Logistic "flat region encountered"

[Gary Fissel]

Stata users,
I am running code that runs a Logistic regression. With my sample and no options I received a "convergence not achieved" error. After I inserted the following options into logistic, technique(dfp 10 nr 10) difficult tolerance(1e-2), I received the "cannot compute an improvement -- flat region encountered" error message. This also occurred when I changed the technique option from 10 to 5. Are there any other alternatives to get this logistic regression to run with the existing sample? Please let me know your ideas on this.

[Carlo Lazzaro]

Gary:
your model might be misspecified.
You may want to add one predictor at time and see when Stata starts warning you about the lack of convergence.

[Gary Fissel]

Carlo,
Thanks for your suggestion. I will try to winsorize my sample to eliminate outliers on variables in my model.

Best,
Gary

[Stephen Jenkins]

Carlo's recommendations referred to inclusion or exclusion of predictors, not to Winsorization. [NB There are recent threads on Statalist arguing the (de)merits of Winsorisation.]

[Joao Santos Silva]

Dear Gary,

ML estimation of a logit model should be reasonably easy even if the model is misspecified. So, I would say that the most likely cause of the difficult convergence is a numerical issue caused either by extreme collinearity or by having regressors in very different scales. Changing the optimization technique in this problem is unlikely to help, and messing with the convergence criteria may be dangerous. Besides making sure that all regressors are in comparable scales, you may start by estimating a model with just a few regressors and then include the others one by one. This should allow you to identify the regressors causing the problem, and that will give you clues about possible solutions.

Best wishes,

Joao