GWAS : normalization of continuous phenotype across different cohorts

Question

[Nicolas Rosewick]

Hello,

Let say I've three independent cohorts on which I want to perform GWAS on a specific continous phenotype. This phenotype is not normally distributed so the idea is to first normalize it using rank-based inverse normal transformaton ; and then perform linear regression on genotypes (with e.g. plink).

As I've three cohorts, the idea will be to perform a meta-analysis combing the three cohorts results.

However before the meta-analysis, even before the individual gwas on each cohort, I've a question regarding the phenotype normalization step.

Should I :

1. Normalize each cohort independently to each other
2. or Merge the clinical table of the three cohorts and normalize the phenotype with all value together.

The resulting normalized phenotypes using either (1 : invRank) or (2 : invRankGlobal) will have different values as cohort_3 has a slightly lower value on average for the phenotype of interest.

Which option should I chose ?

Thanks