GWAS on traits non-normally distributed

Question

[AndrMod]

I'm about to run several GWAS on a set of traits measured on ~160 individuals. I want to use FarmCPU which, as I understand, implements a MLM with pseudo-QTN as coviariates (fixed model); hence it assumes normality of the trait. While analysing the traits I noticed that none of them has a normal distribution. The prevalent shapes are these:

enter image description here

I also assessed the non-normality with a Shapiro-WIlk test. Some of these traits are residuals of a lm(trait ~ covariate), but not further quantile-normalised. Other are parameter of a model predicting the trait values. I'm not very convinced I should transform them more.

I'm aware that most of the trait don't follow a normal distribution; nonetheless, this is concerningme a lot, as -without a reference distribution- I also can't pick an approach to detect phenotypic outliers.

The first histogram can roughly look like a Poisson but "roughly doesn't make it a Poisson proper.

Someone can shed a light on how to deal with them?

Answer 1

[Kevin Blighe]

Hi,

Please be aware that, for a statistical model, it is not the actual values that need to follow a normal distribution; rather, it is the residuals after you have fit the model.

Log-transformation, scaling between certain values (e.g. 0-1), encoding as tertiles or quartiles, et cetera - these are all valid approaches.

Kevin.