Have an idea?

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

How do I interpret results from the interaction search?

I have 4 significant main effects after conducting the interaction search:
Main effects, parameters in model, log-likelihood fit, chi-square value, 2LL p-value, Gain in Pct.Cert.

Access offline    15    -5406.79    42.42940    0.00000    0.32%

Delivery offline    16    -5418.37    19.27594    0.00024    0.15%

Selection & price
Access offline    17    -5422.11    11.79627    0.01893    0.09%

Distribution Channel
Selection & price    15    -5424.92    6.16907    0.04575    0.05%

What do these numbers tell me? How can I interpret the effects?
Sustainability = the retailer acts sustainable

So can I say, for a person that derives a high utility from retailers that act sustainable also possibilities to access the store offline has a high utility?

Thank you!
asked Aug 27, 2019 by Lena

1 Answer

0 votes

These results suggest that some of the interactions are significant but they do not help you interpret the interactions. For that you will need to look at the utilities (for main effects and the additive interaction effects) for interpretation.  

I would be cautious about adding interactions.  Maybe add the first one and then see if the addition of a second one is still significant (since the first two both include one of the same attributes, it may well happen that entering the first one in your model will make the second one non-significant).  Be careful not to include so many interactions that you over-fit your model.
answered Aug 27, 2019 by Keith Chrzan Platinum Sawtooth Software, Inc. (92,075 points)
thank you very much for the quick reply!

How do I add this interaction effect to my model though?
Thank you again!
Keith is right to be cautious.  The interaction test you are reporting here is an Aggregate Logit test.  However, most users will end up conducting HB utility estimation for their final model because of the big benefits HB provides in fitting individual-level utility models.  Once you fit individual utility models, usually the vast majority of the interaction effects that appear to be statistically significant under aggregate logit models end up going away and not being useful.  And, interaction effects in HB often lead to long run times and overfitted models.

If you do want to add interaction terms to either your HB or Logit models, you can do so by clicking the Gear icon (the settings icon) in the Analysis Manager.  For example, you select a logit-type utility run and you click the Gear icon to the right of it.  Then, within that settings dialog you'll go to the Attribute Coding area, then you'll see at the bottom of that dialog a pane that says Interactions with a green plus icon that says "Add...".
Thank you very much!
Unfortunately I do not see a green plus icon that says "add" in the attribute coding area.
I only see the attributes with their respective levels and at the bottom it says "Restore Defaults".
Any other way I could add the interaction effect?
Please call our tech support line at 801-477-4700.  You are probably using a version of our software that's old or different from what I was expecting, based on our most recent versions.