Normalization of Ordinal-Categorical data in GWAS

Question

[kakukeshi]

Hi,

I want to perform a GWAS using the number of events (e.g number of epileptic attacks) as a trait. Do I need to perform normalization of this type of data and if so, which method would you recommend me. The data has the following distribution:

Answer 1

[Kevin Blighe]

Hello, not a common question but I've given 2 answers on it previously:

• A: Should Covariates be log transformed in a GWAS study
• A: Log-tranformation and GWAS

In a nutshell (to summarise), you don't have to normalise provided that your statistical test is set-up to expect the distribution that your variable follows. Your epileptic attacks variable looks like a kind of Poisson or negative binomial distribution, in which case your test would have to be aware of this. Otherwise, you could log-normalise your variable such that it would follow the binomial / normal distribution, which is the default distribution expected for most tests.

Kevin

[kakukeshi]

Thank you very much for your answer Kevin. I noticed your previous post about the normalization of quantitative variables, but I think with categorical data the solution is more complex. I think the data follows something like an exponential or geometric distribution. I've tried to log-normalize the data to get something normally distributed unsuccessfully (being the closest something like log(x)+0.001).

I'm quite curious about this. I thought that this was a common problem in GWAS because of the many traits like this but apparently is not the case. By the way, do you know any GWAS study on traits with a similar distribution?

Many thanks

[Kevin Blighe]

Yes, you're correct that it is more complex for categorical data - my previous answers were more for continuous variables. Also, this is most likely not a common theme on Biostars because it's definitively more in the realm of statistics; so, you would likely get a better response on https://stats.stackexchange.com/

I don't recall reading any published manuscript specifically on the topic of the normaliation of categorical variables. I think that information on what people have done would be buried deep in the supplementary material. I have come across a few publications on this general topic, such as: Are your covariates under control? How normalization can re-introduce covariate effects

I'm trying to think in which ways it would affect the statistical models and to find ways around it. I don't know, exactly, if breaking up your data into strata and conducting conditional logistic regression would help. I don't know what other variables you have in your study, though.