Heritability Estimation

note

Important caveats with our heritability estimation approach can be found here.

note

Due to high computational cost associated with running genotype-based heritability estimation on the EUR ancestry group, we only provide genotype-based heritability estimates for a pilot set of traits in EUR, with results available for all non-EUR ancestry-trait pairs. Summary statistic methods were run for all ancestry-trait pairs.

We have used several methods to estimate heritability across up to 6 ancestry groups and 7,228 traits, totaling over 16,000 ancestry-trait pairs. All results can be found in the phenotype manifest and heritability manifest. The following approaches were used:

• Univariate LD score regression (LDSC; ldsc), run using LD score flat files using high-quality HapMap3 SNPs with MAF $\geq$ 0.01 with summary statistics exported from the results table. See Bulik-Sullivan et al. 2015 Nat Gen for more information on the method.
• Stratified LD score regression (S-LDSC; sldsc_25bin), run using the same summary statistcs as ldsc with LD scores generated from SNPs in 5 MAF and 5 LD score bins. See Finucane et al. 2015 Nat Gen for more information on the method.
• Randomized Haseman-Elston (RHEmc; rhemc_8bin) using genotype data with 2 MAF and 4 LD score bins using default settings. This was predominantly used to analyze non-EUR ancestry groups but includes a small set of estimates for traits in EUR. See Pazokitoroudi et al. 2020 Nat Comm for more information on the method.
• Randomized Haseman-Elston (rhemc_25bin) using genotype data with 5 MAF and 5 LD score bins using default settings. Only run for non-EUR ancestry groups.
• Randomized Haseman-Elston (rhemc_25bin_50rv) using genotype data with 5 MAF and 5 LD score bins using 50 random variables to reduce run-to-run variability at a slightly higher computational cost. Only run for non-EUR ancestry groups.

Multi-component analyses were performed using either a 5x5 or 2x4 grid of MAF and LD score bins. MAF bins were either (0, 0.1), (0.1, 0.2), (0.2, 0.3), (0.3, 0.4), (0.4, 0.5) or (0.0, 0.05), (0.05, 0.5); LD score bins were computed as quantiles of the LD score distribution across all SNPs. LD scores used in creation of SNP bins were exported for all SNPs with MAC > 20; LD scores used for LD score regression were filtered after generation as described here.

All randomized Haseman-Elston runs were performed using the same fixed effects covariates used in the main GWAS; namely the first 20 genotype PCs, $age$, $sex$, $age*sex$, $age^2$, and $age^2*sex$. In the case of sex-specific phenotypes, only the first 20 genotype PCs, $age$, and $age^2$ were included.

Genotype and sample filtering was performed for genotype-based heritability estimation. We filtered to SNPs outside the MHC region that were defined with MAF $\geq 0.01$ for which we did not observe significant deviation from Hardy-Weinberg equilibrium. Importantly, we filtered to SNPs that passed the above criteria for all ancestry groups, ensuring the same set of SNPs for heritability estimation across all ancestry groups. We included only unrelated samples for genotype-based heritability estimation.

We note that our S-LDSC and LDSC-based estimates using summary statistics for individuals of EUR ancestry were highly concordant with prior estimates in UKB for overlapping phenotypes, however we observed very poor power for detection of $h^2 > 0$ for non-EUR ancestry groups. Genotype-based Haseman-Elston regression at scale (RHEmc) showed improved power for heritability dection in non-EUR ancestry groups and showed good concordance with S-LDSC in EUR, as discussed in our post.

We used these heritability estimates, along with other important summary statistics, to build a comprehensive QC pipeline.