P Values In Plink

Question

[viavialin]

I have a simple question about the OR and p values derived from PLINK analysis (with covariates using --logistic and --covar)

Does it make sense that the OR and p values change when I use different number of SNPs for a Fisher's Exact Test?

The OR values are not too different but the p values often become smaller (more significant) when there are lesser SNPs in the analysis?

I can't get my head around it since OR values are supposed to be from 2x2 contingency tables and the p values from Fisher should be exact.

[Philipp Bayer]

I'm slightly confused - are you running a logistic regression using --logistic or Fisher's exact test using --fisher?

(Also, are you using --adjust?)

Answer 1

[viavialin]

I use --logistic with --covar .txt and --covar-name multivariate analysis. I do not use --adjust. Should I be using it?

I am comparing the p values for 400 SNPs vs 70 SNPs with the same commands (logistic + covar) and I get different p values, slightly different ORs for the same SNPs.

I tried doing --fisher with --covar but it does not give me the additive effect of covariate with SNP.

[Philipp Bayer]

On --adjust:

If you use a significance cutoff of 5% then you'd expect 20 SNPs to be significant in your 400 SNP dataset just by chance (400*0.05=20). You can either use a new cutoff of 0.05/400=0.00125, OR you can use --adjust, which gives additionally to your original output files p-values adjusted with a variety of technologies like FDR or Bonferroni (see here). Either are fine to me.

On your question about different p-values:

As far as I know, PLINK can use several SNPs together in one logistic regression's model (have a look at slides 20/21 here. If you remove some of these SNPs, the entire model changes, so that the p-values of the other remaining SNPs change, too.