I have conducted a survey experiment with two treatments and a control group. When I run my bivariate regression one of the treatments is significant. I then add an interaction. The interaction variable turns out not to be significant but the treatment that was significant before is now not significant anymore. How can that be?
Interaction terms can cause both under and overconfidence. This is especially the case when your model uses a categorical dependent variable (see Esearey, Berry, and DeMeritt 2016: “Bias and Overconfidence in Parametric Models of Interactive Processes,” American Journal of Political Science 60(2): 521-539.)
Remember, when you engage in an interaction term you are testing a dependent relationship between two explanatory variables. The only effective way to assess your relationship is by looking at the predictive margins. Reviewers and Editors alike will want to see these margins graphically.
To add to Michael's point, here is a document that can further help.
https://www.stata.com/meeting/india13/materials/in13_rising.pdf
Hope it helps!
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running