"Rethinking Interactions: Seems Like Most Regression Interactions Have Been Misinterpreted"
Hypotheses involving interactions are common in social science. Do returns to education differ by gender? Are unexpected losses more impactful than expected ones? Are extroverts less construal moderate the endowment effect? Etc. Linear regressions, as in y=ax+bz+cxz are the most common (only?) way in which such interactions are tested, and yet such approach is extremely likely to give the wrong answer. The false-positive rate can easily reach over 50%, the sign of the interaction can easily be wrong, the average interaction effect is oddly defined and its estimate often biased, and Johnson-Neyman/simple-slopes/spotlight/floodlight analysis are approximately hopeless. I identify 4 questions we often ask from interaction effects, and explore trustworthy alternatives to the current approaches to answering them.