internal/design adjustments in discretemod

[wesm]

Original Launchpad bug 642240: https://bugs.launchpad.net/statsmodels/+bug/642240
Reported by: josef-pktd (joep).

a few notes on discrete models, while looking how nested logit fits in

• nparams
  aic, bic, ... calculations need the number of params self.df_model+self.model.J-1,
  I need it also for a default starting values and other things in ru2nmnl
  If we set nparams in the model, then bic, aic don't need to be special cased for MNL. We could use df_model as default if nparams doesn't exist in the model
• in MNL.J and .K are not very informative when reading, nchoices and nexog would be more informative
  in ru2nmnl, I started to use nchoices, I don't think I have any nexog.
• llnull: I don't think this will work in nested logit, and maybe not in clogit. Unfortunaltely the the arguments to model.class.init are more complicated than the generic (endog, exog)
  --> possibility to override in model or in a Results subclass ?
• llnull: do we need the null model for anything else ? currently null model is not attached
• Rsquare: for MNL type models, SAS (proc MDC) has six definitions additional to McFadden, they are all easy to calculate given llnull

extras not checked yet
Confusion/classification matrix ?

(I expect to add more as I work my way through this.)

[josef-pkt]

the basic class structure of discrete_model has been changed PR #108

But I didn't check yet and work again on extension to conditional or nested logit

[josef-pkt]

I just found this by chance.

review and close, or add to a SUMMary issue together with other refactoring ideas