OK, 2 questions. I have an African American ancestry GWAS meta-analysis and used cov-adjusted LD scores from UKBB AFR to run LDSC. The attenuation ratio was high. Thinking that gnomAD might better fit the ancestry admixture of the cohorts in the meta, then used those, and things improved, with h2 estimation more consistent with other ancestral grouping metas of the same phenotype. However,
now the ratio <0, which I haven’t seen before. Is this a concern?
my understanding is that the gnomAD LD scores are not covariate-adjusted. Is this a concern?
Just walking this out for anyone else who might take a look at this question:
If attenuation ratio is calculated as:
(LDSC intercept – 1)/(meanχ2 – 1)
and it’s the same African American ancestry GWAS used as input (i.e., same denominator for these two sets of LD scores) then the main shift is happening at the level of the intercept, with intercept estimates well above 1 for the cov-adjusted LD scores and below 1 for the gnomAD scores.
It would be good for someone else to weigh in on this as I could imagine there are several reasons that behavior might be expected. One thought is that gnomAD (as you mention) is providing a better fit to the admixed cohorts. For UKB AFR this would mean some SNPs that are higher LD in your meta-analytic sample are being characterized as low LD by the UKB AFR scores, thereby causing the LDSC intercept to drift upward. If the pattern is also that the gnomAD results produce higher heritability estimates then that would be consistent with what the authors write in the original 2015 LDSC paper:
"First, if LD Scores in the reference population are equal to LD Scores in the target population plus mean-zero noise, then the intercept will be biased upward and the slope will be biased downward. "
I’ve seen intercepts below 1 before, so I wouldn’t say the resulting ratio < 0 is something to be too concerned about. Typically the intercept drifts below 1 when genomic control has already been applied to the GWAS data. I also don’t think the original LD scores put out by the LDSC team were covariate-adjusted so I would not say that’s an issue. If you are also getting h2 estimates more consistent with prior literature using gnomAD it sounds like that’s the way to go.