When is it okay to not use model selection If I have a model in mind, to ask a very specific question, do I have to do some form of model/variable selection?
There are many papers describing different ways to do model selection, why some are bad, etc. However, I struggle to come across a peer-reviewed paper that discusses when it is okay to not use model selection.
I ask this because I have been asked by a reviewer 'how did I do model selection', and they frame it in a way that suggests that they think it is mandatory. My gut intuition tells me that it is not, but I am wondering if anyone can explain why it is okay not to do it or point me to some text that suggests this.
 A: You always do model selection in some ways. You either use statistical methods like optimising AIC or BIC; or you do what everybode else does; or you use some form of theory (could be ad hoc) that tells you which variables are relevant and what is the functional form you are looking at.
The advantages of the first way is that you don't need a good theory. The data "decides". I've seen this approach mainly for prediction exercises. If you care more on the effect of one variable on the other rather than prediction, careful reasoning might be more powerful. For understanding a relationship between two variables it is more important to understand which other variables could confound this relationship rather than maximising the explanatory power of the model. Even if an independent variable adds little to the explanatory power of the model, it can still be an important confounder.
It seems as if you had used the third way. Whether this is seen as appropriate depends a bit on the field. I've seen this mostly in papers that care about a causal relationship between two or more variables. Whether choosing a model based on theory is appropriate depends also on your confidence in the theory you use to set up the model. If your theory is very good it is likely that you coose the right model. There are also mixtures possible. You could explain to the referee why you chose your model, and run alternative models that are at least somewhat resonable as robustness checks.
