30 Mediation, Moderation, and Conditional Process Analysis

Regression-Based Approaches for Clinical Research

NICHOLAS J. ROCKWOOD AND ANDREW F. HAYES

Establishing causal effects is the focus of much clinical conditions, enhancers, and inhibitors of an effect allows
psychology research, such as identifying if a new sub- treatments to be tailored to different people, or groups of
stance abuse treatment program is more effective than people, in an effort to maximize treatment effectiveness
treatment as usual, or if the experience of certain psycho- and reduce wasted resources (Finney, 1995; Magill, 2011;
logical symptoms results in certain negative physical Norcross & Wampold, 2011). These questions address the
health outcomes. But merely identifying the existence of role of individual differences and situational contexts in
such effects is not sufficient. Knowing that, for instance, their impact on causal effects of interest, which can also
some treatment is more effective than another ‒ a question further help test and build theory.
of whether or if ‒ raises the question of how or why this The questions of how and why refer to identifying or
effect exists or operates. For example, perhaps a treatment testing the mechanism, or process, that underlies an
for alcohol abuse reduces substance cravings which trans- effect, whereas the questions of for whom and when refer
late to lower use, or maybe it encourages patients to seek to identifying the contingencies of or conditions on that
support from others, who in turn discourage use or help a effect. When combined with strong theory, statistical
person find alternative means of dealing with their anx- mediation analysis and statistical moderation analysis can
ieties, depression, or other factors that give rise to be used to explore or test hypotheses about the mechan-
substance use. isms and contingencies of effects, respectively.
Several clinical researchers have observed the import- Exemplars of mediation and moderation analysis are in
ance and advantages of understanding how causal effects abundance in published clinical research, yet many
operate. In the context of treatments and interventions, researchers still struggle with the distinction between
these advantages include the ability to target key compon- them, both conceptually and statistically. The purpose of
ents of an intervention while placing less emphasis on this chapter is to provide an overview of mediation and
ineffective components to enhance the treatment as a moderation analysis, mapping the concepts to statistical
whole, as well as better understanding of the underlying models. Further, we describe the integration of mediation
psychological disorders addressed by an effective treat- and moderation into a unifying framework known as con-
ment (Frazier, Tix, & Barron, 2004; Kazdin, 2007; Krae- ditional process analysis (Hayes, 2018). After detailing the
mer et al., 2002; Magill, 2011; Onken et al., 2014; Schmidt framework, we present an example of a conditional pro-
& Schimmelmann, 2015; Windgassen et al., 2016). Fur- cess analysis, with the hope that it will provide researchers
thermore, understanding the nature of effects is important insight into how to construct and test a conditional pro-
for theory building and testing competing theories of psy- cess model related to their own research questions and
chological phenomena and experiences. hypotheses of interest.
Seemingly related questions, but conceptually distinct This is a vast topic, and the methodology literature in
from how an effect operates, are for whom or when does this area is massive, so it is necessary for us to focus our
the effect operate or exist and for whom or when does it discussion, lest it be too vague to be useful. Throughout
not. Is the treatment effective for everyone? Most likely the chapter, we focus on the analysis of continuous out-
not. Maybe it is effective for people with addictions to comes that can be estimated with linear regression analy-
certain classes of drugs, but not for people addicted to sis using the least squares criterion for optimizing fit of a
other kinds of drugs. Or perhaps the effectiveness of the model. Most clinical researchers are familiar with least
treatment depends on the amount of emotional support squares regression analysis, and it is built into most statis-
that the patient receives. Individual and clinician charac- tical software and is a staple of graduate training in the
teristics are also likely to influence the effectiveness of a social and health sciences. Linear models are most com-
particular treatment. Understanding these boundary monly used when answering questions about mediation
396

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 397

and moderation, and understanding the concepts and stat-

A B M1
istical methods in this framework provides a nice founda-
M
tion to build upon. We also limit our discussion to M2
scenarios involving dichotomous (e.g., a single treatment X . Y
versus a control condition, as in a simple clinical trial) or X Y .
.
continuous independent and moderator variables.
Mk
Readers interested in mediation or moderation models
involving multicategorical independent variables can con-
sult Hayes and Preacher (2014), Hayes and Montoya C M1 D
(2017), and Hayes (2018). Although our discussion is W
limited to studies with these data properties, many of the
X Y
concepts introduced here can be extended to more com-
plex types of data and research designs with alterations, X Y
sometimes minor, sometimes more significant. M2

Figure 30.1 A simple mediation model (panel A), parallel mul-

MEDIATION tiple mediator model (panel B), serial multiple mediator
model (panel C) and a conceptual representation of moderation
Statistical mediation analysis is frequently used to test (panel D)
hypotheses about the process by which one variable X
causally transmits its effect on another variable Y through
one or more intervening or mediator variables M. dichotomous and the two categories are coded one unit
A mediation process takes the form of a train of causal apart (e.g., therapy as usual = 0, new treatment = 1), c is
events of the form “X influences M, which, in turn, influ- the mean difference in Y between the two groups. In the
ences Y.” The purpose of mediation analysis is to identify parlance of mediation analysis, c is referred to as the total
and explain how the effect of X on Y operates. Several effect of X.
clinical researchers have implemented mediation analysis Although Equation 30.1 is often of substantive interest
to assess the mechanism by which an effect operates. For to researchers, in mediation analysis it plays only a per-
example, Mason and colleagues (2015) found that relative ipheral role in interpretation. The purpose of mediation
to participants assigned to a control condition, partici- analysis is to better understand the causal process by
pants in a text-based smoking-cessation intervention pro- which, or how, X affects Y through a sequence of causal
gram had enhanced feelings of readiness to stop smoking, events in which X influences at least one mediator variable
which was in turn related to fewer cigarettes smoked M which, in turn, causally influences Y. Some simple rules
during a 30-day period. And Duchesne and colleagues of path analysis algebra tell us that the total effect of X, c in
(2017) identified self-esteem as a mediator of the effect of Equation 30.1, can be partitioned into a direct effect and an
body dissatisfaction on depression and anxiety in adoles- indirect effect of X, estimated using a set of two regression
cents. Higher body dissatisfaction was associated with models
lower self-esteem, which resulted in higher depression
M = iM + aX + eM (30.2)
and anxiety.
Y = iY + c0X + bM + eY (30.3)
Figure 30.1, panel A, displays a path diagram correspond-
THE SIMPLE MEDIATION MODEL ing to this model. The direct effect, represented by c0 in
When M and Y are continuous, most researchers inter- Equation 30.3, quantifies the expected difference in Y
ested in mediation begin an analysis by estimating the between two units that differ by one unit in X, but have
following equation: the same value for M. Put another way, it is the effect of X
on Y while statistically controlling for, or holding con-
Y = iY + cX + eY (30.1)
stant, M. Thus, it is X’s effect on Y that is independent of
where Y is the ultimate outcome of interest (e.g., symp- M. Referring back to the dichotomous X example (therapy
toms experienced, frequency of substance use, etc.) and X as usual = 0, new treatment = 1), the direct effect is an
is the presumed cause (a new treatment versus therapy as adjusted mean difference on Y between the groups. It is
usual, or some continuous individual difference) of differ- an estimate of the difference between the groups on Y,
ences between people on Y. This is a simple linear regres- on average, if they were equal on the proposed mediator.
sion model, where iY is the intercept, eY is the error term, The coefficient a in Equation 30.2 quantifies the effect of
and c represents the association between X and Y. Specif- X on M, and the coefficient b in Equation 30.3 quantifies
ically, c estimates how much two cases that differ by one the effect of M on Y statistically controlling for X. The
unit on X differ on Y. Whether X is continuous or dichot- indirect effect of X on Y is the product of these effects.
omous, this interpretation does not change. But when X is Specifically, it is the product of a and b, ab, and it

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 398 ROCKWOOD & HAYES

represents the expected difference in Y for two cases that effect. This bootstrap distribution of ab values serves as an
differ by one unit on X due to X’s effect on M, which in turn empirical estimate of the sampling distribution of the
affects Y. Thus, ab is a quantification of the mediation indirect effect, and a 95 percent percentile-based boot-
process. When Equations 30.1–30.3 are estimated using strap confidence interval can be computed as 2.5th and
least squares regression analysis, the sum of the indirect 97.5th percentiles of the distribution of k bootstrap esti-
and direct effects is equal to the total effect. That is, mates. Simulation studies have demonstrated that com-
ab + c0 = c. Because the indirect effect is the product of pared to alternative methods, inference using a confidence
a and b, a positive indirect effect can correspond to a and b interval for the indirect effect constructed using this per-
values that are both positive or both negative, and a nega- centile-based bootstrap approach provides a good balance
tive indirect effect corresponds to a and b values that are of power, coverage, and Type I error rate (e.g., Hayes &
opposite in sign. Therefore, the signs of a and b are Scharkow, 2013).
important to interpretation of the sign of the indirect
effect.
Establishing a Total Effect
Historically, mediation analysis has been used only after
Inference about the Indirect Effect
one has established evidence of an association between X
Since the indirect effect quantifies the mediation process, and Y. This is predicated on the widespread belief that
an inferential test of the indirect effect should be under- without association between X and Y, there presumably
taken before concluding whether a pattern of relationships is no effect to explain, and therefore no reason to try to
is consistent with a mediation process. Yet because the establish or test the process by which that (non)effect
indirect effect is the product of the coefficients a and b, operates. That is, how can an effect that seems not to exist
inference about the indirect effect is not as straightforward be mediated? But it is now widely accepted among meth-
as it is for such statistics as a difference between two means odologists that a statistically significant total effect (path c
or a correlation coefficient. Simple, though now outdated, in the mediation model in Figure 30.1, panel A, estimated
methods for making inferences about the indirect effect with Equation 30.1) is not a requirement of mediation or a
include those based on the casual steps criteria and its prerequisite of mediation analysis (Cerin & MacKinnon,
variants (Baron & Kenny, 1986; Kraemer et al., 2008) and 2009; Hayes, 2009; Hayes & Rockwood, 2017; O’Rourke &
the Sobel test (Sobel, 1982). These methods are limited by MacKinnon, 2018; Rucker et al., 2011; Shrout & Bolger,
their reliance on a series of significance tests based on the 2002). As Bollen (1989, p. 52) aptly stated “a lack of correl-
individual a and b coefficients (rather than ab, which is ation does not disprove causation.”
what actually quantifies what matters in a mediation analy- This is counterintuitive. An example will make it less so.
sis) or an assumption that the sampling distribution of ab is But first, recall that although c, the total effect of X, quan-
normal or symmetrical, which typically is not the case tifies differences in Y between cases that differ on X, the
(Aroian, 1947). Today, many methodologists recommend total effect is the sum of the direct and indirect effects of X,
using bootstrap confidence intervals for inference about the and there is no mathematical restriction that these effects
indirect effect (Hayes, 2018; Hayes & Scharkow, 2013; be in the same direction. So the indirect effect of X on Y
MacKinnon, Lockwood, & Williams, 2004; Preacher & may be positive (or negative) while the direct effect is
Selig, 2012; Shrout & Bolger, 2002) as bootstrapping does opposite in sign. If these effects are similar in strength,
not make an assumption about the shape of the sampling their sum will be close to zero. Given this background,
distribution of the indirect effect. now consider the effect of aerobic activity (X) on body
Bootstrapping is a resampling procedure used to gener- weight (Y). Common sense and conventional wisdom
ate an empirical approximation of the sampling distribu- would predict a negative relationship between aerobic
tion of a statistic (Efron & Tibshirani, 1994) that can be activity and body weight. All else being equal, increasing
used for inference. In the context of mediation analysis, aerobic activity should result in a reduction in body
this statistic is the indirect effect, ab. To construct a boot- weight. Yet an increase in aerobic activity may result in
strap confidence interval for an indirect effect, the indirect consumption of more calories (calories consumed, M) due
effect is estimated in many bootstrap samples of the data. to an increase in metabolism, which could increase body
A bootstrap sample is constructed by taking a simple weight, all other things being held fixed. Thus, aerobic
random sample of n cases (or rows) from the existing data, activity could positively influence body weight through
where n is the original sample size, and cases are sampled calorie consumption indirectly, but when holding calorie
with replacement. The resulting bootstrap sample is then consumption constant, more aerobic activity should be
analyzed. Using the bootstrap sample, the indirect effect is negatively correlated with body weight. If the increase in
estimated using Equations 30.2 and 30.3 and then saved. weight resulting from aerobic activity operating indirectly
This process is repeated over and over, for a total of k through calorie consumption is as strong as the weight
times (where k is at least 1,000; more is better), producing reduction from aerobic activity operating independent of
a dataset containing k bootstrap estimates of the indirect calorie consumption, the result is no total effect of aerobic

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 399

activity on weight. Without identifying the mediating role randomly assigned to X. In that case, assuming other good
of calorie consumption in the (apparent lack of ) relation- experimental design practices are followed (such as elimin-
ship between X and Y, you could mistakenly conclude that ating contamination of manipulation by other variables and
aerobic activity is not an effective method of reducing differential attrition or dropout in one condition), the X!M
weight. But knowing the direct and indirect effects are (path a) and X!Y (paths c and c0 ) relationships can be
opposite in sign, this information could be used to develop inferred as causal because random assignment would
an intervention that targets and reduces the added calorie render the correlation between X and any potential con-
consumption that occurs as a result of aerobic activity, founds zero. However, random assignment to X does not
thereby allowing the negative effect of aerobic activity on ensure that M causes Y, as other variables that are correlated
weight to manifest itself. Some people call this phenom- with and influence M and Y could be responsible for the
enon suppression, though labeling it as such does not observed association between M and Y even when X is con-
facilitate and is not the same as understanding. trolled. To deal with this, potential confounding variables
Differences in sign of the direct and indirect effects are can be included as covariates in the equations for M and Y
not a requirement for finding an indirect effect absent a (see Hayes, 2018, for a discussion of the path analysis alge-
statistically significant total effect. Remember that c = c0 + bra). This can also help when X is merely observed rather
ab. Suppose the direct effect c0 equals zero but this is than manipulated. Unfortunately, one can never know for
estimated with considerably more uncertainty than the certain if the true confounds, if any, have been controlled.
indirect effect ab (i.e., the standard error of c0 is larger Cause-effect claims also presume proper causal ordering,
than the standard error of ab). As a result, the total effect with the cause preceding the effect. In the case of mediation,
c is estimated with more uncertainty than is the indirect we assume X temporally precedes M, which in turn precedes
effect ab, and so c is harder to detect than ab. Indeed, Y. Random assignment to X helps to meet this assumption
Kenny and Judd (2014) show that equivalent total and for the X!M and X!Y components of a mediation model.
indirect effects (i.e., when the direct effect is zero) are But it does not ensure that M precedes Y. It could be that Y
tested with inequivalent power. More specifically, tests precedes M, and so Y is actually a mediator of the effect of X
on the indirect effect (such as a bootstrap confidence on M. Temporal precedence is hard to establish in observa-
interval) tend to be conducted with higher power than tional (i.e., purely correlational) studies. For this reason,
tests of the total effect. Since the indirect effect is what some have argued that mediation analysis should not be
carries information about the mediation of the effect of X conducted with cross-sectional data (Maxwell & Cole,
on Y, rejection of the null hypothesis of no total effect is 2007; Maxwell, Cole, & Mitchell, 2011), with longitudinal
not a sensible prerequisite of mediation analysis. or experimental data collection being preferred. While we
agree that longitudinal data can help strengthen a cause-
effect claim, the perspective that mediation cannot or should
not be undertaken with correlational data is overly cynical
Causal Inference
for our taste. We believe that mediation analysis can be a
Mediation is a causal process. One cannot talk about useful tool with cross-sectional data, though the researcher
mediation without using causal language. Yet statistical will need to advance a strong theoretical justification for the
methods are agnostic to cause-effect and say nothing presumed temporal ordering in an effort to mitigate the
about whether an observed association is causal. An infer- limitation with the research design. Further, the possibility
ence is a product of the mind of the researcher interpret- of other causal orderings should be understood and
ing an analysis rather than the mathematics of the analysis addressed in discussion or analysis.
and what is found in a statistical output (see Hayes, 2018). The use of ab and c0 in Equations 30.2 and 30.3 as
The statistical tools presented throughout this chapter can estimates of the indirect and direct effects of X assumes
be used to piece together an argument, but causal claims that the effect of M on Y does not depend on X, which also
must rely on more than statistically significant indirect implies that X’s effect on Y does not depend on M.
effects. Within the mediation literature, there has been Recently, there has been research focused on identifying
growing attention on explicitly stating the assumptions the conditions in which this assumption can be relaxed.
that must hold for valid causal inference of indirect (and Discussions of modeling the interaction between X and M
direct) effects. Although these assumptions are often diffi- in a mediation model can be found in Hayes (2015, 2018),
cult or impossible to test, it is important to be aware of Imai, Keele, and Tingley (2010), Preacher, Rucker, and
them for research design and statistical analysis purposes. Hayes (2007), and Valeri and VanderWeele (2013).
One of the more important assumptions is that there are
no variables that confound the observed associations
between the variables in the putative causal system. An
More than One Mediator
association between two variables is confounded by a third
if that third variable causally influences those two variables. X can and often does transmit its effect on Y through
In mediation, this assumption can be met, but only partly, by multiple mechanisms simultaneously. The simple medi-
using an experimental design in which participants are ation model just described doesn’t capture the complexity

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 400 ROCKWOOD & HAYES

of many real-world processes, but mediation analysis is each of the k mediators, with the indirect effect through
easily extended to include additional mediators. Multiple mediator j quantified as the product of aj from Equation
mediator models are important for clinical research 30.4 and bj from Equation 30.5. Thus, as in the simple
because they allow researchers to test competing theories mediation model, indirect effects are products of the con-
while also facilitating a comparison between the sizes or stituent effects linking X to Y. In models involving multiple
magnitudes of more than one mechanism that may be mediators, the indirect effect through a given mediator is
responsible for the effect of X on Y. When X affects Y termed a specific indirect effect. The sum of the specific
through multiple processes simultaneously, a failure to indirect effects across all k mediators is the total indirect
accurately model such complexities can result in misstate- effect of X. For instance, in a three-mediator model (k = 3),
ments, misunderstandings, or, at best, an oversimplified the total indirect effect is a1b1 + a2b2 + a3b3. The total
characterization of how X affects Y. indirect effect is interpreted as the effect of X on Y operat-
In a model involving more than one mediator, there may ing through all mediators in aggregate, and a specific
or may not be a causal relationship among the mediators. indirect effect is the component of the total indirect effect
When there is no hypothesized causal pathway between that is unique to a particular mediator, independent of the
the mediators, the model is termed a parallel mediator indirect effect through all other mediators in the model.
model. Figure 30.1, panel B, displays an example of a The total indirect effect and the direct effect of X add up to
parallel mediation model with k mediators. For example, the total effect c, estimated with Equation 30.1.
Zhang and colleagues (2013) tested a mediation model in There are a few important properties of a parallel medi-
which wellbeing and coping were hypothesized as medi- ation model worth acknowledging. First, the indirect
ators of the relationships between a culturally sensitive effect of X through a given mediator is expected to change
intervention program for suicidal African-American with the inclusion of additional mediators if the additional
women and depressive symptoms and suicide ideation. mediators are correlated with the mediator in question.
They found that existential wellbeing mediated each of The b path for mediator j in Equation 30.5 (i.e., the esti-
these relationships, while religious wellbeing, adaptive mated effect of Mj on Y) will vary depending on the other
coping, and maladaptive coping did not. mediators in a model, just as regression coefficients in a
A mediation model in which one or more mediators regression analysis are dependent on what other variables
exerts its effect on another mediator is known as a serial are in the model. Second, the specific indirect effects are
mediation model. Figure 30.1, panel C, displays an example not influenced by the scale of the mediators. That is, each
of a serial mediation model with two mediators. These specific indirect effect is scaled only in terms of the scales
models have also been popular in clinical psychology of X and Y. This makes specific indirect effects through
research. For instance, Stanley, Joiner, and Bryan (2017) different mediators directly comparable even when the
found that anger and depression mediated the relationship measurement scales of the mediators are different
between mild traumatic brain injury and suicide risk for a (Preacher & Hayes, 2008). As a consequence, you can
clinical sample of deployed military personnel. In their compare specific indirect effects to test whether they are
model, anger was estimated as causally preceding depres- the same or different. Two different theories may postu-
sion. According to the theory they were testing, brain injury late different mediators of the effect of a therapeutic inter-
causes anger, which in turn causes depression, which sub- vention on some kind of treatment outcome. If both
sequently leads to an elevated risk of suicide. theories’ mediators are in the model simultaneously, you
Models that combine parallel and serial models can also can test whether one theory’s specific indirect effect is
be theorized and estimated. Space precludes an exhaustive statistically different from another theory’s. This is a more
discussion of the mathematics of complex multiple medi- refined test of competing theories than asking which indir-
ator models, so we focus on the parallel multiple mediator ect effect is different from zero, or whether one or both is.
model here, as doing so will help to inform our discussion Difference in significance does not imply statistical differ-
of the example conditional process analysis introduced ence, and equivalence in significance does not equate with
later. For a discussion of the serial and combined parallel statistical equivalence. For a more detailed discussion of
and serial model, see Hayes (2018). mechanics of comparing specific indirect effects, see
The direct and indirect effects of X on Y in a parallel Hayes (2018) or Preacher and Hayes (2008).
multiple mediator model, as in Figure 30.1, panel B, can
be estimated with a set of k + 1 regression equations, k for
the mediators (one for each mediator), and one for Y:
MODERATION
Mj = iMj + ajX + eMj for j = 1 to k (30.4)
Moderation is used to explore and test the conditional
Xk
0 nature of effects. X’s effect on Y is moderated if the strength
Y ¼ iY +c X + bM
j¼1 j j
+eY (30.5)
or direction of the causal effect of X on Y depends on the
The direct effect of X is c0 in Equation 30.5 and quantifies value or level of a third variable, with that third variable
the effect of X on Y operating independent of all k medi- called a moderator. Moderators can be either categorical
ators. Now there are multiple indirect effects, one through or continuous, and they may moderate the effect of a

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 401

categorical or continuous variable. For example, Shanmu- To estimate the model, Equation 30.7 is typically
gam and Davies (2015) were interested in the relationship expanded by distributing X across the linear function,
between self-critical perfectionism and eating pathology resulting in
in athletes. They found that gender moderated this rela-
Y = iY + b1X + b2W + b3XW + eY (30.8)
tionship, where the effect was strong and positive for
females, though largely nonexistent for males. where XW is the product of X and W. Thus, all of the regres-
The utility of moderation analysis in clinical research is sion coefficients can be estimated by including X, W, and
obvious if you are interested in the tailoring of treatments XW as predictors of Y in a linear regression model. The
and therapies to individuals. Researchers can detect and p-value for the weight for XW or a confidence interval, pro-
better understand how the effectiveness of a given treat- vided by standard regression analysis software, is used for
ment varies across individuals, where a particularly effect- inference as to whether X’s effect on Y is linearly moderated
ive treatment for one group may be ineffective, or even by W, or that X and W interact in their influence on Y.
harmful, for another. An additional benefit of moderation It is a widely believed myth that testing for interaction
analysis in a clinical setting is the ability to identify treat- between X and W requires building a regression model
ment inhibitors and enhancers. As an example of such using hierarchical entry of predictors, where X and W
phenomena, Singewald and colleagues (2015) reviewed (and any covariates, if desired) are entered into the first
previous research demonstrating that some pharmaco- model, followed by the XW product. An F-rest for the
logical treatments can enhance the effectiveness of psy- change in R2 when XW is added is used as a test of inter-
chotherapeutic treatments of anxiety, fear, and trauma- action. Although there is nothing wrong with this proced-
related disorders, while other medications, such as anxio- ure ‒ the validity of this procedure is not a myth ‒ it is not
lytics, can inhibit these psychotherapeutic treatments’ necessary to use hierarchical entry in this fashion. The
effectiveness. Moderation can be used in such scenarios p-value for the test of the null hypothesis that the weight
to quantify and test the differential treatment effects when for XW in Equation 30.7 is zero is the same as the p-value
combined with various medications. for the F-test comparing the fit of the model with and
This section will provide a brief overview of moderation without the product, so the same conclusion about inter-
analysis. Topics include conceptualizing the model, action will be reached either way.
making inferences about moderation, and interpreting
the regression coefficients.

Interpretation of Regression Coefficients

The Linear Moderation Model
The regression coefficient b3 quantifies the difference in
The effect of X on Y is moderated by W if the effect of X on the effect of X on Y for each one-unit difference in W.
Y varies as a function of W. Moderation is represented in When the moderator W is dichotomous (e.g., in a clinical
conceptual form in Figure 30.1, panel D. Mathematically, trial, a variable coding whether a participant is assigned to
such a model can be represented as therapy as usual or a new experimental therapy) and the
two values are coded one unit apart (e.g., 0 and 1), then b3
Y = iY + f(W)X + b2W + eY (30.6)
corresponds to the difference between the effect of X on Y
where f(W) is some function of W. The linear function in the two groups.
b1 + b3W is most popular, which, when substituted into The proper interpretation of b1 and b2 is often misun-
Equation 30.6, yields derstood. One of the common misinterpretations of b1 and
b2 is that they represent “main effects” of X and W, respect-
Y = iY + (b1 + b3W)X + b2W + eY (30.7)
ively. This misinterpretation is a result of overgeneralizing
In this model, X’s effect on Y is of most interest and so X is concepts from the factorial ANOVA literature, textbooks,
sometimes called the focal predictor. Its effect is a linear and statistics classrooms to all linear models. ANOVA is
function of W and so depends on W, making W the moder- just a special form of a linear model, but not all ANOVA
ator of the effect of X on Y. The conditional effect of X on Y, concepts generalize to any linear model. In a factorial
which we can denote as θX!Y, is the linear function. In ANOVA, the main effect of X is defined as the effect of X
Equation 30.7, θX!Y = b1 + b3W. Observe that the depend- averaged across the groups that define W, and the main
ency between the effect of X on Y and W is carried in b3. If effect of W is the effect of W averaged across the groups
b3 is zero, the relationship between X and Y is constrained that define X. However, as discussed below, b1 and b2 are
to b1 for all values of W and therefore is linearly independ- actually closer to what are known in ANOVA-speak as
ent of W. Therefore, an inference about whether b3 is simple effects. Coefficients b1 and b2 are main effects only
different from zero is used as a test of linear moderation when X and W are categorical variables and the categories
of the effect of X on Y by W. If it is, it is sometimes said are coded a specific way. See Darlington and Hayes (2017)
that X and W interact. So, interaction and moderation are and Hayes (2018) for a discussion of the difference
synonyms, at least in our discussion. Not everyone uses between a simple effects and a main effects parameteriza-
these terms interchangeably. tion of factorial ANOVA in a regression context.

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 402 ROCKWOOD & HAYES

To better see the interpretation of b1 and b2, suppose XW plugging them into the regression equation, which gener-
were excluded from Equation 30.8. In that case, b1 is the ates estimates of Y for those combinations. When W or X is
effect of X on Y holding W fixed at any value, and b2 is the dichotomous, use the values in the data coding the two
effect of W on Y holding X fixed at any value. These are groups. For continuous X or W, the most commonly used
unconditional effects or partial effects. But when XW is values are the sample mean and one standard deviation
included, as in Equation 30.8, this changes the interpret- above and below the mean, although these are mere con-
ation of b1 and b2. In Equation 30.8, the effect of X on Y ventions and entirely arbitrary. To ensure the values
depends on the value of W, and the effect of W on Y chosen are not outside the range of data (as interpretation
depends on the value of X and so become conditional of a model in a region of space where one has no data is ill-
effects. More specifically, b1 is the estimated effect of X advised), the 25th, 50th, and 75th percentile may be used
on Y when W = 0. Remember that X’s effect is a function f instead. Hayes (2018) recommends the 16th, 50th, and
(W) = b1 + b3W, which equals b1 when W = 0. Similarly, b2 84th percentiles. But other strategies for making the
is the estimated effect of W on Y conditional on X = 0. So choice could be used. Regardless, once the plot of the
b1 and/or b2 are meaningful only if X = 0 and/or W = 0 are model is generated, it becomes easier to see the pattern
meaningful. If X = 0 is not meaningful (e.g., if it is impos- in how variation in X is related to variation in Y differently
sible value on a measurement scale), then b2 is not mean- for different values of W. These are the conditional effects
ingful, and if W = 0 is not meaningful, then b1 is not of Y. Recognizing a pattern is the first step in substantively
meaningful. interpreting it. For more detail on visualizing an inter-
One way to ensure that b1 and b2 are meaningful is to action in a regression model, see Hayes (2018).
mean-center X and W before the creation of XW. A variable In addition to plotting a model to visualize the condi-
is mean-centered by subtracting the sample mean from tional nature of the relationship between X and Y, it is
each value in the data. When this is done, b1 quantifies useful to use inferential methods to test for the presence of
the effect of X on Y when W ¼ W
and b2 quantifies the a conditional effect of X on Y at values of W chosen.
effect of W on Y when X ¼ X.
However, this does not turn A popular method is the pick-a-point approach to probing
b1 and b2 into “main effects” as the term is used in analysis an interaction (Bauer & Curran, 2005; Hayes, 2018), also
of variance, except in certain circumstances. Several called an analysis of simple slopes. The conditional effect
books and journal articles state that mean-centering of X of X on Y is easy to generate for any value of W you choose.
and W in a moderation analysis must be done to reduce It is b1 + b3W. However, the standard error, needed for
multicollinearity, as XW is usually highly correlated with X inference for a given value of W, can be tricky to calculate.
and W. Although mean-centering does reduce multicolli- Several textbooks provide formulas to calculate the stand-
nearity, doing so has no effect on the test of moderation, ard error of a linear function of regression coefficients, but
as b3, its standard error, t- and p-values, and confidence the regression-centering approach makes inference about a
interval will be the same regardless. Mean-centering X and conditional effect in a regression model easy. This method
W is a personal choice one makes. It is not a requirement. is best understood by recalling that the coefficient b1 rep-
See Hayes (2018) for further discussion of this topic as resents the conditional effect of X on Y when W = 0. By
well as other myths of moderation analysis. centering W around the value of interest w (i.e., subtract-
Understanding the correct interpretation of b1 and b2 ing w from every value of W in the data, even if W is just
also sheds light on why the lower order terms (i.e., X and two arbitrary codes for two groups) prior to constructing
W) should be included as predictors in the model in add- XW and estimating the model, then b1 estimates the con-
ition to the interaction term XW. Removing X (or W) from ditional effect of X on Y when W = w, and the standard
the model would be equivalent to constraining b1 (or b2) to error of b1 is a valid estimate of the standard error of this
zero. It is rare that we would have such prior information conditional effect. Thus, the pick-a-point approach can be
on these conditional effects and so the constraints, if conducted by repeating the analysis multiple times with W
inaccurate, can have negative consequences on our esti- centered around each of the values of interest.
mate of the interaction effect and its corresponding infer- It is important to keep in mind that the conclusions
ential test. drawn from the inferential tests may change depending
on the values of the moderator at which the moderation
effect is probed. This is one of the drawbacks of the pick-a-
point approach. For this reason, some recommend using
Plotting and Probing a Moderation Effect
the Johnson-Neyman technique, which does not require
A substantively interesting moderation effect can corres- you to choose values of W but, rather, analytically derives
pond to several different trends in the data. To better values of W that help you find where in the distribution of
understand a moderation effect and what it is telling you W that X is significantly related to Y and where it is not (for
about a phenomenon of interest, you should generate a details, see Bauer & Curran, 2005; Hayes, 2018; Hayes &
visual depiction of the model that allows you to see how Matthes, 2009). Regardless, don’t be lured into thinking
X’s effect on Y varies with the moderator. This involves that if X’s effect on Y is significantly different from zero at
choosing combinations of values of X and W and then some value of W but not significantly different from zero

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 403

at another value, this implies moderation. Differences in

A B
the significance of two conditional effects does not mean M M
they differ from each other. It is the test on b3 in Equation
30.8 that provides formal support for a moderation
X Y X Y
hypothesis, which is about how X’s effect on Y varies with
W. And recall that the test of b3 does not depend on the
W W
choice of centering, so the method used to probe the
interaction should only influence our understanding of
the nature of the interaction, not the existence of it. C M1 W
D M1 W

X Y X Y
CONDITIONAL PROCESS ANALYSIS
To this point we have discussed how statistical mediation W M2 M2
and moderation analysis can be used to help understand
the process by which an effect operates and the contingen-
Figure 30.2 Four examples of conditional process models with a
cies of an effect. Although these methods are useful as single moderator W.
distinct analytic tools with different purposes, they can
be combined to explore and test the conditional nature
contains an indirect effect of X on Y that is specified as
of the process by which one variable influences another.
moderated by W. The examples in panels A and B contain
Several clinical researchers have stressed the importance
only a single mediator, but they differ with respect to
of combining mediation and moderation analyses to
whether W moderates the effect of X on M, the effect of
better understand psychological phenomena and the
M on Y, or the direct effect of X on Y. The models in panels
effectiveness of treatments in various populations. For
C and D of Figure 30.2 are multiple mediator models (in
example, in an overview of methods for improving psycho-
parallel form in panel C, and in serial form in panel D),
therapy and psychological intervention research, Emmelk-
with W moderating one or more of the paths in the indir-
amp and colleagues (2014) ask why and for whom
ect effect of X on Y through at least one mediator. These
cognitive behavior therapy methods are more effective
are only a few of the numerous possibilities.
than other therapeutic methods. They note that “relatively
A growing number of researchers in clinical psychology
little is known about why a certain type of psychotherapy
have used conditional process analysis to explore the
is effective for one child but fails to produce a positive
boundary conditions of a causal mechanism. For instance,
effect in another child” (Emmelkamp et al., 2014, p. 76).
Mason and colleagues (2015, discussed earlier) were inter-
Collins, Murphy, and Bierman (2004) describe how vari-
ested not only in the effectiveness of text-based interven-
ables identified as mediators and moderators in treatment
tions to reduce smoking brought about by increasing
research can be used to develop adaptive interventions,
readiness to stop (the mediator), but also the effect of
where repeated measurements of these variables over time
friends’ smoking behaviors on the strength of this mechan-
are used to adaptively tailor the treatment to the individ-
ism. They found that the size of the indirect effect from
ual. When used in combination, mediation and moder-
treatment (X) to reduction of cigarettes (Y) through readi-
ation analysis can be used to address such questions.
ness to stop smoking (M) depended on the number of
Conditional process analysis is the integration of medi-
friends that smoke (W). That is, this mechanism was in
ation and moderation analyses in a unified statistical
operation more among those who had fewer friends that
model. Thus, mediation and moderation models are the
smoke. Gaume and colleagues (2016) tested a conditional
building blocks of a conditional process model. When con-
process model where motivational interviewing (MI)
ducting a conditional process analysis, interest lies in not
experience (W) moderated the indirect effect of MI-
only identifying the causal mechanisms linking X to Y, but
consistent behaviors (X) on drinking at three-month
also understanding the boundary conditions or condi-
follow-up (Y) through client change talk (M). The indirect
tional nature of these mechanisms. Conditional process
effect was stronger for therapists that had more MI experi-
models subsume models that test for moderated mediation,
ence. And using a Latino sample, Torres and Taknint (2015)
where the strength or direction of an indirect effect
found that ethnic identity and self-efficacy (W) moderated
depends on the value of a moderator. They can take many
the effect of ethnic microaggressions (X) on traumatic stress
forms and are built by combining features from the medi-
symptoms (M), which, in turn, affected depression (Y).
ation and moderation models discussed in the previous
sections. A conditional process model can include one or
more mediators while allowing the strength or direction of
A Simple Conditional Process Model
one or more of the paths linking X to Y (i.e., X’s effect on
M, M’s effect on Y, or both) to depend on a moderator W. Consider a conditional process model that allows the
Some examples can be found in Figure 30.2, each of which indirect effect of X on Y through M to be moderated by

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 404 ROCKWOOD & HAYES

W while the direct effect of X is fixed to be constant. If the mediation is different from zero. A confidence interval for
moderation of the indirect effect operates through the the index of moderated mediation can be constructed
moderation of the X!M path only, the model appears as using the bootstrapping method described earlier. If the
in Figure 30.2, panel A. Using the principles of moderation confidence interval does not include zero, this is evidence
and mediation analysis described earlier, and under the of moderated mediation, but if the confidence interval
assumption that all effects are linear and that W is linearly includes zero, you cannot definitively claim that the indir-
related to the size of the effect of X on M, this model can be ect effect depends on the moderator. Note that with this
specified with a set of two regression equations: approach, a claim of moderated mediation is not reliant
on a series of significance tests of the individual compon-
M = iM + a1X + a2W + a3XW + eM (30.9)
ents, a3 and b. Inference about the index of moderated
0
Y = iY + c X + bM + eY (30.10) mediation is all that is needed when determining if the
indirect effect is moderated by W (see Hayes, 2015, 2018,
Note the only difference between this conditional process
for details).
model and the simple mediation model using Equations
30.2 and 30.3 is the equation for M, which now includes W
and XW as predictors. Including XW as a predictor allows
Probing the Moderation of Mediation
the effect of X on M to depend linearly on W.
By rearranging Equation 30.9 as M = iM + (a1 + a3W) X + After establishing that the indirect effect of X is moderated
a2W + eM, it is clear that X’s effect on M is a linear function by W, this moderation effect can be visualized and probed,
of W: θX!M = a1 + a3W. But the effect of M on Y (control- just as in the simple moderation model. Now, however, the
ling for X) is fixed to be independent of W. It is simply b, interest is on the conditional indirect effect of X on Y
just as in a simple mediation model. And the direct effect through M at different values of W. Equation 30.11 is a
of X on Y remains c0 . It is fixed to be independent of W in linear function of W. A line graph depicting the relation-
this model. ship between W and the size of X’s indirect effect can be
Recall that the indirect effect in a simple mediation generated by plotting the output of Equation 30.11 as a
model is the product of the effect of X on M and the effect function of many values of W. The resulting plot will
of M on Y when statistically controlling for X. In a condi- take the form of a straight line, with the steepness of the
tional process model, an indirect effect is still defined as a slope determined by a3b. When the output of the
product of its constituent effects. So in this model, the function results in a value less than zero for a given value
indirect effect of X on Y through M is of W, then X is negatively related to Y through M when
W is that value, whereas when output of the function is
Indirect effect of X = (a1 + a3W) b = a1b + a3bW (30.11)
greater than zero, X is positively related to Y through M at
(see, e.g., Edwards & Lambert, 2007; Hayes, 2018; that value of W. The moderation of the indirect
Preacher et al., 2007). Notice that Equation 30.11 is a effect can also be probed using an analogue of the pick-
linear function of W. As the value of the indirect effect of a-point approach discussed earlier, choosing various
X will depend on the value of W (unless a3 = 0 or b = 0), the values of W and plugging them into Equation 30.11
output of this function is called a conditional indirect after the regression coefficients are estimated. However,
effect. If W is dichotomous, then there are two conditional this produces only point estimates of the conditional
indirect effects of X, one for each value of W. But if W is a indirect effect at those values of W. Additional work is
numerical continuum, then there are many, or even an required for inference. Because a conditional indirect
infinite number of conditional indirect effects, one for effect is a product of regression coefficients and so its
each possible value of W. sampling distribution will be nonnormal, a bootstrap
confidence interval for the conditional indirect effect
can be used for inference. Special software is needed,
Index of Moderated Mediation such as a good structural equation modeling program
or a special tool such as the PROCESS macro used in the
In Equation 30.11, a1b is the indirect effect of X on Y when
following section (see Hayes, 2018, for a discussion).
W = 0. If W = 0 is not meaningful, neither is a1b, and it
estimates nothing interpretable. The term a3b in Equation
30.11 is analogous to the regression weight for the product
A Worked Example
of X and W in a moderation model, in that it quantifies the
relationship between the effect of X and the moderator. Now that the fundamentals of the framework for integrat-
Except here it is the product of two regression coefficients, ing mediation and moderation analyses have been dis-
and it quantifies the change in the indirect effect of X as W cussed, we walk through an example loosely based on a
changes by one unit. Hayes (2015) calls a3b the index of study conducted by Barnett and colleagues (2010), who
moderated mediation for this model. An inference about implemented a brief intervention targeting alcohol use in
moderation of the indirect effect ‒ moderated mediation ‒ emergency department patients. The participants con-
is undertaken by testing whether the index of moderated sisted of patients admitted to a hospital emergency

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 405

department who had been under the influence of

A Risks
alcohol at the time of admission. Based on various W
diagnostic assessments, participants were classified as M1
Severity
severe (W = 1) or not severe (W = 0) abusers of alcohol.
The participants were then randomly assigned to a
treatment that consisted of a motivational interview with Y
X
a counselor followed by personalized feedback (X = 1) or
Motivational Use
were provided only personalized feedback without
interview
any interaction with a counselor (X = 0). This is the inde-
pendent variable in the analysis below. Two potential M2
mediators of the effect of the intervention, posttreatment Treatment
perceived risk and benefits of alcohol use (M1) and seeking
degree of treatment seeking (M2), were measured.
Higher scores on these represent greater perceived risks B Risks
a12
W M1 eM1
and more treatment seeking. The outcome variable (Y) is a
composite measure of alcohol use (frequency and Severity a11 b1 eY
amount) at 12-month follow-up, with higher scores reflect- a13
ing more use. In this example, the moderator W is a c'
X Y
dichotomous variable. However, the procedure we discuss a22
Motivational Use
here applies whether W is dichotomous or a numerical interview b2
a21
continuum.
Note that the analysis we describe below is not based on XW M2 eM2
a23
the actual data from this study but instead uses simulated Treatment
data constructed for the purpose of this example. Further- seeking
more, although Barnett and colleagues (2010) were inter-
Figure 30.3 A first stage moderated parallel multiple mediator
ested in both moderators and mediators of the effect of
model in conceptual form (panel A) and in the form of a path
motivational interviewing, they did not estimate a condi- diagram (panel B)
tional process model as we do here. We offer no guarantee
that the results of the analysis we present here match what
The Model in Visual and Equation Form
would be found using real data (theirs or anyone else’s),
A visual representation of such a conditional process
and readers of this chapter should not use our analysis to
model can be found in Figure 30.3, panel A. It depicts a
argue for the existence of any empirical support in Barnett
first-stage moderated parallel multiple mediator model, as it
and colleagues (2010) or elsewhere for the findings we
contains two mediators, and only the effect of X on the
report.
mediators (the first stage of the mediation process) is
proposed as moderated. The diagram can be represented
mathematically in the form of a set of three regression
Research Questions
equations, one for each of the mediators and one for Y,
Suppose it is hypothesized that perceived risks of
using the principles described in the prior sections. These
alcohol use and degree of treatment seeking, M1 and M2,
equations are
are mediators of the relationship between the motiv-
ational interviewing treatment (X) and alcohol use (Y) at M1 = iM2 + a11X + a12W + a13XW + eM1 (30.12)
12-month follow-up. Furthermore, perhaps there is
M2 = iM2 + a21X + a22W + a23XW + eM2 (30.13)
reason to believe that the indirect effects through per-
0
ceived risks and treatment seeking will differ in size Y = iY + c X + b1M1 + b2M2 + eY (30.14)
depending on severity of the patient’s abuse of alcohol
and represented in the form of a path diagram in
(W). Specifically, perhaps the indirect effect of motiv-
Figure 30.3, panel B. When specified in this form, the
ational interviewing on alcohol use through perceived
assumption is that all effects are linear. Furthermore, the
risks is proposed to be larger than the indirect effect
XW products in Equations 30.12 and 30.13 specify X’s
through treatment seeking for patients with less
effect on M1 and M2 as linear functions of W.
severe alcohol abuse, while the reverse is expected for
patients with severe alcohol use. This expectation is the
result of the prediction that the difference in effects Model Estimation
between these two types of abusers would stem from dif- The regression coefficients in Equations 30.12‒30.14 can
ferences between them in the effect of the treatment on be estimated using any statistical software that can con-
the mediators (i.e., X!M1 and X!M2). That is, W is pro- duct regression analysis. However, many of the details we
posed as moderating the relationships between X and M1 discussed throughout this paper, such as probing inter-
and X and M2. actions and constructing bootstrap confidence intervals

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 406 ROCKWOOD & HAYES

Table 30.1 Regression coefficients, standard errors, p-values, and confidence intervals from the motivational interviewing
conditional process analysis

95% confidence
Outcome Predictor Parameter Estimate SE p interval

Risks (M1) Intercept iM1 3.228 0.114 <0.001 (3.004, 3.453)

Interview (X) a11 0.453 0.165 0.006 (0.129, 0.777)
Severity (W) a12 0.189 0.162 0.244 (‒0.130, 0.508)
XW a13 ‒0.516 0.236 0.030 (‒0.981, ‒0.051)

Treatment Intercept iM2 3.534 0.115 <0.001 (3.308, 3.761)

seeking
Interview (X) a21 ‒0.127 0.166 0.444 (‒0.453, 0.199)
(M2)
Severity (W) a22 ‒0.308 0.163 0.060 (‒0.629, 0.012)
XW a23 0.867 0.238 <0.001 (0.398, 1.335)

Use (Y) Intercept iY 6.227 0.277 0.006 (0.227, 1.318)

0
Interview (X) c ‒0.416 0.117 <0.001 (‒0.646, ‒0.187)
Risks (M1) b1 ‒0.513 0.056 <0.001 (‒0.624, ‒0.403)
Treat seeking (M2) b2 ‒0.276 0.055 <0.001 (‒0.384, ‒0.168)
Table entries come from the PROCESS output in the Appendix.

for products or functions of regression coefficients, can be about indirect effects, the index of moderated mediation,
quite tedious to implement without the assistance of and conditional indirect effects. SEM programs offer
special tools designed for these purposes. One such tool is more flexibility than some of the other options for condi-
the PROCESS macro for SPSS and SAS (Hayes, 2018). tional process modeling, but they also often require sub-
PROCESS was designed specifically to simplify mediation, stantially more programming effort. And as Hayes,
moderation, and conditional process analysis, as the estima- Montoya, and Rockwood (2017) discuss, SEM and a
tion of indirect effects, the index of moderated mediation, regression-based tool like PROCESS generally produce
and the probing of moderated mediation and estimation of equivalent results for the kinds of observed variable
conditional (indirect) effects is automated with minimal models (i.e., nothing latent) discussed in this chapter, so
syntax or an easy-to-use point-and-click interface (for SPSS the extra programming skills and effort required to use
users). Output for models containing moderators also SEM software may not be the most efficient use of time
includes code that can be used to plot the conditional effects and resources, although we encourage all clinical
and conduct various additional analyses. We used the PRO- researchers to work toward developing such programming
CESS macro for SPSS to conduct the analysis in this skills at some point in their careers.
example. PROCESS is freely available and can be down-
loaded from www.processmacro.org. Guidance on its use Model Components
can be found in Hayes (2018). When working with a complex conditional process model,
Mediation, moderation, and conditional process analy- it is often useful to begin by looking at smaller compon-
sis can also be conducted using structural equation mod- ents of the bigger model. Understanding the smaller com-
eling (SEM) software. An advantage of using SEM is that ponents is accomplished by looking at and interpreting the
doing so allows for the estimation of latent variable regression coefficients from the individual equations
models, which can help reduce the effects of measurement before bringing those results together as needed to calcu-
error on the estimation of effects when a structural model late statistics and test hypotheses that require integration
is combined with a good measurement model of the latent across the equations. The estimated regression coefficients
variables. Although any SEM program can be used, some in Equations 30.12‒30.14, along with standard errors, p-
programs provide features that make conditional process values, and confidence intervals can be found in
modeling easier than do others. For example, Mplus Table 30.1. These come from the PROCESS output in
(Muthén & Muthén, 2015) and lavaan (Rosseel, 2012) Appendix 30.A, sections 1, 2, and 3. In the PROCESS
can construct bootstrap confidence intervals for functions output, Y, X, M1, M2, and W are named “use,” “intview,”
of model coefficients, allowing the user to make inferences “riskbene,” “trtseek,” and “severity,” respectively.

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 407

Table 30.2 Conditional indirect effects of motivational interviewing on alcohol use and the index of moderated
mediation, with bootstrap confidence intervals, from the conditional process analysis

Mediator Group (W) Parameter Estimate 95% bootstrap confidence interval

Risks (M1) Not severe (0) a11b1 ‒0.233 (‒0.408, ‒0.072)

Severe (1) (a11 + a13)b1 ‒0.032 (‒0.149, 0.222)

Index of MM a13b1 0.265 (0.020, 0.533)
Treatment seeking (M2) Not severe (0) a21b2 0.035 (‒0.055, 0.132)
Severe (1) (a21 + a23)b2 ‒0.204 (‒0.340, ‒0.095)
Index of MM a23b2 ‒0.239 (‒0.413, ‒0.098)
Estimates and confidence intervals come from the PROCESS output in the Appendix.

The effect of the motivational interview, X, on each of meaning that, controlling for perceived risks and treat-
the mediators depends on abuse severity, W, and each of ment seeking, those who received the motivational
the M equations take the form of a simple moderation interview used alcohol less at 12-month follow-up. The
model as discussed in the moderation section above. For effects of each of the mediators on later use are both
the risk/benefits measure, M1, the effect of X is θX!M1, = statistically significant (b1 = ‒0.513, p < 0.001; b2 = ‒
a11 + a13W = 0.453 – 0.516W. This conditional effect of 0.276, p < 0.001). Those who perceived alcohol as riskier
X for severe abusers and less severe abusers can be found used alcohol less later, as did those who engaged in more
by plugging in W = 0 and W = 1 into this function. For less treatment seeking.
severe abusers, the effect is a11 + a13(0) = 0.453, whereas
for severe abusers, the effect is a11 + a13(1) = 0.063. Conditional Indirect Effects
Because the treatment indicator X is dichotomous with Having interpreted the components of the larger model,
codes in the data that differ by one unit (0 and 1), these we can now integrate this information across the compon-
effects can be interpreted as mean differences. Specific- ents to calculate the indirect effects of the motivational
ally, less severe abusers who received the motivational interview on later alcohol use through perceived risks (M1)
interviewing treatment scored 0.453 units higher on the and treatment seeking (M2). The indirect effect through
risk and benefits measure, on average, than less severe perceived risks is the product of the effect of X on M1,
abusers who received only personalized feedback. And θX!M1 = a11 + a13W, and the effect of M1 on Y, b1. Multi-
severe abusers who received the motivational interview plying these effects results in (θX!M1) b1 = (a11 + a13W) b1 =
treatment scored slightly lower (‒0.063), on average, than a11b1 + a13b1W = ‒0.233 + 0.265W. So, the indirect effect
severe abusers who received feedback only. The difference through perceived risks depends on alcohol abuse severity,
in the effect of X on M1 between the two types of patients is W. As we did with the conditional effect of X on M1, we can
a13 = ‒0.516, the coefficient for XW. This difference is calculate the conditional indirect effect of X on Y through
statistically significant (p = 0.030), indicating that the M1 for the two abuse severity groups by plugging in values
treatment effect on the risk/benefit measure changes of W into this equation. The indirect effect through per-
depending on the patient type. ceived risk is a11b1 + a13b1W = ‒0.233 + 0.265(0) = ‒0.233
The equation for the effect of the motivational interview for less severe abusers (W = 0). This is the product of the
on treatment seeking, the second mediator M2, is con- conditional effect of X on M1 when W = 0 (0.453) and the
structed similarly. The conditional effect of X on M2 is effect of M1 on Y (‒0.513). But for severe abusers of alco-
θX!M2 = a21 + a23W = ‒0.127 + 0.867W, and the difference hol, the indirect effect through perceived risk is a11b1 +
between the treatment effect for the two severity groups is a13b1W = ‒0.233 + 0.265(1) = 0.032. This is the product of
a23 = 0.867, p < 0.001. Here, the effect of the motivational the conditional effect of X on M1 when W = 1 (‒0.063) and
interview on treatment seeking is larger for severe the effect of M1 on Y (‒0.513).
abusers, ‒0.127 + 0.867(1) = 0.740, than for less severe Table 30.2 includes each of these conditional indirect
abusers, ‒0.127 + 0.867(0) = 0.127. effects as well as 95 percent bootstrap confidence intervals
The equation for drinking at 12-month follow-up, Y, is based on 5,000 bootstrap samples. These can be found in
6.227 – 0.416X – 0.513M1 – 0.276M2. As set up in this section 4 of the PROCESS output. The 95 percent boot-
model, none of the effects of the independent variable or strap confidence interval for less severe abusers is ‒0.408
mediators are expressed as functions of a moderator, so to ‒0.072 and does not contain zero, but the confidence
these effects can be interpreted as are any partial interval of ‒0.149 to 0.222 for severe abusers does. Yet
regression coefficients. The direct effect of the motiv- difference in significance does not necessarily mean that
ational interview on later use is c0 = ‒0.416, p < 0.001, these conditional indirect effects differ from each other

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 408 ROCKWOOD & HAYES

(see Hayes, 2015, 2018). The difference between these two Summary
conditional indirect effects is the index of moderated
Throughout this chapter, we’ve provided an overview of
mediation, which here is a13b1 = 0.265. A 95 percent boot-
mediation, moderation, and conditional process analysis,
strap confidence interval for this index is 0.020–0.533,
with an emphasis on mapping particular research ques-
which is entirely above zero. So, we can say that these
tions of interest to their respective statistical models.
two indirect effects are significantly different from each
We’ve described how statistical mediation analysis can
other. For less severe abusers, motivational interviewing
be used to help identify, quantify, and understand a causal
decreases alcohol use by increasing the perceived risks of
sequence of events in which one variable influences
use, which reduces use. Such a process does not operate
another through one or more intermediary variables,
among severe abusers.
while statistical moderation analysis can be used to test
The same procedure can be used to quantify the condi-
the boundary conditions or contingencies of an effect.
tional indirect effect of motivational interviewing on alco-
These methods have proved useful for clinical research,
hol use through treatment seeking, M2, which is (θX!M2)
both separately and when integrated within a unifying
b2 = (a21 + a23W)b2 = a21b2 + a23b2W = 0.035 – 0.239W. For
conditional process model. After detailing the framework
less severe abusers (W = 0), the conditional indirect effect
of mediation, moderation, and conditional process analy-
is 0.035 – 0.239(0) = 0.035, which is the product of the
sis, with a focus on the substantive meaning of model
conditional effect of X on M2 when W = 0 (‒0.127) and the
parameters and how they relate to hypotheses of interest,
effect of M2 on Y (‒0.276). For severe abusers (W = 1), the
we presented an example analysis that synthesized many
conditional indirect effect is 0.035 – 0.239(1) = ‒0.204,
of the concepts discussed throughout this chapter. We
which is the product of the conditional effect of X on M2
believe this exposition should provide clinical researchers
when W = 1 (0.740) and the effect of M2 on Y (‒0.276).
with a good understanding of the conceptual and statis-
The difference between these two conditional indirect
tical foundations of mediation and moderation analysis
effects is a23b2= ‒0.293, the index of moderated mediation
and also aid them in the development of their own condi-
for this indirect effect, with a 95 percent bootstrap confi-
tional process models for testing and exploring their own
dence interval from ‒0.413 to ‒0.098. So the difference
hypotheses and substantive interests.
between these conditional indirect effects is statistically
significant. Probing this moderation of the indirect effect
shows that the conditional indirect effect of motivational
interviewing on use through treatment seeking is not REFERENCES
definitively different from zero among less severe abusers
Aroian, L. A. (1947). The Probability Function of the Product of
(95 percent bootstrap confidence interval = ‒0.055 to Two Normally Distributed Variables. Annals of Mathematical
0.132), but among severe abusers, this indirect effect is Statistics, 265–271.
significantly negative (95 percent bootstrap confidence Barnett, N. P., Apodaca, T. R., Magill, M., Colby, S. M., Gwaltney,
interval = ‒0.340 to ‒0.095). So, among severe abusers, C., Rohsenow, D. J., & Monti, P. M. (2010). Moderators and
motivational interviewing results in an increase in treat- Mediators of Two Brief Interventions for Alcohol in the Emer-
ment seeking, which in turn is related to a reduction in gency Department. Addiction, 105(3), 452–465.
alcohol use. No such process is in operation among less Baron, R. M., & Kenny, D. A. (1986). The Moderator-Mediator
severe abusers. Variable Distinction in Social Psychological Research: Concep-
tual, Strategic, and Statistical Considerations. Journal of Per-
sonality and Social Psychology, 51(6), 1173–1182.
Bauer, D. J., & Curran, P. J. (2005). Probing Interactions in Fixed
Substantive Conclusion
and Multilevel Regression: Inferential and Graphical Tech-
To summarize, and using a data set fabricated for the niques. Multivariate Behavioral Research, 40, 373–400.
purpose of this example, we have used a combination of Bollen, K. A. (1989). Structural Equation Modeling with Latent
a parallel mediator model and the simple moderation Variables. New York: Wiley.
model to better understand the contingencies of the pro- Cerin, E., & MacKinnon, D. P. (2009). A Commentary on Current
cesses by which motivational interviewing affects alcohol Practice in Mediating Variable Analyses in Behavioural Nutri-
use. Perceived risk of alcohol use and treatment seeking tion and Physical Activity. Public Health Nutrition, 12(8),
were the mediators of interest in this example. The effect 1182–1188.
of the motivational interview on each of the mediators was Collins, L. M., Murphy, S. A., & Bierman, K. L. (2004).
dependent on severity of the patient’s alcohol abuse, and A Conceptual Framework for Adaptive Preventive Interven-
tions. Prevention Science, 5(3), 185–196.
each of the mediators was negatively correlated with later
Darlington, R. B., & Hayes, A. F. (2017). Regression Analysis and
alcohol use. Furthermore, the indirect effect through each
Linear Models: Concepts, Applications, and Implementation.
mediator was conditional on alcohol use severity. The New York: Guilford Press.
indirect effect through perceived risk was larger for less Duchesne, A.-P., Dion, J., Lalande, D., Bégin, C., Émond, C.,
severe abusers than for severe abusers, but the indirect Lalande, G., & McDuff, P. (2017). Body Dissatisfaction and
effect through treatment seeking was larger for severe Psychological Distress in Adolescents: Is Self-Esteem a Medi-
abusers than it was for less severe abusers. ator? Journal of Health Psychology, 22(12), 1563–1569.

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 409

Edwards, J. R., & Lambert, A. L. (2007). Methods for Integrating Differ between the Baron & Kenny and MacArthur Approaches.
Moderation and Mediation: A General Analytical Framework Health Psychology, 27, S101–S108.
Using Moderated Path Analysis. Psychological Methods, 12, Kraemer, H. C., Wilson, G. T., Fairburn, C. G., & Agras, W. S.
1–22. (2002). Mediators and Moderators of Treatment Effects in Ran-
Efron, B., & Tibshirani, R. J. (1994). An Introduction to the Boot- domized Clinical Trials. Archives of General Psychiatry, 59(10),
strap. Boca Raton, FL: Chapman and Hall. 877–883.
Emmelkamp, P. M., David, D., Beckers, T., Muris, P., Cuijpers, P., MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confi-
Lutz, W., . . . Berking, M. (2014). Advancing Psychotherapy and dence Limits for the Indirect Effect: Distribution of the Product
Evidence-Based Psychological Interventions. International and Resampling Methods. Multivariate Behavioral Research, 39
Journal of Methods in Psychiatric Research, 23(S1), 58–91. (1), 99–128.
Finney, J. W. (1995). Enhancing Substance Abuse Treatment Magill, M. (2011). Moderators and Mediators in Social Work
Evaluations: Examining Mediators and Moderators of Treat- Research: Toward a More Ecologically Valid Evidence Base
ment Effects. Journal of Substance Abuse, 7(1), 135–150. for Practice. Journal of Social Work, 11(4), 387–401.
Frazier, P. A., Tix, A. P., & Barron, K. E. (2004). Testing Moder- Mason, M., Mennis, J., Way, T., & Campbell, L. F. (2015). Real-
ator and Mediator Effects in Counseling Psychology Research. Time Readiness to Quit and Peer Smoking within a Text Mes-
Journal of Counseling Psychology, 51(1), 115–134. sage Intervention for Adolescent Smokers: Modeling Mechan-
Gaume, J., Longabaugh, R., Magill, M., Bertholet, N., Gmel, G., & isms of Change. Journal of Substance Abuse Treatment, 59,
Daeppen, J.-B. (2016). Under What Conditions? Therapist and 67–73.
Client Characteristics Moderate the Role of Change Talk in Maxwell, S. E., & Cole, D. A. (2007). Bias in Cross-Sectional
Brief Motivational Intervention. Journal of Consulting and Clin- Analyses of Longitudinal Mediation. Psychological Methods, 12
ical Psychology, 84(3), 211–220. (1), 23–44.
Hayes, A. F. (2009). Beyond Baron and Kenny: Statistical Medi- Maxwell, S. E., Cole, D. A., & Mitchell, M. A. (2011). Bias in Cross-
ation Analysis in the New Millennium. Communication Mono- Sectional Analyses of Longitudinal Mediation: Partial and Com-
graphs, 76, 408–420. plete Mediation under an Autoregressive Model. Multivariate
Hayes, A. F. (2015). An Index and Test of Linear Moderated Behavioral Research, 46(5), 816–841.
Mediation. Multivariate Behavioral Research, 50(1), 1–22. Muthén, L. K., & Muthén, B. (2015). Mplus Version 7.4 [Computer
Hayes, A. F. (2018). Introduction to Mediation, Moderation, and software manual]. Los Angeles: Muthén & Muthén.
Conditional Process Analysis: A Regression-Based Approach (2nd Norcross, J. C., & Wampold, B. E. (2011). What Works for Whom:
edn.). New York: Guilford Press. Tailoring Psychotherapy to the Person. Journal of Clinical
Hayes, A. F., & Matthes, J. (2009). Computational Procedures Psychology, 67(2), 127–132.
for Probing Interactions in OLS and Logistic Regression: SPSS Onken, L. S., Carroll, K. M., Shoham, V., Cuthbert, B. N., &
and SAS Implementations. Behavior Research Methods, 41, Riddle, M. (2014). Re-envisioning Clinical Science: Unifying
924–936. the Discipline to Improve the Public Health. Clinical Psycho-
Hayes, A. F., & Montoya, A. K. (2017). A Tutorial on Testing, logical Science, 2(1), 22–34.
Visualizing, and Probing an Interaction Involving a Multicate- O’Rourke, H. P., & MacKinnon, D. P. (2018). Reasons for Testing
gorical Variable in Linear Regression Analysis. Communication Mediation in the Absence of an Intervention Effect: A Research
Methods and Measures, 11(1), 1–30. Imperative in Prevention and Intervention Research. Journal of
Hayes, A. F., & Preacher, K. J. (2014). Statistical Mediation Analy- Studies on Alcohol and Drugs, 79, 171–181.
sis with a Multicategorical Independent Variable. British Jour- Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and Resampling
nal of Mathematical and Statistical Psychology, 67(3), 451–470. Strategies for Assessing and Comparing Indirect Effects in Mul-
Hayes, A. F., & Rockwood, N. J. (2017). Regression-Based Statis- tiple Mediator Models. Behavior Research Methods, 40(3),
tical Mediation and Moderation Analysis in Clinical Research: 879–891.
Observations, Recommendations, and Implementation. Behav- Preacher, K. J., & Selig, J. P. (2012). Advantages of Monte Carlo
iour Research and Therapy, 98, 39–57. Confidence Intervals for Indirect Effects. Communication
Hayes, A. F., & Scharkow, M. (2013). The Relative Trustworthi- Methods and Measures, 6(2), 77–98.
ness of Inferential Tests of the Indirect Effect in Statistical Preacher, K. J., Rucker, D. D., & Hayes, A. F. (2007). Addressing
Mediation Analysis Does Method Really Matter? Psychological Moderated Mediation Hypotheses: Theory, Methods, and Pre-
Science, 24(10), 1918–1927. scriptions. Multivariate Behavioral Research, 42(1), 185–227.
Hayes, A. F., Montoya, A. K., & Rockwood, N. J. (2017). The Rosseel, Y. (2012). lavaan: An R Package for Structural Equation
Analysis of Mechanisms and Their Contingencies: PROCESS Modeling. Journal of Statistical Software, 48(2), 1–36.
versus Structural Equation Modeling. Australasian Marketing Rucker, D. D., Preacher, K. J., Tormala, Z. L., & Petty, R. E.
Journal, 25, 76–81. (2011). Mediation Analysis in Social Psychology: Current Prac-
Imai, K., Keele, L., & Tingley, D. (2010). A General Approach to tices and New Recommendations. Social and Personality Psych-
Causal Mediation Analysis. Psychological Methods, 15(4), ology Compass, 5(6), 359–371.
309–334. Schmidt, S. J., & Schimmelmann, B. G. (2015). Mechanisms of
Kazdin, A. E. (2007). Mediators and Mechanisms of Change in Change in Psychotherapy for Children and Adolescents: Cur-
Psychotherapy Research. Annual Review of Clinical Psychology, rent State, Clinical Implications, and Methodological and Con-
3, 1–27. ceptual Recommendations for Mediation Analysis. European
Kenny, D. A., & Judd, C. M. (2014). Power Anomalies in Testing Child and Adolescent Psychiatry, 24, 249–253.
Mediation. Psychological Science, 25, 334–339. Shanmugam, V., & Davies, B. (2015). Clinical Perfectionism and
Kraemer, H. C., Kiernan, M., Essex, M., & Kupfer, D. J. (2008). Eating Psychopathology in Athletes: The Role of Gender. Per-
How and Why Criteria Defining Moderators and Mediators sonality and Individual Differences, 74, 99–105.

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 410 ROCKWOOD & HAYES

Shrout, P. E., & Bolger, N. (2002). Mediation in Experimental and A Moderated Mediational Model. Journal of Counseling Psych-
Nonexperimental Studies: New Procedures and Recommenda- ology, 62(3), 393–401.
tions. Psychological Methods, 7(4), 422–445. Valeri, L., & VanderWeele, T. J. (2013). Mediation Analysis
Singewald, N., Schmuckermair, C., Whittle, N., Holmes, A., & Allowing for Exposure-Mediator Interactions and Causal
Ressler, K. J. (2015). Pharmacology of Cognitive Enhancers Interpretation: Theoretical Assumptions and Implementation
for Exposure-Based Therapy of Fear, Anxiety and Trauma- with SAS and SPSS Macros. Psychological Methods, 18(2),
Related Disorders. Pharmacology & Therapeutics, 149, 150–190. 137–150.
Sobel, M. E. (1982). Asymptotic Confidence Intervals for Indirect Windgassen, S., Goldsmith, K., Moss-Morris, R., & Chalder, T.
Effects in Structural Equation Models. Sociological Method- (2016). Establishing How Psychological Therapies Work: The
ology, 13(1982), 290–312. Importance of Mediation Analysis. Journal of Mental Health, 25
Stanley, I. H., Joiner, T. E., & Bryan, C. J. (2017). Mild Traumatic (2), 93–99.
Brain Injury and Suicide Risk among a Clinical Sample of Deployed Zhang, H., Neelarambam, K., Schwenke, T. J., Rhodes, M. N.,
Military Personnel: Evidence for a Serial Mediation Model of Anger Pittman, D. M., & Kaslow, N. J. (2013). Mediators of a
and Depression. Journal of Psychiatric Research, 84, 161–168. Culturally-Sensitive Intervention for Suicidal African-American
Torres, L., & Taknint, J. T. (2015). Ethnic Microaggressions, Women. Journal of Clinical Psychology in Medical Settings, 20
Traumatic Stress Symptoms, and Latino Depression: (4), 401–414.

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 Appendix 30.A
Output from the PROCESS Macro for SPSS

****************PROCESS Procedure for SPSS Release 3.1**********

Written by Andrew F. Hayes,Ph.D. www.afhayes.com

Documentation available in Hayes(2018). www.guilford.com/p/hayes3

****************************************************************

Model:7
Y :use
X :intview
M1:riskbene
M2:trtseek
W :severity
Sample
Size:300
****************************************************************

1 OUTCOME VARIABLE:
riskbene

Model Summary
R R-sq MSE F df1 df2 p
.1617 .0261 1.0422 2.6478 3.0000 296.0000 .0492

Model
coeff se t p LLCI ULCI
constant 3.2281 .1141 28.2826 .0000 3.0035 3.4528
intview .4530 .1647 2.7509 .0063 .1289 .7770
severity .1892 .1619 1.1685 .2435 -.1295 .5079
Int_1 -.5161 .2363 -2.1837 .0298 -.9812 -.0510

Product terms key:

Int_1 : intview x severity

411

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 412 ROCKWOOD & HAYES

Test(s) of highest order unconditional interaction(s):

R2-chng F df1 df2 p
X*W .0157 4.7687 1.0000 296.0000 .0298
----------
Focal predict: intview (X)
Mod var: severity(W)

Conditional effects of the focal predictor at values of the moderator(s):

severity Effect se t p LLCI ULCI

.0000 .4530 .1647 2.7509 .0063 .1289 .7770
1.0000 -.0632 .1696 -.3725 .7098 -.3968 .2705

Data for visualizing the conditional effect of the focal predictor:

Paste text below into a SPSS syntax window and execute to produce plot.

DATA LIST FREE/

intview severity riskbene .
BEGIN DATA.
.0000 .0000 3.2281
1.0000 .0000 3.6811
.0000 1.0000 3.4173
1.0000 1.0000 3.3542
END DATA.
GRAPH/SCATTERPLOT=
intview WITH riskbene BY severity.

****************************************************************

2 OUTCOME VARIABLE:
trtseek

Model Summary
R R-sq MSE F df1 df2 p
.2513 .0631 1.0564 6.6484 3.0000 296.0000 .0002

Model
coeff se t p LLCI ULCI
constant 3.5344 .1149 30.7575 .0000 3.3082 3.7605
intview -.1272 .1658 -.7674 .4435 -.4534 .1990
severity -.3084 .1630 -1.8919 .0595 -.6293 .0124
Int_1 .8665 .2379 3.6415 .0003 .3982 1.3348

Product terms key:

Int_1 : intview x severity

Test(s) of highest order unconditional interaction(s):

R2-chng F df1 df2 p
X*W .0420 13.2608 1.0000 296.0000 .0003
----------

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 MEDIATION, MODERATION, AND CONDITIONAL PROCESS ANALYSIS 413

Focal predict: intview (X)

Mod var: severity(W)

Conditional effects of the focal predictor at values of the moderator(s):

severity Effect se t p LLCI ULCI

.0000 -.1272 .1658 -.7674 .4435 -.4534 .1990
1.0000 .7393 .1707 4.3309 .0000 .4033 1.0752

Data for visualizing the conditional effect of the focal predictor:

Paste text below into a SPSS syntax window and execute to produce plot.

DATA LIST FREE/

intview severity trtseek .
BEGIN DATA.
.0000 .0000 3.5344
1.0000 .0000 3.4072
.0000 1.0000 3.2259
1.0000 1.0000 3.9652
END DATA.
GRAPH/SCATTERPLOT=
intview WITH trtseek BY severity.

****************************************************************

3 OUTCOME VARIABLE:
use

Model Summary
R R-sq MSE F df1 df2 p
.5591 .3126 .9842 44.8607 3.0000 296.0000 .0000

Model
coeff se t p LLCI ULCI
constant 6.2274 .2772 22.4618 .0000 5.6818 6.7730
intview -.4163 .1165 -3.5747 .0004 -.6455 -.1871
riskbene -.5134 .0560 -9.1626 .0000 -.6237 -.4031
trtseek -.2759 .0548 -5.0316 .0000 -.3839 -.1680

******************DIRECT AND INDIRECT EFFECTS OF X ON Y***********

4 Direct effect of X on Y
Effect se t p LLCI ULCI
-.4163 .1165 -3.5747 .0004 -.6455 -.1871

Conditional indirect effects of X on Y:

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press

 414 ROCKWOOD & HAYES

INDIRECT EFFECT:
intview -> riskbene -> use

severity Effect BootSE BootLLCI BootULCI

.0000 -.2325 .0860 -.4080 -.0717
1.0000 .0324 .0934 -.1490 .2215

Index of moderated mediation (difference between conditional in direct effects):

Index BootSE BootLLCI BootULCI
severity .2650 .1297 .0201 .5331
---

INDIRECT EFFECT:
intview -> trtseek -> use

severity Effect BootSE BootLLCI BootULCI

.0000 .0351 .0471 -.0554 .1317
1.0000 -.2040 .0630 -.3398 -.0947

Index of moderated mediation (difference between conditional in direct effects):

Index BootSE BootLLCI BootULCI
severity -.2391 .0817 -.4131 -.0975
---

***********************ANALYSIS NOTES AND ERRORS****************

Level of confidence for all confidence intervals in output:

95.0000

Number of bootstrap samples for percentile bootstrap confidence intervals:

10000

https://doi.org/10.1017/9781316995808.037 Published online by Cambridge University Press