Negative heritability and filtering summary statistics

Question

[gokberk]

Hi all,

I ran partitioned heritability analysis using LDSC and custom annotations and got negative heritability enrichment scores for 3 phenotypes out of 34. We were a bit confused about the negative heritability results and then I was advised to filter my summary statistics to keep only the SNPs which are present in the baseline model which I used to run partitioned heritability.

My first question is, how do you think keeping only the SNPs in my baseline would affect the heritability enrichments? I did a bit of googleing and found out that usually low heritability, sampling errors and low sample size drives negative heritability. How do you think filtering SNPs in my sumstats based on the baseline would change the heritability measure?

Secondly, a technical question: How can I filter my sumstats based on the baseline? Is there a function to use from one of the common tools or should I manually keep the SNPs that fall within genomic intervals in my baseline bed files?

Thanks!

Answer 1

[gokberk]

For the record - I found out that my issue was too short annotations. I found out that your annotations must encompass at least 1% of the whole genome (1% of all SNPs in the 1KG phase 3 reference panel is around ~20k SNPs) in an old post at the LDSC google group.

However, based on my experience, you need at least around 90k-100k SNPs in your annotation for LDSC to work as intended and to minimize negative heritability estimates.

Hope this is helpful for others!

[Tina]

hi gokberk,

I ran partitioned heritability analysis using LDSC with the baseline annotation together with my own annotation. however I found that some of the baseline annotation got a negative heritability and a very substantial negative enrichment (though my own annotation got a normal positive enrichment and heritability), do you know why? Since there are negative heritability and enrichments, I am no sure about the enrichment I get for my own annotation. I have about 592k SNPs in my annotation.

[gokberk]

Hi Tina, I think getting negative heritability enrichments is plausible, as long as enrichment_p is not significant. If you see significant negative heritability enrichments, then I suppose there might be a problem with your model specification. See here for a useful discussion on negative heritability estimates: https://groups.google.com/g/ldsc_users/c/Ne3ifGyjjbg