association analysis with stratification

Question

[alessandrotestori7]

Hello! I want to run a classical genome-wide association analysis (cases vs controls) in order to understand whether particular polymorphisms are relevant in a disease. The problem I have encountered is that cases come from a different population than controls. Is it possible to run an association analysis anyway (with a correction) and get reliable statistics or should I discard a lot of samples? Are there any detailed pipelines or packages I can use? Please let me know. Thanks!

Answer 1

[theobroma22]

You can use the mantelhaen.test function in the stats R package, which conveniently is loaded in the library directory when you installed R. Alternatively, look at the epiR package, although I've never used it.

[alessandrotestori7]

Thanks for your reply, though I'm not sure that is exactly what I need. However, please do let me know if I'm wrong about it. Indeed, I was rather looking for a pipeline or a package which can adjust p-values from an association study in case the controls and the cases have different allele distributions. I think several softwares/tools are available (e.g.: EIGENSTRAT, Price el al 2006, GenABEL?), but I was unable to follow a pipeline or find a recent package.

[theobroma22]

If you need to adjust p-values you can get the q-values and local FDR values, using the qvalue package available on Bioconductor.

[theobroma22]

Now I see your study more clearly, and think the non-population controls were added to the data set as variation, or noise.

[WouterDeCoster]

Interesting, can you elaborate how it deals with population stratification? I considered studies with different control population vs patients dead upon arrival.

[theobroma22]

It would've been correct to consider populations with vastly different polymorphisms as dead upon arrival...but, in statistical modelling you could have also considered 1) having two controls, or 2) contrasting population cases with non-population controls, or 3) stratifying them separately in the contingency table which can help with the analysis like detecting any confounding variables after you obtain the point estimates and chi-squared values.