Have an idea?

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

Segmentation with HB

How segmentation is done applying HB model? Are the parthworths utilities calculated at individual level and then in the second step clustering methods are used to identify the segments? If yes, does the background descriptor (identifier variables) such as age or income included as a variable in cluster analysis or clustering is done only on partworths utilities?
asked May 28, 2017 by Robin59 Bronze (565 points)

1 Answer

+2 votes
Robin,

We see folks do segmentation lots of different ways:
 
Some folks use latent class MNL to create segments but then they use their HB utilities to run simulations and so on.

Other folks put the HB utilities into cluster analysis to create segments, either alone, to create a segmentation that is just needs-based, or in conjunction with other potential segmentation basis variables.   They might use cluster analysis for this or sometimes they use latent class/finite mixture modeling to create the segments.

Sometimes folks do separate segmentations, some with the utilities as basis variables, others with the non-utility variables as basis variables and then they combine all those segmentation solutions into a cluster ensemble.  

Or folks might mix and match the above in an almost endless variety of ways using some other sort of custom ensembles in cluster ensemble analysis.  

Although there are a very large number of ways one COULD go, I find it easier to stick with a simpler method, since ultimately you may be called upon to explain what you've done to a non-technical audience.  For this reason I usually either use latent class MNL to make my utility-based segments or I use cluster analysis  on the HB utilities.  On rare occasions I'll build custom ensembles of segmentations, some done with utilities and some done with other variables.
answered May 28, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (58,750 points)
Thanks Keith,

You mentioned "For this reason I usually either use latent class MNL to make my utility-based segments or I use cluster analysis  on the HB utilities." I think you do not include background variables such as age or gender in forming segments and it is done purely based on utilities. The risk of this approach is that the identified segments may not be able to be explained by background variables. For instance in case of car, you may find a segment that like red. But non of the background variables are dominant to explain this segment. How do you tackle this issue
...