Have an idea?

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

CCEA - Inclusion of the Dual-None Option?

Hello Sawtooth-Community,

I draw on your CCEA software to perform a consensus clustering on part-worth utilities (zero-centered-diffs). So far, I just included the part-worths for the attributes and refrained from taking the utility values of the dual-none option into account.

For my understanding, the utility of the dual-none tells me the threshold value when answering the question whether the respondent buys a product or not. Hence, only if the sum of the individual utilities from the product attributes exceed the utility of the dual-none, the respondent will purchase the product under consideration.

As a consequence, I have no idea how to include the dual-none into the clustering since the utility values differ dramatically. For example, respondent A has a utility for the dual-none of 10.1 while respondent B has 0.4 and respondent C -3.7. Compared to the part-worths of the attributes, it seems that the dual-none has a different scale. Therefore, I have the feeling of comparing apples and pears. Is this correct?

However, since I want to calculate relative attribute importances for the obtained groups afterwards, I stick to the hint in your CCEA manual and do not perform a standardization. Is there any possibility to include the dual-none in the clustering without standardization or is it ok to include it with the current scale?

Many thanks to all of you

asked May 1, 2014 by Arnold

1 Answer

+1 vote
There are several tricky aspects to this question, including:

1.  Do you want to have the None parameter as a driver of the segmentation?  This will tend to lead to groups that differ quite a lot in terms of the way they use the None.

2.  Do you want to be using CCEA on the normalized part-worths resulting from HB instead of directly performing Latent Class analysis on the raw choice data?  Both approaches can tend to work well, but using Latent Class is a more direct approach that accomplishes the segmentation in a single step rather than two steps.  (Statisticians prefer single-step procedures).

3.  Regarding the scaling of the None parameter in zero-centered diffs, we find a multiplier for each respondent such that the differences between best and worst levels (for the attributes NOT including the None parameter) average 100 across the attributes.  Whatever that multiplier is per respondent, we also apply it to the None parameter.  So, the None parameter should be on an appropriate apples-to-apples comparison to the other attribute levels after performing the zero-centered diffs transformation.
answered May 1, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (131,890 points)