This is probably a silly question but i d like to learn the difference between interpretation of results from these commands in plink:
--logistic --condition snp code
--logistic --interaction snp code
i really cant tell the descriptive difference between checking for snp-snp or snp-covariant association and checking for snp-snp or snp-covariant interaction.
Conditioning on a SNP is done when you have two (or more) SNPs and you wish to ask the question "is the effect of SNP two independent of the effect of SNP one?". The PLINK documentation describes conditioning on SNP: "for two conditioning SNPs, rs1001 and rs1002 say, and also a standard covariate, the model would be"
Y = b0 + b1.ADD + b2.rs1001 + b3.rs1002 + b4.COV1 + e
"If the b1 coefficient for the test SNP is still significant after entering these covariates, this would suggest that it does indeeed have an effect independent of rs1001, rs1002 and the other covariate." Here the test SNP b1 is the one where we'd like to test independence.
Testing for interaction is done when you have a covariate (which may not be a SNP) and a SNP and you wish to ask the question, "if I add an interaction coefficient to a model that already has terms for the covariate and the SNP, is that interaction term significant?" The PLINK documentation illustrates this with the following model:
Y = b0 + b1.ADD + b2.COV1 + b3.COV2 + b4.ADDxCOV1 + b5.ADDxCOV2 + e
Here the test is not "is my SNP independent of the covariate", but instead "if I have both my SNP and the covariate as terms, do they together exert a stronger effect on the phenotype than I would expect to see through the linear addition of their individual effects?"
Conditioning on a SNP is like removing its effect. As an example, let's say you find that SNP1 is strongly associated with your trait. Then you find SNP2, nearby, is also strongly associated. You might then ask whether these are 2 separate signals. By conditioning on SNP1 (i.e., holding SNP1 constant for all subjects), you can examine whether SNP2 still has an effect.
Testing for interactions answers a different question: Does the association between SNP1 and my trait differ by SNP2 genotype? This is a gene-gene interaction. Or you might ask if SNP1 behaves differently among statin drug users versus non-users. This would be testing a gene-environment interaction.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
Excellent answer. I wish I could give a double-up here.
Thank you so much for your explanation
@David
"If the b1 coefficient for the test SNP is still significant after entering these covariates, this would suggest that it does indeeed have an effect independent of rs1001, rs1002 and the other covariate."
"still significant" part is confusing, after applying conditioning my b1 drops from 0.091049 to 0.036405. Is 0.036405 still significant?
The effect attributable to your test SNP (reflected in the size of the coefficient) appears to be getting smaller after conditioning, but the passage I was quoting meant significance as assessed by a P value for the coefficient. A small coefficient could still be statistically signifiant with a large dataset. Whether it's actually meaningful is not a strictly statistical judgement.