Have an idea?

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

CCEA clustering for relative importances

Dear all,

I would like to conduct a clustering analysis using CCEA (ensemble). I want to have clusters that show how the relative attribute importances vary between them. Just want to make sure that the procedure is correct:

1. Export utilities in Lighthouse using ZCD with the option "include individual level importances"
2. Clean .csv file: remove all attribute level columns and NONE column
3. Import into CCEA software and use default settings for ensemble solution ("Standardize variables" and "Center cases" not selected).

Is that correct or did I miss something?

Also, I noticed that the best reproducibility figure is best for 4 groups. However the figure seems low (close to 90%). Do you think it is still acceptable or is that figure too poor?

Note: I am not using Latent Class clustering because I am using ACBC
asked Apr 24 by danny Bronze (1,310 points)
edited Apr 24 by danny

1 Answer

0 votes
Danny, your process is correct.  As for the 70%, it’s probably a little lower than I usually see, but segmentation studies are so different one to the next that it’s hard to say for sure.  Check the pseudo F statistics and maybe drop attributes that have low F statistics and rerun.  That often improves reproducibility.  Also profile your segments on other variables to see if they appear valid.  If your sample size is large, there are other ways to validate that we can discuss.
answered Apr 24 by Keith Chrzan Platinum Sawtooth Software, Inc. (105,750 points)
Yes, if you're using ACBC, CCEA is the way to go, not Latent Class
The suggestion about removing low F variables wasn't based on using conjoint importances.  But if you do apply it to conjoint importance data, it's not going to ruin anything.
The only thing is that removing one attribute leads to a sum that is not 100 anymore, in all segments
If I segment on importances I get a reproducibility of >90% for 3 groups, if I segment on part-worths I get only around 70% for the same amount of groups (highest rate among all segment size options). 70% doesn't sound like a good number to me though... (idea was to compute relative importances from the part-worth segmentations).
I guess I should rather run on relative importances then...?
I don't think the basis variables for your segmentation NEED to sum to 100%
...