SUMM: frequency weights, contingency tables, exog with discrete support

[josef-pkt]

objective: better support for datasets that only include categorical exog

motivation: examples like in this http://data.princeton.edu/wws509/stata/c6s5.html have only categorical exog where the data is given in the form of contingency tables.

This would be relatively easy to support if the models had frequency weights fweights.
Related support for the data is mostly available in pandas.

related issues:

...

[josef-pkt]

I'm thinking of using a new model, e.g. OrderedModel or OrdinalModel, for trying out the design for this. with weights and fweights and pandas Categorical support for endog.

[josef-pkt]

implementation idea for fweights:

use new attribute or property wnobs so we don't collide with current nobs.
wnobs could be a property of the model that returns either _wnobs if available or nobs, otherwise we would have to check before usage whether wnobs is available in a model.
The property could be added to LikelihoodModel so it will be inherited automatically.

todo: we still have to check usage of nobs versus len(endog) or exog.shape[0]. I think there is an open issue.