Have an idea?

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

How to calculate level importances for CBC results

I'm conducting a CBC study for a client and they've requested, in addition to atttribute importances, a report of level importances within each attribute, expressed as %'s and totaling to 100% within attribute.  Apparently this is something they've received from other vendors before, but is not something I've ever seen (I typically report out the ZC-diffs or run sensitivity analysis on the base case to highlight relative level values).  

My thinking of how to approach is to exponentiate the raw utility of each level of a single attribute, add them together, and then divide exp(A1L1) by this sum in order to get a % importance of A1L1 within A1.  It seems like one benefit of this approach is the valence of the raw utility would be taken into account, with negative utilities getting the smallest % importance (it seems like often clients are a little confused by valence/directionality when we talk about attribute importance).

Are there downsides to this approach that I'm not thinking through?  Or is there a commonly-accepted method of calculating level importance out there that I just haven't seen before?  Many thanks!
asked Jun 23, 2019 by Liz

1 Answer

+1 vote
Hi, Liz.

I think the downside is just the oddity of calling the resulting percentages "importances," because of course that's not what they are.  But if your client likes that name and finds it useful, it doesn't cause any harm.  Just exponentiate the levels of a given attribute, sum the exponentiated utilities and divide each exponentiated utility by the sum of the exponentiated utilities.  

Say you have a 3-level variable with utilities of -1, 0, and 1.  Exponentiated those are 0.368, 0.00 and 2.72 and the resulting shares would be 9%, 24.5% and 66.5%, and you could report those to your client.  I would probably calculate these at the respondent level and then calculate the average across respondents.   

I suppose you could think of these percentages as normalized odds ratios.  

This is essentially a share simulation as if you were simulating as many alternatives as you had levels of an attribute, where all other attributes were held constant and the only difference between alternatives is the levels of the one attribute.
answered Jun 23, 2019 by Keith Chrzan Platinum Sawtooth Software, Inc. (84,975 points)
edited Jun 23, 2019 by Keith Chrzan