[Joao Santos Silva]

Hello everybody,

If indeed glm does not converge because of perfect predictors, it should be possible to solve the problem by using an adapted version of the method described in
Santos Silva and Tenreyro (2010), On the Existence of the Maximum Likelihood Estimates in Poisson Regression, Economics Letters, 107(2), pp. 310-312.

Assuming that the problem is only caused by perfect predictors of zeros, exactly the same steps can be used:
1 - run a simple ols regression of Y on X using only the observations with 0 < Y.
2 - run the glm model including only the regressors that were not excluded due to perfect collinearity in the ols regression
3 - if any of the regressors excluded are dummies, the observations for which they are equal to 1 should be excluded in the glm regression

[Jaya Jumrani]

Hello everyone,

First and foremost, I'm extremely sorry for the delayed response and thanks a lot for each one of yours' replies on my post.

However, the problem still persists and I've always been using the glm command for my fractional response model. It was only that as a check for perfect predictions that I had estimated the model using both the logit and glm command. I didn't get any such warning messages of perfect prediction via the glm command and the model did not converge. Further, when I ran the normal logit command I did get the error message of some of the independent variables perfectly predicting the data (as indicated above).

On further reading, I came across a paper titled “Georg Heinze and Michael Schemper, A solution to the problem of separation in logistic regression, Statistics in Medicine, 2002, vol. 21 2409-2419” who have suggested a user written Stata command firthlogit (it fits logistic models by penalized maximum likelihood regression). However, when I attempted to use this command for my model, I get the error r(503).

Best,
Jaya

[Joao Santos Silva]

Dear Jaya,

I think you do not want to use -firthlogit- exactly for the same reason you do not want to use -logit-; as far as I know, it treats all positive observations as being equal to 1. Did you try the method I suggested above?

All the best,

Joao

[Jaya Jumrani]

Dear Joao,

I did try the above method as suggested by you. But it leads to the elimination of a lot of key variables (both continuous and dummy) and ultimately glm leads to an error message of no observations when I exclude the collinear variables and dummies' observations which are equal to 1.

Best,
Jaya

[Jeff Wooldridge]

Sorry I confused the issue. What I would do is simplify the functional form of the model to see if that will converge, and then add terms. So start with the covariates (continuous and discrete) just appearing in level form. Then perhaps add quadratics of the continuous variables. Then add some interactions -- the ones you think are most important. Remember, because of the nonlinear in the fractional response model, interactions are built in. Same with diminishing or increasing effects. A simpler model may be sufficient for your purposes. You can see if the simpler fractional logit works better than a more complicated linear model, as in Papke and Wooldridge (1996), by computing an R-squared.

[HANA]

How can I do moderated mediation in Stata?

[Maarten Buis]

HANA:

1) Don't use other threads to ask your question. Instead start your own thread. If you go to the General forum, you can see at the top left a button "+ New Topic". Click on that to start your own thread:



2) Don't shout. If you use all bold letters you leave the impression of shouting; just use normal font to ask your question.

3) If you want a reasonable answer you will need to add a lot more details. There are various tips and tricks on how to ask succesful question in section 9 of the Statalist FAQ. You can find a tab to the FAQ at the top of the page:


4) Posting under our real name has a long tradition on this list. We believe that it has contributed to a friendly and professional atmosphere on this list, so we ask you respect that and change your login name to your real name. You can do so by using the "contact us" button at the bottom of every page:

[Jaya Jumrani]

It's not a problem at all Dr. Wooldridge and in fact I'm nowadays in the process of trying all sorts of permutations and combinations to simplify the model and make it work..

[Jaya Jumrani]

This comment may be uninformed, but I wonder if the problem is due to perfectly predicting some zero outcomes based on certain combinations of the covariates. For binary responses, logit, probit, and glm handle this problem and drop the relevant observations. I don't know whether glm has a way of recognizing perfect predictions with fractional responses. I haven't encountered this problem in the examples I've done.

Hello everybody,

As was first suggested by Dr. Wooldridge and then confirmed in an email conversation with Dr. Silva, the problem in the model is indeed that of perfect predictors and glm is not able to take care of this. A solution is yet to be found for this.

Thanks, once again, Dr. Silva for all the support rendered. Really appreciate it.

Best,
Jaya

[Stephen Jenkins]

Followers of this thread might be interested in

fhetprob: A fast QMLE Stata routine for fractional probit models with multiplicative heteroskedasticity

See http://www.richard-bluhm.com/data/ to net install the program and help file, and net get documentation/extended description, and example data.

[Erjan Toqbolat]

Dear Stata Users,

This is the most relevant thread I have found on the internet to ask my question. I have been reading your papers, posts, comments everywhere on the internet but could not find an answer to my question which is:

My dependent variable is of bounded nature, between 0 and 1, including both in few cases. I use glm Y X1 X2 Xn dummy1 dummy2, link(probit) family(binomial). I would like to include an interaction term in my model by multiplying the two dummy variables - glm Y X1 X2 Xn dummy1 dummy2 dummy1*dummy2, link(probit) family(binomial).

I am aware of the problem of including interaction terms in non-linear models. Some papers discuss how to address them in logit and probit models (for example, “inteff” command by Norton et al 2004 used after logit and probit commands in Stata). According to Jeff and Maarten, I cannot use probit instead of glm with link(probit) family(binomial). Is there anything I could do with interactions terms in glm with link(probit) family(binomial) case?

Thank you very much in advance.
Regards,
Erjan