Forum: open-discussion

RE: mvProbit, plans on beautifying output? [ Reply ]
By: Arne Henningsen on 2015-02-17 23:02

[forum:41917]

Dear Paul and Ott

Thank you for your suggestions for improving mvProbit!

I agree that it would be nice if the names of the coefficients were more informative.

I think that the slowness of mvProbit() comes from calculating the integrals of the multivariate normal distribution.

Currently, I am too busy with other things to work on mvProbit and I would be very happy if you could implement some (or many) improvements :-)

Please contribute your improvements through the SVN repo and please let me know if you have any questions.

Best wishes,
Arne

RE: mvProbit, plans on beautifying output? [ Reply ]
By: Ott Toomet on 2015-02-17 18:43

[forum:41915]

Dear Paul,
thank you for this encouraging feedback :-) I have myself not worked with mvProbit, I am Arne's co-author on several other packages.

I have planned several improvements on mvprobit, including speeding up the 2D case (it should be possible to use 2D normal cdf instead of Poisson integration), making it possible to let the correlation (rho) depend on the explanatory variables, and a review of the classes, methods, tests, examples, etc. However, for various reasons, I started extending sampleSelection package instead. And it is not much one can achieve during evenings and weekends. But it is still on my agenda.

Help is definitely welcome, I think Arne agrees as well :-)

Cheers,
Ott

mvProbit, plans on beautifying output? [ Reply ]
By: Paul Johnson on 2015-02-16 18:26

[forum:41911]

Dear Arne

I used mvProbit a couple of years ago and now I'm back to test it out again. I notice your output does not name the variables, it has b_0_1 and such. I'd like to get the variable names inserted there and I think I can help (we recently worked on a similar exercise). But I don't want to do that if you have already worked on it, or if you think mvProbit is a stalled project.

I'm in this because we were recently presented with some Mplus estimates claiming to be multivariate probit models, but it appears they are not. The Mplus outut has no correlations between the errors in the separate models, much to our surprise. I'm trying to get to the bottom of the question now.

I became aware of the Mplus question when I ran the same data through mvProbit. Your coefficient estimates don't match Mplus, but I'm pretty sure your estimates are correct. For confirmation, we got the Stata addon for multivariate probit (Cappellari and Jenkins, "Multivariate probit regression using simulated maximum likelihood", Stata Journal 3(3) 278-294, http://www.stata-journal.com/sjpdf.html?articlenum=st0045) which uses the simulation BHHH approach that you use, and we get parameter estimates for intercept and predictor coefficients that match mvProbit.

However, I'm a little worried about the estimates of the error correlations. In Stata, we don't get similar estimates for the latent error correlations to mvProbit. I've assigned a GRA to keep working on that, it may be we can see where the mismatch is. Unfortunately, we don't have a copy of limdep for comparison.

We notice that estimations with mvProbit are quite a bit slower than the Stata thing, and we are reminded of your warning in the documentation about mvProbit. We had a little trouble figuring out that the tolerance and iterations values ought to be because your mvProbit examples set tolerance at 0.5. After dialing that back, and tuning up the iterations, we find mvProbit converges eventually. But it is slower, I can't see for sure if the problem is in the initial values or the implementation of the optimizer. If you have guesses about where the slowdown lives, I might put a little effort into speeding it up.

pj