cmset choice models in Stata 16

[Elena Mvk]

I am trying to figure out the new choice model (cmset) analysis for results of the discrete choice experiment (DCE).

I have a panel data set with unlabeled alternatives: that is my respondents saw two options per choice-set, there were no alternative specific variables, the attribute-levels varied across choice sets. Each respondent saw 12 choice sets.

How do I run this command9CMset) with unlabelled choice tasks? I would like to use the cmclogit and cmmixlogit.
So far I tried:
cmset id scenario(the 12 choice sets), noalternatives
cmset id gr-var, no alternatives

Also, is there an option for latent class analysis under cmset ?

Thank you.
Kind regards,
Elena

[Sarah Thorne]

Hi Elena,

I think I am having a similar problem. I have two unlabeled alternatives and have not been able to run cmmixlogit or cmxtmixlogit with the "nocons" option specified. Does anyone know if there is a way to run a random parameters logit model without an alternative-specific constant using the new choice modeling package?

Thanks!

[Seema Kacker]

Hi Elena,

Did you ever figure this out? I am in a similar situation, but am unclear on if the noalternatives option is applicable: I also do not have labeled alternatives, each alternative is simply characterized by its attribute levels. I also have panel data (12 choice sets per person, with 2 alternatives in each choice set). If the noalternatives option does apply, is it possible to use margins to plot predictions across a case-specific and alternative-specific variable?

Thanks
Seema

[Emmanuel Teitelbaum]

Hi Everyone,

I have a similar issue, although in my case I have panel data with three unlabeled alternatives and a "none" option in each choice set. I am not a statistician, but here is what I have been able to learn through my own research and my interactions with one Stata representative via email:

1) For experiments with labeled alternatives you should analyze your data using McFadden's choice model, which includes alternative specific constants, but for unlabeled alternatives you should use conditional logit without alternative specific constants. There is a discussion of this HERE on ResearchGate with a few helpful references;

2) cmclogit runs McFadden's choice and is therefore designed for experiments with labeled alternatives. cmclogit is the new version of asclogit, which stands for "alternative specific conditional logit." If you try to run cmclogit with no alternatives, you get an error message. If you run cmclogit and specify your choice set as the alternatives when they are unlabeled then Stata will perform the estimations; but when you try to use the "margins" post-estimation commands the predictions you get will be meaningless;

3) clogit is the appropriate choice for analyzing unlabeled alternatives, but there are difficulties in using the margins command to interpret the results. There are discussions of this issue on the Statalist HERE and HERE. A solution involving a wrapper for xtlogit (aextlogit) and average semi-elasticities is proposed;

4) For mixed logit models, one would think that the "noalternatives" option would allow you to run models with unlabeled alternatives. However, when I try to use margins after cmxtmixlogit after first specifying the "noalternatives" option in cmset, I get an error message: "alternative variable not found....r(301)." Also, it takes about ten times as long to run my models using cmxtmixlogit as with the older user-written mixlogit by Hole (2007);

5) For calculating predictions following mixlogit, I found it helpful to read and go through the example code offered in THIS publication by Lancsar et. al. (2017). There is also a discussion of how to calculate marginal effects and standard errors following mixlogit HERE. (One thing I don't understand about the examples in this post is why they are restricting the calculation of the marginal effect to the first alternative in their choice sets (e.g. "alt == 1"). I think this restriction is only necessary in the case of labeled alternatives?)

It would be good to get some clarity from an expert on these issues and, specifically, whether the new choice models in Stata 16 should be able to handle DCE data with unlabeled alternatives. If the new choice models are not able to handle unlabeled alternatives, this should probably be stated explicitly in the manual along with suggestions for how to analyze them.

Best wishes,

Manny

[Red Owl]

Emmanuel Teitelbaum

If you have not already seen it, you might be interested in post #2 at https://www.statalist.org/forums/for...fic-regressors

Red Owl
Stata/IC 16.0 (Windows 10, 64-bit)

[Sarah Thorne]

Thanks Emmanuel Teitelbaum

I don't quite follow what "no alternatives mean"; is this the same as having an unlabeled experiment with no alternative-specific constant? And so is the takeaway that cmclogit & cmmixlogit not the right tools for analyzing unlabeled choice experiment data?

Thanks!

[Elena Mvk]

Hi everyone, thank you for all your comments.

I am sorry for not replying sooner. I did figure it out, and it gives me almost identical results to a mixlogit command.

cmset rid seq option

cmxtmixlogit y, r( $_x X1 X2 X3 etc) noconstant

(Note: rid = the respondent's ID, seq = the pair of choice alternative (this DCE had 2 alternatives per choice) for the respondents (there were 12 choice sets per respondent, so 12 pairs), option = whether it was option 1 or 2 within the pair).

[Elena Mvk]

Hi Elena,

I think I am having a similar problem. I have two unlabeled alternatives and have not been able to run cmmixlogit or cmxtmixlogit with the "nocons" option specified. Does anyone know if there is a way to run a random parameters logit model without an alternative-specific constant using the new choice modeling package?

Thanks!

Hi Sarah, the r() in // cmxtmixlogit y, r( $_x X1 X2 X3 etc) noconstant // would specify the random parameters.

[Francois Dessart]

Hi everyone,
Thanks Emmanuel Teitelbaum for your valuable insights into this discussion.
I am trying to estimate willingness to pay with cmxtmixlogit in Stata, but I fail to understand how to do this. Would you or anyone else have experience or guidance regarding how to do this?