Statistics Glossary → Interaction effect
Interaction effectconcept
A situation in which the effect of one variable on the outcome changes with the level of another variable.
When two factors or variables interact, the effect of one cannot be summarized by a single number; its size, and even its direction, depends on the value of the other. In regression an interaction is modelled with a product term: y = b0 + b1·x + b2·z + b3·x·z, so the slope of x becomes b1 + b3·z and b3 measures the interaction. In ANOVA, non-parallel lines in the interaction plot point to an interaction.
When a significant interaction is present, interpreting the main effects on their own can mislead; simple effects (the effect of one variable at each level of the other) are examined instead. Adding a product term changes the meaning of the main-effect coefficients: b1 becomes the effect of x when z = 0, which is why centring continuous variables makes interpretation easier. In models with a nonlinear link such as logistic regression the scale matters: no interaction on the log-odds scale does not imply no interaction on the probability scale.
An everyday example
A fertilizer raises wheat yield by 80 kg per decare in irrigated fields but by only 10 kg in non-irrigated fields. Because the fertilizer's effect depends on irrigation, saying 'the fertilizer raises yield by 45 kg on average' paints a wrong picture for both field types.
In MerQur
The 'Model' option of Two-Way ANOVA (Full factorial A × B or Main effects only); 'Interaction terms' in multiple linear, logistic, weighted and Cox regression ('a*b' adds the main effects and the product, 'a:b' only the product); the 'Interaction' box in linear mixed models; 'Interaction degree' in the log-linear model. Variance Components Analysis can also add the A×B component of two crossed random factors.
Analyses where this concept appears
Two-Way ANOVAMixed-Design ANOVA, Split-PlotANCOVALog-Linear ModelMultiple Linear RegressionLogistic RegressionMultinomial Logistic RegressionLinear Mixed Models (LMM)Generalized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Variance Components Analysis (VARCOMP)Conditional Logit (Choice Model)Bayesian ANOVACox Regression (Proportional Hazards Model)Time-dependent Cox Regression