Journal of Pe~nality and Social Psychology Copyright 1986 by the American Psychological Association, Inc.
1986, Vol. 51, No. 6, 1173-1182 0022-3514/86/$00.75
The Moderator-Mediator Variable Distinction in Social Psychological
Research: Conceptual, Strategic, and Statistical Considerations
Reuben M. Baron and David A. Kenny
University of Connecticut
In this article, we attempt to distinguish between the properties of moderator and mediator variables
at a number of levels. First, we seek to make theorists and researchers aware of the importance of
not using the terms
moderator and mediator
interchangeably by carefully elaborating, both concep-
tually and strategically, the many ways in which moderators and mediators differ. We then go beyond
this largely pedagogical function and delineate the conceptual and strategic implications of making
use of such distinctions with regard to a wide range of phenomena, including control and stress,
attitudes, and personality traits. We also provide a specific compendium of analytic procedures ap-
propriate for making the most effective use of the moderator and mediator distinction, both sepa-
rately and in terms of a broader causal system that includes both moderators and mediators.
The purpose of this analysis is to distinguish between the
properties of moderator and mediator variables in such a way
as to clarify the different ways in which conceptual variables
may account for differences in peoples' behavior. Specifically,
we differentiate between two often-confused functions of third
variables: (a) the moderator function of third variables, which
partitions a focal independent variable into subgroups that es-
tablish its domains of maximal effectiveness in regard to a given
dependent variable, and (b) the mediator function of a third
variable, which represents the generative mechanism through
which the focal independent variable is able to influence the
dependent variable of interest.
Although these two functions of third variables have a rela-
tively long tradition in the social sciences, it is not at all uncom-
mon for social psychological researchers to u, the terms
mod-
erator and mediator
interchangeably. For example, Harkins,
Latan6, and Williams 0980) first summarized the impact of
identifiability on social loafing by observing that it "moderates
social loafing" (p. 303) and then within the same paragraph
proposed "that identifiability is an important mediator of social
loafing:' Similarly, Findley and Cooper (1983), intending a
moderator interpretation, labeled gender, age, race, and socio-
economic level as mediators of the relation between locus of
control and academic achievement. Thus, one largely pedagogi-
This research was supported in part by National Science Foundation
Grant BNS-8210137 and National Institute of Mental Health Grant
R01 MH-40295-01 to the second author. Support was also given to him
during his sabbatical year (1982-83) by the MacArthur Foundation at
the Center for Advanced Studies in the Behavioral Sciences, Stanford,
California.
Thanks are due to Judith Harackiewicz, Charles Judd, Stephen West,
and Harris Cooper, who provided comments on an earlier version of
this article. Stephen P. Needel was instrumental in the beginning stages
of this work.
Correspondence concerning this article should be addressed to Reu-
ben M. Baron, Department of Psychology U-20, University of Connect-
icut, Storrs, Connecticut 06268.
cal function of this article is to clarify for experimental re-
searchers the importance of respecting these distinctions.
This is not, however, the central thrust of our analysis. Rather,
our major emphasis is on contrasting the moderator-mediator
functions in ways that delineate the implications of this distinc-
tion for theory and research. We focus particularly on the
differential implications for choice of experimental design, re-
search operations, and plan of statistical analysis.
We also claim that there are conceptual implications of the
failure to appreciate the moderator-mediator distinction.
Among the issues we will discuss in this regard are missed op-
portunities to probe more deeply into the nature of causal
mechanisms and integrate seemingly irreconcilable theoretical
positions. For example, it is possible that in some problem areas
disagreements about mediators can be resolved by treating cer-
tain variables as moderators.
The moderator and mediator functions will be discussed at
three levels: conceptual, strategic, and statistical. To avoid any
misunderstanding of the moderator-mediator distinction by er-
roneously equating it with the difference between experimental
manipulations and measured variables, between situational and
person variables, or between manipulations and verbal self-re-
ports, we will describe both actual and hypothetical examples
involving a wide range of variables and operations. That is,
moderators may involve either manipulations or assessments
and either situational or person variables. Moreover, mediators
are in no way restricted to verbal reports or, for that matter, to
individual-level variables.
Finally, for expository reasons, our analysis will initially
stress the need to make clear whether one is testing a moderator
or a mediator type of model. In the second half of the article,
we provide a design that allows one to test within the structure
of the same study whether a mediator or moderator interpreta-
tion is more appropriate.
Although these issues are obviously important for a large
number of areas within psychology, we have targeted this article
for a social psychological audience because the relevance of this
distinction is highest in social psychology, which uses experi-
1173
1174 REUBEN M. BARON AND DAVID A. KENNY
mental operations and at the same time retains an interest in
organismic variables ranging from individual difference mea-
sures to cognitive constructs such as perceived control.
The Nature of Moderators
In general terms, a moderator is a qualitative (e.g., sex, race,
class) or quantitative (e.g., level of reward) variable that affects
the direction and/or strength of the relation between an inde-
pendent or predictor variable and a dependent or criterion vari-
able.
Specifically within a correlational analysis framework, a
moderator is a third variable that affects the zero-order correla-
tion between two other variables. For example, Stem, McCants,
and Pettine (1982) found that the positivity of the relation be-
tween changing life events and severity of illness was considera-
bly stronger for uncontrollable events (e.g., death of a spouse)
than for controllable events (e.g., divorce). A moderator effect
within a correlational framework may also be said to occur
where the direction of the correlation changes. Such an effect
would have occurred in the Stern et al. study if controllable life
changes had reduced the likelihood of illness, thereby changing
the direction of the relation between life-event change and ill-
ness from positive to negative.
In the more familiar analysis of variance (ANOVA) terms, a
basic moderator effect can be represented as an interaction be-
tween a focal independent variable and a factor that specifies
the appropriate conditions for its operation. In the dissonance-
forced compliance area, for example, it became apparent that
the ability of investigators to establish the effects of insufficient
justification required the specification of such moderators as
commitment, personal responsibility, and free choice (cf.
Brehm & Cohen, 1962).
An example of a moderator-type effect in this context is the
demonstration of a crossover interaction of the form that the
insufficient justification effect holds under public commitment
(e.g., attitude change is inversely related to incentive), whereas
attitude change is directly related to level of incentive when the
counterattitudinal action occurs in private (cf. Collins & Hoyt,
1972). A moderator-interaction effect also would be said to oc-
cur if a relation is substantially reduced instead of being re-
versed, for example, if we find no difference under the private
condition.
Toward Establishing an Analytic Framework
for Testing Moderator Effects
A common framework for capturing both the correlational
and the experimental views of a moderator variable is possible
by using a path diagram as both a descriptive and an analytic
procedure. Glass and Singer's (1972) finding of an interaction
of the factors stressor intensity (noise level) and controllability
(periodic-aperiodic noise), of the form that an adverse impact
on task performance occurred only when the onset of the noise
was aperiodic or unsignaled, will serve as our substantive exam-
ple. Using such an approach, the essential properties of a mod-
erator variable are summarized in Figure 1.
The model diagrammed in Figure 1 has three causal paths
that feed into the outcome variable of task performance: the
Figure 1.
Moderator model.
impact of the noise intensity as a predictor (Path a), the impact
of controllability as a moderator (Path b), and the interaction
or product of these two (Path c). The moderator hypothesis is
supported if the interaction (Path c) is significant. There may
also be significant main effects for the predictor and the moder-
ator (Paths a and b), but these are not directly relevant concep-
tually to testing the moderator hypothesis.
In addition to these basic considerations, it is desirable that
the moderator variable be uncorrelated with both the predictor
and the criterion (the dependent variable) to provide a clearly
interpretable interaction term. Another property of the moder-
ator variable apparent from Figure 1 is that, unlike the media-
tor-predictor relation (where the predictor is causally anteced-
ent to the mediator), moderators and predictors are at the same
level in regard to their role as causal variables antecedent or
exogenous to certain criterion effects. That is, moderator vari-
ables always function as independent variables, whereas medi-
ating events shift roles from effects to causes, depending on the
focus oftbe analysis.
Choosing an Appropriate Analytic Procedure:
Testing Moderation
In this section we consider in detail the specific analysis pro-
cedures for appropriately measuring and testing moderational
hypotheses. Within this framework, moderation implies that
the causal relation between two variables changes as a function
of the moderator variable. The statistical analysis must measure
and test the differential effect of the independent variable on the
dependent variable as a function of the moderator. The way to
measure and test the differential effects depends in part on the
level of measurement of the independent variable and the mod-
erator variable. We will consider four eases: In Case 1, both
moderator and independent variables are categorical variables;
in Case 2, the moderator is a categorical variable and the inde-
pendent variable a continuous variable; in Case 3, the modera-
1 At a conceptual level, a moderator may be more impressive if we go
from a strong to a weak relation or to no relation at all as opposed to
finding a crossover interaction. That is, although crossover interactions
are stronger statistically, as they are not accompanied by residual main
effects, conceptually no effect shifts may be more impressive.
THE MODERATOR-MEDIATOR DISTINCTION 1175
tor is a continuous variable and the independent variable is a
categorical variable; and in Case 4, both variables are continu-
ous variables. To ease our discussion, we will assume that all the
categorical variables are dichotomies.
Case 1
This is the simplest case. For this case, a dichotomous inde-
pendent variable's effect on the dependent variable varies as a
function of another dichotomy. The analysis is a 2 • 2 ANOVA,
and moderation is indicated by an interaction. We may wish to
measure the simple effects of the independent variable across
the levels of the moderator (Winer, 1971, pp. 435-436), but
these should be measured only if the moderator and the inde-
pendent variable interact to cause the dependent variable.
Case 2
Here the moderator is a dichotomy and the independent vari-
able is a continuous variable. For instance, gender might moder-
ate the effect of intentions on behavior. The typical way to mea-
sure this type of moderator effect is to correlate intentions with
behavior separately for each gender and then test the difference.
For instance, virtually all studies of moderators of the attitude-
behavior relation use a correlational test.
The correlational method has two serious deficiencies. First,
it presumes that the independent variable has equal variance at
each level of the moderator. For instance, the variance of inten-
tion must be the same for the genders. If variances differ across
levels of the moderator, then for levels of the moderator with
less variance, the correlation of the independent variable with
the dependent variable tends to be less than for levels of the
moderator with more variance. The source of this difference is
referred to as a restriction in range (McNemar, 1969). Second,
if the amount of measurement error in the dependent variable
varies as a function of the moderator, then the correlations be-
tween the independent and dependent variables will differ spuri-
ously.
These problems illustrate that correlations are influenced by
changes in variances. However, regression coefficients are not
affected by differences in the variances of the independent vari-
able or differences in measurement error in the dependent vari-
able. It is almost always preferable to measure the effect of the
independent variable on the dependent variable not by correla-
tion coefficients but by unstandardized (not betas) regression
coefficients (Duncan, 1975). Tests of the difference between re-
gression coefficients are given in Cohen and Cohen (1983, p.
56). This test should be performed first, before the two slopes
are individually tested.
If there is differential measurement error in the independent
variable across levels of the moderator, bias results. Reliabilities
would then need to be estimated for the different levels of the
moderator, and slopes would have to be disattenuated. This can
be accomplished within the computer program LISREL-VI
(J6reskog & S6rbom, 1984) by use of the multiple-group op-
tion. The levels of the moderator are treated as different groups.
Case 3
In this case, the moderator is a continuous variable and the
independent variable is a dichotomy. For instance, the indepen-
Figure 2.
Three different ways in which the moderator changes the effect
of the independent variable on the dependent variable: linear (top), qua-
dratic (middle), and step (bottom).
dent variable might be a rational versus fear-arousing attitude-
change message and the moderator might be intelligence as
measured by an IQ test. The fear-arousing message may be
more effective for low-IQ subjects, whereas the rational message
may be more effective for high-IQ subjects. To measure modera-
tor effects in this case, we must know a priori how the effect of
the independent variable varies as a function of the moderator.
It is impossible to evaluate the general hypothesis that the effect
of the independent variable changes as a function of the moder-
ator because the moderator has many levels.
Figure 2 presents three idealized ways in which the modera-
tor alters the effect of the independent variable on the dependent
variable. First, the effect of the independent variable on the de-
pendent variable changes linearly with respect to the moderator.
The linear hypothesis represents a gradual, steady change in the
effect of the independent variable on the dependent variable as
the moderator changes. It is this form of moderation that is gen-
erally assumed. The second function in the figure is a quadratic
function. For instance, the fear-arousing message may be more
generally effective than the rational message for all low-IQ sub-
jects, but as IQ increases, the fear-arousing message loses its ad-
vantage and the rational message is more effective.
The third function in Figure 2 is a step function. At some
critical IQ level, the rational message becomes more effective
than the fear-arousing message. This pattern is tested by dichot-
omizing the moderator at the point where the step is supposed
to occur and proceeding as in Case 1. Unfortunately, theories
in social psychology are usually not precise enough to specify
the exact point at which the step in the function occurs.
The linear hypothesis is tested by adding the product of the
moderator and the dichotomous independent variable to the re-
1176 REUBEN M. BARON AND DAVID A. KENNY
gression equasion, as described by Cohen and Cohen (1983) and
Cleary and Kessler (1982). So if the independent variable is de-
noted as X, the moderator as Z, and the dependent variable as
Y, Y is regressed on X, Z, and
XZ.
Moderator effects are indi-
cated by the significant effect of
XZ
while X and Z are con-
trolled. The simple effects of the independent variable for
different levels of the moderator can be measured and tested by
procedures described by Aiken and West (1986). (Measurement
error in the moderator requires the same remedies as measure-
ment error in the independent variable under Case 2.)
The quadratic moderation effect can be tested by dichotomiz-
ing the moderator at the point at which the function is pre-
sumed to accelerate. If the function is quadratic, as in Figure 2,
the effect of the independent variable should be greatest for
those who are high on the moderator. Alternatively, quadratic
moderation can be tested by hierarchical regression procedures
described by Cohen and Cohen (1983). Using the same notation
as in the previous paragraph, Y is regressed on
X, Z, XZ, Z 2,
and
XZ 2.
The test of quadratic moderation is given by the test
of XZ 2. The interpretation of this complicated regression equa-
tion can be aided by graphing or tabling the predicted values
for various values ofXand Z.
Case 4
In this case both the moderator variable and the independent
variable are continuous. If one believes that the moderator al-
ters the independent-dependent variable relation in a step func-
tion (the bottom diagram in Figure 2), one can dichotomize the
moderator at the point where the step takes place. After dichot-
omizing the moderator, the pattern becomes Case 2. The mea-
sure of the effect of the independent variable is a regression co-
efficient.
If one presumes that the effect of the independent variable
(X) on the dependent variable (Y) varies linearly or quadrati-
cally with respect to the moderator (Z), the product variable
approach described in Case 3 should be used. For quadratic
moderation, the moderator squared must be introduced. One
should consult Cohen and Cohen (1983) and Cleary and Kessler
(1982) for assistance in setting up and interpreting these regres-
sions.
The presence of measurement error in either the moderator
or the independent variable under Case 4 greatly complicates
the analysis. Busemeyer and Jones (1983) assumed that the
moderation is linear and so can be captured by an
XZ
product
term. They showed that measuring multiplicative interactions
when one of the variables has measurement error results in low
power in the test of interactive effects. Methods presented by
Kenny and Judd (1984) can be used to make adjustments for
measurement error in the variables, resulting in proper esti-
mates of interactive effects. However, these methods require
that the variables from which the product variable is formed
have normal distributions.
The Nature of Mediator Variables
Although the systematic search for moderator variables is rel-
atively recent, psychologists have long recognized the impor-
lance of mediating variables. Woodworth's (1928) S-O-R
model, which recognizes that an active organism intervenes be-
tween stimulus and response, is perhaps the most generic for-
mulation of a mediation hypothesis. The central idea in this
model is that the effects of stimuli on behavior are mediated
by various transformation processes internal to the organism.
Theorists as diverse as Hull, Tolman, and Lewin shared a belief
in the importance of postulating entities or processes that inter-
vene between input and output. (Skinner's blackbox approach
represents the notable exception.)
General A nalytic Considerations
In general, a given variable may be said to function as a medi-
ator to the extent that it accounts for the relation between the
predictor and the criterion. Mediators explain how external
physical events take on internal psychological significance.
Whereas moderator variables specify when certain effects will
hold, mediators speak to how or why such effects occur. For
example, choice may moderate the impact of incentive on atti-
tude change induced by discrepant action, and this effect is in
turn mediated by a dissonance arousal-reduction sequence (of.
Brehm & Cohen, 1962).
To clarify the meaning of mediation, we now introduce a path
diagram as a model for depicting a causal chain. The basic
causal chain involved in mediation is diagrammed in Figure 3.
This model assumes a three-variable system such that there are
two causal paths feeding into the outcome variable: the direct
impact of the independent variable (Path c) and the impact of
the mediator (Path b). There is also a path from the independent
variable to the mediator (Path a).
A variable functions as a mediator when it meets the follow-
ing conditions: (a) variations in levels of the independent vari-
able significantly account for variations in the presumed media-
tor (i.e., Path a), (b) variations in the mediator significantly ac-
count for variations in the dependent variable (i.e., Path b), and
(c) when Paths a and b are controlled, a previously significant
relation between the independent and dependent variables is no
longer significant, with the strongest demonstration of media-
tion occurring when Path c is zero. In regard to the last condi-
tion we may envisage a continuum. When Path c is reduced to
zero, we have strong evidence for a single, dominant mediator.
Iftbe residual Path c is not zero, this indicates the operation of
multiple mediating factors. Because most areas of psychology,
including social, treat phenomena that have multiple causes, a
more realistic goal may be to seek mediators that significantly
decrease Path c rather than eliminating the relation between the
independent and dependent variables altogether. From a theo-
retical perspective, a significant reduction demonstrates that a
given mediator is indeed potent, albeit not both a necessary and
a sufficient condition for an effect to occur.
THE MODERATOR-MEDIATOR DISTINCTION 1177
Testing Mediation
An ANOVA provides a limited test ofa mediational hypothesis
as extensively discussed in Fiske, Kenny, and Taylor (1982).
Rather, as recommended by Judd and Kenny (1981 b), a series
of regression models should be estimated. To test for mediation,
one should estimate the three following regression equations:
first, regressing the mediator on the independent variable; sec-
ond, regressing the dependent variable on the independent vari-
able; and third, regressing the dependent variable on both the
independent variable and on the mediator. Separate coefficients
for each equation should be estimated and tested. There is no
need for hierarchical or stepwise regression or the computation
of any partial or semipartial correlations.
These three regression equations provide the tests of the link-
ages of the mediational model. To establish mediation, the fol-
lowing conditions must hold: First, the independent variable
must affect the mediator in the first equation; second, the inde-
pendent variable must be shown to affect the dependent variable
in the second equation; and third, the mediator must affect the
dependent variable in the third equation. If these conditions all
hold in the predicted direction, then the effect of the indepen-
dent variable on the dependent variable must be less in the third
equation than in the second. Perfect mediation holds if the inde-
pendent variable has no effect when the mediator is controlled.
Because the independent variable is assumed to cause the me-
diator, these two variables should be correlated. The presence
of such a correlation results in multicollinearity when the
effects of independent variable and mediator on the dependent
variable are estimated. This results in reduced power in the test
of the coefficients in the third equation. It is then critical that
the investigator examine not only the significance of the co-
efficients but also their absolute size. For instance, it is possible
for the independent variable to have a smaller coefficient when
it alone predicts the dependent variable than when it and the
mediator are in the equation but the larger coefficient is not
significant and the smaller one is.
Sobel (1982) provided an approximate significance test for
the indirect effect of the independent variable on the dependent
variable via the mediator. As in Figure 3, the path from the
independent variable to the mediator is denoted as a and its
standard error is sa; the path from the mediator to the depen-
dent variable is denoted as b and its standard error is sb. The
exact formula, given multivariate normality for the standard er-
ror of the indirect effect or
ab,
is this:
Vb2sa 2 q- a2Sb 2 d- Sa2Sb 2
Sobel's method omits the term
Sa2Sb 2,
but that term ordinarily
is small. His approximate method can be used for more compli-
cated models.
The use of multiple regression to estimate a mediational
model requires the two following assumptions: that there be no
measurement error in the mediator and that the dependent vari-
able not cause the mediator.
The mediator, because it is often an internal, psychological
variable, is likely to be measured with error. The presence of
measurement error in the mediator tends to produce an under-
estimate of the effect of the mediator and an overestimate of
the effect of the independent variable on the dependent variable
when all coefficients are positive (Judd & Kenny, 198 la). Obvi-
ously this is not a desirable outcome, because successful media-
tors may be overlooked.
Generally the effect of measurement error is to attenuate the
size of measures of association, the resulting estimate being
closer to zero than it would be if there were no measurement
error (Judd & Kenny, 198 la). Additionally, measurement error
in the mediator is likely to result in an overestimate in the effect
of the independent variable on the dependent variable. Because
of measurement error in the mediator, effects of the mediator
on the dependent variable cannot totally be controlled for when
measuring the effects of the independent variable on the depen-
dent variable.
The overestimation of the effects of the independent variable
on the dependent variable is enhanced to the extent that the
independent variable causes the mediator and the mediator
causes the dependent variable. Because a successful mediator is
caused by the independent variable and causes the dependent
variable, successful mediators measured with error are most
subject to this overestimation bias.
The common approach to unreliability is to have multiple
operations or indicators of the construct. Such an approach re-
quires two or more operationalizations or indicators of each
construct. One can use the multiple indicator approach and es-
timate mediation paths by latent-variable structural modeling
methods. The major advantages of structural modeling tech-
niques are the following: First, although these techniques were
developed for the analysis of nonexperimental data (e.g., field-
correlational studies), the experimental context actually
strengthens the use of the techniques. Second, all the relevant
paths are directly tested and none are omitted as in ANOVA.
Third, complications of measurement error, correlated mea-
surement error, and even feedback are incorporated directly
into the model. The most common computer program used to
estimate structural equation models is LISREL-VI (JSreskog
& S6rbom, 1984). Also available is the program EQS (Bentler,
1982).
We now turn our attention to the second source of bias in
the mediational chain: feedback. The use of multiple regression
analysis presumes that the mediator is not caused by the depen-
dent variable. It may be possible that we are mistaken about
which variable is the mediator and which is the dependent vari-
able.
Smith (1982) has proposed an ingenious solution to the prob-
lem of feedback in mediational chains. His method involves the
manipulation of two variables, one presumed to cause only the
mediator and not the dependent variable and the other pre-
sumed to cause the dependent variable and not the mediator.
Models of this type are estimated by two-stage least squares or
a related technique. Introductions to two-stage least squares are
in James and Singh (1978), Duncan (1975), and Judd and
Kenny (1981a). The earlier-mentioned structural modeling
procedures can also be used to estimate feedback models.
Overview of Conceptual Distinctions
Between Moderators and Mediators
As shown in the previous section, to demonstrate mediation
one must establish strong relations between (a) the predictor