Gwas non normally distributed-continous or categorical trait

Question

[dp0b]

Hi,

I have a trait that was measured using a assay but a large proportion of the samples where below the threshold for detection of the assay so my phenotype isnt normally distributed and transformation (log, sqrt, box-cox) isn't successful. Is it better then to treat the data as continous for gwas or categorical high-low. The only thing is with high-low, there would be individuals with phenotype values close to the distinction between the two. Advise would be much appreciated. The distribution is below

https://ibb.co/nPTH0w

I was hoping to use gcta software for the gwas.

Thanks

[Kevin Blighe]

Can you show the distribution here? - you can share images by uploading and obtaining a URL from here: https://imgbb.com/

How are you aiming to conduct the analysis - PLINK or SAS or R or ... ? If the distribution follows a Poisson or Inverse Gaussian, you would be able to select that in SAS or R. Cannot confirm if it's possible with PLINK.

[dp0b]

Thanks for getting back to me. I have edited the question and shown the distribution as you suggested. I was hoping to do the gwas using gcta but can use SAS to transform the data.

[Kevin Blighe]

Thanks for that. So, it looks like an Inverse Gaussian. You're using the genotypes to predict the phenotype, I imagine?

I'm not a SAS programmer (more R), so, don't know the exact way to code for Inverse Gaussian but I'm sure that it's not difficult.

In R, it would be something like:

glm(Phenotype ~ genotype, data=MyData, family=inverse.gaussian())

I've put parallelised code for running these models in R on my GitHub page, if you wanted to try that out. Go to my profile here on Biostars and get the link. It's also here: R functions edited for parallel processing

Kevin

[dp0b]

Thanks for you help,Ill have a look but preference is a mixed model that fits the grm