Post by pingye on Jun 2, 2015 4:09:53 GMT
Hi GCTA developers,
I have 2 questions regarding the use of GCTA:
(1) My understanding is that by default the GRM is computed as WW'/N, where W is the standardized genotype design matrix and N is the number of SNPs used? I think if GRM is computed in this way then the estimated h^2 is the proportion of variance explained by the causal variants that are in LD with the genotyped SNPs we have? I also see in your 2010 Nat Genet paper you talked about adjusting the Ajk in order to estimate the total heritability by all causal variants (Multiply the Ajk by the beta). Is GCTA capable of doing that? If so then what options enable to do that? --grm-adj 0.1?
(2) If I pick the top 1,000 SNPs from a GWAS with a quantitative trait , and then use them to estimate the variance explained by the causal variants in LD with those 1,000 SNPs (using the same sample). Often time I see the estimated h^2 is very large. I think in this situation the estimated h^2 is biased upward? My thinking is that the sample already "saw" those 1,000 SNPs in GWAS analysis and "believed" them are in LD with causal SNPs (with relatively large effects). Then in the process of heritability estimation the sample tends to estimate the sigma^2(G) upward and sigma^2(e) downword? Similar to overfitting? Am I on the right track? I would like to hear your comments on this.
Thank you very much for your help and I appreciate your time!!
Best,
Pingye
I have 2 questions regarding the use of GCTA:
(1) My understanding is that by default the GRM is computed as WW'/N, where W is the standardized genotype design matrix and N is the number of SNPs used? I think if GRM is computed in this way then the estimated h^2 is the proportion of variance explained by the causal variants that are in LD with the genotyped SNPs we have? I also see in your 2010 Nat Genet paper you talked about adjusting the Ajk in order to estimate the total heritability by all causal variants (Multiply the Ajk by the beta). Is GCTA capable of doing that? If so then what options enable to do that? --grm-adj 0.1?
(2) If I pick the top 1,000 SNPs from a GWAS with a quantitative trait , and then use them to estimate the variance explained by the causal variants in LD with those 1,000 SNPs (using the same sample). Often time I see the estimated h^2 is very large. I think in this situation the estimated h^2 is biased upward? My thinking is that the sample already "saw" those 1,000 SNPs in GWAS analysis and "believed" them are in LD with causal SNPs (with relatively large effects). Then in the process of heritability estimation the sample tends to estimate the sigma^2(G) upward and sigma^2(e) downword? Similar to overfitting? Am I on the right track? I would like to hear your comments on this.
Thank you very much for your help and I appreciate your time!!
Best,
Pingye
