Does the response need to be stored as an FeMat?

[dmbates]

Currently the response is stored as an FeMat in the allterms field of a LinearMixedModel. That was done for convenience back in the days when the random-effects terms were in a separate reterms field from the feterms. By making the response look like a fixed-effects term the feterms had a concrete type of Vector{FeMat{T}}.

This approach has the disadvantage that a model cannot be constructed from a response that is initialized to zeros. There is a check on the rank of an FeMat during its construction and a column of zeros has rank zero. The reason for such a model is for simulations. The simulate! function overwrites the values of the response vector.

It may make more sense make the type of the allterms field Vector{Union{ReMat{T},FeMat{T},Vector{T}}} and skip that check on the rank.

[palday]

I think changing the response to Vector{T} would also make some part of manipulating allterms or more precisely extracting the response, reterms and feterms from allterms easier. You could even use multiple dispatch and avoid conditionals for some things (as the implementation of isnothing does).

Should we warn when constructing a model with constant response? Or perhaps warn when fitting a model with constant response? I'm a strong believer in garbage-in, garbage out, but this type of basic error checking might still be useful.

[palday]

With #253, you can now handle a rank-zero FeMat, so the the disadvantage goes away. That said, I think it might still make sense to store the response as a simple vector because that ultimately what it is right now. But even though this change isn't hard, it isn't free. I can also think of a few potential extensions where having an FeMat as the response would be useful (true multivariate regression; fitting of multiple univariate models in parallel).

[palday]

I think the conclusion we came to was "this is too much effort and FeMat now works for zero-rank matrices", right?

[kliegl]

Multivariate extension of LMM was one of Harald Baayen's goals for the Cooperation Group. Linked linear mixed models also come to mind. Just saying ...

[dmbates]

Closed by #464