Have an idea?

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

Running HB for different groups together and doing a segmentation


I am doing a Maxdiff analysis for 3 different groups.

I have read that I can do one of both:
1. Do the HB analysis separately for the three groups
2. Do the HB analysis together for the three groups, but use a covariate to indicate the difference between the groups.

I am now wondering, if I do the HB analysis together using the covariate, is it also correct to run a segmentation on all data together?

We suspect that the strategy of the client, for using the segments, will be the same over all groups - as otherwise there will be too much segments to work with, so it might be more logical to do the segmentation over all groups instead of separate per group.

Many thanks
asked Sep 16, 2021 by Tina Van Regenmortel (415 points)

1 Answer

0 votes
I'll answer this question assuming you had a single MaxDiff exercise that all three groups saw, as that would be the normal situation.  (Though in a separate question you posted it seems this was not the case.)

If you are trying to develop segments based only on the importances/preferences of the MaxDiff questions (revealed based on the MaxDiff questions), then the recommended approach is not to use HB followed by cluster.  Rather, the recommended approach is to directly use latent class MNL.

However, if your aim is to segment the data by the three different groups, then running HB with the covariate will bring out more separation between the groups because each group will show Bayesian shrinkage toward that group's preferences rather than toward the generic population including all respondents.
answered Sep 16, 2021 by Bryan Orme Platinum Sawtooth Software, Inc. (198,715 points)
Thank you for your answer Bryan.

I am indeed running a single MaxDiff exercise for the three groups.

I momentarily forgot yesterday that I could indeed just run Latent Class within Sawtooth instead of first running HB and then clustering.
Can I assume that there is no issue in running Latent Class for the three groups together? As respondents from the different groups that answer similarly will just be clustered together? I shouldn't be thinking about adding covariates to that analysis?

Just one question, you say "if your aim is to segment the data by the three different groups", what do you mean by this? Do you mean creating a different segmentation per group? Because that is just what we are NOT looking for.

Many thanks