Overview
Andrew Hayes' PROCESS macro for SPSS, SAS and R has made mediation and moderation analysis accessible, but its numbered templates can be confusing. Choosing Model 4 when your hypothesis implies Model 7, forgetting to mean-center, or misreading the index of moderated mediation are common problems we see in draft results chapters.
We start by writing your hypotheses as a diagram, identify the matching PROCESS model (or a custom specification in lavaan, Mplus or Stata when PROCESS cannot express it), and agree how effects will be tested — usually percentile bootstrap confidence intervals for indirect effects rather than the older Sobel test or the causal-steps approach.
You receive annotated syntax, results tables, simple-slopes or Johnson–Neyman plots, and a plain-language explanation of each effect. We are also direct about the limits: statistical mediation in cross-sectional data is consistent with a causal process but does not prove one.
Who it’s for and the problems it solves
This is for you if you are…
- Master's and doctoral researchers with hypotheses involving mediators, moderators or both
- Management, psychology, education and health researchers using survey data
- Authors whose reviewers questioned their mediation approach (for example, reliance on Baron and Kenny's steps)
- Researchers who need to decide between PROCESS and full structural equation modeling
Common problems we help with
- You do not know which PROCESS model number matches your conceptual framework
- Your total effect is non-significant and you are unsure whether mediation can still be tested
- The interaction is significant but you cannot explain what it means
- You have several mediators in parallel or in sequence and the output is overwhelming
- A reviewer asked for the index of moderated mediation or a Johnson–Neyman analysis
What’s included — and what isn’t
Included
- Translating hypotheses into a statistical diagram and the matching PROCESS model or custom model
- Simple mediation, parallel and serial multiple mediation with bootstrapped confidence intervals for each indirect effect
- Moderation with continuous or categorical moderators, mean-centering decisions and interaction probing
- Simple slopes at chosen moderator values and Johnson–Neyman regions of significance
- Moderated mediation (conditional indirect effects) with the index of moderated mediation
- Covariate selection and reporting
- Assumption and diagnostic checks inherited from regression: linearity, heteroskedasticity (with HC-type standard errors), influential cases
- Latent-variable alternatives in lavaan, Mplus or SmartPLS when measurement error matters
- Publication-quality diagrams, interaction plots and APA-style tables
- Annotated PROCESS syntax, R or Stata code
Not included
- Testing many mediator–moderator combinations and reporting only those that 'worked'
- Claiming causal mechanisms from cross-sectional data without appropriate caveats
- Writing your results or discussion text for submission — we explain so you can write
- Altering data or dropping cases to obtain significant indirect effects
What you receive
- Model map
- Conceptual and statistical diagrams for each hypothesis, with the PROCESS model number or custom specification and the effects to be tested.
- Effects table
- Direct, indirect, total and conditional effects with bootstrapped 95% confidence intervals, and the index of moderated mediation where relevant.
- Interaction figures
- Simple-slopes plots and Johnson–Neyman plots formatted for your thesis or journal.
- Syntax and output
- Commented code with the random seed and number of bootstrap samples recorded, so results can be reproduced exactly.
- Explanation notes
- What each coefficient and interval means in the language of your hypotheses, and how to word supported and unsupported results.
How it works
Diagram the hypotheses
We draw your conceptual model with you and check that each arrow corresponds to a stated hypothesis.
Select the model
We match the diagram to a PROCESS model or custom specification and agree covariates, centering and bootstrap settings.
Check the data
Scale scores, missing data, outliers and regression assumptions are checked before running the models.
Estimate and probe
We estimate the models, then probe interactions and conditional indirect effects.
Deliver and explain
Tables, figures and syntax, followed by a walkthrough session and follow-up questions.
Models and techniques
- PROCESS Model 4
- Simple or parallel mediation: X affects Y through one or more mediators, each indirect effect tested with a bootstrap interval.
- PROCESS Model 6
- Serial mediation, where one mediator affects the next (X → M1 → M2 → Y), with each specific indirect path reported.
- PROCESS Models 1 and 2
- Moderation by one or two moderators, probed with simple slopes and the Johnson–Neyman technique.
- PROCESS Models 7, 8, 14 and 58
- Common first-stage, second-stage and dual-stage moderated mediation templates, tested with the index of moderated mediation.
- Latent-variable mediation
- SEM in lavaan, Mplus or SmartPLS when constructs are measured with several items and measurement error would bias observed-score estimates.
- Percentile bootstrap
- Resampling (commonly 5,000 or more draws) to build confidence intervals for indirect effects, which are rarely normally distributed.
Why we do not rely on the causal-steps approach
Older guidance required a significant total effect before mediation could be tested and judged mediation by a series of significance tests. Current methodological practice tests the indirect effect directly, usually with a bootstrap confidence interval, because an indirect effect can exist even when the total effect is small or non-significant — for example, when two indirect paths act in opposite directions.
We also avoid labeling results as 'full' or 'partial' mediation, which depends heavily on sample size, and instead report the size and precision of each effect.
Making interactions interpretable
A significant interaction coefficient tells you the effect of X depends on W, but not how. We probe it: estimating the effect of X at low, average and high values of W (or at meaningful categories), plotting the lines, and using the Johnson–Neyman technique to find where along W the effect becomes significant. Mean-centering helps interpret lower-order terms, though it does not change the interaction test itself.
Where the moderator is categorical, we report the effect within each group. Where two moderators are involved, we plot the full pattern rather than rely on a three-way coefficient.
Cautions we build into the write-up
Mediation implies a causal order. With data collected at one time point, the order comes from theory, not from the analysis, and reversing X and M may fit equally well. We help you acknowledge this, suggest longitudinal or experimental designs for future work, and consider sensitivity to unmeasured confounding of the mediator–outcome path where appropriate.
Testing several indirect effects and moderators multiplies the chance of false positives. We distinguish pre-registered or hypothesized effects from exploratory ones and report all effects that were tested.
- Same-source, self-report data can inflate relationships (common method variance)
- Observed-score analyses ignore measurement error, which can bias indirect effects
- Bootstrap results depend on the seed and number of resamples, so both are reported
What the work can look like
Illustrative example — fictional data
Sample excerpt: moderated mediation in a workplace survey
A doctoral student tested whether abusive supervision reduces work engagement through lower psychological safety, and whether team support weakens the first link (PROCESS Model 7, n = 358). All numbers are invented.
Conditional indirect effects
The indirect effect via psychological safety was −0.19, 95% bootstrap CI [−0.27, −0.12] at low team support (−1 SD), and −0.08, 95% CI [−0.14, −0.03] at high team support (+1 SD), based on 10,000 resamples.
Index of moderated mediation
Index = 0.07, 95% CI [0.02, 0.12]. Because the interval excludes zero, the indirect effect differs across levels of team support, supporting H4.
Wording note
The write-up describes the result as consistent with a buffering role of team support, and notes that cross-sectional, single-source data limit causal conclusions.
Illustrative example with invented numbers, not a real client case. Your outputs follow your data, model and style guide.
What you’ll need to provide
- Your hypotheses and conceptual framework diagram
- The anonymized dataset with computed or computable scale scores
- Details of each measure, including reverse-coded items
- Which covariates you plan to control for and why
- Your software (SPSS, SAS or R for PROCESS; or R, Stata, Mplus for alternatives)
You don’t need everything ready to ask for a quote — tell us what you have and we’ll say what’s missing.
What affects price and turnaround
Price depends on
- Number of hypotheses and models
- Model complexity: serial mediators, several moderators, moderated mediation
- Whether a latent-variable SEM alternative is needed
- Data preparation, scale construction and missing-data handling
- Delivered analysis versus guided learning sessions
Turnaround depends on
- Clarity of the conceptual model when we start
- Whether scale scores are ready or must be computed and checked
- Number of models to estimate and probe
Every project is quoted individually. How we price our work.
Questions people ask before starting
Which version of PROCESS should I use?
Use the current release from Andrew Hayes' website, which is available for SPSS, SAS and R. Versions change occasionally (version 5 was current when this page was written), so we note the exact version in your methods section and syntax.
Can PROCESS handle latent variables or multiple outcomes?
No. PROCESS works with observed variables such as scale means and one outcome per model. When you need latent constructs, several outcomes or a full theoretical model, we recommend SEM in lavaan, Mplus or SmartPLS.
My total effect is not significant. Is there no mediation?
Not necessarily. Modern practice tests the indirect effect directly. We report it with a confidence interval and explain what a non-significant total effect means for your argument.
Should I mean-center my variables?
Centering makes lower-order coefficients easier to interpret and is often recommended, but it does not change the interaction test. We decide with you and report the choice.
Can you help if I have already run the analysis?
Yes. We check your model choice, settings and interpretation, point out anything that needs to change and explain why.