Cmp command very slow!

[Chiara Natalie Focacci]

Dear all,

I am running the cmp command since last night and still got no results (this means 10 hours) . Is it common? How much does it take to have some results? My data set is not that large. I have 16.000 observations, my dependent variable is ordered with 5 and the regressor is endogenous so I also use an instrument.

Thanks in advance

[Nick Cox]

David Roodman (the author of cmp, which is a community-contributed command from SSC, and earlier the Stata Journal, as you are asked to explain) monitors the forum for threads concerning his program. I think he would want much more detail.

With 16000 observations any issue is likely to be in the model and how difficult it is to get estimation to converge. I think you should give the exact command you used and you may need to say much more about the data. Have you tried simpler models and found that they converge easily?

[Chiara Natalie Focacci]

Dear Nick,

thanks a lot for your reply. My outcome of interest is Yi={0,1,2,3,4,5} as indicating different intensities of a certain characteristic in the contracts offered to unemployed subjects. I control for gender, nation and age dummies. My main regressor is participation in a certain programme for the unemployed, but this is endogenous so I use eligibility as an instrument. To sum up, this is the command I am trying to run since yesterday:

cmp (Y=participation sex nationdummy1-nationdummy103 agedummy1-15) (participation=eligible sex nationdummy1-nationdummy103 agedummy1-15), ind($cmp_probit $cmp_oprobit) qui robust.

I use cmp_probit for the participation / eligibility story with participation being (0,1) and eligibility being (0,1); and oprobit for Y as it is ordered according to different intensities. I observe 16,000 individuals overall, with some of them participating in the programme and others not. I observe the type of job offered at the end of the programme, so it is a standard policy evaluation.

Thanks,
Chiara

[Nick Cox]

Thanks for the detail. I will leave the field open for experts (especially David himself) to make well-informed comments. Yet I am not surprised that with (it may be guessed) well over 100 parameters being estimated that the model is proving hard to fit.

[Chiara Natalie Focacci]

Thanks Nick. I changed my model and only used the most important dummy for nation. With fewer parameters it works!

What I was wondering is how to interpret the results and most of all where to find the different classes of my Y. I can only see what the effect of X is on a general Y but not the effect of X for Y when Y=0, Y=1, Y=2, Y=3, Y=4, Y=5.

I look forward to receiving a comment from David!
Best,
Chiara