Logit regression do not converge

[Inês Gomes]

Hello,

i have a dataset containing information about entrepreneurs in the field of gender, education level, age, industry of start-up, year of creation , economic activity controls (gdp, population and unemployment).

My main regression contains

dependent variable : gender (1 for female/ 0 for male)

independent variables : dummies for education level, age, entry dummies ( if firm was established in 80s, 90s or 00s), log of control variables

+ year, industry and county fixed effects

When i only use the dependent and independent variables, stata runs the regression perfectly and without problems. Unfortunately, when i add the fixed effects , the regressions does not converge.

How can i solve this problem, using the fixed effects? I need this specification in logit and probit.

[ben earnhart]

Please provide exact commands you used, and exact output. If you click on the "A" right above the box where you type, you will get the full menu for formatting. Choose "#" for code, and paste within the code tags. Then we might have a clue what is going on. In general, I'm iffy on *ever* using gender as a dependent variable. I suppose in a multi-level context, given your independent variables, it might make sense, though; the log odds of an *organization* choosing or being founded by a female... Huh.

[Attaullah Shah]

You can force the number of iteration by specifying the option iter(30) or higher. This does not guarantee the equation will converge, but at least stata will not remain in lingo and will produce results after 30 iterations.

[Andrew Lover]

Are you getting any error messages? Otherwise, you can investigate adding the -difficult- or -technique- options.

[Carlo Lazzaro]

Inès:
you give the list too few details for getting helpful replies.
I second the previous advices about reporting exactly what you typed and what you Stata gave you back (as per FAQ). Otherwise, the best you can get (from me, at least), is one or more blind guesses about what went wrong with your model.
If your model doesn't converge (by the way, what is the messages that appears when the regression does not converge: back-up? concavity warning?) after you add fixed effects, it would be interesting to see how fixed effects were included in the right hand side of the regression (e.g.: have you cross-checked that all the dummies are correct?).

Kind regards,
Carlo

[Maarten Buis]

What puzzles me is that you use gender as your dependent variable. If you mean with that the biological sex of the respondent, than how can that be explained by education, age, etc.? Does a male (in recent cohorts) suddenly become female when he attains more education? If you mean the socially constructed gender using it as a dependent variable can make more sense, but how can that be a 0/1 variable?

[Richard Williams]

I can see situations where gender is the dependent variable. In particular, suppose respondents are given a choice between a man and a woman. What makes them more likely to choose a woman?

If you were analyzing corporations, what sorts of corporations are more likely to choose women leaders?

It does seem trickier or more problematic in this case. It seems like it would make more sense to have the dv be entrepreneurial/ not entrepreneurial, with gender as one of the explanatory variables. But that doesn't seem to be the nature of the data set.

We don't have any explanation of the theory behind the problem. Maybe it does make good sense in this case.

But in any event, we can't help Ines with anything until we have more information!

[Nick Cox]

I'd ask also about the proportions male and female. Logit models don't get easier to estimate if the mean response is very near 0 or very near 1. In this case I fear the former.

[Richard Williams]

If it is a rare events problem, firthlogit (available from SSC) may be good. Allison discusses this at http://www.statisticalhorizons.com/l...or-rare-events.