Mediation analysis asks through which mechanism an independent variable (X) affects an outcome (Y); moderation analysis asks under which conditions that effect strengthens or weakens. The PROCESS macro for SPSS and R has become the standard postgraduate tool because it runs both families of models with bootstrap confidence intervals in a single command. This guide walks through the conceptual distinction, the model choices and the reporting templates step by step.
Mediator or moderator? The conceptual distinction
A mediator (M) answers the why/how question: X changes M, and M in turn changes Y (for example, workload → burnout → turnover intention). A moderator (W) answers the when/for whom question: the strength or direction of the X–Y relationship varies with the level of W (the effect of workload on burnout is weaker for those with high social support). Confusing the two is the most common source of misspecified models and faulty theoretical conclusions: a mediator is a mechanism, a moderator is a condition.
Why the causal steps approach is outdated
The classical Baron-Kenny causal steps approach ties mediation to the significance of a series of separate regressions, most notably requiring a significant total effect of X on Y as a precondition. The methods literature has moved past this for two reasons: first, under suppression a genuine indirect effect can exist while the total effect is non-significant; second, the steps never test the indirect effect itself — the a×b product. The same criticism applies to the Sobel test, which assumes normality even though the distribution of a product term is skewed. Today's standard is to test the indirect effect directly with a bootstrap confidence interval: compute the 95% interval from at least 5,000 resamples, and conclude the indirect effect is significant when the interval excludes zero.
Mediation analysis with PROCESS Model 4
Simple mediation is specified as Model 4 and reported through four paths: a (X→M), b (M→Y controlling for X), c (the total effect of X on Y) and c′ (the direct effect of X on Y with M in the model). The indirect effect is the product a×b, and the identity c = c′ + a×b holds. The labels "full mediation" and "partial mediation" are being abandoned in the current literature: the distinction is overly sensitive to sample size and reduces a theoretical question to whether c′ happens to be significant. Report instead the magnitude of the indirect effect, its confidence interval and, where possible, its standardised value.
PROCESS estimates the paths with ordinary least squares regression, so linearity, independence of errors and homoscedasticity should be checked (heteroscedasticity-consistent standard errors can be requested within the macro). X need not be continuous: a binary X enters directly, and a multicategorical X enters via indicator coding. If the outcome is binary, PROCESS switches automatically to logistic regression. Covariates such as age and gender are added to both the M and Y equations so that the indirect effect is purged of their influence.
Moderation with PROCESS Model 1
Moderation is tested with Model 1: Y is regressed on X, W and the X×W interaction term, and the evidence for moderation is a significant interaction coefficient. Mean-centring continuous predictors eases interpretation (centring does not change the significance of the interaction; it makes the main effects readable as effects at the mean of W). A significant interaction must always be probed: simple slopes analysis gives the X→Y slope at low (−1 SD), mean and high (+1 SD) levels of W, while the Johnson-Neyman technique identifies the range of W over which the effect is significant without arbitrary cut-points — and is increasingly the preferred approach. Interaction effects are typically small (f² is often around 0.02), so moderation tests demand noticeably larger samples than main effects; plan the sample with G*Power against the interaction term.
Moderated mediation: Models 7 and 14
When the mechanism and condition questions combine, moderated mediation comes into play: does the indirect effect itself vary with the level of W? Model 7 places the moderator on the a path (X→M); Model 14 places it on the b path (M→Y). The key output is the index of moderated mediation, which summarises in a single coefficient how the indirect effect changes per unit of W; if its bootstrap confidence interval excludes zero, the conditional indirect effects differ significantly across levels of W. When the index is significant, report the conditional indirect effects at selected levels of W separately. For latent-variable versions of these models, the approaches in our SEM guide are preferable.
| Model | What it tests | Key output |
|---|---|---|
| Model 1 | Simple moderation (X×W) | Interaction coefficient, simple slopes, Johnson-Neyman |
| Model 4 | Simple mediation (X→M→Y) | a×b indirect effect, 95% bootstrap CI |
| Model 6 | Serial mediation (chain of two+ mediators) | Separate indirect effect for each chain |
| Model 7 | Moderated mediation (a path) | Index of moderated mediation |
| Model 14 | Moderated mediation (b path) | Index of moderated mediation |
Reporting template
The minimum reporting set: the PROCESS model number and version used, the number of resamples (5,000), all path coefficients with standard errors, the indirect effect with its 95% bootstrap confidence interval, the interaction coefficient with simple slopes or Johnson-Neyman results for moderation, the covariates, and the centring decision. A model sentence: "The indirect effect was significant, a×b = 0.16, 95% CI [0.09, 0.24], 5,000 bootstrap samples." For formatting conventions, see our guide to reporting statistics in APA 7.
Mediation is the statistics of 'why'; moderation is the statistics of 'when' — a model that confuses them is wrong before the data arrive.
Frequently Asked Questions
Does mediation analysis require a significant total effect?
No. The current approach tests the indirect effect directly with a bootstrap confidence interval, and under suppression an indirect effect can be significant while the total effect is not. The precondition imposed by the causal steps approach is therefore no longer required.
How many bootstrap resamples should I use?
The common standard is 5,000 resamples, and 10,000 is also accepted. Fewer resamples make the confidence interval less stable. Report the 95% percentile or bias-corrected interval and interpret whether it excludes zero.
Should I report the full versus partial mediation distinction?
The current literature recommends moving away from these labels because the distinction is sensitive to sample size and reduces to whether c′ is significant. It is sounder to discuss the magnitude of the indirect effect, its confidence interval and its theoretical meaning.
What does Celsus offer for mediation and moderation analyses?
We help with model selection (the right PROCESS model number), data preparation and centring, analysis in SPSS or R, simple slopes and Johnson-Neyman plots, interpretation of the index of moderated mediation, and write-ups formatted for theses or journals. Outputs are delivered with the syntax files.