Journal of Operations Management 32 (2014) 99113

Contents lists available at ScienceDirect

Journal of Operations Management

journal homepage: www.elsevier.com/locate/jom

Technical note

Theorizing, testing, and concluding for mediation in SCM research:

Tutorial and procedural recommendations
M. Rungtusanatham , J.W. Miller 1 , K.K. Boyer 2
The Ohio State University, Department of Management Sciences, 2100 Neil Avenue, Columbus, OH 43210, United States

a r t i c l e i n f o a b s t r a c t

Article history: Empirical research in Supply Chain Management is increasingly interested in complex models involving
Received 8 January 2014 mediation effects. We support these endeavors by directing attention to the practices for the theorizing
Accepted 8 January 2014 of, the testing for, and the drawing of conclusions about mediation effects. Our paper synthesizes diverse
Available online 20 January 2014
literature in other disciplines to provide an accessible tutorial as to the mathematical foundation of
mediation effects and the various methods available to test for these effects. We also provide guidance to
Keywords:
SCM scholars in the form of eight recommendations aimed at improving the theorizing of, the testing for,
Mediation
and the drawing of conclusions about mediation effects. Recommendations pertaining to how mediation
Indirect effect
SCM empirical research
effects are hypothesized and stated and how to select among methods to test for mediation effects are
novel contributions for and beyond the Supply Chain Management discipline.
2014 Elsevier B.V. All rights reserved.

1. Introduction
Supply Chain Management (SCM) scholars are increasingly
interested in complex models involving mediation processes or
effects, wherein an independent variable X inuences one or more
mediators (M1 , M2 , . . ., Mp ) which, in turn, inuence a dependent
variable Y. Consider, for example, the topic of supply chain integra-
tion. Frohlich and Westbrook (2001) demonstrated how different
integration patterns (X) relate to operational performance (Y). Since
then, Rosenzweig et al. (2003) have analyzed whether or not, and
how, competitive capabilities mediate the impact of supply chain
integration on operational performance.
Conceptualizing and evaluating models with mediation effects
exemplies research efforts to cultivate a more nuanced scien-
tic understanding of the mechanisms through which X affects Y
(MacKinnon, 2008; Mathieu et al., 2008; Preacher and Hayes, 2008).
These research efforts are deliberate searches to uncover causal
processes, perhaps overlooked when the initial scientic focus was
simply to establish the existence of a relationship between X and
Y (Bollen, 1989). Successful searches result in increased precision
in the theoretical explanations as to how X inuences Y (Edwards
and Berry, 2010). As such, SCM research theorizing, testing, and
concluding for mediation processes can be one fruitful avenue to
Corresponding author. Tel.: +1 614 292 0680.
E-mail addresses: rungtusanatham.1@osu.edu (M. Rungtusanatham),
miller 5350@sher.osu.edu (J.W. Miller), boyer 9@sher.osu.edu (K.K. Boyer).
1
Tel.: +1 614 292 0680.
2
Tel.: +1 614 292 4605.
respond to disciplinary calls for more and better theories about SCM
phenomena (Carter, 2011; Schroeder, 2008), provided that these
endeavors are properly executed.
This provision, however, may not be perfectly accurate as it
relates to SCM research involving mediation processes. In Appendix
A, we summarize the design and discuss the results of an exemplary
(i.e., not exhaustive) review of 81 SCM articles involving mediation
processes that were published, between 20082011, in the Journal
of Business Logistics, the Journal of Operations Management, and the
Journal of Supply Chain Management. Our review highlights three
shortcomings with respect to how SCM research has been theo-
rizing, empirically testing, and concluding for mediation effects.
One shortcoming is that we rarely hypothesize mediation effects
even when our conceptual models, described pictorially or in prose,
depict mediation processes. A second shortcoming is that we often
draw erroneous conclusions about mediation effects based on sta-
tistical results stemming from applying problematic methods or,
more critically, on ad hoc interpretations of statistical results. A
third shortcoming is that when our conceptual models incorporate
multiple (e.g., three) mediation effects, we sometimes draw erro-
neous conclusions about all three mediation effects by relying on
an omnibus test only.
These three shortcomings, we believe, reect an incomplete
exposure by SCM scholars to recent developments regarding the
theorizing and testing of mediation processes. Many SCM scho-
lars are undoubtedly familiar with what mediation is and how to
test for mediation effects via such familiar methods as the Baron
and Kenny (1986) Method, the James et al. (2006) Method, or
the Sobel Test (Sobel, 1982). They are, however, likely to be less
conversant about other methods (e.g., Bootstrapping, Monte Carlo

0272-6963/$ see front matter 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jom.2014.01.002
100 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
Fig. 1. A multiple-mediator reference model.
Simulation, Bayesian Estimation) that are currently being discussed
and debated in journals dedicated to the disciplines of quantita-
tive psychology, quantitative sociology, or statistics and that, more
importantly, are being introduced to overcome inherent limita-
tions in the more familiar methods. Our paper seeks to remedy this
knowledge deciency to ensure that the practices for the theorizing
of, the testing for, and the drawing of conclusions about mediation
effects in SCM research agree with established norms.
To this end, we begin, in Section 2, with a tutorial on what
mediation effects are, what methods there are to test for and draw
conclusions regarding mediation effects, and what the advantages
and disadvantages of these various methods are. Our tutorial syn-
thesizes across classical and recently-published sources, including
Baron and Kenny (1986); Bollen (1989); Bollen and Stine (1990);
Cheung and Lau (2008); Collins et al. (1998); Fritz and MacKinnon
(2007); Fritz et al. (2012); Hayes (2009); Hayes and Preacher
(2010); James et al. (2006); Lau and Cheung (2012); Macho and
Ledermann (2011); MacKinnon et al. (2000, 2002, 2004, 2007);
Preacher and Hayes (2004, 2008); Preacher and Kelly (2011);
Preacher and Selig (2012); Rucker et al. (2011); Shrout and Bolger
(2002); Sobel (1982, 2008); Stone and Sobel (1990); Taylor et al.
(2008); Toghi and MacKinnon (2011); Williams and MacKinnon
(2008); Yuan and MacKinnon (2009). Our tutorial, as such, mimics
Flynn et al. (1990), who synthesized diverse published sources to
introduce survey research into the SCM discipline and, like Flynn
et al. (1990), hopefully summarizes relevant content about medi-
ation effects in a manner that is more readily accessible to SCM
scholars at large.
Section 3 continues the tutorial to offer guidance to SCM scholars
with respect to theorizing, testing, and concluding for mediation
effects. We present this guidance as eight procedural recommenda-
tions three pertaining to how hypotheses about mediation effects
are to be developed and stated, three pertaining to how mediation
effects are to be empirically estimated and tested for, and two per-
taining to how conclusions about mediation effects are to be drawn.
These recommendations, once adopted, overcome the three short-
comings identied in our review of SCM research in Appendix A
and, moreover, strengthen the validity and robustness of statis-
tical results for mediation effects. We acknowledge that many of
our recommendations have been similarly raised by scholars in
communications (Hayes, 2009), management (Wood et al., 2008),
marketing (Zhao et al., 2010), and psychology (Rucker et al., 2011) in
their attempts to guide research within their respective disciplines.
We believe, however, that the recommendations pertaining to how
hypotheses about mediation effects are to be stated in light of the
research agenda and how methods to test for mediation effects are
to be selected reect novel contributions for and beyond the SCM
discipline.
In Section 4, we briey recap the contributions of our paper.
We then conclude by highlighting topics (e.g., comparing the
strength of one mediation effect against another, mediation effects
in multi-level models, etc.) that are not addressed in our treatise
on theorizing, testing for, and drawing conclusions about media-
tion processes and by referring those interested in these topics to
appropriate expertise.
2. Theorizing, testing, and concluding for mediation
effects: a tutorial
2.1. Understanding mediation effects: basic mathematical
representation
As a reference model for this tutorial, consider Fig. 1 depict-
ing an independent variable, X, affecting a dependent variable, Y,
directly (Panel A) and, alternatively, through two parallel mediation
processes, M1 and M2 (Panel B).3 Fig. 1 (Panel B), moreover, assumes
explicitly that:
3
A model with mediators in parallel involves two or more mediators on different
pathways linking X to Y, similar to Fig. 1 (Panel B); one with mediators in series
involves two or more mediators along the same pathway linking X to Y, such as
X M1 M2 Y.
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
1. M1 , M2 , and Y are continuous variables with normally distributed
error terms of e2 , e3 , and e4 respectively;
2. X, M1 , and M2 are observed variables measured without error;
3. Each of the three error terms of e2 , e3 , and e4 has a conditional
expectation of zero;
4. e2 and e3 are correlated; and
5. e2 and e4 are uncorrelated, as are e3 and e4 .
Assumption 1 is stated in order to conne our tutorial to the
case of mediators and outcome variables not being dichotomous,
counts, or censored in nature. Assumption 4 is stated because the
error terms for parallel mediators, e2 and e3 , are likely to be corre-
lated since they are predicted by the same X variable and should be
estimated in order for the presence or absence of mediation effects
to be correctly inferred from statistical results (MacKinnon, 2008;
Preacher and Hayes, 2008). Finally, Assumptions 2, 3, and 5 are
stated to facilitate a discussion about mediation effects without
having to be concerned with the threat of endogeneity.4
The relationships in Fig. 1 can be mathematically represented by
a set of equations. Using the notation summarized in Table 1 con-
sistent with MacKinnon (2008), Eq. (1) expresses the relationship
depicted in Fig. 1 (Panel A) while Eqs. (2)(4) capture the relation-
ships depicted in Fig. 1 (Panel B):
Y = i + cX + e1
M1 = i + a1 X + e2
M2 = i + a2 X + e3

Y = i + c X + b1 M1 + b2 M2 + e4
The parameter c in Eq. (1) for Fig. 1 (Panel A) denotes the total
effect of X on Y, such that a unit change in X results in an expected
c unit change in Y. When X can be theorized to affect Y through the
two mediation processes of M1 and M2 as shown in Fig. 1 (Panel
B), the total effect of X on Y can be decomposed into two portions.
One portion, represented by c in Eq. (4), is the direct effect of X on
Y, such that, holding M1 and M2 constant, a one unit change in X
results in an expected c unit change in Y. The second portion is the
total indirect effect of X on Y (Bollen, 1989), which we denote here
as  X . For Fig. 1 (Panel B),  X can be further disintegrated into two
portions: one for the specic indirect effect of X on Y through M1 and
one for the specic indirect effect of X on Y through M2 .5 Because a
4 Explicit Procedures, refers to methods that conduct formal statis-
Briey, with respect to Fig. 1, endogeneity is said to be present if X were cor-
related with e1 for Fig. 1 (Panel A). For Fig. 1 (Panel B), endogeneity is said to be
present if X were correlated with any of the three error terms of e2 , e3 , or e4 or if M1
or M2 were correlated with e4 . Endogeneity arises (a) when a model omits a variable
(e.g., Z) that is correlated with X and, at the same time, Z is correlated with M1 or
M2 holding X constant; (b) when a model omits a variable (e.g., Z) that is correlated
with X holding M1 and M2 constant and, at the same time, Z is correlated with Y
holding X, M1 , and M2 constant; (c) when a model omits a variable (e.g., Z) that is
correlated with M1 (or with M2 ) holding X and M2 (or M1 ) constant and, at the same
time, Z is correlated with Y holding X, M1 , and M2 constant; (d) when the variables
predicting Y (i.e., X, M1 , or M2 in Fig. 1 (Panel B)) are measured imperfectly; and (e)
when X (or M1 or M2 ) affects Y and simultaneously Y affects X (or M1 or M2 ); or (f)
when X affects M1 (or M2 ) and simultaneously M1 (or M2 ) affects X. If endogeneity
were indeed present in Fig. 1 (Panel B), then conclusions regarding the mediation
effects of M1 or M2 could be invalid (Bascle, 2008; Shaver, 1998).
Whereas Assumption 2 safeguards against the threat of endogeneity due to X, M1 ,
and M2 being measured imperfectly, Assumptions 3 and 5 safeguard against the
threat of endogeneity due to omitted variables. The threat of endogeneity due to
simultaneity is mitigated by the specied structure of the linkages in Fig. 1 (Panel
B). By stating these assumptions explicitly, our tutorial on theorizing, testing, and
concluding for mediation effects no longer confounds mediation effects with endo-
geneity. For a deeper introduction to endogeneity and its implications, please see
Wooldridge (2009).
5
For models with only one mediator, the indirect effect, the total indirect effect,
and the specic direct effect are different concepts with numerically-identical val-
ues.
101
specic indirect effect is a theoretically-justied mediation process
whose magnitude can be estimated (Williams et al., 2009), as the
number of mediators grows in a research model of interest, so does
the number of specic indirect effects to be theoretically justied
and estimated.6 With respect to Fig. 1 (Panel B), the specic indirect
effect of X on Y through M1 is estimated by a1 b1 and that for M2 by
a2 b2 . The total indirect effect of X on Y,  X , and the total effect of X
on Y, c, can then be computed as Eqs. (5) and (6) respectively:
x = a1 b1 + a2 b2 (5)
 
c = c + x = c + [a1 b1 + a2 b2 ] (6)
Note that Eqs. (5) and (6) have two implications in terms of the-
orizing, testing, and concluding for mediation effects. For one, if
the direct effect of X on Y, c , is signicant and positive (i.e., c > 0)
while the total indirect effect of X on Y through one or more medi-
ation processes,  X , is signicant and negative (i.e.,  X < 0) or vice
versa, then the total effect, c, may conceivably not differ from zero.
MacKinnon et al. (2000) refer to this phenomenon as inconsistent
mediation. For another, if the sign of one specic indirect effect is
signicant and positive (e.g., a1 b1 > 0) while the sign of a second
specic indirect effect is signicant and negative (e.g., a2 b2 < 0) or
vice versa, then erroneous conclusions about mediation effects, in
aggregate, may arise from examining  X only (Collins et al., 1998).
(1) Hence, in models in which X inuences Y through more than one
mediator, not only must the specic indirect effect for each medi-
(2)
ator be theorized (Williams et al., 2009) but each specic indirect
(3) effect must be statistically evaluated before conclusions be drawn
as to the presence or absence of mediation effects.
(4)
2.2. Testing for mediation effects: from traditional to recent
methods
The empirical testing of mediation effects involves procedures
that can be classied into two broad categories. The rst category,
labeled as Implicit Procedures, refers to the more traditional and
familiar methods of the Baron and Kenny (1986) Method and the
James et al. (2006) Method; both infer the presence (or absence)
of mediation processes by examining the joint signicance of var-
ious constituent paths connecting X to Y (Cohen and Cohen, 1983;
MacKinnon et al., 2002). For Fig. 1 (Panel B), the mediation effect
of M1 is deemed to be established when the direct path from X to
M1 (i.e., a1 ) is signicant and the direct path from M1 to Y is sig-
nicant (i.e., b1 ) is also signicant. The second category, labeled as
tical tests for the presence (or absence) of mediation effects and
include the Sobel Test (Sobel, 1982), the Distribution of Product
Method (MacKinnon et al., 2002, 2007; Williams and MacKinnon,
2008), Bootstrapping (Bollen and Stine, 1990; Shrout and Bolger,
2002), Monte Carlo Simulation (MacKinnon et al., 2004; Preacher
and Selig, 2002), and Bayesian Estimation (Yuan and MacKinnon,
2009). With respect to Fig. 1 (Panel B), the mediation effect of M1
is determined by testing formally whether or not a1 b1 differs from
zero.
As a quick reference, we offer two tables regarding the various
methods to test for mediation effects. Table 2 provides a summary
6
According to Mathieu and Taylor (2006), M1 (or M2 ) is said to have a mediation
effect when the bivariate relationship between X and Y is signicant, while M1 (or
M2 ) is said to have an indirect effect when the bivariate relationship between X and Y
is not. This nomenclature distinction as to whether M1 (or M2 ) is a mediation effect
versus an indirect effect has been challenged by Hayes (2009), who demonstrate
that the signicance of the bivariate relationship between X and Y is not relevant
when discussing mediation processes. We concur with Hayes (2009) and, there-
fore, use the terms indirect effect, mediation effect, and mediation process
interchangeably.
102 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113

Table 1
Mediation effects: basic mathematical representation for Fig. 1.

Parameter (where in Fig. 1?) What it represents What it means?

c Total effect of X on Y A one unit change in X leads to an expected c unit change in Y

c Direct effect of X on Y when M1 and M2 mediators are A one unit change in X leads to an expected c unit change in Y, holding
hypothesized M1 and M2 constant
a1 b1 Specic indirect effect of X on Y through the M1 mediator A one unit change in X leads to an expected a1 b1 change in Y by X
inuencing M1 , holding M2 constant
a2 b2 Specic indirect effect of X on Y through the M2 mediator A one unit change in X leads to an expected a2 b2 change in Y by X
inuencing M2 , holding M1 constant
a1 b1 + a2 b2 Total indirect effect of X on Y through both the M1 A one unit change in X inuences both M1 and M2 which, in turn, leads
mediator and the M2 mediator, or  X to an expected  X , or a1 b1 + a2 b2 , change in Y

Table 2
Methods to test for mediation effects: advantages and disadvantages.

Method Advantages Disadvantages

BK Method Can be easily implemented with regression analysis or
using Structural Equation Modeling (SEM) software
packages
JMB Method Can be easily implemented using SEM software packages
Can accommodate multiple mediators in parallel or in a
series
Sobel Test Can be easily implemented using SEM software packages
Can accommodate multiple mediators in parallel or in a
series
Can be used to test contrastsa
DOP Method Can accommodate multiple mediators in parallel
Corrects for the non-normality of the sampling
distribution of a specic indirect effect
Bootstrapping Can accommodate multiple mediators in parallel or in a
series
Corrects for the non-normality of the sampling
distribution of a specic indirect effect
Offers a greater degree of exibility to test for contrasts
Monte Carlo Simulation Same as those for Bootstrapping
Bayesian Estimation Same as those for Bootstrapping
Can incorporate priors to improve precision of parameter
estimation
Accommodates testing of mediation effects with small
samples
a
Contrasts are statistical comparisons of specic indirect effects linking X to Y (MacKinnon, 2000; Preacher and Hayes, 2008). For Fig. 1 (Panel B), assuming that a1 b1 and
a2 b2 have the same sign, whether a1 b1 = / a2 b2 is a contrast that can be statistically compared.
b
Bias-Corrected Bootstrapping to test for mediation effects can result in elevated Type I error rates in small sample sizes when the magnitude of one path of a specic
indirect effect (e.g., the X M1 for Fig. 1 (Panel B) or a1 ) is zero and the magnitude of the other path in the specic indirect effect (e.g., M1 Y for Fig. 1 (Panel B) or b1 ) is
moderate to large (see Fritz et al., 2012).
Has low statistical power
Does not provide a quantication of a specic indirect effect
Does not directly test the signicance of a specic indirect effect
Does not accommodate models with inconsistent mediation
Can be cumbersome with multiple mediators in parallel or in a
series
Does not provide a quantication of a specic indirect effect
Does not directly test the signicance of a specic indirect effect
Has low statistical power due to the assumption that the
sampling distribution of a specic indirect effect is normal,
especially in small samples
Cannot be utilized to test contrasts
Is not as widely implemented in statistical software
May result in elevated Type I error rates, in the case of
Bias-Corrected Bootstrapping under specic conditionsb
Cannot be utilized to test for mediation effects in multi-level
models
Requires knowledge of Monte Carlo Simulation
Requires knowledge of Bayesian statistics and inference
Can result in biased parameter estimates when
overly-informative priors are used

of the advantages and the disadvantages of each method. In addi-
tion, Table 3 provides step-by-step instructions for implementing
each method in order to test for the mediation effects depicted in
Fig. 1 (Panel B).
2.2.1. Implicit procedures
2.2.1.1. The Baron and Kenny (1986) Method. The Baron and Kenny
(1986) Method (henceforth, BK Method) is based on the original
guidelines provided in Judd and Kenny (1981). Since its introduc-
tion, the BK Method has been the predominant procedure for the
empirical testing of mediation effects in the social sciences (Zhao
et al., 2010). Bhattacharyya et al. (2010); Ellinger et al. (2010); and
Hill et al. (2009) are examples of SCM research that have deployed
the BK Method to test for mediation effects.
For Fig. 1, to determine whether or not M1 mediates the rela-
tionship between X and Y, the BK Method estimates and interprets
three regression models in four steps. First, Y is regressed on X to
obtain an estimate of c (i.e., estimate Eq. (1)); c must be signi-
cant in order to proceed. Second, M1 is regressed on X to obtain
an estimate of a1 (i.e., estimate Eq. (2)); a1 must be signicant
in order to proceed. Third, Y is regressed on X and M1 to obtain
estimates of c and b1 (i.e., estimate Eq. (4) but without M2 ); b1
must be signicant in order to proceed. Fourth, c is evaluated rel-
ative to c; if c is non-signicant, then the relationship between X
and Y is said to be completely mediated by M1 ; otherwise, if c is
signicant but |c | < c, then the relationship between X and Y is said
to be partially mediated by M1 . These four steps are then repeated
in order to determine whether or not the M2 mediation effect is
present (absent) for the inuence of X on Y.7
Despite its popularity, the BK Method has a number of limita-
tions. For one, because each hypothesized mediation effect in Fig. 1
(Panel B) has to be analyzed separately, the BK Method is cum-
bersome when the relationship between X and Y involves many
mediation processes (Preacher and Hayes, 2008). For another, the
BK Method does not quantify the magnitude of the mediation effect
(Preacher and Kelly, 2011), nor can the BK Method accommodate
7
Note that Kenny et al. (1998) relaxed the requirement that c must be signicant
when testing for the presence (absence) of mediation processes linking X to Y. How-
ever, per MacKinnon et al. (2002), empirical testing for mediation effects via the BK
Method continues to comply with the original stipulation that the total effect of X
on Y, c, must be signicant in order for mediation effects to exist.
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113 103

Table 3
Methods to test for mediation effects in Fig. 1 (Panel B): step-by-step instructions.

BK Method
Step 1 Estimate Eq. (1) to obtain c; if c is non-signicant, stop.
Step 2 Estimate Eq. (2) to obtain a1 for M1 ; if a1 is non-signicant, stop.
Step 3 Estimate Eq. (4) but without M2 to obtain b1 ; if b1 is non-signicant, stop.
Step 4 Examine c from estimating Eq. (4) in the previous step; if c is non-signicant, conclude in favor of complete mediation by
M1 ; if c is signicant but |c | < |c|, conclude in favor of partial mediation by M1 .
Step 5 Repeat steps 24 for M2 .
JMB Method
Step 1 Specify, a priori, whether complete mediation or partial mediation is to be expected.
Step 2 If complete mediation is hypothesized in Step 1, estimate Eqs. (2)(4) simultaneously using SEM software but with c
constrained to zero; if partial mediation is hypothesized in Step 1, estimate Eqs. (2)(4) using SEM software but with c freely
estimated.
Step 3 Examine the signicance of a1 , a2 , b1 , and b2 ; all must be signicant for M1 and M2 to have mediation effects.
Step 4 If complete mediation is hypothesized in Step 1, examine the model 2 and conclude in favor of (against) complete mediation
by M1 and M2 if the model 2 is non-signicant (signicant); if partial mediation is hypothesized in Step 1, conclude in favor
of (against) partial mediation if c is signicant (non-signicant).
Sobel Test
Step 1 Estimate Eqs. (2)(4) simultaneously using SEM software; alternatively, estimate Eqs. (2)(4) separately.
Step 2 Quantify the specic indirect effect of M1 by multiplying a1 and b1 .
Step 3 Compute Sa1 b1 , the standard error for the point estimate of the specic indirect effect of M1 , or a1 b1 ; see MacKinnon et al.
(2002) for details on the different formulae for computing the standard error of a specic indirect effect.
Step 4 Test for the signicance of a1 b1 using a z-test or by examining whether or not the condence interval of a1 b1 contains 0;
mediation by M1 is supported if the z-value of a1 b1 (a1 b1 Sa1 b1 ) > |critical value| or if the condence interval for a1 b1 does
not contain 0; repeat Steps 24 for a2 b2 and the mediation effect of M2 .
DOP Methoda
Step 1 Estimate Eqs. (2)(4) simultaneously using SEM software or, alternatively, estimate Eqs. (2)(4) separately to obtain a1 , a2 , b1 ,
and b2 and their respective standard errors of Sa1 , Sa2 , Sb1 , and Sb2 .
Step 2 Test for the signicance of a1 b1 by examining whether or not the adjusted condence interval of a1 b1 contains 0; mediation
by M1 is supported if the condence interval for a1 b1 does not contain 0; repeat Step 2 for a2 b2 and the mediation effect of M2 .
Bootstrappingb
Step 1 Randomly draw k bootstrap samples of sample size N with replacement from the data of an original sample.
Step 2 For each kth bootstrap sample, estimate Eqs. (2)(4) separately or simultaneously via SEM software to obtain a1 b1 .
Step 3 Derive an empirical sampling distribution for a1 b1 based on the k estimates of a1 b1 from Step 2.
Step 4 Test for the signicance of a1 b1 by examining whether or not interval bounded by a specied lower percentage and a
specied upper percentage for the empirical distribution of a1 b1 contains 0; mediation by M1 is supported if the condence
interval for a1 b1 does not contain 0; repeat Steps 24 for a2 b2 and the mediation effect of M2 .
Monte Carlo Simulationc
Step 1 Estimate Eqs. (2)(4) simultaneously using SEM software or, alternatively, estimate Eqs. (2)(4) separately to obtain a1 , a2 , b1 ,
and b2 and their respective standard errors of Sa1 , Sa2 , Sb1 , and Sb2 .
Step 2 Simulate k samples of a1 from a normal distribution with mean = a1 (from Step 1) and standard deviation = Sa1 (from Step 1)
to generate a11 , a12 , . . ., a1k ; simulate k samples of b1 from a normal distribution with mean = b1 (from Step 1) and
standard deviation = Sb1 (from Step 1) to generate b11 , b12 , . . ., b1k .
Step 3 Multiply (a11 b11 ), (a12 b12 ), . . ., (a1k b1k ) and use these multiplication values to derive an empirical sampling
distribution for a1 b1 .
Step 4 Test for the signicance of a1 b1 by examining whether or not the interval bounded by a lower cutoff percentile (2.5th) and an
upper cutoff percentile (97.5th) contains 0; mediation by M1 is supported if the interval for a1 b1 does not contain 0; repeat
Steps 24 for a2 b2 and the mediation effect of M2 .
Bayesian Estimationd
Step 1 Specify informative or diffuse priors for all model parameters in Fig. 1 (Panel B), including a1 , a2 , b1 , b2 , c , e2 , e3 , e4 , the
covariance between e2 and e3 , and the variance of X.
Step 2 Deploy the Gibbs sampler in conjunction with Markov Chain Monte Carlo to draw a large number of samples for each model
parameter.
Step 3 Retain every nth sample and compute the desired mediation effects of a1 b1 and a2 b2 ; use the respective sets of computed
values to approximate 95% credibility intervals for a1 b1 and a2 b2 .
Step 4 Examine whether or not the 95% credibility interval for a1 b1 includes 0; if not, mediation by M1 is support; repeat for a2 b2
and the mediation effect of M2 .
a
The adjusted condence intervals for a1 b1 and a2 b2 can be readily obtained by inputting the estimates of a1 , a2 , b1 , and b2 and their respective standard errors of Sa1 ,
Sa2 , Sb1 , and Sb2 into PRODCLIN (MacKinnon et al., 2007) or into RMediation (Toghi and MacKinnon, 2011);
b
INDIRECT is a tool offered by Preacher and Hayes (2008) to facilitate the deployment of Bootstrapping to test for mediation effects using regression analysis. Moreover,
Appendix B documents the annotated syntax in Mplus to implement Bias-Corrected Bootstrapping to test for mediation effects given a conceptual model analogous to Fig. 1
(Panel B).
c
Preacher and Selig (2012) provide an online calculator that implements testing for mediation effects via Monte Carlo Simulation.
d
Bayesian Estimation to test for mediation effects for a model analogous to Fig. 1 (Panel B) can be implemented in Mplus via the annotated syntax documented in Appendix
C.

models with inconsistent mediation (MacKinnon et al., 2000). More
critically, because of the requirement that the relationship between
X and Y has to statistically exist (i.e., that c is signicant), the BK
Method suffers from low statistical power (Fritz and MacKinnon,
2007; MacKinnon et al., 2002).
2.2.1.2. The James et al. (2006) Method. The James et al. (2006)
Method (henceforth, JMB Method) is an extension of the James and
Brett (1984) method. Examples of SCM research deploying the JMB
Method to test for mediation effects are Paulraj et al. (2008) and
Kristal et al. (2010).
For Fig. 1 (Panel B), the JMB Method begins by specifying, a pri-
ori, either complete mediation or partial mediation (i.e., M1 and M2
are both either complete mediators or partial mediators). Next, if
complete mediation is hypothesized a priori, a structural equation
model is then tted to data to estimate Eqs. (2)(4) but with c in
Eq. (4) constrained to zero. Complete mediation by both M1 and
M2 simultaneously is then supported when the constituent paths
104 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
connecting X, M1 , M2 , and Y (i.e., a1 , a2 , b1 , and b2 ) are all signicant
and the model 2 is non-signicant.8 Alternatively, if partial medi-
ation is hypothesized a priori, a fully saturated structural equation
model is tted to estimate Eqs. (2)(4) with c allowed to be freely
estimated. Partial mediation by both M1 and M2 simultaneously is
then supported when a1 , a2 , b1 , and b2 are all signicant and c is
signicant.9
Compared to the BK Method, the JMB Method is better able
to simultaneously evaluate models with multiple mediation pro-
cesses, whether in parallel or in series. More importantly, because
the JMK Method does not stipulate the existence of a relationship
between X and Y as a necessary condition, it has more statistical
power than the BK Method (Mathieu and Taylor, 2006). However,
like the BK Method, the JMB Method does not provide numerical
estimates of the magnitude of the constituent mediation effects.
2.2.2. Explicit Procedures
2.2.2.1. The Sobel Test. The Sobel Test (Sobel, 1982) is a formal
method to testing for the signicance of a mediation effect that
has become the default in many SEM software, including AMOS,
EQS, LISREL, and Mplus (MacKinnon, 2008).10 Zhou et al. (2011) is
a recent example of SCM research that deployed the Sobel Test to
examine mediation effects.
With respect to Fig. 1 (Panel B), the Sobel Test proceeds by deriv-
ing the standard error for a specic indirect effect, such as a1 b1 for
the M1 mediation process. The standard error of a1 b1 , denoted as
Sa1 b1 , is approximated as follows:11
 one disadvantage vis--vis the Sobel Test is that the DOP Method
Sa1 b1 = a21 Sb2 + b21 Sa21 (7)
1 there are relatively limited implementations of the DOP Method in
Dividing the M1 specic indirect effect, a1 b1 , by Sa1 b1 computes
a z-value that inputs into a z-test or facilitates constructing the con-
dence interval for a1 b1 at a given level of signicance. The z-test
result or the condence interval is then interpreted to support the
presence (absence) of the M1 mediation effect. The steps are then
repeated to determine the presence (absence) of the M2 mediation
effect.
The Sobel Test, besides being readily available in SEM software,
has the advantage of being able to accommodate models with
multiple mediation processes that are in parallel or in a series.
It, moreover, can be used to conduct tests of contrasts compar-
ing the strength of one specic indirect effect (e.g., a1 b1 ) against
another (e.g., a2 b2 ), to see if the different specic indirect effects are
equal in magnitude or if one is greater than another (MacKinnon,
8
When a1 , a2 , b1 , and b2 are all signicant and the 2 is also signicant, complete
mediation by both M1 and M2 simultaneously is rejected. This rejection, however,
does not automatically support the case of both M1 and M2 simultaneously being
partial mediators. Instead, Eqs. (2)(4) have to be re-estimated via SEM but with c
freely estimated.
9
Because the structural equation model of partial mediation effects being esti-
mated has no degrees of freedom, a model 2 cannot be computed.
10
In a model with just one mediator, M1 , connecting X to Y, the Sobel Test is a
formal test of the total indirect effect of X on Y through the M1 mediator and, at the
same time, is a formal test of the specic indirect effect of X on Y through the M1
mediator.
 2
11
In general, besides Sa1 b1 = a21 Sb2 + b1 Sa21 offered by Sobel (1982), there are
1
two other formulations to compute the standard error for the specic indirect
 is offered by Aroian (1947) and
effect represented by a1 b1 . One formulation
2
computes an exact standard error, Sa1 b1 = a21 Sb2 + b1 Sa21 + Sa21 + Sb2 ; the Aroian
1 1
(1947) formulation is popularized by Baron and Kenny 
(1986) as the Sobel Test.
2 dence intervals for a specic indirect effect of interest to facilitate the drawing of
Goodman (1960) provides another formulation, Sa1 b1 = a21 Sb2 + b1 Sa21 Sa21 Sb2 ,
1 1
which yields an unbiased estimate of the standard error. Hayes and Scharkow (in
press) and MacKinnon et al. (2002) demonstrate through simulation studies that
all three formulations display similar performance. Since many SEM software com-
pute the standard error without Sa21 Sb2 , references to the Sobel Test in this paper
1
will point specically to the Sobel (1982) formulation.
2000; Preacher and Hayes, 2008). A disadvantage of the Sobel Test,
however, is the assumption that the sampling distribution of a spe-
cic indirect effect, whether a1 b1 or a2 b2 , is normally distributed
(MacKinnon et al., 2002). With small sample sizes, this assumption
is typically violated (Aroian, 1947; Craig, 1936), producing Sobel
Test results that have been demonstrated to be low in statistical
power (Cheung and Lau, 2008; MacKinnon et al., 2002, 2004; Taylor
et al., 2008).12
2.2.2.2. The Distribution of Product Method. The Distribution of
Product Method (henceforth, DOP Method) is a method explained
in MacKinnon et al. (2002, 2004, 2007) to account for the non-
normal sampling distribution of a specic indirect effect.13 We are
unaware of denitive examples of SCM research showcasing the
DOP Method to test for mediation effects.
Procedurally, the DOP Method is similar to the Sobel Test in that
it begins by obtaining a1 , a2 , b1 , b2 , and their respective standard
errors of Sa1 , Sa2 , Sb1 , and Sb2 , for Fig. 1 (Panel B) via separate
regression models or simultaneously using SEM software. A con-
dence interval for a specic indirect effect of interest (e.g., a1 b1 )
is then constructed and interpreted to determine if it contains zero.
The procedural difference between the DOP Method and the Sobel
Test is that an adjustment is made in constructing the condence
interval to account for the sampling distribution of a1 b1 being non-
normal.
Like the Sobel Test, the DOP Method is able to accommodate
models with multiple mediation processes in parallel. However,
cannot conduct tests of contrasts. Moreover, with noted exceptions,
statistical packages.
2.2.2.3. Bootstrapping. Bollen and Stine (1990) pioneered the
deployment of Bootstrapping to analyze mediation processes. Since
then, MacKinnon et al. (2004), Preacher and Hayes (2004, 2008),
and Shrout and Bolger (2002) have contributed to the popularity
of Bootstrapping as a method for examining mediation effects.14
A recent example of SCM research illustrating how to bootstrap to
test for mediation effects is Saldanha et al. (2013).
For Fig. 1 (Panel B), Bootstrapping to test for mediation effects
proceeds as follows. First, k bootstrap samples of sample size N
with replacement are randomly from the data of an original sample,
where N is identical across the k bootstrap samples and is equal to
the sample size of the original sample (Efron and Tibshirani, 1993).
For each kth bootstrap sample, Eqs. (2)(4) are estimated to calcu-
late a specic indirect effect of interest, such as a1 b1 . Second, the k
estimates of a1 b1 are used to generate an empirical sampling distri-
bution. Third, the signicance of a1 b1 is determined by examining
12
The Sobel Test results produced by LISREL or AMOS from analyzing models with
two or more parallel mediators pertain to the total indirect effect of X on Y through
all mediation processes, as opposed to through a specic mediation process (Macho
and Ledermann, 2011). In other words, for Fig. 1 (Panel B), the LISREL-generated or
AMOS-generated Sobel Test results address the signicance of  X in Eq. (5) and not
the signicance of a1 b1 or a2 b2 . The signicance of the specic indirect effects of
a1 b1 or a2 b2 can be determined by following the procedure in Bollen (1989).
13
For ease of implementation of the DOP Method, the adjusted condence inter-
vals can be constructed by inputting point estimates of the constituent paths of a
specic indirect effect and corresponding standard errors into PRODCLIN, an add-
on procedure into SAS, SPSS, and R (MacKinnon et al., 2007) or into RMediation
(Toghi and MacKinnon, 2011). These tools provide the critical values for the con-
appropriate inferences about the presence of absence of mediation effects.
14
INDIRECT is a tool offered by Preacher and Hayes (2008) to facilitate the
deployment of Bootstrapping to test for mediation effects using regression analy-
sis. Moreover, Appendix B documents the annotated syntax in Mplus to implement
Bias-Corrected Bootstrapping to test for mediation effects given a conceptual model
analogous to Fig. 1 (Panel B).
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
whether or not an interval bounded by a specied lower percent-
age and a specied upper percentage for the empirical sampling
distribution of a1 b1 contains zero. One way to specify the critical
percentages (aka Percentile Bootstrapping) is to compute the 95%
condence interval-based percentile values of /2 and (1 )/2,
where = Type 1 error. An alternative way (aka Bias-Corrected
Bootstrapping) is to adjust for the non-symmetric shape of the
empirically-derived sampling distribution by shifting the critical
percentile values of the 95% condence interval (Efron, 1987; Lau
and Cheung, 2012; MacKinnon et al., 2004). The appropriate steps
are then repeated for a2 b2 .
Bootstrapping, as a method to test for mediation effects, has
several advantages. Not only can it correct for the non-normality of
the sampling distribution of a specic indirect effect but it can also
accommodate models with multiple mediation processes in paral-
lel or in a series. Bootstrapping, moreover, offers exibility to test
for specic indirect effects, as well as for contrasts (Lau and Cheung,
2012; MacKinnon, 2000; Preacher and Hayes, 2008). In fact, Boot-
strapping to test for mediation effects has been demonstrated to
have the greatest statistical power to detect signicant mediation
processes while maintaining acceptable Type I error rates (Cheung
and Lau, 2008; MacKinnon, 2004; Taylor et al., 2008). One, albeit
very nuanced, disadvantage relates particularly to Bias-Corrected
Bootstrapping. Bias-Corrected Bootstrapping has been shown to
result in elevated Type I error rates in small sample sizes when
one constituent path in a specic indirect effect (e.g., a1 for Fig. 1
(Panel B)) is non-signicant and the other constituent path in the
same specic indirect effect (e.g., b1 in Fig. 1 (Panel B)) is not only
signicant but moderate to large in magnitude (Fritz et al., 2012).
2.2.2.4. Monte Carlo Simulation. Monte Carlo Simulation tests for
mediation effects by producing and evaluating an empirical samp-
ling distribution for a specic indirect effect (MacKinnon et al.,
2004).15 In this regard, Monte Carlo Simulation and Bootstrapping
are similar methods. However, unlike Bootstrapping, Monte Carlo
Simulation produces an empirical sampling distribution for a spe-
cic indirect effect (e.g., a1 b1 ) indirectly. We did not nd examples
of SCM research using Monte Carlo Simulation to test for mediation
effects.
For Fig. 1 (Panel B), Step 1 in Monte Carlo Simulation is to esti-
mate Eqs. (2)(4) using SEM software or separately via regression
to obtain a1 , a2 , b1 , and b2 and their respective standard errors of
Sa1 , Sa2 , Sb1 , and Sb2 . For the a1 b1 specic indirect effect, Step 2 sim-
ulates k samples of a1 from a normal distribution, with mean = a1
and standard error = Sa1 from Step 1, to generate a11 , a12 , . . ., a1k ;
this is repeated to simulate k samples of b1 from a normal dis-
tribution, with mean = a1 and standard error = Sa1 from Step 1, to
generate b11 , b12 , . . ., b1k . Step 3 multiplies the simulated a1 and b1
values; these product values are then used to derive an empirical
sampling distribution for a1 b1 . Step 4 determines the signicance
of a1 b1 by examining whether or not an interval bounded by the
2.5th percentile and 97.5th percentile contains zero, with M1 hav-
ing a mediation effect if the interval does not contain zero. The steps
are then repeated to examine the a2 b2 specic indirect effect for
M2 .16
As a method to test for mediation effects, Monte Carlo Sim-
ulation enjoys the same advantages as Bootstrapping (i.e., can
accommodate complex models with multiple mediators, is not
15
Preacher and Selig (2012) provide an online calculator to implement the testing
of mediation effects via Monte Carlo Simulation.
16
For Fig. 1 (Panel B), if there is interest in statistically comparing the strength of
the M1 mediation effect and that of the M2 mediation effect, besides their presence,
then Step 1 of Monte Carlo Simulation should also compute the covariance among
a1 , a2 , b1 , and b2 (Lau and Cheung, 2012; Preacher and Selig, 2012).
105
affected by the non-normality of the sampling distribution for the
specic indirect effect of interest, and can exibly conduct tests of
contrasts). The disadvantage with this method to test for mediation
effects, however, is that it requires sufcient user knowledge about
simulation generally and Monte Carlo Simulation specically.
2.2.2.5. Bayesian Estimation. Yuan and MacKinnon (2009)
introduced and demonstrated Bayesian Estimation as a method to
test for mediation effects. Two examples of Bayesian Estimation
to evaluate mediation processes are Zhang et al. (2009) from
marketing and Pirlott et al. (2012) from nutrition; no examples
of SCM research deploying Bayesian Estimation for analyzing
mediation processes are available.
Bayesian Estimation as a method to test for mediation effects
differs from other methods discussed above in the fundamen-
tal manner of statistical inference.17 Statistical inference refers
generally to the process by which conclusions about population
parameters are made given a random sample of data from the pop-
ulation of interest. The methods described above are associated
with a frequentist perspective, which treats population parameters
as having unknown but xed values. Based on this perspective, a
percentage-based interval (e.g., 95% condence interval) for an esti-
mate of the population parameter from sample data means that
when repeated samples are drawn from the population and a con-
dence interval is computed for the estimate of the population
parameter in each sample, then, on average, 95% of these computed
intervals contains the true value for the population parameter. In
contrast, Bayesian Estimation as a method for analyzing mediation
effects is associated with the Bayesian perspective, which treats
population parameters as random variables. Based on this perspec-
tive, a percentage-based interval (e.g., 95% credibility interval) for
a population parameter signals that there is a percentage-based
probability (i.e., 95%) of observing the true value of the population
parameter within the upper and lower bounds of the computed
credibility interval.
Given Fig. 1 (Panel B), Bayesian Estimation to analyze media-
tion processes, described most generally, proceeds as follows. Step
1 determines whether or not to incorporate prior knowledge about
each of the 10 model parameters in Fig. 1 (Panel B): a1 , a2 , b1 , b2 ,
c , e2 , e3 , e4 , the covariance between e2 and e3 , and the variance
of X. For example, if strong evidence exists to suggest that a1 has
a standardized estimate of 0.30 (i.e., a1 has an informative prior
probability distribution or simply informative prior), then the prior
probability distribution of a1 can be informatively specied to have
a mean of 0.30 and a standard deviation based on the standard
errors of past estimates. If an informative prior for a particular
model parameter is not available, then a diffuse prior, such as a
mean of 0.00 and an arbitrary standard deviation, can be speci-
ed. Step 2 deploys the Gibbs sampler (Casella and George, 1992;
Geman and Geman, 1984) in conjunction with Markov Chain Monte
Carlo (Edwards, 2010) to draw a large number of samples (e.g.,
100,000) for each model parameter. The large number of samples is
to allow subsequent manipulations to discard the rst portions of
a Markov chain which may not represent steady-state conditions.
Step 3 retains only every nth sample and computes the desired
mediation effects of a1 b1 and a2 b2 ; the sets of computed values for
a1 b1 and a2 b2 are then used to approximate their respective credi-
bility intervals and facilitate statistical inference. Finally, in Step 4,
the 95% credibility interval for a1 b1 is examined, with the absence
of zero in the credibility interval indicating a 95% probability that
that a1 b1 = / 0 and that M1 has a mediation effect. Step 4 is then
17
For a more detailed but non-technical treatment of statistical inference from a
Bayesian perspective, please refer to Zyphur and Oswald (in press).
106 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113

Table 4
Recommendations for theorizing of, testing for, and drawing conclusions regarding mediation processes.

When theorizing for mediation processes

Recommendation 1 Remember to hypothesize explicitly for mediation effects whenever the inuence of X (an independent variable)
on Y (a dependent variable) is discussed or depicted as being transmitted through one or more mediators, M1 , M2 ,
. . ., Mp .
Recommendation 2 Either
i. Develop hypotheses for the effect of X on M1 and the effect of M1 on Y, and conclude by stating the hypothesis
for the mediation effect of M1 (i.e., the Segmentation Approach); repeat for other mediators or
ii. Develop the hypothesis that M1 mediates the effect of X on Y or that X has an indirect effect on Y through M1
without needing to articulate hypotheses relating X to M1 and M1 to Y (i.e., the Transmittal Approach); repeat for
other mediators
With either approach, hypothesize the direct effect of X on Y when other mediators, besides M1 , may theoretically
or empirically explain the total effect of X on Y or when the direct effect of X on Y may have a sign that is opposite
of the sign for a specic indirect effect.
Recommendation 3 Avoid the nomenclature of complete mediation or partial mediation when hypothesizing mediation processes.
When testing for mediation processes
Recommendation 4 Include the direct effect of X on Y when testing for mediation effects.
Recommendation 5 Abandon the BK Method when testing for mediation effects.
Recommendation 6 Choose Explicit Procedures (i.e., Bayesian Estimation, Bootstrapping, Monte Carlo Simulation, and the Distribution
of Product Method over Implicit Procedures (i.e., the BK Method, the JMB Method, or the Sobel Test) to test for
mediation effects.
Before drawing conclusions for or against mediation processes
Recommendation 7 Interpret non-signicant Sobel Test results cautiously because the assumption of normality for the sampling
distribution of a specic indirect effect is often violated.
Recommendation 8 Examine the signicance of specic indirect effects above and beyond the total indirect effect.

repeated to determine whether or not M2 has a mediation effect.18
Appendix C shows the annotated syntax in Mplus to implement
Bayesian Estimation for analyzing mediation effects.
Bayesian Estimation as a method to analyze mediation effects
has similar advantages as those for Bootstrapping. It can, moreover,
analyze mediation effects in small samples or models involving
mediation effects across levels (Scheines et al., 1999; Yuan and
MacKinnon, 2009). Bayesian Estimation is also able to improve
the precision of estimates about model parameters by incorpo-
rating prior information about model parameters (Muthn and
Asparouhov, 2012; Yuan and MacKinnon, 2009), as long as overly-
informative priors are not used since these can bias model
parameter estimates. However, Bayesian Estimation as method to
analyze mediation effects is extremely complex and, under specic
circumstances (e.g., when sample sizes are large, when diffuse pri-
ors are utilized, or when priors cannot be informatively specied),
may yield results that are no different from those based on the more
familiar, maximum-likelihood estimation (Rupp et al., 2004).
3. SCM research involving mediation effects: procedural
recommendations
To ensure that the theorizing of, the testing for, and the draw-
ing of conclusions about mediation effects in SCM research comply
with established norms, we offer the eight recommendations in
Table 4. The rst three recommendations, which pertain to the task
of theorizing for mediation effects, (1) emphasize the importance
of hypothesizing for mediation effects, (2) provide advice as to the
manner by which hypotheses for mediation effects are to be devel-
oped, and (3) argue against hypothesizing mediation effects via the
nomenclature of complete versus partial mediation. The next three
recommendations, which concern the task of testing for mediation
effects, (4) caution against excluding the estimation of the direct
effect of X on Y (i.e., c in Eq (4)), (5) encourage abandoning the
BK Method, and (6) advocate for the choice of Explicit Procedures
over Implicit Procedures. The last two recommendations, which
18
Our description of Bayesian Estimation to analyze mediation effects is very gen-
eral and, admittedly incomplete. For a detailed, more technical treatment of this
topic, we suggest Yuan and MacKinnon (2009); for a better appreciation of Markov
Chain Monte Carlo, we suggest Edwards (2010); and for technical information about
the Gibbs sampler, we suggest Casella and George (1992).
pertain to the task of drawing conclusions about mediation effects,
(7) qualify the interpretation of non-signicant Sobel Test results
and (8) argue for the evaluation and reporting of the signicance of
specic indirect effects above and beyond the total indirect effect.
These eight recommendations, we trust, effectively overcome the
three shortcomings in SCM research involving mediation processes
that are presented in Appendix A.
3.1. Theorizing for the presence of mediation effects
3.1.1. Remember to hypothesize explicitly for mediation effects
Our rst recommendation is for SCM scholars examining
complex models involving mediation processes to remember to
hypothesize explicitly for mediation effects of interest. This rec-
ommendation may seem obvious and unnecessary since there is
common understanding among SCM researchers as to mediation
effects being theoretical processes that describe, explain, and pre-
dict how X inuences Y through one or more mediators (Williams
et al., 2009). However, as our exemplary review of articles pub-
lished in JBL, JOM, and JSCM between 2008 and 2011 in Appendix A
nds, this common understanding is not consistently being trans-
lated into research practice; a high percentage of published SCM
research (75% of articles) did not engage in hypothesizing for
mediation effects despite invoking mediation processes in prose or
in diagrammatic form. Recommendation 1 is, therefore, offered not
as new advice but as a reminder for SCM scholars to formally engage
in a theorizing exercise to formulate hypotheses about mediation
effects before proceeding to test for and draw conclusions about
them.
3.1.2. Hypothesize mediation effects via the segmentation or the
transmittal approach
Recommendation 2 concerns how formal hypotheses for medi-
ation effects are developed and articulated. Referring back to Fig. 1
(Panel B), if the intent of the research is to examine (i) the effects of
X on M1 and M2 , (ii) the effect of M1 on Y in the presence of X and
holding M2 constant, (iii) the effect of M2 on Y in the presence of X
and holding M1 constant, and (iv) the effect of X on Y while holding
M1 and M2 constant, then determining the statistical signicance of
a1 , a2 , b1 , b2 , and c is of primary interest and whether or not a1 b1
or a2 b2 is signicant is of secondary interest. For this scenario, the
theorizing task should focus rstly on hypothesizing a1 (the X M1
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
path), a2 (X M2 path), b1 (the M1 Y path), b2 (the M2 Y path),
and, nally, c (i.e., the X Y path or the direct effect of X on Y).
Hypothesizing c , as suggested in this recommendation, signals that
the effect of X on Y is deemed to be present beyond the M1 and the
M2 mediation processes and should, therefore, be explicitly spec-
ied when prior theoretical or empirical evidence justies doing
so. Moreover, hypothesizing c is especially critical in the case of
inconsistent mediation (MacKinnon et al., 2000), wherein the sign
of the c estimate is the opposite of the sign for  X , the total indirect
effect of X on Y through both mediation processes. To complete this
theorizing task, hypotheses that M1 and M2 conceptually represent
mediation processes should then be explicitly stated. For brevity,
we label this approach of hypothesizing individual paths en route
to theorizing for mediation effects as the Segmentation Approach.
Examples of SCM research following the Segmentation Approach
include Zhou et al. (2011), Wu et al. (2010), and Paulraj (2011).
An alternative approach is what we label as the Transmit-
tal Approach. Examples of SCM research following this theorizing
approach include Rungtusanatham (2001) and Sarkis et al. (2010).
For this scenario, if the intent of the research is to examine the
mediating processes of M1 and M2 and how M1 and M2 explain the
documented inuence of X on Y, then demonstrating the statistical
signicance of a1 b1 and a2 b2 is the primary interest. In this case,
the total effect of X on Y, c in Fig. 1 (Panel A), is documented to exist
by default and, as such, the theorizing task should focus on hypoth-
esizing how and why M1 and M2 are mediators that transmit the
effect of X onto Y, without having to articulate specic hypothe-
ses regarding to a1 , a2 , b1 , and b2 (Cheung and Lau, 2008; Hayes,
2009). Nonetheless, to complete this theorizing task, a determina-
tion needs to be made as to whether or not the direct effect of X on Y
(i.e., the X Y path or c in Fig. 1 (Panel B)) should be hypothesized.
Again, theoretical or empirical evidence as to the existence of medi-
ation processes beyond M1 and M2 or the specter of inconsistent
mediation should then justify the hypothesizing of c .
3.1.3. Avoid the nomenclature of compete versus partial
mediation
For Recommendation 3, we advise against the nomenclature
of complete mediation versus partial mediation when hypothe-
sizing mediation effects for various reasons. Preacher and Kelly
(2011: p. 96) note that complete mediation is concluded based
on the absence of statistical signicance of the direct effect, or
in other words . . . using the absence of evidence. . . as evidence
of absence. . .; complete mediation, as such, can never be truly
tested, which runs counter to the philosophy of scientic inquiry.
Hayes (2009) argues, moreover, that the nomenclature of complete
mediation versus partial mediation is imprecise as to its ability to
quantify the strength of a specic indirect effect. Also, the effect
size of a specic indirect effect should be sample-size independent
(Cohen, 1988); unfortunately, this is not the case, with the likeli-
hood of concluding in favor of complete mediation being higher in
small samples (Preacher and Kelly, 2011). Lastly, the nomenclature
of complete mediation versus partial mediation is not apropos for
explaining models with inconsistent mediation (MacKinnon et al.,
2000). If Figure 1 (Panel B) were to involve inconsistent mediation,
then |c | is de facto larger than |c| and, more critically, c may not
be detected when c and  X are signicant, equal in effect sizes, but
opposite in signs.
3.2. Testing for the presence of mediation processes
3.2.1. Include the direct effect of X on Y when testing for
mediation effects
Recommendation 4 advocates for the inclusion and estima-
tion of the direct effect of X on Y when conducting empirical
tests for mediation effects (i.e., estimate c in Eq. (4)); otherwise,
107
erroneous conclusions about mediation effects may be drawn (Zhao
et al., 2010). Theoretically, excluding c signies that a given model
of interest has exhaustively identied all mediation processes
explaining the effect of X on Y (Morgan and Winship, 2007; Zhao
et al., 2010). In other words, excluding c theoretically and statisti-
cally from Fig. 1 (Panel B) effectively signals that M1 and M2 are the
only mediators of the relationship between X and Y. Unfortunately,
because most theories tend to be incomplete in terms of their arti-
culated mechanisms (Machamer et al., 2000), positing that M1 and
M2 are the only two mediation processes connecting X to Y is dif-
cult to defend theoretically. Statistically, when M1 and M2 are only
two of a number of possible mediators explaining how X inuences
Y, failure to estimate c in the course of testing for the M1 and the M2
mediation effects, especially when c is signicant, yields inated
estimates of the effect of M1 on Y, or b1 , and of the effect of M2 on
Y, or b2 , in Eq. (4) (Preacher and Hayes, 2008).19
Consider, for example, Rao et al. (2011) who hypothesized and
tested for the effects of physical distribution service price (PDSP)
and physical distribution service quality (PDSQ) on repurchase
behavior (RB) through the customer satisfaction (CS) mediator.
In analyzing this mediation model, Rao et al. (2011, p. 174) ini-
tially tted a structural equation model without the two paths from
PDSQ or PDSP to RB (i.e., the direct effect of PDSQ on RB and that of
PDSP on RB) and reported signicant effects for all three structural
paths: PDSQ CS (p < 0.001), PDSQ CS (p < 0.001), and CS RB
(p < 0.001). However, as part of post hoc analyses, Rao et al. (2011,
p. 175) tted a fully-saturated model to include the two direct
effects of PDSQ and PDSP on RB and found the path from CS to
RB to no longer be signicant (p > 0.10). Hence, excluding the paths
for the direct effects of PDSQ to RB and of PDSP to RB would have
led to an erroneous conclusion in support of the presence of a CS
mediation process; whereas, properly including and estimating the
PDSQ RB and PDSP RB direct effects would have resulted in the
opposite and correct conclusion against a CS mediation process.
3.2.2. Abandon the use of the BK Method to test for mediation
effects
Our fth recommendation is to abandon the BK Method in order
to properly test for mediation effects. Recommendation 5 concurs
with advice already provided by other scholars including Hayes
(2009), MacKinnon et al. (2002), and Zhao et al. (2010), and is partic-
ularly relevant for single-level models, wherein variables and their
effects are dened at the same conceptual level. The BK Method
requires that the total effect of X on Y exists (i.e., is statistically
signicant). Preacher and Hayes (2004) demonstrated that such an
association between X and Y is not required in order for a media-
tion process to exist between X and Y. Moreover, the BK Method
suffers from additional inadequacies, including the inabilities (i) to
quantify, or explicitly test the signicance of, the effect size of a
mediation process (Hayes, 2009; Preacher and Kelly, 2011; Zhao
et al., 2010); (ii) to accommodate models with inconsistent media-
tion (MacKinnon et al., 2000); and (iii) to truly detect for mediation
processes (Fritz and MacKinnon, 2007; MacKinnon et al., 2002).
Consider, for example, Wu et al. (2010) who theorized that
buyer inuence (BI) impacts supplier performance (SP) through
the supplier co-opetition (SCO) mediator. Following the BK
Method, Wu et al. (2010) found BI SCO to be signicant, SCO SP
to be signicant, and BI SP to be non-signicant. With BI SP
19
Implementing this recommendation in regression analysis entails including X as
a predictor variable in Eq. (4). In structural equation modeling, this recommendation
can be implemented via a 2 test to determine whether or not a model with the
direct effect of X on Y freely estimated produces a better model t than one with the
direct effect of X on Y constrained to zero. If model t improves by freely estimating
the direct effect of X on Y, then X does have an inuence on Y beyond the mediation
processes reected in the model.
108 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113

Table 5
Reanalyzing Wu et al. (2010): Bayesian Estimation, Percentile Bootstrapping, and Sobel Test results.a

Parameter Percentile Bootstrapping b Bayesian Estimation c Sobel Testd

Estimate (mean) 95% Credibility Interval Estimate (median) 95% Condence Interval

c: BI SP 0.090 [0.232, 0.345] 0.088 [0.248, 0.418]

a1 : BI SCO 0.452 [0.130, 0.739] 0.452 [0.158, 0.738]
b1 : SCO SP 0.460 [0.776, 0.105] 0.461 [0.792, 0.123]
a1 b1 : BI SCO SP 0.208 [0.459, 0.022] 0.197 [0.422, 0.023]
X : Total indirect effect 0.208 [0.459, 0.022] 0.197 [0.422, 0.023] p = 0.065
a
Wu et al. (2010) provided the original data for reanalysis via Mplus (version 7.0).
b
Parameter estimates are mean values from the empirical sampling distribution generated through Percentile Bootstrapping with maximum likelihood estimation. The
specic indirect effect of BI on SP through SCO is 0.208 and signicant since the corresponding 95% condence interval [0.459, 0.022], does not contain zero.
c
Bayesian Estimation used the default PX1 Gibbs sampler implemented in Mplus (Asparouhov and Muthn, 2010). Diffuse (non-informative) priors were specied for all
model parameters. Results are based on 25,000 iterations. Diagnostics examined for model convergence include the potential scale reduction (PSR), parameter Markov chain
trace plots, and parameter autocorrelation plots (Kaplan and Depaoli, 2012; Muthn and Asparouhov, 2012). Model t was evaluated using posterior predictive checking
(PPC) (Levy, 2011). The specic indirect effect of BI on SP through SCO is 0.197 and signicant since the corresponding 95% credibility interval [0.422, 0.023], does not
contain zero.
d
Fig. 2 shows that the assumption of normality for the sampling distribution of the specic indirect effect, BI SCO SP, does not hold.

being non-signicant, testing for the mediation effect of SCO was
then terminated to suggest the absence of an SCO mediation effect.
Reanalyzing the same data provided by Wu et al. (2010) using the
Sobel Test, Percentile Bootstrapping, and Bayesian Estimation (see
results in Table 5), we nd BI to have a signicant specic indi-
rect effect on SP through the SCO mediator but, contrary to what
had been hypothesized initially in Wu et al. (2010), SCO has a sig-
nicant negative mediation effect instead of a signicant positive
mediation effect for the inuence of BI on SP.*
3.2.3. Choose Explicit Procedures over implicit procedures to test
for mediation effects
Given the availability of Explicit Procedures that provide for-
mal statistical tests for mediation effects and that also do not
require the assumption of normality, our seventh recommenda-
tion is to test for mediation effects using Bayesian Estimation (Yuan
and MacKinnon, 2009), Bootstrapping (Cheung and Lau, 2008),
Monte Carlo Simulation (Preacher and Selig, 2012), or the DOP
Method (MacKinnon et al., 2002, 2004, 2007). More specically,
of the four recommended Explicit Procedures, Bayesian Estimation
to test for mediation effects is particularly useful and preferred
when prior information is available to improve the precision of
parameter estimates (Muthn and Asparouhov, 2012; Yuan and
MacKinnon, 2009) or when sample sizes are small (Scheines et al.,
1999). However, when sample sizes are large, when diffuse priors
are utilized, or when priors cannot be informatively specied, then
Bayesian Estimation results essentially converge on results based
on maximum-likelihood estimation (Rupp et al., 2004). Given these
circumstances, Bayesian Estimation, therefore, leads to similar
conclusions about mediation effects as those from Bootstrapping,
Monte Carlo Simulation, and the DOP Method.
With large samples, preference should be given to Bootstrap-
ping since this method is more exible than the DOP Method,
does not require additional knowledge about Monte Carlo Simu-
lation, and can be easily coded into common statistical packages
for analyzing structural equation models, such as Mplus and AMOS.
Percentile Bootstrapping, the DOP Method, and Monte Carlo Sim-
ulation have been shown to perform similarly well in detecting
mediation effects without elevating Type I error rates (Fritz et al.,
2012; MacKinnon et al., 2004; Williams and MacKinnon, 2008).
Bias-Corrected Bootstrapping, on the other hand, has the most
power to detect mediation effects but, as previously explained, may
result in elevated Type I error rates in small samples when one con-
stituent path in a specic indirect effect is non-signicant and the
other constituent has a moderate-to-large effect size (Fritz et al.,
2012). Hence, with samples sizes being small (N < 100), Percentile
Bootstrapping should be preferred over Bias-Corrected Bootstrap-
ping but with sample sizes being large (N > 100) and per Fritz et al.
(2012), Bias-Corrected Bootstrapping should be preferred over Per-
centile Bootstrapping when Type II error concerns outweigh Type
I error concerns and vice versa.
3.3. Concluding for the presence (or absence) of mediation effects
3.3.1. Draw conclusions from non-signicant Sobel Test results
cautiously
Recommendation 6 concerns conclusions against mediation
effects drawn from non-signicant Sobel Test results. When test-
ing for mediation effects using the Sobel Test, the results should
be interpreted cautiously given the low statistical power of this
test (Cheung and Lau, 2008; Yuan and MacKinnon, 2009). The low
statistical power of the Sobel Test stems from the assumption that
the sampling distribution of a specic indirect effect is normal. This
assumption is typically violated in small samples (MacKinnon et al.,
2002) because the product of two normally-distributed random
variables is non-normal (Aroian, 1947). As a result of this assump-
tion violation, the Sobel Test may lead to incorrect conclusions
about mediation effects.
Consider again the results from reanalyzing the data from Wu
et al. (2010) in Table 5. The Sobel Test results for the specic indi-
rect effect of BI on SP through the SCO mediator is non-signicant
at the = 0.05 level (p = 0.065) and suggests the absence of an SCO
mediation process for the inuence of BI on SP. However, as Fig. 2,
the histogram of the draws from the posterior distribution of this
specic indirect effect, reveals, the specic indirect effect of BI on
SP through the SCO mediator is skewed and non-normal. Hence,
while Percentile Bootstrapping and Bayesian Estimation nds this
specic indirect effect to be signicantly negative, the Sobel Test is
unable to detect this negative mediation effect because the normal-
ity assumption of the sampling distribution for this specic indirect
effect does not hold.
3.3.2. Forget the total indirect effect, examine specic indirect
effects
Recommendation 8 advocates for the evaluation and repor-
ting of specic indirect effects above and beyond an evaluation
and reporting of the total indirect effect. Our last recommenda-
tion applies, in principle, to models involving multiple mediators
in parallel but applies by default to models involving only one
mediator. This recommendation is particularly relevant when
structural equation modeling packages, such as LISREL and AMOS,
are deployed to analyze models with multiple parallel media-
tors. LISREL and AMOS automatically report the signicance of
 X , the total indirect effect, but do not report the signicance of
the constituent specic indirect effects comprising  X (Macho and
Ledermann, 2011). Since the total indirect effect for a model with
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113 109

Fig. 2. Histogram of the posterior distribution for the indirect effect of BI SCO SP in Wu et al. (2010).

Table 6
Results for a multiple-mediator model identical to Fig. 1 (Panel B) based on simulated data.a

Parameter Bias-Corrected Bootstrapping Bayesian Estimation

Estimate 95% Credibility Interval Estimate 95% Condence Interval

c : XY 0.078 [0.051, 0.198] 0.078 [0.058, 0.212]

a1 : X M1 0.388 [0.516, 0.260] 0.388 [0.516, 0.257]
a2 : X M2 0.185 [0.051, 0.322] 0.185 [0.044, 0.323]
b1 : M1 Y 0.248 [0.123, 0.370] 0.248 [0.111, 0.380]
b2 : M2 Y 0.441 [0.312, 0.564] 0.441 [0.314, 0.566]
a1 b1 : X M1 Y 0.096 [0.163, 0.048] 0.094 [0.158, 0.036]
a2 b2 : X M2 Y 0.082 [0.026, 0.156] 0.080 [0.018, 0.152]
X : Total indirect effect 0.015 [0.095, 0.072] 0.015 [0.106, 0.076]
a
A multivariate normal data set of N = 200 observations was simulated from a correlation matrix in Hayes (2009, p. 413) for a model with two parallel mediators identical
to Fig. 1 (Panel B). The total indirect effect from Bayesian Estimation is non-signicant because the 95% credibility interval contains zero; likewise, the total indirect effect
from Bias-Corrected Bootstrapping is also non-signicant because the 95% condence interval contains zero. With both procedures, the specic indirect effect through M1
(i.e., X M1 Y) is signicant and negative while the specic indirect effect through M2 (i.e., X M2 Y) is signicant and positive and the sum of these two specic indirect
effects equal 0.015. With the total indirect effect approximating zero, it becomes non-signicant.

only one mediator is equivalent to the specic indirect effect of
the single mediator, the signicance of  X allows conclusions to
be drawn as to the signicance of the specic indirect effect for
the single mediator. However, for a model with multiple parallel
mediators, this is no longer true and erroneous conclusions about
each individual mediation process may be reached based simply
on an evaluation of the total indirect effect, particularly when the
specic indirect effects have equal effect sizes but opposite signs
(MacKinnon, 2008; Preacher and Hayes, 2008).
To illustrate the danger of evaluating the total indirect effect
only and forgetting to examining the individual specic indirect
effects, consider an example multivariate data set of N = 200 obser-
vations simulated from a correlation matrix provided in Hayes
(2009, p. 413) for a model with two parallel mediators identical
to Fig. 1 (Panel B). Bayesian Estimation and Bias-Corrected Boot-
strapping produce the results in Table 6. Looking appropriately at
the two specic indirect effects, the conclusions are (i) that X inu-
ences Y through M1 in a signicant and negative manner and (ii)
that X inuences Y through M2 in a signicant but positive manner.
Because the magnitudes of these two opposing mediation pro-
cesses are approximately equal, they cancel each other out in Eq.
(5) when computing the total indirect effect of X on Y through the
M1 and the M2 mediators,  X . Analyzing  X , without examining the
individual specic indirect effects of a1 b1 and a2 b2 , leads, therefore,
to the erroneous conclusion that no mediation effects are present.
4. Conclusions
Empirical research in SCM is increasingly going beyond estab-
lishing the relationship between an independent variable, X, and a
dependent variable, Y, to examine complex models involving medi-
ation effects. These efforts, aptly aimed at increasing the precision
of theoretical explanations as to how X inuences Y through one
or more mediation processes (Edwards and Berry, 2010), are not
only welcome but also signal a maturing of SCM as a scientic dis-
cipline. Our paper supports these efforts by directing attention to
the practices for the theorizing of, the testing for, and the drawing
of conclusions about mediation processes in SCM research.
We begin by synthesizing across diverse literature bases to
provide an accessible tutorial on not only the mathematical foun-
dation for what mediation effects are but also what methods there
are to test for mediation effects. This tutorial, we hope, caters to
both SCM scholars who have little expertise with respect to medi-
ation effects and how to test for them and those who are familiar
110 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
with how to test for mediation effects via the more familiar meth-
ods of the BK Method, the JMB Method, or the Sobel Test but who
have not kept up-to-date with more recent and more sophisti-
cated methodological developments. We continue this tutorial to
offer recommendations to improve the theorizing of, the testing
for, and the drawing of conclusions about mediation processes in
SCM research. Many of our recommendations have been similarly
raised by scholars in communications (Hayes, 2009), management
(Wood et al., 2008), marketing (Zhao et al., 2010), and psychol-
ogy (Rucker et al., 2011) in their attempts to guide researchers
within their respective disciplines. However, those pertaining to
how hypotheses about mediation effects are to be stated in light
of the research agenda and how methods to test for mediation
effects are to be selected reect novel contributions for and beyond
the SCM discipline. These recommendations, we trust, ameliorate
current shortcomings in SCM research involving mediation effects.
In conclusion, we emphasize that our treatise on theoriz-
ing of, testing for, and drawing conclusions about mediation
processes does not differentiate between continuous versus non-
continuous mediators, nor does it differentiate between linear
versus non-linear mediation effects. A mediator can be continuous,
dichotomous, counts, or censored (MacKinnon, 2008; Muthn and
Muthn, 2012); mediation effects do not have to be linear (Hayes
and Preacher, 2010) and may potentially be moderated by other
variables (Edwards and Lambert, 2007; Preacher et al., 2007). These
differences, however, do not change the essence of our eight rec-
ommendations. Moreover, while our treatise answers the question
of whether or not specic indirect effects are signicant, it does not
answer the question of which specic indirect effect among several
has the stronger or weaker effect size. We refer those interested
in this question to MacKinnon (2000) who explain how to com-
pare the statistical strengths of different specic indirect effects.
Finally, our treatise limits the discussion to single-level models,
with variables (i.e., independent, dependent, and mediating) all
dened at, and their effects conned to, the same conceptual level.
We recognize that multi-level models involving mediation effects
can be analyzed but note that advice as to how to appropriately
analyze mediation effects in multi-level models is only beginning
to emerge. For those interested in testing for mediation effects in a
multi-level structural equation model, we suggest Preacher et al.,
(2010); for those interested in Bayesian Estimation as a method to
test for mediation effects in multi-level models, we suggest Kaplan
and Depaoli (2012) and Yuan and MacKinnon (2009).
Appendix A. SCM research involving mediation effects: an
exemplary review
Appendix A describes the design and discusses the results of
a review of 81 SCM articles involving mediation processes that
were published, between 2008 and 2011, in the Journal of Business
Logistics (JBL), the Journal of Operations Management (JOM), and the
Journal of Supply Chain Management (JSCM). This review is exem-
plary and not exhaustive. The objective is to identify shortcomings
with respect to how SCM research has been theorizing, empirically
testing, and concluding for mediation effects.
A.1. Journal and Article Selection
JBL, JOM, and JSCM are agship journals of the three professions
of logistics, operations, and sourcing. These three professions, in
turn, are often considered to be denitional elements of what con-
stitutes SCM (cf., Schroeder et al., 2013). The high Impact Factor
ratings for these three journals also suggest that there is signicant
visibility among scholars who conduct research relevant to these
three professions.
We constrained our review to articles published after 2007
to allow for the cross-disciplinary diffusion of relatively-recent
articles that seek to update theoretical and methodological under-
standing pertaining to methods to test for mediation effects (e.g.,
MacKinnon et al., 2002, 2004; Preacher and Hayes, 2004; Shrout
and Bolger, 2002). For our evaluation period, JBL, JOM, and JSCM
published a total of 335 articles (excluding editorials). Of the 335
articles, we retained 81 relevant articles for review by eliminat-
ing conceptual or theory-building articles; articles of an analytical
nature; case-based research articles; and articles involving survey
data, experimental data, or secondary data that did not involve
mediation processes.
A.2. Article coding
We coded the 81 articles for ve specic pieces of information.
First, we determined the number of mediation processes (=1 or
>1) subsumed in the conceptual models of the 81 articles. Second,
we veried whether mediation effects had been formally hypothe-
sized (YES or NO). For articles with conceptual models having more
than one mediator, at least one mediation effect had to be formally
hypothesized in order for the article to receive a YES coding. Third,
we coded the 81 articles to identify the method(s) deployed to test
for mediation effects. Articles in which no formal mediation testing
had been conducted receive a code of 0 for this evaluation cate-
gory. When multiple methods were used in an article, the article
is coded with each method contributing equally. For example, sup-
pose article A tests for mediation processes with the Baron and
Kenny (1986) Method (henceforth, BK Method) and the Sobel Test
(Sobel, 1982). In coding this article A, we therefore assigned 0.5
for the BK Method and 0.5 for the Sobel Test. Fourth, we determined
whether or not the 81 articles drew conclusions about mediation
effects (YES or NO). Finally, for articles with conceptual models
involving more than one mediation effect, we checked whether or
not the article drew explicit conclusions about multiple mediation
effects based on reporting or interpreting the total indirect effect
only (YES or NO).
A.3. Coding results
The coding results for the 81 articles in our review are tabulated
in Tables A1A3. Table A1 describes the 81 articles with respect
to the number of mediation processes subsumed in the concep-
tual models underlying the SCM research, as well as whether or
not mediation effects had been formally hypothesized and conse-
quently tested. Table A2 considers the 54 out of 81 articles that drew
conclusions about mediation effects and identies the method(s), if
any, that had been deployed to support these conclusions. Table A3
examines articles involving conceptual models with multiple medi-
ation processes and documents the manner in which conclusions
regarding individual specic mediation effects were drawn.
Table A1 reveals an uneven distribution of the 81 articles across
JBL, JOM, and JSCM, with more than 50% (i.e., 46 articles) appearing
in JOM. During the same period, JBL published 22 articles (27%)
while JSCM published the remaining 13 articles (16%). More-
over, 48 of the 81 reviewed articles (59%) involve more than one
mediation process, suggesting that the conceptual models being
evaluated in SCM research are increasingly complex. However, only
19 of the 81 articles (23.5%) formally hypothesized for mediation
effects. We compared SCM research examining conceptual models
with only one mediation process (six out of 48 articles (12.5%))
against SCM research examining conceptual models with more
than one mediation process (13 out of 33 articles (39.4%)) and nd
the latter to be statistically more likely to hypothesize for media-
tion effects (z = 2.81, p < 0.01). Interestingly, two of the 19 articles
that formally hypothesized for multiple mediation processes failed
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113 111

Table A1
Formal theorizing and testing for mediation in articles published in SCM journals (20082011).

Journal Number of articles with conceptual models involving mediation processes

>1 Mediation processes =1 Mediation process

Formally hypothesized Not formally hypothesized Formally hypothesized Not formally hypothesized

Formally Not formally Formally Not formally Formally Not formally Formally Not formally
tested tested tested tested tested tested tested tested

JBL 2 0 1 10 1 0 0 8
JOM 2 0 3 20 11 0 2 8
JSCM 0 2 1 7 1 0 1 1
Total 4 2 5 37 13 0 3 17
6 42 13 20
48 33
81

Table A2
Testing for mediation effects by method in articles published in SCM journals (20082011).a

Journal Number of articles drawing conclusions about mediation effects

Formally hypothesized Not formally hypothesized

Formally tested Not formally tested Formally tested Not formally tested

BK Method JMB Method Sobel Test DOP Method BK Method JMB Method Sobel Test DOP Method

JBL 2 0 1 0 0 0 0 1 0 7
JOM 5.5 4.5 3 0 0 0 0.5 4.5 0 16
JSCM 0 1 0 0 2 1 1 0 0 3
Total 17 2 8 26
19 34
53
a
BK Method = Baron and Kenny (1986) Method; JMB Method = James et al. (2006) Method; DOP Method = Distribution of Product Method.

Table A3
Conclusions drawn about specic indirect effects in articles published in SCM journals (20082011).

Journal Number of articles involving conceptual models with > 1 mediation process

Formally hypothesized Not formally hypothesized

Concluded by examining No conclusion Concluded by examining No conclusion

Specic indirect effects Total indirect effect only Specic indirect effects Total indirect effect only

JBL 2 0 0 3 2 6
JOM 1 1 0 12 2 9
JSCM 2 0 0 4 0 4
Total 6 0 23 19
6 42
48

to subsequently conduct formal tests for these mediation effects;
both appeared in JSCM. In addition, eight out of the 62 articles that
did not hypothesize but should have hypothesized for mediation
effects had proceeded to conduct formal tests for mediation effects.
53 of the 81 articles in our review drew conclusions about
mediation effects, even though only 17 articles had formally
hypothesized for mediation effects and subsequently conducted
formal testing for these mediation effects. Of the remaining 36 arti-
cles, two articles had formally hypothesized but did not formally
test for mediation effects, eight articles had not formally hypoth-
esized but did formally test for mediation effects, and 26 articles
drew conclusions about mediation without formal hypothesizing
or formal testing, a priori. Considering only the 25 articles that
drew conclusions about mediation effects based on formal tests
for mediation effects, Table A2 reveals a prevalence of Implicit Pro-
cedures over Explicit Procedures (BK Method or JMB Method = 62%,
Sobel Test = 38%). Implicit Procedures refer to methods that infer
the presence (or absence) of mediation effects by examining the
joint signicance of various constituent paths connecting X to Y
(Cohen and Cohen, 1983; MacKinnon et al., 2002); Explicit Proce-
dures refer to methods that conduct formal statistical tests for the
presence (or absence) of mediation effects. None of the 25 articles
considered the Explicit Procedures of Bootstrapping or Monte Carlo
Simulation. Table A2, moreover, reveals that 28 out of 53 articles
drew conclusions about mediation effects based on an ad hoc exam-
ination of the simultaneous signicances of individual constituent
paths.
48 of the 81 articles in our review involved conceptual mod-
els with > 1 mediation process. Table A3 shows that 42 out of 48
articles (87%) did not offer formal hypotheses about mediation
processes. 19 of these 42 articles were, however, technically cor-
rect in that they did not draw conclusions regarding mediation
processes. Of the 29 articles that did draw conclusions about medi-
ation processes, ve articles (17%) drew their conclusions about
mediation processes by interpreting the total indirect effect. The
remaining 24 articles correctly examined specic indirect effects
for each individual mediation process.
A.4. Shortcomings in SCM research involving mediation processes
Our brief review indicates that SCM research involving medi-
ation processes face three shortcomings. First, formal hypotheses
concerning mediation processes of interest, even when models are
described in prose or pictorially drawn to depict mediation effects,
112 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113

are often not fully developed. Second, mediation effects are typi- MODEL CONSTRAINT: This block of code identies other effects to
cally tested using Implicit Procedures (i.e., the BK Method or the NEW (X M1 Y X M2 Y be estimated
Total);
JMB Method) and have not progressed toward using Explicit Pro- X M1 Y = a1 * b1;
cedures, despite theoretical critiques against Implicit Procedures X M2 Y = a2 * b2;
(Hayes, 2009) and demonstrated statistical concerns with their Total = (a1 * b1) + (a2 * b2);
applications (Fritz and MacKinnon, 2007; MacKinnon et al., 2002). OUTPUT: This block of code requests that results be
CINTERVAL (HPD); reported using the highest posterior density
Third, specic indirect effects are frequently not examined when
(HPD) interval
multiple mediation processes are involved. These shortcomings PLOT: This block of code requests plots of the
have to be overcome in order for SCM research involving medi- TYPE = PLOT2; posterior parameter distributions, posterior
ation processes to be executed in accordance with established parameter trace plot to assess whether the
norms. Marko Chains were operating properly, and
the autocorrelations plots for each parameter

Appendix B.

References
Bias-Corrected Bootstrapping to test for mediation effects in
Mplus: Aroian, L.A., 1947. The probability function of the product of two normally dis-
An annotated syntax for Fig. 1 (Panel B) tributed variables. The Annals of Mathematical Statistics 18 (2), 265271.
DATA: This code species the name of the le Asparouhov, T., Muthn, B., 2010. Bayesian Analysis Using Mplus: Technical Imple-
mentation. Muthn & Muthn, Los Angeles, CA.
File = c:\Multi Med.txt; containing the data
Baron, R.M., Kenny, D.A., 1986. The moderatormediator variable distinction in
VARIABLE: This block of code identies the variables to
social psychological research: conceptual, strategic, and statistical consider-
NAMES ARE Y X M1 M2; be used in analyses
ations. Journal of Personality and Social Psychology 51 (6), 11731182.
USEVARIABLES ARE Y X M1 Bascle, G., 2008. Controlling for endogeneity with instrumental variables in strategic
M2; management research. Strategic Organization 6 (3), 285327.
ANALYSIS: This block of code species that 5000 Bhattacharyya, K., Datta, P., Offodile, O.F., 2010. The contribution of third-party indi-
ESTIMATOR = ML; bootstraps samples be generated with cies in assessing global operational risks. Journal of Supply Chain Management
BOOTSTRAP = 5000; resampling 46 (4), 2543.
MODEL: This block of code denes the structure of Bollen, K.A., 1989. Structural Equations with Latent Variables. Wiley, New York, NY.
M1 ON X (a1); model to be tested Bollen, K.A., Stine, R., 1990. Direct and indirect effects: classical and bootstrap esti-
M2 ON X (a2); mates of variability. Sociological Methodology 20 (1), 115140.
Y ON M1 (b1); Carter, C.R., 2011. A call for theory: the maturation of the supply chain management
Y ON M2 (b2); discipline. Journal of Supply Chain Management 47 (2), 37.
Casella, G., George, E.I., 1992. Explaining the Gibbs sampler. The American Statisti-
Y ON X;
cian 46 (3), 167174.
M1 WITH M2;
Cheung, G.W., Lau, R.S., 2008. Testing mediation and supression effects of latent vari-
MODEL CONSTRAINT: This block of code identies other effects to
ables: bootstrapping with structural equation models. Organizational Reseach
NEW (X M1 Y X M2 Y be estimated Methods 11 (2), 296325.
Total); Cohen, J., 1988. Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Erl-
X M1 Y = a1 * b1; baum, Mahwah, NJ.
X M2 Y = a2 * b2; Cohen, J., Cohen, P., 1983. Applied Multiple Regression/Correlation Analysis for the
Total = (a1 * b1) + (a2 * b2); Behavioral Sciences, 2nd ed. Erlbaum, Hillsdale, NJ.
OUTPUT: This block of code asks that the Collins, L.M., Graham, J.W., Flaherty, B.P., 1998. An alternative framework for den-
CINTERVAL bias-corrected condence interval be ing mediation. Multivariate Behavioral Research 33 (2), 295312.
(BCBOOTSTRAP); calculated Craig, C.C., 1936. On the frequency function of xy. The Annals of Mathematical Statis-
tics 7 (1), 115.
Edwards, M.C., 2010. A Markov chain Monte Carlo approach to conrmatory factor
Appendix C. analysis. Psychometrika 75 (3), 474497.
Edwards, J.R., Berry, J.W., 2010. The presence of something or the absence of noth-
ing: increasing theoretical precision in management research. Organizational
Bayesian Estimation to test for mediation effects in Mplus: Research Methods 13 (4), 668689.
An Annotated Syntax for Fig. 1 (Panel B) Edwards, J.R., Lambert, L.S., 2007. Methods for integrating moderation and
DATA: This code species the name of the le mediation: a general analytical framework using moderated path analysis. Psy-
chological Methods 12 (1), 122.
File = c:\Multi Med.txt; containing the data
Efron, B., 1987. Better bootstrap condence intervals. Journal of the American Sta-
VARIABLE: This block of code identies the variables to
tistical Association 82 (397), 171185.
NAMES ARE Y X M1 M2; be used in analyses
Efron, B., Tibshirani, R., 1993. An Introduction to the Bootstrap. Chapman & Hall,
USEVARIABLES ARE Y X M1 New York, NY.
M2; Ellinger, A.E., Keller, S.B., Bas, A.E., 2010. The empowerment of frontline service staff
ANALYSIS: This block of code species using two in 3PL companies. Journal of Business Logistics 31 (1), 7998.
ESTIMATOR = BAYES; Markov Chains to conduct analyses Flynn, B.B., Sakakibara, S., Schroeder, R.G., Bates, K.A., Flynn, E.J., 1990. Empirical
CHAINS = 2; research methods in operations management. Journal of Operations Manage-
ALGORITHM = GIBBS (PX1); This code species that the default Gibbs ment 9 (2), 250284.
sampling algorithm be utilized Fritz, M.S., MacKinnon, D.P., 2007. Required sample size to detect the mediated
THIN = 20; This code species that only every 20th effect. Psychological Science 18 (3), 233239.
iteration of the Markov Chain be retained to Fritz, M.S., Taylor, A.B., MacKinnon, D.P., 2012. Explanation of two anomalous results
avoid autocorrelation of the parameters in statistical mediation analysis. Multivariate Behavioral Research 47 (1), 6187.
BITERATIONS = 100,000 This code species that the maximum Frohlich, M.T., Westbrook, R., 2001. Arcs of integration: an international study of
supply chain strategies. Journal of Operations Management 19 (2), 185200.
(25,000); number of Markov Chain iterations before
Geman, S., Geman, D., 1984. Stochastic relaxation, gibbs distributions and the
reaching the Gelman-Rubin convergence
Bayesian restoration of images. IEEE Transactions on Pattern Analysis and
criteria is 100,000 iterations and that Machine Intelligence 6 (6), 721741.
convergence is not allowed before the Goodman, L.A., 1960. On the exact variance of products. Journal of the American
25,000th iteration Statistical Association 55 (292), 708713.
MODEL: This block of code denes the structure of Hayes, A.F., 2009. Beyond Baron and Kenny: statistical mediation analysis in the new
M1 ON X (a1); model to be tested mellennium. Communication Monographs 76 (4), 408420.
M2 ON X (a2); Hayes, A.F., Preacher, K.J., 2010. Quantifying and testing indirect effects in sim-
Y ON M1 (b1); ple mediation models when the constituent paths are nonlinear. Multivariate
Y ON M2 (b2); Behavioral Research 45 (4), 627660.
Y ON X; Hayes, A.F., Scharkow, M., 2013. The relative trustworthiness of inferential tests of
M1 WITH M2; the indirect effect in statistical mediation analysis: does method really matter?
Psychological Science, http://dx.doi.org/10.1177/0956797613480187 (in press).
 M. Rungtusanatham et al. / Journal of Operations Management 32 (2014) 99113
Hill, J.A., Eckerd, S., Wilson, D., Greer, B., 2009. The effect of unethical behavior
on trust in a buyersupplier relationship: the mediating role of psychological
contract violation. Journal of Operations Management 27 (4), 281293.
James, L.R., Brett, J.M., 1984. Mediators, moderators, and tests for mediation. Journal
of Applied Psychology 69 (2), 309321.
James, L.R., Mulaik, S.A., Brett, J.M., 2006. A tale of two methods. Organizational
Research Methods 9 (2), 233244.
Judd, C.M., Kenny, D.A., 1981. Process analysis estimating mediation in treatment
evaluations. Evaluation Review 5 (5), 602619.
Kaplan, D., Depaoli, S., 2012. Bayesian structural equation modeling. In: Hoyle, R.H.
(Ed.), Handbook of Structural Equation Modeling. Guilford Press, New York, NY.
Kenny, D.A., Kashy, D.A., Bolger, N., 1998. Data analysis in social psychology. In:
Gilbert, D., Fiske, S.T., Lindzey, G. (Eds.), The Handbook of Social Psychology, vol.
1, 4th ed. McGraw-Hill, New York, NY, pp. 223265.
Kristal, M.M., Huang, X., Roth, A.V., 2010. The effect of an ambidexterous supply chain
strategy on combinative competitive capabilities and business performance.
Journal of Operations Management 28, 415429.
Lau, R.S., Cheung, G.W., 2012. Estimating and comparing specic mediation effects in
complex latent variable models. Organizational Research Methods 15 (1), 316.
Levy, R., 2011. Bayesian data-model t assessment for structural equation modeling.
Structural Equation Modeling 18 (4), 663685.
Machamer, P., Darden, L., Craver, C.F., 2000. Thinking about mechanisms. Philosophy
of Science 67 (1), 125.
Macho, S., Ledermann, T., 2011. Estimating, testing, and comparing specic effects in
structural equation models: the phantom model approach. Psychological Meth-
ods 16 (1), 3443.
MacKinnon, D.P., 2000. Contrasts in multiple mediator models. In: Rose, J.S., Chas-
sin, L., Presson, C.C., Sherman, S.J. (Eds.), Multivariate Applications in Substance
Use Research: New Methods for New Questions. Lawrence Erlbaum Associates,
Mahwah, NJ, pp. 141160.
MacKinnon, D.P., 2008. Introduction to Statistical Mediation Analysis. Taylor & Fran-
cis Group, New York, NY.
MacKinnon, D.P., Fritz, M.S., Williams, J., Lockwood, C., 2007. Distribution of the
product condence limits for the indirect effect: program PRODCLIN. Behavior
Research Methods 39 (3), 384389.
MacKinnon, D.P., Krull, J.L., Lockwood, C.M., 2000. Equivalence of the mediation,
confounding, and suppression effect. Prevention Science 1 (4), 173181.
MacKinnon, D.P., Lockwood, C.M., Williams, J., 2004. Condence limits for the indi-
rect effect: distribution of the product and resampling methods. Multivariate
Behavioral Research 39 (1), 99128.
MacKinnon, D.P., Lockwood, C.M., Hoffman, J.M., West, S.G., Virgil, S., 2002. A com-
parison of methods to test mediation and other intervening variable effects.
Psychological Methods 7 (1), 83104.
Mathieu, J.E., Taylor, S.R., 2006. Clarifying conditions and decision points for medi-
ational type inferences in organizational behavior. Journal of Organizational
Behavior 27 (8), 10311056.
Mathieu, J.E., DeShon, R.P., Bergh, D.D., 2008. Mediational inferences in organiza-
tional research: then, now, and beyond. Organizational Research Methods 11
(2), 203223.
Morgan, S.L., Winship, C., 2007. Counterfactuals and Causal Inference: Methods and
Principles for Social Research. Cambridge University Press, New York, NY.
Muthn, B., Asparouhov, T., 2012. Bayesian structural equation modeling: a more
exible representation of substantive theory. Psychological Methods 17 (3),
313335.
Muthn, L.K., Muthn, B.O., 2012. Mplus Users Guide, 7th ed. Muthn & Muthn, Los
Angeles, CA.
Paulraj, A., 2011. Understanding the relationships between internal resources and
capabilities, sustainable supply management and organizational sustainability.
Journal of Supply Chain Management 47 (1), 1937.
Paulraj, A., Lado, A.A., Chen, I.J., 2008. Inter-organizational communication as a
relational competency: antecedents and performance outcomes in collabo-
rative buyersupplier relationships. Journal of Operations Management 26 (1),
4564.
Pirlott, A.G., Kisbu-Sakarya, Y., Defrancesco, C.A., Elliot, D.L., Mackinnon, D.P., 2012.
Mechanisms of motivational interviewing in health promotion: a Bayesian
mediation analysis. International Journal of Behavioral Nutrition and Physical
Activity 9 (1), 69.
Preacher, K.J., Hayes, A.F., 2004. SPSS and SAS procedures for estimating indirect
effects in simple mediation models. Behavior Research Methods, Instruments,
& Computers 36 (4), 717731.
Preacher, K.J., Hayes, A.F., 2008. Asymptotic and resampling strategies for assessing
and comparing indirect effects in multiple mediator models. Behavior Research
Methods 40 (3), 879891.
Preacher, K.J., Kelly, K., 2011. Effect size measures for mediation models: quanti-
tative strategies for communicating indirect effects. Psychological Methods 16
(2), 93115.
Preacher, K.J., Selig, J.P., 2012. Advantages of Monte Carlo condence intervals for
indirect effects. Communication Methods and Measures 6 (1), 7798.
113
Preacher, K.J., Rucker, D.D., Hayes, A.F., 2007. Addressing moderated media-
tion hypotheses: theory, methods, and prescriptions. Multivariate Behavioral
Research 42 (1), 185227.
Preacher, K.J., Zyphur, M.J., Zhang, Z., 2010. A general multilevel SEM framework for
assessing multilevel mediation. Psychological Methods 15 (3), 209233.
Rao, S., Goldsby, T.J., Grifs, S.E., Iyengar, D., 2011. Electronic logistics service qual-
ity (e-LSQ): its impact on the customers purchase satisfaction and retention.
Journal of Business Logistics 32 (2), 167179.
Rosenzweig, E.D., Roth, A.V., Dean, J.W., 2003. The inuence of an integration strategy
on competitive capabilities and business performance: an exploratory study of
consumer products manufacturers. Journal of Operations Management 21 (4),
437456.
Rucker, D.D., Preacher, K.J., Tormala, Z.L., Petty, R.E., 2011. Mediation analysis in
social psychology: current practices and new recommendations. Social and Per-
sonality Psychology Compass 5 (6), 359371.
Rungtusanatham, M., 2001. Beyond improved quality: the motivational effects
of statistical process control. Journal of Operations Management 19 (6),
653673.
Rupp, A.A., Dey, D.K., Zumbo, B.D., 2004. To bayes or not to bayes, from whether to
when: applications of bayesian methodology to modeling. Structural Equation
Modeling 11 (3), 424451.
Saldanha, J.P., Hunt, C.S., Mello, J.E., 2013. Driver management that drives carrier
performance. Journal of Business Logistics 34 (1), 1532.
Sarkis, J., Gonzalez-Torre, P., Adenso-Diaz, B., 2010. Stakeholder pressure and the
adoption of environmental practices: the mediating effect of training. Journal of
Operations Management 28 (2), 163176.
Scheines, R., Hoijtink, H., Boomsma, A., 1999. Bayesian estimation and testing of
structural equation models. Psychometrika 64 (1), 3752.
Schroeder, R.G., 2008. Introduction to the special issue on theory develop-
ment in operations management. Journal of Operations Management 17 (3),
354356.
Schroeder, R.G., Meyer, S.M., Rungtusanatham, M.J., 2013. Operations Management
in the Supply Chain: Decisions and Cases. McGraw-Hill, New York, NY.
Shaver, J.M., 1998. Accounting for endogeneity when assessing strategy perfor-
mance: does entry mode choice affect FDI survival. Management Science 44
(4), 571585.
Shrout, P.E., Bolger, N., 2002. Mediation in experimental and nonexperimental
studies: new procedures and recommendations. Psychological Methods 7 (4),
422445.
Sobel, M.E., 1982. Asymptotic condence intervals for indirect effects and their
standard errors in structural equation models. Sociological Methodology 13,
290312.
Sobel, M.E., 2008. Identication of causal parameters in randomized studies with
mediating variables. Journal of Educational and Behavioral Statistics 33 (2),
230251.
Stone, C.A., Sobel, M.E., 1990. The robustness of estimates of total indirect effects
in covariance structure models by maximum likelihood. Psychometrika 55 (2),
337352.
Taylor, A.B., MacKinnon, D.P., Tein, J.-Y., 2008. Tests of the three-path mediated
effect. Organizational Research Methods 11 (2), 241269.
Toghi, D., MacKinnon, D.P., 2011. RMediation: an R package for mediation analysis
condence intervals. Behavior Research Methods 43 (3), 692700.
Williams, J., MacKinnon, D.P., 2008. Resampling and distribution of the product
methods for testing indirect effects in complex models. Structural Equation
Modeling 15 (1), 2351.
Williams, L.J., Vandenberg, R.J., Edwards, J.R., 2009. Structural equation modeling in
management research: a guide for improved analysis. The Academy of Manage-
ment Annals 3 (1), 543604.
Wood, R.E., Goodman, J.S., Beckmann, N., Cook, A., 2008. Mediation testing in man-
agement research: a review and proposals. Organizational Research Methods 11
(2), 270295.
Wooldridge, J.M., 2009. Introductory Econometrics: A Modern Approach, 4th ed.
South-Western Cengage Learning, Mason, OH.
Wu, Z., Choi, T.T., Rungtusanatham, M.J., 2010. Suppliersupplier relationships in
buyersuppliersupplier triads: implications for supplier performance. Journal
of Operations Management 28 (2), 115123.
Yuan, Y., MacKinnon, D.P., 2009. Bayesian mediation analysis. Psychological Meth-
ods 14 (4), 301322.
Zhang, J., Wedel, M., Pieters, R., 2009. Sales effects of attention to feature adver-
tisements: a Bayesian mediation analysis. Journal of Marketing Research 46 (5),
669681.
Zhao, X., Lynch, J.G., Chen, Q., 2010. Reconsidering Baron and Kenny: myths and
truths about mediation analysis. Journal of Consumer Research 37 (2), 197206.
Zhou, H., Benton, W.C., Schilling, D.A., Milligan, G.W., 2011. Supply chain integration
and the SCOR model. Journal of Business Logistics 32 (4), 332344.
Zyphur, M.J., Oswald, F.L., 2014. Bayesian estimation and inference: a users
guide. Journal of Management, http://dx.doi.org/10.1177/0149206313501200
(in press).