Q: So, he wants to do shrinkage and set how much to shrink, so he splits the data in half to see what works. I always thought that was a little bit confusing. You have all of this Bayesian machinery with shrinkage in mind. Then they don't know how to set the prior so they try it and see what works. In principle you're taking advantage of having a prior. You should think about how people or stores are different--whether each part differs or the whole thing. You have some effects in the model and you are not using it. You've got something in your model--the data, the situations, and it's not clear to me that you are using it.

A:

Yes, it's an inefficient way to do it.

Q:

It always seemed to confusing to me.

A:

You're right. I would agree, but the problem is that if I want to set some kind of prior on k, then I can set some kind of noninformative on k -- set some kind of uniform. It's not that I have no ideas -- the problem is that before going into this, I really did think that demographics were important and I really did want to have a small k. The problem is if I just set that without regard to what happens in terms of sensitivity, unless I do something like this, people are very critical about why I chose k. It's just arbitrary. So I'm thinking about some better ways to set this prior here. It'd be good to try to have a debate to try to infer more about it. I really don't want to put this informative prior but I don't really want to set where the k boundaries are. So, yeah, I think you're right, that this is sort of a strange way to justify what prior you should use, but ..

Back