MAINT: simplify discrete calls to get_robustcov_results

[jbrockmendel]

get_robustcov_results is called in LikelihoodModelResults.__init__. Avoid re-calling it to make things less-stateful.

[coveralls]

Coverage decreased (-0.002%) to 84.2% when pulling f32994b on jbrockmendel:discrete_calls into a764908 on statsmodels:master.

[coveralls]

Coverage increased (+0.001%) to 88.395% when pulling 8dda7c7 on jbrockmendel:discrete_calls into 52b142c on statsmodels:master.

[codecov]

Codecov Report

Merging #5234 into master will decrease coverage by 1.98%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #5234      +/-   ##
==========================================
- Coverage   86.07%   84.09%   -1.99%     
==========================================
  Files         698      613      -85     
  Lines      118309    97619   -20690     
  Branches    13131    10657    -2474     
==========================================
- Hits       101840    82093   -19747     
+ Misses      13707    13118     -589     
+ Partials     2762     2408     -354     

Impacted Files	Coverage Δ
statsmodels/discrete/discrete_model.py	90.64% <100.00%> (+2.34%)	⬆️
statsmodels/discrete/tests/test_discrete.py	99.85% <100.00%> (+<0.01%)	⬆️
statsmodels/stats/descriptivestats.py	23.68% <0.00%> (-63.45%)	⬇️
statsmodels/stats/base.py	52.94% <0.00%> (-14.41%)	⬇️
statsmodels/nonparametric/tests/test_kernels.py	89.13% <0.00%> (-10.87%)	⬇️
statsmodels/tools/docstring.py	71.12% <0.00%> (-8.30%)	⬇️
statsmodels/graphics/gofplots.py	83.95% <0.00%> (-7.64%)	⬇️
statsmodels/stats/tests/test_diagnostic.py	92.65% <0.00%> (-7.19%)	⬇️
statsmodels/stats/_lilliefors.py	91.13% <0.00%> (-6.23%)	⬇️
statsmodels/graphics/correlation.py	84.84% <0.00%> (-6.07%)	⬇️
... and 376 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 52b142c...8dda7c7. Read the comment docs.

[jbrockmendel]

All four of the use_t = True, cov_type='nonrobust' cases fail under master. @josef-pkt is that intentional?

[jbrockmendel]

@josef-pkt insights here?

[jbrockmendel]

@bashtage any idea if the existing behavior is correct? If so, what am I missing?

[jbrockmendel]

@josef-pkt when you get a chance this is a specifically-your-attention-needing thing. Is the existing behavior intentional?

[josef-pkt]

This might be a bug as in #5331

[jbrockmendel]

This might be a bug as in #5331

Neat. Before I dig into this, is the simplification+unification in this PR something you're open to?

[josef-pkt]

The general idea looks good.
What I don't know is whether we want to have the cov_type keywords twice, once in the super call and once when creating the model specific results instance.
My "impression" is that it might be enough to use it once or that it should be used only once, but I didn't go through those details to see how this works.
I don't remember whether there was something tricky in the computation, but this needs to be careful because the extra dispersion parameter in GP and NB is transformed during optimization but we use the untransformed parameterization for the results. This means that the super(..).fit call is not using the "results parameterization" and cannot compute the appropriate cov_params.

There are some tricky parts to the internal temporary transformation, but I don't remember at the moment what the status is. e.g. #3747

[jbrockmendel]

@bashtage not-just-cln PRs this (#5234), #4936, #5205, #5241, #5785, #5678, #5264. With the exception of the first and last, these are all about code de-duplication.

[josef-pkt]

I'm pretty sure this is either incorrect or fragile because of the extra dispersion parameter

[josef-pkt]

and I guess it computes the sandwich cov_params twice, the first time incorrectly