Convergence not achieved - mlogit

[Chris Rooney]

Hello all

I ran a multinomial logit, and received the message that "convergence was not achieved" (after 16 000 iterations - took a while!). However, after removing the time dummies (d1-d8), the multinomial logit ran perfectly. Why would this be the case? And is it possible to include them so that the mlogit runs smoothly? The time dummies are set up in a such a way that if a person is in period 1, d1=1, and for every other time period d1=0, etc.

[Jason Cruso]

Hi,
Just bumped into this old thread.
I face the same issue. My mlogit does not converge if i include year dummy. Specifically, for one particular year. Wonder what could be the problem?

Thanks,

J

[Clyde Schechter]

-mlogit- is a complicated and somewhat brittle estimator. What does the distribution of that year look like and how does it relate to the outcome variable? If that year variable is nearly always 0, or if it comes close to perfectly predicting the outcome, then that can be a cause of non-convergence. You might explore this by running:

tab year_dummy outcome_variable if e(sample)

after the -mlogit- runs.

Also take a close look at the output from -mlogit- itself. These results are not valid and cannot be used, but they may offer a clue about what is going wrong. See if there are other variables whose parameter estimates are ridiculous, or have missing or ridiculous standard erorrs. It could be that this particular dummy is nearly colinear with those variables, another possible cause of non-convergence.

Those things are worth a try. In the end, you may or may not find out what the problem is.

[Richard Williams]

Sometimes adding the -difficult- option to a command works wonders. More often you will have to do things like Clyde suggests.

Whatever you do, I would add the option -iterate(1000)- or even -iterate(200)-. I've never seen a command converge on iteration 15,997. Usually if it doesn't converge after a few dozen iterations it never converges at all. I don't know why Stata picked such a huge default value.