Have an idea?

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

understanding CBC results with interactions

hey there,

My CBC consists out of 4 Attributes. Attribute 1 (landscape types) has 5 levels, whereas Attributes 2 (wind), 3 (pv), and 4 (powerlines) are converted to linear Attributes.  I also integrated an interaction effect of Attribute A1 and A2. Lets assume that everything is significant. The aggregate Logit results show me following effects (e.g):
A1.level1: -0.7
A1.level2:  +0.4
A1.level3: -0.2
A1.level4: -0.1
A1.level5: +0.6

A2: -0,4
A3: -0.1
A4: -0,3

Interaction: A1.level1 x A2: +0,4
...

It is shown, that although A1.level1 and A2 show negative effect on the utility, the interaction of both provide positive values. How is that possible and how to deal with it?

Thanks four your support!
Regards,
Boris
asked Dec 12, 2018 by bs77 Bronze (730 points)

1 Answer

0 votes
The interaction effect is the adjustment to be made after considering the main effects.  So, for A1_Level1, the main effect of attribute 2 plus its interaction effect leads to no net utility change due to changes in attribute 2.  That's because the -0.4 of the main effect for A2 is adjusted by +0.4 for the interaction effect.  

You'll want to check the precision of your estimates, of course, before drawing firm conclusions.  From aggregate logit, you'd hope that the standard error for the main effects was 0.05 or less.  And, you'd hope that the standard error for the interaction effects is 0.1 or less.

Also, recognize that interaction effects found in aggregate models are often due to unexplained heterogeneity in preferences.  Once HB estimation is employed, these interaction effects usually are explained (captured) via the heterogeneity of tastes in the main effects.
answered Dec 12, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (163,615 points)
Thanks a lot. That was really helpful.
So for instance, if the interaction effect A1.level1 x A2 would be -0.2 instead of -0.4 the utility would rise for 0.2 in total, which means the interaction effect would reduce the negative preference. right?
...