Have an idea?

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

Model efficiency for allocation based conjoint with one concept (and none option) per task


I'm trying to use Sawtooth to create a conjoint exercise which will show one concept per task and a none option. The respondent can allocate a 100% between the two - so that means the 'None' option will be 100%-(the % he/she gives to the concept shown). The total number of respondents are 200.

When we are trying to generate and test the design with (say 200 versions) the std. error is ~0.08. The number of tasks is around 10-15.  This standard error is quite high and we are aiming for something below 0.05.
Is this because of the fact that we are showing just one concept per task?

Also, where can we see the 'D-efficiency' in the 'Test Design' window?
asked Oct 31, 2018 by Shruti

1 Answer

0 votes
Interesting questions.  First off, doing allocation for a one-concept CBC vs. the None is very similar to doing Paul Green style ratings-based conjoint from the 1970s.  But, doing this within our CBC software will use the MNL (logit) equation and maximum likelihood estimation rather than using OLS with R-squared (which you could do using our separate CVA conjoint analysis package).  The results should be pretty similar.  And, because you are using multiple versions (blocks) of the design, it will support the estimation of first-order interaction effects, whereas the original approach by Paul Green typically featured just one version of the questionnaire and could only support main effects.

Please note that when you run "Test Design" in our CBC software, we generate random responders to estimate the standard errors and these random responders are generated to be making discrete 1/0 choices of the alternatives.  We ignore whether you are doing chip allocation or discrete choice and assume the worst case scenario that people pile all their chips on their preferred offer.  So, the 0.08 standard errors you are seeing are really worst case situation where respondents do not use a continuous probability scale, but answer either left or right.

We report relative D-efficiency in our design report, listed as "Strength of Design".  This value is proportional to D-efficiency, but is not placed on a 0-100 scale.
answered Oct 31, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (195,590 points)