Have an idea?

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

What values are displayed in the 'Charts' Tab of Interaction Search Tool

Hey all,

I am just looking at the Output of my Interaction Search of ACBC data. In the 'Charts' tab there are values given for each product of attribute levels of significant attribute interactions. I realized for each attribute combination the 9 panels add up to 0, so, they are probably some sort of zero-centered values. I am having difficulties in interpreting these values, though, when I compare them to the values of HB estimation including the significant interaction effects.

One example from my data:
Interaction search:
Salary_low x Task_unattractive= - 2.68
HB estimation:
Salary_low x Task_unattractive= 29.58

How can I interpret the effect on this level now, given that one time it's a negative value and the other time it's highly positive?

I hope someone can help me with this.

Thank you in advance!
asked Jan 31 by Helen

1 Answer

0 votes
A few preliminary things to know about this:

1.  Interaction search tool uses aggregate (pooled) logit.  HB is individual-level estimation.

2.  Raw utilities for logit or HB are on the logit scale (meaning that exponentiating the scores leads to values that are proportional to choice likelihood).

3.  Other reports from our utility estimation routines can show "zero-centered diffs scaling" which places each respondent's utilities on a much larger scale (compared to raw logit scaling) such that the average difference between best and worst levels across attributes is 100 points for each respondent.

Now, to more directly answer your query:

The Charts tab from the Interaction Search Tool shows the net raw utility of the main effect of the row attribute level plus the main effect of the column attribute level plus the interaction effect between those two attribute levels (three utility terms summed).

However, the second thing you are showing (the interaction effect from HB) is just the interaction effect (the adjustment after considering the main effects) and it is obviously on the zero-centered diffs scaling.

If you were to look at the HB results (either raw or zero-centered diffs scaling) and add the main effect for Salary_Low plus the main effect for Task_Unattractive plus the interaction effect between those two levels, you should see a net negative utility similar to the pattern seen in the interaction search tool Charts tab.
answered Jan 31 by Bryan Orme Platinum Sawtooth Software, Inc. (134,015 points)