Dealing with confounders

[oharar]

One topic that isn't touched upon is confounding variables. I've use models selection on them in the past, because I don't really care about them, but some might be important and dropping minor confounders should give more power elsewhere. Of course, regularisation does the same thing.

I guess my main point is that confounders are slightly different, so might have to be dealt with, even if you end up with 2 lines saying "treat them like everything else in the model, but interpret them differently".

[BertvanderVeen]

To add to this; this came to mind for me as one of the few scenarios in which information criteria could actually be useful. Of course penalized regression would help with this too, but in cases where n>p and one fits a couple of competing models without penalization, information criteria could be used instead.