Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Can I study interaction between attribute levels and respondents' characteristics?

Hi, I would like to study the interaction between preferences for attribute levels and socioeconomic characteristic (e.g. ethnicity, income, age group etc). We hypothesize that preference utilities are different among ethnic and income groups.

Can I perform this in CBC/HB or logit? Or do I need to export the utilities and run the analyses using a statistical package for instance STATA?

Thank you!
asked Jan 13, 2017 by Xin Yi

1 Answer

0 votes
Dear Xin,

You could actually do this using CBC/HB or logit (you would run logit in our latent class module as a model with one class).  

The latent class model with a single class is a general purpose logit package.  You will need to export your data and add effects or dummy-coded demographic variables, which will vary across but not within individual respondents.  Then just run the model and interact the demographics with your design variables.  

With HB it might be easier to be more exploratory.  First, run your HB analysis.  Now, using the zero-centered diffs utilities as your dependent variables, run analyses of variance to see whether those utilities differ significantly by demographics.
answered Jan 13, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (62,700 points)
Dear Keith,

I have done exactly what you recommended for HB, but I used Kruskal Wallis instead of ANOVA because the zcd utilities do not appear to be normally distributed. However, my total sample size is 300. For ethnicity, it's 100 per group. For income, min 49 and max 163. By central limit theorem, can I safely assume them to be normal?

Also, I do not quite know what to select to run latent class as a single class and interact demographics with attribute levels (what I can see is interaction between attribute levels). Will you kindly advise? Thank you very very much.
Dear Xin,

For the logit, what software are you using?  It would be easier to run if you do it through our Latent Class software, because then you need only recode the design matrix to add the additional demographic variables.  I am not sure there is an easy way to do this within our integrated Lighthouse software.
Dear Keith,

I was using latent class in Analysis Manager of Lighthouse Studio. I have since downloaded latent class module. I created a new project with the design file, data/response file as well as demographic file. How do I recode the design matrix to add the additional demographic variables? Also, do I fix the min and max number of groups as "1" since I'm running it as single clasS? Thank you!
Let's say you have gender as a demographic variable.  Add a column for gender, coded 1 if male, 2 if female.  Respondent Han is a male, so for all alternatives in all of his choice sets, the value is 1.  If Jen is a female, then for all of her alternatives in all of her choice sets, she gets a value of 2.  If you have a labeled design, then it may make sense to get a main effect for gender.  If not, then you will have main effects for all of your other design variables but for your demographics, just compute their interactions with the non-demographic design variables.  Then see if these interactions add significantly to the log likelihood of your model.  I would do an omnibus test first to see if ANY are significant before running a separate test for each demographic variable, to control experimentwise error.  

And yes, fix the minimum and maximum number of groups at 1 - that's how you trick the latent class software into running an aggregate MNL model.
Thanks so much for your advice! Will work on it.