Hayes’ PROCESS Macro: Model Numbers, Syntax, and Reporting

Hayes’ PROCESS macro runs mediation, moderation, and combined moderated-mediation (“conditional process”) models in SPSS, SAS, and R without hand-coding the regression equations or the bootstrap resampling loop. The single most confusing part of using it is the model number: a single integer argument that tells PROCESS which path diagram to fit — which variable plays which role, and which paths, if any, are moderated. Enter the wrong model number and PROCESS will run without complaint; it just tests a different hypothesis than the one you meant to test. This page is a lookup reference for the model numbers actually used in applied research, the syntax to run each one, and how to translate PROCESS’s console output into a manuscript-ready report.

This page assumes you already know which kind of relationship you’re testing. For the underlying statistics — why bootstrapped indirect effects replaced the causal-steps method, and how to report a mediation result in general — see mediation analysis: methods and reporting. For building and interpreting a moderation (interaction) model, see moderation analysis: building and interpreting an interaction term. If you’re still deciding whether your third variable is a mediator or a moderator in the first place, start with mediator vs. moderator. For the potential-outcomes formalization beneath PROCESS’s regression-based approach, see causal mediation analysis.

What the model number actually specifies

Every PROCESS model is a path diagram with up to three roles beyond the outcome (Y): the independent variable (X), one or more mediators (M), and one or more moderators (W or Z). The model number is Hayes’ index into a fixed library of these diagrams, documented with a picture of each one in Appendix A of his book, Introduction to Mediation, Moderation, and Conditional Process Analysis (Guilford Press, 3rd ed.). The number does not encode effect sizes or which variables you assign to which role — you still supply x=, m=, y=, and w= yourself — it encodes the structure: how many mediators, how many moderators, and which specific paths (X→M, M→Y, or the direct X→Y path) each moderator is attached to.

Two models with the same variables can produce completely different results if the number is wrong, because PROCESS builds a different set of regression equations for each one. A first-stage moderated-mediation model and a second-stage moderated-mediation model both use exactly one X, one M, one Y, and one W — but they test whether the moderator changes a different link in the chain, and mixing them up is a real, recurring error in submitted manuscripts, not a hypothetical one.

The core model numbers

These are the templates that cover the large majority of applied uses. PROCESS’s full library (documented in Appendix A, and viewable directly in the model-number dropdown of the SPSS custom-dialog installation) extends further, including designs with two different moderators acting on separate stages — if your design needs that, work from the diagram library directly rather than guessing a number from a table.

• Model 1 — simple moderation. X, W, Y only, no mediator. Tests whether the effect of X on Y depends on W (the X×W interaction). This is the model behind moderation analysis and interaction terms, run through PROCESS instead of a manual regression syntax.
• Model 4 — simple (or parallel multiple) mediation. X, one or more M, Y, no moderator. With a single M this is the textbook indirect-effect model; specifying more than one M variable does not change the model number, it runs a parallel multiple-mediator model (up to 10 mediators) whose paths are estimated simultaneously, each with its own bootstrapped indirect effect.
• Model 6 — serial multiple mediation. X, two or more M variables entered in a specified causal order (M1 → M2 → …), Y. Tests the indirect effect through each mediator individually and through the full chain, distinct from Model 4’s parallel structure where the mediators are not ordered relative to each other.
• Model 7 — first-stage moderated mediation. X, M, Y, W, where W moderates only the X→M path (the a path). The indirect effect through M becomes conditional on the value of W; the M→Y (b) path and the direct effect are not moderated.
• Model 14 — second-stage moderated mediation. Same variables as Model 7, but W moderates only the M→Y path (the b path) instead of the a path. The a path and direct effect are unmoderated.
• Model 8 — first-stage moderation with a moderated direct effect. W moderates both the a path and the direct X→Y (c′) path; the b path is left unmoderated.
• Model 15 — second-stage moderation with a moderated direct effect. The mirror of Model 8: W moderates both the b path and the direct effect, leaving the a path unmoderated.

Two variables described as “moderating mediation” almost never means the same thing across two papers unless the model number matches — a reviewer who cites “Model 7” and one who cites “Model 14” are describing structurally different hypotheses about where in the chain the moderation happens, not two ways of saying the same thing.

Choosing the right model number

Work from the conceptual diagram you already have in mind, not from a keyword search for your effect size:

• If there’s no mediator at all and you’re asking whether X’s effect on Y depends on W — Model 1.
• If there’s a mediator and no moderator — Model 4 (or Model 6 if the mediators have a specified causal order).
• If there’s a mediator and exactly one moderator, decide which link the moderator theoretically changes: the X→M link (Model 7), the M→Y link (Model 14), or both the relevant stage and the direct effect (Model 8 or 15). This is a substantive decision your hypothesis has to justify, not a default PROCESS can pick for you — misassigning first-stage vs. second-stage moderation is the single most common model-selection error reviewers flag.
• If the design needs two different moderators on two different stages simultaneously, or a moderated moderator (three-way interaction feeding into mediation), the model number lives outside this core set — open the full diagram library in Appendix A or the SPSS dialog’s model list before choosing.

Running PROCESS: SPSS, SAS, and R syntax

The macro is installed once (as an SPSS custom dialog, a SAS macro file, or an R script/package) and then called with a consistent argument pattern across all three: which columns are y, x, m, and w, which model number to run, how many bootstrap resamples, and (recommended for reproducibility) a fixed seed. The example below is Model 4 (simple mediation, X→M→Y) with 5,000 bootstrap resamples:

• SPSS (Syntax Editor, after installing the PROCESS custom dialog): process y=Y/x=X/m=M/model=4/boot=5000/seed=12345.
• R (after sourcing or loading the process() function that ships with PROCESS for R, released 2020): process(data = df, y = "Y", x = "X", m = "M", model = 4, boot = 5000, seed = 12345)
• SAS (after %include-ing the macro file): %process (data=df, y=Y, x=X, m=M, model=4, boot=5000, seed=12345);

For a moderated-mediation model, add a w= argument and change only the model= number — the same X/M/Y/W column names feed Model 7, 8, 14, or 15; the model number is what tells PROCESS which path to attach W to, not a separate argument you set per path. Argument names have been stable across recent PROCESS versions, but always confirm the full option list (centering, HC-robust standard errors, multicategorical-X coding) against the documentation shipped with your installed version before relying on a default you haven’t checked.

PROCESS also has built-in support for a multicategorical X (more than two groups) via indicator coding, following the method in Hayes and Preacher’s 2014 British Journal of Mathematical and Statistical Psychology paper on mediation analysis with a multicategorical independent variable — specified with a mcx= option rather than manually dummy-coding X yourself beforehand.

Worked example: Model 4 output, translated

The table below is simulated data, not a real study — it exists only to show what PROCESS’s console output looks like and how each line maps onto a manuscript sentence. It was generated by a seeded script (N = 300, fixed seed, structural paths a = 0.40, b = 0.35, c′ = 0.15 plus Gaussian noise) so the numbers below are exactly reproducible, not illustrative guesses: a percentile bootstrap with 5,000 resamples was run on the same simulated sample.

PROCESS output line	Value	Reported as
Model 1 (outcome: M), effect of X — coeff, se, t, p	a = .510, se = .058, t = 8.77, p < .001	“X significantly predicted M, a = .51, SE = .06, t(298) = 8.77, p < .001”
Model 2 (outcome: Y), effect of M — coeff, se, t, p	b = .311, se = .055, t = 5.66, p < .001	“Controlling for X, M significantly predicted Y, b = .31, SE = .05, t(297) = 5.66, p < .001”
Model 2 (outcome: Y), direct effect of X — coeff, se, t, p	c′ = .137, se = .062, t = 2.23, p = .026	“The direct effect of X on Y remained significant, c′ = .14, SE = .06, t(297) = 2.23, p = .026”
Total effect of X on Y (no M in the model)	c = .296, se = .058, t = 5.11, p < .001	“The total effect of X on Y was c = .30, SE = .06, t(298) = 5.11, p < .001”
Indirect effect(s) of X on Y — Effect, BootLLCI, BootULCI	ab = .158, 95% CI [.092, .234]	“The indirect effect of X on Y through M was ab = .16, 95% bootstrap CI [.09, .23] (5,000 resamples), which excludes zero”

Note the arithmetic check that also holds in PROCESS’s real output: the direct effect plus the indirect effect reproduces the total effect (c′ + ab = .137 + .158 = .295 ≈ c = .296, within rounding), because PROCESS decomposes the total effect exactly this way rather than estimating it as a separate, unrelated quantity.

Reporting PROCESS results in a manuscript

A methods and results section built from PROCESS output generally needs, at minimum:

• The model number itself, stated explicitly (reviewers use it to check your path diagram matches your hypothesis) — “using Model 4 of Hayes’ PROCESS macro (v…)”.
• The software and PROCESS version (SPSS/SAS/R, and the PROCESS version number, since argument behavior and defaults have changed across releases).
• The number of bootstrap resamples (5,000 is the common default reported in current practice; note explicitly whether the interval is a percentile or bias-corrected bootstrap CI, since PROCESS documents both as available and which one your specific output used should be stated rather than assumed by the reader).
• Coefficients with standard errors for every path in the diagram (a, b, and, where applicable, the moderation coefficients), not just the headline indirect effect.
• The indirect effect and its bootstrap CI, reported as the primary test of mediation — there is no separate p-value for ab under the bootstrap approach; the interval excluding zero is the test.
• For a moderated-mediation model, the index of moderated mediation (PROCESS’s own summary statistic for whether the indirect effect significantly changes across levels of W) with its own bootstrap CI, plus the conditional indirect effects at representative values of W (commonly ±1 SD and the mean) rather than the index alone.

Hayes does not publish PROCESS through a standalone journal article; the citation convention is to cite the book itself (Introduction to Mediation, Moderation, and Conditional Process Analysis) rather than searching for a “PROCESS macro” paper to reference.

Frequently asked questions

Does the model number change if I add covariates?

No. Covariates are added as a separate argument (commonly cov=) and are partialled out of every equation in the diagram; they don’t change which diagram you’re fitting, only what each path controls for. Choose the model number from the X/M/W structure alone, then add covariates on top of it.

How many bootstrap resamples should I use?

5,000 is a common default in current applied practice and is what PROCESS itself defaults to in recent versions for most models; it’s sufficient for a stable interval in typical single-mediator and single-moderator designs. Raise it for a more complex model (multiple mediators or moderators) or if interval bounds visibly shift between reruns at a lower count.

Is PROCESS the same thing as fitting the model in a general SEM package?

They test the same underlying paths but aren’t identical tools. PROCESS fits each equation in the diagram as a separate OLS (or logistic, for a binary Y) regression and bootstraps the products of coefficients across those separate models; a full SEM package estimates every path simultaneously in one model, which matters once your design has correlated errors, latent variables, or multiple outcomes that PROCESS’s separate-equations approach doesn’t represent. For a single mediator or moderator with observed (non-latent) variables, the two approaches typically produce very similar indirect-effect estimates.

Can X have more than two groups (a multicategorical predictor)?

Yes — PROCESS supports multicategorical X natively via indicator coding (the mcx= option), following Hayes and Preacher’s (2014) method for mediation analysis with a multicategorical independent variable, rather than requiring you to dummy-code X by hand before running the macro.

Where do I find the diagram for a model number that isn’t in this list?

Appendix A of Hayes’ book reproduces the path diagram for every model number PROCESS supports, and the SPSS custom-dialog installation shows the same diagrams inline in its model-selection dropdown — check the diagram directly rather than inferring an unfamiliar number’s structure from its neighbors in a table like this one.