Dear Sawtooth Forum, dear Bryan, dear Keith,
I am conducting an ACBC and want to cluster / segment the data. However, I am unsure which method to use for this. I read Keith's article on segmentation here (https://sawtoothsoftware.com/resources/blog/posts/segmentation-how-to-do-it-badly-and-well), manuals, several forum posts, but they don't clarify the problem fully for me.
My data: the attribute levels of my variables are all either ordered or unordered categorical (e.g., i) feature does not exist vs. feature does exist; ii) similar vs. different but complementary vs. different but not complementary; or iii) low vs. medium vs. high). I let respondents choose attributes regarding a situation and not a product or other object, hence the need for categorical variables. These variables are my basis variables for the cluster analysis.
I have heard from other researchers that conducted ACBC clustering in 2017-ish that the CCEA's ensembles technique was their preferred choice for such data. They basically followed the steps outlined by Bryan here (https://legacy.sawtoothsoftware.com/forum/7858/categorical-variables-in-ccea?show=7858#q7858) and here (https://legacy.sawtoothsoftware.com/forum/30269/demographics-in-ccea-clustering?show=30269#q30269).
However, I read opposing views that suggested Latent Class Clustering instead:
a) ""Categorical variables are generally not appropriate for use within CCEA. For example, a variable such as "preferred_color" where 1=blue, 2=red, 3=green, etc. would not be appropriate for use in CCEA. CCA expects increasing values to indicate "more" of the variable and decreasing values to mean "less." Such is not the case with categorical variables.
Although one can cluster on individual-level utilities resulting from an HB analysis of Choice-Based Conjoint or MaxDiff (best-worst scaling), it is probably more appropriate to utilize Latent Class procedures for these cases. Using CCEA would involve the two-stage procedure of first computing utilities using HB, and then secondly using those data within CCEA (where any errors in the first stage would be accepted as "truth" in the second phase). Latent Class provides a way to simultaneously estimate part-worth utilities and divide the sample into meaningful segments." (CCEA v3 Manual, Software for Convergent Cluster & Ensemble Analysis (Updated June 9, 2008))
b) "I think this approach [CCEA's ensembles technique] is somewhat "clugey" compared to the more sound Latent Class (via Latent Gold) or PAM solutions for this." (Bryan in one of the forum posts I linked to above).
So, my questions are twofold:
1) Which clustering method is most suitable for the data (ordered and unordered categorical) I have?
2) Do you have an approximation / orientation value from past experiences about the differences between the two approaches, i.e., do they yield different results or are the clusters basically the same, just with slightly different values for the attribute importance.
Side-information: I have an Academic Research Grant and am using Lighthouse 9. I have access to Sawtooth's CCEA software, but unsure, if the Latent Gold access is included, too.
If you need any more information to answer the question, please let me know.
Thanks a lot for your help in advance!
Best,
Nathan