Model Selection

The formal theory of model selection is less thoughtful and less practical than many contributors would have you believe, but there is still some material that one must learn just to be part of the club. As an antipasto, try Zucchini:

Zucchini, Walter, An Introduction to Model Selection, J. Mathematical Psychology 44, 41--61, 2000.

This is well-written introduction for non-statisticians is at a slightly more advanced level than 434. Nevertheless, it is readable, provided that you skip a little --- but not too much. For us, the discussion of the bootstrap is most informative, even though the discussion needs to be modified for time series applications.

The Elephant in the Room

The theory of model selection is limited by its generality. It tries to give advice about the superiority of one model over another without ever asking how one plans to use the model. Moreover, the theory is mute when it comes to advising us when either (1) our model is adequate for the purpose we have in mind, or (2) exhaustive of the data we have at hand.

A genuine theory of model adequacy requires honest attention to the subject matter. Many statisticians find this to be too much work, but for honest scientists, this is where all the fun begins.