Article
Improving Our Understanding
of Moderation and Mediation in
Strategic Management Research
Herman Aguinis
1
, Jeffrey R. Edwards
2
,
and Kyle J. Bradley
1
Abstract
We clarify differences among moderation, partial mediation, and full mediation and identify meth-
odological problems related to moderation and mediation from a review of articles in Strategic
Management Journal and Organization Science published from 2005 to 2014. Regarding moderation,
we discuss measurement error, range restriction, and unequal sample sizes across moderator-based
subgroups; insufficient statistical power; the artificial categorization of continuous variables; assumed
negative consequences of correlations between product terms and its components (i.e., multi-
collinearity); and interpretation of first-order effects based on models excluding product terms.
Regarding mediation, we discuss problems with the causal-steps procedure, inferences about
mediation based on cross-sectional designs, whether a relation between the antecedent and the
outcome is necessary for testing mediation, the routine inclusion of a direct path from the ante-
cedent to the outcome, and consequences of measurement error. We also explain how integrating
moderation and mediation can lead to important and useful insights for strategic management theory
and practice. Finally, we offer specific and actionable recommendations for improving the appro-
priateness and accuracy of tests of moderation and mediation in strategic management research.
Our recommendations can also be used as a checklist for editors and reviewers who evaluate
manuscripts reporting tests of moderation and mediation.
Keywords
moderating effect, mediating effect, contingency, interactionism, interaction
For decades, hypotheses that involve moderation and mediation have been central to strategic
management research. Moderation represents the idea that the magnitude of the effect of an
antecedent (e. g., orga niz at ional str uct ure or strateg y) on firm outcomes depen ds on conting enc y
1
Department of Management and Entrepreneurship, Kelley School of Business, Indiana University, Bloomington, IN, USA
2
Kenan-Flagler Business School, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA
Corresponding Author:
Herman Aguinis, Department of Management and Entrepreneurship, Kelley School of Business, Indiana University, 1309 E.
10th Street, Bloomington, IN 47405-1701, USA.
Email: haguinis@indiana.edu
Organizational Research Methods
1-21
ª The Author(s) 2016
Reprints and permission:
sagepub.com/journalsPermissions.nav
DOI: 10.1177/1094428115627498
orm.sagepub.com
at INDIANA UNIV on January 28, 2016orm.sagepub.comDownloaded from
factors, such as the uncertainty and instability of the e nvi ronme nt and the products and service s
produced by the firm (e. g., Chand ler, 1962; La wrenc e & Lorsch, 19 67; Sc hoonhov en, 1981;
Thompson, 1967). On the other hand, mediation points to the presence of an intervening variable
or mechanism that transmits the effect of an antecedent variable on an outcome (MacCorquodale
& M eehl, 1948; Mathieu, DeShon, & Bergh, 2008; Ndofor, Sirmon, & He, 2011). For instance,
mediation is captured by the notion that the effect of the competitive environment on firm
performance is transmitted by firm strategy, such that the environment influences strategic choices
that in turn affect performance (Child, 1972). In a nutshell, moderation refers to the conditions
under which an effect varies in size, whereas mediationreferstounderlyingmechanismsand
processes that connect antecedents and outcomes. Clearly, both of these pursuits are critical for
advancing strategic management theory and practice.
In spite of their centrality, the assessment and interpretation of moderation and mediation are
undermined by several problems. We reached this conclusion after systematically reviewing
articles published in Strategic Management Journal (SMJ)andOrganization Science (OS)
between January 2005 and December 2014 that assessed moderation, mediation, or both. Our
review of the 205 articles that assessed moderation revealed seven key problems. Overall, these
demonstrated an average of 2.57 of the seven problems we identified, with only one article
avoiding the probl ems enti rel y. In simila r fashi on, our re view of the 62 artic le s that addr esse d
mediation revealed six key problems, and on avera ge , the articles exhibited 3.52 of the pr oble ms
each, with none of the articles being problem-free. Accordingly, there is a need to identify and
describe the se problems and explain h ow they can be ameliorated or avoided in the future. Such
treatment would contribute to ‘‘the ongoing stream of methodological inquiry in strategy
research’’ (Wiersema & Bowen, 2009, p. 688) and benefit strategic ma nagement researchers as
they pursue answers to questions that are important to field.
The goal of this article is to advance our understanding of the meaning, analysis, and inter-
pretation of moderation and mediation in strategic management research. We do s o utilizing a
five-pronged approach. First, we clarify the conceptual nature and distinctions among moderation,
partial media tion, and full med iati on. Se co nd, we identify key problems regarding moderation and
mediation and report results of our literature review regarding their relative frequency of occur-
rence in articles in SMJ and OS. Third, we illustrate t he detrimental impact of these issues by
referring t o s pe cific substantive domains a nd studies. Because the problems we identified are so
pervasive, we believe it would be inappropriate to identify specific articles by name. Rather, we
use various research domains for illustration to engage the reader in the substantive importance of
the issues involved. Fourth, we offer proposed solutions to address these problems in future
research. Finally, we go beyond the more traditi onal treatment of moderation and me di ation to
explain how integra ting moderation and mediation can lead to important and useful insights for
strategic m anage ment theory and practice .
Moderation and Mediation: Conceptual Distinctions
A moderator variable influences the nature (e.g., magnitude and/or direction) of the effect of an
antecedent on an outcome.
1
Moderation is illustrated graphically in Figure 1a, which shows that the
moderator variable Z influences the path relating X to Y. When the moderator variable is categorical
(e.g., industry type), the traditional data-analytic approach is subgrouping analysis, which consists of
comparing correlation or regression coefficients across the various subgroups or categories (Aguinis
& Pierce, 1998; Boyd, Haynes, Hitt, Bergh, & Ketchen, 2012). When the moderating effect is
continuous (e.g., firm resources), studies typically rely on moderated multiple regression (Aiken
& West, 1991; Coh en, 1978), which con sists o f cre ating a regression model that predicts the
outcome based on a predictor X, a second predictor Z hypothesized to be a moderator, and the
2 Organizational Research Methods
at INDIANA UNIV on January 28, 2016orm.sagepub.comDownloaded from
product term between X and Z, which carries information on the moderating effect of Z on the X-Y
relation. The regression coefficient for the XZ product term from which X and Z have been partialed
out offers information on the presence as well as magnitude of the moderating effect.
A mediator variable transmits the effect of the antecedent on the outcome, either in part or whole
(Baron & Kenny, 1986; MacKinnon, 2008). Figure 1b shows a no-mediation model in which there is
only a direct effect of X on Y. Mediation is illustrated graphically in Figure 1c, which shows that X
affects Y both directly (i.e., path c
0
) and indirectly (i.e., the combination of paths a and b) through the
mediator M. The indirect effect represents that part of the effect of X on Y that is mediated by M, with
the magnitude of this effect represented by the product of the paths a and b. A full mediation model is
one in which ab 6¼ 0 and c
0
¼ 0, whereas partial mediation exists when ab 6¼ 0 and c
0
6¼ 0.
Literature Review
As mentioned earlier, we conducted a literature review of empirical articles published in SMJ and
OS between January 2005 and December 2014. During this decade, SMJ published 794 articles, and
OS published 717 articles, for a total of 1,511 articles. Our review included all ‘‘Research Articles’’
and ‘‘Research Notes and Commentaries’’ sections of both journals and excluded articles published
in the ‘‘From the Editor’’ section. We used Google Scholar to search for articles that reported
moderation or mediation tests. For moderation, we used the terms interaction, interacting, modera-
tion, and moderating. This search resulted in 775 articles from SMJ and 696 articles from OS.We
then manually examined these articles to determine which ones included tests of moderation using
multiple regression because the majority of moderation and mediation articles use this data-analytic
XY
Z
a. Moderation
XY
c
b. Direct effect (i.e., no mediation)
X
Y
M
ab
c'
c. Mediation
X Y
M
a
1
b
2
b
1
d. Moderated mediation
Z
a
3
b
5
b
4
Figure 1. Graphic representation of moderation and mediation models.
Aguinis et al. 3
at INDIANA UNIV on January 28, 2016orm.sagepub.comDownloaded from
approach.
2
The final count of articles addressing moderation was 205, which included 126 in SMJ
and 79 in OS.
We conducted a similar search for mediation articles. Specifically, we used Google Scholar with
the terms indirect, mediate, mediation, and mediator. This initial search resulted in a total of 315
articles published in SMJ and 385 articles published in OS. We then manually searched these articles
to find those that actually conducted a mediation analysis, which yielded a final count of 62 articles,
including 24 in SMJ and 38 in OS.
In the next two sections, we describe problems regarding moderation and mediation, including
illustrations of the impact of these issues on substantive conclusions as well as suggested solutions
for each of the problems we identified. We first address moderation, and then we turn to mediation.
Moderation: Problems and Solutions
Problem 1: Lack of Attention to Measurement Error
The most prevalent problem in strategic management studies that examine moderation concerns the
effects of measurement error. Specifically, 62.44% of articles in our review did not identify mea-
surement error as a potential problem, as evidenced by the fact that they did not mention measure-
ment error at all. Our results specific to moderation are consistent with the finding that most articles
published in SMJ do not report reliability estimates (Boyd, Gove, & Hitt, 2005). The reasons for this
omission are unclear. It could reflect an implicit assumption that the effects of measurement error are
negligible, lack of knowledge rega rding the biasing effects of measurement error on parameter
estimates and hypothesis tests, or prevailing norms in the domains represented by the articles.
On the surface, the lack of attention to measurement error might seem understandable for
certain constructs. For example, measures of pe rfo rmanc e for public firms must go through an
audit process, which leaves little room for subjectivity that might introduce measurement e rror
(Boyd, Bergh, Ireland, & Ketchen, 2013; Dalton & Aguinis, 2013; Godfrey & Hill, 1995). How-
ever, many other constructs involve ratings of beliefs and opinions collected using self-report
surveys, which are measured with error (Boyd et al., 2005). Measurement error is problematic
because when independent and moderator variables are measured with error, unstandardized
coefficient estimates will be biased, and this bias is particularly pronounced for moderating
effects. In contrast, measurement error in outcome variables does not bias coefficient estimates,
but it will attenuate estimates of explained v ariance, making it seem that predictors have less
explanatory power than is actually the case.
Busemeyer and Jones (1983) provided the following expression, which estimates the reliability
for the product term XZ based on the reliabilities of the predictor X and moderator Z variables when
both are standardized:
r
XZ;XZ
¼
r
2
XZ
þ r
XX
r
ZZ
r
2
XZ
þ 1
: ð1Þ
Equation 1 indicates that when the predictor X and the moderator Z are uncorrelated (i.e., r
XZ
¼ 0),
the reliability of the product term is reduced to the product of the reliabilities of the predictors. For
example, if the reliability of X is .70 and the reliability of Z is also .70, the resulting reliability of the
product term is only .49! It seems safe to assume that few, if any, strategic management researchers
would find it acceptable that 50% of the variance in a measure is random error.
There is good reason to believe that the deleterious effects of measurement error are pervasive in
articles reporting tests of moderation. In fact, the vast majority of these articles report very small
moderating effects across various domains, such as the moderating effect of headquarters embedd-
ednes s on the relation between subsidiary embeddedness and headquarters value-add ed, or the
4 Organizational Research Methods
at INDIANA UNIV on January 28, 2016orm.sagepub.comDownloaded from
moderating effect of a firm’s resources and capabilities to deal with natural gas deregulation on the
relation between managerial domain-specific experience and opportunity interpretation (i.e., rang-
ing from threat to opportunity). In these examples, reliabilities for the product terms were not
reported. Moreover, given that, when reported, reliabilitie s for the components are often in the
.70s, it is likely that about 50% of variance in product terms is random error.
In short, tests of many moderator variable hypotheses have been undermined due to the deleter-
ious impact of measurement error. Future research should, at a minimum, report reliability estimates
for all predictors, including product components.
3
Reporting reliability is particularly important for
situations when a hypothesized moderating effect is not found because if reliability is low, an
existing moderating effect is likely underestimated and, in some cases, might go undetected.
Problem 2: Variable Distributions Are Assumed to Include the Full Range of Possible Values
A second important problem is that samples of firms used in strategic management research usually
do not represent the full range of possible scores on the variables under consideration that might
exist in the population. For example, studies regarding the resource-based theory of the firm rarely
include the full range of resources (Crook, Ketchen, Combs, & Todd, 2008). Similarly, firms with
poor performance in the population might not be represented in the sample, which could instead
consist mostly of firms with high scores on performance and related variables (Bergh et al., in press).
These mechanisms lead to range restriction, meaning that the variance of variables is smaller in the
sample compared to the variance in the population.
Although rarely acknowledged, range restriction has an adverse impact on tests of moderation
(i.e., 34.15% of articles in our review seemed to include scores that did not span the full possible
range). Specifically, Aguinis and Stone-Romero’s (1997) Monte Carlo study revealed that when
sample variance is less than population variance, even by what may be considered a small amount,
the statistical power for detecting moderating effects is substantially diminished. For example, in a
situation with a total sample size of 300 and no truncation on X scores, the statistical power to detect
a medium-size moderating effect was an acceptable .81. However, when the scores were sampled
from the top 80% of the distribution of the population scores, power decreased to .51. In other words,
assuming that moderation exists, the accuracy of the moderating effect test is tantamount to flipping
a coin. Thus, given the realistic conditions simulated by Aguinis and Stone-Romero, even a rela-
tively mild degree of range restriction (i.e., just the bottom 20% of the distribution is truncated) can
markedly decrease statistical power and threaten the validity of conclusions regarding moderating
effect hypotheses. And, as with measurement error, even if a moderating effect is statistically
significant, range restriction can bias the observed effect size downward. Drawing from the exam-
ples mentioned earlier, it is unlikely that variables such as a firm’s resources and capabilities to deal
with natural gas deregulation and managerial domain-specific experience included the full range of
scores that might exist in the population of firms. Hence, most of this research has likely under-
estimated the true size of moderating effects.
In short, range restriction makes population-level moderating effects seem smaller than they
actually are or might even render them statistically nonsignificant. Future research should attempt
to capture the full range of scores of all variables involved in the analysis. When this is not feasible,
and if moderating effects are small or nonsignificant, the estimated population variance should be
provided to rule out range restriction as a plausible alternative explanation for the results obtained.
Problem 3: Unequal Sample Size Across Moderator-Based Categories
A third problem is that when the moderator variable is inherently categorical (e.g., industry type,
firm type), the number of firms across categories is usually not equal. This issue was apparent in
Aguinis et al. 5
at INDIANA UNIV on January 28, 2016orm.sagepub.comDownloaded from
