Have an idea?

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

Clustering individual-level utilities from CBC/HB


I used CBC/HB to compute the individual-level utilities (at respondent level). My question is: is it possible to use this data to compute segment-level utilities ?

For example, if I filter all the rows from the "*_utilities.csv" file where respondents are male, how can I compute the segment-level utilities for all male respondents ? Would taking an average work ?
asked Jun 12, 2014 by anonymous

1 Answer

0 votes
There are a few layers to this question, from the quick-and-easy answer to deeper statistical issues.  I doubt you want to hear the deep issues, but I'll post a few comments in case others are tuning in.

Easy answer: just take the average across the respondents who are males.  That's the segment summary utility for males.

More complete answers...

1.  The "scale factor" is a thorny issue involving CBC data and can make it hard to compare one respondent or group of respondents to another.  If respondents have high error, their utilities are uniformly squished somewhat closer toward zero.  If respondents have low error, their utilities are uniformly expanded larger.  So, if you try to compare a utility from one respondent group to another respondent group (and if their response error differs substantially) then what you think you see as a utility difference may just owe to the scale factor expansion/contraction.

So, to help reduce the possibility of running into these troubles, we'd recommend you compare respondents using the normalized utilities (the zero-centered diffs) as can be exported from our SMRT simulator package (under Analysis + Run Manager + Export).  The normalized utilities put all respondents on the same scale (or at least the scale that is assumed if we make sure that the sum of the differences between best and worst levels across respondents are constant).

2.  Some academics would prefer that if you plan to compare respondent groups that you include those respondent groups as covariates in the HB run.  This is an advanced option.
answered Jun 12, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (172,790 points)
Clustering choice data to fit HB ?
Thanks a lot !
I followed your recommendation for 1 and used the zero-centered diff utilities, but now have a follow-up question. If I decide to filter respondents in one segment (e.g. males) and do the average utility / importance, how would I check for the statistical significance of the result ? What I mean by that is: is there a way to check for the quality of the sample / result, i.e. if the respondents from the segment that I choose truly have similar preferences ?
Thank you !