Have an idea?

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

Can I combine HB utilities with segmentation from LC and compare segments?

We ran a conjoint analysis and used LC to find two groups. Then with the segmentation, we used the HB utilities generated and made a boxplot to compare both segments.  However we got misleading results:

LC (part worth utilities  Rescaled)
Attr 1.  Level 1 -> G1: 22.43, G2 : 39.21
Attr 1.  Level 2-> G1: -22.43, G2 : -39.21

HB (Av. Utility Values, utility scaling zero-centered diff)
Attr 1.  Level 1 -> G1: 30.07, G2 : 25.49
Attr 1.  Level 2-> G1: -30.07, G2 : -25.49

LC (part worth utilities  Rescaled)
Attr 2.  Level 1: A + B -> G1: 14.9, G2 : 7.08
Attr 2.  Level 2: A + B + C -> G1: 16.63, G2 : 7.02
Attr 2.  Level 3: A + B + C  + D -> G1: 28.26, G2 : 35.79
Attr 2.  Level 4: No -> G1: -59.8, G2 : -49.8

HB (Av. Utility Values, utility scaling zero-centered diff)
Attr 2.  Level 1: A + B -> G1: 6.13, G2 : 6.58
Attr 2.  Level 2: A + B + C -> G1: 29.27, G2 : 0.92
Attr 2.  Level 3: A + B + C  + D -> G1: 26.65, G2 : 27.65
Attr 2.  Level 4: No -> G1: -62.05, G2 : -35.15

With the LC results, it seems attribute 1, level 1 has a higher utility for G2 but with HB it has a higher utility for G1.

Additionally, for attribute 2, using LC it seems that the difference between level 1 and level 2 for G2 is quite low but with HB it isn't the same result. Likewise, for G1 with LC, Level 3 > Level 2 > Level 1 but this is not the case with HB.

Taking this into account, which utilities should I use? Can I compare segments with HB utilities?
asked Aug 5, 2020 by Maria Peña (140 points)

2 Answers

0 votes

I wouldn't expect the LC utilities and the HB utilities to match perfectly.  For one thing, when you take an average of the HB utilities for the members of a class, you're putting whole respondents each in a single class.  So Jones goes in class 1 and Smith you assign to class 2 and so on.  But in latent class each respondent is classified partly in class 1 and partly in class 2.  In other words, the latent classes are weighted averages of the respondents, but in your HB mean utilities each respondent falls wholly into a single class.  

It depends a little on what your objectives are.  If you want to report in an academic way the results of latent class analysis, then I would use the LC-MNL utilities.  In marketing practice, we often use LC-MNL to identify segments, then assign each respondent wholly to her modal segment  and focus our report on the HB utilities and simulations based on the HB utilities.
answered Aug 5, 2020 by Keith Chrzan Platinum Sawtooth Software, Inc. (107,050 points)
thanks Keith.
0 votes
First, let's make sure we're on the same page regarding the potential differences between a latent class MNL run and an HB MNL run.

With Latent Class MNL, respondents are probabilistically assigned to groups.  So, for example, respondent #1 might have an 88% likelihood of belonging to group 1 and a 12% likelihood of belonging to group 2.  That respondent's data are used in a weighted way, proportional to probabilities of membership, to contribute to each group's utilities.  However, when we create segmentation variables for you to use as filters for post hoc segmentation using the individual-level utilities from an HB run, each respondent is assigned wholly into the group they are most likely to belong to.  So, respondent 1 would be wholly assigned into group 1 for the purposes of running an HB MNL model for the group 1 respondent.

From an HB perspective, each respondent's utilities are influenced not only by the respondent's own choices but by the utilities and covariances across the population.   So, there is some natural smoothing of the data for respondents or individual groups toward population tendencies.  The more choice tasks you have relative to the parameters to be estimated, the less the Bayesian smoothing.  

The issues above might be responsible for the differences you are seeing in the interpretation of the segments' utility scores when comparing LC MNL to HB MNL.

Next, I don't have a sense for sample sizes involved.  So, we don't know how much precision or not we're seeing in the utilities as we compare.

In sum, the result should be similar when comparing LC MNL utilities for groups to HB utilities for the same respondent groups (when respondents have been assigned to the segment they most likely belong to).  But, we shouldn't expect them to be identical.
answered Aug 5, 2020 by Bryan Orme Platinum Sawtooth Software, Inc. (191,015 points)
edited Aug 5, 2020 by Bryan Orme
Thanks Bryan.
The sample size was 495 respondents, G1: 311 and G2: 184.

I actually wasn't expecting results to be identical, but at least with a similar tendency. The fact that one group preferred more something and then it was the other group created the doubt.