Multivariate Probit Set-up using mvprobit

[Samuel Welch]

Dear Users,

I have been doing some research through past literature into how a multivariate probit model could be applied to my investigation to analyse the adoption of multiple adaptation strategies by farmers. I have binary variables for my the uptake of my three key dependent variables, as well as some other relevant independent variables to include in the model. I was trying to look through this forum to understand how to set-up a model like this in stata. Also my data is weighted (pweight) data which could also complicate matters further.

When attempting to use the mvprobit command in stata I received the following error message:

unknown function *normprob()
r(133);


Any help with structuring the set-up of this model or troubleshooting the above error would be greatly appreciated!

Thanks in advance.

[John Mullahy]

Samuel, it would be helpful if you could provide your estimation command and indicate what version of Stata you are running. Also tagging Stephen Jenkins who is likely to have definitive insights.

[Samuel Welch]

Thank you for the response John!

As recommended I will tag Stephen Jenkins to potentially assist with my issue.

I am relatively new to stata but have been trying to use information on this forum and the wider internet to guide my work.

The estimation command which gave the above error was as follows:

Code:
mvprobit (DepVar1 = IndepVar1 IndepVar2 ...) (DepVar2 = IndepVar1 IndepVar2 ...) (DepVar3 = IndepVar1 IndepVar2 ...)

I am running Stata version 17.

I am also interested as to whether I could include weights in the mvprobit command, however I wanted to understand how to format the estimation command prior to addressing this.

Thanks in advance for any help!

[Stephen Jenkins]

I don't why there is a problem, to be honest. (Note too that the help-file tells you which weights are allowed.) The program works for me:

. ssc d mvprobit

----------------------------------------------------------------------------------------------------------------------------------------------------------
package mvprobit from http://fmwww.bc.edu/repec/bocode/m
----------------------------------------------------------------------------------------------------------------------------------------------------------

TITLE
      'MVPROBIT': module to calculate multivariate probit regression using simulated maximum likelihood

DESCRIPTION/AUTHOR(S)
      
       mvprobit estimates M-equation probit models, by the method of
      simulated maximum likelihood (SML). (Cf. probit and biprobit
      which estimate 1-equation and 2-equation probit models by maximum
      likelihood.) The variance-covariance matrix of the
      cross-equation error terms has values of  1 on the leading
      diagonal, and the off-diagonal elements are correlations  to be
      estimated. mvprobit uses the Geweke-Hajivassiliou-Keane (GHK)
      simulator  to evaluate the M-dimensional Normal integrals in the
      likelihood  function. For each observation, a likelihood
      contribution is calculated  for each replication, and the
      simulated likelihood contribution is the  average of the values
      derived from all the replications. The simulated  likelihood
      function for the sample as a whole is then maximized using
      standard methods (ml in this case).
      
      KW: multivariate probit
      KW: simulated maximum likelihood
      
      Requires: Stata version 7.0
      
      Distribution-Date: 20240306
      
      Author: Lorenzo Cappellari, Université del Piemonte Orientale and University of Essex
      Support: email [email protected]
      
      Author: Stephen P. Jenkins, London School of Economics and Political Science
      Support: email
      

INSTALLATION FILES                             (type net install mvprobit)
      mvprobit.ado
      mvprob_ll.ado
      mvprobit.hlp
      mvppred.ado
      mvppred.hlp
----------------------------------------------------------------------------------------------------------------------------------------------------------
(type ssc install mvprobit to install)

. which mvprobit
d:\home\stephenj\ado\stbplus\m\mvprobit.ado
*! version 3.0.1  05may2024 Cappellari & Jenkins   
*!  fix rarely-biting bug in Replay (thanks to Jeff Pitblado for suggested fix)
*! version 3.0.0  12april2006 Cappellari & Jenkins            
*!      revised: aa algorithm, draw vble naming scheme
*! version 2.1.0  19may2005  Cappellari & Jenkins            
*!      update to version 8.2
*! version 1.0.0  15jan2003  Cappellari & Jenkins            (SJ3-3:st0045)
*! Multivariate probit by method of MSL.  

. use http://www.stata-press.com/data/r7/school.dta, clear

. mvprobit (private = years logptax loginc) (vote = years logptax loginc) (pub12 = years loginc), draws(53)

Iteration 0:  Log likelihood = -155.61255  
Iteration 1:  Log likelihood = -153.95274  
Iteration 2:  Log likelihood = -153.82632  
Iteration 3:  Log likelihood = -153.82562  
Iteration 4:  Log likelihood = -153.82562  

Multivariate probit (MSL, # draws = 53)           Number of obs   =         95
                                                  Wald chi2(8)    =      10.02
Log likelihood = -153.82562                       Prob > chi2     =     0.2636

------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
private      |
       years |  -.0137918   .0261964    -0.53   0.599    -.0651359    .0375523
     logptax |  -.1699739   .6602698    -0.26   0.797    -1.464079    1.124131
      loginc |   .3723182   .5128901     0.73   0.468     -.632928    1.377564
       _cons |  -3.690357   4.837828    -0.76   0.446    -13.17233    5.791611
-------------+----------------------------------------------------------------
vote         |
       years |  -.0165344   .0147786    -1.12   0.263       -.0455    .0124312
     logptax |  -1.281263   .5680567    -2.26   0.024    -2.394633    -.167892
      loginc |   1.006309   .4406539     2.28   0.022     .1426436    1.869975
       _cons |  -.6689479   4.042113    -0.17   0.869    -8.591345    7.253449
-------------+----------------------------------------------------------------
pub12        |
       years |   .0046203   .0144897     0.32   0.750    -.0237789    .0330195
      loginc |   .1668134   .3437143     0.49   0.627    -.5068542    .8404809
       _cons |  -1.740017   3.475208    -0.50   0.617    -8.551299    5.071264
-------------+----------------------------------------------------------------
    /atrho21 |  -.2282406   .2386533    -0.96   0.339    -.6959924    .2395113
-------------+----------------------------------------------------------------
    /atrho31 |  -.2959931   .2197236    -1.35   0.178    -.7266435    .1346573
-------------+----------------------------------------------------------------
    /atrho32 |    -.08106   .1694425    -0.48   0.632    -.4131612    .2510412
-------------+----------------------------------------------------------------
       rho21 |  -.2243581   .2266403    -0.99   0.322    -.6018178    .2350341
-------------+----------------------------------------------------------------
       rho31 |  -.2876415   .2015442    -1.43   0.154    -.6210076    .1338493
-------------+----------------------------------------------------------------
       rho32 |  -.0808829    .168334    -0.48   0.631    -.3911535    .2458971
------------------------------------------------------------------------------
Likelihood ratio test of  rho21 = rho31 = rho32 = 0:  
             chi2(3) =  3.57387   Prob > chi2 = 0.3113

. version 17: mvprobit (private = years logptax loginc) (vote = years logptax loginc) (pub12 = years loginc), draws(53)

Iteration 0:  Log likelihood = -155.61255  
Iteration 1:  Log likelihood = -153.95274  
Iteration 2:  Log likelihood = -153.82632  
Iteration 3:  Log likelihood = -153.82562  
Iteration 4:  Log likelihood = -153.82562  

Multivariate probit (MSL, # draws = 53)           Number of obs   =         95
                                                  Wald chi2(8)    =      10.02
Log likelihood = -153.82562                       Prob > chi2     =     0.2636

------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
private      |
       years |  -.0137918   .0261964    -0.53   0.599    -.0651359    .0375523
     logptax |  -.1699739   .6602698    -0.26   0.797    -1.464079    1.124131
      loginc |   .3723182   .5128901     0.73   0.468     -.632928    1.377564
       _cons |  -3.690357   4.837828    -0.76   0.446    -13.17233    5.791611
-------------+----------------------------------------------------------------
vote         |
       years |  -.0165344   .0147786    -1.12   0.263       -.0455    .0124312
     logptax |  -1.281263   .5680567    -2.26   0.024    -2.394633    -.167892
      loginc |   1.006309   .4406539     2.28   0.022     .1426436    1.869975
       _cons |  -.6689479   4.042113    -0.17   0.869    -8.591345    7.253449
-------------+----------------------------------------------------------------
pub12        |
       years |   .0046203   .0144897     0.32   0.750    -.0237789    .0330195
      loginc |   .1668134   .3437143     0.49   0.627    -.5068542    .8404809
       _cons |  -1.740017   3.475208    -0.50   0.617    -8.551299    5.071264
-------------+----------------------------------------------------------------
    /atrho21 |  -.2282406   .2386533    -0.96   0.339    -.6959924    .2395113
-------------+----------------------------------------------------------------
    /atrho31 |  -.2959931   .2197236    -1.35   0.178    -.7266435    .1346573
-------------+----------------------------------------------------------------
    /atrho32 |    -.08106   .1694425    -0.48   0.632    -.4131612    .2510412
-------------+----------------------------------------------------------------
       rho21 |  -.2243581   .2266403    -0.99   0.322    -.6018178    .2350341
-------------+----------------------------------------------------------------
       rho31 |  -.2876415   .2015442    -1.43   0.154    -.6210076    .1338493
-------------+----------------------------------------------------------------
       rho32 |  -.0808829    .168334    -0.48   0.631    -.3911535    .2458971
------------------------------------------------------------------------------
Likelihood ratio test of  rho21 = rho31 = rho32 = 0:  
             chi2(3) =  3.57387   Prob > chi2 = 0.3113

[Samuel Welch]

Dear Stephen Jenkins,

Thank you very much for the response Professor.

Even using your above code I still receive the same issue:

Code:
ssc install mvprobit
which mvprobit
/Users/samuelwelch/Library/Application Support/Stata/ado/plus/m/mvprobit.ado
*! version 3.0.1 05may2024 Cappellari & Jenkins
*! fix rarely-biting bug in Replay (thanks to Jeff Pitblado for suggested fix)
*! version 3.0.0 12april2006 Cappellari & Jenkins
*! revised: aa algorithm, draw vble naming scheme
*! version 2.1.0 19may2005 Cappellari & Jenkins
*! update to version 8.2
*! version 1.0.0 15jan2003 Cappellari & Jenkins (SJ3-3:st0045)
*! Multivariate probit by method of MSL.

. d:\home\stephenj\ado\stbplus\m\mvprobit.ado
: invalid name
r(198);

. *! version 3.0.1 05may2024 Cappellari & Jenkins

. *! fix rarely-biting bug in Replay (thanks to Jeff Pitblado for suggested fix)

. *! version 3.0.0 12april2006 Cappellari & Jenkins

. *! revised: aa algorithm, draw vble naming scheme

. *! version 2.1.0 19may2005 Cappellari & Jenkins

. *! update to version 8.2

. *! version 1.0.0 15jan2003 Cappellari & Jenkins (SJ3-3:st0045)

. *! Multivariate probit by method of MSL.

.
. . use http://www.stata-press.com/data/r7/school.dta, clear

. mvprobit (private = years logptax loginc) (vote = years logptax loginc) (pub12 = years loginc), draws(53)

unknown function *normprob()
r(133);

Hence I believe it may not be an issue with my data or code but something else.

[Joseph Coveney]

I can replicate this error.

Using set trace on the error arises in the program mvprob_ll when a global macro (in bold font below) evaluates to an empty string.

                  - replace `d`i'' = invnorm(${S_MLE_z`i'`d'}*normprob((`arg`i'')/`c`i'`i''))
                  = replace __00009T = invnorm(*normprob((__00009V)/__0000AH))
unknown function *normprob()

Do-file and log file are attached for reference. The log file is ca. 360 kilobytes (contains the output from trace), the forum software apparently cannot accommodate a file that large as an attachment and so I zipped it before attaching it. The forum software won't permit attaching a filename with a .zip extension and so I changed it to .txt. (You can rename its file extension back to .zip after downloading it.)

I didn't install the package using ssc install, but rather copied the files to the present working directory (see do-file). I don't know whether that has anything to with causing the error to happen.

[Samuel Welch]

Dear Joseph,

Thank you for your response.

However, I am a bit confused about what you are trying to say I should do when faced with this error?

Thanks in advance!

[Stephen Jenkins]

Thanks for the further investigations, Samuel Welch and Joseph Coveney I cannot reproduce this problem. Relatedly, I can't see how/why the global macro that Joseph cites is empty. The code block in mvprobit.ado in which it/they are created is:

        forval i = 1/$S_MLE_M {
                forval d = 1/`D' {
                        tempvar z a
                        draws `z' `touse'
                        global S_MLE_z`i'_`d' "`z'" 
                        if "`aa'" != "" {
                                ge double `a'`z' = 1 - `z'
                                global S_MLE_z`i'_`=`D'+`d'' "`a'`z'" 
                        }
                }

The other global macros that are required (like S_MLE_M, the number of equations) are created earlier on. The error CODE extract that Joseph shows is from a call to mvprob_ll.ado. That's a version 8.2 program (if that matters).

So, I am stumped. (Also puzzled because other users have been using -mvprobit- without problems within the last year or so.)

[Joseph Coveney]

So, I am stumped. (Also puzzled because other users have been using -mvprobit- without problems within the last year or so.)

Is that when this last revision was made, May of last year? (The revision date is accidentally recorded as May of this year.)

Anyway, I can see a problem in that the underscore is missing when that global macro is called in the line that I show above: the code that you post (that creates the macro) shows its having an underscore

global S_MLE_z`i'_`d' "`z'"

and

global S_MLE_z`i'_`=`D'+`d'' "`a'`z'"

but what's called (in the snippet of trace output shown in my post) lacks that underscore.

replace `d`i'' = invnorm(${S_MLE_z`i'`d'}*normprob((`arg`i'')/`c`i'`i''))

and so no such global macro exists.

[Stephen Jenkins]

Joseph Coveney Thanks very much again. Well spotted. Strange, nonetheless, that the line you refer to (line 90 in mvprob_ll.ado) is the following in my (master) copy, and I note the underscore present!

                replace `d`i'' = invnorm(${S_MLE_z`i'_`d'}*normprob((`arg`i'')/`c`i'`i''))

[The very minor update to -mvprobit- recently was to deal with a rare problem arising with replay-ing the command and I changed mvprobit.ado.]

Whatever, it seems that the version of mvprob_ll.ado on SSC has a problem. I don't know how it arose, but I'll ask Kit Baum to replace the version there with what I have here.

thanks again for your help with this.

[Samuel Welch]

Thank you very much for all your help in investigating this and trouble shooting the issue Joseph Coveney and Stephen Jenkins, it is much appreciated.

Will it just be a case of re installing the updated version of mvprobit when it is replaced by the new version?

Thanks again for your help with this!

[Stephen Jenkins]

-ssc install mvprobit, replace- should do it. If you're in a hurry, you could edit line 90 in mvprob_ll.ado to fix the apparent problem with the SSC version of that file, noting the commentary above. (I've sent my version of that file to Kit/SSC.)

[Samuel Welch]

Dear Stephen Jenkins,

Thank you again for all your help, I have implemented the recommended changes into line 90 of the mvprob_ll.ado and run the command and it now works.

Also thank you very much Joseph Coveney for you help with finding the issue with the command!

[Stephen Jenkins]

with thanks as ever to Kit Baum, the "mvprob_ll.ado" file in the "mvprobit" package on SSC has been replaced. (See #12)

[Adama Touray]

Dear Stephen Jenkins

I am trying to apply mvprobit to my current research, but I need a little clarification on the factor variable operators with mvprobit (it's completely new to me). I tried to incorporate the operator (e.g. i.sex) as is usually the case with probit/logit but it is not working. However, when I removed the operator, I was able to estimate the model. Thus, how do I interpret such a categorical variable, especially when the categories are more than 2?

The second question is, aside from mvpred, are there any other postestimation tests I need to do? The third question is can I have the same exact variables in all four equations (I have four equations and I want to use the same covariates)? And finally, what is the command to create the correlation table after the estimation? I would also appreciate it if you could suggest more reading materials on mvprobit aside from Cappellari and Jenkins (2003, 2004, 2006) and Greene (2003). Thank you in advance.