You are on page 1of 6

Research in Nursing & Health, 2000, 23, 415±420

Focus on Research Methods


Mediator and Moderator
Variables in Nursing Research:
Conceptual and Statistical
Differences
Jill A. Bennett

School of Nursing, San Diego State University, 5500 Campanile Drive, San Diego, CA
Received 6 July 1999; accepted 17 April 2000

Abstract: Mediators and moderators are variables that affect the association
between an independent variable and an outcome variable. Mediators
provide additional information about how or why two variables are strongly
associated. In contrast, moderators explain the circumstances that cause a
weak or ambiguous association between two variables that were expected to
have a strong relationship. Mediators and moderators are often overlooked in
research designs, or the terms are used incorrectly. This article summarizes
the conceptual differences between mediators and moderators. The statistical
analysis of moderators and mediators in multiple regression is brie¯y
described and two examples are presented. ß 2000 John Wiley & Sons, Inc.
Res Nurs Health 23:415±420, 2000

Keywords: mediator; moderator; data analysis

Mediator and moderator variables provide useful that purpose, mediators and moderators are
information about how, why, or when a phenom- de®ned and differentiated. In addition, statistical
enon occurs. Unfortunately, many nurse research- methods for analysis of mediator and moderator
ers fail to test for mediators or moderators in their effects are described and examples presented.
data or use the terms incorrectly. For example, In intervention studies, a mediator or mod-
``mediator'' is often used as if it meant ``pre- erator can explain why a nursing intervention
dictor.'' A mediator or moderator is a third works. For example, Kwekkeboom, Huseby-
variable that changes the association between an Moore, and Ward (1998) hypothesized that a
independent variable and an outcome variable moderator, imaging ability, could explain why
(Baron & Kenny, 1986). Thus, consideration of a guided imagery (an intervention) had produced
mediator or moderator allows a more precise mixed results in reducing cancer pain in earlier
description of the relationship between indepen- studies. The researchers did an exploratory study
dent and outcome variables. If a researcher fails in which guided imagery was used to alleviate
to consider the possibility of a mediator or mode- anxiety and, in addition, they measured a
rator effect in the data, a more exact explanation proposed moderator, individual ability to gener-
for an outcome may be missed. The purpose of ate realistic images. The results showed that
this paper is to discuss the importance of media- participants who had high levels of imaging
tors and moderators in nursing research and to ability (the moderator variable) had a reduction in
encourage more widespread analysis of these anxiety when the guided imagery intervention
effects in answering nursing questions. To achieve was used, but those with low levels of imaging
*Assistant Professor.

ß 2000 John Wiley & Sons, Inc. 415


416 RESEARCH IN NURSING & HEALTH

ability did not reduce their anxiety. Thus, the


effect of the intervention on the outcome
depended on the level of the moderator. The
researchers hope to continue their research on this FIGURE 2. Conceptual model of a mediator
effect.
moderator that may explain the previously
reported weak association between guided ima- independent variable with the outcome variable
gery and cancer pain. ``depends on'' the value (or level) of the
Likewise, the analysis of mediators and moderator variable (Cohen & Cohen, 1983).
moderators in descriptive studies can elicit On the other hand, a mediator is a variable that
information about why or how a direct associa- speci®es how the association occurs between an
tion occurs between an independent variable and independent variable and an outcome variable. A
an outcome variable. For example, Schumacher, mediator effect is only tested when there is a
Dodd, and Paul (1993) studied caregivers of signi®cant direct effect between the independent
persons receiving chemotherapy for cancer and variable and the outcome variable, but there is a
found that caregiver strain had a signi®cant single possibility that a mediator variable conceptually
order correlation with depression. However, the occurs ``between'' the two variables. A mediator
researchers also investigated possible mediators effect, shown in Figure 2, exists if the following
and moderators. They found that coping was a conditions are met: (a) variations in the indepen-
mediator in the association and thus provided a dent variable predict variations in the mediator
fuller picture of how caregiver strain can develop variable, (b) variations in the mediator variable
(or not) into depression. predict variations in the outcome variable, and (c)
When the associations in (a) and (b) are
controlled in the model, the direct relationship
THE DEFINITIONS OF MEDIATOR between the independent variable and the out-
AND MODERATOR EFFECTS come variable becomes nonsigni®cant (Baron &
Kenny, 1986).
Mediators are different from moderators and
many people are confused by the distinctions
between the two. It is important to understand THE DIFFERENCES
whether a mediator or moderator effect is BETWEEN MEDIATOR AND
hypothesized, because they require different MODERATOR EFFECTS
statistical analyses.
A moderator is an independent variable that Sometimes it is dif®cult to distinguish mediators
affects the strength and/or direction of the and moderators when forming hypotheses about
association between another independent variable variables. The de®nitional difference, that a
and an outcome variable. Figure 1 illustrates a mediator is predicted by the independent variable
moderator effect. A moderator variable may and a moderator is a separate independent
initially be analyzed in a multiple regression variable, is important but not always obvious. A
model as one of the independent variables. variable such as ``coping'' is an example. Some
However, subsequent steps in the analysis may researchers conceptualize ``coping'' as a media-
uncover two distinct regression slopes in the data tor because it cannot occur unless a stressor has
that depend on the value of the moderator. The occurred; an independent variable ``stressor''
moderator interacts with the independent variable predicts the mediator ``coping,'' and together they
of interest so that the independent variable's predict an outcome. In contrast, some researchers
association with the outcome variable is stronger conceptualize ``coping'' as a moderator variable.
or weaker at different levels of the moderator In this case, ``coping'' is an independent variable
variable. In other words, the association of the that affects the stressor±outcome relationship;
when coping ability is high, the stressor±outcome
association is weak, whereas when coping ability
is low, the stressor±outcome association is strong
(Holmbeck, 1997). The decision about whether a
variable is a mediator or moderator should be
based on theory and the conceptual framework
that guides the research.
FIGURE 1. Conceptual model of a moderator Baron and Kenny (1986) offer some guidelines
effect. for determining whether a proposed variable
MEDIATOR AND MODERATOR VARIABLES / BENNETT 417

represents a hypothesized mediator or moderator independent variables are categorical or contin-


effect. Mediator-oriented research is usually con- uous, the general strategy is to test for an
cerned with the mechanism of the relationship interaction using hierarchical multiple regression
between the independent variable and the out- analysis. In the ®rst step (or steps) of the
come variable. In other words, the ``how'' and regression, the independent variables (including
``why'' is more interesting to the researcher than the moderator) are entered into the model as
the independent variable itself. In contrast, a predictors of the outcome variable. The indepen-
researcher who includes a moderator in the model dent variables do not have to be signi®cant
wants to know ``when'' the relationship occurs predictors of the outcome variable in order to test
between the independent and outcome variables. for an interaction in the next step. In a separate
Therefore, the researcher is usually more inter- step, an interaction term (the product of two
ested in the independent variable than the independent variables, which represents the
moderator. moderator effect) is entered. If the interaction
It follows then that mediators are usually term explains a statistically signi®cant amount of
investigated when the relationship between the variance in the dependent variable, a moderator
independent variable and the outcome variable is effect is present. Figure 3 show a diagram of this
statistically signi®cant. If this relationship is procedure.
strong, the researcher is likely to be interested The interaction term represents a joint relation-
in ®nding a mediator that explains how or why the ship between the two independent variables and
independent variable predicts the outcome. On this relationship accounts for additional variance
the other hand, if the association between the in the outcome variable beyond that explained by
independent variable and the outcome variable is either single variable alone. In other words,
weak or inconsistent, the researcher may several different regression slopes represent the
hypothesize a moderator, the values of which association, rather than just one, and the associa-
could explain the circumstances that strengthen tion of the independent variable with the outcome
or weaken the association. variable depends on the value of the moderator
A mediator is likely to be an internal property variable. It should be noted that this conditional
of the person or group being studied. If the relationship is symmetrical; it can also be said
individual is the unit of analysis, the mediator that the association of the moderator and the
variable is often an ``in the head'' mechanism that outcome variable depends on the value of the
elaborates the meaning of the relationship independent variable (Cohen & Cohen, 1983).
between the independent variable and the out- A hierarchical regression in which the interac-
come variable. For example, Song and Lee (1996) tion term is entered in its own step allows the
found that depression mediated the relationship researcher to see the main effects of the
between sensory de®cits in elders and their independent variables (the independent variable
functional capacity. However, mediators are not of interest and the moderator) in the earlier steps
always individual-level characteristics. If a group separately from the effect of the moderator in the
is the unit of analysis, group-level characteristics, ®nal step. It is also possible to use a regression
such as cohesiveness or role con¯icts, could be model in which all the variables, including the
mediators. interaction term, are entered in a single step. In
In contrast, the investigation of a moderator this case, the signi®cance of the semi-partial
effect often includes experimental manipulation
of the moderator variable and, therefore, a mode-
rator may be a nursing intervention. However,
moderators may also be naturally occurring vari-
ables. For example, social resources were inves-
tigated by Hall, Sachs, and Rayens (1998) as a
moderator in the relationship between women's
childhood abuse and their potential for abuse of
their own children.

STATISTICAL TESTS FOR A


MODERATOR EFFECT

Though the approach to testing for statistical FIGURE 3. Statistical model of a moderator
signi®cance of a moderator effect varies if the effect.
418 RESEARCH IN NURSING & HEALTH

correlation of the interaction term will show if a outcome variable and (b) the direct relationship
moderator effect is present. However, the main of the independent variable to the outcome
effects of the independent variables cannot be variable is less signi®cant than it was in the
seen unless they are entered in a separate step, second equation (Baron & Kenny, 1986).
because the presence of the interaction term in the
same step alters the variance explained by the
independent variables alone. Thus, the usual AN EXAMPLE OF A
procedure is to use a multi-step hierarchical MEDIATOR EFFECT
regression.
An interaction effect may be dif®cult to detect Yarcheski, Scoloveno, and Mahon (1994)
statistically if the sample does not have a full reported their investigation of a mediator effect
range of values for the independent variable and in an article titled ``Social Support and Well-
the moderator variable. Thus, a homogeneous Being In Adolescents: The Mediating Role of
sample may weaken detection of a moderator Hopefulness.'' The authors provided a theoretical
effect. In addition, measurement error in either of basis for hopefulness as a mediating variable by
the variables is compounded in the interaction citing earlier research studies that provided
variable. This error can be dif®cult to detect using theoretical and empirical evidence for an associa-
multiple regression analysis and may result in an tion between social support and well-being,
underestimation of the moderator effect (Holm- between social support and hope, and between
beck, 1997). hope and well-being. They then hypothesized
that, in some instances, individuals may believe
that hopefulness is important to their well-being
STATISTICAL TESTS FOR A because of their relationships with other people.
MEDIATOR EFFECT Thus, the purpose of the study was correctly and
clearly de®ned as testing the hypothesis that
A mediator effect can also be tested with multiple hopefulness mediated, and therefore helped
regression analysis. However, mediator effects explain, the relationship between perceived social
should only be tested if there is a signi®cant direct support and general well-being in adolescents.
association between an independent variable and Data from 99 participants between 15 and 17
an outcome variable, otherwise there is no years old were analyzed to test the hypothesis.
relationship to mediate. An important assumption First, the single order relationships among the
of this method is that the outcome variable does variables were con®rmed by statistically signi®-
not predict the mediator variable. cant Pearson correlations in the expected direc-
The method, shown in Figure 4, uses three tion. As predicted, perceived social support and
regression equations to test for the statistical well-being were correlated (r ˆ .55), perceived
signi®cance of a mediator effect. The ®rst social support and hopefulness were correlated
equation should show that the independent (r ˆ .57), and hopefulness and well-being were
variable is a signi®cant predictor of the mediator. correlated (r ˆ .60).
The second equation should show that the Next, three regression analyses were per-
independent variable is a signi®cant predictor of formed, following the method speci®ed by Baron
the outcome variable. The third equation should and Kenny (1986) and depicted in Figure 5. The
contain both the independent and mediator ®rst equation regressed hopefulness on perceived
variables entered simultaneously with the out- social support (F(1,97) ˆ 45.90, p < .001). Social
come variable. Two conditions must be met in the support explained 32% of the variance in hope-
third equation if a mediator effect is present: (a) fulness. The second equation regressed well-
the mediator is a signi®cant predictor of the being on social support (F(1,97) ˆ 41.84,
p < .001). Social support explained 30% of the
variance in well-being. The third equation
regressed well-being on both hopefulness and
perceived social support. This ®nal equation met
the two requirements for a mediator effect: (a)
The hypothesized mediator, hopefulness, was a
signi®cant predictor (t ˆ 4.69, p < .001) and
explained 19% of the variance in well-being,
and (b) the variance in well-being explained by
FIGURE 4. Statistical model of a mediator effect. perceived social support was reduced from 30%
MEDIATOR AND MODERATOR VARIABLES / BENNETT 419

FIGURE 5. Example of statistical analysis of a mediator effect (data


from Yarcheski, Scoloveno, and Mahon, 1990).

in the second equation to 9% in the third teristic) was entered as a block in the fourth step.
equation. Thus, the reduced direct association The dependent variable was coping effectiveness.
between perceived social support and well-being Thus, the analysis followed the correct procedure
when hopefulness was in the model supported the for testing moderator effects by entering the
hypothesis that hopefulness was at least one of predictors, including the proposed moderator
the mediators in the relationship between per- variables, into the model before the interaction
ceived social support and general well-being. terms were entered in a separate step.
The results showed that some interaction terms
were signi®cant for some coping strategy models,
AN EXAMPLE OF A indicating that particular demographic variables
MODERATOR EFFECT moderated the effectiveness of the strategy. For
example, the strategy of active cognitive coping
Spitzer, Bar-Tal, and Golander (1995) investi- was more effective for younger subjects than for
gated the hypothesized moderator effect of older (i.e., age was a moderator in the relationship
demographic variables on the relationship between coping strategy and coping effective-
between stress and coping effectiveness. In their ness).
paper, ``The Moderating Effect of Demographic
Variables on Coping Effectiveness,'' the research-
ers stated that the reason for conducting the INCORPORATING MEDIATORS
analysis was the lack of empirical evidence in the AND MODERATORS IN
literature for the theoretical proposition that NURSING RESEARCH
coping is a moderator in the relationship between
stress and adaptation. Therefore, they proposed Analysis of mediator or moderator effects may
that personal and situational variables may supply more in-depth information about a
moderate the effectiveness of coping strategies research phenomenon than can be explained by
in a stressful situation. In other words, a particular direct effects alone. Four elements that should be
coping strategy would be more or less effective included in reports of mediator or moderator
depending on the level of a given demographic research are (a) correct de®nition and use of the
variable. Data from 78 adults who were treated at terms mediator or moderator, (b) a rationale for
an outpatient clinic (mean age 54.2 years) were the hypothesized mediator or moderator effect
analyzed with eight hierarchical regressions for and evidence for the hypothesis based on
each of three different coping strategies. literature and/or conceptual framework, (c)
For each coping strategy that might predict statistical analysis that is matched to the
coping effectiveness, the following hierarchical hypothesized mediator or moderator effect, and
steps were followed. Stress was entered alone in (d) interpretation of the mediator or moderator
the ®rst step. One demographic variable and effect in the ®ndings. Nurse scientists who are
one coping strategy were entered as a block interested in exploring more than just the direct
in the second step. Three interaction terms effects of predictor variables on outcome vari-
(stress  coping strategy, stress  demographic ables may want to consider hypotheses about
characteristic, demographic characteristic  mediators and moderators that could provide
coping strategy) were entered as a block in the additional information about why an observed
third step. A second level interaction term phenomenon occurs or under what circumstances
(stress  coping strategy  demographic charac- a nursing intervention has the greatest effect.
420 RESEARCH IN NURSING & HEALTH

REFERENCES Kwekkeboom, K., Huseby-Moore, K., & Ward, S.


(1998). Imaging ability and effective use of guided
Baron, R.M., & Kenny, D.A. (1986). The moderator- imagery. Research in Nursing & Health, 21, 189±
mediator variable distinction in social psychological 198.
research: Conceptual, strategic, and statistical con- Schumacher, K.L., Dodd, M.J., & Paul, S.M. (1993).
siderations. Journal of Personality & Social Psychol- The stress process in family caregivers of persons
ogy, 51, 1173±1182. receiving chemotherapy. Research in Nursing &
Cohen, J., & Cohen, P. (1983). Applied multiple Health, 16, 395±404.
regression/correlation analysis for the behavioral Song, M., & Lee, E.O. (1996). Development of a
sciences. (2nd ed.). Hillsdale, NJ: Lawrence Erl- functional capacity model for the elderly. Research
baum Associates. in Nursing & Health, 19, 173±181.
Hall, L.A., Sachs, B., & Rayens, M.K. (1998). Mothers' Spitzer, A., Bar-Tal, Y., & Golander, H. (1995). The
potential for child abuse: The roles of childhood moderating effect of demographic variables on
abuse and social resources. Nursing Research, 47, coping effectiveness. Journal of Advanced Nursing,
87±95. 22, 578±585.
Holmbeck, G.N. (1997). Toward terminological, con- Yarcheski, A., Scoloveno, M.A., & Mahon, N.E.
ceptual, and statistical clarity in the study of media- (1994). Social support and well-being in adoles-
tors and moderators: Examples from the child- cents: The mediating role of hopefulness. Nursing
clinical and pediatric psychology literatures. Journal Research, 43, 288±292.
of Consulting & Clinical Psychology, 65, 599±610.

You might also like