Cmroprobit - Issues of convergence

[Jean Dewals]

Dear colleagues,

I'm currently working on a consumers survey. In this survey, we asked to seafood consumers to rank their 3 preferred labelling schemes among a pool on five alternatives (Origin France, Ecolabel, Health claims, Animal Welfare and a No label alternative).

After that, i want to study the preferences expressed regarding different consumers' characterisitics. To do so, i used the cmroprobit command in STATA 17.

I have a sample of approx 900 individuals, and my data are organised as follow (exemple for 3 individuals)

ID = identifies each individual in our sample
Genre = indentifies the gender of our individuals
Univ and sante = Measures consumers' motivations regarding their altruism and their implication in diet
altern~s = identifies our labelling schemes alternatives
choix = identifes the ranking attributes by the consumers to each labelling scheme


I then used the follwing code :

cmset ID alternatives
cmroprobit choix, casevars (genre univ sante) reverse basealternative(5)

I also have increased my maximum iteration to 3000 before lauching my model.

Unfortunately, i didn't succed to make my model convege to a solution. I tried to change my basealternative, to change my explanatory variables, but i never succeed.

I had the following error message from STATA:

cannot compute an improvement -- flat region encountered
r(430)

I don't know if some of you have experience with this kind of model, or have already encountered the same problem as me. If possible I would like to have your opinions on this issue.

- What could be the problem in my data or my model choice ?

- How to overcome this problem?

I'm not an expert in STATA, sorry if I'm making big mistakes in the way to proceed. I remain available for any further information !

Regards,

DEWALS Jean-François

[Hong Il Yoo]

Jean Dewals: As Keane (1992, [Link]) illustrates, it's difficult to identify the multinomial probit model with an unrestricted covariance matrix empirically when you don't have any attribute that varies over alternatives within a choice set. Since the rank-ordered probit model assumes the same random utility specification as the multinomial probit model, I suspect that it's similarly difficult for you to identify the rank-ordered probit model empirically. You can consider constraining the covariance matrix to an identity matrix, as commonly done with the multinomial probit model when one doesn't have alternative-specific attributes.

[Jean Dewals]

Der Hong Il Yoo, thanks for your reply!

I understand that the problem encountered is therefore the lack of a specific alternative variable in my data. I thought this model could handle such limits (through the use of "casevar" option).

My problem is that i'm not sure to understand how i can translate my covariance matrix into and identity matrix in a good way. Isn't there a risk of losing information through such conversion?

My question is : Do you have any exemple to perform this conversion correctly in STATA? Or a reference that could help me to better understand the purpose of this conversion and its execution?

Sorry, I am not a specialist in statistical issues, but I would like to have a greater understanding of the problems I encounter.

Best regards,

Jean-François

[Hong Il Yoo]

Jean Dewals: I haven't used the command before, but if my reading of -help cmroprobit- is correct you can add -correlation(independent)- and -stddev(homoskedastic)- to your options statement to impose the relevant constraints. For example, -cmroprobit y, casevars(x) reverse correlation(independent) stddev(homoskedastic)-. Your concern is valid, your empirical results will not allow you to say anything about correlation and heteroskedasticity in your latent errors. Having said that, this is not necessarily a loss of information since the information in question cannot be obtained to begin with: Given your data, a model with a flexible covariance matrix is unidentified empirically.

[Jean Dewals]

Hong Il Yoo Thanks a lot for your help, i will try this solution !