Reconciling S-LDSC and LDAK models and functional enrichment estimates

Recent work has highlighted the importance of accounting for linkage disequilibrium (LD)-dependent genetic architectures in analyses of heritability, motivating the development of the baseline-LD model used by stratified LD score regression (S-LDSC) and the LDAK model. Although both models include LD-dependent effects, they produce very different estimates of functional enrichment (with larger estimates using the baseline-LD model), leading to different interpretations of the functional architecture of complex traits. Here, we perform formal model comparisons and empirical analyses to reconcile these findings. First, by performing model comparisons using a likelihood approach, we determined that the baseline-LD model attains likelihoods across 16 UK Biobank traits that are substantially higher than the LDAK model. Second, we determined that S-LDSC using a combined model (unlike methods that use the LDAK or baseline-LD models) produces robust enrichment estimates in simulations under both the LDAK and baseline-LD models, validating the combined model as a gold standard. Third, in analyses of 16 UK Biobank traits, we determined that enrichment estimates obtained by S-LDSC using the combined model were nearly identical to those obtained by S-LDSC using the baseline-LD model (concordance correlation coefficient{rho} c = 0.99), but were larger than those obtained using LDAK ({rho}c = 0.54). Notably, LDAK enrichment estimates were much higher for a non-default version of LDAK that models SNPs in perfect LD differently by assigning non-zero weights to all SNPs. Our results support the use of the baseline-LD model and confirm the existence of functional annotations that are highly enriched for complex trait heritability.