Hi there,

I am running a mixed logit model using the user written command mixlogit on Stata 16/MP8 with only one random coefficients and around 100 additional (constant) parameters to estimate. I have a very large number of observations (NXJ is more than 200 millions). Convergence is not an issue. The main problem is the time it takes to compute each iteration of the Sim Lik. I have tried with a few random sub-samples and have some questions on how to improve on speed for the overall sample:

1. I noticed that speed is quite sensitive to the number of halton draws -- nrep(#) option. How small is too small in order to increase speed at the cost of inducing noise and bias in the estimated coefficients?

2. Does the number of halton draws affect only the estimates of the random coefficients or also the other estimates of the model? I did not find any monte-carlo result on this, any suggestions?

3. Any other trick to improve speed in this case? What about the number of initial sequence elements to drop when creating Halton sequence? Can I reduce this to a very small number through the burn(#) option since I have only random coefficient?

Thanks a lot in advance for your attention and time. Best,

Matteo