Have an idea?

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

How many choices are needed to reliably calculate an interaction effect and how can reversals of partworth be explained?


my question concerns the calculation of interaction effects with CBC-HB.

I ran an HB analysis including an interaction of two attributes (A, 4 level and B, 2 level).  The results of all main effects as well the most interaction terms were as expected unless for one interaction term (say A1 and B1/2). Here the results are reversed, meaning that while the utility should be positive for A1-B1 and negative for A1-B2 it is the other way round. For all other levels of A the interaction with B1 is always positive and with B2 always negative.

As A1 is the most unattractive level of the four, is it possible that the respondents did not choose a stimulus with the A1 level  frequently enough to calculate an interaction effect? I counted how often the A1 level was actually chosen from a respondent and endet up with 711. Is that a sufficient base to calculate the interaction? Could this explain the reversed utilities? And if not, which other explanations could be possible?

Thanks a lot for your advise!
asked Mar 14, 2017 by anonymous

1 Answer

0 votes
These are difficult questions to answer.  The amount of data needed to successfully measure an interaction effect depends strongly on how important the interaction effect is to begin with.

Sorry to state the obvious, but you need to interpret interaction effects in light of main effects.  So, please consider the net effect of main effects for the attribute levels involved plus their interaction effects before drawing conclusions regarding whether you think the interaction effect has face validity.
answered Mar 14, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (148,340 points)