Psychological Methods Copyright 2006 by the American Psychological Association

2006, Vol. 11, No. 2, 142–163 1082-989X/06/$12.00 DOI: 10.1037/1082-989X.11.2.142

Conceptualizing and Testing Random Indirect Effects and Moderated

Mediation in Multilevel Models: New Procedures
and Recommendations
Daniel J. Bauer, Kristopher J. Preacher, and Karen M. Gil
University of North Carolina at Chapel Hill

The authors propose new procedures for evaluating direct, indirect, and total effects in
multilevel models when all relevant variables are measured at Level 1 and all effects are
random. Formulas are provided for the mean and variance of the indirect and total effects and
for the sampling variances of the average indirect and total effects. Simulations show that the
estimates are unbiased under most conditions. Confidence intervals based on a normal
approximation or a simulated sampling distribution perform well when the random effects are
normally distributed but less so when they are nonnormally distributed. These methods are
further developed to address hypotheses of moderated mediation in the multilevel context. An
example demonstrates the feasibility and usefulness of the proposed methods.

Keywords: multilevel model, hierarchical linear model, indirect effect, mediation, moderated
mediation

In psychology and other social sciences, hypotheses often diagram, the product of the paths labeled a and b represents
concern the causal pathways through which key predictors the indirect effect of X on Y, the path labeled c⬘ represents
transmit their effects to specific outcomes. For example, the direct effect of X on Y, and c ⫽ ab ⫹ c⬘ is the total effect
Wei, Mallinckrodt, Russell, and Abraham (2004) examined of X on Y (Alwin & Hauser, 1975; Bollen, 1987, 1989).
the extent to which attachment anxiety and avoidance were Depending on the pattern of these effects, the variable M
related to maladaptive perfectionism and, in turn, depressive may be called a mediator, a suppressor, or simply an inter-
mood. Similarly, Catanzaro and Laurent (2004) hypothe- vening variable (MacKinnon, Krull, & Lockwood, 2000).
sized that alcohol expectancies play a role in drinking as a We generally use the terms mediator and mediation to be
means of coping, which then predicts a number of drinking consistent with the focus of much of the literature on indi-
behaviors. In both of these cases, the causal effects of the rect effects, although our exposition also pertains to other
original predictor are transmitted at least partially through patterns of indirect effects.
an intervening variable, as diagrammed in Figure 1. In the Within the context of linear regression and path analysis,
a number of methods have been proposed for evaluating
mediation, and this remains an active area of research (e.g.,
Daniel J. Bauer, Kristopher J. Preacher, and Karen M. Gil, MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002;
Department of Psychology, University of North Carolina at Chapel MacKinnon, Lockwood, & Williams, 2004; Shrout &
Hill. Bolger, 2002). Such methods are, however, inappropriate if
Additional materials are on the Web at http://dx.doi.org/ the analyzed data are actually hierarchical in nature (i.e.,
10.1037/1082-989X.11.2.142.supp composed of two or more nested levels). Two forms of
This work was funded in part by National Institute on Drug hierarchical data are common in psychological research.
Abuse Grant F32 DA16883 awarded to Kristopher J. Preacher. In First, individuals may be assessed from a number of groups.
addition, the study of sickle cell disease that formed the basis of Individuals are then said to be nested within groups. To the
the applied analyses was supported by National Institutes of Health extent that individuals within a group share common expe-
Grant R01 HL6217 awarded to Karen M. Gil, principal investiga-
riences, we would expect their scores on the outcome vari-
tor. We also thank Greg Stonerock for assisting in the preparation
of the sickle cell disease data and Li Cai for assisting with the
able to be correlated across members of the group, which
technical appendix. violates the independence assumption of many statistical
Correspondence concerning this article should be addressed to models. Second, repeated observations may be made on the
Daniel J. Bauer, Department of Psychology, Davie Hall 337-A, same individuals. In this case, the repeated measurements
University of North Carolina at Chapel Hill, Chapel Hill, NC are said to be nested within individuals. Repeated measure-
27599-3270. E-mail: [email protected] ments are typically correlated within persons, which again
142
 RANDOM INDIRECT EFFECTS 143

compromises the independence assumption. For both types of

hierarchical data, we refer to the lower level as Level 1 and the
upper level as Level 2. More than two levels of data are
possible, but we restrict our attention to two-level models.
Because the independence assumption is violated for
these data structures, multiple linear regression and path
analysis will produce biased tests of the effects in the model
(Hox, 2002; Kreft & de Leeuw, 1998; Raudenbush & Bryk,
2002). One way to appropriately model such data is to use
a multilevel model, also known as a hierarchical linear
model or a mixed-effects model. A key advantage of the
multilevel model is that it captures the correlations among
the Level 1 observations through the estimation of random
effects. These random effects can take the form of random
intercepts, reflecting differences in the overall level of the
outcome variable across Level 2 units; random slopes, re-
flecting differences in the effects of predictors across Level
2 units; or both. For instance, a multilevel model with a
random intercept produces a compound symmetric correla-
tion structure for the Level 1 observations. In addition to the
statistical advantages of multilevel models for correlated
data, random effects can have quite interesting substantive
interpretations. A random slope, for example, indicates that
the causal effect of the predictor on the outcome differs over
Level 2 units. This may then initiate the search for potential
moderators of the effect.
Another important feature of hierarchical data is that
predictors can reside at different levels of the data (e.g.,
individual- vs. group-level characteristics). Given this, me-
diation in multilevel models may take several forms, as Figure 2. Upper and lower level mediation in a two-level model:
shown in Figure 2 (Kenny, Kashy, & Bolger, 1998; Krull & Nested frames indicate levels of sampling, boxes indicate vari-
MacKinnon, 2001). Upper level mediation exists when the ables, arrows without circles represent fixed effects, arrows with
effect of a Level 2 predictor on a Level 1 outcome is circles represent random effects, and arrows not originating from
mediated by another Level 2 predictor (2 3 2 3 1 medi- variables indicate residuals.
ation). Lower level mediation exists when the mediator is a
Level 1 variable. In some cases of lower level mediation the
effect of a Level 2 predictor is mediated (2 3 1 3 1 Raudenbush and Sampson (1999) and Bauer (2003). Krull
mediation), and in other cases the effect of a lower level and MacKinnon (1999, 2001) offered an alternative method
predictor is mediated (1 3 1 3 1 mediation).1 Methods for applicable to all three types of multilevel mediation that is
assessing upper level mediation have been proposed by similar to the causal steps approach of Baron and Kenny
(1986) for ordinary regression models. More recently,
Pituch, Wittaker, and Stapleton (2005) examined several
approaches for testing indirect effects in 2 3 2 3 1 and 2 3
1 3 1 models for cluster-randomized treatment designs.
What all of these approaches hold in common is an
assumption that the causal effects are fixed (not random),

1
As a reviewer pointed out, one could also consider other forms
Figure 1. Diagram of a simple mediation model. Variables are of mediation for multilevel data. For instance, one might hypoth-
indicated in boxes: X is the primary predictor, M is the mediator, esize that an individual-level predictor affects a group-level me-
and Y is the distal outcome variable. Arrows originating from diator, which, in turn, affects a group- or individual-level outcome.
variables indicate hypothesized causal effects. Labels for these In this article, we consider only models in which the causal effects
effects are indicated next to the arrows. Arrows not originating reside at the same level or are transmitted from a higher to a lower
from variables indicate residuals. level variable.
144 BAUER, PREACHER, AND GIL

meaning that the magnitude of the effects is equal for all The terms eMij and eYij are Level 1 residuals for M and Y,
Level 2 units. In lower level mediation, however, the causal respectively. The other five terms are interpreted similarly
effects can be random because some predictors reside at to the intercepts and slopes of a standard regression model,
Level 1. In 2 3 1 3 1 mediation, the effect of the mediator with the caveat that each coefficient is random, meaning that
on the outcome may be random; in 1 3 1 3 1 mediation, the value of the coefficient varies across Level 2 units (as
all three causal effects can be random. These random effects indicated by the j subscript). That is, the intercepts for M
are indicated in Figure 2 as circles on the causal paths. They and Y are designated dMj and dYj, respectively, the effect of
represent heterogeneity in the causal effects across Level 2 X on M is designated aj, the effect of M on Y is designated
units and may be of considerable substantive interest. For bj, and the direct effect of X on Y is designated c⬘j . The
instance, in the empirical analysis to be presented later, we random effects of the model permit heterogeneity in the
analyze daily diary data to assess whether emotional reac- causal effects. For instance, if Level 1 represents repeated
tions to pain mediate the relation between physical pain and measures, indexed by i, and Level 2 represents persons,
stress in patients with sickle cell disease (SCD) and whether indexed by j, then this model can be used to assess how the
the strength of mediation differs across participants. strength of the hypothesized causal relations among X, M,
Kenny, Korchmaros, and Bolger (2003) first called atten- and Y vary across individuals. For simplicity, we assume for
tion to the importance of random effects in lower level the time being that there are no Level 2 predictors of these
mediation models, in particular in 1 3 1 3 1 models. random effects.
Furthermore, they proposed a method for fitting 1 3 1 3 1 A number of assumptions are required to make the model
models when all three causal pathways are random effects, estimable and ensure that the effects are unbiased. These
as depicted in the lower panel of Figure 2. The purpose of assumptions are inclusive of those that apply to all multi-
the present article is to expand on this prior research to level models (as conventionally estimated) but include
improve the conceptualization and estimation of lower level features that are unique to multivariate multilevel models,
mediation with random effects in multilevel models. We such as the lower level mediation model in Equation 1.
begin by showing how one can estimate the 1 3 1 3 1 These unique features are italicized in the following list of
model simultaneously using conventional multilevel or assumptions:2
mixed modeling software. We then clarify that, in the pres- 1. The predictors are uncorrelated with the random effects
ence of random indirect effects, one may be interested in (intercepts and slopes) and residuals, both within and across
two related but different questions. First, is there heteroge- equations (e.g., Xij must be uncorrelated with dMj, aj, and
neity in the strength of the indirect, direct, and total effects eMj and uncorrelated with dYj , bj , c⬘j , and eYij).
across units of the population? Second, how precisely can 2. The residuals eMj and eYij are each normally distributed
we estimate the average effects in the population? After with an expected value of zero, and they are uncorrelated
distinguishing these questions, we proceed to a small sim- with one another. Typically, the residuals for each outcome
ulation study designed to evaluate the quality of the average are assumed to be independent and homoscedastic across i
effect estimates and their confidence intervals (CIs). Next, within j, but these restrictions can and should be relaxed in
we extend the model to include predictors of heterogeneity certain circumstances (e.g., when residuals are expected to
in the indirect and direct effects. Last, we demonstrate the autocorrelate with repeated measures). The additional as-
value of estimating and predicting random indirect effects in sumption that the residuals are uncorrelated across out-
1 3 1 3 1 models with an empirical example. Our focus comes is required to identify the effect of M on Y.
throughout is on the estimation and testing of the indirect, 3. The random effects are normally distributed with
direct, and total effects of the distal predictor on the out- means equal to the average effects in the population. Al-
come, not on the broader causal steps approach to mediation though other assumptions are possible, we also assume that
outlined by Baron and Kenny (1986).
2
Another assumption that is sometimes listed for these models
Lower Level Mediation Models With Random is that the predictors are fixed. This assumption seems to present a
Indirect Effects problem for the model in Equation 1 because M appears in the
second equation as a fixed predictor but also appears in the first
Loosely following the notation of Kenny et al. (2003), equation as a random outcome variable. Like ordinary regression,
however, the assumption that the predictors are fixed is unneces-
one can write the lower level mediation model depicted in
sary as long as the predictors are uncorrelated with the errors, as
the lower panel of Figure 2 with two Level 1 equations as
we note in Assumption 1 (Demidenko, 2004, p. 143). We can thus
consider M to be a random variable in both equations. Conse-
M ij ⫽ dMj ⫹ ajXij ⫹ eMij quently, we must now assume a distribution for M, namely that it
is conditionally normal (conditional on the random effects and X),
Y ij ⫽ dYj ⫹ bjMij ⫹ c⬘j Xij ⫹ eYij . (1) as indicated in Assumption 5.
 RANDOM INDIRECT EFFECTS 145

all of the random effects covary with one another. (Note that regression models for M and Y separately for each unit j,
the random effects of a multilevel model are often expressed collecting the slope estimates, and then calculating the sam-
in deviation form, with the means subtracted, so that the ple covariance of the slope estimates. They acknowledged,
expected values of the random effects are zero.) however, that this was not an optimal strategy. A second
4. The Level 1 residuals are uncorrelated with the random limitation of the two-step approach is that the asymptotic
effects both within and across equations (e.g., eMj is uncor- covariance matrix of the model estimates is similarly in-
related with dMj, aj, dYj, bj, and c⬘j ). complete. This matrix captures the sampling variability in
5. Assumptions 2 and 3 imply that the distributions of M the estimates, and some of its missing elements are needed
and Y are conditionally normal. That is, the distribution of to capture the standard errors of the average indirect and
M is normal conditional on X and the distribution of Y is total effects in the model.
normal conditional on M and X.3 The solution that we propose is to formulate the model
We show that Assumption 3 is particularly important for with a single Level 1 equation through the use of selection
deriving estimates of the indirect effects. We can write (or indicator) variables. Researchers have used this strategy
Assumption 3 more formally as in the past to aid in the estimation of other kinds of multi-
variate multilevel models (e.g., MacCallum, Kim, Malar-

冢冣
d Mj key, & Kiecolt-Glaser, 1997), and it is equally useful in this
aj context. The basic idea is to form a new outcome variable—
dYj for instance, Z— by stacking Y and M for each unit i within
bj j. This single outcome variable allows us to fit a “multivar-
c⬘j iate” model using univariate multilevel modeling software.
To distinguish the two variables stacked in Z, we also create

冤冢 冣 冢 冣冥
dM ␴d2Mj two selection variables—for instance, SM and SY. The vari-
a ␴d Mj,a j ␴a2 j able SM is set equal to 1 when Z refers to M and is 0
⬃ N dY , ␴dMj,dYj ␴aj,dYj ␴d2Yj . (2) otherwise. Similarly, the variable SY is set equal to 1 when
b ␴dMj,bj ␴aj,bj ␴dYj,bj ␴b2j Z refers to Y and is 0 otherwise. We retain the variables X
c⬘ ␴dMj,c⬘j ␴aj,c⬘j ␴d Yj,c⬘j ␴b j,c⬘j ␴c⬘
2 and M in the new data set, as they are needed as predictors
j
of Z. An example of this data rearrangement is presented in
The means of the random effects are designated with the Figure 3.
same symbols as the random effects, absent the subscript j The purpose of rearranging the data in this fashion is that
to indicate they are constant over Level 2 units (e.g., a is the we can now specify the lower level mediation model with a
mean of aj). These are the fixed effects of the model. The single equation:
covariance matrix for the random effects captures the non-
independence of the observations collected at Level 1 and, Z ij ⫽ SMij共dMj ⫹ ajXij兲 ⫹ SYij共dYj ⫹ bjMij ⫹ c⬘j Xij兲 ⫹ eZij. (3)
together with the variances of the Level 1 residuals, consti-
tutes the variance components of the model. The diagonal Notice that the two selection variables essentially toggle
elements of this matrix, the variances, characterize hetero- from the model for M to the model for Y in Equation 3. For
geneity in the causal effects of the model across Level 2 instance, when Z is a value of M, then SM ⫽ 1 and SY ⫽ 0,
units. For instance, ␴c⬘ 2
j
is the variance of the direct effect and Equation 3 simplifies to Zij ⫽ dMj ⫹ aj Xij ⫹ eZij . This
共c⬘j ), indicating the extent to which the strength of the direct is the model for M from Equation 1. Similarly, when Z is a
effect differs over Level 2 units. The off-diagonal elements value of Y, then SY ⫽ 1 and SM ⫽ 0, and Equation 3
of the matrix capture the covariance among the random simplifies to Zij ⫽ dYj ⫹ bj Mij ⫹ c⬘j Xij ⫹ eZij , or the model
effects. For instance, a positive value for ␴aj ,bj would indi- for Y from Equation 1. Thus, although we have stacked Y
cate that for Level 2 units for which the effect of X on M (aj) and M into a single dependent variable for the purpose of
is high, the effect of M on Y (bj) is often also high.
To examine lower level mediation, Kenny et al. (2003)
3
estimated each Level 1 model in Equation 1 separately. A reviewer correctly noted that the unconditional joint distri-
Unfortunately, this two-step approach sacrifices some im- bution of M and Y would not be normal, given the contribution of
the ajbj product to Y. However, it is important to note that the
portant information. First, one cannot estimate directly the
conditional distribution of M given X and Y given M and X can still
covariance of random effects residing in different Level 1 be normal and that it is this conditional distribution for which
models, including all of the elements enclosed in the dotted normality is assumed. Although this assumption is common to all
square in Equation 2. As we show, this makes it difficult to linear multilevel models, it may often be unrealistic in practice.
capture the dispersion of the indirect and total effects of the Therefore, in our simulation study, we consider the consequences
model. Kenny et al. (2003) proposed an ad hoc solution to of violating this assumption when the random effects are nonnor-
the problem that involved estimating ordinary least squares mally distributed.
146 BAUER, PREACHER, AND GIL

article). Generic syntax for rearranging the data and fitting

the model in Equation 4 in SAS is provided on the Web
at http://dx.doi.org/10.1037/1082-989X.11.2.142.supp, al-
though other programs, such as SPSS, HLM, or MLwiN,
could also be used to fit the model.
Note that this specification strategy generalizes to models
with more than one mediator through the creation of addi-
tional selection variables. The key is to “trick” the software
into estimating a multivariate system of equations through
the creation of a single outcome variable Z. Given this
set-up, one can estimate the entire model simultaneously
using conventional software for univariate multilevel mod-
els. The resulting output includes the full covariance matrix
of the random effects and the full asymptotic covariance
matrix of the fixed effects estimates and variance compo-
nents, both of which are critical for our next developments.
To summarize to this point, we note that by using the
selection variable approach, we can estimate the complete
lower level mediation model simultaneously, providing all
of the necessary information for evaluating the hypothe-
sized causal effects of the model. We now turn to methods
for quantifying and testing variability in these causal effects.

Figure 3. In the top diagram, data are shown for two upper level
units (j ⫽ 1 and j ⫽ 2) for an outcome (Y), a mediator (M), and a Evaluating Random Indirect and Total Effects in
predictor (X) whose effect on Y is thought to be mediated by M. Multilevel Models
This data set is then rearranged into the lower diagram through the
creation of a new variable Z that represents Y whenever the In this section, we consider two related questions: Given
selection variable SY ⫽ 1 (and SM ⫽ 0) and represents M whenever the presence of random effects in a lower level mediation
the selection variable SM ⫽ 1 (and SY ⫽ 0). For each observation model, how does one quantify variability in the indirect and
in the top diagram, there are then two rows in the bottom diagram. total effects across Level 2 units of the population, and how
The values of M and X are repeated for these two rows. does one obtain CIs for the average indirect and total
effects? Although these are related questions, they are quite
different in purpose. For the first question, we are interested
fitting the model, the two outcomes continue to be distin- in variability across Level 2 units (i.e., heterogeneity in
guished in the model equations by the selection variables. causal effects). For the second question, we are interested in
To aid in specifying the model, it is helpful to distribute variability in our estimates of the average effects across
the selection variables in Equation 3 as follows: Level 2 units (i.e., the precision of the average causal
effects). This distinction is analogous to the difference be-
Z ij ⫽ dMjSMij ⫹ aj共SMijXij兲 ⫹ dYjSYij
tween the standard deviation of a variable and the standard
⫹ bj共SYijMij兲 ⫹ c⬘j (SYijXij) ⫹ eZij . (4) error of its mean. We now present the technical details for
evaluating each question. Additional details on the deriva-
Equation 4 shows that one would specify a model for Z with tion of the results presented in this section are provided in
no intercept but with random effects for SM and SY (dMj and the Appendix. Readers who wish to see only the conclusions
dYj, respectively) and with random effects for the product of these developments can skip to the Summary of Proce-
variables SMX, SYM, and SYX (aj, bj, and c⬘j , respectively). In dures subsection.
addition, one must use some method to allow the residual
variance Var(eZij) to differ depending on SM (or, equiva-
lently, SY). This represents a form of heteroscedasticity
Investigating Heterogeneity in Causal Effects
because the residual variance for Z is then conditional on The first question concerns how to characterize heteroge-
SM. Fortunately, most multilevel modeling software pro- neity in the causal effects over Level 2 units. Let us consider
grams offer one or more options for modeling heterosce- the indirect effect first. The indirect effect for a given unit j
dasticity. Depending on the substantive context, more com- is ajbj. As Kenny et al. (2003) noted, because aj and bj are
plicated Level 1 residual variance (or covariance) structures not necessarily independent, the expected value or average
can also be entertained (see the example we provide later in the of ajbj is (Goodman, 1960, p. 712):
 RANDOM INDIRECT EFFECTS 147

E共ajbj兲 ⫽ ab ⫹ ␴aj ,bj . (5) the precision of the estimated average indirect effect, âb̂ ⫹
␴ˆ aj ,bj , and estimated average total effect, âb̂ ⫹ ␴ˆ aj ,bj ⫹ ĉ⬘j ,
As such, the average indirect effect in the population is a then we can calculate the sampling variances of these esti-
function of the average effect of X on M (or a), the average mates as
effect of M on Y (or b), and the covariance between the two
random effects (or ␴aj ,bj ). One interesting implication of this Var共âb̂ ⫹ ␴ˆ aj ,bj 兲 ⫽ b2 Var共â兲 ⫹ a2 Var(b̂)
formula is that even if a and b are zero, the average indirect
effect can be nonzero through the contribution of ␴aj ,bj . We ⫹ Var共â兲Var共b̂兲 ⫹ 2abCov共â, b̂兲
note that this formula does not require any assumptions on ⫹ Cov共â, b̂兲2 ⫹ Var共␴ˆ aj ,bj 兲 (9)
the distributions of the random effects (Kendall & Stuart,
1969, p. 283). and
Making the assumption that aj and bj are normally dis-
tributed, Kenny et al. (2003) showed that the variance of
Var共âb̂ ⫹ ␴ˆ aj ,bj ⫹ ĉ⬘兲 ⫽ b2 Var共â兲 ⫹ a2 Var共b̂兲
aj bj is
⫹ Var共â兲Var共b̂兲 ⫹ 2abCov共â, b̂兲 ⫹ Cov共â, b̂兲2 ⫹ Var共ĉ⬘兲
Var共ajbj兲 ⫽ b2 ␴a2j ⫹ a2 ␴b2j ⫹ ␴a2j␴b2j ⫹ 2ab␴aj ,bj ⫹ ␴a2j ,bj. (6)
⫹ 2bCov共â, ĉ⬘兲 ⫹ 2aCov共b̂, ĉ⬘兲 ⫹ Var 共␴ˆ aj ,bj 兲. (10)
This variance quantifies heterogeneity in the strength of the
indirect effect at Level 2. This value may be of considerable The variances and covariances in these equations (desig-
interest, as it indicates the extent to which the indirect effect nated Var and Cov) represent the asymptotic sampling
of X on Y varies across the Level 2 sampling units. Simi- variances and covariances of the fixed effect estimates â, b̂,
larly, Kenny et al. (2003) noted that the average total effect, and ĉ⬘ and the covariance estimate ␴ˆ aj ,bj .
aj bj ⫹c⬘j , can be expressed as In practice, one must replace the population values in
Equations 9 and 10 with their sample estimates (the model
E共aj bj ⫹ c⬘j 兲 ⫽ ab ⫹ ␴aj ,bj ⫹ c⬘. (7) estimates and estimated asymptotic variances and covari-
ances of the estimates) to obtain the estimated sampling
Again with the assumption of normality of the random
variances of the average direct and indirect effects. Most
effects, the variance of the random total effect can be
multilevel or mixed modeling software programs provide an
expressed as
option to output the estimated variance– covariance matrix
Var共aj bj ⫹ c⬘j 兲 ⫽ b2 ␴a2j ⫹ a2 ␴b2j ⫹ ␴a2j ␴b2j ⫹ 2ab␴aj ,bj of the model estimates. Note that only by estimating the full
model simultaneously can one estimate covariances be-
⫹ ␴a2j ,bj ⫹ ␴c⬘
2
j
⫹ 2b␴aj ,c⬘j ⫹ 2a␴bj ,c⬘j . (8) tween fixed effects residing in different equations—for ex-
ample, Cov(â, b̂) and Cov(â, ĉ⬘).
This variance, in turn, quantifies heterogeneity at Level 2 in To make inferences concerning the average indirect and
the strength of the total effect. total effects, we can form CIs for the estimates. One method
One nice property of maximum likelihood estimation, the for constructing CIs assumes normality for the sampling
typical method for fitting multilevel models, is that any distributions of the estimates. Under this assumption, 95%
function of maximum likelihood estimates (MLEs) is itself CIs for the average indirect effect and average total effect
an MLE (Raudenbush & Bryk, 2002, p. 52). As such, we are obtained as
can obtain MLEs for the quantities in Equations 5 through
8 by inserting the estimates from the fitted model in place of ˆ共âb̂ ⫹ ␴ˆ a ,b 兲兴1/ 2
their population values. The simultaneous modeling ap- 共âb̂ ⫹ ␴ˆ aj ,bj 兲 ⫾ 1.96 关Var j j
(11)
proach we have described generates all of these estimates
from a single model. Of particular interest may be the and
estimated average indirect and total effects, which we can
obtain by inserting the model estimates into Equations 5 and ˆ共âb̂ ⫹ ␴ˆ a ,b ⫹ ĉ⬘兲兴1/ 2 ,
共âb̂ ⫹ ␴ˆ aj ,bj ⫹ ĉ⬘兲 ⫾ 1.96 关Var (12)
j j

7. Our second question is concerned with the precision of

these estimated average indirect and total effects. where ⫾1.96 is the critical value of the z distribution and
ˆ
Var is used to indicate the estimated sampling variance
Quantifying the Precision of the Estimated Average obtained when the sample-based estimates are inserted into
Causal Effects
We begin by emphasizing that Equations 6 and 8 capture 4
Kenny et al. (2003) sometimes interpreted these equations to
heterogeneity in the indirect effects and total effects in the give the sampling variances of the estimated average indirect and
population of Level 2 units.4 If we are instead interested in total effects.
148 BAUER, PREACHER, AND GIL

Equation 9 or 10. Alternatively, one could perform a null Equations 9 and 10. One can then use these standard errors
hypothesis test by forming the critical ratio of each estimate to form CIs (or significance tests) by assuming that the
to its standard error and comparing the result with the estimates of the average indirect and total effects have
critical value of the z distribution. In either case, the as- normal sampling distributions, as shown in Equations 11
sumption of normality for the sampling distributions will and 12. A second approach is to simulate the sampling
not hold exactly, given that âb̂ is a product of normally distributions of the average indirect and total effects. This
distributed estimates (and hence will not be normal). The second approach is more difficult to implement, but it has
deviation from normality may be small enough, however, the advantage that the simulated sampling distributions of
that the CIs or significance tests will still be reasonably the average indirect and total effect estimates will have the
accurate. theoretically correct (nonnormal) forms. Both tests are im-
An alternative method for constructing CIs that may hold plemented in a SAS macro available on the Web at http://
promise is the Monte Carlo (MC) method of MacKinnon et dx.doi.org/10.1037/1082-989X.11.2.142.supp. Given the
al. (2004). In this approach, the sampling distribution for the trade-offs involved, we now turn to a comparison of the
effect of interest is not assumed to be normal and is instead performance of these two alternative procedures for making
simulated from the model estimates and their asymptotic inferences about the average indirect and total effects.
variances and covariances (a form of parametric bootstrap-
ping). For instance, to simulate the sampling distribution of Performance With Simulated Data
the average indirect effect, one would define a multinormal
distribution with means equal to â, b̂, and ␴ˆ aj ,bj and covari- As a preliminary investigation of the accuracy of the
ance matrix equal to the estimated covariance matrix of estimates and CIs just described, we conducted a modest
these estimates. One would then take random values from simulation study based on the example presented in Kenny
this multinormal distribution and plug them into Equation 5 et al. (2003, pp. 122–123). In their example, the random
to compute the average indirect effect. Collecting the results intercept for the M equation, dMj, had a mean of 0 and a
over many draws provides a simulated sampling distribution variance of .6, and the random intercept for the Y equation,
for the average indirect effect. One would then obtain con- dYj, had a mean of zero and a variance of .4. These two
fidence limits for the average indirect effect by taking the random effects were normally distributed and were uncor-
corresponding percentiles of this simulated sampling distri- related with each other and with the other random effects in
bution (e.g., 2.5th and 97.5th for a 95% CI). The advantage the model. For the M equation, the Level 1 residual variance
of this approach is that it does not assume normality for the was set to ␴e2M ⫽ .65; for the Y equation, it was set to
sampling distributions of the average indirect and total ␴e2Y ⫽ .45. The causal paths were simulated as follows: aj
effects. and bj were both normally distributed with means of a ⫽
b ⫽ .6 and variances of ␴a2j ⫽ ␴b2j ⫽ .16 and c⬘j was
normally distributed with a mean of c⬘ ⫽ .2 and a variance
Summary of Procedures
of ␴c⬘2 j ⫽ .04. The covariance between aj and bj was ␴aj ,bj
In conclusion, when one estimates lower level mediation ⫽ .113, yielding a correlation of .706. Neither aj nor bj was
models with random effects, interest may center on both the correlated with c⬘j . Last, the predictor X was simulated from
average causal effects and potential heterogeneity in these the equation Xij ⫽ X៮ j ⫹ eXij , where X៮ j ⬃ N 共0, 1兲 and
causal effects across the Level 2 units of the population. The eXij ⬃ N 共0, 1兲.
average indirect and total effects can be computed via To generate our own simulated data, we used this model
Equations 5 and 7. Estimates obtained from these equations as a base from which we varied four design factors. In
indicate the strength of the indirect and total effects for an varying these factors, we focused specifically on how they
average Level 2 unit in the population. Allowing the indirect might influence the average indirect effect estimate, as this
and total effects to be random, however, also implies that effect is typically of most interest (results differed little for
there is variability about these average values across Level the average total effect). We chose the first two factors of
2 units. This heterogeneity in the strength of the indirect and the simulation study to manipulate the effect size of the
total effects in the population can be characterized by the average indirect effect as expressed in Equation 5. First, the
variance formulas given in Equations 6 and 8. Therefore, for magnitude of the a and b parameters was set to either a ⫽
instance, the variance obtained from Equation 6 indicates b ⫽ .3 or a ⫽ b ⫽ .6. The higher values for a and b are
the extent to which the magnitude of the indirect effect identical to those used by Kenny et al. (2003), whereas the
differs across Level 2 units. lower values are close to the smaller effect sizes considered
We have also suggested two possible ways to make by Krull and MacKinnon (2001) for similar models. The
inferences for the average indirect and total effect estimates. second design factor was the magnitude of the covariance
In the first approach, the standard errors of the average between aj and bj. This parameter was set to one of three
indirect and total effects are computed as the square roots of values, ␴aj ,bj ⫽ ⫺ .113, 0, or .113 (correlations of ⫺.706,
 RANDOM INDIRECT EFFECTS 149

.000, and .706, respectively). The next two design factors replications resulted in nonconvergence and more replica-
manipulated the total sample size. We first set the number of tions produced unconstrained solutions. The two sample
Level 2 units to N ⫽ 25, 50, 100, or 200. We then set the size factors did not have an equal influence, however. Given
number of observations per Level 2 unit to nj ⫽ 4, 8, 16, or a finite number of total observations, the model becomes
32. These sample sizes are consistent with those used by more stable if there are more observations per Level 2 unit
Krull and MacKinnon (2001). The final design factor con- at the expense of the number of Level 2 units. For instance,
cerned the distributions of the random effects aj and bj. To with 800 possible observations, setting nj ⫽ 16 and N ⫽ 50
evaluate the robustness of the estimates, we simulated aj and resulted in 92% unconstrained solutions, setting nj ⫽ 8 and
bj either from normal distributions or from ␹2(3) distribu- N ⫽ 100 resulted in 87% unconstrained solutions, and
tions, having skew 1.63 and kurtosis 4. setting nj ⫽ 4 and N ⫽ 200 resulted in 78% unconstrained
Together, these four factors combined to yield 144 con- solutions. In addition to these results, small effects of the
ditions in a fully factorial design. In addition, to investigate other factors were also observed: The models were more
Type I errors, we simulated supplemental data where a ⫽ likely to produce unconstrained solutions if all of the ran-
b ⫽ 0 and ␴aj ,bj ⫽ 0 at each sample size and for both dom effects were normally distributed, if the a and b values
distributional conditions, which resulted in an additional 24 were small (.3), and if the covariance of aj and bj was zero.
conditions. Within each cell of the design, we simulated 500 The latter effect can be understood as a consequence of the
samples of data, which resulted in a total of 84,000 repli- population value for the corresponding correlation being
cations (further details on the data generation are available
farther away from the boundary values of ⫺1 and 1. Why
from Daniel J. Bauer on request). We then fitted the model
smaller values of a and b produced more unconstrained
specified in Equation 4 to each data set using SAS PROC
solutions is less clear.
MIXED with the restricted maximum likelihood estimator
and a maximum of 1,000 iterations.
We now consider four aspects of the results in turn. First, Bias of the Estimate
given the complexity of the model, we evaluate potential
We next evaluated the bias of the average indirect effect
problems of estimation by examining the rate of model
estimate. Bias was computed as the difference between the
nonconvergence and boundary solutions (e.g., solutions for
mean estimate and the corresponding population value. The
which a variance estimate is constrained to zero because it
would otherwise become negative). Second, we consider average indirect effect in the population was computed from
under what conditions the estimate of the average indirect the true parameter values and Equation 5. Across all 144
effect is appreciably biased. As part of this evaluation, we conditions in the factorial design, the mean bias for the
consider potential bias in the estimate of ␴aj ,bj in particular, average indirect effect estimate was ⫺.001, with a mini-
as we could not estimate this part of the average indirect mum of ⫺.076 and a maximum of .027. Further, the abso-
effect directly using the previous two-step approach (but it lute bias was less than .01 in all but 18 conditions.5 These
is provided by the simultaneous modeling approach advo- 18 conditions included 14 conditions with nonnormal ran-
cated here). Third, we examine Type I error rates for the dom effects and 9 conditions with the minimum total sam-
null hypothesis test of the average indirect effect. Last, we ple size of N ⫽ 25 and nj ⫽ 4. A similar evaluation of the
compare the coverage rates and power of CIs constructed bias in the covariance estimate for aj and bj revealed that,
with the normal approximation and the MC method. across the 144 conditions, the mean bias was .0002, with a
minimum of ⫺.033 and a maximum of .041. The absolute
bias was less than .01 in all but 12 conditions, all of which
Difficulty of Estimation were conditions for which nj ⫽ 4. Ten of the 12 conditions
The outcome of fitting each model was categorized as included nonnormal random effects, and 7 of the 12 condi-
either nonconverged, a boundary solution, or an uncon- tions had the minimum number of Level 2 units (N ⫽ 25).
strained (nonproblematic) solution. Boundary solutions Overall, these results indicated that the estimate of the
were defined as solutions in which one or more random average indirect effect (and the covariance estimate for aj
effect was constrained to have a variance of zero or the
correlation between two random effects was constrained to 5
We also evaluated the relative bias of the estimates but quickly
be 1 to obtain a solution in which all estimates attained
ran into difficulty interpreting these values. That is, relative bias is
permissible values. Though constraints to boundary values
computed as the ratio of bias to the true parameter value, but if the
do not invalidate the model results, they are typically seen true parameter value is small even relatively trivial levels of bias
as an undesirable result and suggest that the available data will result in high relative bias. Consistent with this, we found that
are not sufficient to support the complexity of the model. It the relative bias was highest for conditions in which the population
is not surprising that the results shown in Table 1 indicate parameter value was lowest. For this reason, we focus in the text
that as the sample size increased, by increased N or nj, fewer on bias as opposed to relative bias.
150 BAUER, PREACHER, AND GIL

Table 1
Result of Fitting the Lower Level Mediation Model With Random Causal Effects as a Function
of the Number of Level 2 Units (N) and the Number of Level 1 Units Within Each Level 2 Unit
(nj)
Result of model fitting (% of replications)
Failed to Boundary Unconstrained
Observations converge solution solution
N ⫽ 25

nj ⫽ 4 93.03 4.90 2.07

nj ⫽ 8 56.63 23.10 20.27
nj ⫽ 16 16.18 23.02 60.80
N ⫽ 50

nj ⫽ 4 54.70 26.93 18.37

nj ⫽ 8 13.93 28.10 57.97
nj ⫽ 16 1.18 6.72 92.10
N ⫽ 100

nj ⫽ 4 16.52 34.13 49.35

nj ⫽ 8 1.80 10.97 87.23
nj ⫽ 16 0.37 0.67 98.97
N ⫽ 200

nj ⫽ 4 2.50 19.52 77.98

nj ⫽ 8 0.58 1.70 97.72
nj ⫽ 16 0.27 0.00 99.73
Note. Results are collapsed over conditions that vary in the population parameter values and in the distributions
of the random effects (normal or nonnormal).

and bj) was largely unbiased except for a few of the non- procedure is known to underestimate the sampling variabil-
normal random effects or very small sample conditions. It is ity of the fixed effect estimates because the variance com-
noteworthy that these are also the conditions for which it ponents of the model are treated as known (when they are,
was difficult to estimate the model, which provides con- in fact, estimated). Kackar and Harville (1984) provided a
verging evidence that we need larger samples to obtain method for inflating the elements of the covariance matrix
reliable results. of the fixed effects to correct for this bias. We compare the
CIs we obtained by using the default covariance matrix of
Type I Errors the fixed effects versus the inflated Kackar–Harville (K-H)
Using the 24 supplementary cells of the simulation design covariance matrix of the fixed effects. To do so, we calcu-
for which the population value of the average indirect effect lated Type I error rates as the percentage of replications in
was zero, we next considered the Type I error rates for tests which the 95% CI for the average indirect effect failed to
of the average indirect effect (setting the nominal error rate cover zero.
at 5%). We calculated 95% CIs using the normal approxi- Tables 2 and 3 compare the Type I error rates for the four
mation in Equation 11 and the MC method of MacKinnon et types of CIs for normally and nonnormally distributed ran-
al. (2004). The latter method involved taking 50,000 ran- dom effects, respectively. The results indicate that all four
dom draws from the estimated sampling distribution of the CIs tended to produce Type I error rates below the nominal
estimates (i.e., the multivariate normal sampling distribu- level, especially at the lower sample sizes. CIs based on the
tion of â, b̂, and ␴ˆ aj ,bj ), calculating Equation 5 for each MC method, however, reflected the nominal Type I error
draw, and then computing the 2.5th and 97.5th percentiles rate better than the other CIs. Further examination of the
of the resulting values. Both the normal approximation and data indicated that, at the smallest sample sizes, the standard
the MC method require an estimate of the covariance matrix errors of the average indirect effect estimate were overesti-
of the fixed effects (e.g., a and b). Typically, one computes mated, producing CIs that were too wide and thus more
these values by taking the inverse of the information matrix likely to include zero. We now consider whether this is also
for the estimates. For multilevel models, however, this the case when the null hypothesis is false.
 RANDOM INDIRECT EFFECTS 151

Table 2
Type I Error Rates: Normal Random Effects
Normal approximation Monte Carlo method
a b c
Condition Replications Usual cov K-H cov Usual covb K-H covc
N ⫽ 25

1. nj ⫽ 4 70 0.00 0.00 0.00 0.00

2. nj ⫽ 8 337 2.97 2.97 5.34 5.34
3. nj ⫽ 16 484 2.48 2.48 4.34 4.13
N ⫽ 50

4. nj ⫽ 4 351 4.27 4.27 4.84 4.56

5. nj ⫽ 8 486 3.70 3.70 4.12 4.12
6. nj ⫽ 16 500 3.20 3.20 4.20 4.20
N ⫽ 100

7. nj ⫽ 4 482 4.98 4.98 5.39 5.39

8. nj ⫽ 8 499 4.41 4.41 4.61 4.61
9. nj ⫽ 16 500 6.00 6.00 6.20 6.00
N ⫽ 200

10. nj ⫽ 4 500 4.00 4.00 4.00 4.00

11. nj ⫽ 8 500 6.00 6.00 6.20 6.20
12. nj ⫽ 16 500 4.80 4.60 5.00 4.60
Collapsing

1 through 12 5,209 4.24 4.22 4.86 4.76

2 through 12 5,139 4.30 4.28 4.92 4.83
4 through 12 4,318 4.61 4.59 4.96 4.86
Note. Results are presented by the number of Level 2 units (N) and the number of Level 1 units per Level 2
unit (nj). The nominal Type I error rate was set at 5%.
a
Number of replications resulting in a converged solution from which Type I error rates were calculated (out
of 500, except for results collapsed over conditions). b Covariance matrix of a and b estimated by the inverse
of the information matrix. c Covariance matrix of a and b estimated by the method of Kackar and Harville
(1984).

CI Coverage Rates Tables 4 and 5 also point to some clear differences between
methods for computing the CIs. Across replications, the CIs
Coverage rates for the CIs are presented in Table 4 by
constructed via Equation 11 performed slightly better than
sample size and in Table 5 by the magnitude of the param-
those computed via the MC method. This advantage of the
eter values in the population model. Consistent with the
results on Type I errors, the CIs were too wide in the normal approximation method was seen for 105 of 144
smallest sample condition (N ⫽ 25, nj ⫽ 4) when the conditions (73%) but was stronger when the random effects
random effects were normal. Outside of this condition, were nonnormal. Additionally, Tables 4 and 5 show that the
however, the CIs actually tended to be too narrow, produc- K-H covariance matrix of the fixed effects produced CIs
ing slightly lower than 95% coverage of the true population with better coverage rates. Overall, we attained the best
parameter values. The dominant factor determining the cov- coverage rates by combining the normal approximation of
erage rates of the CIs was the distribution of the random Equation 11 with the K-H procedure. The MC method with
effects. In general, regardless of the method for computing the K-H procedure performed nearly as well.
the CIs, the coverage rates hovered around 94% when the Finally, Tables 4 and 5 indicate that the effect of nonnor-
random effects were normally distributed and dropped to mal random effects on the coverage rates of the CIs was
about 90% when the random effects were nonnormally moderated by several factors. First, the coverage rates were
distributed. The lower coverage rates for the nonnormal slightly better at the smallest sample sizes (low N combined
conditions reflected a general underestimation of the sam- with low nj). Second, the coverage rates were least affected
pling variances of all of the estimates in the model. by nonnormality when the covariance of the random effects
152 BAUER, PREACHER, AND GIL

Table 3
Type I Error Rates: Nonnormal Random Effects
Normal approximation Monte Carlo method
a b c
Condition Replications Usual cov K-H cov Usual covb K-H covc
N ⫽ 25

1. nj ⫽ 4 22 0.00 0.00 0.00 0.00

2. nj ⫽ 8 198 1.01 1.01 2.53 2.53
3. nj ⫽ 16 435 2.07 2.07 3.45 3.45
N ⫽ 50

4. nj ⫽ 4 205 1.46 1.46 2.44 2.44

5. nj ⫽ 8 450 4.67 4.44 4.67 4.67
6. nj ⫽ 16 495 4.65 4.65 5.25 5.25
N ⫽ 100

7. nj ⫽ 4 428 5.14 5.14 5.61 5.61

8. nj ⫽ 8 495 3.43 3.43 4.04 3.84
9. nj ⫽ 16 500 4.20 4.20 4.80 4.80
N ⫽ 200

10. nj ⫽ 4 497 4.83 4.63 5.03 5.03

11. nj ⫽ 8 500 4.00 4.00 4.40 4.20
12. nj ⫽ 16 500 4.60 4.60 5.00 5.00
Collapsing

1 through 12 4,725 3.92 3.87 4.49 4.44

5 through 12 3,865 4.42 4.37 4.84 4.79
Note. Results are presented by the number of Level 2 units (N) and the number of Level 1 units per Level 2
unit (nj). The nominal Type I error rate was set at 5%.
a
Number of replications resulting in a converged solution from which Type 1 error rates were calculated (out
of 500, except for results collapsed over conditions). b Covariance matrix of a and b estimated by the inverse
of the information matrix. c Covariance matrix of a and b estimated by the method of Kackar and Harville
(1984).

was zero, because of bias in the estimate of the covariance method, the MC method also resulted in superior power in
when it was not zero. As would be expected, this effect was 136 of the 144 conditions in the simulation study (94.4%).
greatest when the a and b parameters were small (reflecting These differences in power were, however, also quite small,
the then proportionally larger contribution of the covariance never differing by more than .0045.
to the expected value). In contrast, these same factors min-
imally affected the coverage rates when the random effects Summary
were all normally distributed.
In total, both the normal approximation method and the
MC method for constructing CIs performed well when the
Power random effects were normally distributed and the K-H pro-
To evaluate power, we calculated the proportion of rep- cedure was used to estimate the sampling covariance matrix
lications with CIs for the average indirect effect excluding of the fixed effects. Between the two, the MC method was
zero for each of the 144 cells in the factorial design. The slightly more powerful and attained nominal Type I error
results were consistent with the CI coverage rates. Because rates at smaller sample sizes but also had slightly lower CI
the K-H procedure necessarily results in larger estimates of coverage rates. When the random effects were nonnormal,
sampling variability, this procedure produces wider CIs and, effect estimates remained unbiased in most cells of the
hence, lower power. Across conditions, however, the aver- design, but the CIs did not have accurate coverage rates.
age loss in power due to the use of the K-H procedure never One exception was when the null hypothesis was true, in
exceeded .0022 for either the normal approximation or the which case the CIs covered zero approximately 95% of the
MC method. Similarly, because the MC method produced time, yielding roughly 5% Type I error rates, except in the
consistently narrower CIs than the normal approximation smallest samples. We now turn to a slightly more compli-
 RANDOM INDIRECT EFFECTS 153

Table 4
Confidence Interval Coverage by Distribution and Sample Size
Normally distributed aj and bj Nonnormally distributed aj and bj
Normal approximation Monte Carlo method Normal approximation Monte Carlo method
Condition Usual cova K-H covb Usual cova K-H covb Usual cova K-H covb Usual cova K-H covb
N ⫽ 25

nj ⫽ 4 96.43 97.92 95.83 97.02 92.68 95.12 91.46 95.12

nj ⫽ 8 93.75 94.55 93.62 94.43 92.71 93.82 91.89 93.31
nj ⫽ 16 93.74 93.93 93.36 93.66 89.64 90.10 89.18 89.68
N ⫽ 50

nj ⫽ 4 93.51 94.70 93.57 94.52 91.43 92.97 91.04 92.49

nj ⫽ 8 94.12 94.26 94.01 94.37 90.90 91.46 90.16 90.82
nj ⫽ 16 94.63 94.63 94.59 94.69 91.02 91.09 90.41 90.55
N ⫽ 100

nj ⫽ 4 93.04 93.49 92.93 93.49 91.22 91.87 90.92 91.61

nj ⫽ 8 93.85 93.98 93.88 93.98 90.13 90.37 89.85 90.02
nj ⫽ 16 94.76 94.83 94.73 94.76 90.27 90.34 89.74 89.87
N ⫽ 200

nj ⫽ 4 94.08 94.35 94.18 94.15 89.29 89.50 88.91 89.05

nj ⫽ 8 94.63 94.63 94.33 94.46 90.67 90.77 90.64 90.74
nj ⫽ 16 94.60 94.60 95.00 95.00 89.68 89.75 89.78 89.88
All replications 94.14 94.39 94.10 94.35 90.44 90.77 90.06 90.43
Note. Results are presented by the number of Level 2 units (N) and the number of Level 1 units per Level 2 unit (nj).
a
Covariance matrix of a and b estimated by the inverse of the information matrix. b Covariance matrix of a and b estimated by the method of Kackar and
Harville (1984).

cated multilevel mediation model, one in which the strength as well as our preceding developments to propose a strategy
of the causal effects depends on other predictors in the for investigating such effects in multilevel models.
model. The moderated multilevel mediation model diagrammed in
Figure 4 has the same Level 1 equations as before, namely,
Moderated Multilevel Mediation
M ij ⫽ dMj ⫹ aj Xij ⫹ eMij ,
Given evidence of significant random effects for either
the aj or bj paths, one may wish to add factors to the model Y ij ⫽ dYj ⫹ bj Mij ⫹ c⬘j Xij ⫹ eYij (13)
to explain this variability (Kenny et al., 2003). A simple
but now each random coefficient is expressed as a linear
version of this model, with a single Level 2 predictor W, is
function of W in the Level 2 model:
diagrammed in Figure 4. Significant prediction of either aj
or bj by W would represent a case of moderated mediation,
d Mj ⫽ ␥dM0 ⫹ ␥dM1 Wj ⫹ ud Mj
in which the strength of the indirect effect of the Level 1
predictor X depends on the Level 2 predictor W. The Level a j ⫽ ␥a0 ⫹ ␥a1 Wj ⫹ uaj
2 predictor may also impact c⬘j . Though this is not a case of
moderated mediation as usually defined, significant predic- d Yj ⫽ ␥dy0 ⫹ ␥dy1 Wj ⫹ udyj
tion solely of c⬘j still alters the balance of direct and indirect
effects relative to the total effects and may thus be of b j ⫽ ␥b0 ⫹ ␥b1 Wj ⫹ ubj
substantive interest. The topic of moderated mediation (of
several forms) has been explored more extensively for sin- c⬘j ⫽ ␥c⬘0 ⫹ ␥c⬘1 Wj ⫹ uc⬘j . (14)
gle-level regression models (e.g., James & Brett, 1984; Judd
& Kenny, 1981; Lance, 1988; Morgan-Lopez, 2003; Muller, The assumption of normality previously made for aj , bj , and
Judd, & Yzerbyt, 2005; Preacher, Rucker, & Hayes, 2006; c⬘j in Equation 2 now shifts to the residuals of the random
Wegener & Fabrigar, 2000). We now draw on this literature coefficients, designated by the symbol u. In particular, these
154 BAUER, PREACHER, AND GIL

Table 5
Confidence Interval Coverage by Distribution and Effect Size
Normally distributed aj and bj Nonnormally distributed aj and bj
Normal approximation Monte Carlo method Normal approximation Monte Carlo method
Condition Usual cova K-H covb Usual cova K-H covb Usual cova K-H covb Usual cova K-H covb
a ⫽ b ⫽ .3

␴aj ,bj ⫽ ⫺.113 94.95 95.14 94.75 94.85 91.12 91.41 90.51 95.12
␴aj ,bj ⫽ 0 95.07 95.27 94.78 94.90 93.72 93.96 93.34 93.31
␴aj ,bj ⫽ .113 93.60 93.72 94.23 94.31 84.33 84.62 83.65 89.68
a ⫽ b ⫽ .6

␴aj ,bj ⫽ ⫺.113 93.78 94.21 93.42 93.92 91.90 92.41 91.53 92.49
␴aj ,bj ⫽ 0 93.91 94.17 93.79 94.19 93.69 94.21 93.71 90.82
␴aj ,bj ⫽ .113 93.51 93.80 93.57 93.92 87.49 87.76 87.26 90.55
All replications 94.14 94.39 94.10 94.35 90.44 90.77 90.06 90.43
Note. Results are presented by the value of fixed effects in the population (a and b) and the covariance of the random effects in the population (␴aj ,bj).
a
Covariance matrix of a and b estimated by the inverse of the information matrix. b Covariance matrix of a and b estimated by the method of Kackar
and Harville (1984).

residuals are assumed to be normal with an expected value (Raudenbush & Bryk, 2002, pp. 58 – 61). Note that this joint
of zero and a full covariance matrix. The symbol ␥ is test is only possible with the simultaneous estimation strat-
reserved for the fixed effects of the model. Note that one egy advocated in this article. Regardless of which testing
may or may not wish to include W as a predictor of all of the strategy is taken, rejection of the null hypothesis would
random slopes (aj, bj, and c⬘j ). However, when W is included provide evidence of moderated mediation.
as a predictor of one of the random slopes for an equation, Given evidence of moderated mediation, we may wish to
we strongly recommend that it be included as a predictor of evaluate the simple indirect effects, that is, the value of
the random intercept for the same equation.6 Equation 15 at various levels of W. Inference tests of the
The conditional expected value of the indirect effect is simple indirect effects are also possible. Note that in the
E共aj bj 兩Wj ⫽ w兲 ⫽ 共␥a0 ⫹ ␥a1 w兲共␥b0 ⫹ ␥b1 w兲 ⫹ ␴uaj ,ubj. (15) special case in which W ⫽ 0, the simple indirect effect
obtained from Equation 15 is simply
To test moderation of the indirect effect, we can either test
␥a1 and ␥b1 separately (the default of most software) or, E共aj bj 兩Wj ⫽ 0兲 ⫽ ␥a0 ␥b0 ⫹ ␴uaj ,ubj . (16)
perhaps preferably, perform a joint test of ␥a1 and ␥b1
In this special case, Equation 16 is of the same basic form
as Equation 5, and we can calculate the sampling variance
of the simple indirect effect with the analogue to Equation
9, or

Var共␥ˆ a0 ␥ˆ b0 ⫹ ␴ˆ u aj ,u bj 兲 ⫽ ␥b0
2
Var共␥ˆ a0 兲 ⫹ ␥a0
2
Var共␥b0 兲
⫹Var共␥ˆ a0 兲Var共␥ˆ b0 兲 ⫹ 2␥a0␥b0Cov共␥ˆ a0,␥ˆ b0兲
⫹ Cov共␥ˆ a0,␥ˆ b0兲2 ⫹ Var共␴ˆ uaj ,ubj 兲. (17)

6
Figure 4. Moderated lower level mediation model: Nested When a Level 2 covariate predicts a random slope, this results
frames indicate levels of sampling, boxes indicate variables, ar- in a cross-level interaction. Likewise, when a Level 2 covariate
rows without circles represent fixed effects, arrows with circles predicts a random intercept, this results in a main effect. One can
represent random effects, and arrows not originating from vari- see this by forming the reduced form equation for the model (i.e.,
ables indicate residuals. Note that the arrows from W to the random by substituting the Level 2 equations into the Level 1 equations).
effects indicate moderation of the hypothesized causal effects. Typically, one should always include main effects when evaluating
Predictions of M and Y by W are not shown on the diagram but interactions. See Bauer and Curran (2005) for additional discus-
would be included in the model. sion of these issues for multilevel models.
 RANDOM INDIRECT EFFECTS 155

Replacing the population values in Equation 17 with our sample average physical sensation of pain (PHYS), emotional dis-
estimates, we can use the normal approximation or MC method comfort of the pain (EMOT), and level of stress (STRESS)
to construct a CI for the simple indirect effect at W ⫽ 0. that the participant experienced over the day. Scores were
Given these simplifications, one way to conduct tests of scaled to range from 0 to 10, with higher values indicating
simple indirect effects at other levels of W is to rescale W so greater pain or stress.
that the value of interest is equal to zero, much as is often Given the study design, the Level 1 units are the daily
done to evaluate other types of simple effects (e.g., see reports of pain and stress by the participants, and the Level
Aiken & West, 1991). For example, suppose w is the 2 units are the participants. The model under investigation
conditional value of W that is of interest. We can form a new posits that the experience of physical pain will increase
variable WCV ⫽ W ⫺ w that is centered on this value. If we stress but that this effect will be largely mediated by the
reestimate the model, replacing W with WCV, the coeffi- emotional response to the pain, that is, its perceived un-
cients for the fixed effects will adjust to account for the pleasantness. Further, given evidence of individual differ-
change in scale, but the model will be equivalent. With ences in the direct and indirect effects of physical pain, we
these new coefficients, we can now apply Equations 16 and seek to explain these differences on the basis of Level 2
17 to obtain the estimate and sampling variance of the covariates collected at the baseline interview (i.e., extending
simple indirect effect at W ⫽ w (or WCV ⫽ 0). In this way, to a moderated mediation model). To simplify our analysis
we can compute a CI for the simple indirect effect at W ⫽ of the data, the model we fit assumes that missing data,
w. The simulation study conducted in the prior section including data that are missing due to attrition, are missing
suggests that these CIs will be reasonably accurate when the at random (Raudenbush & Bryk, 2002, pp. 199 –200). In
distributions of the random effects are fairly normal but less actuality this may be untrue, and a more complex analysis
so when these distributions are nonnormal. may be required (Schafer & Graham, 2002, provided a
Similarly, the simple direct effect is review of current approaches to modeling nonignorably
missing data). Finally, we used the K-H method for esti-
E共c⬘j 兩Wj ⫽ w兲 ⫽ ␥c 0⬘ ⫹ ␥c 1⬘ w. (18) mating the covariance matrix of the fixed effects and the
MC method for estimating 95% CIs for the average indirect
Here, the test of moderation is carried by a single parameter and total effects, although in all cases the inferences made
estimate, ␥ˆ c⬘1. If this estimate is significantly different from would have been identical if we had used the normal ap-
zero, we may wish to probe the simple direct effects of X on proximation method. SAS code for implementing the mod-
Y. Again, by forming a new variable WCV, centered at a els presented here is listed on the Web at http://dx.doi.org/
conditional value of interest for W, we can obtain tests of the 10.1037/1082-989X.11.2.142.supp.
simple direct effects through the test of ␥ˆ c 0⬘. The simple total For our initial model, the Level 1 equations were speci-
effects of X on Y are the sum of the simple indirect and fied as
direct effects, and inference tests can be conducted on the
summed estimates in a similar fashion. More complex methods Emotij ⫽ dEj ⫹ aj Physij ⫹ eEij
for probing moderation effects in multilevel models, including
the calculation of regions of significance, are given in Bauer Stressij ⫽ dSj ⫹ bj Emotij ⫹ c⬘j Physij ⫹ eSij (19)
and Curran (2005) and could also be extended to moderated
mediation models (Preacher et al., 2006). where the subscript i references the repeated assessment and
the subscript j references the participant. The coefficients aj,
bj, and c⬘j were all also allowed to be random, and all of the
Empirical Example random effects were allowed to covary. Homogeneity of
variance of the Level 1 residuals was assumed within each
We now implement the procedures we have outlined with
equation, but the residual variances were allowed to differ
data from a study concerning perceptions of pain and stress
across equations, as shown here:
in African American adolescents and adults with SCD (Gil
et al., 2003, 2004). As noted in the introduction, multilevel Var共eEij 兲 ⫽ ␴E2
models are applicable both to the case of individuals nested
within groups and to the case of repeated measures nested Var共eSij 兲 ⫽ ␴S2 . (20)
within individuals. In this application, the latter structure is
present. The 94 participants first completed a baseline in- Aside from the random effects of the model, serial autocor-
terview and were then asked to complete a daily diary each relation is a second common source of dependence in daily
evening for up to 6.5 months. The number of days diaries diary data (Schwartz & Stone, 1998; West & Hepworth,
were completed ranged from 2 to 196, with a median of 69 1991). That is, ratings made close in time to one another are
days (a 75% completion rate). Diaries included three typically more highly correlated than ratings made farther
100-mm visual analogue scales designed to measure the apart in time. Although a variety of methods can be used to
156 BAUER, PREACHER, AND GIL

account for serial autocorrelation (Beal & Weiss, 2003; tively. The formulas for the variances require normality of the
West & Hepworth, 1991), in the present case we assumed a random effects to be exact, so we first checked this assumption
continuous-time autoregressive structure for both outcomes by plotting the distributions of the empirical Bayes estimates
in Equation 19 following the recommendation of Schwartz for the random coefficients. We noted some nonnormality, as
and Stone (1998). This error structure assumes that the depicted in Figure 5 for the aj and bj estimates.
correlation between the residuals across assessments de- Next, noting that the covariance between aj and bj was
clines as an exponential function of the time lag between estimated as .01 (r ⫽ .26), we determined the estimated
assessments and allows for varying time intervals between average indirect effect to be .29 (95% CI ⫽ .16, .41) and the
assessments. More formally, for any two observations i and estimated average total effect to be .43 (95% CI ⫽ .36, .51).
i⬘, the covariance between the residuals was specified as Given the nonnormality of the random effects, the results of
our simulation study suggest that these CIs may be too
Cov共eEij ,eEi⬘j 兲 ⫽ ␴E2 ␳Edii⬘
narrow and should be interpreted with caution. In contrast,
the effect estimates are probably unbiased, indicating that,
Cov共eSij ,eSi⬘j 兲 ⫽ ␴S2 ␳Sdii⬘ , (21)
on average, about 67% of the total effect of physical pain on
where dii⬘ is the number of days elapsed between observa- stress was indirect (mediated by the emotional reaction to
tions i and i⬘, ␳E is the autoregressive parameter for emo- pain). Individual differences in the indirect effect were
tional reactions to pain, and ␳S is the autoregressive param- characterized by Equation 6 to have a variance of .11 (SD ⫽
eter for stress ratings. We estimated the full model with .33). The variance of the total effect across participants was
restricted maximum likelihood using the specification strat- estimated from Equation 8 to be .08 (SD ⫽ .29). Although
egy detailed earlier in the article. it may at first appear odd that the variance of the total effect
The results for the initial model are reported in Table 6. All was less than the variance of the indirect effect, the reason
three causal paths, aj, bj, and c⬘j , varied significantly across for this finding is that the direct effect was negatively
persons and, on average, were positive and significantly dif- correlated with both components of the indirect effect. That
ferent from zero. To better understand this pattern of effects, is, as the indirect effect went up, the direct effect went
we next estimated the expected value and variance of the down, and vice versa, resulting in less variability in the total
random indirect and total effects, aj bj and aj bj ⫹ c⬘j , respec- effect than in the indirect effect.

Table 6
Empirical Analysis Results
95% confidence limits
Effect Estimate SE Lower Upper
Fixed (average) effects

dE 0.028 0.026 ⫺0.025 0.081

dS 1.050 0.013 0.794 1.306
a 0.909 0.014 0.880 0.937
b 0.302 0.069 0.163 0.441
c⬘ 0.148 0.059 0.029 0.267
Level 1 residual structure

␴E2 0.206 0.003 0.199 0.212

␴S2 2.501 0.046 2.417 2.596
␳E 0.059 0.014 0.031 0.087
␳S 0.296 0.012 0.272 0.320
Covariance/correlation matrix of random effectsa
1 2 3 4 5
1. dEj 0.050 0.275 ⫺0.276 ⫺0.140 0.321
2. dSj 0.070 1.319 ⫺0.040 ⫺0.346 0.104
3. aj ⫺0.007 ⫺0.005 0.014 0.258 ⫺0.228
4. bj ⫺0.011 ⫺0.137 0.011 0.119 ⫺0.495
5. c⬘j 0.016 0.026 ⫺0.006 ⫺0.037 0.048
a
The variances of the random effects are shown on the diagonal, the covariances of the random effects are shown
below the diagonal, and the correlations among the random effects are shown above the diagonal.
 RANDOM INDIRECT EFFECTS 157

Figure 5. Diagnostic plots for judging the normality assumption for the random slopes involved
in the indirect effect estimate. Top panels are histograms for the empirical Bayes (EB) estimates of
the random effects obtained by fitting the initial model to the sickle cell disease data; bottom panels
are normal quantile– quantile plots for the EB estimates.

To better understand individual differences in the Given that the moderating effects of ACUTE on the three
strength of the direct, indirect, and total effects, we next causal paths were nonsignificant, the analysis would typi-
extended the model we have described by adding a Level 2 cally end here. However, simply to demonstrate the meth-
predictor. The predictor of interest was the number of major odology described in the preceding section, we proceeded to
acute complications of SCD (ACUTE) the patient reported probe the simple direct, indirect, and total effects of phys-
having experienced within the past year. This predictor was ical pain on stress at each observed count of acute compli-
added to the equation for each random effect. Of greatest cations, zero through six. As can be seen in Figure 6,
interest was whether acute complications of SCD would although the total effect was relatively unaffected by the
moderate the indirect effect, or aj bj. The coefficient esti- moderator, the balance of the direct and indirect effects
mates indicated that higher levels of acute complications shifted markedly. For participants reporting zero acute SCD
reduced the role of emotional reactions to pain as a mediator complications in the past year, the effect of physical pain on
between physical pain and stress; however, the simulta- stress was essentially all indirect. As the number of acute
neous test of these coefficients was not significant, ␹2(2) ⫽ SCD complications reported by the participant rose, how-
4.75, p ⫽ .093. In contrast, higher levels of acute compli- ever, this indirect effect diminished to nearly zero, and the
cations appeared to increase the direct effect of physical direct effect of physical pain on stress increased to higher
pain on stress, but this was also not significant, t(24.4) ⫽ levels. Inferential tests of the simple direct and indirect
1.82, p ⫽ .080. effects revealed significant indirect effects for participants
158 BAUER, PREACHER, AND GIL

Figure 6. Bars indicate the decomposition of the average causal effect of physical pain on stress
into a direct effect and an indirect effect through emotional reactions to pain. The total height of each
column conveys the magnitude of the total effect, with the exception of the first column, where the
direct effect is negative and the total effect is indicated with a white horizontal line. Note that the
total effect of physical pain on stress is relatively constant but that the mediation of this effect by
emotional reactions to pain wanes as the severity of sickle cell disease (SCD) increases.

reporting four or fewer acute SCD complications in the past fication. The result may be bias in the indirect effect esti-
year and significant direct effects for participants reporting mate. Confidence in the independence assumption can be
three or more complications. The simple total effects were increased through the inclusion of potential common causes
relatively stable and statistically significant across all ob- of Y and M (other than X) in the model and through the use
served levels of acute SCD complications. of multiple methods of data collection. Further recommen-
dations for improving causal inferences can be found in
Epistemological Issues in Assessing Mediation Berk (1988, 2004), Cole and Maxwell (2003), Gollob and
Reichardt (1987, 1991), and Hoyle and Robinson (2004).
Having established and demonstrated procedures for eval- With regard to model specification, one issue that is
uating hypotheses of lower level mediation in multilevel unique to the lower level mediation model is the partitioning
models, we now review several important epistemological of the covariance structure for the Level 1 observations. If
issues that must be considered whenever mediation is to be the Level 1 residual covariance structure is not correctly
assessed. The foremost assumption of any mediation model specified, then this may lead one to overestimate the vari-
is that the distal predictor and the mediator exert causal ance components for the random effects. As a consequence,
effects on their respective dependent variables. Necessary one could conclude that there is more heterogeneity in the
preconditions for causal inference are that (a) the variables hypothesized causal effects of the model than is in fact
involved must covary with each other, (b) causes must occur present in the population of Level 2 units. A second spec-
before their presumed effects, and (c) sources of spurious ification issue unique to the multilevel setting is which
covariation should be eliminated (Frazier, Tix, & Barron, effects should have random versus fixed components. This
2004). To satisfy the second condition, X should be mea- issue is important because misspecification of the random
sured prior to M and M should be measured prior to Y effects could also bias the indirect effect estimate. Ideally,
whenever possible (Gollob & Reichardt, 1987, 1991). Ex- the theoretical model of the causal processes will indicate
perimental manipulation of X can greatly strengthen the which of the causal effects may show heterogeneity across
causal inference, as can secondary manipulation of M Level 2 units. If, however, theory does not provide strong
(Spencer, Zanna, & Fong, 2005). With respect to the third guidance, then researchers may wish to empirically evaluate
condition, MacKinnon et al. (2002) and Holland (1988) the issue by including random effects for each causal path
noted that the independence assumption made for the cross- and assessing the magnitude of the estimated variance com-
equation residuals in Equation 1 is particularly question- ponents for these effects.
able. Violation of this assumption can occur in a number of In our view, it is rarely the case that an investigator is able
ways, including the omission of important variables, the to address all of the aforementioned issues simultaneously.
presence of common method variance, and model misspeci- For instance, for the SCD analyses we have presented, the
 RANDOM INDIRECT EFFECTS 159

measures are contemporaneous, and questions could be vances in multilevel structural equation modeling, however,
raised about whether the causal effects are in the correct it is currently infeasible to estimate a 1 3 1 3 1 mediation
direction (or bidirectional), whether the estimates are in- model with random direct and indirect effects among the
flated by common method variance, and whether an autore- latent variables. The principal difficulty is that the typical
gressive structure was sufficient to capture the over-time estimators for multilevel SEMs allow for random intercepts
residual correlations. Like the SCD analyses, definitive ev- but not random slopes (Bentler & Liang, 2003; Goldstein &
idence of mediation is often not obtained from a single MacDonald, 1988; B. O. Muthén, 1994; B. O. Muthén &
model; instead, multiple studies are typically necessary. Satorra, 1995). In principle, maximum likelihood with numer-
ical integration (L. K. Muthén & Muthén, 2004) or Bayesian
Limitations and Future Directions estimation techniques (Ansari, Jedidi, & Dube, 2002) can be
used to include random slopes in a multilevel SEM. Future
Although we believe that the simultaneous approach we research should explore the potential use of these new methods
have proposed to modeling random indirect effects in 1 3 with 1 3 1 3 1 models for latent variables.
1 3 1 multilevel mediation models has a number of A number of practical difficulties may also be encoun-
strengths, we would be remiss if we did not also note its tered by applied researchers wishing to fit these models.
limitations. First, our simulation study considers only a Foremost, in prototypical form, the model includes two
small set of the possible conditions that might be encoun- random intercepts and three random slopes. In general, the
tered in practice. Indeed, our own analysis of the SCD data estimation of a multilevel model becomes more difficult and
involves unbalanced data, potentially nonrandom attrition, computationally demanding as the number of random ef-
and serial dependence, none of which is considered in the fects increases, and the present case is no exception. The
simulation study. We thus view our results as a preliminary identification of the variance components depends heavily
indication of when and where the performance of the pro- on the number of Level 1 observations per Level 2 unit,
posed procedures can be expected to suffer when the as- whereas the accuracy with which they are estimated de-
sumptions of the model are unmet, but more studies are pends on the number of Level 2 units (Hox, 2002). In our
needed to make more definitive conclusions. simulation study, we encountered serious difficulty estimat-
One clear implication of our simulation study is that ing the model when the number of Level 1 observations was
additional thought must be given to the best way to accom- small (e.g., four). Given this, certain kinds of study designs
modate nonnormal random effects. One possibility is to are more likely to permit the estimation of this model than
retain the current model but implement a nonparametric others. For instance, studies using ecological momentary
bootstrap. Recently, Pituch, Stapleton, and Kang (in press) assessment (e.g., diary studies) typically yield many observa-
evaluated bootstrapping for conducting tests of mediation in tions per unit and hence may be ideally suited for the estima-
multilevel models and obtained promising results. Their tion of lower level mediation models. In contrast, studies with
research was, however, restricted to models with normally fewer observations per unit may not support the estimation of
distributed random effects and without random indirect five random effects, forcing the investigator to simplify the
effects. Unfortunately, the computational burden involved model by removing one or more random effects.
in bootstrapping the more complex models we consider may
be prohibitive. For instance, the initial model we fitted to the Conclusions
SCD data took 40 min to estimate. Running the same model
on 1,000 resampled data sets would thus take about 28 days The ability to investigate heterogeneity in causal effects is
of computing time. Some way must be found to improve the one of the most attractive features of multilevel models.
efficiency of the process if bootstrapping is to be a feasible Most approaches for investigating mediation in multilevel
option for these models. models have, however, exclusively focused on fixed causal
An additional limitation of the present approach is that the effects. In an important next step, Kenny et al. (2003)
model requires that the residuals for the mediator and the proposed an approach to evaluating 1 3 1 3 1 mediation
distal outcome be uncorrelated, an assumption that may not when the causal paths are random. In this model, the direct,
hold for reasons discussed in the prior section. For single- indirect, and total effects vary in strength across the Level 2
level models, one can potentially address this assumption by units of the population. The present article extends the work
using a structural equation model (SEM) in which M and Y of Kenny et al. (2003) in two important ways. First, we
are latent variables and the residuals of the indicators for the present a simultaneous modeling approach that provides all
latent variables covary across constructs (see, e.g., Bollen, of the necessary information to calculate the variances of
1989, p. 324). SEMs also account for potential measure- the random indirect and total effects as well as the standard
ment error in the observed variables and can be used to errors of the average indirect and total effects. We believe
partial out common method variance (Bollen & Paxton, that the formulas for calculating these standard errors are
1998; Kenny & Zautra, 2001). Despite many recent ad- new results. Additionally, this simultaneous modeling ap-
160 BAUER, PREACHER, AND GIL

proach offers the opportunity to conduct tests of moderated Bollen, K. A., & Paxton, P. (1998). Detection and determinants of
mediation when the components of the random indirect bias in subjective measures. American Sociological Review, 63,
effect are predicted by a Level 2 variable. Second, we 465– 478.
provide an initial study of the robustness of the estimates Catanzaro, S. J., & Laurent, J. (2004). Perceived family support,
and CIs for the average indirect effect using simulation negative mood regulation expectancies, coping, and adolescent
methodology. This study indicates that the estimate of the alcohol use: Evidence of mediation and moderation effects.
average indirect effect was unbiased under most conditions. Addictive Behaviors, 29, 1779 –1797.
The CIs also provided fairly accurate coverage rates when Cole, D. A., & Maxwell, S. E. (2003). Testing mediational models
the random effects were normal, although they were too with longitudinal data: Questions and tips in the use of structural
narrow when the random effects were nonnormal. With this equation modeling. Journal of Abnormal Psychology, 112, 558 –
caveat in mind, we believe that the developments we have 577.
presented offer applied researchers an improved approach Demidenko, E. (2004). Mixed models: Theory and applications.
for estimating indirect effects in 1 3 1 3 1 multilevel Hoboken, NJ: Wiley.
mediation models. Future developments such as those we Frazier, P. A., Tix, A. P., & Barron, K. E. (2004). Testing mod-
have suggested may offer further improvements, enabling erator and mediator effects in counseling psychology research.
applied researchers to more fully address hypotheses of Journal of Counseling Psychology, 51, 115–134.
mediation in multilevel data structures. Gil, K. M., Carson, J. W., Porter, L. S., Ready, J., Valrie, C.,
Redding-Lallinger, R., & Daeschner, C. (2003). Daily stress and
References mood and their association with pain, health-care use and school
activity in adolescents with sickle cell disease. Journal of Pe-
Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing diatric Psychology, 28, 363–373.
and interpreting interactions. Newbury Park, CA: Sage. Gil, K. M., Carson, J. W., Porter, L. S., Scipio, C., Bediako, S. M.,
Alwin, D. F., & Hauser, R. M. (1975). The decomposition of & Orringer, E. (2004). Daily mood and stress predict pain,
effects in path analysis. American Sociological Review, 40, health care use, and work activity in African American adults
37– 47.
with sickle-cell disease. Health Psychology, 23, 267–274.
Ansari, A., Jedidi, K., & Dube, L. (2002). Heterogeneous factor
Goldstein, H. I., & McDonald, R. P. (1988). A general model for
analysis models: A Bayesian approach. Psychometrika, 67, 49 –78.
the analysis of multilevel data. Psychometrika, 53, 455– 467.
Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator
Gollob, H. F., & Reichardt, C. S. (1987). Taking account of time
variable distinction in social psychological research: Concep-
lags in causal models. Child Development, 58, 80 –92.
tual, strategic, and statistical considerations. Journal of Person-
Gollob, H. F., & Reichardt, C. S. (1991). Interpreting and estimat-
ality and Social Psychology, 51, 1173–1182.
ing indirect effects assuming time lags really matter. In L. M.
Bauer, D. J. (2003). Estimating multilevel linear models as struc-
Collins & J. L. Horn (Eds.), Best methods for the analysis of
tural equation models. Journal of Educational and Behavioral
change (pp. 243–259). Washington, DC: American Psycholog-
Statistics, 28, 135–167.
ical Association.
Bauer, D. J., & Curran, P. J. (2005). Probing interactions in fixed
Goodman, L. A. (1960). On the exact variance of products. Jour-
and multilevel regression: Inferential and graphical techniques.
Multivariate Behavioral Research, 40, 373– 400. nal of the American Statistical Association, 55, 708 –713.
Beal, D. J., & Weiss, H. M. (2003). Methods of ecological mo- Holland, P. W. (1988). Causal inference, path analysis, and recur-
mentary assessment in organizational research. Organizational sive structural equations models. Sociological Methodology
Research Methods, 6, 440 – 464. 1988, 18, 449 – 484.
Bentler, P. M., & Liang, J. (2003). Two-level mean and covariance Hox, J. (2002). Multilevel analysis: Techniques and applications.
structures: Maximum likelihood via an EM algorithm. In S. P. Mahwah, NJ: Erlbaum.
Reise & N. Duan (Eds.), Multilevel modeling: Methodological Hoyle, R. H., & Robinson, J. C. (2004). Mediated and moderated
advances, issues, and applications (pp. 53–70). Mahwah, NJ: Erl- effects in social psychological research: Measurement, design,
baum. and analysis issues. In C. Sansone, C. C. Morf, & A. T. Panter
Berk, R. A. (1988). Causal inference for sociological data. In N. J. (Eds.), The Sage handbook of methods in social psychology (pp.
Smelser (Ed.), Handbook of sociology (pp. 155–172). Newbury 213–233). Thousand Oaks, CA: Sage.
Park, CA: Sage. James, L. R., & Brett, J. M. (1984). Mediators, moderators, and tests
Berk, R. A. (2004). Regression analysis: A constructive critique. for mediation. Journal of Applied Psychology, 69, 307–321.
Thousand Oaks, CA: Sage. Judd, C. M., & Kenny, D. A. (1981). Process analysis: Estimating
Bollen, K. A. (1987). Total, direct, and indirect effects in structural mediation in treatment evaluations. Evaluation Review, 5, 602–
equation models. Sociological Methodology 1987, 17, 37– 69. 619.
Bollen, K. A. (1989). Structural equations with latent variables. Kackar, R. N., & Harville, D. A. (1984). Approximations for
New York: Wiley. standard errors of estimators of fixed and random effects in
 RANDOM INDIRECT EFFECTS 161

mixed linear models. Journal of the American Statistical Asso- Muthén, B. O. (1994). Multilevel covariance structure analysis.
ciation, 79, 853– 862. Sociological Methods & Research, 22, 376 –398.
Kendall, M. G., & Stuart, A. (1969). The advanced theory of Muthén, B. O., & Satorra, A. (1995). Complex sample data in
statistics (3rd ed., Vol. 2). London: Charles Griffen & Co. structural equation modeling. Sociological Methodology, 1995,
Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data analysis in 216 –316.
social psychology. In D. Gilbert, S. T. Fiske, & G. Lindzey Muthén, L. K., & Muthén, B. O. (2004). Mplus user’s guide (3rd
(Eds.), The handbook of social psychology (4th ed., Vol. 1, pp. ed.). Los Angeles: Muthén & Muthén.
223–265). New York: McGraw-Hill. Pituch, K. A., Stapleton, L. M. & Kang, J. Y. (in press). A
Kenny, D. A., Korchmaros, J. D., & Bolger, N. (2003). Lower comparison of single sample and bootstrap methods to assess
level mediation in multilevel models. Psychological Methods, 8, mediation in cluster randomized trials. Multivariate Behavioral
115–128. Research.
Kenny, D. A., & Zautra, A. (2001). Trait-state models for longi- Pituch, K. A., Whittaker, T. A., & Stapleton, L. M. (2005). A
tudinal data. In L. M. Collins & A. G. Sayer (Eds.), New comparison of methods to test for mediation in multisite exper-
methods for the analysis of change (pp. 243–269). Washington, iments. Multivariate Behavioral Research, 40, 1–24.
DC: American Psychological Association. Preacher, K. J., Rucker, D. D., & Hayes, A. F. (2006). Suggested
Kreft, I., & de Leeuw, J. (1998). Introduction to multilevel mod- procedures for addressing moderated mediation hypotheses.
eling. London: Sage. Manuscript submitted for publication.
Krull, J. L., & MacKinnon, D. P. (1999). Multilevel mediation Rao, C. R. (1965). Linear statistical inference and its applications.
modeling in group-based intervention studies. Evaluation Re- New York: Wiley.
view, 23, 418 – 444. Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear
Krull, J. L., & MacKinnon, D. P. (2001). Multilevel modeling of models: Applications and data analysis methods (2nd ed.).
individual and group level mediated effects. Multivariate Be- Newbury Park, CA: Sage.
havioral Research, 36, 249 –277. Raudenbush, S. W., & Sampson, R. (1999). Assessing direct and
Lance, C. E. (1988). Residual centering, exploratory and confir- indirect effects in multilevel designs with latent variables. So-
matory moderator analysis, and decomposition of effects in path ciological Methods & Research, 28, 123–153.
models containing interactions. Applied Psychological Mea- Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of
surement, 12, 163–175. the state of the art. Psychological Methods, 7, 147–177.
MacCallum, R. C., Kim, C., Malarkey, W., & Kiecolt-Glaser, J. Schwartz, J. E., & Stone, A. A. (1998). Strategies for analyzing
(1997). Studying multivariate change using multilevel models ecological momentary assessment data. Health Psychology, 17,
and latent curve models. Multivariate Behavioral Research, 32, 6 –16.
215–253. Shrout, P. E., & Bolger, N. (2002). Mediation in experimental and
MacKinnon, D. P., Krull, J. L., & Lockwood, C. M. (2000). nonexperimental studies: New procedures and recommenda-
Equivalence of the mediation, confounding, and suppression tions. Psychological Methods, 7, 422– 445.
effect. Prevention Science, 1, 173–181. Spencer, S. J., Zanna, M. P., & Fong, G. T. (2005). Establishing a
MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., causal chain: Why experiments are often more effective than
& Sheets, V. (2002). A comparison of methods to test mediation mediational analyses in examining psychological processes.
and other intervening variable effects. Psychological Methods, Journal of Personality and Social Psychology, 89, 845– 851.
7, 83–104. Wegener, D. T., & Fabrigar, L. R. (2000). Analysis and design for
MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). nonexperimental data: Addressing causal and noncausal hypoth-
Confidence limits for the indirect effect: Distribution of the eses. In H. T. Reis & C. M. Judd (Eds.), Handbook of research
product and resampling methods. Multivariate Behavioral Re- methods in social and personality psychology (pp. 412– 450).
search, 39, 99 –128. Cambridge, England: Cambridge University Press.
Morgan-Lopez, A. A. (2003). A simulation study of the mediated Wei, M., Mallinckrodt, B., Russell, D. W., & Abraham, W. T.
baseline by treatment interaction effect in preventive interven- (2004). Maladaptive perfectionism as a mediator and moderator
tion trials. Unpublished doctoral dissertation, Arizona State between adult attachment and depressive mood. Journal of
University. Counseling Psychology, 51, 201–212.
Muller, D., Judd, C. M., & Yzerbyt, V. Y. (2005). When moder- West, S. G., & Hepworth, J. T. (1991). Statistical issues in the
ation is mediated and mediation is moderated. Journal of Per- study of temporal data: Daily experiences. Journal of Person-
sonality and Social Psychology, 89, 852– 863. ality, 59, 609 – 662.

(Appendix follows on next page)

162 BAUER, PREACHER, AND GIL

Appendix

Derivation of Variance Estimates

This appendix addresses the derivation of four variances men- so

tioned in the text:
Var共g兲 ⫽ a2 ␴b2j ⫹ b2 ␴a2j ⫹ ␴a2j ␴b2j ⫹ 2ab␴aj ,bj ⫹ ␴a2j ,bj . (A5)
1. variance of the indirect effect among Level 2 units
(Equation 6),
Heterogeneity of the Total Effect Among Level 2
2. variance of the total effect among Level 2 units (Equa- Units (Equation 8)
tion 8),
Kenny et al. (2003) represented variability in the total effect by
3. sampling variance of the mean estimated indirect effect
(Equation 9), and Var共aj bj ⫹ c⬘j 兲 ⫽ b2 ␴a2j ⫹ a2 ␴b2j ⫹ ␴a2j ␴b2j

4. sampling variance of the mean estimated total effect ⫹ 2ab␴aj ,bj ⫹ ␴a2j ,bj ⫹ ␴c⬘
2
j
⫹2b␴ aj ,c⬘j ⫹2a␴ bj ,c⬘j . (A6)
(Equation 10).
Again, in matrix notation,
All derivations that follow rely on a large statistics literature
stretching back to the 1930s addressing the distribution of the Var共g兲 ⫽ D⬘⌺(␪)D⫹1⁄2tr兵共H⌺共␪ 兲兲2 其. (A7)
product of two random variables. The variances may be derived in
a number of ways; for the sake of brevity, we rely on proofs using When g ⫽ aj bj ⫹c⬘j , 共␦g/␦a兲 兩a,b,c⬘ ⫽ b, 共␦g/␦b兲 兩a,b,c⬘ ⫽ a, and
a second-order elaboration of the well-known multivariate delta 共␦g/␦c⬘兲 兩a,b,c⬘ ⫽ 1, and hence D⬘ ⫽ [b a 1]. Here,
method (Rao, 1965) for deriving the variance of an approximated

冋 ␴ a2j ␴aj ,bj ␴aj ,c⬘j

册
function of variables. The idea behind this method is to expand the
function of interest using a second-order Taylor series around the ␴ aj ,bj ␴ b2j ␴ bj,c⬘j
⌺共 ␪ 兲 ⫽ (A8)
population means of the variables involved and then to find the ␴ aj,cj ␴ bj,c⬘j ␴ c⬘
2
j
variance of this expression. Under the assumption of normality,
this method yields an exact result for the variances needed here. and

Heterogeneity of the Indirect Effect Among Level 2

Units (Equation 6) H⫽ 冋 0
1
0
1
0
0
0
0
0
册 , (A9)

Kenny et al. (2003, Equation 11) quantified the variability of

this indirect effect in multilevel modeling as so

Var共aj bj 兲 ⫽ b2 ␴a2 ⫹ a2 ␴b2 ⫹ ␴a2 ␴b2 ⫹ 2ab␴aj ,bj ⫹ ␴a2j ,bj . (A1) Var共g兲 ⫽ a2 ␴b2j ⫹ b2 ␴a2j ⫹ ␴a2j ␴b2j

In what follows, we use a second-order elaboration of the multi- ⫹ 2ab␴aj ,bj ⫹ ␴a2j ,bj ⫹ ␴c⬘
2
j
⫹ 2b␴aj ,c⬘j 2a␴bj ,c⬘j . (A10)
variate delta method for deriving the variance of a function of
random variables. In compact matrix notation, Sampling Variability of the Mean Estimated Indirect
Effect (Equation 9)
Var共g兲 ⫽ D⬘⌺共␪ 兲D ⫹ 1⁄2tr兵共H⌺共␪ 兲兲2 其, (A2)
The sampling variance of the mean estimated indirect effect
where ␪ is a vector of the variable elements of g, D is the gradient
may be obtained by similar procedures. The mean indirect effect is
(vector of first derivatives) of g with respect to ␪ evaluated at the
means of aj and bj, H is the Hessian (matrix of second derivatives)
g ⫽ ab ⫹ ␴aj bj . (A11)
of g with respect to ␪, and ⌺(␪) is the covariance matrix of ␪. When
g ⫽ aj bj, 共␦g/␦a兲兩a,b ⫽ b and 共␦g/␦b兲兩a,b ⫽ a, and hence D⬘ ⫽
Using the second-order delta method requires the gradient and
[b a]. Here,
Hessian of g and the asymptotic covariance matrix of parameter

冋 册
estimates. Here, D⬘ ⫽ [b a 1],
␴ a2j ␴aj ,bj
⌺共 ␪ 兲 ⫽

冋 册
␴aj ,bj ␴b2j (A3)
0 1 0
and H⫽ 1 0 0 , (A12)
0 0 0

H⫽ 冋 01 10 册, (A4) and
 RANDOM INDIRECT EFFECTS 163

⌺共 ␪ 兲 ⫽ 冋 ␴ â2 ␴ â,bˆ
␴ â,bˆ ␴ˆb
0
2
0
0
0 ␴ ␴2ˆ a ,b
j j
册 (A13) H⫽
冤
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
冥 , (A16)

(asymptotic covariances between fixed and random effect param- and

eters are zero). Thus,

冤 冥
␴ â2 ␴â,bˆ ␴â,ĉ⬘ 0
Var共g兲 ⫽ D⬘⌺(␪)D ⫹ ⁄ tr兵共H⌺共␪ 兲兲 其 ⫽ b ␴
12 2 2 2
â ␴â,bˆ ␴ˆb
2
␴b,ĉ⬘
ˆ 0
⌺共 ␪ 兲 ⫽ . (A17)
⫹ a2␴ˆb ⫹ ␴â2 ␴ˆb ⫹ 2ab␴â,bˆ ⫹ 共␴â,bˆ兲2 ⫹ ␴␴2ˆ a ,b .
2 2
(A14) ␴â,ĉ⬘ ␴b,ĉ⬘
ˆ ␴ĉ⬘2 0
j j
0 0 0 ␴ 2
␴ˆ aj ,bj

In practice, the sample estimates of these quantities are substituted Thus,

for their population counterparts.
Var共g兲 ⫽ D⬘⌺共␪ 兲D ⫹ 1⁄2tr兵共H⌺共␪ 兲兲2 其
Sampling Variability of the Mean Estimated Total ⫽ b2 ␴â2 ⫹ a2 ␴ˆb ⫹ 2ab␴â,bˆ ⫹ 2b␴â,ĉ⬘ ⫹ 2a␴ˆb,ĉ⬘
2

Effect (Equation 10)

⫹ ␴␴2ˆ a ,b ⫹ ␴ĉ⬘2 ⫹ ␴â2 ␴ˆb ⫹ 共␴â,bˆ兲2 .
2
j j
(A18)
The mean estimated total effect is
In practice, the sample estimates are substituted for their popula-
g ⫽ ab ⫹ c⬘ ⫹ ␴aj ,bj . (A15) tion counterparts.

Using the second-order delta method requires the gradient and Received March 16, 2005
Hessian of g and the asymptotic covariance matrix of parameter Revision received March 1, 2006
estimates. Here, D⬘ ⫽[b a 1 1], Accepted March 1, 2006 䡲