-
Notifications
You must be signed in to change notification settings - Fork 2.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
SUMM/ENH: two-part models with heteroscedasticity, dispersion modelling #7639
Comments
I'm not sure what the best way to get started.
For simplicity we use full, joint MLE estimation/optimization which means we don't need to use any of the old computational tricks in the |
I think I'll do this next Also as examples to work out design issues location: first candidates: GLM families for continuous endog, gaussian and gamma (and maybe a reason to look at IG) One advantage of weight/precision parameterization is that we can have obs with weight = 0, as in RLM. not at first: dispersion models for discrete data |
relevant here (parking a reference when I remember, I don't have the article right now) Romano, Joseph P., and Michael Wolf. "Resurrecting weighted least squares." Journal of Econometrics 197, no. 1 (2017): 1-19. |
update I had a bug in loglike, using squrt(scale) instead of scale in one term now, without fixing scale=1 in WLS, the scale estimate differs from one only by df-correction, That should make a good unit test.
initial comment belowsomething puzzling: scale estimate in WLS compared to GaussianHet, WLS scale is around 0.5, but should be 1, I thought. As consequence bse are only around half of het MLE. If I fix WLS scale=1, then WLS cov_params agrees with the the inverted negative mean sub-matrix of full MLE hessian. This would correspond to EIM, while OIM differs a bit (hessian/cov_params is not block-diagonal in a smaller sample, nobs=59)
(maybe I still have a bug somewhere ) |
We need also GMMHet, e.g. for quasi-likelihood models with variance functions #1777 example excess dispersion for Binomial counts.
Outside of GLM/LEF (and elliptical, symmetric distributions), mean is not (asymptotically) orthogonal to variance. Even if we don't use it as estimation model, we can use it for specification tests and robustness checks. more general: classes of moment conditions that use (i)link functions. I guess we need weight functions in (core) moment condition and not in "instruments" One limitation of GMM, if we have multiple local minima, then GMM doesn't provide a (parametric) objective function to choose between those local optima. |
another thought:
Then, with exp function (log-link)
I don't have added offsets yet to MultiLinkModel. |
(I didn't find a general issue for this)
issues
Beta regression has been merged. It includes two link functions, one for mean and one for precision (1/variance_scale)
The same pattern can be used for other distributions, e.g. gaussian, ...
This can include discrete models that have a dispersion parameter, e.g. Negbin and generalized Poisson.
possible alternative/complement is to add a subclass of GLM, but GLM implementation does not model scale as MLE.
We have many issues on diagnostics for heteroscedasticity, but not much on modelling heteroscedasticity.
The text was updated successfully, but these errors were encountered: