I'm working with a validation dataset, testing 7 previous SNP association. A few of these SNPs is validated as associated with a binary phenotype of interest.
Now I want to calculate the amount of phenotype variance that is explained by these SNPs.
I would like to calculate the amount of phenotype variance explained by
a) all 7 SNPs in the validation cohort
b) the few SNPs that are significant in the validation cohort only
I'm particularly interested in knowing how to carry out this calculation in "Stata" software.
Secondly, the SNPs also associates with a similar phenotype, however, a phenotype explained by a numeric scale - does this change the method?
I have read the article by Wray et al. Pitfalls of predicting complex traits from SNPs…however, i'm still not sure about how to calculate this heritability.
I hope someone can guide me?
/cheers
http://www.nature.com/nrg/journal/v14/n7/full/nrg3457.html
Your phenotype is binary (a disease, perhaps), so it makes more sense to talk about how much "risk" or "latent risk," rather than "variance," is explained by the 7 SNPs. There exist several "pseudo-R2s" for logistic or probit regression, though you'll probably want to go with Nagelkerke's generalized R2, which has a lot of really nice theoretical properties. In particular, it is equivalent to the ordinal R2 if you happen to be doing OLS regression with a quantitative phenotype. I don't know Stata at all, but if it outputs the loglikelihood of your regression model, you could calculate the Nagelkerke R2 by hand if need be.
I'd imagine that the analysis you'd actually do would be to construct a "genetic score" from the 7 SNPs or the subset that replicate, as the case may be. This would be a new variable, and each participant's genetic score would be the sum of their reference allele counts at each SNP, with each count weighted by the SNP's regression coefficient. You'd then use this new variable, along with the usual covariates, in a logistic/probit regression to predict the phenotype; to compute the Nagelkerke R2, you'll need the loglikelihood from one regression model fit with the genetic score, and from another fit without it.
To weight the SNPs, you could use the regression coefficients calculated in the validation sample, but it might be more interesting to use the coefficients calculated in the discovery sample. Using the coefficients from the validation sample will give you a descriptive statistic for how much risk for the disease is explained by those SNPs in that sample, which will be unrealistically optimistic about how well the phenotype would be predicted using those same weights in a different sample. If you use the coefficients from the discovery sample (which is probably bigger than the validation sample), you would be getting a reasonable estimate of how well those SNPs can predict the phenotype in practical scenarios.
BTW, the genetic scoring I describe dates back to a 2009 paper from the International Schizophrenia Consortium: http://www.nature.com/nature/journal/v460/n7256/abs/nature08185.html
http://www.cesarzamudio.com/uploads/1/7/9/1/17916581/nagelkerke_n.j.d._1991_-_a_note_on_a_general_definition_of_the_coefficient_of_determination.pdf
- Badges
- Science topic
- Similar topics
- Correction