com/
Published by:
http://www.sagepublications.com
On behalf of:
Additional services and information for Organizational Research Methods can be found at:
Subscriptions: http://orm.sagepub.com/subscriptions
Reprints: http://www.sagepub.com/journalsReprints.nav
Permissions: http://www.sagepub.com/journalsPermissions.nav
Citations: http://orm.sagepub.com/content/9/2/233.refs.html
What is This?
Organizational
Research Methods
Volume 9 Number 2
The structural equation modeling approach to testing for mediation is compared to the Baron and
Kenny approach. The approaches are essentially the same when the hypothesis being tested pre-
dicts partial mediation. The approaches differ, however, in how each tests for complete media-
tion. Disparities in both theory and statistical estimators are identified and discussed. A strategy
for future tests of mediation is recommended.
T he primary problem with tests for mediation evolves from the fact that different statistical
strategies are available. These strategies are similar in many respects but can differ in
complex and subtle ways. The result is that somewhat different tests are being used across
studies to confirm/disconfirm mediation hypotheses. In this article, we shall first describe
how primary statistical strategies for testing for mediation differ. We shall then outline a gen-
eral strategy to test for mediation. We begin with a brief illustration of different strategies.
AuthorsNote: The authors thank the following individuals for their helpful suggestions and advice: Jack Feldman,
Randy Engle, Emily Chia-Huei Ko, and Matthew Minton.
233
complete mediation; the regression weights for M1, M2, and M3 were significant in the first
analysis, and the regression weight for X was not significant in the second analysis. In the
second set of equations (Set 2), Y was regressed on X in the first analysis. M1, M2, and M3
were then added to the predictor set, and a second analysis was conducted (this is actually not
necessary because it is redundant with the second regression analysis in Set 1). Complete
mediation was again supported in that the regression weight for X was significant in the first
analysis. The weights for M1, M2, and M3 were also significant in the second analysis,
although this is necessary only for partial mediation (see later discussion). Relationships
between X and the three mediators were not discussed, but these relationships should have
been significant. If this was the case, then complete mediation was supported.
It is noteworthy that James and Brett (1984) strongly recommended the use of confirma-
tory analytic techniques to test mediation models. Our third and final illustration is based on
the use of confirmatory procedures (LISREL) to test the hypothesis that leader-member
exchange (M) mediates the relationship between transformational leadership (X) and both
organizational citizenship behavior (Y1) and task performance (Y2; Wang, Law, Hackett,
Wang, & Chen, 2005). Structural parameters were estimated for several alternative partial
mediation models as well as for a complete mediation model. Inasmuch as the models formed
a nested sequence, it was possible to compare the goodness of fit of alternative models using
difference chi-squares. These tests, when combined with the parameter estimates, indicated
that (a) M was a function of X, (b) both Y1 and Y2 were functions of M, and (c) with M con-
trolled, X did not affect Y1 or Y2 directly. Overall, a complete mediation model was confirmed.
It is not easy even for statisticians to decipher the similarities and differences among the
three strategies illustrated above because although the tests are similar in many respects, they
are not identical. It would seem useful, therefore, to compare and contrast primary strategies
for testing for mediation, pointing out the similarities and differences. Where differences
occur, it would also seem useful to make recommendations about what we believe to be the
preferred stratagems. Accordingly, we shall compare and contrast the first strategy illustrated
above with a more straightforward version of the third strategy. The hierarchical regression
tests described in James and Brett (1984) are not included because they are subsumed by the
confirmatory analytic approach.
Figure 1
Complete Mediation Model
eM eY
b mx b ym
X M Y
The second article by Baron and Kenny (1986), later extended by Kenny, Kashy, and
Bolger (1998), recommended to social psychologists a test of mediation based on a set of
steps involving correlations and regression weights. We describe this approach below. In the
years since these two publications, published tests of mediation tend to adopt either the SEM
procedure recommended by James and Brett (1984) or what has come to be known as the
Baron-Kenny approach. Some studies use some combination of the two procedures (e.g.,
Harrison, Price, Gavin, & Florey, 2002) or something distinctive.
Here we describe the differences between the Baron and Kenny approach, or simply the
B-K approach, and the SEM approach to testing for mediation. We begin by summarizing
the SEM approach. (Extended treatments of statistical procedures are available in the origi-
nal articles and the references.) We then overview the B-K approach as it is presently con-
ceived (see Kenny et al., 1998; MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002;
Shrout & Bolger, 2002). We highlight similarities and differences between the two
approaches. Where differences occur, we make recommendations about what we believe to
be the preferred stratagems. We conclude with recommendations for future research.
We assume that readers are familiar with the basic ideas of confirmatory approaches to
research, including the key condition that confirmatory approaches test a priori theoretical
models. If not, we refer readers to our earlier discussion of this issue (James & Brett, 1984).
We also assume familiarity, or at least we will not reiterate, many of the excellent points pre-
sented in recent treatments of mediation regarding modeling and statistical testing (e.g.,
MacKinnon et al., 2002; Pearl, 2000; Shrout & Bolger, 2002). This allows us to focus on our
specific concern here, which is comparing what is and what is not estimated in the SEM
approach and the B-K approach to testing for mediation. Such comparisons are aided by
focusing the presentation on statistical estimating equations for both the SEM approach and
the B-K approach. We assume manifest variable designs and ordinary least squares (OLS)
estimation.
SEM Approach
In accordance with most SEM texts, which employ the rule of parsimony from the philoso-
phy of science (e.g., Simon, 1977), James and Brett (1984) used the complete mediation
model as the focal or baseline paradigm for mediation. This is the most basic and parsimoni-
ous mediation model and is shown in Figure 1, where X is the antecedent, M is the mediator,
and Y is the consequence. In the complete mediation model, all the effects of antecedent X on
the consequence Y are transferred through the mediator M.
The SEM-based test of complete mediation is straightforward. There are two endogenous
variables in the model (M and Y) and thus two structural equations. The (OLS) estimating
equations are (variables may be thought of as being in either deviation or standardized form)
m = bmxx + e. (1)
y = bymm + e. (2)
Equation 1 tests the hypothesized relationship between X and M; a significant bmx corrobo-
rates the hypothesis that M is a probabilistic function of X. Equation 2 tests the hypothesized
relationship between Y and M; a significant bym corroborates the hypothesis that Y is a proba-
bilistic function of M. Both coefficients must be significant for the model to be supported.
The complete mediation model has one degree of freedom because the direct path from X
to Y has been set theoretically, which is to say a priori and not empirically based on the data at
hand, to zero. This degree of freedom affords the opportunity to conduct a goodness-of-fit
test, which is a test to ascertain if the hypothesized lack of a direct effect of X on Y can be con-
firmed empirically. There are a number of ways to conduct the goodness-of-fit test, all of
which should converge to the same inference. We shall use the test that underlies much of the
work in SEM. (See MacKinnon et al., 2002, for discussion of goodness-of-fit tests and signif-
icance tests for mediation.)
First, we compute the indirect effect of X on Y via M, which is
bmxbym. (3)
The goodness-of-fit test for the one degree of freedom associated with the complete media-
tion model can now be calculated by
where the reproduced r given by Equation 4 is compared to the observed ryx. If the reproduced
ryx and the observed ryx are not significantly different, then one infers that the only path from X
to Y is via M.
Equations 1, 2, and 4 provide the basic SEM analyses necessary to test for complete medi-
ation. It is also possible to test the indirect effect given in Equation 3 for significance.
MacKinnon et al. (2002) and Shrout and Bolger (2002) discussed the issues involved in
deciding whether to use this test (or its alternatives) to evaluate the adequacy of the model.
Partial mediation in SEM analysis. What if the a priori model to be tested is partial media-
tion? Or, what if the goodness-of-fit test above fails by indicating a direct X to Y relationship
as well as perhaps a mediated relationship? The answer to these questions is often to base an
SEM analysis on the partial mediation model illustrated below. Note that partial mediation is
one of several possibilities when a complete mediation model is disconfirmed, and a new
sample is needed to test for partial mediation if the course to the partial mediation model was
Figure 2
Partial Mediation Model
eM
M eY
b mx b ymx
X Y
b yxm
empirical disconfirmation of a complete mediation model. If this is the case, or if a priori the-
ory proposes a partial mediation model, the SEM steps to follow are shown in Figure 2.
The partial mediation model differs from the complete mediation model by the addition of
a direct effect from X to Y. This addition connotes that part of the causal effect of X on Y is
direct, whereas a separate part of the X to Y causal effect passes through the mediator M.
The partial mediation model is just identified, meaning that it has no degrees of freedom.
The model is less parsimonious than the complete mediation model and thus harder to reject
(or easier to accept). We may, however, test to determine if the paths that are predicted to be
present are corroborated by significant parameter estimates. The estimating equations asso-
ciated with the two endogenous variables in the model are
m = bmxx + e. (5)
Equation 5 is identical to Equation 1. Equation 6 tests the hypotheses that both X and M
have direct and unique effects on Y. A significant byx.m corroborates the hypothesis that X has a
direct effect on Y after M is controlled, whereas a significant bym.x corroborates the hypothesis
that part of the effect of X on Y passes through the mediator M after X is controlled (we
assume signs of coefficients are in the anticipated directions; suppression is not considered
here, nor are nonlinear or nonadditive relationships). The partial mediation model is con-
firmed if all three coefficients in Equations 5 and 6 are significant. (By confirmed, we mean
the model provides a useful and plausible explanation; it is assumed that other, untested mod-
els may be equally useful and explanatory; see Duncan, 1975, p. 19, for alternative three-
variable models). If partial mediation is confirmed, then the indirect effect of X on Y via M is
estimated by bmxbym.x.
A caveat. There are books written on what we have not addressed in SEM in this brief
treatment of complete and partial mediation models. It is especially noteworthy that the
results of the SEM analyses just described for both complete and partial mediation models do
not imply that a given model is true even though the pattern of parameter estimates is consis-
tent with the predictions of the model. There are always other equivalent models implying
different causal directions or unmeasured common causes that would also be consistent with
the data. With this caution in mind, we will proceed with the discussion. The prior six equations
are sufficient to conduct the comparison of what is and is not estimated in the SEM and B-K
approaches to testing for mediation. The overview of the B-K approach is presented below.
B-K Approach
The critical difference between the SEM approach and the B-K approach is the choice of
focal or baseline model for mediation. As noted above, the SEM approach follows the parsi-
mony principle by employing the complete mediation model. The B-K approach adopts the
partial mediation model as its focal paradigm. The B-K approach adopts the partial media-
tion model because it is presumed to be the primary explanatory model in psychology (i.e., a
majority of mediated relationships in psychology are assumed to follow the partial mediation
paradigm).
The B-K approach to testing for mediation is based on the following four steps.
y = byxx + e, (7)
where byx should differ significantly from zero (in the expected direction). This step is equiva-
lent to testing the hypothesis that the correlation between X and Y is zero.
Step 2. M is regressed on X to test the hypothesis that the antecedent X is a cause of the
mediator M. The estimating equation is
m = bmxx + e, (8)
Step 3. The following estimating equation is used to address the questions raised in Steps 3
and 4:
Note that this equation is identical to Equation 6 in the SEM approach to testing for partial
mediation.
In Step 3, bym.x is tested for significance. If significant, one infers that M is uniquely related
to Y when X is controlled. The B-K approach indicates that a nonsignificant bym.x given a sig-
nificant rym suggests that the M-Y relationship is spurious when X is included in the model as
a direct cause of both M and Y. Or, X is a common cause of M and Y, and M has no relation-
ship with Y other than that attributable to X. A nonsignificant bym.x attributable to a
nonsignificant rym indicates that M is not a mediator.
Step 4. Step 4 consists of a significance test of byx.m, which is a test of the direct path
between X and Y with M controlled. If significant, a partial mediation model is indicated,
conditional on satisfaction of the criteria for mediation in Steps 1 through 3. If nonsignifi-
cant, a complete mediation model is indicated, conditional again on satisfaction of the crite-
ria for mediation in the first three steps.
If either partial or complete mediation is found, the indirect effect of X on Y is estimated
by bmxbym.x. This indirect effect can be tested for significance using procedures presented in
MacKinnon et al. (2000) and Shrout and Bolger (2002).
Partial mediation. In the context of what is and what is not estimated, there is little differ-
ence in how the two approaches test for partial mediation. Step 1 from the B-K approach is
not represented in the SEM approach, but it is generally implicit. Otherwise, Equations 5 and
6 that comprise the basis for the test of partial mediation in the SEM approach are the same as
Equations 8 and 9 that are used to test for partial mediation in the B-K approach. Thus, indi-
rect effects are similarly estimated.
Complete mediation. In the test of complete mediation, there are similarities and differ-
ences between the SEM and B-K approaches. That Step 1 does not appear in the SEM
approach and is controversial in the B-K approach is not a major difference in our opinion.
Furthermore, Step 2 from the B-K approach uses Equation 8 to test for an X to M relation-
ship, which is replicated by Equation 1 in the SEM approach. The test to determine if byx.m is
not significantly different from zero in Step 4 of the B-K approach is equivalent to the test of
fit given in Equation 4 in the SEM approach.
However, the approaches differ if the goodness-of-fit test for complete mediation fails. In
the SEM approach, a significant direct path from X to Y with M controlled disconfirms the
complete mediation model, but it does not necessarily imply what the proper fall-back causal
model should be (albeit it is usually) partial mediation. In the B-K approach, there is no
opportunity to fail Step 4 because the significance/nonsignificance of byx.m determines whether
the partial or complete mediation model is adopted to explain the results. This aspect of the
B-K approach presents an anomaly in the midst of what is otherwise a confirmatory analysis.
Use of bym.x when bym is the statistical estimator. If the correlation between X and Y is equal
to the product of the correlation between X and M and the correlation between Y and M, then
bym.x is often equal to bym (see the exception below). Thus, if the complete mediation model is
precisely upheld by statistical estimators, then frequently, there is no pragmatic difference
between using bym.x in place of bym. However, we may expect that an ryx precisely equal to rmxrym
will be a rare occurrence, and thus the estimated values of bym and bym.x can be expected to dif-
fer. More important, there is no reason, practical or otherwise, to engage in a debate about
which estimator to use when theory and mathematics clearly indicate that bym should be
estimated in complete mediation models.
There is an important condition in which bym is not equal to bym.x in complete mediation
models. This is the case in which the mediated model is for all intents and purposes determin-
istic. This condition occurs when both rmx and rym, and therefore bmx and bym, are essentially
equal to 1.0. For example, when driving a vehicle to work, pressing on the accelerator causes
the vehicle to go faster. Stronger presses (X) produce greater accelerations (Y). Among the
many mediators (M) are linkages from the accelerator pedal to the fuel system, the fuel sys-
tem to the engine, and the engine to the wheels. On arriving at work, an initial task is usually
to respond to correspondence. Thoughts (X) are converted into a letter (Y) via the mediating
personal computer (M1) and printer (M2). (Number of letters written per day varies. The cor-
relation between the daily frequency of thoughts about what to write and the daily use of a PC
to convert the thoughts to a written file is essentially 1.0, as are the correlations between the
points in the rest of the chain.) The letters (Y) are sent to recipients (Z) via the post office (a
new M). (Correlations are again based on daily frequencies.) Later, in passing a colleague in
the hall, a ritual is followed in which the colleague pats one on the back. Over days, pats (X)
vary from mere taps to hearty thumps, which engender sensations varying from touch to near
pain (Y), the sensations being mediated by the autonomic nervous system (M).
There are, of course, random shocks in each mediation system. Pressing on the accelerator
will not cause the car to go faster if it is out of gas, thoughts will not be transferred into letters
if the printer is broken, letters will not be transferred on federal holidays, and a synapse may
misfire randomly in the autonomic nervous system. Nonetheless, each mediated relation
occurs with sufficient dependability to render, based on many replications, a correlation of
essentially 1.0 for each link in the causal chain. We rely on dependable relations such as
these to give our lives predictability and stability. Indeed, we plan our lives on the presump-
tion that mediated relations such as these (and many others) will occur with essentially
perfect consistency.
In each mediation model, (each) M adds nothing unique to the prediction of Y in relation
to X. This is not a problem for the SEM approach to testing for complete mediation. That is, if
we use bmx and bym to estimate the mediation relations in each illustration, the values would
correctly be essentially 1.0. However, these mediation models place a serious strain on the B-
K approach to testing for complete mediation. If we use bym.x to estimate each M to Y media-
tion relation, the estimator is undefined. This is because with correlations equal essentially to
1.0, the denominator of bym.x is equal to zero.
Thus, attempts to use bym.x to estimate the M to Y relation in complete mediation models
may result in an intractable solution. We are unlikely to encounter this problem in disciplines
such as social psychology, industrial-organizational psychology, and organizational behav-
ior because deterministic models are rare. However, deterministic models are not at all rare in
other disciplines such as neurophysiology (Shepherd, 2004), and tests for mediation should
apply here as well. The ensuing inference is that it is time to reconsider Step 3 of the B-K
approach and dispense with the requirement that M add uniquely to X in the prediction of Y
in tests of complete mediation. We empathize with and support Baron and Kennys concern
for testing for spuriousness in the M, Y relationship. However, as we have noted, Step 3 is
redundant in complete mediation models that satisfy the tests described earlier. Moreover, it
creates an impossible condition in deterministic models.
Recommendations
We conclude with advice outlining general strategies for testing for mediation.
Step 2. Once the form of mediation has been hypothesized, our second recommendation is
to test the hypothesis using SEM techniques. If complete mediation is predicted, then the
equations and tests are as follows:
m = bmx x + e, (1)
y = bym m + e, (2)
where bmx in Equation 1 and bym in Equation 2 should both be significant and the reproduced r
given by Equation 4 should not be significantly different from the observed ryx.
These are reproductions of earlier equations, as indicated by the primes. We will not reiter-
ate the logic of the tests here, nor will we attempt to present statistical tests or other more
advanced statistical treatments cited earlier in this article. Our objective here is simply to out-
line general strategy.
If partial mediation is hypothesized, then the equations and tests are as follows:
m = bmx x + e, (5)
where all three parameter estimates in the three equations should be significant. These equa-
tions furnish the same test as the B-K approach with the exception that here, byx.m has to be sig-
nificant to support partial mediation.
In sum, the fundamentals of tests of mediation are not complex. The tests can be made
complex by adding variables, embedding mediation relationships in larger causal models,
testing alternative causal orders, testing alternative causal directions, framing tests in terms
of the nesting and sequencing of models, (as in the third illustration of mediation at the begin-
ning of this article), adding additional manifest indicators of X, M, and Y, and conducting the
SEM in terms of a latent variable analysis, building time into the design, and so on. One form
of additional complexity that we strongly recommend is to add tests of alternative causal
models to a basic mediation analysis. The test of alternative causal orders in mediation mod-
els by Griffin (1977) remains a classic. The tests of mediation in leadership by Wang et al.
(2005) used to illustrate SEM also included tests of alternative causal models. As we said
more than 20 years ago, the objective of SEM is not to confirm or disconfirm a specific model
(James et al., 1982). The objective is to contrast alternative models and identify those that
appear to offer useful explanations versus those that do not.
References
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological
research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology,
51, 1173-1182.
Duncan, O. D. (1975). Introduction to structural equations. New York: Academic Press.
Early, P. C., & Mosakowski, E. (2000). Creating hybrid team cultures: An empirical test of transnational team
functioning. Academy of Management Journal, 43, 26-49.
Griffin, L. J. (1977). Causal modeling of psychological success in work organizations. Academy of Management
Journal, 20, 6-33.
Harrison, D. A., Price, K. H., Gavin, J. H., & Florey, A. T. (2002). Time, teams, and task performance: Changing
effects of surface- and deep-level diversity on group functioning. Academy of Management Journal, 45, 1029-
1045.
James, L. R., & Brett, J. M. (1984). Mediators, moderators, and test for mediation. Journal of Applied Psychology,
69, 307-321.
James, L. R., Mulaik, S. A., & Brett, J. M. (1982). Causal analysis: Assumptions, models, and data. Beverly Hills,
CA: Sage.
Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data analysis in social psychology. In D. Gilbert, S. T. Fiske, &
G. Lindzey (Eds.), Handbook of social psychology (Vol. 1, pp. 233-265). New York: McGraw-Hill.
MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002). A comparison of methods
to test mediation and other intervening variable effects. Psychological Methods, 7, 83-104.
Mulaik, S. A. (2002). The curve-fitting problem: An objectivist view. Philosophy of Science, 68, 218-241.
Pearl, J. (2000). Causality: Models, reasoning and inference. Cambridge, UK: Cambridge University Press.
Rozeboom, W. W. (1956). Mediation variables in scientific theory. Psychological Review, 63, 249-264.
Shepherd, G. M. (2004). The synaptic organization of the brain (5th ed.). New York: Oxford University Press.
Shrout, P. E., & Bolger, N. (2002). Mediation in experimental and nonexperimental studies: New procedures and
recommendations. Psychological Methods, 7, 422-445.
Simon, H. A. (1977). Models of discovery. Dordrecht, the Netherlands: R. Reidel.
Thomas-Hunt, M. C., & Phillips, K. W. (2004). When what you know is not enough: Expertise and gender dynam-
ics in task groups. Personality and Social Psychology Bulletin, 30, 1585-1598.
Wang, H., Law, K. S., Hackett, R. D., Wang, D., & Chen, Z. X. (2005). Leader-member exchange as a mediator of
the relationship between transformational leadership and followers performance and organizational citizen-
ship behavior. Academy of Management Journal, 48, 420-432.
Lawrence R. James is a professor of psychology at the Georgia Institute of Technology in Atlanta. He received
his BS, MS, and PhD degrees from the University of Utah. He has been active in studying the effects of organiza-
tional environments on individual adaptation, motivation, and productivity. His methodological contributions
have made possible tests of new models in areas such as personality, organizational climate, leadership, and per-
sonnel selection. He is a fellow in Division 5 (Evaluation and Measurement) and Division 14 (Industrial and Orga-
nizational Psychology) of the American Psychological Association.
Stanley A. Mulaik is a professor emeritus of psychology at the Georgia Institute of Technology in Atlanta. He
received his BS, MA, and PhD degrees from the University of Utah. He is a quantitative psychologist who special-
izes in factor analysis and structural equation modeling and also has published articles in the field of philosophy of
science and causality. Prior to retirement, he was a fellow of Division 5 (Evaluation and Measurement) of the
American Psychological Association.
Jeanne M. Brett is the DeWitt W. Buchanan Jr. Distinguished Professor of dispute resolution and organizations
and director of the Dispute Resolution Research Center at the Kellogg School of Management, Northwestern Uni-
versity. Her PhD in psychology is from the University of Illinois. Her current areas of research are cross-cultural
negotiations, the resolution of disputes, and the performance of multicultural teams. In 2003, she received the
Academy of Managements Outstanding Educator Award.