mvprobit or cmp?

[Mario Amato Meno]

Dear STATA Users,

I am trying to estimate a multivariate probit for a given outcome that would model jointly the three main dependent in my equation, so to have a mv probit with 4 dimensions.

To be more clear: an individual i could decide between three, not exclusive, different choices (x1, x2, x3) on an outcome y that could be 0 or 1, given a set of covariates (X), and with x1, x2, x3 and y being all dummies.

My questions are:

- should I use the mvprobit command or the Roodman's cmp?

- How should I build the command to take into account that the three dependent should be considered jointly?

Any hint would be more than appreciated!

Best regards,

Amato

[Stephen Jenkins]

take into account that the three dependent should be considered jointly

I don't understand what you mean by this. In a multivariate probit model, there are K binary outcomes each with a set of predictors, with the error terms of the equations assumed K-variate normal.

mvprobit and cmp are each fine programs, and each can be used to fit a multivariate probit model. The former is more specialised (and older) than the latter, which is a very general (and hence very powerful) tool. mvprobit uses a plug-in rather than Mata (so differences in speed are unclear ex ante). There are differences in post-estimation functionality. Over to you to do some research to decide what suits your project best.

[Mario Amato Meno]

Dear Mr Jenkins,
thanks for you reply.

Maybe I was unclear, but I am quite new to STATA. I understand what you said about the two commands and I have read the features both commands have. My question on the point that was not clear enough is: how to estimate the model taking into account that there can be correlated errors between the strategies?

[Stephen Jenkins]

how to estimate the model taking into account that there can be correlated errors between the strategies?

Sorry, but I have no idea what you mean. As stated in #2, both programs allow correlations between the errors in the different equations -- that's the essence of the multivariate probit model.

One could fit K univariate probit equations, i.e. ignoring the cross-equation error correlations, and your estimates of the coefficients on the regressors in each equation would be consistent (in statistical sense). However, if in fact, the cross-equation correlations are non-zero, the multivariate probit approach (which estimates the correlations assuming a multivariate normal distribution) produces consistent and more efficient estimates.

I recommend some proper reading. Perhaps study the manual discussion of the bivariate probit model first.

PS it's "Stata" (not "STATA") -- please see Forum FAQ about this