How to pass extra information to a maximum-likelihood evaluator when using ml?

[Chi-lin Tsai]

Hi,

I know that, if I directly use Mata's moptimize, I can use moptimize_init_userinfo and moptimize_util_userinfo to pass extra information to the evaluator programme. However, I prefer to use Stata's ml to call Mata's moptimize, so I would like to know whether there is any option in ml for doing that? Specifically, I need an extra dummy variable and a scalar to specify the log-likelihood function in the Mata evaluator programme.

I am also aware that ml's option group can pass an additional variable to the evaluator, and the variable will be treated as a cluster indicator in the calculation of robust errors. However, this is not a solution for me, because the extra variable I need is not a cluster indicator. Using ml's option group will seriously underestimate robust errors in my model.

A temporary solution found so far is specifying the dummy variable as an additional dependent variable in Stata's ml, and using it by moptimize_init_depvar in Mata. However, I don't really like this method, since it messed up the code a bit, and also the dummy variable is not really a dependent variable in the model. More importantly, this method cannot pass a scalar to the evaluator.

Overall, I would like to know whether there is any documented or undocumented option in Stata's ml for doing what moptimize_init_userinfo and moptimize_util_userinfo do in Mata.

(PS1. I consider this this question mainly as a question about Stata rather than Mata, so I post it here.)

(PS2. I am using Stata 14, but hope that my programmes can be backward compatible to at least Stata 11.)



Thank you very much.

Chi-lin

[Maarten Buis]

Ehen writing commands using Stata's ml I typically pass on extra information using global macros. This can also be the name of (temporary) variables or (temporary) scalars.

[Chi-lin Tsai]

Maarten, thanks for this advice. Using global macros is indeed a way out, and I did use this method for passing scalars to Mata evaluators. However, if I understand correctly, when my programme creates a global macro for its own purpose, there is always a possibility of overwriting users' macro. Certainly, I can choose a rarely used macro name to reduce that possibility, or backup users' macro and restore it after my programme executed. However, I still wonder whether there is undocumented options in ml for making everything even easier.

Anyways, thank you very much.

[Maarten Buis]

I understand your misgivings about using globals, I share them. I try to avoid them whenever possible, and it feels very awkward using them whenever I failed to find another solution.

I tend to use names that start with S_ML. Using such conventions is I think more reasonable than trying to guess which names are rarely used.

[Paul Wileyto]

Hoping someone is still paying attention, even though the last post was 8 years back. Now with version 17, are there any ways to get around using global macros for passing ancillary information to ml evaluators? It works, but feels clunky.

Thanks

Paul

[Maarten Buis]

Nope, your choice is still either use global macros or go to Mata.