cmmixlogit non-convergence and speed

[Jana Hilsenroth]

I'm running a mixed multinomial logit model using cmmixlogit. I have ~82,000 observations (~4,500 individuals) in my dataset. When I ran my model with all of my attributes as random parameters it took more than one month to run (Stata 17 MP) and did not converge. I'm in the process of adding one random parameter at a time and saving the previous model's values as starting points for the next iteration. I'm hoping by 'guiding' the model this way, it might converge. However, this process is again taking a long time. Are there any suggestions to speed this process up? I do have access to a supercomputer, however, I'm not sure if additional CPUs would speed up the process- i.e. am I limited by computing power or by my code?

[Hong Il Yoo]

Jana Hilsenroth: You may find my earlier reply to a related question useful.

https://www.statalist.org/forums/for...84#post1590384

If you have access to a multi-core computer and Stata MP to take advantage of it, you can definitely speed things up by using the combination of the two resources. But you should not expect a miracle: It may still take several days or weeks before you see a single set of estimates. You may consider applying specialist algorithms such as Ken Train's EM (https://doi.org/10.1016/S1755-5345(13)70022-8) and Jonathan James's MM (https://doi.org/10.1002/jae.2532) for mixed logit models which are known to speed things up considerably, but these algorithms are not without limitations (e.g., generalizability) and you'll have to do a fair amount of programming as they're not available as canned estimation commands in Stata.

If you're happy to assume that your random parameters follow a categorical distribution, you may consider applying my -lclogit2- command (https://journals.sagepub.com/doi/10....36867X20931003) to estimate latent class logit models instead of multivariate normal mixed logit models. Looking at your sample size, the estimation run may still take a few days but my guess is that it's not going to take over a month.