code review: GLM

[josef-pkt]

• why do we calculate pinv_wexog in GLM.initialize? It's not used AFAICS.
  normalized_cov_params is calculated with it, but that's also unused.
• fit "state" attached to the model: several fit results are attached to model (e.g. mu). It's part of the design for the iterative estimation. I don't see a problem, but it would be better to detach those from the model, once they are attached to the results instance.

naming

• family predict versus fitted: not clear which is which. ?
  (difficult because statistics and econometrics differ, link versus inverse-link)

[josef-pkt]

minor: copy offset only if we have to change it:

        if hasattr(self, 'offset'):
            #offset = self.offset.copy()
            offset = self.offset
        if hasattr(self, 'exposure'):
            #offset += self.exposure
            offset = offset + self.exposure

[josef-pkt]

bump for 0.10

I just saw again the first why? I don't see any reason.

in GLM.initialize

        self.pinv_wexog = np.linalg.pinv(self.exog)
        self.normalized_cov_params = np.dot(self.pinv_wexog,
                                            np.transpose(self.pinv_wexog))

And the GLMResults.__init__ attaches pinv_wexog for no reason that I can see

my guess is that these are leftovers from earlier versions of fit.

[josef-pkt]

pinv in initialize has been removed in #4624