A thesis hypothesised that “lecturer support moderates the relationship between academic pressure and burnout”, then ran a three-step mediation-style analysis: pressure predicts support, support predicts burnout. The examiners pointed out that the analysis did not test the stated hypothesis. Moderation and mediation sound alike, both involve a “third variable”, but they answer entirely different questions.
The confusion is common and leads to errors in design, analysis and interpretation. This article separates the two concepts and sets out an analysis routine for each.
Two different questions
| Moderator | Mediator | |
|---|---|---|
| Question | When, for whom, is the effect stronger or weaker? | How, through what pathway, does the effect occur? |
| Example | Pressure leads to less burnout among students with strong lecturer support | Pressure reduces sleep, and poor sleep leads to burnout |
| Relation to the predictor | Usually not affected by it | Affected by it |
| Analysis | Interaction term X × W | Indirect effect a × b |
| Time order | Not required | X before M, M before Y |
A quick test: if you can write the hypothesis with “depends on” or “stronger among”, it is moderation. If you can write it with “through” or “because”, it is mediation.
Analysing moderation: interaction, centring, simple slopes
- Centre continuous predictors and moderators (subtract the mean). This does not change the interaction test but makes main coefficients readable: the coefficient of X becomes its effect when W is at its mean.
- Include the interaction X × W alongside X and W. Keep X and W in the model even if they are not significant.
- If the interaction is significant, do not stop at the coefficient. Compute simple slopes: the effect of X at low, average and high W, or at practically meaningful values.
- Plot the regression lines at those levels of W. The plot usually says more than the table.
- For continuous W, the Johnson–Neyman technique shows the range of W over which the effect of X is statistically significant.
A note on power: detecting an interaction needs a much larger sample than detecting a main effect of similar size. Many small studies “find no moderation” simply because they lacked the power to find it.
Analysing mediation: the indirect effect is central
The classic three-step approach (X predicts Y, X predicts M, M predicts Y controlling for X) was once standard but is no longer recommended as the main test. It requires X to predict Y significantly at the first step, although indirect effects can exist even when the total effect is not significant, and it never tests the indirect effect directly.
Current practice:
- Estimate path a (X → M) and path b (M → Y, controlling for X).
- The indirect effect = a × b.
- Test it with a bootstrap confidence interval (typically several thousand resamples). If the interval excludes 0, there is evidence of an indirect effect.
- Report both the direct effect (c′) and the total effect (c).
Labels such as “full” and “partial” mediation are increasingly discouraged, because they depend on sample size and invite the reading that the whole mechanism has been found.
Mediation with cross-sectional data: state the limits
Mediation is a causal claim about the chain X → M → Y. When all three variables are measured at once in one questionnaire, the data cannot distinguish that chain from others, such as Y → M, or M causing both X and Y. The model still produces tidy numbers, but the causal argument rests entirely on theory.
What to do:
- Where possible, measure X, M and Y at different time points.
- State in the limitations that the direction of relationships is assumed.
- Consider confounders of the M → Y path, which are often overlooked.
- Use cautious language: “consistent with a mediation hypothesis” rather than “demonstrates the mechanism”.
When the two combine: moderated mediation
Combined models ask whether an indirect effect differs across levels of a moderator. They are powerful but demand large samples and clear theory. The PROCESS macro for SPSS, SAS and R numbers many such models, while the lavaan package in R lets you specify models more flexibly. Do not pick model 58 just because it exists; draw your theoretical model first, then find the tool.
In practice: a checklist before analysis
- Write each hypothesis as a sentence and look for “depends on” or “through”.
- Draw it: moderation is an arrow pointing at another arrow; mediation is a chain of arrows.
- For moderation: centre, include the interaction, compute simple slopes, plot.
- For mediation: estimate a, b and a × b, bootstrap the interval, report c and c′.
- Check whether your sample is large enough for the expected interaction or indirect effect.
- Write limitations on time order and unmeasured confounders.
Sample sentence: “The indirect effect of academic pressure on burnout through sleep quality was 0.12 (95% bootstrap CI 0.06–0.19), consistent with a mediation hypothesis; because the data are cross-sectional, the direction of effects should be tested with a longitudinal design.”
Next step: open the hypotheses section of your current paper, underline every hypothesis involving a third variable and write MOD (moderation) or MED (mediation) beside it. Then check the analysis section: if the label and the analysis do not match, fix it now.
Câu hỏi thường gặp
What is the difference between a moderator and a mediator?
A moderator tells you when or for whom an effect is stronger or weaker; a mediator explains the pathway through which it occurs. They are analysed with interactions and indirect effects respectively.
Do I need to centre variables when testing an interaction?
Centring does not change the interaction test, but it makes the main coefficients easier to interpret as effects at the mean of the other variable.
Is the classic three-step approach to mediation still valid?
It is no longer recommended as the main test. Estimate the indirect effect directly and test it with a bootstrap confidence interval.
Can I test mediation with cross-sectional data?
You can run the analysis, but it cannot establish causal order. State the limitation clearly, use cautious language and measure variables at different times where possible.
What are simple slopes?
They are the effects of the predictor at specific values of the moderator, used to interpret a significant interaction rather than relying on the interaction coefficient alone.
