Logistic Regression with more than one dependent variable

[Swapnil Shekhar]

Hi, I need to run a logistic regression with more than one dependent variable however, the set of independent variable is same for all the dependent variables. So, I want to estimate a multivariate logit model which basically estimates the equations simultaneously for the same set of independent variables. From what I understand, there is no way to estimate that. The closest that I got to multivariate quality response models was biprobit model. Please guide me to estimate this equation.

Thanks,
Swapnil

[Maarten Buis]

The fact that your independent/explanatroy/right-hand-side/x-variables are the same is not enough to justifiy simulataneous estimation. You also need to care about (potential) correlation among the error terms. You could estimate such a model with gsem. However, these correlations are hard to estimate / identify as you don't observe them directly, and the number of these hard to estimate correlations increases very fast with the number of depedent/explained/left-hand-side/y-variables. So my very first step would be to ask whether I really need to do this, and I would seriously consider not doing this simultaneously. If I were to continue, I would expect to run into trouble very quickly, and I would start thinking about whether there are additional identification possibilities with this dataset-model combination at a very early stage.

[Swapnil Shekhar]

Hi Mr. Buis. Thank you for your reply. I referred to certain papers which suggest that in the given case, where I am trying to determine the factors influencing strategies adopted by the farmers to adapt to changing climate, use of multivariate logit instead of multinomial logit. However, I would like to know how does one decide about choosing among the two models?

[Maarten Buis]

A multivariate logit would allow multiple strategies for the same farmer. The real question with such models is: do you want to allow correlations among the error terms, and when you do, can you identify such a model?

With a multinomical logit each farmer can choose only one strategy.

[Nathan E. Fosse]

Hi, there! It sounds as if you're interested in multivariate regression (not to me confused with multiple regression, i.e. OLS). I believe it is for continuous outcomes only, but you may want to look at "mvreg." mvreg is only really helpful if you want to compare coefficients across models (it's otherwise same to specifying outcome variables singly).

If you only have two outcomes, seemingly unrelated regression might work ("sureg").
Hope this helps!
- Nate

[Beth Feingold]

Hi there - I am trying to do a similar multivariate *logistic* regression and have not found a way to do it yet. at least, not without assuming proportional hazards, which i cannot assume with my data. did you find a work-around with your question here?

Thanks,
Beth

[email protected]

[Stephen Jenkins]

There are multivariate probit models -- that is, you can proceed straightforwardly if you don't insist on the logit link. As Maarten indicated a while ago the key issue is whether you allow for correlations in error terms across equations. On MV probit models, see the articles by Cappellari and Jenkins in the Stata Journal in 2003 and 2006 (articles freely downloadable from SJ site), with associated code provided on SSC and/or SJ. See also David Roodman's very versatile cmp program on SSC