Mediation and moderation are two concepts that are constantly confused in psychological research, despite the fact that they address fundamentally different questions. Mediation asks why or through what mechanism a variable X affects a variable Y. Moderation asks when or under what conditions that relationship changes. They are distinct questions that require distinct models, and confusing them can lead to erroneous conclusions about the psychological mechanisms under study.
An example helps clarify the distinction. Suppose you are studying the relationship between job stress and depressive symptoms. A mediation model would ask: does job stress produce depressive symptoms through rumination? That is, does stress lead to rumination, and is rumination what actually produces depression? A moderation model would ask something different: is the relationship between stress and depression stronger in individuals with low social support than in those with high social support? In the first case, rumination is the explanatory mechanism. In the second, social support is a condition that modifies the strength of the relationship.
Mediation analysis and its assumptions
The classic mediation model, popularized by Baron and Kenny in 1986, established four conditions for demonstrating mediation: that X predicts Y, that X predicts the mediator M, that M predicts Y while controlling for X, and that the direct relationship between X and Y decreases (partial mediation) or disappears (full mediation) when M is included in the model. This approach has serious problems that modern methodology has been correcting. The most important is that the Baron and Kenny method has very low statistical power and may fail to detect indirect effects that truly exist.
The currently preferred approach is the test of the indirect effect using bootstrap, popularized by Preacher and Hayes. Instead of checking the four conditions sequentially, this method directly estimates the product of the coefficients for path a (from X to M) and path b (from M to Y), and uses bootstrap resampling to construct a confidence interval for that product. If the interval does not include zero, one concludes that a significant indirect effect exists. This method is more powerful, does not require the assumption of normality for the indirect effect (which is almost never met), and can be easily implemented using Andrew Hayes' PROCESS macro in SPSS or the lavaan package in R.
However, there is a fundamental problem that no statistical method can solve: mediation implies a causal model, and cross-sectional data cannot confirm causal relationships. When you measure stress, rumination, and depression at the same point in time, you cannot know whether stress causes rumination which causes depression, whether depression causes rumination which amplifies the perception of stress, or whether all three variables influence each other in complex ways. Maxwell and Cole (2007) demonstrated that cross-sectional mediation models can produce severely biased estimates of the true longitudinal indirect effect, even with the sign reversed. This should give pause to any researcher who interprets a cross-sectional mediation analysis as evidence of a causal mechanism.
Before you submit: reading a cross-sectional indirect effect as a causal mechanism is one of the first things a methodological reviewer underlines, together with an interaction reported without simple slopes. If your manuscript is already written, run it through the AI Paper Reviewer: a free Reviewer 2 style pre-review that tells you which claims your design cannot support, while you can still soften them.
Moderation and interactions
Moderation analysis is conceptually simpler, as it reduces to an interaction analysis. If the relationship between X and Y depends on the level of a third variable W, then there is an interaction between X and W in predicting Y. In a multiple regression model, this translates to including the product X*W as an additional predictor. If the coefficient for that product is significant, there is moderation.
Interpreting the interaction requires more than looking at the p-value of the interaction term. It is necessary to decompose the interaction to understand how the relationship between X and Y changes at different levels of the moderator. The most common technique is simple slopes analysis, which estimates the relationship between X and Y at specific values of W (typically the mean and one standard deviation above and below). The Johnson-Neyman technique goes a step further and identifies the exact values of the moderator at which the relationship between X and Y transitions from significant to non-significant, offering more precise information than simple slopes analysis.
Moderated mediation and mediated moderation
In practice, many models in psychology combine mediation and moderation. Moderated mediation occurs when the indirect effect of X on Y through M depends on the level of a moderator W. For example, the effect of stress on depression through rumination might be stronger in individuals with a tendency toward neuroticism. Mediated moderation, in turn, occurs when an interaction effect between X and W on Y is explained by a mediator. These models, known as conditional process models, can be implemented with relative ease using PROCESS or lavaan (see our SEM model comparator), but their interpretation is considerably more complex than that of simple models.
My recommendation for researchers working with these models is threefold. First, think about theory before technique. The statistical model should derive from a clear theoretical hypothesis about the mechanisms, not the other way around. Second, be honest about the limitations of your data. If your data are cross-sectional, do not speak of causal mechanisms or temporal processes, because your data cannot support those claims. And third, report all relevant information, including direct, indirect, and conditional effects, bootstrap confidence intervals, and simple slopes analyses when there are significant interactions. A transparent and well-reported analysis will always be more convincing than one that only shows favorable results. If you need support with these models, check our thesis analysis service.