Have an idea?

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

Latent Class Segments & Demographics


I have conducted a maxdiff latent class analysis using the stand-alone program on my profile case (case 2) maxdiff study and still can't decide whether to choose 2 groups or 3 groups as a better explanation of the sample. Can you help me with that?

Also later on when this is decided, how can I extract the demographics related to each group separately and then compare them together to check if there any statistically significant difference between them? is there an option or function that can separate them in the Lighthouse program or the standalone program? would chai square test between these two groups be enough for the comparison or regression is required?

N.B. can I use the Reported importance of attributes to understand how each segment weighs the importance of each item, then add the values for the items that represent the same dimension to get how important this dimension for these segments directly?

Thanks alot
asked Feb 25, 2020 by AMYN Bronze (2,980 points)

1 Answer

+2 votes

For the 2 groups vs 3 groups solution, you need to look at the goodness of fit measures (particularly BIC and CAIC) and then at the "story" that the between-segments differences in utilities tell in both of your solutions and decide which makes most sense for your study.   

Once you've decided on a solution, then you can export the modal segment memberships, which will be categorical variables, and then run chi-squared tests on any categorical profiling variables and ANOVAs (or t-tests) on any metric profiling variables to see how strongly the latent classes relate to the profiling variables.  

Be careful about how much adding you do, because even the profile case of MaxDiff doesn't necessarily support an additive composition rule.
answered Feb 25, 2020 by Keith Chrzan Platinum Sawtooth Software, Inc. (116,875 points)
Thank you very much, Keith, for your reply.

I had also watched your webinar about maxDiff on Youtube yesterday, and it was beneficial and to the point, especially The questions asked.

I have probably decided on the three groups solution, but I am not sure what do you mean by "modal segment memberships"?

What I understand is that I need to export the respondents to be classified into three groups (like adding a new column indicating which group they are in), then collect the averages (or counts) of all the profiling variables for each group, and finally test for differences using ANOVA or Chi-square. The issue here is in the first step, what is the quickest way to do this?

On the other hand, Can I use the (Rescaled for comparison utilities) to compare the differences between the utilities of the 3 groups using ANOVA, or should I use a non-parametric test like K-W ranked test?


The software already produces a single variable telling you which segment has the highest probability for each respondent - that's what I meant by the modal segment membership.  

And yes, you could use the utilities rescaled for comparison using ANOVA if you had run HB and had respondent-level utilities.  For LC MNL utilities you have your coefficients and you can ask for standard errors and you can look for differences as you would with t-tests.