Significant negative heritability in LDSC

I wondered whether someone could help me understand negative (and significant) h2snp estimates (estimated via either LDSC or stratified LDSC). And if there is some good reference on how to deal with them.

Many thanks!

Hi Giaco,

Great question! Raymond Walters previously posted an answer to this on the LDSC github:

Negative h2SNP estimates are much more common with stratified LDSC due to the greater sampling variation observed for heritability estimates stratified across the functional annotations. I would say if you are seeing a sizable negative h2SNP estimate for LDSC (e.g., < -.3) then it would also be good to look at a manhattan and/or QQ-plot of the hapmap3 SNPs used for LDSC estimation to make sure nothing stands out as potentially off, such as lots of floating points in the manhattan.


Hi Andrew,
I was wondering how can this negative heritability be statistically significant?