Have an idea?

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

CBC Latent class - how to segment best?

Hello, I've got a problem to chose the best segmentation for my CBC via latent class estimation.

My report says that:

Minimum number of groups    2
Maximum number of groups    5
Number of replications    5
Maximum number of iterations    100
Convergence limit for log likelihood    0,01000
Random number seed    1
Null log-likelihood    -3465,73590

The summary of best replication includes the groups (2,3,4,5), replication (1,4,3,1), log-likelihood (constantly decreasing from group 2 to 5), percent certainty (increasing), AIC (decreasing), AIC (decreasing), CAIC (increasing), BIC, ABIC (decreasing), Chi-Square (increasing) and relative Chi-Square (decreasing).
How can I derive which segmentation is the best for my study?   

Thank you very much!
asked Aug 27, 2015 by anonymous

1 Answer

0 votes
"Best" is really a subjective term :)  Clustering is almost always a mixture of art and science (unfortunately!).  The Latent Class technical paper (http://www.sawtoothsoftware.com/support/technical-papers/sawtooth-software-products/cbc-latent-class-technical-paper-2004) has a section with a numerical example and a discussion about choosing the number of segments that might be helpful.  A multi-pronged approach of looking at the fit statistics, checking to see in the group counts if you are just splitting previous groups up, whether or not one group has more managerial relevance, etc. is probably the "best" approach to clustering.
answered Aug 27, 2015 by Brian McEwan Gold Sawtooth Software, Inc. (38,015 points)