kc
New Member
Posts: 9
|
Post by kc on Mar 1, 2014 0:05:38 GMT
Hi there,
I got the following output for my analysis on the variance explained by my SNPs on my trait of interest: Source Variance SE V(G) 0.002753 0.120219 V(e) 0.161132 0.114409 Vp 0.163885 0.013478 V(G)/Vp 0.016799 0.732900 The estimate of variance explained on the observed scale is transformed to that on the underlying scale: (Proportion of cases in the sample = 0.204134; User-specified disease prevalence = 0.083000) V(G)/Vp_L 0.025638 1.118523 logL 154.422 logL0 154.422 LRT 0.000 df 1 Pval 0.4912 n 387
I just wanna confirm the interpretation of the finding. Does it mean the heritability of my trait is 2.6% (0.025638) using my list of snps?
Thank you!
|
|
|
Post by Jian Yang on Mar 1, 2014 11:37:59 GMT
The sample size is too small so that the standard error is huge. It means that you don't have any power to detect the genetic variance captured by the SNPs.
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 1, 2014 18:35:00 GMT
I got about 3000 individuals in my dataset. May I know the required/min. required sample size for the heritability estimation? Thanks!
|
|
|
Post by Jian Yang on Mar 2, 2014 5:00:11 GMT
SE is roughly = 300/n so I would expect that you'll get a SE of ~0.1 using 3000 unrelated samples.
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 3, 2014 23:08:06 GMT
Thank you for the information. Now, I got a SE of 1.118523 vs 0.1. Does it mean there's some errors of my analysis? Any hint/help would be greatly appreciated. Thank you!
|
|
|
Post by Zhihong Zhu on Mar 4, 2014 6:11:22 GMT
Did you remove related samples?
|
|
|
Post by Jian Yang on Mar 4, 2014 9:59:44 GMT
Thank you for the information. Now, I got a SE of 1.118523 vs 0.1. Does it mean there's some errors of my analysis? Any hint/help would be greatly appreciated. Thank you! But the result shows that there are only 387 included in the analysis. You may need to have a look at the log output carefully.
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 5, 2014 5:18:23 GMT
I think only 387 individuals retained due to removal of cryptic relatedness. I used a cutoff of 0.025 (--grm-cutoff 0.025). Is it too much? It would be great if I can have some hints about a better cutoff of the relatedness. Thank you!
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 5, 2014 22:51:17 GMT
I also tried another sample so that more sample size retained for the heritability estimation. Here's the result: Source Variance SE V(G) 0.048005 0.012945 V(e) 0.099811 0.011791 Vp 0.147817 0.003784 V(G)/Vp 0.324763 0.084267 The estimate of variance explained on the observed scale is transformed to that on the underlying scale: (Proportion of cases in the sample = 0.176672; User-specified disease prevalence = 0.083000) V(G)/Vp_L 0.553581 0.143639 logL 1644.697 logL0 1635.483 LRT 18.429 df 1 Pval 8.816e-06 n 3498
From this analysis, can I conclude that the heritability of my trait is 55% (with SE of 0.14 and p-value =9e-06) using my snp list?
Thank you!
|
|
|
Post by Jian Yang on Mar 6, 2014 5:59:43 GMT
If there are close relatives in your sample, the estimate can be interpreted as the variance explained by SNPs because the pedigree structure can be picked up by SNPs so that your result will be similar as that from a pedigree analysis. You may have a read of the following paper, which proposes a method of how to use GCTA to estimate variance explained by SNPs in family data. www.plosgenetics.org/article/info%3Adoi%2F10.1371%2Fjournal.pgen.1003520
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 6, 2014 19:50:13 GMT
Thank you for the paper! I will look into it.
I also tried to run the same analysis on a limited no. of SNPs which believe to be important functionally for my trait. There are about 30 of them. However, the resulting heritability is low compared to using the whole set of SNPs in my GWAS (~900K).
Source Variance SE V(G) 0.000086 0.000241 V(e) 0.144014 0.003453 Vp 0.144100 0.003448 V(G)/Vp 0.000594 0.001669 The estimate of variance explained on the observed scale is transformed to that on the underlying scale: (Proportion of cases in the sample = 0.176672; User-specified disease prevalence = 0.083000) V(G)/Vp_L 0.001012 0.002845 logL 1635.552 logL0 1635.483 LRT 0.139 df 1 Pval 0.3549 n 3498
Therefore, can I say that my selected list of snps (~30) may not help explain the variance of my trait significantly. May I know if you think this observation makes sense? Thank you!
|
|
|
Post by Jian Yang on Mar 7, 2014 6:35:21 GMT
Alternative explanation is that it's lack of power because the sample size is too small.
|
|
kc
New Member
Posts: 9
|
Post by kc on Mar 7, 2014 23:33:14 GMT
Just want to clarify one thing: the sample size doesn't only count the no. of individuals, but the no. of snps as well? because in my latest analysis resulting in insignificant results (shown above) involve ~30snps with ~3500 individuals. Thank you!
|
|
|
Post by Jian Yang on Mar 8, 2014 0:39:16 GMT
The sample size only counts the number of individual. However, the SE is a function of sample size and the variance GRM which depends on the number of SNPs used.
|
|