Change dof_residual definition to not depend on dispersion parameter

[andreasnoack]

Fixes #509

There was actually already a test of a model affected by this but the reference value was wrong. I checked the value in R.

@nalimilan how do you feel about dof after all these years? After preparing this, I think we might be better of from deprecating dof (which would have to happen in StatsAPI). I tend to think that it brings more confusion than convenience.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.85%. Comparing base (e2f6c98) to head (5c26648).
Report is 3 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #560      +/-   ##
==========================================
- Coverage   89.86%   89.85%   -0.01%     
==========================================
  Files           8        8              
  Lines        1125     1124       -1     
==========================================
- Hits         1011     1010       -1     
  Misses        114      114              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nalimilan]

Sorry for the delay. Sounds reasonable to me to include this as part of 2.0 given that it's breaking.

Regarding dof, you mean we should deprecate it because people confuse it with dof_residual?

[andreasnoack]

Regarding dof, you mean we should deprecate it because people confuse it with dof_residual?

Yeah. The confusion seems to show up repeatedly and I think that dof_residual is all you really need.

[nalimilan]

I see. But it can be useful in some cases, notably it's used in the definition of aic, bic and adjr2. In R you have to access private fields of model objects to get that information, which is bad.

Maybe we could put a warning in the docstring? Or we could have a single dof function with an argument to choose whether you want the consumed or residual DOF, but that would be very breaking.