See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228079476

An Introduction to Statistical Mediation Analysis

Article · January 2008

DOI: 10.4324/9780203809556

CITATIONS READS

3,990 14,439

1 author:

David MacKinnon
Arizona State University
303 PUBLICATIONS   47,451 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SHIELD View project

STD primary prevention View project

All content following this page was uploaded by David MacKinnon on 13 November 2015.

The user has requested enhancement of the downloaded file.


'
I
I
I
STATISTICAL MEDIATION ANALYSIS
CHAPTER
David P. MacKinnon,]eeWon Cheong, and Angela G. Pirlott
Mediating variables are central to psychology
because they explain the processes of psychological
phenomena. As a field, psychology focuses on how
an organism is intermediate in the link between a
stimulus and the response to that stimulus. This
focus on the organism that intervenes between stim-
ulus and behavior was recognized early in psychol-
ogy in the stimulus to organism to response (S-0-R)
model (Woodworth, 1928). In this model, the
organism, a person for example, translates a stimu-
lus into a response by means of mediating processes
within the individual. For example, when a list of
words (S) is presented, the person (0) memorizes
them and then later recalls (R) the words. This
S-0-R model has been extended to understand medi-
ating processes for other units besides individuals-
such as schools, teams, and communities-and is
now widely used to develop and refine prevention
and treatment programs (Kazdin, 2009; MacKinnon,
2008).
Psychological theories specify mediating mecha-
nisms that may explain psychological phenomena.
For example, the theory of reasoned action (Fish-
bein & Ajzen, 1975) in social psychology postulates
that attitudes cause intentions, which in turn cause
behavior. Applying this theory to intervention
research for smoking, an intervention must first
change the attitudes toward the consequences of
smoking, intentions to smoke, and perceptions of
efficacy toward quitting, so that the person can
eventually stop smoking. In cognitive psychology,
memory processes mediate the transmission of
information into a response. When a number of
DO!: 10.1037/13620-018
APA Handbook of Research Methods in Psychology: Vol. 2. Research Designs, H. Cooper (Editor-in-ChieD
Copyright© 2012 by the American Psychological Association. All rights reserved.
•
18
words are presented, using pictorial cues may be
more effective for word recall than memorizing the
words in the presented order. Social learning theory
describes how various behaviors are learned in
social settings. For example, when a child watches a
model being reinforced for performing a certain
behavior, the child will later produce the same
behavior under the same circumstances as a result of
this learning process (Bandura, Ross, & Ross, 1963).
In clinical psychology, a cognitive theory of depres-
sion suggests that changing cognitive attributions
about the self or the world reduces depression
(Beck, Rush, Shaw, & Emery, 1979). In develop-
mental psychology, a theory of attachment postu-
lates that deprivation at birth leads to developmental
deficits, which lead to poor subsequent parenting
behavior (Arling & Harlow, 1967).
In the simplest mediation theory, the investiga-
tion of mediation specifies a chain of relations by
which an antecedent variable affects a mediating
variable, which in turn affects a dependent variable.
Mediating variables can be behavioral, biological,
psychological, or social constructs that transmit the
effect of one variable to another variable. There are
two overlapping applications of mediation theory.
One major application of mediating variables is after
an effect is observed and researchers investigate how
this effect occurred. This application arises from
Hyman's (1955) and Lazarsfeld's (1955) outlines of
elaboration methodologies. In this framework, a
third variable is inserted into the analysis of an X~ Y
relation to improve the understanding of the rela-
tion, that is, to determine whether the relation is
313
MacKinnon, Cheong, and Pirlott
due to a mediator or is spurious. The most notable
citation for this approach to mediation theory is
the classic Baron and Kenny (1986) article, which
clarified the steps to assess mediation described in
earlier references (Hyman, 1955; Lazarsfeld, 1955).
Another type of application of mediation theory is
selecting the mediating variables for intervention on
the basis of theories specifying the causes of the
dependent variable or on prior research demonstrat-
ing that these are candidate causal variables of the
dependent variable. If the mediating variables are
causally related to the dependent variable, then
changing the mediating variables will change the
dependent variable. For example, in drug preven-
tion programs, mediating variables such as social
norms or expectations about drug use are targeted
to change a dependent variable such as drug use.
Many researchers have emphasized the importance
of considering mediation in treatment and preven-
tion research (Baranowski, Anderson, &: Carmack,
1998;Judd &: Kenny, 1981a, 1981b; Kazdin, 2009;
Kraemer, Wilson, Fairburn, &: Agras, 2002; MacKin-
non, 1994; Weiss, 1997). Evaluating mediation to
explain an observed effect is probably more suscep-
tible to chance findings than evaluating mediation
by design because the mediators in the former case
are often selected after the study, whereas the medi-
ators in the latter case are selected in advance on the
basis of theory and prior empirical research. Most
programs of research investigating mediating vari-
ables employ both mediation by design and media-
tion for explanation approaches (MacKinnon, 2008,
Chapter 2).
Because of the importance of identifying mediat-
ing variables in psychological research, methods to
assess mediation are an area of active research. The
purpose of this chapter is to outline current think-
ing about mediation analysis in psychology, but the
length of the chapter precludes addressing all new
developments, which can be found in other sources
(MacKinnon, 2008; MacKinnon, Fairchild, &: Fritz,
2007). This chapter first defines mediation and
other third-variable effects. Statistical mediation
methods using a single mediator case are then
described to clarify the extensions of the single
mediator model discussed in the rest of the chapter.
Assumptions of the single mediator model are then
314
_a~dressed, followed by sections on the latest
,.
I
research on statistical testing of mediated effects,
such as longitudinal mediation models, mediator
and moderator models, and causal inference for
mediation models. Last, directions for future
research are discussed.
DEFINITIONS
The simplest mediation model involves an indepen-
dent variable, X, a mediating variable, M, and
dependent variable, Y. As described elsewhere
(MacKinnon, 2008), there are several different types
of third variable effects. In its simplest form, media-
tion represents the addition of a third variable to an
X --1 Y relation so that the causal sequences can be
modeled such as X causes the mediator, M, and M
causes Y, that is, X --1 M --1 Y. Although the media-
tion relation may appear simple it has several com-
plications. One of these complications is that
mediation is one of several relations that may be
present when a third variable is included in the anal-
ysis of a two-variable relationship. The third variable
is a confounding variable if it causes both X and Y;
ignoring the third variable leads to an inaccurate
inference about the relation of X and Y. If the third
variable is related to X or Y so that information
about the third variable improves prediction of Y by
X, but including the third variable in the analysis
does not substantially alter the relation of X to Y, the
third variable is a covariate. If the third variable
modifies the relation of X to Y such that the X toY
relation differs at different values of the third vari-
able, the third variable is a moderator. A mediator
differs from each of these other third variable effects
in that the mediator is in a causal sequence such that
X causes M and M causes Y (Kraemer et al., 2002;
MacKinnon, Krull, &: Lockwood, 2000; Robins &:
Greenland, 1992, for more information on the third
variable effects).
MEDIATION REGRESSION EQUATIONS
The single mediator model or X _, M --1 Y is shown
in Figure 18.1 and in Equations l to 3:
(l)

Y = i 2 + c' X+ b M + e2,
M = i 3 + a X+ e3,
where Y is the dependent variable, X is the indepen-
dent variable, and M is the mediator; the coefficients
i 1, h, and i 3 are intercepts in each equation; and e1,
e2 , and e3 are residuals. In Equation 1, the coefficient
c represents the total effect (i.e., the total effect that
X can have on Y). In Equation 2, the coefficient c'
denotes the relation between X and Y controlling
forM, representing the direct effect (i.e., the effect
of X on Y that is not intervened by M). The coeffi-
cient b denotes the relation between M and Y con-
trolling for X. Finally, in Equation 3, the coefficient
a indicates the relation between X and M. Equations
2 and 3 are represented in Figure 18.1, which shows
how the total effect of X on Y is separated into a
direct effect relating X toY and a mediated effect
where X has an indirect effect on Y through M. The
current practice of statistical mediation analysis can
be grouped into three approaches: (a) causal steps,
(b) difference in coefficients, and (c) product of
coefficients (MacKinnon, Lockwood, et al., 2002),
which are all based on the information from the
regression equations for testing the single mediator
model.
The first approach to statistical mediation analy-
sis, called the causal steps approach, is based on
the influential work of Baron and Kenny (1986;
Kenny, Kashy, & Bolger, 1998) and judd and Kenny
(1981a, 1981b), originating in Hyman (1955) and
Lazarsfeld (1955). In this approach, a researcher
Mediator
a b
Independent
Variable (X)
FIGURE 18.1. Single mediator model.
·~i~:;:-
Statistical Mediation Analysis -
(2) conducts four steps of analyses to establish media-
tiorl. and estimate Equations 1 to 3. First, the inde-
(3)
pendent variable X should be significantly related to
the dependent variable Y, resulting in the significant
coefficient cin Equation 1. Second, the independent
variable X should be significantly related to the
hypothesized mediating variable M, producing a sig-
nificant coefficient ain Equation 3. Third, the medi-
ating variable M must be significantly related to the
dependent variable Y, controlling for the indepen-
dent variable X, thus finding a significant coefficient
b, in Equation 2. Finally, the relation between the
independent variable X and the dependent variable
Y should be weaker when the mediating variable M
is added to the model. Thus, the coefficient C' should
be smaller than the coefficient c(i.e., c- C' > 0). In the
causal steps approach, the conditions by which a
potential mediator is identified as a significant medi-
ator are clearly established, but the mediated effect
is not directly estimated.
The other two approaches, the difference in coeffi-
cients and the product of coefficients approaches,
involve estimation of the mediated or indirect effect
and its standard error, allowing formal tests for sig-
nificance of the mediated effects. In the difference in
coefficients approach, the mediated effect is esti-
mated by comparing the relations between the inde-
pendent variable X and the dependent variable Y
from Equations 1 and 2, where the effect of X on Y is
estimated with and without adjusting for the media-
tor M. The idea is that the mediated effect can be
estimated by the difference between the total effect
(M)
Dependent
Variable (Y)
c'
315
MacKinnon, Cheong, and Pirlott

and the direct effect that is not attributable to the mediation must come from a program of research,
,.
!

mediator, c - c'. In the product of coefficients
approach, the mediated effect is estimated by the
a
product of and b, ab (Alwin & Hauser, 1975),
from Equations 2 and 3. Thus, the mediated effect
reflects the extent to which the independent variable
X changes the mediator M and the extent to which
the mediator changes the dependent variable Y. The
quantities in Equations 1 to 3 can also be presented
in a plot as described elsewhere (MacKinnon, 2008).
The mediated effect estimated in the difference in
coefficients approach, c- C', is algebraically equiva-
lent to the mediated effect obtained in the product
of coefficients approach, ab, under the normal the-
including a variety of designs and approaches to
uncovering true mediation relations.
Standard Error of the Mediated Effect
One of the most frequently used standard errors of
the mediated effect is the first-order solution derived
by Sobel (1982, 1986) using the multivariate delta
method (Bishop, Fienberg, & Holland, 1975) as
follows:
(4)
a
where and bare estimated regression coefficients

ory ordinary least squares and maximum likelihood

and (]"~ and ut
are the squared standard error ofa
and b from Equations 2 and 3. The formula shown
estimation, as long as the same sample is used for
in Equation 4 is implemented in various structural
Equations 1 to 3 (MacKinnon, Warsi, & Dwyer,
equation modeling programs, such as EQS (Bentler,
1995). When the mediated effect is assessed in mul-
1997) and Mplus (L. K. Muthen & Muthen, 1998-
tilevel models (Krull & MacKinnon, 1999) and
2007), for testing significance of mediated or indi-
logistic or probit regression (MacKinnon & Dwyer,
rect effects. Alternative formulas for the standard
1993), the two estimates of the mediated effect, ab
error for different approaches can be found in
and c- C', are not always equivalent.
MacKinnon (2008) and MacKinnon, Lockwood,
Although the statistical mediation analysis is
Hoffman, West, and Sheets (2002).
straightforward under the assumption that the
model is correctly specified, the identification of
Confidence Limits for the
mediation relations has several complications which
Mediated Effect
can only be addressed in a program of research
Confidence limits of the mediated effects can be
(MacKinnon, 2008). For example, like all statistical
constructed using the estimated mediated effect and
analyses, it is often helpful to consider two models,
its standard error. The upper and lower confidence
the population or true model and the sample model.
limits are as follows:
The population model represents the true relations
among variables, those relations that we would see if Upper confidence limit (UCL) =mediated
we had the entire population. As a sample of data
effect + Zrypel error ( 0"ai,). (5)
represents one of many different samples from the
population, sample estimates of population parame- Lower confidence limit (LCL) = mediated
effect - Zrype I error ( 0"af,). (6)
ters vary from sample to sample. Generally, it is
assumed that the equations and variables specified
For example, when the mediated effect is calculated
in the sample model are the same equations in the
by the product of the coefficients, ab, and the stan-
population model. There are qualifications even to
dard error is estimated by Equation 4, the confi-
this sample-population dichotomy. For example,
dence limits are obtained as follows:
causal inference approaches to mediation suggest
another superpopulation model that represents the ::.£. + *,. .
UV- ZType I error v iib" (7)
true causal relations among variables with all
assumptions validated, similar to the theoretical As the mediated effect is the product of two regres-
mediation model described by MacCorquodale and sion coefficients, its distribution most closely fol-
Meehl (1948). The challenge of using sample data to lows the distribution of the product of two normally
uncover mediation relations is that evidence for distributed random variables, which may not be

316

normally distributed (Springer, 1979). As a result,
more accurate confidence limits can be obtained by
using the distribution of the product of the two ran-
dom variables to obtain critical values (MacKinnon,
Fritz, Williams, & Lockwood, 2007). PRODCLIN
(MacKinnon, Fritz et al., 2007) is a new program
that provides critical values of the distribution of the
product of the two random variables and computes
confidence intervals for the mediated effects
(PRODCLIN can be found at http://www.public.asu.
edu/-davidprn/ripl!Prodclin/).
Alternatively, resampling methods such as boot-
strapping can be used to compute confidence inter-
vals that adjust for the nonnormal distribution of the
product of coefficients (Bollen & Stine, 1990; Mac-
Kinnon, Lockwood, & Williams, 2004; Preacher &
Hayes, 2004; Shrout & Bolger, 2002). From the orig-
inal sample data, repeated samples are obtained, and
on the basis of the mediation effects estimated from
these repeated samples, an empirical sampling distri-
bution is formed to determine the significance of the
mediated effect and to construct the confidence
intervals (Efron & Tibshirani, 1993; Manly, 1997).
Resampling methods are also useful when assump-
tions of statistical methods are not met, such as data
with nonnormal distributions, and also provide con-
fidence limits for more complex mediation models.
More on resampling methods for mediation is
described in MacKinnon (2008, Chapter 12).
Significance Testing
A statistical significance test for the mediated effect
can be conducted in several ways. One way to test
the mediated effect is to assess whether the confi-
dence interval includes zero. When zero is not
included in the confidence interval, the mediated
effect is statistically significant or the hypothesized
mediator is considered a statistically significant
mediator. Another way to test the significance of the
mediation effect is to obtain the ratio of the esti-
mated mediated effect to its standard error and com-
pare the ratio with the critical values under the
normal distributions. For example, ab is divided by
the standard error in Equation 4 and this ratio is
compared with ±1.96. If the ratio is greater than
1.96 or smaller than -1.96, the mediated effect
is significantly different from zero. In simulation
Statistical Mediation Analysis
studies (MacKinnon, Lockwood et al., 2002; Mac-
Kinnon et al., 2004) of the statistical performance of
14 commonly used methods for testing the signifi-
cance of mediation, the most widely used causal step
methods showed quite low power and low Type I
error rates. Power was also low for the methods that
assume the normal distribution of mediated effect,
in which the estimated mediated effect (ab or c- c')
is divided by its respective standard error and the
ratio is compared with the critical values of z or t
distr~bution. A method testing the significance of a
and b and considering there to be significant media-
tion if both tests are statistically significant (i.e.,
joint significance test) showed a good balance
between Type I and Type II errors, as did methods
based on the distribution of the product of two
regression coefficients or bootstrap resampling.
The low power of most methods for testing medi-
ation can be explained in several ways. The low
power of the causal steps approach results from the
requirement of the first step that the relation
between X and Y be significant, especially in the case
of complete mediation (i.e., direct effect c' is zero).
There are cases in which the mediated effect is sta-
tistically significant, even if the relation between X
and Yin Equation 1 is not significant, such as in
inconsistent mediation, in which the direction of the
mediated effect is the opposite of the direction of the
direct effect. Investigating the overall relation of X
and ¥provides useful information for some
research; however, requiring the significant relation
between X and Y for mediation to exist substantially
reduces power to detect real mediation effects for
the causal steps approach. The low power of the
methods using the product of the coefficients and
the difference in the coefficients is due to the non-
normality of the distribution of the mediated effects
(MacKinnon et al., 2004). This nonnormality issue
can be resolved with resampling methods and the
methods based on the distribution of the product of
ab described earlier.
Complete Versus Partial Mediation
The total effect c is the sum of the mediated effect
(ab or c - c') and the direct effect c'. Complete
mediation is defined as a case where the total effect
is completely explained by the mediator. In sample
317
MacKinnon, Cheong, and Pirlott
data, complete mediation is observed when the_
mediated effect is statistically significant (i.e., iib -:t. 0)
and the direct effect is not statistically significant
(i.e., c'= 0), implying that for the population model
the mediated effect and total effect are equal (i.e.,
ab =c). Partial mediation refers to the case where
the relation between the independent and the
dependent variables is not completely accounted for
by the mediator. In sample data, both the mediated
effect ~nd the direct effect are statistically significant
(i.e., ab"' 0 and c' "'0), indicating that the mediator
significantly accounts for part of the relation
between the independent and the dependent vari-
ables, consistent with mediated and direct effects in
the population model (i.e., ab -:t. 0 and c' -:t. 0). Con-
sidering the complexity of the behaviors studied in
social science research, there may be a variety of
causes of those behaviors, and thus it is often more
realistic to expect that a single mediator would only
partially explain the relation between the indepen-
dent and the dependent variables (Baron & Kenny,
1986). james, Mulaik, and Brett (2006) argued for
the specification of complete or partial mediation
prior to study and first testing for complete media-
tion. The test for complete mediation does not
include the c' parameter in the model, and as a
result, inadequate model fit in covariance structure
analysis suggests incomplete mediation. If theory
predicts complete mediation, this approach provides
a straightforward way to test mediation.
MEDIATION EQUATIONS WITH
TWO MEDIATORS
The two-mediator model helps to compare the
causal steps, difference in coefficients, and product
of coefficients approaches to testing mediation in
more complicated models. When the mediation
model includes two mediators, the regression equa-
tions can be expanded as follows (MacKinnon,
2008):
Y = i1 + c X+ e1 , (8)
Y = i2 + c' X+ b1M1 + b2M2 + e2, (9)
M1 = i3 + a1X + e3, (10)
M 2 = i 4 + a2X + e4, (11)
318
where X is the single independent variable, M1 and
I !
M2 are two mediators, andY is the dependent vari-
able. The coefficients a 1 and a2 represent the rela-
tions between the independent variable and the two
mediators, respectively. The coefficients b1 and h
represent the relations between each mediator and
the dependent variable controlling for the indepen-
dent variable and the other mediator. Again, an esti-
mator of the mediated effect can be obtained by the
difference between the total effect and the direct
effect, that is, c- C' or by the product of the relevant
coefficients (i.e., a1b1 and a2b2). The estimator of
c- C' is the total mediated effect, that is, the mediated
effect via both M1 and M2 . On the other hand, the
estimators of a1b1 and a2b2 are the mediated effects
uniquely attributable to M1 and M2 , respectively.
The two-mediator Equations 8, 9, 10, and 11
demonstrate the superiority of the product of coeffi-
cients approach compared with the causal steps and
difference in coefficients tests. Each method is rela-
tively easily applied to the case of a single mediator;
however, only the product of coefficients method is
directly applicable to more complicated models.
Causal steps and difference score methods could be
applied but are more cumbersome. With two media-
tors, there are now three mediated effects, the effect
of X on Y through M1 , the effect of X on Y through
M2 , and the total mediated effect of X on Y through
M1 and M2 , called the total indirect effect. With the
causal steps method, it is possible to test the signifi-
cance of the aand bpaths corresponding to each
mediated effect, but there is not a direct test of the
total mediated effect, the sum of a1b1 and a2b2 . The
other limitation of the causal steps method is that
there is not a direct point estimate and standard
error useful for confidence limits and effect size
calculation. In the product of coefficients method,
however, it is much more straightforward to test
whether the total mediated effect, the sum of a1b1
and a2b2 , is statistically significant either by forming
the ratio of the estimate to estimated standard error
or by creating confidence limits via a resampling
method. The difference score method is appropriate
to test the total indirect effect because it is the differ-
ence between the relation of X on Y before and after
adjustment for both mediators, c - C'. The standard
error of this difference for the total indirect effect is

given in MacKinnon (2008). However, the differ-
ence in coefficients method is quite cumbersome for
testing the specific indirect effects, alb] or a2b2,
because it is not clear which coefficients would be
compared. These problems with causal steps and
difference in coefficients methods are magnified in
more complicated models. These quantities can be
obtained in a straightforward manner with the prod-
uct of coefficients method.
CONSISTENT AND INCONSISTENT
MODELS
Consistent or inconsistent mediation models are
determined by the signs of the mediated and the
direct effects. For consistent mediation models, all
the mediated and the direct effects in a model
have the same sign. Inconsistent mediation models,
on the other hand, include at least one mediated
effect having the opposite sign to the other mediated
or direct effects in the model (Blalock, 1969; Davis,
1985; MacKinnon et al., 2000). When the signs of
the mediated effect(s) and the direct effect are
inconsistent, the overall relation between the inde-
pendent and the dependent variables (i.e., the total
effect) may not be statistically significant, as
described in MacKinnon et al. (2000).
The two-mediator model also is helpful to dem-
onstrate inconsistent mediation effects. Sheets and
Braver (1999) hypothesized that an overall relation
between social dominance and sexual harassment
would be zero because of opposing mediated effects
via the harasser's power (M 1) and desirability (M 2).
There are other hypothetical examples of possible
counterproductive effects. For example, an absti-
nence program may increase intentions to abstain
but participating in the program may also increase
interest in sexual activity leading to a null program
effect because of opposing mediators. For another
example, incarceration in prison may lead to reha-
bilitation, which reduces recidivism, but exposure
to fellow inmates in prison may also engender a
norm more favorable toward criminal activity,
which then increases recidivism. It is possible that
any intervention would have opposing mediation
effects when the program is composed of multiple
components designed to change the outcome. The
Statistical Mediation Analysis
opposing mediation relations would be observed
more easily if multiple mediator models are
estimated.
EFFECT SIZE MEASURES OF MEDIATION
Effect size in mediation models can be specified for
each path involved in the mediation pathway and
for the entire mediated effect. To specify the effect
size of each path, correlations, partial correlations,
and standardized coefficients are used. There are
several ways to define effect size of the entire medi-
ated effect. One of the most commonly used effect
size measures is the proportion mediated, which is
obtained by the ratio of the mediated effect to the
total effect, that is, ab!(ab + C'). The proportion
mediated has heuristic value, in that researchers can
gauge the effect size in terms of the proportion of
the total effect that is mediated. Limitations of the
proportion mediated effect include the large sample
size requirements (MacKinnon et al., 1995), ambi-
guity when effects are small, and ambiguity in inter-
pretation for inconsistent mediation models when
the mediated and the direct effects have the opposite
signs (taking absolute values of all the effects before
calculating the proportion mediated may be helpful
in this case; see Alwin & Hauser, 1975). Other mea-
sures of effect size for the entire mediated effect are
the ratio of the mediated effect to direct effect, ab!C',
R2 measures (Fairchild, MacKinnon, Taborga, &
Taylor, 2009), and effect standardized by variance of
the dependent variable (see MacKinnon, 2008, for
information on mediation effect size measures).
ASSUMPTIONS OF THE SINGLE
MEDIATOR MODEL
Most current developments in mediation analysis
address statistical and infer~ntial assumptions of the
mediation model. For the ab estimator of the medi-
ated effect, several simultaneous regression analysis
assumptions are required, including that the mediator
and the residual in Equation 2 are independent and
also that the residuals in Equations 2 and 3 are inde-
pendent (MacKinnon, 2008; McDonald, 1997). It is
also assumed that there is not an interaction between
X and Min Equation 3, although this interaction can
319
MacKinnon, Cheong, and Pirlott
be tested and in some cases may be expected on the
basis of the theory outlined in the next section. The
temporal order of the variables in the model is also
assumed to be correctly specified (e.g., X---? M---? Y
rather than X---? Y---? M). Several other types of
model specification are assumed to be correct, includ-
ing self-containment-that no variables related to the
variables in the mediation equations are left out of the
estimated model and that coefficients estimate causal
effects. It is also assumed that the model has minimal
errors of measurement (Holland, 1988;James &
Brett, 1984; McDonald, 1997).
Assumption of No Interaction Between X
andM
The X x M interaction could be included in Equa-
tion 2, which would suggest a moderator effect such
that the b coefficient differs across the levels of X.
Different b coefficients across levels of X may reflect
that an experimental manipulation may have
changed the relation of M to Y. For example, a
smoking prevention program may remove a relation
between offers to use tobacco (M) and subsequent
tobacco use (Y) in the program group but not for
participants in the control group, because persons
exposed to the program learned skills to refuse
offers of tobacco so that offers are not significantly
related to tobacco use Qudd & Kenny, 1981a). Sig-
nificant XM interactions may also be obtained when
there are other nonlinear relations in the model. If a
program increases M to a value so that the relation
between MandY differs from the relation at other
levels of M, the XM interaction would be statistically
significant because of a nonlinear relation between X
and Y. If there is some other variable that is an
important mediator it is possible that this variable
may predict both MandY, leading to a statistically
significant XM interaction.
Assumption of Correct Model
Specification
There have been many important recent extensions
to address limitations of the standard mediation
analysis described thus far. First, more complicated
models are often hypothesized. These models may
include multiple independent variables, multiple
mediators, and multiple outcomes. With these more
320
comprehensive models, the relations among vari-
ables in the mediation model may be more explicitly
specified and the mediation effects may be more
accurately estimated. Second, when the data are
clustered, methods for testing mediation within and
across levels have been developed to accommodate
the statistical issues in multilevel analysis and also
to explore the rich information in multiple levels of
analysis (Krull & MacKinnon, 1999, 2001; Preacher,
Zyphur, & Zhang, 2010). Third, mediation effects
may differ by subgroups defined by moderator vari-
ables both within (such as M or Y) the mediation
model and outside (variables other than M
or Y) the mediation model (Fairchild & MacKinnon,
2009; Muller ,Judd, & Yzerbyt, 2005; Preacher et al.,
2010). Fourth, mediation requires temporal prece-
dence clarifying that X affects M that affects Yin a
longitudinal or temporal order (Gollob & Reichardt,
1991; Kraemer et al., 2002; MacKinnon, 2008).
Finally, developments in the causal interpretation of
mediation studies (Holland, 1988; Robins & Green-
land, 1992) provide a useful framework to describe
the strengths and limitations of possible causal
inferences from a mediation study.
MEDIATION WITH CATEGORICAL
OUTCOMES
In some mediation studies the dependent variable
is categorical, such as whether a person suffered a
heart attack or died or not. In such cases, Equa-
tions 1 and 2 must be rewritten for logistic or pro-
bit regression, so that the dependent variable is a
latent continuous variable that has been dichoto-
mized for the analysis. Because the residual vari-
ances in logistic or probit regression are fixed, the
parameters c, c', and b depend on the other predic-
tor variables in the model. So, the differences in
coefficients across models could reflect real differ-
ences, but they could also be artificial effects
caused by the fixed error variance (MacKinnon,
2008). For example, theestimatorc-c' for media-
tion may be incorrect because the parameter esti-
mate of c' depends on the true relation of the
mediator to the outcome and the scaling of Equa-
tions 1 and 2 (MacKinnon & Dwyer, 1993). A solu-
tion to this discrepancy is to standardize regression

coefficients before estimating mediation (MacKin-
non, Lockwood, Brown, Wang, & Hoffman, 2007;
Winship & Mare, 1983), which will bring the two
estimates of the mediated effect closer in value.
LONGITUDINAL MEDIATION MODELS
When the same variables are measured repeatedly,
researchers can examine more complex questions
regarding mediating mechanisms. The time ordering
of X, M, andY can be specified in the mediation
model with longitudinal data and researchers can be
more confident in the causal sequence of the media-
tion effect than in cross-sectional studies. Longitudi-
nal data also allow researchers to investigate
mediation in terms of within-individual changes.
For example, mediation can be modeled as X causes
changes in M between pre- and post-test, which, in
turn, causes changes in Y between pre- and post-
test. In addition, one can examine the stability of
effects across time, such as whether the effect of X
on M or the effect of M on Yare stable across time,
and test the stability of the mediation effect across
time. However, it is challenging to determine the
optimal timing of measurements to accurately assess
when longitudinal relations occur. Also, researchers
need to pay more attention to the potential mis-
specification of the model, such as omitted variables
or paths, and hypothesizing correct mediation
pathways (Cheong, 20ll; Cole & Maxwell, 2003;
Collins, Graham, & Flaherty, 1998).
Although there are various ways to model longi-
tudinal mediation relations (MacKinnon, 2008), we
present four approaches: autoregressive models,
latent growth curve models, latent difference scores
models, and person-oriented longitudinal models.
Besides these approaches, one can combine the
autoregressive model and latent growth curve model
(Bollen & Curran, 2004) or specify the model
parameters in a continuous time metric to reflect
different time intervals between measurements
(Boker & Nesselroade, 2002; Fritz, 2007).
Autoregressive Mediation Model
In the typical autoregressive mediation model, rela-
tions among X, M, and Y one measurement occasion
apart are specified. When the mediation is evaluated
Statistical Mediation Analysis
in autoregressive models, researchers have several
options for modeling mediation. First, one can focus
only on the relations consistent with longitudinal
mediation, such as Xn -7 M12 -7 Yn, assuming that
mediation would not occur within the same wave.
Another possibility is to add contemporary media-
tion relations (e.g., X12 -7 M12 , M12 -7 Yu) to the
longitudinal autoregressive mediation model and
estimate the contemporary mediation effect within
each wave, except the first wave, for which the rela-
tions among X, M, and Yare typically specified as
correlated. A third type of autoregressive mediation
model includes any possible relations among X, M,
and Y, including longitudinal relations based on
time ordering (e.g., Mn -7 X12 , Y12 -7 Mn), which
may be counterintuitive in that the directions of the
relations among X, M, andY are the opposite to the
hypothesized mediation model. This type of model,
however, may be more realistic, considering that X,
M, andY are interrelated. It is possible that M12 is
predicted by Yn because Yn is related to Mn.
Regardless of the type of autoregressive model,
the estimated mediation effect and its standard error
can be obtained in the usual way, using the relevant
path coefficients and their estimates of standard
errors, to test the significance of the point estimate
of mediation and construct confidence intervals.
More details on autoregressive mediation models
can be found in MacKinnon (2008) and Gollob and
Reichardt (1991).
Autoregressive mediation models are beneficial
because they can provide information about time-
specific mediation effects, such as when the media-
tion effects start to occur or when they stop
working. It is a common practice, however, to esti-
mate autoregressive models only using covariance
structure without mean structure, resulting in reli-
ance on the order of individuals in the variables and
ignoring the level of longitudinal changes (see also
Dwyer, 1983; Rogosa, 1988). In addition, as with
many potential mediation effects, it may not be easy
to determine which mediation relation represents
the true model.
Latent Growth Mediation Model
When latent growth modeling (LGM) is applied for
testing mediation, the mediation process is typically
321
MacKinnon, Cheong, and Pirlott
modeled using parallel process models (see B. 0.
Muthen & Curran, 1997; Singer & Willett, 2003, for
more on LGM), where the growth trajectories of X,
M, and Yare estimated in three distinctive growth
trajectories and the mediation is hypothesized in the
relations among the growth factors. Typically,
researchers examine whether the slope of X affects
the slope of M and whether the slope of M, in turn,
affects the slope of Y. Alternatively, mediation can
be evaluated in the relations among the initial level
of X, the slope of M, and the slope of Y. As in the
mediation models described in earlier sections, the
relation between the trajectory of X and the trajec-
tory of Y has two sources: the indirect effect via the
trajectory of M and the direct effect on the trajectory
of Y. The unique aspect of testing mediation in the
LGM framework is that the mediation is modeled
for individual changes, rather than the levels of indi-
viduals on the variables. Thus, one can examine
whether the greater changes in Yare the results of
the greater or smaller changes in M.
When the mediated effect is examined on the
basis of the relations of the slope factors of X, M,
andY that are measured across the same time peri-
ods, a causality explanation is limited because the
relations among the slopes of the three trajectories
are correlational. When the variable X represents
randomized group status, the effect of X on the
slope of M can be interpreted as causal, but the rela-
tion between the slope of M and the slope of Y is
still correlational. One way to improve causal
explanation in the LGM mediation model is using
the two-stage piecewise parallel process model
(Cheong, MacKinnon, & Khoo, 2003), in which X
affects the growth of Mat an earlier phase, which
then affects the growth of Y at a later phase. Mea-
sures are very important in LGM mediation models.
If the measures change over time, the interpretation
of the change will be confounded. For example, if
the items measuring aggression are modified to
measure aggressive behaviors that are age appropri-
ate, it is difficult to determine whether the observed
change is due to the changes in aggression or due to
the changes in measures. For this reason, measure-
ment invariance across time may need to be tested
before modeling the mediational process in LGM
framework.
322
Latent Change Score Model
As in the LGM approach to mediation, latent change
score (LCS) models also examine the relations
among the changes in X, M, and Y. Although the
changes in the LGM approach are estimated on the
basis of several waves across time, the changes in the
LCS approach are estimated on the basis of pairs of
adjacent waves (Ferrer & McArdle, 2003; McArdle,
2001; McArdle & Nesselroade, 2003). Again, the
relation between the change in X and the change in
Yare composed of two parts (i.e., the indirect effect
via the change in M and the direct effect) and the
interpretation of mediation is similar to the LGM
approach: the change in X affects the change in M,
which then affects the change in Y.
In the LCS modeling, the true scores at each
wave and the change scores between waves are esti-
mated using latent variables and fixed parameters.
Once the change scores are obtained, these change
scores are then analyzed using the same equations as
for cross-sectional models. LCS mediation models
with more than two waves of data are particularly
informative when researchers expect different medi-
ation mechanisms for changes at different waves of
measurement. For example, the treatment program
may change adolescent drug use via change in
parental monitoring at early adolescence but via
change in peer norms at later adolescence. In addi-
tion, researchers can test time-specific mediation, as
in autoregressive mediation models, by investigating
when the changes in X start to affect the changes in
M or when the program effects start to decay. In
addition, one can examine the change in the differ-
ence scores in LCS models and specify the models to
represent moving averages.
Person-Oriented Longitudinal Models
Several different approaches have been suggested to
identify subgroups of persons on the basis of their
values of the independent variable, mediating vari-
able, and dependent variable. Typically, binary
latent variables are created to indicate individuals'
status, that is, to signify whether the individual's
responses are consistent with hypothesized media-
tion pattern or not. Two of these original approaches
are based on trajectory classes (L. K. Muthen &
Muthen, 1998-2007) and staged responses across
 .
.
I-G:
~
trials (Collins et al., 1998). These models represent
several new ways to understand mediational pro-
cesses at both the individual level as well as the
group level (von Eye, Mun, & Mair, 2009). A related
approach to identifying mediational processes is to
focus on single-subjects data with repeated mea-
sures. In this approach, mediation relations observed
with one subject are tested with future subjects to
provide cumulative evidence for a mediation relation.
In some research areas where sample sizes are small,
such as some clinical populations, single-subject
methods may be the only reasonable approach.
MODERATION AND MEDIATION
The strength and form of mediation relations may
depend on other variables. For example, the relation
of X to M (a path) or M to Y (b path) may differ
across levels of a moderator variable, resulting in
different mediated effect (ab) across levels of moder-
ator variable. The moderator variable may be either
an experimentally manipulated factor or a naturally
occurring variable, such as gender or ethnicity.
These types of models have been an active area of
research in the past few years (Fairchild & MacKin-
non, 2009; MacKinnon, 2008; Muller et al., 2005).
One way to organize the different types of mediation
analysis with moderator variables is to consider two
different types of cases: (a) moderation of a medi-
ated effect and (b) mediation of a moderator effect.
Moderation of a Mediated Effect
In the first case, moderation of a mediation relation,
the X toM, M toY, or the entire mediation relation
may differ across levels of a moderator variable,
such as subgroups of participants (e.g., cohorts,
ages, or sexes). For a single mediator case, the appli-
cation of moderation of a mediation analysis con-
sists of estimating the same mediation model for
each subgroup and then comparing the X to M rela-
tion, theM to Y relation, and the mediated effect
across subgroups. The equivalence of the mediated
effect across groups can be tested (MacKinnon,
a,
2008). Tests of the equality of b, and c' coeffi-
cients provides information about the invariance
of action theory (how the program changes media-
tors) and the invariance of conceptual theory
Statistical Mediation Analysis
(how mediators are related to the outcome) across
groups. The moderation of a mediated effect is more
complex when the moderator variable is continuous.
Although the regression equations are the same as
for the categorical moderator case, the interpretation
of results may be complicated for the continuous
moderator case because of the large number of val-
ues of the moderator at which mediation relations
may potentially differ. One approach is to apply a
multilevel mediation model to these data
(Asparouhov & Muthen, 2008; L. K. Muthen &
Muthen, 1998-2007).
Mediation of a Moderator Effect
In a second type of mediation and moderation anal-
ysis, mediation of a moderator effect may be investi-
gated. In this situation, a mediating variable is
sought to explain how an interaction between two
variables is related to a dependent variable, that is,
to investigate evidence that the mediator transmits
the relation from an interaction to a dependent vari-
able. One common example from the treatment and
prevention literature is that program effects are
greater for high-risk subjects so that the interaction
effect is of program exposure and risk status. In this
case, the interaction may affect a mediating variable
of social norms that then affects drug use. The pur-
pose of mediation of a moderator analysis is to
assess whether the mediating variable(s) explains
the interaction effect. Investigation of these effects
consists of estimating a series of regression equa-
tions for which the main effect of a covariate and the
interaction of the covariate and the independent
variable are included in both Equations 2 and 3.
To date, models investigating mediation and
moderation have been largely independent, which is
not surprising given the complexity of investigating
mediation and moderation alone. This separation in
the theory and statistical testing between modera-
tion and mediation has contributed to some ambigu-
ity regarding the substantive motivation and
statistical testing of these models. A critical goal of
future research for mediation and moderation mod-
els will be to further develop and evaluate a general
model in which each of the models is a special case
(Fairchild & MacKinnon, 2009; MacKinnon, 2008;
Muller et al., 2005).
323
MacKinnon, Cheong, and Pirlott

CAUSAL INFERENCE on M, ii, are estimators of the true causal effect

because of the randomization of units to treatment. In
Methods for testing mediation that are based on
contrast, the regression coefficient, b, is not an accu-
regression and structural equation modeling
rate causal effect estimator because this relation is
approaches have been criticized from the perspec-
correlational-participants are not directly random-
tive of counterfactual approaches to causal analysis
ized to scores on the mediator. Along the same logic,
of the relations among variables (MacKinnon, 2008,
the estimator, c', is also not an accurate causal estima-
Chapter 13). Limitations to mediation models in
tor of the direct effect because this relation is also cor-
particular and structural equation models in general
relational. These approaches provide an instrumental
have been outlined in the social science (Holland,
variable approach, whereby an estimate of the true
1988;james et al., 2006; Morgan & Winship, 2007)
causal relation between M and Y is the extent to
and epidemiological literature (Robins & Green-
which the predicted scores in the X toM relation M'
land, 1992). One widely known limitation of these
are related to Y, assuming a linear additive relati;n of
models is the equivalent model criticism. Applied to
M toY and assuming that there is no direct effect of X
the mediation model, if X, M, and Yare measured at
on Y (i.e., complete mediation). Several new
the same time, other equivalent models (e.g., M ~X
approaches to causal inference for mediation have
~ Y or M ~X t- Y) would explain the data equally
added to this original work. These methods use addi-
well and are often indistinguishable without more
tional information, such as covariates or specifying
information (Spirtes, Glymour, & Scheines, 1993).
types of persons on the basis of their response to any
Another limitation is the temporal precedence
intervention, to improve causal inference of media-
assumption. Cross-sectional data require additional
tion effects. One of these promising alternatives is
assumptions if they are to be interpreted in a causal
based on ways in which participants could respond to
fashion. A third limitation is the aforementioned
an intervention and estimates the mediated effect
assumption of no omitted variables. When X repre-
within these stratifications (Angrist, Imbens, &
sents randomization to conditions, inference regard-
Rubin, 1996; Frangakis & Rubin, 2002;]o, 2008). For
ing X toM and X toY relations is less problematic.
example, different types of persons are identified,
With randomization of a sufficient number of units
such as persons who would acquire the mediator if
to levels of X, all unmeasured covariates are
exposed to the intervention and persons who would
assumed to be balanced between groups at baseline
not get the mediator whether exposed to it or
and, thus, the relation of X toM and the relation of
not. Sobel (2007) enhanced the Holland (1988)
X toY (that is not adjusted forM) can be attributed
instrumental variable method to further investigate
to the randomized manipulation. However, infer-
assumptions of the method. Other alternatives use
ence regarding theM to Yand X to ¥adjusted forM
randomization approaches on the basis of covariates
relations is problematic because individuals are not
to model observed and counterfactual data (Lynch,
randomly assigned to the levels of M and thus omit-
Cary, Gallop, & Ten Have, 2008; Pearl, 2009; Robins,
ted var~ables may seriously affect the interpretation
1994; Ten Have et al., 2007).
of the band c' coefficients.
One of the most important contributions of these
Modem causal inference approaches suggest strat-
causal inference approaches is that they provide
egies for investigating assumptions of the statistical
alternative ways to improve causal interpretation in
methods, especially the ambiguity regarding the inter-
mediation analysis especially for theM toY relation.
pretation of the M to Y relation. When X represents
However, the fact that theM toY relation in the
random assignment to conditions, causal interpreta-
mediation model is considered the weakest link is
tion of mediating variables is improved (Holland,
interesting from the perspective of theories in psy-
1988; Robins & Greenland, 1992) because X must
chology. In most psychological studies, theM to Y
causally precede M and Y. Holland showed that under
relation is specified on the basis of theory, extensive
some assumptions, the regression coefficient for the
prior research, and a variety of information besides
intervention effect on Y, c, and the intervention effect
statistical analysis, and mediators are selected

324

because of this information. The emphasis on ambi-
guity of the M to Y relation may be reduced in psy-
chology because it is generally much easier to
conduct replication and extension experiments in
psychology compared with some other disciplines,
such as medicine, for which conducting studies is
often more expensive and requires extensive fol-
low-up for disease processes to emerge. The applica-
tion of causal inference approaches in psychology
will remain an important and active area of research
in part because it illuminates untested assumptions
in mediation theory. Regarding the causal interpre-
tation of mediation relations, researchers have sev-
eral options based on current research practice.
First, new models of causal inference can be applied
to mediation analysis, although this may be difficult
given the paucity of clear, concrete examples of their
application in psychology. Second, one can treat the
results of the mediation analysis as descriptive infor-
mation, rather than true underlying causal media-
tion relations, especially for theM to Y relation, and
address the variety of limitations of such a media-
tion analysis. Third, one can plan future experimen-
tal studies to provide evidence for the consistency
and specificity of mediation relations as described in
the next section. These future studies may also
include qualitative methods and clinical observa-
tions. In particular, a program of research that
repeatedly tests mediator theory, including testing
mediation theory in other contexts, provides the
most convincing evidence for mediation.
EXPERIMENTAL DESIGNS TO
INVESTIGATE MEDIATION
Most mediation designs in psychology use random
assignment of participants to experimental conditions
with the goal of obtaining results consistent with one
theory and inconsistent with another (MacKinnon,
Taborga, & Morgan-Lopez, 2002; Spencer, Zanna, &
Fang, 2005; West & Aiken, 1997). In these studies,
group mean differences in the outcome are then
attributed to the experimental manipulation of the
mediator, without measuring these mediators. Experi-
mental results in conjunction with theoretical predic-
tions provide support for mediation hypotheses and
recommend further studies to specify and validate the
Statistical Mediation Analysis
mediating process. In this section, we first discuss
double randomization designs and describe two
domains of experimental mediation: specificity and
consistency. More examples of these designs can be
found in MacKinnon (2008).
Double Randomization Design
A double randomization design involves two sepa-
rate randomized studies for investigating mediation
relations. In Study 1, participants are randomized to
the levels of X to determine the causal relation of X
toM. On the basis of the results of Study 1, the lev-
els of M are defined and participants are then ran-
domly assigned to these levels in Study 2. If a
significant relation of M to Y exists in Study 2, there
is more evidence for causality in the mediation rela-
tions (MacKinnon, Lockwood et al., 2002; Spencer
et al., 2005; West & Aiken, 1997).
Double randomization experiments provide use-
ful information about the mediation relations and
reduce the plausibility of alternative explanations;
however, the practicality of implementing double
randomization experiments is often limited. The
greatest drawback is the requirement of random
assignment of participants to the levels of the media-
tor so that theM~ Y relation can be experimentally
tested, although there may be cases in which the
mediator is external so this may be more feasible.
The extent to which the manipulation can directly
change the mediator is an important aspect of these
designs as well. If a manipulation is able to change a
mediator such that the relation between X and M is
close to perfect, then the manipulation does indeed
directly change the mediator. As the dependency
between X and M approaches perfection, however,
there will be statistical problems owing to multicol-
linearity between X and M.
Specificity Designs
Mediation studies with specificity designs focus on
observing the hypothesized mediation relations only
for certain predicted variables or groups to ascertain
that the mediation relations are unique or specific to
those target variables or groups. For example, exper-
imental mediation designs assessing the specificity
of X examine the extent to which the mediation
relation between X and M is specific to certain
325
MacKinnon, Cheong, and Pirlott
manipulations of X but not others. Thus, different
versions of X are manipulated to determine which
specific component of X is responsible for the
change in M. Similarly, studies assessing the specific-
ity ofY examine whether the mediation relation is
observed for some dependent measures (Y) but not
others. Experimental designs focusing on the speci-
ficity of moderators examine the extent to which the
mediation relations are observed for certain groups
but not others (e.g., culture, ethnicity or race, socio-
economic status, region, and gender). The specific-
ity of a moderator could also examine specific levels
of a continuous moderator or individual difference
variable, such as the need for cognition or level of
depression.
Experimental designs examining the specificity of
the mediator demonstrate that mediation relations
are observed for some mediators (M) but not others,
therefore providing evidence that the specific com-
ponents of the mediator drive the relation. Three
major types of specificity of mediator designs are
multiple mediator, blockage, and enhancement designs.
Multiple mediator designs test multiple mediators
within an experiment to determine which mediator
is responsible for the change in Y. These designs
provide evidence of the specificity of M to determine
what mediating component drives the mediation
relation. Blockage designs involve manipulations
that block the mediator from operating. To deter-
mine whether a certain variable actually mediates
the relation between X andY, one would block the
mediator and examine whether the X toY relation
still occurs. If the X to Y relation no longer exists,
this provides evidence that the X to Y relation is
causally dependent on the~ mediator. Enhancement
designs involve interventions that enhance the
effects of a hypothesized mediator, thus providing
further support for the mediational process. To
determine whether a certain variable actually medi-
ates the X to Y relation, one would include a manip-
ulation that would enhance the mediator and
examine whether the X to Y relation occurs in
greater magnitude.
Consistency Designs
Consistency designs replicate mediation relations in
new settings, groups, species (animals, humans),
326
and times to provide evidence that the mediation
relation is consistently observed across many
domains and variables. Consistency designs for X
seek to replicate the mediation relation with alterna-
tive independent variables. This provides evidence
of the generalizability of the X~ M relation across
settings, time, and populations. Consistency designs
forM seek to replicate the mediation relation with
alternative mediators. Consistency designs for Y repli-
cate the mediation relations with alternate depen-
dent variables. Again, this provides support of the
generalizability of the X and M variables in the
mediation relation. Consistency designs replicate
the mediation relation across different groups, for
example, across race and ethnicity, socioeconomic
status, regions, gender, and individual difference
measures. These designs would also hint at evidence
for the universality of psychological phenomena.
Statistical tests of mediation relations are needed
in all of the experimental designs. For example in
the double randomization design, the manipulation
of M in the M to Y study may not be perfect, so this
model is again a mediator model with the manipula-
tion of X, the mediator M, and the dependent vari-
able Y. Statistical analysis of mediation relations are
needed in this design as they are in other research
studies. Similarly, tests of moderation of a mediation
relation are important in specificity designs to dem-
onstrate different mediation relations across groups.
Consistency designs are also improved with analyses
demonstrating failure to reject the hypotheses that
effects are consistent across groups.
GUIDELINES FOR REPORTING
MEDIATION ANALYSIS
The purpose of this section is to provide a number
of guidelines for reporting the research study on
mediation relations in psychology. Most of the char-
acteristics of mediation studies are the same as for
any research study in psychology. First, describe the
theoretical and empirical basis for hypothesized
mediation relation before conducting the study.
Describing the mediation theory will clarify the
overall purpose of the study and will likely force
consideration of alternative interpretations of the
results of the study, leading to better research

design. If the mediation theory is complex, explicitly
indicate which mediated effect or combinations of
mediated effects will be investigated in the study as
well as the pattern of effects in the mediation model.
Discuss how the mediators targeted for study are the
critical mediators on the basis of prior research. Sec-
ond, describe results of each link in the mediation
chain and report the estimated mediated effect, stan-
dard errors, and confidence limits as well as effect
size measures. If there are more links in the chain,
then significance tests for each link in the chain
clarify the accuracy of mediation theory. Calcula-
tions of confidence limits and significance tests for
the mediated effect should be conducted using a
method that incorporates the nonnormal distribu-
tion of the product, either on the basis of the distri-
bution of the product or the resampling methods.
Third, clear discussion of how the study assessed
the specified temporal relation among variables is
necessary. If longitudinal data are not available,
defense of the ordering in the analysis may be more
difficult but potentially more important both to bol-
ster evidence for a mediation relation and to guide
the design of future longitudinal studies. Fourth,
discuss how omitted variables may alter conclusions
and provide some indication of the sensitivity of the
observed results to additional confounding vari-
ables. Fifth, directly address problems with inter-
preting the M to Y relation. As emphasized in causal
inference approaches, randomization is central to
defending a hypothesized mediation relation. Thus,
the extent to which M can be considered random-
ized should be addressed by considering counterfac-
tual cases, such as how the relation between M and
Y may differ across experimental groups. Also,
detailed discussion of the mediator investigated in
the study is useful. Is the mediator you measured
the actual mediator? Is there a more fine-grained
mediator that may actually be the most important in
changing Y? Similarly, if you could measure addi-
tional mediators what would they be? Sixth,
describe additional designs and research findings
that could be used to clarify a mediation relation. In
particular, future experimental studies to investigate
the consistency and specificity of the mediation rela-
tion are necessary to provide more convincing evi-
dence for mediation. Overall, the identification of
a
Statistical Mediation Analysis
mediation relations requires a sustained program of
research, including many different types of informa-
tion including qualitative, quantitative, and clinical
information. During this process, it may be useful
to incorporate prior information on mediation
relations in a Bayesian perspective (Yuan &
MacKinnon, 2009).
CONCLUSION
Extensive interest in statistical analysis of mediation
relations is understandable considering that psycho-
logical theories focus on the process by which phe-
nomena occur. Mediation relations are also of
interest in many other fields, including epidemiol-
ogy, public health, and medicine. Mediating pro-
cesses are critically important for intervention and
prevention research because they provide informa-
tion that can be used to make these interventions
more efficient and powerful. Demonstration of
mediation relations is a difficult and challenging
process, but there has been considerable recent
development in methods to assess mediation. Signif-
icance tests for mediation on the basis of the non-
normal distribution of the product are most
accurate, including tests directly based on the distri-
bution of the product and methods that model the
distribution of the product, such as the bootstrap or
resampling methods. Multiple mediator and more
comprehensive models allow for consideration of
omitted and additional variables that may be central
to test mediation. Alternative longitudinal mediation
models provide important opportunities to test tem-
poral relations among variables in the mediation
model. Complementary investigation of mediation
relations with person-oriented models provides
more evidence for true mediation relations. Devel-
opments in causal inference for mediation relations
are rapidly increasing, thereby providing an accurate
assessment of the limitations and strengths of con-
temporary mediation methods. Approaches to test
the sensitivity of tests of mediation to violations of
assumptions should add greatly to the identification
of true mediation relations. An important character-
istic of psychological research is that it is often eas-
ier to conduct randomized replication studies than
in other fields, such as sociology or epidemiology.
327
MacKinnon, Cheong, and Pirlott
This opportunity for replication is ideal for testing
mediation theory in the variety of applications nec-
essary to demonstrate the consistency and specificity
of theoretical mediating process.
References
Alwin, D. F.,&: Hauser, R. M. (1975). The decomposi-
tion of effects in path analysis. American Sociological
Review, 40,37-47. doi:10.2307/2094445
Angrist,J. D., Imbens, G. W., &: Rubin, D. B. (1996).
Identification of causal effects using instrumen-
tal variables (with commentary).]ournal of the
American Statistical Association, 91, 444-4 72.
doi:l0.2307/2291629
Arling, G. L., &: Harlow, H. F. (1967). Effects of social
deprivation on maternal behavior of rhesus monkeys.
journal of Comparative and Physiological Psychology,
64,371-377. doi:l0.1037/h0025221
Asparouhov, T., &: Muthen, B. 0. (2008). Multilevel mix-
ture models. In G. R. Hancock&: K. M. Samuelsen
(Eds.), Advances in latent variable mixture models
(pp. 27-51). Charlotte, NC: Information Age.
Bandura, A., Ross, D.,&: Ross, S. A. (1963). Vicarious
reinforcement and imitative learning. The journal
of Abnormal and Social Psychology, 67, 601-607.
doi:10.1037!h0045550
Baranowski, T., Anderson, C.,&: Carmack, C. (1998).
Mediating variable framework in physical activ-
ity interventions: How are we doing? How might
we do better? American journal of Preventive
Medicine, 15,266-297. doi:l0.1016/S0749-
3797(98)00080-4
Baron, R. M., &: Kenny, D. A. (1986). The moderator-
mediator variable distinction in social psychologi-
cal research: Conceptual, strategic, and statistical
considerations. journal of Personality and Social
Psychology, 51,1173-1182. doi:10.1037/0022-3514.
51.6.1173
Beck, A. T., Rush, A.]., Shaw, B. F.,&: Emery, G. (1979).
Cognitive therapy of depression. New York, NY:
Guilford Press.
Bentler, P.M. (1997). EQS for Windows (Version 5.6).
Encino, CA: Multivariate Software.
Bishop, Y. M. M., Fienberg, S. E., & Holland, P. W.
(1975). Discrete multivariate analysis: Theory and
practice. Cambridge, MA: MIT Press.
Blalock, H. M. (1969). Theory construction: From verbal
to mathematical formulations. Englewood Cliffs, NJ:
Prentice-Hall.
Boker, S.M.,&: Nesselroade,J. R. (2002). A method for
modeling the intrinsic dynamics of intraindividual
variability: Recovering the parameters of simulated
oscillators in multi-wave panel data. Multivariate
328
Behavioral Research, 37, 127-160. doi:l0.l207/
S15327906MBR3701_06
Bollen, K. A.,&: Curran, P.j. (2004). Autoregressive
latent trajectory (ALT) models: A synthesis of two
traditions. Sociological Methods and Research, 32
336-383. doi:10.ll77/0049124103260222 '
Bollen, K. A.,&: Stine, R. A. (1990). Direct and indirect
effects: Classical and bootstrap estimates of vari-
ability. Sociological Methodology, 20, 115-140.
doi:10.2307/271084
Cheong,]. (2011). Accuracy of estimates and
statistical power for testing mediation in latent
growth modeling. Structural Equation Modeling 18
195-211. ' '
Cheong,]., MacKinnon, D.P.,&: Khoo, S. T. (2003).
Investigation of mediational processes using paral-
lel process latent growth curve modeling. Structural
Equation Modeling, 10, 238-262. doi:10.1207/
S15328007SEM1002_5
Cole, D. A.,&: Maxwell, S. E. (2003). Testing media-
tional models with longitudinal data: Questions
and tips in the use of structural equation model-
ing.]ournal of Abnormal Psychology, 112, 558-577.
doi: 10.1037/0021-843X.112.4 .558
Collins, L. M., Graham,]. W., &: Flaherty, B. P. (1998).
An alternative framework for defining mediation.
Multivariate Behavioral Research, 33, 295-312.
doi: 10.1207/s15327906mbr3302_5
Davis,]. A. (1985). The logic of causal order. Beverly Hills,
CA: Sage.
Dwyer,]. H. (1983). Statistical models for the social and
behavioral sciences. New York, NY: Oxford University
Press.
Efron, B.,&: Tibshirani, R.]. (1993). An introduction to
the bootstrap. Boca Raton, FL: Chapman &: Hall!CRC
Press.
Fairchild, A.].,&: MacKinnon, D.P. (2009). A general
model for testing mediation and moderation effects.
Prevention Science, 10, 87-99. doi:l0.1007/s11121-
008-0109-6
Fairchild, A.]., MacKinnon, D.P., Taborga, M.P.,&:
Taylor, A. B. (2009). R2 effect size measures for
mediation analysis. Behavior Research Methods 41
486-498. doi:10.3758/BRM.41.2.486 ' '
Ferrer, E., &McArdle,].]. (2003). Alternative struc-
tural models for multivariate longitudinal data
analysis. Structural Equation Modeling, 10, 493-524.
doi:10.1207/Sl5328007SEM1004_l
Fishbein, M., &: Ajzen, I. (1975). Belief, attitude, intention,
and behavior: An introduction to theory and research.
Reading, MA: Addison-Wesley.
Frangakis, C. E.,&: Rubin, D. B. (2002). Principal strati-
fication in causal inference. Biometrics, 58, 21-29.
doi: l0.111l!j.0006-341X.2002.00021.x

Fritz, M. S. (2007). An exponential decay model for media-
tion (Unpublished doctoral dissertation). Arizona
State University, Tempe.
Gollob, H. F., & Reichardt, C. S. (1991). Interpreting
and estimating indirect effects assuming time lags
really matter. In L. M. Collins & ]. L. Horn (Eds.),
Best methods for the analysis of change: Recent
advances, unanswered questions, future directions
(pp. 243-259). Washington, DC: American
Psychological Association. doi:10.1037/10099-015
Holland, P. W. (1988). Causal inference, path analy-
sis, and recursive structural equation models.
Sociological Methodology, 18,449-484. doi:10.2307/
271055
Hyman, H. H. (1955). Survey design and analysis:
Principles, cases, and procedures. Glencoe, IL: Free
Press.
James, L. R., & Brett,]. M. (1984). Mediators, mod-
erators, and tests for mediation. ]oumal of Applied
Psychology, 69,307-321. doi:10.1037/002l-9010.
69.2.307
James, L. R., Mulaik, S. A., & Brett,]. M. (2006). A tale
of two methods. Organizational Research Methods, 9,
233-244. doi:10.117711094428105285144
]o, B. (2008). Causal inference in randomized experi-
ments with mediational processes. Psychological
Methods, 13,314-336. doi:10.1037/a0014207
judd, C. M., & Kenny, D. A. (l981a). Estimating the
effects of social interventions. Cambridge, England:
Cambridge University Press.
Judd, C. M., & Kenny, D. A. (1981b). Process
analysis: Estimating mediation in treatment
evaluations. Evaluation Review, 5, 602-619.
doi:10.1177/0l93841X8100500502
Kazdin, A. E. (2009). Understanding how and why psy-
chotherapy leads to change. Psychotherapy Research,
19,418-428. doi:10.1080/l0503300802448899
Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data
analysis in social psychology. In D. I. Gilbert, S. I.
Fiske, & G. Lindzey (Eds.), The handbook of social
psychology (4th ed., Vol. 1, pp. 233-265). Boston,
MA: McGraw Hill.
Kraemer, H. C., Wilson, I., Fairburn, C. G., & Agras,
S. (2002). Mediators and moderators of treat-
ment effects in randomized clinical trials. Archives
of General Psychiatry, 59,877-883. doi:10.10011
archpsyc.59.10.877
Krull,]. L., & MacKinnon, D.P. (1999). Multilevel
mediation modeling in group-based interven-
tion studies. Evaluation Review, 23, 418-444.
doi:10.1177/0193841X9902300404
Krull,]. L., & MacKinnon, D.P. (2001). Multilevel mod-
eling of individual and group level mediated effects.
Statistical Mediation Analysis
Multivariate Behavioral Research, 36, 249-277.
doi: 10.1207/S1532 7906MBR3602_06
Lazarsfeld, P. F. (1955). Interpretation of statistical rela-
tions as a research operation. In P. F. Lazarsfeld &
M. Rosenberg (Eds.), The language of social research:
A reader in the methodology of social research
(pp. ll5-l25). Glencoe, IL: Free Press.
Lynch, K. G., Cary, M., Gallop, R., & Ten Have, I. R.
(2008). Causal mediation analyses for randomized tri-
als. Health Services and Outcomes Research Methodology,
8, 57-76. doi:10.1007/s10742-008-0028-9
MacCorquodale, K., & Meehl, P. E. (1948). On a distinc-
tion between hypothetical constructs and interven-
ing variables. Psychological Review, 55, 95-107.
doi: 10.103 7/h0056029
MacKinnon, D.P. (1994). Analysis of mediating variables
in prevention intervention studies. In A. Cazares &
L. A. Beatty (Eds.), Scientific methods for prevention
intervention research (pp.12 7-153). Washington,
DC: U.S. Department of Health and Human Services.
MacKinnon, D.P. (2008). Introduction to statistical media-
tion analysis. New York, NY: Erlbaum.
MacKinnon, D.P., & Dwyer,]. H. (1993). Estimation of
mediated effects in prevention studies. Evaluation
Review, 17,144-158. doi:10.1177/0193841X9
301700202
MacKinnon, D.P., Fairchild, A.j., & Fritz, M.S. (2007).
Mediation analysis. Annual Review of Psychology, 58,
593-614. doi:l0.1146/annurev.psych.58.110405.
085542
MacKinnon, D.P., Fritz, M.S., Williams,]., & Lockwood,
C. M. (2007). Distribution of the product confidence
limits for the indirect effect: Program PRODCLIN.
Behavior Research Methods, 39, 384-389. doi:lO.
3758/BF03193007
MacKinnon, D.P., Krull,]. L., & Lockwood, C. M.
(2000). Equivalence of the mediation, confounding,
and suppression effect. Prevention Science, 1,
173-181. doi:10.1023/A:1026595011371
MacKinnon, D.P., Lockwood, C. M., Brown, C. H.,
Wang, W., & Hoffman,]. M. (2007). The intermedi-
ate endpoint effect in logistic and probit regression.
Clinical Trials, 4, 499-513. doi:10.1177/174077
4507083434
MacKinnon, D.P., Lockwood, C. M., Hoffman,]. M.,
West, S. G., & Sheets, V. (2002). A comparison of
methods to test mediation and other intervening
variable effects. Psychological Methods, 7, 83-104.
doi: 10.103 7/1082-989X. 7.1.83
MacKinnon, D.P., Lockwood, C. M., & Williams,].
(2004). Confidence limits for the indirect effect:
Distribution of the product and resampling meth-
ods. Multivariate Behavioral Research, 39, 99-128.
doi: 10.1207/sl5327906mbr3901_4
329
MacKinnon, Cheong, and Pirlott
MacKinnon, D.P., Taborga, M.P.,&: Morgan-Lopez, A.
A. (2002). Mediation designs for tobacco prevention
research. Drug and Alcohol Dependence, 68, S69-S83.
doi: 10.10 16/S03 76-8716(02)00216-8
MacKinnon, D.P., Warsi, G.,&: Dwyer,]. H. (1995).
A simulation study of mediated effect measures.
Multivariate Behavioral Research, 30, 41-62.
doi: 10.1207/s15327906mbr3001_3
Manly, B. F.]. (1997). Randomization and Monte Carlo
methods in biology (2nd ed.). New York, NY:
Chapman &: Hall.
McArdle,].]. (2001). A latent difference score approach
to longitudinal dynamic structural analysis. In
R. Cudeck, S. du Toit, &: D. Sorbom (Eds.),
Structural equation modeling: Present and future. A
Festschrift in honor of Karl]oreskog (pp. 341-380).
Lincolnwood, IL: Scientific Software International.
McArdle,].].,&: Nesselroade,j. R. (2003). Growth curve
analysis in contemporary research. In]. A. Schinka
&: W. F. Velicer (Eds.), Handbook of psychology:
Vol. 2. Research methods in psychology (pp. 447-480).
New York, NY: Wiley.
McDonald, R. P. (1997). Haldane's lungs: A case study in
path analysis. Multivariate Behavioral Research, 32,
1-38. doi:10.1207/s15327906mbr3201_1
Morgan, S. L., &: Winship, C. (2007). Counterfactuals
and causal inference: Methods and principles for social
research. New York, NY: Cambridge University Press.
Muller, D.,judd, C. M., &: Yzerbyt, V. Y. (2005). When
moderation is mediated and mediation is moder-
ated. journal of Personality and Social Psychology, 89,
852-863. doi:10.1037/0022-3514.89.6.852
Muthen, B. 0., &: Curran, P.]. (1997). General longitudi-
nal modeling of individual differences in experimen-
tal designs: A latent variable framework for analysis
and power estimation. Psychological Methods, 2,
371-402. doi:10.1037/1082-989X.2.4.37l
Muthen, L. K., &: Muthen, B. 0. (1998-2007). Mplus
user's guide (5th ed.). Los Angeles, CA: Authors.
Pearl,]. (2009). Causality: Models, reasoning, and infer-
ence (2nd ed.). New York, NY: Cambridge University
Press.
Preacher, K.]., &: Hayes, A. F. (2004). SPSS and
SAS procedures for estimating indirect effects
in simple mediation models. Behavior Research
Methods, Instruments, and Computers, 36, 717-731.
doi: 10.3 758/BF03206553
Preacher, K. ]., Zyphur, M. ]., &: Zhang, Z. (2010). A gen-
eral multilevel SEM framework for assessing multi-
level mediation. Psychological Methods, 15, 209-233.
Robins,]. M. (1994). Correcting for non-compliance in
randomized trials using structural nested mean mod-
els. Communications in Statistics Theory and Methods,
23,2379-2412. doi:10.1080/03610929408831393
330
Robins,]. M., &: Greenland, S. (1992). Identifiability
and exchangeability for direct and indirect effects.
Epidemiology, 3, 143-155. doi:10.1097/00001648-
199203000-000 13
Rogosa, D. R. (1988). Myths about longitudinal research.
InK. W. Schaie, R. T. Campbell, W. M. Meredith,&:
S.C. Rawlings (Eds.), Methodological issues in aging
research (pp. 171-209). New York, NY: Springer.
Sheets, V. L., &: Braver, S. L. (1999). Organizational
status and perceived sexual harassment: Detecting
the mediators of a null effect. Personality and Social
Psychology Bulletin, 25,1159-1171. doi:10.1177/
01461672992512009
Shrout, P. E.,&: Bolger, N. (2002). Mediation in experi-
mental and nonexperimental studies: New proce-
dures and recommendations. Psychological Methods,
7, 422-445. doi:10.1037/1082-989X.7.4.422
Singer,]. D.,&: Willett,]. B. (2003). Applied longitudinal
data analysis: Modeling change and event occurrence.
London, England: Oxford University Press.
Sobel, M. E. (1982). Asymptotic confidence inter-
vals for indirect effects in structural equation
models. Sociological Methodology, 13, 290-312.
doi:10.2307/270723
Sobel, M. E. (1986). Some new results on indirect effects
and their standard errors in covariance structure
models. Sociological Methodology, 16, 159-186.
doi:10.2307/270922
Sobel, M. E. (2007). Identification of causal parameters in
randomized studies with mediating variables. Journal
of Educational and Behavioral Statistics, 33, 230-251.
doi:10.3102/1076998607307239
Spencer, S.j., Zanna, M.P.,&: Fong, G. T. (2005).
Establishing a causal chain: Why experiments are
often more effective than mediational analyses
in examining psychological processes. journal of
Personality and Social Psychology, 89, 845-851.
doi: 10.1037/0022-3514.89.6.845
Spirtes, P., Glymour, C.,&: Scheines, R. (1993).
Causation, prediction, and search. New York, NY:
Springer-Verlag.
Springer, M.D. (1979). The algebra of random variables.
New York, NY: Wiley.
Ten Have, T. R.,joffe, M. M., Lynch, K. G., Brown, G. K.,
Maisto, S. A.,&: Beck, A. T. (2007). Causal mediation
analyses with rank preserving models. Biometrics, 63,
926-934. doi:10.1111/j.1541-0420.2007.00766.x
von Eye, A., Mun, E. Y., &: Mair, P. (2009). What carries
a mediation process? Configura! analysis of media-
tion. Integrative Psychological and Behavioral Science,
43,228-247. doi:10.1007/sl2124-009-9088-9
Weiss, C. H. (1997). How can theory-based evaluation
make greater headway? Evaluation Review, 21,
501-524. doi:10.1177/0193841X9702100405
 Statistical Mediation Analysis

West, 5. G., & Aiken, L. 5. (1997). Toward understand-
ing individual effects in multicomponent prevention
programs: Design and analysis strategies. In K. Bryant,
M. Windle, & 5. West (Eds.), The science of prevention:
Methodological advances from alcohol and substance abuse
research (pp. 167-209). Washington, DC: American
Psychological Association. doi: 10.103 7/10222-006
Winship, C., & Mare, R. D. (1983). Structural
equations and path analysis for discrete data.
American]oumal of Sociology, 89, 54-llO.
doi: 10.1086/227834
Woodworth, R. 5. (1928). Dynamic psychology. In
C. Murchison (Ed.), Psychologies of 1925 (pp. 111-126).
Worchester, MA: Clark University Press.
Yuan, Y., & MacKinnon, D.P. (2009). Bayesian media-
tion analysis. Psychological Methods, 14, 301-322.
doi: 10.103 7/aOO 16972

I
't
I

I
u 331

L View publication stats