Multivariate Ordered Logit/Probit
Message/Author
 Armen posted on Monday, May 15, 2006 - 5:53 pm
Dear Professor/s,

I am trying to fit a Multivariate ordered logit/probit model using Mplus. I have 6 equations (with same 8 variables in each) that I would like to estimated jointly.

Here is a part of the code in Mplus.

ANALYSIS:
!ESTIMATOR=WLSMV;
ESTIMATOR=ML;
!iterations=10000;

MODEL:

y1 on x1 x2 x3 x4 x5 x6 x7 x8;
y2 on x1 x2 x3 x4 x5 x6 x7 x8;
y3 on x1 x2 x3 x4 x5 x6 x7 x8;
y4 on x1 x2 x3 x4 x5 x6 x7 x8;
y5 on x1 x2 x3 x4 x5 x6 x7 x8;
y6 on x1 x2 x3 x4 x5 x6 x7 x8;

Question 1. Will this code estimate the model jointly?

Question 2. Does ML use logit by default? If it does, what kind of reference would you suggest for Multivariate logit models?

Question 3.How can I specify probit option for ML?

Thank you,
Armen
 Linda K. Muthen posted on Tuesday, May 16, 2006 - 6:58 am
1. Yes.
2. Yes.
Maddala, G.S. (1983). Limited-dependent and qualitative variables in econometrics. Cambridge: Cambridge University Press.
3. LINK=PROBIT; in the ANALYSIS command.
 Armen posted on Wednesday, May 17, 2006 - 8:44 pm
Thank you very much! I have run the models. One little confusion has occured. Using WLSMV Mplus reports the cross-equation correlations. However, it does not report those using ML (logit and probit). What can I do to get the cross-equation correlations using ML option?

Armen
 Linda K. Muthen posted on Thursday, May 18, 2006 - 6:45 am
I'm not sure what you mean by cross-equation correlations. Do you mean residual covariances?
 Armen posted on Wednesday, May 31, 2006 - 5:09 pm
Yes. How can I get residual covariances across equations under ML? Also, how can I compute marginal effects in Mplus for multivariate ordered probit model?

Thank you!
 Linda K. Muthen posted on Thursday, June 01, 2006 - 6:45 am
You need to use a factor as shown in Example 7.16. Each residual covariance is one dimension of integration.

I think by marginal effects you mean derivatives or elasticities. These are not provided by Mplus.
 Rick Sawatzky posted on Wednesday, August 16, 2006 - 7:26 pm
When specifying an ordered logistic model with categorical explanatory variables, is my interpretation correct that the reported coefficients for the explanatory variables represent the log odds ratios of "0" versus "1" for the dummy-coded explanatory variables (i.e., the reference category for the dummy-coded explanatory variable is "1")? The signs of the coefficients are opposite from what I would expect if the reference category were "0" for the explanatory variables.
 Bengt O. Muthen posted on Thursday, August 17, 2006 - 10:51 am
No, the reference category for the 0/1 x variable is 0 like in regular regression on a dummy variable. Perhaps there is another reason for the sign change.
 Kyle Stiegert posted on Tuesday, June 05, 2007 - 6:20 am
The problem I am working on involves correlated ordinal responses to numerous survey questions. All the survey questions are driving toward a similar theme (effectiveness of technical assistance) but they ask about various skills and services the TA provided. It appears multivariate ordered probit or logit is one way to proceed.

The above setup by Armen looks very user-friendly. Will Mplus use markov-chain monte carlo simulation to maximize the likelihood function. How can I read more on what mplus is doing in the background?
 Linda K. Muthen posted on Tuesday, June 05, 2007 - 9:02 am
Mplus does not do Bayesian estimation using Markov-Chain Monte Carlo. For ordinal responses, probit regression is available using a weighted least squares estimator or a maximum likelihood estimator and logistic regression is available using a maximum likelihood estimator. See the technical appendices on the website for further details.
 Sarah Ryan posted on Wednesday, February 29, 2012 - 11:50 am
Linda stated, in 2006:

"I think by marginal effects you mean derivatives or elasticities. These are not provided by Mplus."

Is this still the case? A reviewer is asking for average marginal effects (my work is based on a structural regression model using WLSMV).
 Linda K. Muthen posted on Wednesday, February 29, 2012 - 6:14 pm
There have been no changes in this regard.
 Sindra Sharma-Khushal posted on Monday, January 27, 2014 - 7:07 am
Hi,

I have a control and treatment group with multiple binary outcomes and a mix of categorical and continuous IVs. I was hoping to run a multivariate probit and have specified the following code:

GROUPING IS Trmt (1 = CLIMATE 2 = CONTROL);
CATEGORICAL ARE y1 y2;

ANALYSIS:
ESTIMATOR IS WLSMV;
ITERATIONS = 10000;
CONVERGENCE = 0.01;
PARAMETERIZATION=DELTA;

MODEL:

y1 y2 ON x1 x2 x3 x4 x5 x6;
'x1-x3 are categorical'

Am I on the right track?
 Linda K. Muthen posted on Monday, January 27, 2014 - 9:00 am
It looks that way. If x1, x2, or x3 has three or more unordered categories, you need to create a set of dummy variable.
 Sindra Sharma-Khushal posted on Monday, January 27, 2014 - 10:51 am
Thank you