ENH: Quasi-Likelihood Ratio test for overdispersed GLM, discrete

[josef-pkt]

If we have misspecification in the form of simple under/overdispersion, like in GLM or Poisson model with scale not equal to 1, then the Quasi-LR test can still be used to test competing nested models.

see end of section 7.5.1 in Cameron/Trivedi Microeconometrics, which reference
Wooldridge, J. M. (2002), Econometric Analysis of Cross Section and Panel Data, p 370

for general QMLE the LR statistic is weighted sum of chi-squares distributed
Cameron Trivedi Microeconometrics section 8,5,3
(I have already written the distribution for weighted sum of chi-squares random variables, the problem is getting the case specific weights, which are eigenvalues of some matrix products)