collinearity problems with mixed logit

[Miriam Prades]

Hi,
I have a problem with the mixed logit model to assess the answers from a discrete choice experiment (n=90). I have 7 attributes: 5 attributes with 4 levels and 2 attributes with 2 levels. We constructed an orthogonal and balanced design composed by 8 scenarios. I use the mixlogit command and STATA returns the following error message: "Some variables are collinear - check your model specification".

It means that attributes are highly correlated and could be due to the amount of levels from attributes. My question is, how can I manage properly this situation to solve the problem? And another important question, could I have avoided this problem?

Has anyone had the same problem with mixed logit?

Thanks

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of a helpful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

The first thing to do is to diagnose the colinearity. Since it is probably a problem with the x's , you can look into colinearity using regress which is easier that mixlogit and has more diagnostics available. Try running this as a regression and see what happens. If you find a variable that doesn't estimate, try regressing that on all the other x's. There are also colinearity diagnostics available in regression. Also, look at the correlation matrix. While it is possible that mixlogit generates colinearity some way that regress doesn't, I'd bet on regress helping you diagnose your problem.

Once you figure out the problem, then you can worry about how to fix it. You might also check if there is a Stata routine that would suffice - they are often better tested than user written routines.

[Miriam Prades]

Thanks Phil

[Karina Breis]

Hey there,
I am facing a similar problem at the moment so I was wondering if you Miriam have a hint on how to solve this.
I already ran a regression to detect the problematic variables.
Any hint would be appreciated.
Thank you!