Best implementation of negative binomial association test for GWAS data?

Question

[LauferVA]

I am being asked to implement a negative binomial association test on Immunochip data.

I ran the logistic model on a dichotimization of the response variable, but we did not find any associations at genome-wide levels of significance, so I am now being ask to consider the response variable as a continuous variable and test again using the negative binomial.

Does anyone know which are the best implementations of such a test for genetic data? Currently the files are in Plink format, although I can convert them if necessary.

Thank you.

Answer 1

[brentp]

Negative binomial is for modeling count data. What are your counts?

If you dichotomized the response variable as you say, then logistic regression makes sense. If you want to re-try on the continuous response variable, then just run a linear model.

Is your model like:

response ~ genotype + age + ...

?

[LauferVA]

Yes, it is:

radiographic severity ~ genotype + age + sex + PC1 + ...

So I gather from your answer that you would just go with a standard R package, in other words, there is not a significantly better test that's adapted to genetic data?

Thank you

Answer 2

[jstats]

Not sure if you still want this, but I have R code to perform a neg binomial GWAS. Our data had lots of zero counts, but we felt that going to the full zero inflated models was a step too far.

[LauferVA]

Hi Jstats,

Just seeing this. May I ask you why you thought it was going too far?

Also, I would take the R code for the association testing although I already wrote my own .. but it would be good for comparison.

However, if you can, what I would be more interested in is your procedure in model selection and decision making process in using negative binomial regression instead of ZINB. I don't have access to any code written in R that walks me through that part of it ...

Thanks very much.