Hi!

I've been working on a discrete choice experiment with a mixed logit model. The issue is, however, I'm getting the message saying "some variables are collinear - check your model specification" and I have no idea why I am getting this issue.

My model has the following format:

mixlogit choice, rand(brandA brandB brandC exc percentage price nochoice) group(group) id(qid) burn(10000) nrep(100)

each attribute's levels are as such:
brand A -> 1,0
brand B -> 1,0
brand C -> 1,0
exc -> 1,0
percentage -> 0.4, 0.7, 1.0
price -> 0.6, 1.0, 1.4
no choice -> 1,0

Here, each respondent answered 8 choice situations, each composed of 3 alternatives with 1 opt-out option (i.e., no-choice, "I rather not choose any of the alternatives"). In the case of no-choice, I assumed "0" value for all attributes to represent a situation with no utility change. and this is why I did not use a reference attribute in the brands. (i.e., not including brand A to set it as the base level)

Thus, the regression for no-choice situation would be:

1 = 0*brandA + 0*brandB + 0*brandC + 0*exc + 0*percentage + 0*price + 1*nochoice

where the choice variable, y, is 1-0 dummy.

So, the coding for brands looks like this:

d1, d2, d3
A 1 0 0
B 0 1 1
C 0 0 1
X 0 0 0

where X represents the opting-out situation.

Well, this strategy doesn't really work because of unusually high VIFs (according to "collin", the VIFs are well over zillions)

I feel like I'm missing something here. What do you think I should look for?

(Also, I switched the format to effect-type coding as well, and the collinearity issue was still present.)
(Note that I haven't included all the variables for simplicity; actually, I have total 4 categorical variables with 3 levels each, 2 dummies (including "no choice"), and 2 linear variables)

Apologies if my posting is too messy to comprehend. I tried my best to make it as intuitive as possible.

Thanks!