Have an idea?

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

CBC analysis based on groups

Dear all,

fpr an acamedic project we conducted a CBC study and analyzed the results from HB estimation pooling all respondents together. Although we have a small sample size (120) we would like to have a closer look at potential respondents groups. We read that in a state-of-the-art approach we would need to conduct Cluster analysis but we wonder if there is a less "demanding" and more practical approach to show what you can do with CBC data in addition to what we have done so far (analyzing the data of all respondents simultaneously). As it is a pilot study we are not interested in testing significant differences between groups like Cluster analysis does. However, we hypothesize there might be a group of respondents that is more brand-oriented and a group that is more price-sensitive. What could we do with our data now?

Thanks a lot for your help!
asked May 2, 2016 by Joanna

2 Answers

+1 vote
I would strongly recommend latent class to you.  Our CBC software (under either SSI Web or its new name "Lighthouse") has latent class built right into the interface.  It's one of the analysis options.  It finds groups of respondents who are similar within groups but different between groups.  It reports logit utilities and importance scores broken out by the groups.  It generally has stronger statistical foundations than cluster for choice data.  You should investigate 2-group, 3-group, 4-group etc. solutions.  With only 120 respondents, I probably wouldn't investigate solutions beyond that.

Would be very appropriate for developing hypotheses regarding n=120.
answered May 2, 2016 by Bryan Orme Platinum Sawtooth Software, Inc. (154,405 points)
Interpretation of Latent Class results
0 votes
A closely related strategy, which might be useful given your relatively small sample, is mentioned in Bryan Orme's method text.  Accordingly, use HB to compute individual utility coefficients.  Separately, use Latent Class to generate segments of respondents with similar preferences.   Select a Latent Class model (probably 2 or 3 classes) which minimizes BIC and yields a conceptually useful solution.  Then use the posterior probability of group membership to determine the most probable segment for each participant.
answered May 5, 2016 by cunnic Bronze (1,440 points)