Have an idea?

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

Calculating D-efficiency for CBC

Hey there,

I conducted both an ACBC and CBC study for a research project. I am now  required to document the D-efficiencies separately for both methods. I used the Sawtooth Software's "Test Design" procedure and found D-efficiencies for ACBC (for each respondent, which is absolutely fine), but no corresponding value for CBC.

For CBC I also included the OLS efficiency test and received efficiency-values for all attribute levels. So my question is: Is it possible to use these efficiency values to calculate a corresponding D-efficiency parameter for CBC? And if not, how might I get this?

I would be very pleased to receive your help!
related to an answer for: Efficiency of CBC Design
asked Jan 18, 2019 by Verena

1 Answer

0 votes
Hi, Verena,

At the bottom of the results print out for Test Design you will see a measure called "strength of design."  This is our D-efficiency number.  It is NOT a relative D-efficiency (which would be the D-efficiency of your design compared to the D-efficiency of a theoretical ideal design).
answered Jan 18, 2019 by Keith Chrzan Platinum Sawtooth Software, Inc. (95,775 points)
Hey Keith,
Thank you very much for your fast response (and sry for my delayed one)!
I already thought that this value might be the D-efficiency number, however, I have trouble to interpret this value. When having relative D-efficiencies it seems plausible that, for example, a modified survey version  reaches only 80% of another one's efficiency. But how might I use this value for comparing CBC's efficiency with ACBC's efficiency? Is a comparison of these two different methods only feasible by comparing standard errors?
Thanks again for your help! :-)

ACBC D-efficiencies are for individual respondents, so they are not comparable to the D-efficiency of a CBC design, which is the efficiency of the overall design (all versions, all respondents).

One uses D-efficiency in a CBC design to anticipate which designs will have lower standard errors than other designs.  With ACBC, standard errors make more sense at the respondent level (I may discriminate very little in a color attribute and you might discriminate very little on price, so in my design we might expect larger standard errors on color (where they are relatively harmless) and on yours we might expect larger ones on price (where, again, they are not very harmful, since price isn't  something you care that much about.  The goals in an ACBC (minimizing the important errors for each person) are different than for a CBC (minimizing the total error by equalizing standard errors for all attributes) so it really doesn't make much sense to compare them in the first place.
Alright, thank you very much for your detailed responses!
I will keep that in mind when further working on the project.