"Revert" defaults to mirror GLM

[ablaom]

Context

@OkonSamuel Can you please address the question posed in the linked comment above?

[ablaom]

cc @tiemvanderdeure

[OkonSamuel]

(I also just noticed that neither LinearCountRegressor nor LinearBinaryClassifier have a dropcollinear field. I think this might be an oversight?)

It wasn't an oversight. At that time the version of GLM.jl (i.e v1.4.0) only supported the dropcolinear argument when calling GLM.lm function and not GLM.glm function.

@OkonSamuel Do you remember a reason for making the new default false?

There wasn't any reason for overriding the GLM defaults. It wasn't my intent to override the defaults. My aim at that time was to replace the previous deprecated kwarg allowrankdeficient with the new kwarg dropcolinear . Both allowrankdeficient and dropcolinear keyword arguments weren't supported by GLM.glm function. So it was causing errors on something I was working on at that time.
So we can stick to the GLM defaults and mention in the docs that by default the behavior is to drop collinear columns.

Secondly, it appears the position of intercept in a formula matters. StatsModels.@formula will always put intercept as the first term. In this package, we put it at the end in glm_formula. Reversing this so the intercept is the first term will actually also fix the error (I have no clue why, but it does).

@tiemvanderdeure What Error where you referring to here?

[tiemvanderdeure]

@tiemvanderdeure What Error where you referring to here?

It was an error where fit would throw an error when it should have dropped collinear variables, because the position of the intercept term does not correspond to what is expected by StatsModels. But it got fixed by #45, so we don't need to worry about it.

dropcollinear got implemented for generalized models in GLM v1.8.1, but I see no reason why we couldn't require that if we make a breaking version.