Moderation asks when or for whom an effect occurs. Does workload increase burnout for everyone, or less so for employees with strong social support? If the strength (or direction) of the X → Y relationship changes with the level of a third variable W, then W is a moderator.
This guide shows how to test moderation with an interaction term, how to probe it with simple slopes and the Johnson-Neyman technique, and how to report it.
In this guide
The moderation model
Moderation is tested by adding the product of X and W to a regression:
Y = b₀ + b₁X + b₂W + b₃(X × W) + e
- b₃, the interaction coefficient, is the key test. If it is significant, the effect of X on Y depends on W.
- The effect of X on Y at a given value of W is b₁ + b₃W. This is called the conditional effect or simple slope.
In a model with an interaction, b₁ is the effect of X when W = 0, not an average "main effect". If zero is not a meaningful value of W (for example, a 1–7 scale), b₁ is not meaningful either. Mean-centering fixes this.
Step 1: Mean-center X and W (recommended)
Subtract each variable's mean from every score. After centering:
- b₁ becomes the effect of X at the average level of W, and b₂ the effect of W at the average level of X;
- the interaction coefficient b₃ and its p-value do not change;
- correlations between X, W and X × W go down, which makes the output easier to read.
Centering does not "fix" multicollinearity in any statistical sense, but it makes the lower-order coefficients interpretable. The free moderation calculator can center for you.
Step 2: Test the interaction
Look at three things:
- b₃ and its p-value (or 95% CI).
- ΔR² due to the interaction: how much extra variance the X × W term explains. Interaction effects are typically small in field data; ΔR² of .01–.03 is common and can still matter.
- Heteroscedasticity. Robust (HC3) standard errors are a good default for interaction models.
Step 3: Probe the interaction
A significant b₃ tells you the slope changes, not how. There are two standard ways to probe it.
Simple slopes (pick-a-point)
Estimate the effect of X at chosen values of W, usually the mean and ±1 SD (or the 16th, 50th and 84th percentiles). For categorical moderators, use each group.
The Johnson-Neyman technique
Instead of picking points, Johnson-Neyman finds the exact value(s) of W where the effect of X becomes, or stops being, significant. It gives a "region of significance", which is more informative than three arbitrary points.
Worked example
A researcher surveys 300 nurses. X = workload, W = social support, Y = burnout. X and W are mean-centered, and the SD of social support is 0.90.
| Predictor | b | SE | t | p |
|---|---|---|---|---|
| Constant | 2.10 | 0.04 | 52.50 | < .001 |
| Workload (X) | 0.45 | 0.05 | 9.00 | < .001 |
| Social support (W) | −0.20 | 0.05 | −4.00 | < .001 |
| Workload × Support | −0.15 | 0.05 | −3.00 | .003 |
R² = .34; the interaction adds ΔR² = .02, F(1, 296) = 9.00, p = .003.
Simple slopes (effect of workload on burnout = 0.45 − 0.15 × W):
| Social support | Value of W | Simple slope |
|---|---|---|
| Low (−1 SD) | −0.90 | 0.45 + 0.135 = 0.59 |
| Average (mean) | 0.00 | 0.45 |
| High (+1 SD) | +0.90 | 0.45 − 0.135 = 0.32 |
Workload increases burnout at every level of support, but the effect is almost twice as strong for nurses with low support as for nurses with high support. Social support buffers the effect of workload.
Johnson-Neyman: the workload effect stays positive and significant across nearly all observed values of support, and becomes non-significant only at very high support levels. Report the exact boundary value and the percentage of the sample above it, as given in the output.
Plot the interaction
Always show a plot: three lines (low, average and high W) with X on the horizontal axis and predicted Y on the vertical axis. Lines that fan out or cross make the moderation obvious at a glance. The calculator draws this plot automatically.
Moderation with categorical variables
If W is a group variable (for example, gender), the logic is the same: code it as a dummy (0/1), multiply it by X, and test the interaction. The simple slopes are the X → Y effects within each group. For three or more groups, use several dummy variables and test them together with an F-test on ΔR².
Two moderators or three-way interactions
- Two moderators (W and Z, each interacting with X): PROCESS Model 2.
- Moderated moderation (X × W × Z): PROCESS Model 3, or the three-way interaction calculator.
- Moderated mediation: see the moderated mediation tool and our mediation guide.
- Latent moderators measured with several items: CB-SEM Studio supports latent interactions with simple slopes; PLS-SEM Studio has a two-stage moderation approach.
How to report moderation in APA 7
A moderation analysis with mean-centered predictors and HC3 standard errors showed that social support moderated the relationship between workload and burnout, b = −0.15, SE = 0.05, t(296) = −3.00, p = .003, ΔR² = .02. Simple slopes analysis showed that workload was more strongly related to burnout when social support was low (−1 SD; b = 0.59) than when it was high (+1 SD; b = 0.32). Figure 1 shows the interaction.
Common mistakes
- Interpreting b₁ as a "main effect" without centering.
- Forgetting to include X and W themselves in the model alongside X × W.
- Testing moderation by comparing correlations or running separate regressions in subgroups (that loses power and does not test the difference).
- Claiming a buffering effect without probing and plotting the interaction.
- Dichotomizing a continuous moderator with a median split.
Frequently asked questions
Do I need to mean-center variables for moderation analysis?
It is not required for the interaction test, because b₃ and its p-value are identical with or without centering. It is recommended because it makes the coefficients of X and W interpretable as effects at the average of the other variable.
What is the difference between simple slopes and Johnson-Neyman?
Simple slopes test the effect of X at a few chosen values of the moderator, such as the mean and ±1 SD. Johnson-Neyman finds the exact moderator values where the effect becomes significant or non-significant.
What is a good ΔR² for an interaction?
Interaction effects in observational data are usually small. ΔR² values of .01 to .03 are common and can still be theoretically and practically important.
Is moderation the same as PROCESS Model 1?
Yes. PROCESS Model 1 is a simple moderation model with one moderator. The free moderation calculator runs models similar to PROCESS Models 1 to 3, as an independent implementation.