I am currently observing an interesting phenomenon in my analysis. I have a simple logistic regression model for independent Inds. The model is as follows:
$$\operatorname{logit}(Y) = \beta_0+\beta_1X $$
My data $X$ can be stratified by gender (2 groups), each about equal sample size. I have tried two methods:
- OLS using all data
- OLS for each gender and Fisher-combine p-values
There are lots of the same model run (all SNPs on a Chromosome), and it seems that Method 2 would result in a more conservative p-value, ~ 10 times larger than the p-value from Method 1. And this results is true for all the SNPs (~20) that I am interested in (SNPs with a p-value less than $5 \cdot 10^{-8}$).
I'm not sure if this is true for all, and if there's a rigorous way to show it.
