Have an idea?

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

ACBC Interaction: Significance between which levels?

Dear Forum,

I conducted an ACBC and analyzed attribute interactions with the 2LL Likelihood test. Three significant interactions were found and in the export I can also see the chart with the additional part worths for each two-way combination of attribute levels.

My question is, how can I see, which levels interact significantly?
So the 2LL test tells me, that there is at least one two-way group with significant different utilities, or am I missinterpreting the result?
Is there something like a post-hoc test module to investigate which two-way interactions are significant and which are not?

THANK YOU very much for your help.

Best regards,

asked Sep 26, 2018 by Chris Berlin Bronze (570 points)
edited Sep 27, 2018 by Chris Berlin

1 Answer

+1 vote
The 2LL test tells you whether adding an interaction between two attributes adds significant additional fit to an aggregate logit model.  It does not tell you which levels within this attribute have the most significant interaction effect (though I describe further below how to do this).  Furthermore, just because we might see a significant interaction effect in an aggregate logit model does not necessarily mean we will see a significant interaction effect once heterogeneity of tastes across respondents is captured via HB.  

If you are interested in knowing which specific levels that are interacting have the most significant statistical significance (largest t-value) under aggregate logit, then you should click on the "Runs" tab in the interaction search output.  Go to the run involving the interaction between two attributes that interests you, and then scan that logit report for the interaction parameter t-Ratios (they are found beneath the Main Effects).  The larger the absolute magnitude of the t-Ratio, the more statistically significant the parameter is (compared to the zero, or no effect).  Absolute value of 1.96 or larger is significant to at least the 95% confidence level.
answered Sep 27, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (163,515 points)