st: mvprobit Cholesky factorization

["Meulemann Max" <mmeulemann@ethz.ch>]

Hello everyone, 

I am using Stat 12 SE and I am having problems with mvprobit estimations. My data is from a larger worldwide survey and has roughly 350 observations.
Using mvprobit I often get the message "Warning: cannot do Cholesky factorization of rho matrix" and finally "could not calculate numerical derivatives flat or discontinuous region encountered"

I have a larger set of responses on an expected leadership of a country i for an issue j. There are 5 countries i=1,2,3,4,5 and a category None i=6 and there are 6 issues j=1,2,3,4,5,6. Respondents were allowed to pick several countries as leading for an issue. Otherwise I would have probably liked to use multinominal models. 

I could fit univariate probit models for each of the 6*6 possible answer or account for correlation among those items across countries or issues and additionally use multivariate probit and hence mvprobit.

In principle, this works fine but for some exceptions. I have problems to do estimations across regions for each issue when I include the variables for the category "None". I decided to exclude this item when I estimate across regions for each issue as the tetrachoric and polychoric correlations are close to -1 or cannot be estimated. From the help on mvprobit I gather it is problematic when the rhos get close to one and cause the problem with the Cholesky factorization. I guess it makes intuitively sense that this correlation is -1 as whenever they give a positive answer for any region they should give a negative answer to no region by design. 

I have more trouble when I estimate across issues for each region and especially for the category "None". I tried larger number of draws up to 50 (using more draw takes forever). Pre-estimation of the atrho0 matrix. Options aa and difficult. Excluded regressors that perfectly predicted failure/success in the univariate probit models. Changed seeds. I checked for high correlations among items with tetrachoric and polychoric commands, but I am not sure what I could do then? There is a high degree of correlation but normally not close to one. Is there any way I can avoid this error or can investigate what cause the correlation problem? 

Best

Max

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/