Have an idea?

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

Interaction effects in ACA


I want to do an ACA, but I'm not sure about how to best account for possible interaction effects in the research design.

As I'm expecting many different interactions between attribute(s) (levels) composite attribute levels are not an option. A technical paper on ACA says that in that case "the most common solution is to evaluate interactions by pooling data from many respondents". However, I'm not quite sure about what this means and if there are any other recommendations for the research design in favor of calculating interaction effects - I'm thinking about measures like deliberately fostering overlaps between hold outs (as it is recommended for ACBC) and/or including a certain number of attributes in the hold outs.

Does anyone have any experiences with such a case and can give me recommendations on that issue? Thank you very much for your help in advance!
asked Apr 7, 2016 by anonymous

2 Answers

0 votes
Sorry, but there isn't a way to turn on interactions for an ACA exercise in our software.  I'm guessing that technical paper is referring to doing some reworking of the data where instead of treating something like 10 respondents answered 10 questions, you rework the data so it looks like 1 respondent answered 100 questions and you estimate a single set of utilities for everything.  Since ACA is a partial-profile method, the data would be very sparse for looking at interaction effects, so I'm guessing this is probably the only way to do it.

In general modifying ACA data is very, very messy and not something I would recommend to someone.
answered Apr 7, 2016 by Brian McEwan Gold Sawtooth Software, Inc. (37,085 points)
0 votes
I would add that since ACA is a partial profile method, there is some doubt about what an interaction would even mean.  If Attribute A is present in a question and Attribute B is not, the value for the A-B interaction will depend on some assumption the respondent makes about the level of Attribute B.  Different respondents might assume different (within-question constant) levels for B, and these are not captured anywhere in the data set.  So the by their very nature partial profiles cause spareness and indeterminacy in defining interactions.
answered Apr 7, 2016 by Keith Chrzan Gold Sawtooth Software, Inc. (48,525 points)