A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Moderation asks whether the effect of X on Y depends on the value of a third variable, M (the moderator). Interaction terms in regression capture this dependence. Social processes often vary by context (geography, time, organizational type), and moderation analysis reveals who effects work for. Interpreting interactions requires careful plotting and simple slope tests; centering predictors aids interpretation. Moderation is foundational to understanding heterogeneous treatment effects and conditional relationships.
From linear regression, you know how to estimate the effect of a predictor X on an outcome Y while holding other variables constant. That coefficient tells you the average relationship across your sample. But "average across your sample" can obscure what's really happening: the effect of X on Y might be strong for some people and weak — or even opposite — for others. Moderation analysis is the formal way to ask: does the relationship between X and Y change depending on some third variable M?
The concrete setup is simple. You want to know whether the effect of X differs across levels of M. To test this, you add a product term — X × M — to your regression. The model becomes: Y = b₀ + b₁X + b₂M + b₃(X×M). The coefficient b₃ on the product term is the moderation effect. It tells you how much the slope of X changes for each one-unit increase in M. If b₃ is positive, the X→Y relationship gets stronger as M increases. If b₃ is negative, the relationship weakens (or reverses). If b₃ is near zero, there's no moderation — the effect of X is consistent across levels of M.
The hardest part of moderation analysis is not estimation — it's interpretation. The coefficient on X (b₁) is now a conditional coefficient: it tells you the effect of X when M = 0. If M has no natural zero, that coefficient may be meaningless. This is why centering matters: by subtracting the mean of M before computing the product term, you make M = 0 correspond to a meaningful value (the average case), so b₁ gives you the effect of X at the mean of M. Always center continuous moderators. Always report the conditional effects explicitly — do not rely on a reader to mentally reconstruct them from the raw coefficients.
The best way to communicate a moderation result is visually. Plot the predicted values of Y against X for several values of M (typically low, mean, and high — or the quartiles). If lines converge, the moderator matters. If lines are parallel, it doesn't. Simple slopes — the slope of X at a specific value of M — can then be tested individually for statistical significance. A common error is declaring moderation "significant" based on b₃ alone, without checking whether the effect of X is actually meaningful at the substantively interesting values of M. The interaction coefficient says the slopes differ; the simple slopes tell you what those slopes actually are.
Finally, moderation is a claim about for whom an effect operates — a question of scope conditions. If education improves earnings more for men than women, gender moderates the education-earnings relationship. If political messaging is more persuasive among low-information voters, political knowledge moderates persuasion. These are not statistical curiosities — they are theoretically important claims about how social processes are structured by context. The best moderation analyses begin with a theoretical account of why the moderator should matter, test it explicitly, and use simple slopes and plots to translate the results back into substantive, interpretable findings.
Topics in reflective domains aren't scored by quiz answers. Read, reflect, and mark when you've thought it through.