model, by showing that estimates of SNP heritability can be highly sensitive to which model is assumed. Then we empirically tested the GCTA and LDAK Models on GWAS data for a wide variety of complex traits. We found that the LDAK Model fits real data both significantly and substantially better than the GCTA Model, indicating that LDAK estimates more accurately describe the genetic architecture of complex traits than those from GCTA or LDSC.\n\nSome of our most striking results were our revised estimates of functional enrichments (the heritability enrichments of SNP categories defined by functional annotations). In general, estimates from LDAK were substantially more modest than previous estimates based on the GCTA Model. For example, we estimated that DNase I hypersensitive sites (DHS) were 1.4-fold (SD 0.1) enriched, whereas a study using GCTA had found they were 5.1-fold (SD 0.5) enriched,5 and we estimated that conserved SNPs were 1.3-fold (SD 0.3) enriched, whereas a study using S-LDSC (stratified LDSC) had found they were 13.3-fold (SD 1.5) enriched.6\n\nIn their correspondence, Gazal et al. dispute our findings. They assert that the heritability model assumed by LDSC is more realistic than the LDAK Model, and that estimates of enrichment from S-LDSC7 are more accurate than those from LDAK. Here, we explain why their justification for preferring the model used by LDSC is incorrect, and provide a simple demonstration that S-LDSC produces unreliable estimates of enrichment.