cmp for instrumental multinomial probit when there is cluster, baseoutcome, if conditions

[James Park]

I have been running non-instrumental multinomial probit that roughly looks like

mprobit Y X C

where Y is dep var, X is indep var, C are controls.

Now I wasted add an instrument.

So after reading some posts from Statalist, it seems typical command for instrumental multinomial probit is the following (is this correct?)

cmp (Y=X C, iia) (X=Z C), ind($cmp_mprobit $cmp_cont)

where Z is the instrument. I added iia because there is no alternative (Y)-specific controls.

Thank you for all the posters who made me come this far.

But what if original non-instrumental multinomial probit model was more complex? It had cluster, baseoutcome, if conditions.

mprobit Y X C if C<100, baseoutcome(1) vce(cluster D)

Then how should I write instrumental multinomial probit that corresponds to this?

The following two attempts failed.

cmp (Y=X C, iia) (X=Z C) if C<100, ind($cmp_mprobit $cmp_cont) baseoutcome(1) vce(cluster D)

cmp (Y=X C, iia baseoutcome(1)) (X=Z C) if C<100, ind($cmp_mprobit $cmp_cont) vce(cluster D)

Webuse example

By arbitrarily adding fem_work=2, I get a data that I can meaningfully run multinomial probit.

webuse laborsup, clear
replace fem_work=2 if fem_work==1 & _n<200
cmp (fem_work = other_inc fem_educ kids) (other_inc = fem_educ kids male_educ), ind($cmp_mprobit $cmp_cont) qui

This code does not just have cluster, if, baseoutcome. But still this is taking forever to run.. still not finished.

[David Roodman]

cmp does not accept a baseoutcome() option. As explained in the help file, the base outcome must always be first. So if you want another outcome to be the base, recode your outcome variable so the base value gets the lowest number.
If it's getting stuck on the constants-only model, try adding the nolr option.
cmp accepts if clauses and the cluster() option per usual syntax.

[James Park]

Thank you so much! So I tried running

 
 cmp (Y=X C, iia) (X=Z C) if C<100, ind($cmp_mprobit $cmp_cont) vce(cluster D)

It doesn't produce any error. But is this applying if condition and cluster to both regressions?

hat I want is that both 1st regression (Y=X C, iia) and 2nd regression (X=Z C) are clustered by D, and both are limited to observations s.t. C<100. Is that what the above line is doing? Or is "C<100" or "vce(cluster D)" being applied to only one of the two regressions?

I also tried the following two but they throw me error messages so perhaps the above code is something I should do anyways (Perhaps?). But I would like to double-check what it is really doing.. Thank you again so much! cmp is so useful.

 
 cmp (Y=X C, iia if C<100) (X=Z C if C<100), ind($cmp_mprobit $cmp_cont)

 
 cmp (Y=X C, iia vce(cluster D)) (X=Z C, vce(cluster D)) if C<100, ind($cmp_mprobit $cmp_cont)

[James Park]

And another things is that, cmp doesn't seem to recognize _b[X] to save coefficient of X, or _se[X] to save standard error of X. Does cmp have different way of saving them?

replace coef=   _b[X]

[David Roodman]

Yes the if clause and cluster option apply across equations.
The question about accessing coefficients is a general one about how to refer to results from multi-equation models. See https://www.stata.com/manuals13/u13.pdf

[James Park]

Yes the if clause and cluster option apply across equations.
The question about accessing coefficients is a general one about how to refer to results from multi-equation models. See https://www.stata.com/manuals13/u13.pdf

Thank you so much again!

And meanwhile I encountered another problem. I did some research so my question became long so I posted it as a separate thread: https://www.statalist.org/forums/for...any-categories

Could you please take a look when you have time?

[David Roodman]

James, I just fixed a bug in cmp that was creating problems for multnomial probit, rank-ordered probit, and/or "gamma" models (that use the # suffix) when the nolrtest option was not in the command line. It should be on SSC soon. You can also access it now via "net from https://raw.github.com/droodman/cmp/v8.3.0". Then follow links to force installation.

--David

[James Park]

James, I just fixed a bug in cmp that was creating problems for multnomial probit, rank-ordered probit, and/or "gamma" models (that use the # suffix) when the nolrtest option was not in the command line. It should be on SSC soon. You can also access it now via "net from https://raw.github.com/droodman/cmp/v8.3.0". Then follow links to force installation.

--David

Thank you very much for letting me know. I just ran

ssc install cmp, replace
which cmp

and it shows me

c:\ado\plus\c\cmp.ado
*! cmp 8.3.3 23 May 2019

This is the right way to update, correct? And was there another update in May?

[David Roodman]

Yes, that should be fine.