PROCESS Documentation Addendum

PROCESS is documented in Appendices A and B of the third edition of Introduction to Mediation,

Moderation, and Conditional Process Analysis. The addendum to the documentation describes
options and features added to PROCESS since the first printing of the book in December 2021. This
version of this addendum was finalized for distribution on February 21, 2025, and describes
features available in PROCESS 4.1, 4.2, 4.3, and 5.0 beta. All features added in earlier releases are
available in later releases.

Missing Data Identification

(Added in version 4.1)

PROCESS uses listwise deletion prior to analysis, meaning that any case in the data file that has
missing data on any of the variables in the model will be deleted from the analysis. The
resulting sample size after listwise deletion is provided at the top of the PROCESS output, and at
the bottom PROCESS will provide how many cases with missing data were deleted prior to
analysis. However, the user is left in the dark about which cases were deleted from the analysis
as a result of missing data. With the release of PROCESS version 4.1, a new option is available
that provides information about which cases were deleted. When this option is turned on by
adding listmiss=1 to the PROCESS command, PROCESS will list the case numbers at the bottom
of the output, identified by row in the data file, that were deleted from the analysis.

Total Effect of X in Mediation Models

(Added in version 4.1)

In the typical mediation model, the total effect of X on Y, which is the sum of the direct and
indirect effects of X, is estimated in a regression model of Y regressed on X but not the
mediators. But the regression coefficient for X in this model will not always be equal to the sum
of the direct and indirect effects of X, such as when a model includes covariates and the
covariates are not included in all equations that are used to estimate the direct and indirect
effects of X. For this reason, PROCESS will not always produce the total effect of X or the model
that estimates it when the total option is used.

As of the release of version 4.1, for models such as these when the total effect cannot be
estimated by regressing Y on X, PROCESS will now produce the sum of the direct and indirect
effects in the output along with a bootstrap confidence interval for inference when the total
option is used. This option will only generate a point and interval estimate of this sum as well as
a bootstrap estimate of the standard error of the sum. Standardized metrics of the sum are not
available in this release.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

Estimation of and Inference for a Weighted Sum of Regression
Coefficients
(Added in version 4.1)

Any statistic that can be calculated as a weighted sum of the regression coefficients in a model
can be generated using a new linsum option in PROCESS available as of version 4.1. This option
is available only for models 0, 1, 2, and 3. The weighted sum takes the form
𝑘

∑ 𝜆𝑖 𝑏𝑖
𝑖=0

where b0 is the regression constant, b1 through bk are the regression coefficients for the k
variables in the model in the order they appear in the regression output for the model of Y, and
𝜆i are the weights. These weights are listed in sequence 0 to k from left to right following
linsum= in the PROCESS command and in the same order as the regression weights appear in
the PROCESS output from top to bottom.

The estimated value of the consequent variable Y given values on a set of predictor variables in
the model is an example of a weighted sum of regression coefficients. For example, using the
DISASTER data in Chapter 7 and the PROCESS output in Figure 7.6, the estimated justification for
withholding aid for a person in the disaster frame condition (frame = 1) with a score of 3 on
the skepticism scale (skeptic = 3) is

Ŷ = 1(2.4515) + 1(-0.5625) + 3(0.1051) + 3(0.2012) = 2.8079

where the numbers in parentheses are the regression constant and regression coefficients for
frame, skeptic, and the product of frame and skeptic, and the weights are 𝜆0 = 1, 𝜆1 = 1,
𝜆2 = 3, 𝜆3 = 3, for the regression constant and regression coefficients in this same order. In
PROCESS, this weighted sum is generated by adding the linsum option and the sequence of
weights, as in

process y=justify/x=frame/w=skeptic/model=1/linsum=1,1,3,3.

%process(data=disaster,y=justify,x=frame,w=skeptic,model=1,linsum=1 1 3 3)

process(data=disaster,y="justify",x="frame",w="skeptic",model=1,linsum=c(1,1,3,3))

which toward the bottom of the output produces

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

Linear Combination Estimate and Hypothesis Test
Weight vector:
weight
constant 1.0000
frame 1.0000
skeptic 3.0000
Int_1 3.0000

Estimate se t p LLCI ULCI

2.8079 .0826 33.9834 .0000 2.6450 2.9708

showing the estimate of justification for withholding aid from the model for such a person is
2.8079. The t -statistic and p-value for the test of the null hypothesis that this weighted sum
equals zero, provided in the output, would not be of much interest in this example, but the
standard error and confidence interval for the estimate may be. Here, the estimated standard
error of the weighted sum is 0.0826 and the 95% confidence interval for the weighted sum is
[2.6450, 2.9708].

It is very important that the weights following linsum= be in the same order from left to right as
the predictors in the model are displayed in PROCESS output from top to bottom, otherwise the
weighted sum will not be the sum you wish to construct. In SPSS and R, the weights should be
separated by a comma. In SAS, the weights are separated by a space. In R, the comma-
delimited sequence of weights should be enclosed in the c() operator.

The linsum option can also be used to compare two regression coefficients in a model. For
example, in Chapter 2, support for government action is estimated from negative emotions,
positive emotions, ideology, sex, and age. The linsum option can be utilized to test whether the
regression coefficient for negative emotions is equal to the regression coefficient for positive
emotions. This comparison would be a weighted sum of regression coefficients of the form

(0)b0 + (1)b1 + (-1)b2 + (0)b3 + (0)b4 + (0)b5 = b1 – b2

where b0 through b5 are the regression constant and regression coefficients for negative
emotions, positive emotions, ideology, sex, and age, respectively. In terms of the regression
coefficients from the model on pages 51-52,

0(4.064) + 1(0.441) + (-1)(-0.027) + 0(-0.218) + 0(-0.010) + 0(-0.001) = 0.468

In this weighted sum, the weights are 0, 1, -1, 0, 0, and 0 for the regression constant and
coefficients for negative emotions, positive emotions, ideology, sex, and age, respectively. In
PROCESS, the model and weighted sum is estimated with the command

process y=govact/x=negemot/cov=posemot ideology sex age/model=0/

linsum=0,1,-1,0,0,0.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

 %process(data=glbwarm,y=govact,x=negemot,cov=posemot ideology sex age,model=0,
linsum=0 1 -1 0 0 0)

process(data=glbwarm,y="govact",x="negemot",cov=c("posemot","ideology","sex",
"age"),model=0,linsum=c(0,1,-1,0,0,0))

which generates in the output

Linear Combination Estimate and Hypothesis Test

Weight vector:
weight
constant .0000
negemot 1.0000
posemot -1.0000
ideology .0000
sex .0000
age .0000

Estimate se t p LLCI ULCI

.4676 .0411 11.3856 .0000 .3870 .5482

showing that the difference between these regression coefficients is 0.4676 and statistically
significant, t(809) = 11.3856, p < .0001, with a 95% confidence interval of [0.3870, 0.5482]. The
degrees of freedom for the t statistic is the residual degrees of freedom for the model,
displayed in the PROCESS output in the model summary section under “df2.”

The linsum option expects that in a regression model with k predictors (including products
created by PROCESS to capture linear moderation), the sequence should contain k + 1 weights
(the extra weight is for the regression constant). However, when m covariates are listed
following cov= , weights for all the m covariates can be left out of the sequence if desired.
When weights for covariates are not included, PROCESS will automatically set the weights for
each covariate to the arithmetic mean of that covariate. When the number of weights in the
sequence is neither k + 1 nor k + 1 – m, a note will be displayed in the output stating that the
vector of weights is not correct and no output for the weighted sum will be generated.

The linsum option is not available for logistic regression models (i.e., when Y is dichotomous).

Counterfactually-Defined Indirect, Direct, and Total Effects of X in

Mediation Models with X by Mediator Interaction
(added in version 4.2)

With the release of version 4.2, PROCESS can estimate a mediation model that allows
interaction between X and mediator(s). When the xmint option is toggled on (by adding
xmint=1 to the PROCESS command), PROCESS will generate counterfactually defined natural

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

indirect, natural direct, and total effects of X, as well as the controlled direct effect. This option,
which replaces the now discontinued model 74, is available only for simple (one mediator) and
parallel multiple mediator models (model 4). For details on the estimation of such a model, the
mathematics of counterfactually-defined effects of X, as well as several options available for
use in tandemwith the xmint option, see CCRAM Technical Report 022-04 available through the
Resource Hub at the Canadian Centre for Research Analysis and Methods at
[Link]/ccram.

Selective Exclusion of Cases from the Analysis

(added in version 4.3)

With the release of version 4.3, it is possible to selectively exclude observations in the data
from the analysis. This is accomplished by adding the exclude option to the PROCESS command,
specifying the row number in the data file you would like to exclude following an equal sign. For
example, to exclude the observation in the 12th row, add exclude=12 to the PROCESS
command. To exclude more than one observation, list the row numbers of the observations to
be excluded. For example, to exclude the observations in rows 12, 14, and 36, use
exclude=12,14,36 in SPSS, exclude=12 14 36 in SAS, or exclude=c(12,14,36) in R.

Regression Diagnostics
(Added in version 5.0)

Regression diagnostics have many uses in regression analysis, from checking for data entry or
other forms of clerical errors, to finding cases that are high in influence or that are in some way
distorting the analysis, to checking regression assumptions.

Saving Regression Diagnostic Statistics for Examination and Analysis

Most good regression programs can save various regression diagnostic statistics for each case in
the analysis. As of the release of this version, so does PROCESS. This is accomplished using save
option 4, adding save=4 to the PROCESS command. In the R version of PROCESS, you must also
send the output of the save option to an object for storage. For example, in PROCESS for R

diagnostics<-PROCESS(data=…,save=4)

sends the diagnostic statistics to an object named “diagnostics”.

The regression diagnostics PROCESS generates are discussed in Darlington and Hayes (2017)
and other good treatments of regression analysis. These include, for each case in the data:

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

 Column name Diagnostic statistic

pred the estimate of the outcome (i.e., variable on the left side of the equation)
resid residual
dresid deleted residual
stresid standardized residual,
tresid t-residual (aka “externally studentized deleted residual”).
h leverage (aka “hat” value)
mahal Mahalanobis’ distance
cook Cook’s distance
dmsres change in MSresidual as a result of the case being included in the analysis.
drsq change in R2 as a result of the case being included in the analysis.
dskew This will contain all 99999 and is a placeholder for a future diagnostic.
dfb_# dfbetas for the regression constant and each regression coefficient (dfb_0,
dfb_1, etc, appearing in order from left to right as the model variables
appear in the output from top to bottom, starting with the regression
constant).
dfb_ie# Change in the indirect effect with and without the case in the analysis,
calculated as dfb_ie = IEwith – IEwithout . There will be as many dfb_ie
columns in the diagnostics file as there are indirect effects in the model.
This statistic will not be calculated for the total indirect effect.

In models with a mediation component (but no moderation), PROCESS also generates statistics
labeled “dfie_#”, which is the change in the indirect effect that results when a case is included
in the analysis. When a model contains more the one indirect effect, PROCESS will generate as
many columns of these dfie statistics for each case as there are indirect effects, with the
columns corresponding in order from left to right as the indirect effects appear in the output
from top to bottom. In multiple mediator models, there is no dfie calculated or saved for the
total indirect effect.

Cases that were excluded from the analysis as a result of missing data will not be included in
the diagnostics file. Case numbers in the diagnostics file are numbered in a variable name
“casenum” with values that correspond to the row numbers in the original data file being
analyzed. If you are not sure which cases PROCESS deleted as a result of missing data, use the
listmiss option in the PROCESS command. The diagnostics file also contains each case’s value(s)
for the regressor(s) in the model, making it easier to determine if there is a systematic
relationship between any of the diagnostic statistics and the variables in the model.

As a PROCESS command may generate many different regression equations in the output, the
save=4 option may generate more than one file or collection of regression diagnostics. In the
SAS version of PROCESS, these resulting files will be named “diagfile#”, with “#” sequencing

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

upward from 1 to k where k is the number of regression models in the output. The value of “#”
tells you from which equation, in the order the equations appear in the output, the diagnostics
come from. In the case of a model with no mediation component, only a single equation is
generated by PROCESS and so the diagnostic file is called “diagfile” by default. In SAS, you can
change the base name of the file using the diagfile option. For example, adding
diagfile=diagstat to the command will name the file(s) “diagstat” (e.g., diagstat1, diagstat2,
etc). These files are stored as temporary SAS work files in the current SAS session. To save them
for use later, you must save the diagnostic file(s) as SAS data files.

The R version of PROCESS will save the regression diagnostics as a data frame in the named
object as a list, unless the PROCESS command generates only one model. For example, if you
named the object diagnostics and PROCESS generated three regression equations for your
model, the diagnostic statistics for the three models will be held in diagnostics[[1]],
diagnostics[[2]], and diagnostics[[3]], with the numbers corresponding to the
regression equations as they appear in the PROCESS output from top to bottom. If the model
requires only one equation, the object will be a simple data frame rather than a list.

The SPSS version generates only one data file in the active SPSS session, with sets of regression
diagnostics for each model stacked on top of each other and numbered in the data file in a
variable named “equation”. Because not all variables in a PROCESS command will be in every
equation, when a variable is not in an equation, all the cases values will be set to 999999 in the
rows corresponding to that equation. Likewise, not all equations will have the same number of
regressors, so the number of dfbeta statistics will vary from equation to equation. Values in
dfbeta columns that have no corresponding values in an equation are set to 999999. Note that
this data file that PROCESS produces will not be permanently saved until you manually save the
data using the graphical user interface or a SAVE command in SPSS syntax. Because the name of
variable on the left side of the equation will vary from equation to equation, in the SPSS version
of PROCESS, the values of the variable on the right-hand side of the equation will be found in
the column labelled “dv”. In addition, unlike the SAS and R versions, the SPSS version will not
generate regression diagnostic statistics for the total effect model when that model is
requested in the output using the total option.

If any of the variables used in the PROCESS command are the same as the column names
PROCESS tries to use for the diagnostic statistics file (see the table on the prior page), the
diagnostics file will not be created. In that case, change the duplicate variable name in the data
to avoid the naming conflict and rerun the PROCESS command.

Diagnostics, Assumption Tests, and Casewise Influence

Whereas the save=4 option saves a set of diagnostic statistics for each case in the analysis to a
file, as discussed above, including diagnose=1 in the PROCESS command generates a section of
output for each regression equation containing information useful for testing assumptions and
flagging influential cases. Excerpts of an example output are provided below along with an
explanation of the information contained in the excerpt.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

 Min. Max.
fitted 2.7207 6.4109
residual -4.8381 3.4266
t-resid -4.6146 3.2464

This section contains the smallest (Min.) and largest (Max.) estimates of the outcome from the
model (fitted), residual, and t-residual.

Shape of residuals
Skewness Kurtosis
Value -.2708 .5165
se .0856 .1711

This section contains the skew and kurtosis of the residuals along with an estimate of the
standard error of skew and kurtosis. A ratio of “Value” relative to its standard error (“se”) that
exceeds two is diagnostic of a violation of the assumption of normality of the errors in
estimation.
Bonferroni-corrected p for largest t-residual
t-resid p-value casenum
-4.6146 .0037 139.0000

This section provides a general test of model assumptions. Under the standard assumptions of
regression, the t-residuals should follow a t(dfresidual) distribution, each of which has a two-tailed
p-value under the null hypothesis that a case’s measurement on the outcome variable comes
from a normal distribution around the regression line. Because this test is conducted for all
cases in the analysis without any a priori expectations as to which cases might be responsible
for an assumption violation, a Bonferroni correction to the two-tailed p-value is applied to
correct for multiple tests. The output shows the case number in the data file with the smallest
Bonferroni-corrected p-value for its t-residual. A p-value less than .05 (or whatever level of
significance or alpha-level you desire for the test) leads to a rejection of the null hypothesis that
all the regression assumptions are met. Note that as discussed in Darlington and Hayes (2017),
this can be quite low in power relative to tests of specific assumptions. A small p-value is
diagnostic of an assumption violation of some kind without identifying which assumption, but a
large p-value doesn’t necessarily mean all assumptions are met. In this example, case 139 is
contributing most to the assumption violation. This does not mean this is the only potentially
problematic case, however, as output only shows the Bonferroni-corrected p-value for the
case’s t-residual that is most distant from zero.

Most influential observations

casenum dfbeta
constant 142.0000 .0365
negemot 35.0000 -.0054
posemot 139.0000 -.0060
sex 139.0000 .0118
age 139.0000 -.0004
ideology 139.0000 .0084

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

This section of output identifies, for each regression coefficient (as well as the constant), the
case that when included in the analysis changes that regression coefficient the most and by
how much (i.e., cases with the dfbeta for that regression coefficient most distant from zero).
This output shows the regression coefficient for posemot, sex, age, and ideology change the
most when case 139 is included. However, the coefficient for negemot changes the most with
the inclusion of case 35, and the regression constant changes the most when case 142 is
included. Another way of interpreting these dfbeta values is by multiplying them by -1, which
then quantifies how much the regression coefficient changes when that case is deleted from
the analysis. Cases with especially large dfbeta values (ignoring sign) relative to others can be
said to be more or highly influential.
Variable tolerance and VIF
Tol. VIF
negemot .8574 1.1664
posemot .9743 1.0264
sex .9498 1.0529
age .9337 1.0710
ideology .8372 1.1944

This section of output provides the tolerance (Tol.) and variance inflation factor (VIF) for each
variable in the model. These are both sensitive to the strength of the association between a
variable and all of the other variables on the right-hand side of the regression equation. Note
that VIF is just the inverse of tolerance (i.e., VIF = 1/Tol.)

Breusch-Pagan test of heteroskedasticity

Chi-sq df p
Normal 18.5281 5.0000 .0024
Robust 14.7868 5.0000 .0113

This section of output provides the Breusch-Pagan test of heteroskedasticity in two forms. The
null hypothesis tested is that the homoskedasticity assumption is met. The row labeled
“Normal” is the traditional test that assumes the errors in estimation (manifested as the
residuals in the model) are normally distributed. As this test is sensitive to violations of this
assumption, the test is the second row labelled “Robust” is more trustworthy when the errors
in estimation are not normally distributed. As can be seen, both versions of this test suggest the
errors in estimation are heteroskedastic, which is a violation of the assumption of
homoskedasticity.

Indirect effect(s) of X on Y:
Effect BootSE BootLLCI BootULCI
TOTAL -.0029 .0727 -.1449 .1424
resource -.1175 .0464 -.2122 -.0302
workload .1146 .0390 .0444 .1973

Cases with greatest influence on indirect effect(s):

casenum dfb_ie
resource 37.0000 -.0165
workload 37.0000 -.0131

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

This section of output, found in the summary section toward the bottom of a mediation
analysis output, identifies the case(s) that as a result of inclusion in the analysis change the
indirect effect(s) the most and by how much. The dfb_ie statistics are calculated as

dfb_ie = IEwith – IEwithout

where IEwith is the indirect effect with the case included and IEwithout is the indirect effect
without the case included. Thus, the indirect effect if the case is excluded is

IEwithout = IEwith – dfb_ie

The dfb_ie statistics are thus like dfbetas for the regression coefficients but are for the indirect
effects, which are products of regression coefficients. Cases with especially large dfb_ie values
(ignoring sign) relative to others can be said to be more or highly influential. In this example,
case 37, by its inclusion in the analysis, is changing the indirect effect through both resource
and the indirect effect through workload the most. The negative values here mean that the
inclusion of case 37 in the analysis moves the indirect effects through resource and workload to
the left on the number line. So if case 37 were excluded from the analysis, the indirect effects
through resource and workload would be -0.1010 and 0.1276, respectively. In models with
more than one mediator, often one case will be more influential on one indirect effect but a
different case will be more influential on another indirect effect.

The dfb_ie statistics are not provided for conditional indirect effects in a conditional process
analysis.

Sums of Squares, Mean Squares, and Adjusted R2

(Added in version 5.0)

In a regression analysis, total variation in the outcome variable is broken into regression and
residual components. These sources of variation are the total, regression, and residual sums of
squares. With the release of version 5.0, these sources of variation will be displayed in the
output (under the column SS) along with corresponding degrees of freedom (df) and mean
squares (MS) when ssquares=1 is added to the PROCESS command. An example of the resulting
output is below.

Model Summary
R R-sq Adj R-sq F p SEest
.6232 .3883 .3845 102.7169 .0000 1.0673

SS df MS
Regress 585.0188 5.0000 117.0038
Residual 921.5233 809.0000 1.1391
Total 1506.5421 814.0000 1.8508

The use of this option also adds adjusted R2 to the model summary section of output, as above.
To avoid the line of output being excessively wide, “df1” and “df2” for the F-ratio ordinarily in

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

the output when this option is not used is eliminated and can instead by found in the sources of
variation table as the regression and residual degrees of freedom.

Note that ssquares=1 is the default when estimating a model without a mediation or
moderation component (i.e., model=0). To eliminate the printing of adjusted R2 and the sums
of squares in this case, add ssquares=0 to the PROCESS command.

The ssquares option is not available in conjunction with errors-in-variables regression.

Crossvalidation Multiple Correlations/"Shrunken R"

(Added in version 5.0)

Adding crossv=1 to the PROCESS command will produce the three estimates of “Shrunken R”
discussed in Darlington in Hayes (2017, pp. 181-186). “LvOut1” and “LvOut2” are discussed on
page 184 in that order, and “Browne” is discussed on page 185 (see equation 7.1).

Shrunken R estimates
Browne LvOut1 LvOut2
.6177 .6152 .6182

Scale-Free and Standardized Measures of Association

(Added in version 5.0)

Until recently, PROCESS had limited features for producing standardized regression weights or
measures of association, and then only for mediation models. With the release of version 5.0,
various scale-free and standardized measures of association are available as an option for every
model that PROCESS can estimate. These are accessed by adding stand=1 to the PROCESS
command. When this is done, each regression equation will include an output such as below.
The rows are the variables on the right-hand side of a model equation, and the columns are
various measures of scale free, partially, and completely standardized weights for those
variables in the model.

Scale-free and standardized measures of association

r sr pr standYX standY standX
negemot .5777 .4585 .5058 .4952 .3240 .6737
posemot .0430 -.0262 -.0334 -.0265 -.0197 -.0361
ideology -.4183 -.2219 -.2730 -.2425 -.1604 -.3300
sex -.0986 -.0036 -.0046 -.0037 -.0074 -.0050
age -.0971 -.0152 -.0194 -.0157 -.0010 -.0214

eta-sq p_eta-sq f-sq

negemot .2103 .2558 .3437
posemot .0007 .0011 .0011
ideology .0493 .0745 .0805
sex .0000 .0000 .0000
age .0002 .0004 .0004

The statistics available include the simple or zero-order correlation (r) with the outcome (the
variable on the left-hand side of the model equation), the semipartial (sr) and partial

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

correlation (pr), the completely standardized regression weight (StandYX), and two partially or
“semi-“ standardized regression weights (StandY and StandX). The two partially standardized
weights differ with respect to which variable is standardized. For StandY, only the variable on
the left-side of the model is standardized, and for StandX, only the variable on the right-hand
side of the equation is standardized (both the left and right-hand variables are standardized for
StandYX).

PROCESS also generates three commonly-used quantifications of partial association that are
sometimes used as measures of “effect size.” These are 2 (eta-sq) and partial 2 (p_eta-sq),
and Cohen’s f2 (f-sq). Note that 2 and partial 2 are just the squares of the semipartial and
partial correlations also provided by PROCESS. See Darlington and Hayes (2017) for a discussion
of these.

For mediation models, the stand option continues to produce complete and partially
standardized direct, indirect, and total effects as in prior releases. The completely standardized
total, direct, and indirect effects are not provided when X is dichotomous, and the partially
standardized effects are produced only when X is dichotomous or multicategorical.

The stand option is not available when the Y variable in the model is dichotomous or when
used in conjunction with the robustse option or in errors-in-variables regression.

Bootstrap Estimate File Column Names

(added in version 5.0)

In earlier releases of PROCESS, save option 1 (save=1) produces a data file containing all the
bootstrap estimates of every regression coefficient (plus the regression constant). The
bootstrap estimates are in the columns of this data set and labeled “col1,” “col2,” “col3,” etc.,
bootstrap samples are the rows, and a map in the PROCESS output provides a key for knowing
which columns corresponds to which regression coefficients in the model equations. See
Appendix A of Introduction to Mediation, Moderation, and Conditional Process Analysis for a
discussion of this save option.

In the SAS and R versions of PROCESS version 5.0, the column names in this file now provide the
information needed to know which columns contains which bootstrap estimates. The generic
“col1,” “col2,” “col3” column names are no longer used, and the map has been eliminated in
the output. In version 5.0, the column names are now in the format “left_right,” where “left” is
the variable name on the left side of the equation and “right” is the variable name on the right
side of the equation.

For example, adding save=1 to the SAS or R commands on page 93 of Introduction to

Mediation, Moderation, and Conditional Process Analysis (3rd edition) generates a data file
containing 5000 bootstrap estimates of the regression coefficients and constants from the
models of pmi and reaction. The column names in the file are labeled

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

 pmi_constant pmi_cond reaction_constant reaction_cond reaction_pmi

Thus, the first and second column contains the bootstrap estimates of the regression constant
and regression coefficient for cond, respectively, in the model of pmi. The third, fourth, and
fifth columns contain the regression constant and regression coefficients for cond and pmi,
respectively, in the model of reaction. In the R version of PROCESS, the case (i.e., upper or
lower) of the column labels will be consistent with the case of the variable names in the data
frame being analyzed.

Note that the SPSS version of PROCESS 5.0 still uses the old column naming system and
generates the column map in the output, just as in prior releases.

With this new column labeling format, some of the code printed in the 3 rd edition of
Introduction to Mediation, Moderation, and Conditional Process Analysis needs to be modified
to work with PROCESS version 5. For the R version, the following changes are needed:

Line 9 of the code starting on page 446:

bootind<-boots$negtone_dysfunc*(boots$perform_negtone+boots$perform_int_1*
modval[i])

Line 9 of the code on page 479:

bootind<-(boots$justify_frame+boots$justify_int_1*modval[i])*boots$donate_justify

Page 614:

result<-process(data=pmi,y="reaction",x="cond",m="pmi",total=1,normal=1,model=4,
seed=31216,save=1)
ab<-result$pmi_cond*result$reaction_pmi
hist(ab,breaks=25)
diff<-result$reaction_cond-ab
quantile(diff,c(.025,.975))

For the SAS version, the second line of code on page 618 should now be:

data boots;set boots;ab=pmi_cond*reaction_pmi;diff=reaction_cond-ab;run;

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

Regression Analysis without Moderation or Mediation
(Added in version 5.0)

PROCESS version 4.1 added some limited features for the estimation of regression models
without a moderation or mediation component. The release of version 5.0 both simplifies the
command line and greatly expands the ability of PROCESS to estimate ordinary regression
models and various extensions. Although formally designated as model 0, all that is needed in
the PROCESS command is a single outcome variable after y= and at least one variable after x=,
as below

process y=govact/x=negemot posemot ideology sex age/model=0.

%process(data=glbwarm,y=govact,x=negemot posemot ideology sex age,model=0)

process(data=glbwarm,y="govact",x=c("negemot","ideology","sex","age"),model=0)

The inclusion of “model=0” in the PROCESS command is optional, as PROCESS will understand
what to do when no mediator or moderator is specified in the command line.

When estimating a regression model with no moderation or mediation component, the default
setting for the ssquares option (described earlier) is 1, meaning that PROCESS will generate a
sum of squares table as well as adjusted R2 for the model. To eliminate this from the output, set
the ssquares option to 0 (i.e., ssquares=0).

The variables on the right side of the regression equation need not all be entered following x=.
An alternative option is to include at least one variable in the x= list and the remaining
regressors as covariates following cov= as below.

process y=govact/x=negemot posemot/cov=ideology sex age.

%process(data=glbwarm,y=govact,x=negemot posemot/cov=ideology sex age)

process(data=glbwarm,y="govact",x=c("negemot","negemot"),cov=c("ideology","sex",
"age"))

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

The model and corresponding output will be identical to when all variables are listed following
x=. Breaking the variables up in this fashion into x= and cov= sets can be useful when using the
settest option discussed below to conduct a test of equality of fit of the models that include
and exclude the variables in the x= list.

The mcx option can be used in model 0, in which case the multicategorical variable should be
listed first in the x= list, and PROCESS will automatically create category codes as described in
Appendix A of Introduction to Mediation, Moderation, and Conditional Process Analysis. Any
other multicategorical variables on the right side of the equation would have to be already
represented in the data with such codes. Multicategorical variables listed in cov= must be
properly represented with a categorical coding system with the codes (e.g., indicator, Helmert,
etc.) generated outside of PROCESS.

Many other features available in all models that PROCESS can estimate are available in model 0,
including heteroskedasticity-consistent inference (using the hc option), bootstrapping (with the
modelbt option), as week as cluster-robust standard errors and errors-in-variables regression,
discussed later in this document. Some additional features described next are also available in
model 0.

All Subsets Regression

The subsets option conducts all subsets regression. When this option’s toggle is set to 1 (i.e.,
subsets=1), PROCESS generates output containing R2 and adjusted R2 for all possible models
containing at least one regressor. The output takes the form of a table with the variable names
at the top and models occupying the rows, as below. The table entries for each row contain
zeros and ones under the variable name. A one in the column designates that the variable in
that column is included in the model, and a zero means that variable is excluded. The table
rows are sorted in ascending order of the adjusted R2 for the model.

All subsets regression results

negemot posemot ideology sex age R-sq Adj R-sq
.0000 1.0000 .0000 .0000 .0000 .0019 .0006
.0000 .0000 .0000 .0000 1.0000 .0094 .0082
.0000 .0000 .0000 1.0000 .0000 .0097 .0085
.0000 1.0000 .0000 .0000 1.0000 .0117 .0092
.0000 1.0000 .0000 1.0000 .0000 .0123 .0098
.0000 .0000 .0000 1.0000 1.0000 .0164 .0140
.0000 1.0000 .0000 1.0000 1.0000 .0193 .0156
.0000 .0000 1.0000 .0000 .0000 .1750 .1740
.0000 .0000 1.0000 .0000 1.0000 .1751 .1730
.0000 1.0000 1.0000 .0000 .0000 .1759 .1739
.0000 1.0000 1.0000 .0000 1.0000 .1760 .1730
.0000 .0000 1.0000 1.0000 .0000 .1769 .1748
.0000 .0000 1.0000 1.0000 1.0000 .1769 .1738
.0000 1.0000 1.0000 1.0000 .0000 .1781 .1750
.0000 1.0000 1.0000 1.0000 1.0000 .1781 .1740
1.0000 .0000 .0000 .0000 .0000 .3338 .3330
1.0000 .0000 .0000 1.0000 .0000 .3348 .3331
1.0000 1.0000 .0000 .0000 .0000 .3348 .3331
1.0000 1.0000 .0000 1.0000 .0000 .3356 .3331

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

 1.0000 .0000 .0000 .0000 1.0000 .3379 .3363
1.0000 .0000 .0000 1.0000 1.0000 .3384 .3359
1.0000 1.0000 .0000 .0000 1.0000 .3387 .3363
1.0000 1.0000 .0000 1.0000 1.0000 .3391 .3358
1.0000 .0000 1.0000 .0000 .0000 .3873 .3858
1.0000 .0000 1.0000 1.0000 .0000 .3874 .3851
1.0000 .0000 1.0000 .0000 1.0000 .3876 .3853
1.0000 .0000 1.0000 1.0000 1.0000 .3876 .3846
1.0000 1.0000 1.0000 .0000 .0000 .3881 .3858
1.0000 1.0000 1.0000 1.0000 .0000 .3881 .3851
1.0000 1.0000 1.0000 .0000 1.0000 .3883 .3853
1.0000 1.0000 1.0000 1.0000 1.0000 .3883 .3845

The number of possible models explodes as the number of regressors increases, and computing
time and memory requirements increase accordingly. For this reason, all subsets regression is
available only for models that include 15 or fewer regressors. All subsets regression is not
available for models that specify moderation or mediation, models with a dichotomous Y, or
when used in conjunction with the cluster option.

The subsets option is available only for model 0.

Dominance Analysis

Dominance analysis is a method for determining the relative importance of regressors in a

model. Output from a dominance analysis is requested by specifying dominate=1 in the
PROCESS command line. PROCESS will display a dominance table, as below, discussed in section
8.4 of Darlington and Hayes (2017). The entries in the dominance table are the proportion of
the possible subset models in which the variable in the row contributes more to prediction
accuracy than the variable in the column. The diagonals of the dominance table are zero, and
the cells symmetrically located around the diagonal usually sum to one.

Dominance matrix
negemot posemot ideology sex age
negemot .000 1.000 1.000 1.000 1.000
posemot .000 .000 .000 .500 .500
ideology .000 1.000 .000 1.000 1.000
sex .000 .500 .000 .000 .500
age .000 .500 .000 .500 .000

Dominance analysis requires a lot of computations that require time and memory.
Consequently, dominance analysis is available only for models with 15 or fewer regressors. In
addition, dominance analysis is not available for models that specify moderation or mediation,
models with a dichotomous Y, or when used in conjunction with the cluster option.

Spline Regression

PROCESS can conduct spline regression, discussed in section 12.3 of Darlington and Hayes
(2017), wherein separate linear models relating one variable to the outcome are estimated
between joints defined by user-specified values on the measurement scale. Spline regression is

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

conducted by using the spline option, followed by a list of joint values separated by commas (in
SPSS), spaces (in SAS), or using the c() operator (in R). The joint values should be values on the
measurement scale of the variable listed first in the x= list. For example, the command below

process y=govact/x=age negemot posemot ideology sex/spline=30,40,50.

%process(data=govact,y=govact,x=age negemot posemot ideology sex,spline=30 40 50)

process(data=glbwarm,y="govact",x=c("age","negemot","negemot","ideology","sex"),
spline=c(30,40,50))

specify splines for the age variable, with the joints defined at ages 30, 40, and 50. Up to 10
joints may be specified when using the spline option. Joint locations must be listed in ascending
order of value, with no ties, and all spline segments must contain at least two cases. The
variable listed first following x= cannot be multicategorical, and so the spline option is
incompatible with the mcx option. To get a test for the set of variables that define the spline
function, use the settest option described next. The features of the spline option cannot be
accessed through the PROCESS dialog box in SPSS.

The spline option can also be used in a mediation analysis without a moderation component
(e.g., models 4, 6, 80, 81).

Inference for Sets of Regressors

PROCESS can provide a test that all of the regression coefficients for a subset of the regressors
in the model are zero. In a regression model that includes any covariates listed following cov=
PROCESS automatically provides a test that the partial regression coefficients for all the
variables in the x= list are equal to zero. This is equivalent to a test of equality of fit of two
models, one that includes only the variables in the cov= list and a second that includes variables
in the cov= and the x= list. For example, the command

process y=govact/x=negemot posemot/cov=ideology sex age.

%process(data=glbwarm,y=govact,x=negemot posemot,cov=ideology sex age)

process(data=glbwarm,y="govact",x=c("negemot","posemot"),cov=c("ideology","sex",
"age"))

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

estimates govact from negemot, posemot, ideology, sex, and age while also
providing a test of the null hypothesis that the regression coefficients for both negemot and
posemot are both equal to zero. The difference in R2 between the models with and without
negemot and posemot is converted to an F-ratio for the test the null hypothesis, as below.
PROCESS will also provide the difference in R2 between the full model and the model excluding
posemot and negemot. This will show up in the output next to the F test as “R2-chng.”

Hypothesis test for variables in X set:

R2-chng F df1 df2 p
.2114 139.8217 2.0000 809.0000 .0000

When only one variable is listed for X and that variable is specified as multicategorical using the
mcx option, the test is equivalent to a single factor analysis of covariance comparing the group
means, adjusting for differences between the groups on all variables following cov=. When the
variable listed following y= is dichotomous, the test printed by PROCESS will be in the form of a
likelihood ratio test.

When using the spline option, all of the variables that define the spline function for the first
variable following x= are included in the set, but this test is only conducted when adding
settest=1 to the PROCESS command.

This test for sets of variables is available only for model 0.

Cluster-Robust Standard Errors

(added in version 5.0)

PROCESS cannot do multilevel or “random effects” regression analysis. However, as of the

release of PROCESS version 5.0, PROCESS can generate standard errors and conduct inferential
tests that account for the within-cluster nonindependence that is characteristic of data with a
hierarchical or nested structure. A variable defining the clusters must be specified following
cluster= and cluster robust standard errors requested by adding robustse=1 to the PROCESS
command. For example, the command

process y=vote/x=ideology/m=conflict/model=4/cluster=country/robustse=1.

%process(data=civic,y=vote,x=ideology,m=conflict,cluster=country,robustse=1)

process(data=civic,y="vote",x="ideology",m="conflict",cluster="country",robustse=1)

identifies country as the clustering variable while requesting cluster robust standard errors
for inference.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

The algorithm for computation of cluster robust standard errors that is implemented in
PROCESS is described in Cameron and Miller (2005, pp. 323-325) and mimics the vce(robust)
option in Stata and PROC SURVEYREG in SAS. Like Stata, when using the robustse option,
degrees of freedom for t-statistics, residual degrees of freedom (denominator) of F-tests, and
for the construction of confidence intervals is the number of clusters g minus 1 (PROC
SURVEYREG uses the number of clusters g).

F-tests for the model or subsets of variables in the model are conducted using the cluster-
robust covariance matrix of the regression coefficients. However, if the number of clusters is
too small relative to the number of variables (the numerator degrees of freedom) used for the
test, PROCESS may not be able to conduct the test. In that case, F-ratios, degrees of freedom,
and p-values for F-tests will be listed as “99999” in the output. These should not be interpreted.
Consider this a warning that the number of clusters is far too small for reliable inference.

Cluster-robust inference is not available for models that include a dichotomous Y. As cluster-
robust standard errors also account for heterogeneity of variance in the errors in estimation,
the hc option cannot be used in conjunction with robustse.

Clustered and Stratified Bootstrapping

(added in version 5.0)

By default, PROCESS uses the casewise bootstrap when generating bootstrap estimates and
confidence intervals. With the casewise bootstrap, each case in the data has equal probability
of being included in a bootstrap sample. With the release of version 5.0, two new
bootstrapping options are available. With both options, the user specifies a single clustering/
stratification variable in the data following “cluster=” in the PROCESS command that identifies
in which cluster or stratum a case resides. A cluster or stratum is operationalized in the data as
cases with a common numerical value on the clustering variable. In the rest of this discussion,
the term “cluster” is used to refer to both clusters and strata, as the distinction between these
often made in the sampling literature is not pertinent to the mechanics of the bootstrapping
procedure described below.

With a cluster variable specified, one of two bootstrapping options is implemented depending
on the argument following clusboot=. Let N be the sample size, k be the number of clusters,
and nj be the number of cases in cluster j. Adding clusboot=1 to the PROCESS command
implements a bootstrapping procedure such that each bootstrap sample will contain cases from
all k clusters and with exactly nj cases from cluster j. Within cluster j, cases in that cluster are
randomly sampled with replacement and have the same probability of inclusion in a bootstrap
sample as do other cases in cluster j. This procedure ensures that all k clusters are represented
in every bootstrap sample, with each bootstrap sample containing exactly nj cases from cluster j
while also ensuring that each bootstrap sample has exactly N cases.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

An example of when this bootstrapping procedure can be useful is when an analysis includes
groups whose sample sizes were fixed by the researcher in advance. For example, suppose X is
a multicategorical variable defining three groups and the investigator conducted the study so
that 50 cases would be in each group. If 50 cases per group was fixed by design, it would make
sense to restrict the bootstrap sampling to ensure that each bootstrap sample also contains 50
cases from each group. Treating the group variable as the cluster variable and using clusboot
option 1 will guarantee that each bootstrap sample contains 50 cases from group 1, 50 cases
from group 2, and 50 cases from group 3.

A second cluster bootstrapping option is available that randomly chooses k clusters with
replacement and then includes all cases in each randomly selected cluster in the bootstrap
sample. This option is requested by adding clusboot=2 to the PROCESS command line. Unlike
when using clusboot option 1, there is no guarantee that any cases from cluster j will appear in
a bootstrap sample. Furthermore, a bootstrap sample may contain more or fewer than N cases,
depending on the size of the clusters that were randomly selected for inclusion.

When using either of these cluster bootstrapping options, the computation of bootstrap
confidence intervals (as well as bootstrap standard errors) is conducted in exactly the same
manner as when using the casewise bootstrap. For a discussion of the mechanics of
bootstrapping and the construction of bootstrap confidence intervals, see chapter 3 of
Introduction to Mediation, Moderation, and Conditional Process Analysis.

Note that just as is true for the casewise bootstrap, when using the clusboot option, the
standard errors and confidence intervals for each model in the output are still computed using
ordinary OLS regression formulas unless the robustse or hc options are also used. Bootstrap
results (confidence intervals, standard errors, and the mean of the bootstrap estimates) are
displayed in the output only in those output columns with “Boot” in the label.

Errors-in-Variables Regression
(Added in version 5.0)

As of PROCESS version 5.0, errors-in-variables regression is available for the estimation of some
models PROCESS can estimate. Errors-in-variables regression can be used to reduce or
eliminate the bias in the estimation of regression coefficients as well as statistical inference
when variables on the right side of a regression equation contain random measurement error.
For the formulas used by PROCESS for estimation of errors-in-variables regression coefficients
and various standard error options, see Appendix A of Hayes, Allison, and Alexander (2024).

Errors-in-variables regression is estimated by PROCESS whenever the relx, relm, or relcov

options are used in the PROCESS command. These options allow the user to enter the assumed
or estimated reliability of the variables on the right-hand side of regression equations. For
variables in the x= list, provide the reliabilities for each variable in the list following option relx=
and in the same order the variables are listed in the x= list. If no reliabilities are provided (that

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

is, if relx is not included in the command), PROCESS assumes the variables in the x= list are
measured without error (i.e., with reliabilities equal to 1). In SPSS, the reliabilities in the list
should be separated by a comma. In SAS, separate the reliabilities with a space. In R, separate
the reliabilities with a comma and enclose the entire list in the c() operator. Unknown
reliabilities would ordinarily be set to 1 (which in effect treats that variable as it would be
treated in ordinary regression analysis assuming no measurement error). The same rules apply
for relm and relcov when entering the reliabilities for the M variables in the m= list using relm
and covariates in the cov= list using relcov. As with relx, when the relm or relcov options are
not used, PROCESS assumes these variables contain no random measurement error.

For example, the PROCESS command below

process y=withdraw/x=estress/m=affect/cov=ese sex tenure/relx=0.72/relm=0.88

/relcov=0.94,1,1.

%process(data=estress,y=withdraw,x=estress,m=affect,cov=ese sex tenure,relx=0.72,

relm=0.88,relcov=0.94 1 1)

process(data=estress,y="withdraw",x="estress",m="affect",cov=c("ese","sex",
"tenure"),relx=0.72,relm=0.88,relcov=c(0.94,1,1))

estimates the economic stress mediation analysis described in Chapter 4, section 4.2, of
Introduction to Mediation, Moderation, and Conditional Process Analysis. The reliability
estimates discussed below and in the PROCESS command above are provided in the original
Journal of Organizational Behavior article. The reliability of the data for economic stress
(estress), which is X in the model, is set to 0.72, and for business related depressed affect
(affect), the mediator M, reliability is set to 0.88. The model includes three covariates. Sex
and years in business (tenure) are assumed to be measured without any random
measurement error and so the reliabilities are set to 1. But the reliability of entrepreneurial
self-efficacy (ese) is set to its estimated value of 0.94.

The errors-in-variables option can also be useful to ascertain how vulnerable an analysis that
assumes perfect reliability is to unaccounted-for measurement error. This can be accomplished
by setting the reliabilities to plausible values or values lower than are likely and executing the
analysis to see if the results substantively change. If not, then one can conclude that the results
that assume perfect reliability are likely invulnerable to unaccounted-for measurement error.

Errors-in-variables regression makes an adjustment to the variances of variables on the right

sides of regression equations prior to estimating the regression coefficients and standard
errors. This adjustment can produce a variance-covariance matrix of those variables that could
not actually exist in nature. This typically occurs when one or more of the assumed reliabilities

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

entered is small, though it can happen in other circumstances as well. When it does, PROCESS
will not estimate the model and an error message is generating stating that one or more of the
assumed reliabilities is too small to estimate the model. Note that an impossible data matrix
after adjustment can also occur during bootstrapping. When this occurs, the bootstrap sample
will be replaced. A warning message at the bottom of the output will indicate how many
bootstrap samples were replaced during the bootstrapping procedure. If this number is large,
interpret bootstrap results with caution. There are no guidelines or rules of thumb for what
counts as “large” in this situation.

The adjustment to the data requires a different approach to estimating standard errors. These
approaches, described in Hayes, Allison, and Alexander (2024), are available by including the eiv
option in the PROCESS command. By default (if no eiv option is used or by adding eiv=3 to the
command), PROCESS implements a method that accounts for the sampling variance that results
when adjusting for random measurement error and also includes a heteroskedasticity-
consistent component based on the HC3 estimator discussed in Long and Ervin (2000). When all
the reliabilities are set to 1, the regression coefficients will be the same as those estimated with
ordinary least squares, and the standard errors will be equivalent to the heteroskedasticity-
consistent HC3 or “MacKinnon-White” standard error estimator.

Some alternative standard error estimators are also available. By including eiv=0 in the
PROCESS command along with estimated reliabilities, PROCESS uses the method implemented
in Stata15 and later releases and discussed in StataCorp (2023). This method includes a
heteroskedasticity-consistent component based on the HC0 estimator, also known as the
“Huber-White” estimator. A third standard error option implemented in Stata prior to version
15 and discussed in Lockwood and McCaffrey (2020) is available using the eiv=5 option in the
PROCESS command. Like the default approach, this alternative approach adjusts the standard
errors for unreliability but does not include a heteroscedasticity-consistent component. When
all reliabilities are set to 1, the standard errors produced by this approach will be equivalent to
regular OLS standard errors.

Errors-in-variables regression estimation is not available in moderation models (models 1, 2, or

3), in conditional process models that combine moderation and mediation, or in models with a
dichotomous Y. Furthermore, the following options are not available for use in conjunction with
errors-in-variables regression: crossv, diagnose, dominate, effsize, hc, modelres, robustse,
spline, stand, spline, ssquares, subsets, and save option 4 (saving regression diagnostics).

Automatic Plot Generation in PROCESS for R

(added in version 5.0)

When estimating a model with a moderation component, the plot option generates a table of
estimates of the outcome variable (on the left side of an equation) from various combinations
of focal predictor and moderator(s). In the R version of PROCESS version 5 or later, the plot
option will now also automatically generate a visual depiction of the corresponding model. In

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

addition, when using the Johnson-Neyman technique (with option jn=1) in conjunction with the
plot option will produce a Johnson-Neyman plot visually depicting the relationship between the
moderator and the effect of the focal predictor along with confidence interval endpoints. Note
that the production of the Johnson-Neyman plot is sensitive to the probing filter and will only
be generated when the p-value for the corresponding interaction is below the filter setting
(0.10 by default; see the documentation for more information about filter implemented with
the intprobe option).

For example, executing the command below using the teams data from Chapter 11 of
Introduction to Mediation, Moderation, and Conditional Process Analysis

process(data=teams,x="dysfunc",m="negtone",y="perform",w="negexp",plot=1,jn=1,
model=14)

automatically generates the plots below:

There is no way of modifying the axis labels, scaling of the axes, style or colour of lines, or
specifying which variable is placed on the horizontal axis of the plots produced. To customize a
plot, paste it into a graphics program and manually modify sections of the plot you wish to
modify. For moderation models, PROCESS will always place the focal predictor on the horizontal
axis and values of the moderator will determine the lines in the plot, unless the focal predictor
is dichotomous or multicategorical, in which case the moderator will be placed on the
horizontal axis and groups define the lines.

Note that a visual depiction of the model will only be generated for models or sections of a
model with a single moderator. In other words, if more than one variable is specified as
moderating a focal predictor’s effect in a moderation-only model (i.e., models 2 or 3) or

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

sections of models (e.g., the effect of X on M in models 9-13, or the effect of M on Y in models
16-20), no plot is generated for that effect.

Model Number Defaults

(added in version 5.0)

Prior to the release of version 5.0, a PROCESS command must always contain a model number
unless a custom model was being constructed with the use of the bmatrix option. With the
release of version 5.0, PROCESS will assume model 0, model 1, or model 4 in some
circumstances and depending on your PROCESS command, eliminating the need to specify a
model number for these models.

If your PROCESS command does not specify a mediator variable M or moderator variable Z but
does include a moderator variable W, it will assume you want to estimate a simple moderation
model (model 1). Thus, a command such as below will work without a model number:

process y=justify/x=frame/w=skeptic.

%process(data=disaster,y=justify,x=frame,w=skeptic)

process(data=disaster,y="justify",x="frame",w="skeptic")

If your PROCESS command includes no moderator variables (i.e., no W or Z variable is specified)

but at least one mediator variable M is specified, PROCESS will assume you want to estimate a
simple or parallel multiple mediator model (model 4). If more than one variable is listed as M,
PROCESS will estimate a parallel multiple mediator model. Thus, a command such as below will
work without a model number:

process y=reaction/x=cond/m=import pmi.

%process(data=pmi,y=reaction,x=cond,m=import pmi)

process(data=pmi,y="reaction",x="negemot",m=c("import","pmi"))

If your PROCESS command includes no mediator (M) or moderator variables (W and Z),
PROCESS will assume you are estimating a regular OLS or logistic regression model without a

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

moderation or mediation component (model 0). Thus, the command below works without a
model number:

process y=govact/x=negemot posemot ideology sex age.

%process(data=glbwarm,y=govact,x=negemot posemot ideology sex age)

process(data=glbwarm,y="govact",x=c("negemot","ideology","sex","age"))

New Bootstrap Performance Information When Saving Output

(added in version 5.0)

As discussed in the documentation, save option 2 produces a data file containing the numerical
information in the PROCESS output. With the release of version 5, and only when bootstrapping
is used to generate any section of the output, the last row of this data file will contain
information about the performance of the bootstrapping algorithm. The first column will
contain the number of bootstrap samples that had to be replaced during the bootstrapping
procedure. The second column contains how many samples were replaced due to a singularity
in the bootstrap sample. The last column indicates how many samples were replaced as a result
of not being able to apply the errors-in-variables computations on a bootstrap sample.

Output Formatting in SPSS

(added in version 5.0)

By default, the SPSS version of PROCESS produces output in text format. With the release of
version 5, a new display option is available. By adding display=tables to the PROCESS command
in SPSS, certain sections of the output will be in the form of table objects rather than text that
can more easily be edited and resized in other documents if desired.

Note that with the release of SPSS v29, IBM changed the default output font for text output
such as generated by PROCESS. The new default font will produce sloppy-looking output, with
information not properly formatted and spaced. To return the format of the output to pre-v29
form, follow the directions below.

Under the “Edit” menu in SPSS, Choose “Options”. The window below will open. Change the
font under “Text Output” to “Courier New” and click the “Apply” button and then “OK” at the
bottom of the window.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION

New PROCESS menu is an Extension file
(added in version 5.0)

With the release of version 5.0, SPSS users interested in installing the PROCESS dialog box to set
up a model must install a custom dialog extension file (“.spe”) rather than a custom dialog
builder file (“.spv”). To do so, select “Extensions”-> “Install Local Extension Bundle..” and
choose the .spe file that comes in the PROCESS v5 archive. After doing so, the PROCESS menu
can be found under “Analyze”->”Regression”. The custom dialog builder file (.spv) has been
discontinued as of version 5 and is no longer available.

References

Cameron, A. C., & Miller, D. L. (2015). A practitioner’s guide to cluster-robust inference. Journal
of Human Resources, 50, 317-382.

Hayes, A. F., Allison, P. D., & Alexander, S. M. (2024). Errors-in-variables regression as a viable
approach to mediation analysis with random error-tainted measurements: Estimation,
effectiveness, and an easy-to-use implementation. Manuscript submitted for
publication.

Long, J. S., & Ervin, L. H. (2000). Using heteroskedasticity-consistent standard errors in the
linear regression model. American Statistician, 54, 217-224.

Lockwood, J. R., & McCaffrey, D. F. (2020). Recommendations about estimating errors-in-

variables regression in Stata. The Stata Journal, 20, 116-130.

StataCorp (2023). Stata 18 Base Reference Manual. College Station, TX: Stata Press.

Copyright 2023-2025 by Andrew F. Hayes DO NOT POST ONLINE WITHOUT PERMISSION