Interaction Effects

Concept and Logic

An interaction is modeled by adding a product term to a regression equation. With two continuous predictors X and Z the model takes the form: Y = b0 + b1X + b2Z + b3XZ + error. Here b3 is the interaction coefficient and indicates how much the slope of X changes for each one-unit increase in Z. With categorical predictors the test asks whether regression slopes differ across groups. The interaction term captures the extra effect produced by the joint action of the two predictors beyond their separate contributions — this is the standard statistical route for testing moderation hypotheses.

Computation and Interpretation: Simple Slopes and Centering

When continuous predictors are mean-centered — that is, each variable has its mean subtracted before forming the product XZ — multicollinearity between the product term and its components is reduced. More importantly, b1 and b2 become interpretable as the effect of X when Z equals zero, which now corresponds to the average of Z. To interpret a significant interaction, simple slopes of X are computed and tested at low (-1 SD), mean, and high (+1 SD) values of Z. This probing procedure makes the nature of the interaction concrete and directly reportable.

Common Misconceptions and Misuses

A frequent error is dismissing a significant interaction coefficient simply because the main effects are non-significant. Main effects need not be significant for an interaction to be meaningful. A second misconception conflates interaction with correlation: correlation describes the association between two variables, while an interaction tests whether the slope of one variable changes across levels of another. Third, forming a product term without centering can produce main-effect coefficients that are difficult to interpret because they apply only when the other predictor equals zero in its raw metric. Finally, testing interactions in small samples carries low statistical power, making underpowered null results uninterpretable.

Reporting and Importance

When reporting interactions, the full regression table should present all terms (b0, b1, b2, b3) with standard errors and p-values. If the interaction is significant, a simple slopes table and an interaction plot — showing the regression lines of X on Y at different values of Z — are expected. Effect size is best communicated as ΔR², the incremental variance explained by adding the interaction term. Interactions are essential for determining whether an intervention works differently across subgroups, whether a risk factor varies by context, and for confirming theoretical moderation hypotheses. They are therefore a core analytic tool in behavioral science, medicine, and the social sciences.

Sources