mediation analysis - university of british columbiaschaller/528readings/mackinnon... · ·...

ANRV296-PS58-23 ARI 17 November 2006 1:36

Mediation AnalysisDavid P. MacKinnon, Amanda J. Fairchild,and Matthew S. FritzDepartment of Psychology, Arizona State University, Tempe, Arizona 85287-1104;email: [email protected], [email protected], [email protected]

Annu. Rev. Psychol. 2007. 58:593–614

First published online as a Review inAdvance on August 29, 2006

The Annual Review of Psychology is online athttp://psych.annualreviews.org

This article’s doi:10.1146/annurev.psych.58.110405.085542

Copyright c© 2007 by Annual Reviews.All rights reserved

0066-4308/07/0203-0593$20.00

Key Words

intervening variable, indirect effect, third variable, mediator

AbstractMediating variables are prominent in psychological theory and re-search. A mediating variable transmits the effect of an independentvariable on a dependent variable. Differences between mediatingvariables and confounders, moderators, and covariates are outlined.Statistical methods to assess mediation and modern comprehensiveapproaches are described. Future directions for mediation analysisare discussed.

593

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Contents

INTRODUCTION. . . . . . . . . . . . . . . . . 594Definitions . . . . . . . . . . . . . . . . . . . . . . . 595Mediation in Psychological

Research . . . . . . . . . . . . . . . . . . . . . . 596Experimental Approaches to

Mediation. . . . . . . . . . . . . . . . . . . . . 597SINGLE-MEDIATOR MODEL . . . . 598

Mediation Regression Equations . . 598Plotting the Mediation

Equations . . . . . . . . . . . . . . . . . . . . . 599Standard Error of the Mediated

Effect . . . . . . . . . . . . . . . . . . . . . . . . . 599Confidence Limits for the

Mediated Effect . . . . . . . . . . . . . . . 600Significance Testing . . . . . . . . . . . . . . 601Distribution of the Product . . . . . . . 601Computer-Intensive Analysis . . . . . . 601Assumptions of the

Single-Mediator Model . . . . . . . . 602Complete Versus Partial

Mediation. . . . . . . . . . . . . . . . . . . . . 602Consistent and Inconsistent

Models . . . . . . . . . . . . . . . . . . . . . . . 602Effect Size Measures of

Mediation. . . . . . . . . . . . . . . . . . . . . 603EXTENSIONS OF THE

SINGLE-MEDIATORMODEL. . . . . . . . . . . . . . . . . . . . . . . . . 603Multilevel Mediation Models . . . . . 603Mediation with Categorical

Outcomes . . . . . . . . . . . . . . . . . . . . . 604Multiple Mediators . . . . . . . . . . . . . . . 604Longitudinal Mediation Models . . . 604Moderation and Mediation . . . . . . . 605Causal Inference . . . . . . . . . . . . . . . . . 607

SUMMARY AND FUTUREDIRECTIONS . . . . . . . . . . . . . . . . . . 608

INTRODUCTION

Mediating variables form the basis of manyquestions in psychology:

� Will changing social norms about sci-ence improve children’s achievement inscience?

� If an intervention increases secure at-tachment among young children, do be-havioral problems decrease when thechildren enter school?

� Does physical abuse in early childhoodlead to deviant processing of social in-formation that leads to aggressive be-havior?

� Do expectations start a self-fulfillingprophecy that affects behavior?

� Can changes in cognitive attributionsreduce depression?

� Does trauma affect brain stem activa-tion in a way that inhibits memory?

� Does secondary rehearsal increase im-age formation, which increases wordrecall?

Questions like these suggest a chain ofrelations where an antecedent variable af-fects a mediating variable, which then af-fects an outcome variable. As illustrated inthe questions, mediating variables are behav-ioral, biological, psychological, or social con-structs that transmit the effect of one vari-able to another variable. Mediation is oneway that a researcher can explain the processor mechanism by which one variable affectsanother.

One of the primary reasons for the popu-larity of mediating variables in psychology isthe historical dominance of the stimulus or-ganism response model (Hebb 1966). In thismodel, mediating mechanisms in the organ-ism translate how a stimulus leads to a re-sponse. A second related reason for the im-portance of mediating variables is that theyform the basis of many psychological theories.For example, in social psychology, attitudescause intentions, which then cause behavior(Fishbein & Ajzen 1975), and in cognitivepsychology, memory processes mediate howinformation is transmitted into a response.A newer application of the mediating vari-able framework is in prevention and treat-ment research, where interventions are de-signed to change the outcome of interest bytargeting mediating variables that are hypoth-esized to be causally related to the outcome.

594 MacKinnon · Fairchild · Fritz

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


A third reason for interest in mediation ismethodological. Mediation represents theconsideration of how a third variable affectsthe relation between two other variables. Al-though the consideration of a third variablemay appear simple, three-variable systems canbe very complicated, and there are many al-ternative explanations of observed relationsother than mediation. This methodologicaland statistical challenge of investigating me-diation has made methodology for assessingmediation an active research topic.

This review first defines the mediatingvariable and the ways in which it differs fromother variables, such as a moderator or a con-founder. Examples of mediating variables usedin psychology are provided. Statistical meth-ods to assess mediation in the single-mediatorcase are described, along with their assump-tions. These assumptions are addressed in sec-tions describing current research on the statis-tical testing of mediated effects, longitudinalmediation models, models with moderators aswell as mediators, and causal inference for me-diation models. Finally, directions for futureresearch are outlined.

Definitions

Most research focuses on relations betweentwo variables, X and Y, and much has beenwritten about two-variable relations, includ-ing conditions under which X can be consid-ered a possible cause of Y. These conditionsinclude randomization of units to values of Xand independence of units across and withinvalues of X. Mediation in its simplest formrepresents the addition of a third variable tothis X → Y relation, whereby X causes themediator, M, and M causes Y, so X→ M →Y. Mediation is only one of several relationsthat may be present when a third variable, Z(using Z to represent the third variable), is in-cluded in the analysis of a two-variable system.One possibility is that Z causes both X and Y,so that ignoring Z leads to incorrect inferenceabout the relation of X and Y; this would be anexample of a confounding variable. In another

situation, Z may be related to X and/or Y, sothat information about Z improves predictionof Y by X, but does not substantially alter therelation of X to Y when Z is included in theanalysis; this is an example of a covariate. Zmay also modify the relation of X to Y suchthat the relation of X to Y differs at differentvalues of Z; this is an example of a moderatoror interaction effect. The distinction betweena moderator and mediator has been an ongo-ing topic of research (Baron & Kenny 1986,Holmbeck 1997, Kraemer et al. 2001). A me-diator is a variable that is in a causal sequencebetween two variables, whereas a moderator isnot part of a causal sequence between the twovariables. More detailed definitions of thesevariables in a three-variable system may befound in Robins & Greenland (1992).

The single-mediator model is shown inFigure 1, where the variables X, M, and Yare in rectangles and the arrows represent re-lations among variables. Figure 1 uses the no-tation most widely applied in psychology, witha representing the relation of X to M, b rep-resenting the relation of M to Y adjusted forX, and c ′ the relation of X to Y adjusted forM. The symbols e2 and e3 represent residu-als in the M and Y variables, respectively. Theequations and coefficients corresponding toFigure 1 are discussed below. For now, notethat there is a direct effect relating X to Y anda mediated effect by which X indirectly affects

Figure 1Mediation model.

www.annualreviews.org • Mediation Analysis 595

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Table 1 Subject area coverage in currentmediation research

Subject area # Articles citedSocial psychology 98Clinical psychology 70Health psychology 29Developmental psychology 27IO psychology 24Cognitive psychology 18Quantitative psychology(methods)

12

Program evaluation 8Educational psychology 3Environmental psychology 1Evolutionary psychology 1

Y through M. Given that most prior media-tion research has applied this single-mediatormodel, this review starts with this model. Lim-itations and extensions of the model are de-scribed in subsequent sections.

When thinking of mediation, it is help-ful to understand that two models exist: Oneis theoretical, corresponding to unobservablerelations among variables, and the other isempirical, corresponding to statistical analy-ses of actual data (MacCorquodale & Meehl1948). The challenging task of research isto infer the true state of mediation fromobservations. There are qualifications evento this simple dichotomy, and in general, itwill take a program of research to justifyconcluding that a third variable is a mediatingvariable.

Mediation in Psychological Research

In order to ascertain how often mediation isused in psychology, a search was conductedusing the PsycInfo search engine for articlescontaining the word “mediation” in the titleand citing the most widely cited article formediation methods, Baron & Kenny (1986).This search yielded 291 references. Of thesearticles, 80 came from American Psycholog-ical Association (APA) journals. Publications

earlier than the year 2000 were primarilyAPA sources, but there was a surge in non-APA articles after that time. The majority ofthese sources (239 citations) examined medi-ation alone, and 52 investigated both medi-ation and moderation effects. These studiesincluded a mix of cross-sectional and longitu-dinal data, and ordinary least squares regres-sion and structural equation modeling werethe primary analytic methods. The articlescovered a wide range of substantive areas,including social psychology (98 articles) andclinical psychology (70); a complete break-down is listed in Table 1.

Mediation studies, such as those discussedabove, are of two general but overlappingtypes. One type consists of investigating howa particular effect occurs. These studies usu-ally occur after an observed X → Y relation isfound. This approach stems from the elabo-ration methodologies outlined by Lazarsfeld(1955) and Hyman (1955). In this framework,a third variable is added to the analysis of anX → Y relation in order to improve under-standing of the relation or to determine if therelation is spurious. A mediating variable im-proves understanding of such a relation be-cause it is part of the causal sequence of X →M → Y. For example, physical abuse in earlychildhood is associated with violence later inlife. One explanation of this pattern is thatchildren exposed to physical violence acquiredeviant patterns of processing social informa-tion that lead to later violent behavior. Dodgeet al. (1990) found evidence for this theo-retical mediating process because social pro-cessing measures explained the relation be-tween early childhood physical abuse and lateraggressive behavior.

The second type of study uses theory re-garding mediational processes to design ex-periments. Some of the best examples of thisapproach are found in the evaluation of treat-ment and prevention programs. In this re-search, an intervention is designed to changemediating variables that are hypothesized tobe causally related to a dependent variable.If the hypothesized relations are correct, a


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


prevention or treatment program that sub-stantially changes the mediating variables willin turn change the outcome. Primary pre-vention programs, such as drug preventionprograms, are designed to increase resistanceskills, educate, and change norms to reducedrug use. Secondary prevention programssuch as campaigns to increase screening ratesfor serious illness (Murray et al. 1986) edu-cate and change norms regarding health toincrease screening rates. In both of these ex-amples, a mediator that transmits the effect ofan independent variable on a dependent vari-able is first identified by theory and later testedin an experiment. Researchers from manyfields have stressed the importance of assess-ing mediation in treatment and preventionresearch (Baranowski et al. 1998; Donaldson2001; Judd & Kenny 1981a,b; Kraemer et al.2002; MacKinnon 1994; Shadish 1996; Weiss1997). First, mediation analysis provides acheck on whether the program produced achange in the construct it was designed tochange. If a program is designed to changenorms, then program effects on normativemeasures should be found. Second, mediationanalysis results may suggest that certain pro-gram components need to be strengthened ormeasurements need to be improved, as fail-ures to significantly change mediating vari-ables occur either because the program wasineffective or the measures of the mediat-ing construct were not adequate. Third, pro-gram effects on mediating variables in the ab-sence of effects on outcome measures suggestthat program effects on outcomes may emergelater or that the targeted constructs were notcritical in changing outcomes. Fourth, media-tion can sometimes be used to discover prox-imal outcomes that can be used as a surro-gate for an ultimate outcome. For example,in medical studies to reduce death owing to adisease, instead of waiting until death, a moreproximal outcome such as disease symptomsmay be identified. Finally, and most impor-tantly, mediation analysis generates evidencefor how a program achieved its effects. Identi-fication of the critical ingredients can stream-

line and improve these programs by focusingon effective components.

Experimental Approachesto Mediation

Many psychological studies investigating me-diation use a randomized experimental design,where participants are randomized to levelsof one or more factors in order to demon-strate a pattern of results consistent with onetheory and inconsistent with another theory(MacKinnon et al. 2002a, Spencer et al. 2005,West & Aiken 1997). Differences in meansbetween groups are then attributed to theexperimental manipulation of the mediator.The results of the randomized study alongwith the predictions of different theories areused to provide evidence for a mediation hy-pothesis and suggest further studies to local-ize and validate the mediating process. Forexample, a researcher may randomize indi-viduals to conditions that will or will not in-duce cognitive dissonance. In one such study,Sherman & Gorkin (1980) randomly assignedsubjects to solve either (a) a sex-role relatedbrainteaser, or (b) a brainteaser not relatedto sex roles. The sexist brainteaser conditionwas designed to evoke cognitive dissonance inthe self-identified feminist subjects, while thenonsex-role related condition was not. Par-ticipants were then asked to judge the fairnessof a legal decision made in an affirmative ac-tion trial. The results were consistent with theprediction that participants with strong femi-nist beliefs were more likely to make extremefeminist judgments in the trial if they failedthe sexist brainteaser task, in an attempt toreduce cognitive dissonance. Although resultsof this experiment were taken as evidence of acognitive dissonance mediation relation, themediating variable of cognitive conflict wasnot measured to obtain more information onthe link between the manipulation, cognitivedissonance, and feminist judgments.

Double randomization. In some designs itmay be possible to investigate a mediational


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


process by a randomized experiment to inves-tigate the X → M relation and a second ran-domized experiment to investigate the M → Yrelation (MacKinnon et al. 2002a, Spenceret al. 2005, West & Aiken 1997). Spenceret al. (2005) recently summarized two experi-ments reported by Word et al. (1974) that ex-ecuted this design in a study of self-fulfillingprophecy for racial stereotypes. In study 1,white participants were randomly assigned tointerview a black or white confederate. Us-ing measures from the participants, black ap-plicants received less immediacy, higher ratesof speech errors, and shorter interviews thandid white confederates. This part of the studydemonstrated that race of applicant (X) signif-icantly affected interview quality (M). In study2, confederate white interviewers interviewedthe participants from study 1. The confed-erate interviewers either gave interviews likewhite applicants were given in study 1 orthey interviewed applicants with less immedi-acy, higher rates of speech errors, and shorteramounts of interviewer time, like black appli-cants. Here the M variable, type of interview,was randomized and the behavior of the ap-plicants, the Y variable, was measured. Theresults of study 2 indicated that participantstreated like blacks in study 1 performed lessadequately and were more nervous in the in-terview than participants treated like whitesin study 1. Although this type of experimentdoes much to reduce alternative explanationsof the mediation hypothesis, it may be dif-ficult to implement double randomization inother research contexts. Generally, the mostdifficult aspect of the design is the ability torandomly assign participants to the levels ofthe mediator so that the M → Y relation canbe studied experimentally.

SINGLE-MEDIATOR MODEL

Mediation Regression Equations

Experimental studies in psychology rarely in-volve both manipulation of the mediator andmeasurement of mediating variables. If a re-

search study includes measures of a mediat-ing variable as well as the independent anddependent variable, mediation may be inves-tigated statistically (Fiske et al. 1982). In thisway, mediation analysis is a method to increaseinformation obtained from a research studywhen measures of the mediating process areavailable.

There are three major approaches to sta-tistical mediation analysis: (a) causal steps, (b)difference in coefficients, and (c) product ofcoefficients (MacKinnon 2000). All of thesemethods use information from the followingthree regression equations:

Y = i1 + c X + e1, 1.

Y = i2 + c ′X + bM + e2, 2.

M = i3 + aX + e3, 3.

where i1 and i2 and i3 are intercepts, Y is thedependent variable, X is the independent vari-able, M is the mediator, c is the coefficientrelating the independent variable and the de-pendent variable, c ′ is the coefficient relat-ing the independent variable to the depen-dent variable adjusted for the mediator, b isthe coefficient relating the mediator to the de-pendent variable adjusted for the independentvariable, a is the coefficient relating the inde-pendent variable to the mediator, and e1, e2,and e3 are residuals. Equations 2 and 3 are de-picted in Figure 1. Note that the mediationequations may be altered to incorporate linearas well as nonlinear effects and the interactionof X and M in Equation 2, as described laterin this review.

The most widely used method to assessmediation is the causal steps approach out-lined in the classic work of Baron & Kenny(1986; also Kenny et al. 1998) and Judd &Kenny (1981a, 1981b). Four steps are in-volved in the Baron and Kenny approach toestablishing mediation. First, a significant re-lation of the independent variable to the de-pendent variable is required in Equation 1.Second, a significant relation of the indepen-dent variable to the hypothesized mediatingvariable is required in Equation 3. Third, the


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


mediating variable must be significantly re-lated to the dependent variable when boththe independent variable and mediating vari-able are predictors of the dependent variablein Equation 2. Fourth, the coefficient relat-ing the independent variable to the depen-dent variable must be larger (in absolute value)than the coefficient relating the independentvariable to the dependent variable in the re-gression model with both the independentvariable and the mediating variable predict-ing the dependent variable. This causal stepsapproach to assessing mediation has been themost widely used method to assess mediation.As discussed below, there are several limita-tions to this approach.

The mediated effect in the single-mediatormodel (see Figure 1) may be calculated in twoways, as either a b or c − c ′ (MacKinnon &Dwyer 1993). The value of the mediated orindirect effect estimated by taking the differ-ence in the coefficients, c − c ′, from Equations1 and 2 corresponds to the reduction in theindependent variable effect on the dependentvariable when adjusted for the mediator. Totest for significance, the difference is then di-vided by the standard error of the differenceand the ratio is compared to a standard normaldistribution.

The product of coefficients method, in-volves estimating Equations 2 and 3 and com-puting the product of a and b , a b , to form themediated or indirect effect (Alwin & Hauser1975). The rationale behind this method isthat mediation depends on the extent to whichthe program changes the mediator, a, and theextent to which the mediator affects the out-come variable, b. To test for significance, theproduct is then divided by the standard errorof the product and the ratio is compared to astandard normal distribution.

The algebraic equivalence of the a b andc − c ′ measures of mediation was shown byMacKinnon et al. (1995) for normal theory or-dinary least squares and maximum likelihoodestimation of the three mediation regressionequations. For multilevel models (Krull &MacKinnon 1999), logistic or probit regres-

sion (MacKinnon & Dwyer 1993), and sur-vival analysis (Tein & MacKinnon 2003), thea b and c − c ′ estimators of the mediated effectare not always equivalent, and a transforma-tion is required for the two to yield similarresults (MacKinnon & Dwyer 1993).

Plotting the Mediation Equations

The quantities in Equations 1–3 can also bepresented geometrically, as shown in Figure 2(MacKinnon 2007; R. Merrill, unpublisheddissertation). Artificial data are plotted inFigure 2, where the independent variable, X,is dichotomous (to simplify the plot), the me-diator, M, is on the horizontal axis, and thedependent variable, Y, is on the vertical axis.The two slanted lines in the plot representthe relation of M to Y in each X group, oneline for the control group and one line for thetreatment group. The two lines are parallel(note that if there were an XM interaction inEquation 2, then the slopes would not be par-allel), with the slope of each line equal to theb coefficient (b = 0.91, s e b = 0.18). The dis-tance between the horizontal lines in the plotsis equal to the overall effect of X on Y, c (c =1.07, s e c = 0.27), and the distance betweenthe vertical lines is equal to the effect of X onM, a (a = 0.87, s ea = 0.23). The mediatedeffect is the change in the regression line re-lating M to Y for a change in M of a units asshown in the graph. The indirect effect, a b , isequal to c − c ′(c ′ = 0.23, s e c ′ = 0.24). Plotsof the mediated effect may be useful to inves-tigate the distributions of data for outliers andto improve understanding of relations amongvariables in the mediation model.

Standard Error of the MediatedEffect

Sobel (1982, 1986) derived the asymptoticstandard error of the indirect effect using themultivariate delta method (Bishop et al. 1975)in Equation 4. This is the most commonlyused formula for the standard error of the


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


1

43 5 6 7

1

2

3

4

5

6

7

M

Y

aM=i3 (Equation 3 for X=0) M=i3+a (Equation 3 for X=1)

c Total Effect

Y=i1 (Equation 1 for X=0)

Y=i1+c (Equation 1 for X=1)

ab=c-c' Mediated Effect

c'

Y=i2+c'X+bM (Equation 2 for X=0)

Y=i2+c'X+bM (Equation 2 for X=1)

Figure 2Plot of the mediated effect. To simplify figure, no hats are included above coefficient estimates.

mediated effect.

σa b =√

σ 2a b2 + σ 2

ba2 4.

Other formulas for the standard error ofa b and c − c ′ are described in MacKinnonet al. (2002a).

Simulation studies indicate that the esti-mator of the standard error in Equation 4shows low bias for sample sizes of at least 50in single-mediator models (MacKinnon et al.1995, 2002a). In models with more than onemediator, the standard error is accurate forminimum sample sizes of 100–200 (Stone &Sobel 1990). Similar results were obtained forstandard errors of negative and positive pathvalues, and larger models with multiple medi-ating, independent, and dependent variables

(MacKinnon et al. 2002a, 2004; J. Williams,unpublished dissertation).

Confidence Limits for the MediatedEffect

The standard error of a b can be used to test itsstatistical significance and to construct confi-dence limits for the mediated effect as shownin Equation 5:

a b ± z1−ω/2∗σa b . 5.

Confidence limits based on the normal dis-tribution for the mediated effect are ofteninaccurate as found in simulation studies(MacKinnon et al. 1995, 2002a; Stone & Sobel1990) and from bootstrap analysis of the me-diated effect (Bollen & Stine 1990, Lockwood


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


& MacKinnon 1998). These mediated effectconfidence intervals tend to lie to the left ofthe true value of the mediated effect for posi-tive mediated effects and to the right for neg-ative mediated effects (Bollen & Stine 1990,MacKinnon et al. 1995, Stone & Sobel 1990).Asymmetric confidence limits based on thedistribution of the product and bootstrap es-timation have better coverage than these tests(MacKinnon et al. 2004).

Significance Testing

A simulation study of 14 methods to assessthe mediated effect found that the power todetect mediated effects using the most widelyused causal step methods was very low, as weretype I error rates (MacKinnon et al. 2002a,2004). Low power was also observed for testsbased on the normal distribution for mediatedeffect estimators (i.e., a b and c − c ′) dividedby their respective standard errors (Hoyle &Kenny 1999). A joint test of the significanceof a and b was a good compromise betweentype I and type II errors.

There are several explanations for the lowpower of most tests for mediation. First of all,the requirement that there be a significant Xto Y relation in the Baron and Kenny causalsteps test severely reduces power to detect me-diation, especially in the case of complete me-diation (i.e., direct effect is zero). There aremany cases where significant mediation existsbut the requirement of a significant relationof X to Y is not obtained. A recent study usingempirical approaches to determine requiredsample size for 0.8 power to detect a mediatedeffect with small effect size values of the a andb path required approximately 21,000 subjectsfor the causal steps test (Fritz & MacKinnon2007). As the size of the direct effect getslarger, the power to detect mediation usingthe causal steps approach approximates powerto detect mediation by testing whether boththe a and the b paths are statistically signifi-cant. It is important to note that the overall re-lation of X and Y represents important infor-mation for a research study, and in some stud-

ies it may be useful to require an overall X to Yrelation. The point is that requiring an X to Yrelation substantially reduces power to detectreal mediation effects. An explanation for thelow power of tests of mediation based on di-viding an estimator, either a b or c − c ′, of themediated effect by its corresponding standarderror is that the resulting ratio does not alwaysfollow a normal distribution (MacKinnonet al. 2004). Resampling methods and meth-ods based on the distribution of the productof ab address these sampling problems and aredescribed below.

Distribution of the Product

The product of two normally distributed ran-dom variables is normally distributed onlyin special cases (Springer 1979), which ex-plains the inaccuracy of methods of assess-ing statistical significance of mediation basedon the normal distribution. For example, fortwo standard normal random variables witha mean of zero, the excess kurtosis is equalto six (Meeker et al. 1981) compared to an ex-cess kurtosis of zero for a normal distribution.MacKinnon et al. (2002a, 2004a) showed thatin comparison with commonly used meth-ods, significance tests for the mediated effectbased on the distribution of the product hadmore accurate type-I error rates and statis-tical power. A new program, PRODCLIN(MacKinnon et al. 2006a; program down-load available at http://www.public.asu.edu/∼davidpm/ripl/Prodclin/), can now beused to find critical values of the distributionof the product and to compute confidence lim-its for the mediated effect.

Computer-Intensive Analysis

Computer-intensive methods use the ob-served data to generate a reference distribu-tion, which is then used for confidence inter-val estimation and significance testing (Manly1997, Mooney & Duval 1993, Noreen 1989).Programs to compute confidence limits ofthe mediated effect for bootstrap methods is


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


described in Preacher & Hayes (2004) andLockwood & MacKinnon (1998); the AMOS(Arbuckle 1997), EQS (Bentler 1997), LIS-REL (Joreskog & Sorbom 1993), and Mplus(Muthen & Muthen 1998–2006) programsalso conduct bootstrap resampling for the me-diated effect.

Computer-intensive methods, also calledresampling methods, for mediation are im-portant for at least two reasons (Bollen &Stine 1990, MacKinnon et al. 2004, Shrout &Bolger 2002). First, these methods provide ageneral way to test significance and constructconfidence intervals in a wide variety of sit-uations where analytical formulas for quanti-ties may not be available. Second, the methodsdo not require as many assumptions as othertests, which is likely to make them more ac-curate than traditional mediation analysis.

Assumptions of the Single-MediatorModel

There are several important assumptions fortests of mediation. For the a b estimator ofthe mediated effect, the model assumes thatthe residuals in Equations 2 and 3 are inde-pendent and that M and the residual in Equa-tion 2 are independent (McDonald 1997; R.Merrill, unpublished dissertation). It is alsoassumed that there is not an XM interactionin Equation 3, although this can and shouldbe routinely tested. The assumptions of a cor-rectly specified model include no misspeci-fication of causal order (e.g., Y → M → Xrather than X → M → Y), no misspecifica-tion of causal direction (e.g., there is recip-rocal causation between the mediator and thedependent variable), no misspecification dueto unmeasured variables that cause variablesin the mediation analysis, and no misspecifica-tion due to imperfect measurement (Holland1988, James & Brett 1984, McDonald 1997).These assumptions may be difficult to test andmay be untestable in most situations so thatproof of a mediation relation is impossible.A more realistic approach is to incorporateadditional information from prior research,

including randomized experimental studies,theory, and qualitative methods to bolster thetentative conclusion that a mediation relationexists.

Complete Versus Partial Mediation

Researchers often test whether there is com-plete or partial mediation by testing whetherthe c ′ coefficient is statistically significant,which is a test of whether the association be-tween the independent and dependent vari-able is completely accounted for by the me-diator (see James et al. 2006). If the c ′ coef-ficient is statistically significant and there issignificant mediation, then there is evidencefor partial mediation. Because psychologicalbehaviors have a variety of causes, it is of-ten unrealistic to expect that a single mediatorwould be explained completely by an indepen-dent variable to dependent variable relation(Judd & Kenny 1981a).

Consistent and Inconsistent Models

Inconsistent mediation models are modelswhere at least one mediated effect has adifferent sign than other mediated or directeffects in a model (Blalock 1969, Davis 1985,MacKinnon et al. 2000). Although knowledgeof the significance of the relation of X to Yis important for the interpretation of results,there are several examples in which an overallX to Y relation may be nonsignificant, yet me-diation exists. For example, McFatter (1979)described the hypothetical example of work-ers making widgets, where X is intelligence, Mis boredom, and Y is widget production. Intel-ligent workers tend to get bored and produceless, but smarter workers also tend to makemore widgets. Therefore, the overall relationbetween intelligence and widgets producedmay actually be zero, yet there are two oppos-ing mediational processes. A number of otherresources provide examples of these incon-sistent effects (Paulhus et al. 2004, Sheets &Braver 1999). Inconsistent mediation is morecommon in multiple mediator models where


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


mediated effects have different signs. Incon-sistent mediator effects may be especiallycritical in evaluating counterproductive ef-fects of experiments, where the manipulationmay have led to opposing mediated effects.

Effect Size Measures of Mediation

The raw correlation for the a path and thepartial correlation for the b path are effect sizemeasures for mediation models. Standardizedregression coefficients may also serve as ef-fect size measures for individual paths in themediated effect. There are other effect sizemeasures of the entire mediated effect ratherthan individual paths. The proportion medi-ated, 1 − ( c ′

c ) = a b(a b+c ′) , is often used, but val-

ues of the proportion mediated are often verysmall and focusing on an overall proportionmediated may neglect additional mediatingmechanisms (Fleming & DeMets 1996). Theproportion mediated is also unstable unlesssample size is at least 500 (Freedman 2001,MacKinnon et al. 1995). Alwin & Hauser(1975) suggest taking the absolute values ofthe direct and indirect effects prior to calcu-lating the proportion mediated for inconsis-tent models. More work is needed on effectsize measures for mediation.

EXTENSIONS OF THESINGLE-MEDIATOR MODEL

Many important extensions have addressedlimitations of the mediation approach de-scribed above. First, many studies hypoth-esize more complicated models includingmultiple independent variables, multiple me-diators, and multiple outcomes. These modelsmay include hypotheses regarding the com-parison of mediated effects. Second, media-tion in multilevel models may be especiallyimportant, as mediation relations at differ-ent levels of analysis are possible (Krull &MacKinnon 1999, 2001; Raudenbush &Sampson 1999). Third, mediation effects maydiffer by subgroups defined by variables bothwithin the mediation model and outside

the mediation model. Fourth, mediation re-quires temporal precedence from X to M toY, and longitudinal mediation models havebeen developed (Gollob & Reichardt 1991,Kraemer et al. 2002). Finally, developmentsin the causal interpretation of research results(Holland 1988, Robins & Greenland 1992)provide a general framework to understandthe limitations and strengths of possible causalinferences from a mediation study. Each ofthese extensions is described below.

Multilevel Mediation Models

Many studies measure data clustered at sev-eral levels, such as individuals in schools, class-rooms, therapy groups, or clinics. If mediationanalysis from these types of studies is analyzedat the individual level, ignoring the clustering,then type I error rates can be too high (Krull &MacKinnon 1999, 2001). These problems oc-cur because observations within a cluster tendto be dependent so that the independent ob-servations assumption is violated. The inves-tigation of mediation effects at different levelsof analysis also may be important for substan-tive reasons (Hofmann & Gavin 1998). Forexample, a mediated effect present at the ther-apy group level may not be present at the in-dividual level. Similarly, it is possible that themechanism that mediates effects at the schoollevel, such as overall norms, may be differentfrom the mechanism that mediates effects atthe individual level.

Kenny et al. (2003) demonstrated that insome cases the a and the b paths may rep-resent random effects. For example, assumethat X, M, and Y are measured from indi-viduals in schools and that the researcher isinterested in the mediation effect, but the a,b, and c ′ coefficients may vary significantlyacross schools rather than having a single fixedeffect. If a and b are random effects, they maycovary, and an appropriate standard error andpoint estimate for the mediated effect mustallow for this covariance between random ef-fects to be applied. Kenny et al. used a re-sampling method to obtain a value for this


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


covariance. Other methods that have been re-cently proposed to assess this covariance con-sist of combining Equations 2 and 3 into thesame analysis (Bauer et al. 2006) and directlyestimating the covariance among the randomeffects in the Mplus program (Muthen &Muthen 1998–2006).

Mediation with CategoricalOutcomes

In some mediation analyses, the dependentvariable is categorical, such as whether a per-son used drugs or not. In this case, Equations1 and 2 must be rewritten for logistic or pro-bit regression, where the dependent variableis typically a latent continuous variable thathas been dichotomized in analysis. Becausethe residual in each logistic or probit equationis fixed, the parameters c, c ′, and b depend onthe other independent variables in the model.Therefore, the c − c ′ method of estimatingmediation is incorrect because the parameterestimate of c ′ depends on the effect explainedby the mediator and the scaling of Equations1 and 2 (MacKinnon & Dwyer 1993). Onesolution to this problem is to standardize re-gression coefficients prior to estimating me-diation (Winship & Mare 1983). If the me-diator is treated as a continuous variable, aproduct of coefficients test of the mediatedeffect may be obtained using a from ordinaryleast squares regression and b from logistic re-gression. Again, better confidence limits andstatistical tests are obtained if critical valuesfrom the distribution of the product or boot-strap methods are used (D.P. MacKinnon,M. Yoon, C.M. Lockwood, & A.B. Taylor,unpublished manuscript).

Multiple Mediators

Mediating processes may include multiplemediators, dependent variables, and/or inde-pendent variables. In school-based drug pre-vention, for example, prevention programstarget multiple mediators such as resistanceskills, social norms, attitudes about drugs, and

communication skills. The multiple-mediatormodel is likely to provide a more accurateassessment of mediation effects in many re-search contexts. Models with more than onemediator are straightforward extensions ofthe single-mediator case (MacKinnon 2000).Several standard error formulas for compar-ing different mediated effects are given byMacKinnon (2000), and the methods are illus-trated with data from a drug prevention study.

Longitudinal Mediation Models

Longitudinal data allow a researcher to ex-amine many aspects of a mediation modelthat are unavailable in cross-sectional data,such as whether an effect is stable acrosstime and whether there is evidence for one ofthe important conditions of causality, tempo-ral precedence. Longitudinal data also bringchallenges, including nonoptimal measure-ment times, omitted variables or paths, anddifficult specification of the correct longitudi-nal mediated effect of interest (Cheong et al.2003, Cole & Maxwell 2003, Collins et al.1998).

There are three major types of longitudinalmediation models. The autoregressive modelwas described by Gollob & Reichardt (1991)and elaborated by Cole & Maxwell (2003). Inthe basic autoregressive model, relations thatare one measurement occasion (wave) apartare specified, and the relation between thesame variable over time is specified to assessstability, as are covariances among the vari-ables at the first wave and the covariancesamong the residual variances of X, M, and Yat later waves. The covariances among X, M,and Y at the same wave of measurement re-flect that the causal order of these measures isunknown. Only relations consistent with lon-gitudinal mediation are estimated among thevariables.

A second form of the autoregressive medi-ation model includes contemporaneous medi-ation relations among X, M, and Y, such thatmediation can occur within the same wavein addition to longitudinal mediation across


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


waves. Practically, this would occur if therewere a change in the mediator that led tochange in the outcome between the first andsecond wave of measurement. Still anotherform of the autoregressive longitudinal medi-ation model allows for cross-lagged relationsamong variables, where the direction of therelations among X, M, and Y are all free tovary. Freeing the directions of the relation-ships violates the temporal precedence speci-fied by the mediation model but allows possi-ble cross-lagged relations among variables tobe investigated, making it a more reasonablemodel than assuming relations are zero amongthe variables. Limitations of the autoregres-sive models include the cross-lagged modelwhere many true models may yield the samecross-lagged coefficients and the frequent ex-clusion of individual differences in mean level(see Dwyer 1983, Rogosa 1988).

Another model that can be used with lon-gitudinal mediation data is the latent-growthmodeling (LGM) or parallel-process model(Muthen & Curran 1997, Singer & Willet2003). The LGM mediation model examineswhether the growth in X affects the growthtrajectory of M, which affects the growth tra-jectory of Y. As in the nonlatent framework,the relation between X and the growth tra-jectory of Y has two sources: the indirect ef-fect via the growth of M and the direct effect.One limitation of the parallel-process modelis that the mediation relation is correlational:the slope in X is correlated with the slope inM, and the slope in M is correlated with theslope in Y. The interpretation of this correla-tion is that change in M is related to changein Y at the same time, not that change inM is related to change in Y at a later time.An alternate way to specify the LGM media-tion model is the two-stage piecewise parallel-process model (Cheong et al. 2003). In thetwo-stage parallel-process model, the growthof the mediator and the outcome process ismodeled for earlier and later times separately,allowing the mediated effects to be investi-gated at different periods. Measurement in-variance is very important in LGM, because

changes in the measure over time will con-found the interpretation of change over time.

In the difference score approach to lon-gitudinal mediation, differences between themediator and dependent variables scores aretaken, as is the independent variable if it doesnot reflect assignment to treatment condition.These difference scores are then analyzed us-ing the same equations as those used for cross-sectional models. The latent difference score(LDS) model can also be applied to three ormore waves using a latent framework (Ferrer& McArdle 2003, McArdle 2001, McArdle &Nesselroade 2003). In the LDS model, fixedparameters and latent variables are used tospecify latent difference scores, such that themodel represents differences between wavesas dynamic change. The LDS model can beespecially useful in situations where it is ex-pected there will be different predictors atdifferent measurement occasions.

In addition to the autoregressive, LGM,and LDS models, other models can be usedto analyze longitudinal mediation data, in-cluding a combination of the autoregressiveand LGM models (Bollen & Curran 2004)and specification of model parameters in acontinuous time metric to address the prob-lem of different time intervals of measurement(Arminger 1986, Boker & Nesselroade 2002,Dwyer 1992).

Moderation and Mediation

The strength and form of mediated effectsmay depend on other variables. Variables thataffect the hypothesized relation among a setof variables in such ways are known as mod-erators and are often tested as interaction ef-fects (Aiken & West 1991, Baron & Kenny1986). A nonzero XM interaction in Equation2 discussed above is an example of a moder-ator effect that suggests that the b coefficientdiffers across levels of X. Different b coeffi-cients across levels of X may reflect mediationas a manipulation and may alter the relationof M to Y. For example, a smoking preven-tion program may remove a relation between


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


tobacco offers (M) and tobacco use (Y) be-cause persons exposed to the program learnedskills to refuse tobacco offers so that offersare significantly related to use in the controlgroup but not in the program group (Judd& Kenny 1981a). The presence of modera-tor effects indicates that the modeled functionchanges across different levels of the moder-ator variable, where moderators may be ei-ther a manipulated factor in an experimentalsetting or a naturally occurring variable suchas gender. The examination of these variablesand their impact on mediation models is use-ful in psychological research to address thequestion of how an experiment achieved its ef-fects. However, by also examining moderatoreffects, one is able to investigate whether theexperiment differentially affects subgroupsof individuals (Donaldson 2001, MacKinnon2001, MacKinnon & Dwyer 1993, Sandleret al. 1997). Three potential models in whichthis examination may take place are (a) moder-ated mediation, (b) mediated moderation, and(c) mediated baseline by treatment modera-tion models.

Moderated mediation. The moderated me-diation model is the simplest statistical modelwith moderator and mediation effects ( Juddet al. 2001). In this model, a variable medi-ates the effect of an independent variable ona dependent variable, and the mediated effectdepends on the level of a moderator. Thus, themediational mechanism differs for subgroupsof participants (e.g., across cohorts, ages, orsexes; James & Brett 1984).

The single-mediator version of this modelconsists of estimating the same mediationmodel for each subgroup and then compar-ing the mediated effect across subgroups. Astatistical test of the equivalence of the me-diated effect across groups was described inMacKinnon (2007), and tests of the equal-ity of a , b , and c ′ can provide informationon the invariance-of-action theory (how theprogram changes mediators) and conceptual

theory (how mediators are related to the out-come) across groups.

The moderated mediation model is morecomplex when the moderator variable is con-tinuous. Although the regression equationsrequired to estimate the continuous moder-ated mediation model are the same as for thecategorical case, the interpretation of resultsis complicated because of the large number ofvalues of a continuous moderator. In this case,researchers may choose to analyze simple me-diation effects (see Tein et al. 2004).

Mediated moderation. Mediated modera-tion (Baron & Kenny 1986, Morgan-Lopez& MacKinnon 2001) occurs when a mediatoris intermediate in the causal sequence froman interaction effect to a dependent variable.For example, the effect of a prevention pro-gram may be greater for high-risk subjects,and the interaction effect of program expo-sure and risk-taking may then affect a mediat-ing variable of social norms that then affectsdrug use. The purpose of mediated modera-tion is to determine the mediating variable(s)that explain the interaction effect. This modelconsists of estimating a series of regressionequations where the main effect of a covari-ate and the interaction of the covariate andprogram exposure are included in both mod-els. Morgan-Lopez & MacKinnon (2001) de-scribe an estimator of the mediated moderatoreffect that requires further development andevaluation.

Mediated baseline by treatment moder-ation. The mediated baseline by treatmentmoderation model is a special case of themediated moderation model. The substantiveinterpretation of the mediated effect in thismodel is that the mediated effect depends onthe baseline level of the mediator. This sce-nario is a common result in prevention andtreatment research, where the effects of anintervention are often stronger for partici-pants who are at higher risk on the mediatingvariable at the time they enter the program


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


(Khoo 2001, Pillow et al. 1991). These treat-ment condition by baseline interactions havebeen found in numerous areas of research,ranging from universal prevention programswith elementary school children to selectiveprevention interventions with the various at-risk groups (e.g., Ialongo et al. 1999, Martinez& Forgatch 2001, Stoolmiller et al. 2000). In-formation provided in these models may in-dicate for whom an intervention is ineffectiveor even counterproductive and may be usedto screen future participants into more effec-tive programs based on their baseline char-acteristics. Various authors have outlined theequations and rationale for the mediated base-line by treatment moderator model (Baron &Kenny 1986, Morgan-Lopez & MacKinnon2001, Tein et al. 2004).

To date, models with moderators and me-diators have remained largely independent.This separation in their presentation has con-tributed to confusion in the understanding ofeach relative to the others. A critical goal offuture research in this area will be to developand test a general model in which each of themodels is a special case. One such model is de-scribed in Muller et al. (2005). Another modelis in development but has not yet been empir-ically tested in applied research (MacKinnon2007):

Y = i + c ′1X + c ′

2Z + c ′3XZ + b1M

+ b2MZ + hXM + jXMZ + e . 6.

In this model, the XM and XMZ interac-tions are added to the individual mediationand moderation equations to form a generalmodel that includes all effects (including ad-ditional c ′ and b effects). Here the h coeffi-cient represents the test of whether the M toY relation differs across levels of X, and the jcoefficient represents the three-way interac-tion effect whereby the relations between Zand M and Y differ across levels of X. If astatistically significant j coefficient is found,further simple interaction effects and simplemediated effects are explored.

Causal Inference

Methods based on the observed regressionapproach to estimating mediation have beencriticized based on causal analysis of the rela-tions among variables. One of these criticismsaddressed above is the equivalent model crit-icism. For example, if X, M, and Y are mea-sured simultaneously, there are other modelsthat would explain the data equally well (e.g.,X is the mediator of the M to Y relationship orM and Y both cause X), and in many situationsit is not possible to distinguish these alterna-tives without more information (Spirtes et al.1993).

The case in which X represents randomassignment to conditions improves causal in-terpretation of mediating variables (Holland1988, Robins & Greenland 1992) becauseX precedes M and Y. Holland appliedRubin’s (1974) causal model to a mediationand showed, under some assumptions, thetypical regression coefficient for the groupeffect on test score, c , and the group effecton number of hours studied, a , are valid es-timators of the true causal effect because ofthe randomization of units to treatment. Theregression coefficient, b , relating X to Y ad-justed for M, is not an accurate estimator ofthe causal effect because this relation is corre-lational, not the result of random assignment.The estimator, c ′, is also not an accurate causalestimator of the direct effect.

Several new approaches to causal infer-ence for mediation have begun to appear. Onepromising alternative is based on principalstratifications of the possible relations of X toM to Y where the mediated effect is estimatedwithin these stratifications (Angrist et al. 1996,Frangakis & Rubin 2002). B. Jo (unpublishedmanscript) has proposed a latent class versionof this model, and M.E. Sobel (unpublishedmanuscript) has proposed an enhancement ofthe Holland instrumental variable method.

The most important aspect of the causalinference methods is the illustration of theproblems interpreting the M to Y relationas a causal relation. Researchers have several


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


options in this situation. First, apply someof the new models to increase evidence forcausal inference. Second, treat the results ofthe mediation analysis as descriptive informa-tion that may not reflect the true underlyingcausal mediation relation, especially for the Mto Y relation, even when advanced causal in-ference models are applied. Third, future ex-perimental studies (perhaps double random-ization, described above) as well as qualitativeand clinical information are required to vali-date a mediation relation. In particular, a pro-gram of research that sequentially tests pre-dictors of the mediator theory provides themost convincing evidence for mediation.

SUMMARY AND FUTUREDIRECTIONS

There is broad and sustained interest in medi-ation analysis from many areas of psychologyand other fields: Begg & Leung (2000), Botvin(2000), Kristal et al. (2000), Petrosino (2000).Tests for mediation differ considerably in typeI error rates and statistical power (MacKinnonet al. 2002a, 2004). The recommended test ofmediation assesses the statistical significanceof the X to M relation, a path, and then the Mto Y relation, b path. If both are statisticallysignificant, there is evidence of mediation. Be-cause confidence limits are important for un-derstanding effects, confidence limits basedon the distribution of the product or the boot-strap are recommended. This approach alsoapplies to mediated effects in more compli-cated models. It is also important to consider

opposing mediated effects and more compli-cated models such that overall relations maynot be statistically significant yet mediationmay still exist in a research study. These op-posing effects or mediated effects that coun-teract each other resulting in a nonsignificantX to Y relation may be of substantive inter-est. Several effect size measures for mediationmodels have been proposed (A.J. Fairchild,D.P. MacKinnon, & M.P. Taborga, unpub-lished manuscript; Taborga et al. 1999), butthese require more development.

Person-oriented approaches based on tra-jectory classes (Muthen & Muthen 2000) andstaged responses across trials (Collins et al.1998) represent new ways to understand me-diational processes consistent with the goalof examining individual-level processes andgroup-level processes. Longitudinal data pro-vide rich information for the investigation ofmediation. In particular, latent growth curveand latent difference score models may be es-pecially suited to the examination of media-tion chains across multiple waves of data be-cause of the ability to investigate the effect ofprior change on later change. The usefulnessof causal inference models and different alter-natives to learning more about mediation arean important topic for future research. Ad-ditionally, experimental designs to investigatemediation require further development. Sim-ilarly, methods to combine qualitative as wellas quantitative information about mediationalprocesses should clarify mediation relations.These developments will advance our abilityto answer mediation questions in psychology.

ACKNOWLEDGMENTS

The authors acknowledge David Kenny and Helena Kraemer for their comments on thismanuscript and thank Hendricks Brown, Bengt Muthen, and other members of the PreventionScience Methodology Group for comments on presentations related to this review. This articlewas supported by the National Institute of Drug Abuse grant DA09757.

LITERATURE CITED

Aiken LS, West SG. 1991. Multiple Regression: Testing and Interpreting Interactions. NewburyPark, CA: Sage


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Alwin DF, Hauser RM. 1975. The decomposition of effects in path analysis. Am. Sociol. Rev.40:37–47

Angrist JD, Imbens GW, Rubin DB. 1996. Identification of causal effects using instrumentalvariables (with commentary). J. Am. Stat. Assoc. 91:444–72

Arbuckle JL. 1997. AMOS User’s Guide: Version 3.6. Chicago: SmallwatersArminger G. 1986. Linear stochastic differential equation models for panel data with unob-

served variables. Sociol. Methodol. 16:187–212Baranowski T, Anderson C, Carmack C. 1998. Mediating variable framework in physical ac-

tivity interventions: How are we doing? How might we do better? Am. J. Prev. Med.15(4):266–97

Baron RM, Kenny DA. 1986. The moderator-mediator variable distinction in social psycholog-ical research: conceptual, strategic, and statistical considerations. J. Personal. Soc. Psychol.51:1173–82

Bauer DJ, Preacher KJ, Gil KM. 2006. Conceptualizing and testing random indirect effectsand moderated mediation in multilevel models: new procedures and recommendations.Psychol. Methods. 11:142–63

Begg CB, Leung DHY. 2000. On the use of surrogate end points in randomized trials. J. Roy.Statist. Soc. 163:15–28

Bentler PM. 1997. EQS for Windows (Version 5.6) [computer program]. Encino, CA: Multivar.Softw.

Bishop YMM, Fienberg SE, Holland PW. 1975. Discrete Multivariate Analysis: Theory andPractice. Cambridge, MA: MIT Press

Blalock HM. 1969. Theory Construction: From Verbal to Mathematical Formulations. EnglewoodCliffs, NJ: Prentice-Hall

Boker SM, Nesselroade JR. 2002. A method for modeling the intrinsic dynamics of intraindi-vidual variability: recovering the parameters of simulated oscillators in multi-wave paneldata. Multivar. Behav. Res. 37:127–60

Bollen KA, Curran PJ. 2004. Autoregressive latent trajectory (ALT) models: a synthesis of twotraditions. Sociol. Methods Res. 32:336–83

Bollen KA, Stine RA. 1990. Direct and indirect effects: classical and bootstrap estimates ofvariability. Sociol. Methodol. 20:115–40

Botvin GJ. 2000. Preventing drug abuse in schools: social and competence enhancement ap-proaches targeting individual-level etiologic factors. Addict. Behav. 25:887–97

Cheong J, MacKinnon DP, Khoo ST. 2003. Investigation of mediational processes usingparallel process latent growth curve modeling. Struct. Equat. Model. 10:238–62

Cole DA, Maxwell SE. 2003. Testing mediational models with longitudinal data: questions andtips in the use of structural equation modeling. J. Abnorm. Psychol. 112:558–77

Collins LM, Graham JW, Flaherty BP. 1998. An alternative framework for defining mediation.Multivar. Behav. Res. 33:295–312

Davis JA. 1985. The Logic of Causal Order. Sage University Paper Series on Quantitative Applicationsin the Social Sciences, Series No. 07–055. Beverly Hills, CA: Sage

Dodge KA, Bates JE, Pettit GS. 1990. Mechanisms in the cycle of violence. Science 250:1678–83Donaldson SI. 2001. Mediator and moderator analysis in program development. In Handbook of

Program Development for Health Behavior Research and Practice, ed. S Sussman, pp. 470–96.Thousand Oaks, CA: Sage

Dwyer JH. 1983. Statistical Models for the Social and Behavioral Sciences. New York: OxfordDwyer JH. 1992. Differential equation models for longitudinal data. In Statistical Models for

Longitudinal Studies of Health, ed. JH Dwyer, M Feinleib, P Lippert, H Hoffmeister,pp. 71–98. New York: Oxford Univ. Press


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Fairchild AJ, MacKinnon DP, Taborga MP. 2006. R2 Effect-Size Measures for the Mediated Effect.Unpubl. manuscr.

Ferrer E, McArdle JJ. 2003. Alternative structural models for multivariate longitudinal dataanalysis. Struct. Equat. Model. 10:493–524

Fishbein M, Ajzen I. 1975. Belief, Attitude, Intention, and Behavior: An Introduction to Theory andResearch. Reading, MA: Addison-Wesley

Fiske ST, Kenny DA, Taylor SE. 1982. Structural models for the mediation of salience effectson attribution. J. Exp. Soc. Psychol. 18:105–27

Fleming TR, DeMets DL. 1996. Surrogate endpoints in clinical trials: Are we being misled?Ann. Intern. Med. 125:605–13

Frangakis CE, Rubin DB. 2002. Principal stratification in causal inference. Biometrics 58:21–29Freedman LS. 2001. Confidence intervals and statistical power of the “validation” ratio for

surrogate or intermediate endpoints. J. Statist. Plan. Inference 96:143–53Fritz MS, MacKinnon DP. 2007. Required sample size to detect the mediated effect. Psychol.

Sci. In pressGollob HF, Reichardt CS. 1991. Interpreting and estimating indirect effects assuming time

lags really matter. In Best Methods for the Analysis of Change: Recent Advances, UnansweredQuestions, Future Directions, ed. LM Collins, JL Horn, pp. 243–59. Washington, DC: Am.Psychol. Assoc.

Hebb GO. 1966. A Textbook of Psychology. Philadelphia, PA: Saunders. 2nd ed.Hofmann DA, Gavin MB. 1998. Centering decisions in hierarchical linear models. Implications

for research in organizations. J. Manage. 24:623–41Holland PW. 1988. Causal inference, path analysis, and recursive structural equation models.

Sociol. Methodol. 18:449–84Holmbeck GN. 1997. Toward terminological, conceptual, and statistical clarity in the study

of mediators and moderators: examples from the child-clinical and pediatric psychologyliteratures. J. Consult. Clin. Psychol. 65:599–610

Hoyle RH, Kenny DA. 1999. Statistical power and tests of mediation. In Statistical Strategiesfor Small Sample Research, ed. RH Hoyle, pp. 195–222. Newbury Park: Sage

Hyman HH. 1955. Survey Design and Analysis: Principles, Cases, and Procedures. Glencoe, IL:Free Press

Ialongo NS, Werthamer L, Kellam SG, Brown CH, Wang S, Lin Y. 1999. Proximal impactof two first-grade preventive interventions on the early risk behaviors for later substanceabuse, depression, and antisocial behavior. Am. J. Community Psychol. 27:599–641

James LR, Brett JM. 1984. Mediators, moderators, and tests for mediation. J. Appl. Psychol.69:307–21

James LR, Mulaik SA, Brett JM. 2006. A tale of two methods. Org. Res. Methodol. 9:233–44Jo B. 2006. Causal Inference in Randomized Trials with Mediational Processes. Unpubl. manuscr.Joreskog KG, Sorbom D. 1993. LISREL (Version 8.12) [computer program]. Chicago: Sci.

Software Int.Judd CM, Kenny DA. 1981a. Estimating the Effects of Social Interventions. Cambridge, UK:

Cambridge Univ. PressJudd CM, Kenny DA. 1981b. Process analysis: estimating mediation in treatment evaluations.

Eval. Rev. 5:602–19Judd CM, Kenny DA, McClelland GH. 2001. Estimating and testing mediation and moderation

in within-subject designs. Psychol. Methods 6:115–34Kenny DA, Bolger N, Korchmaros JD. 2003. Lower-level mediation in multilevel models.

Psychol. Methods 8:115–28


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Kenny DA, Kashy DA, Bolger N. 1998. Data analysis in social psychology. In The Handbook ofSocial Psychology, Volume 1, ed. DT Gilbert, ST Fiske, G Lindzey, pp. 233–65. New York:Oxford Univ. Press

Khoo ST. 2001. Assessing program effects in the presence of treatment-baseline interactions:a latent curve approach. Psychol. Methods 6:234–57

Kraemer HC, Stice E, Kazdin A, Offord D, Kupfer D. 2001. How do risk factors work to-gether? Mediators, moderators, and independent, overlapping, and proxy risk factors. Am.J. Psychiatry 158:848–56

Kraemer HC, Wilson T, Fairburn CG, Agras WS. 2002. Mediators and moderators of treat-ment effects in randomized clinical trials. Arch. Gen. Psychiatry 59:877–83

Kristal AR, Glanz K, Tilley BC, Li S. 2000. Mediating factors in dietary change: understandingthe impact of a worksite nutrition intervention. Health Educ. Behav. 27(1):112–25

Krull JL, MacKinnon DP. 1999. Multilevel mediation modeling in group-based interventionstudies. Eval. Rev. 23:418–44

Krull JL, MacKinnon DP. 2001. Multilevel modeling of individual and group level mediatedeffects. Multivar. Behav. Res. 36:249–77

Lazarsfeld PF. 1955. Interpretation of statistical relations as a research operation. In The Lan-guage of Social Research: A Reader in the Methodology of Social Research, ed. PF Lazarsfeld, MRosenberg, pp. 115–25. Glencoe, IL: Free Press

Lockwood CM, MacKinnon DP. 1998. Bootstrapping the standard error of the mediatedeffect. In Proceedings of the Twenty-Third Annual SAS Users Group International Conference,pp. 997–1002. Cary, NC: SAS Inst.

MacCorquodale K, Meehl PE. 1948. Operational validity of intervening constructs. Psychol.Rev. 55:95–107

MacKinnon DP. 1994. Analysis of mediating variables in prevention intervention studies.In Scientific Methods for Prevention Intervention Research: NIDA Research Monograph 139,DHHS Pub. 94–3631, ed. A Cazares, LA Beatty, pp.127–53. Washington, DC: U.S. Dept.Health Human Serv.

MacKinnon DP. 2000. Contrasts in multiple mediator models. In Multivariate Applications inSubstance Use Research: New Methods for New Questions ed. JS Rose, L Chassin, CC Presson,SJ Sherman, pp. 141–60. Mahwah, NJ: Erlbaum

MacKinnon DP. 2001. Mediating variable. In International Encyclopedia of the Social and Behav-ioral Sciences, ed. NJ Smelser, PB Baltes, pp. 9503–7. Oxford, UK: Pergamon

MacKinnon DP. 2007. Introduction to Statistical Mediation Analysis. Mahwah, NJ: Erlbaum. Inpress

MacKinnon DP, Dwyer JH. 1993. Estimation of mediated effects in prevention studies. Eval.Rev. 17:144–58

MacKinnon DP, Fritz MS, Williams J, Lockwood CM. 2006a. Distribution of the productconfidence limits for the indirect effect: program PRODCLIN. Behav. Res. Methods. Inpress. Download available at http://www.public.asu.edu/∼davidpm/ripl/Prodclin/

MacKinnon DP, Krull JL, Lockwood CM. 2000. Equivalence of the mediation, confounding,and suppression effect. Prev. Sci. 1:173–81

MacKinnon DP, Lockwood CM, Hoffman JM, West SG, Sheets V. 2002a. A comparison ofmethods to test mediation and other intervening variable effects. Psychol. Methods 7:83–104

MacKinnon DP, Lockwood CM, Williams J. 2004. Confidence limits for the indirect effect:distribution of the product and resampling methods. Multivar. Behav. Res. 39:99–128

MacKinnon DP, Taborga MP, Morgan-Lopez AA. 2002b. Mediation designs for tobaccoprevention research. Drug Alcohol Depend. 68: S69–83


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


MacKinnon DP, Warsi G, Dwyer JH. 1995. A simulation study of mediated effect measures.Multivariate Behav. Res. 30:41–62

MacKinnon DP, Yoon M, Lockwood CM, Taylor AB. 2006b. A Comparison of Methods to Testthe Mediated and Other Intervening Variable Effects in Logistic Regression. Unpubl. manuscr.

Manly BFJ. 1997. Randomization and Monte Carlo Methods in Biology. New York: Chapman &Hall. 2nd ed.

Martinez CR, Forgatch MS. 2001. Preventing problems with boys’ noncompliance: effects ofa parent training intervention for divorcing mothers. J. Consult.Clin. Psychol. 69(3):416–28

McArdle JJ. 2001. A latent difference score approach to longitudinal dynamic structural analy-sis. In Structural Equation Modeling: Present and Future. A Festschrift in Honor of Karl Joreskog,ed. R Cudeck, S du Toit, D. Sorbom. pp. 341–80. Lincolnwood, IL: Sci. Softw. Int.

McArdle JJ, Nesselroade JR. 2003. Growth curve analysis in contemporary research. In Com-prehensive Handbook of Psychology, Vol. II: Research Methods in Psychology, ed J Schinka, WVelicer, pp. 447–80. New York: Pergamon

McDonald RP. 1997. Haldane’s lungs: a case study in path analysis. Multivar. Behav. Res. 32:1–38

McFatter RM. 1979. The use of structural equation models in interpreting regression equationsincluding suppressor and enhancer variables. Appl. Psychol. Meas. 3:123–35

Meeker WQ, Cornwell LW, Aroian LA. 1981. The product of two normally distributed randomvariables. In Selected Tables in Mathematical Statistics, ed. WJ Kennedy, RE Odeh, JMDavenport, Vol. 7, pp. 1–256. Providence, RI: Am. Math. Soc.

Merrill R. 1994. Treatment effect evaluation in nonadditive mediation models. PhD thesis. Tempe:Ariz. State Univ.

Mooney CZ, Duval RD. 1993. Bootstrapping: A Nonparametric Approach to Statistical Inference.Newbury Park, CA: Sage

Morgan-Lopez AA, MacKinnon DP. 2001. A mediated moderation model simulation: mediationalprocesses that vary as a function of second predictors. Presented at 9th Annu. Meet. Soc. Prev.Res., Washington, DC

Morgan-Lopez AA, MacKinnon DP. 2006. Demonstration and evaluation of a method toassess mediated moderation. Behav. Res. Methods 38:77–87

Muller D, Judd CM, Yzerbyt VY. 2005. When moderation is mediated and mediation ismoderated. J. Personal. Soc Psychol. 89(6):852–63

Murray DM, Luepker RV, Pirie PL, Grimm RH, Bloom E, et al. 1986. Systematic risk factorscreening and education: a community-wide approach to prevention of coronary heartdisease. Prev. Med. 15:661–72

Muthen BO, Curran PJ. 1997. General longitudinal modeling of individual differences in ex-perimental designs: a latent variable framework for analysis and power estimation. Psychol.Methods 2:371–402

Muthen LK, Muthen BO. 1998–2006. Mplus: The Comprehensive Modeling Program for AppliedResearchers. User’s Guide. Los Angeles, CA: Muthen & Muthen

Muthen B, Muthen L. 2000. Integrating person-centered and variable-centered analysis:growth mixture modeling with latent trajectory classes. Alcohol. Clin. Exp. Res. 24:882–91

Noreen EW. 1989. Computer-Intensive Methods for Testing Hypotheses: An Introduction. New York:Wiley

Paulhus DL, Robins RW, Trzesniewski KH, Tracy JL. 2004. Two replicable suppressor situa-tions in personality research. Multivar. Behav. Res. 39:303–28

Petrosino A. 2000. Mediators and moderators in the evaluation of programs for children. Eval.Rev. 24(1):47–72


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Pillow DR, Sandler IN, Braver SL, Wolchik SA, Gersten JC. 1991. Theory-based screening forprevention: focusing on mediating processes in children of divorce. Am. J. Comm. Psychol.19:809–36

Preacher KJ, Hayes AF. 2004. SPSS and SAS procedures for estimating indirect effects insimple mediation models. Behav. Res. Methods Instrum. Comput. 36:717–31

Raudenbush SW, Sampson R. 1999. Assessing direct and indirect effects in multilevel designswith latent variables. Sociol. Methods Res. 28:123–53

Robins JM, Greenland S. 1992. Identifiability and exchangeability for direct and indirect effects.Epidemiology 3:143–55

Rogosa DR. 1988. Myths about longitudinal research. In Methodological Issues in Aging Research,ed. KW Schaie, RT Campbell, WM Meredith, SC Rawlings, pp. 171–209. New York:Springer

Rubin DB. 1974. Estimating causal effects of treatments in randomized and nonrandomizedstudies. J. Educ. Psychol. 66:688–701

Sandler IN, Wolchik SA, MacKinnon DP, Ayers TS, Roosa MW. 1997. Developing linkagesbetween theory and intervention in stress and coping processes. In Handbook of Children’sCoping: Linking Theory and Intervention, ed. SA Wolchik, IN Sandler, pp. 3–40. New York:Plenum

Shadish WR. 1996. Meta-analysis and the exploration of causal mediating processes: a primerof examples, methods, and issues. Psychol. Methods 1:47–65

Sheets VL, Braver SL. 1999. Organizational status and perceived sexual harassment: detectingthe mediators of a null effect. Personal. Soc. Psychol. Bull. 25:1159–71

Sherman SJ, Gorkin L. 1980. Attitude bolstering when behavior is inconsistent with centralattitudes. J. Exp. Soc. Psychol. 16:388–403

Shrout PE, Bolger N. 2002. Mediation in experimental and nonexperimental studies: newprocedures and recommendations. Psychol. Methods 7:422–45

Singer JD, Willett JB. 2003. Applied Longitudinal Data Analysis: Modeling Change and EventOccurrence. London: Oxford Univ. Press

Sobel ME. 1982. Asymptotic confidence intervals for indirect effects in structural equationmodels. Sociol. Methodol. 13:290–312

Sobel ME. 1986. Some new results on indirect effects and their standard errors in covariancestructure models. Sociol. Methodol. 16:159–86

Sobel ME. 2006. Identification of Causal Parameters in Randomized Studies with Mediators. Unpubl.manuscr.

Spencer SJ, Zanna MP, Fong GT. 2005. Establishing a causal chain: why experiments are oftenmore effective than mediational analyses in examining psychological processes. J. Personal.Soc. Psychol. 89(6):845–51

Spirtes P, Glymour C, Scheines R. 1993. Causation, Prediction, and Search. New York: Springer-Verlag

Springer MD. 1979. The Algebra of Random Variables. New York: WileyStone CA, Sobel ME. 1990. The robustness of estimates of total indirect effects in covariance

structure models estimated by maximum likelihood. Psychometrika 55:337–52Stoolmiller M, Eddy JM, Reid JB. 2000. Detecting and describing preventive intervention

effects in a universal school-based randomized trial targeting delinquent and violent be-havior. J. Consult. Clin. Psychol. 68:296–306

Taborga MP, MacKinnon DP, Krull JL. 1999. A simulation study of effect size measures in mediationmodels. Poster presented at 7th Annu. Meet. Soc. Prev. Res., New Orleans, LA


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Tein JY, MacKinnon DP. 2003. Estimating mediated effects with survival data. In New De-velopments in Psychometrics: Psychometric Society Proceedings, ed. H Yanai, AO Rikkyo, KShigemasu, Y Kano, JJ Meulman, pp. 405–12. Tokyo: Springer-Verlag

Tein JY, Sandler IN, MacKinnon DP, Wolchik SA. 2004. How did it work? Who did it workfor? Mediation in the context of a moderated prevention effect for children of divorce. J.Consult. Clin. Psychol. 72:617–24

Weiss CH. 1997. How can theory-based evaluation make greater headway? Eval. Rev. 21:501–24

West SG, Aiken LS. 1997. Toward understanding individual effects in multiple componentprevention programs: Design and analysis strategies. In The Science of Prevention: Method-ological Advances from Alcohol and Substance Abuse Research, ed. K Bryant, M Windle, S West,pp. 167–209. Washington, DC: Am. Psychol. Assoc.

Williams J. 2004. Resampling and distribution of products methods for testing indirect effects in complexmodels. PhD thesis. Tempe: Ariz. State Univ.

Winship C, Mare RD. 1983. Structural equations and path analysis for discrete data. Am. J.Sociol. 89:54–110

Word CO, Zanna MP, Cooper J. 1974. The nonverbal mediation of self-fulfilling propheciesin interracial interaction. J. Exp. Soc. Psychol. 10:109–20


Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.

Contents ARI 8 November 2006 21:2

Annual Review ofPsychology

Volume 58, 2007

Contents

Prefatory

Research on Attention Networks as a Model for the Integration ofPsychological ScienceMichael I. Posner and Mary K. Rothbart � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 1

Cognitive Neuroscience

The Representation of Object Concepts in the BrainAlex Martin � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 25

Depth, Space, and Motion

Perception of Human MotionRandolph Blake and Maggie Shiffrar � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 47

Form Perception (Scene Perception) or Object Recognition

Visual Object Recognition: Do We Know More Now Than We Did 20Years Ago?Jessie J. Peissig and Michael J. Tarr � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 75

Animal Cognition

Causal Cognition in Human and Nonhuman Animals: A Comparative,Critical ReviewDerek C. Penn and Daniel J. Povinelli � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 97

Emotional, Social, and Personality Development

The Development of CopingEllen A. Skinner and Melanie J. Zimmer-Gembeck � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 119

vii

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Biological and Genetic Processes in Development

The Neurobiology of Stress and DevelopmentMegan Gunnar and Karina Quevedo � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 145

Development in Societal Context

An Interactionist Perspective on the Socioeconomic Context ofHuman DevelopmentRand D. Conger and M. Brent Donnellan � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 175

Culture and Mental Health

Race, Race-Based Discrimination, and Health Outcomes AmongAfrican AmericansVickie M. Mays, Susan D. Cochran, and Namdi W. Barnes � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 201

Personality Disorders

Assessment and Diagnosis of Personality Disorder: Perennial Issuesand an Emerging ReconceptualizationLee Anna Clark � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 227

Social Psychology of Attention, Control, and Automaticity

Social Cognitive Neuroscience: A Review of Core ProcessesMatthew D. Lieberman � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 259

Inference, Person Perception, Attribution

Partitioning the Domain of Social Inference: Dual Mode and SystemsModels and Their AlternativesArie W. Kruglanski and Edward Orehek � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 291

Self and Identity

Motivational and Emotional Aspects of the SelfMark R. Leary � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 317

Social Development, Social Personality, Social Motivation,Social Emotion

Moral Emotions and Moral BehaviorJune Price Tangney, Jeff Stuewig, and Debra J. Mashek � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 345

viii Contents

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


The Experience of EmotionLisa Feldman Barrett, Batja Mesquita, Kevin N. Ochsner,

and James J. Gross � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 373

Attraction and Close Relationships

The Close Relationships of Lesbian and Gay MenLetitia Anne Peplau and Adam W. Fingerhut � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 405

Small Groups

OstracismKipling D. Williams � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 425

Personality Processes

The Elaboration of Personal Construct PsychologyBeverly M. Walker and David A. Winter � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 453

Cross-Country or Regional Comparisons

Cross-Cultural Organizational BehaviorMichele J. Gelfand, Miriam Erez, and Zeynep Aycan � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 479

Organizational Groups and Teams

Work Group DiversityDaan van Knippenberg and Michaéla C. Schippers � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 515

Career Development and Counseling

Work and Vocational Psychology: Theory, Research,and ApplicationsNadya A. Fouad � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 543

Adjustment to Chronic Diseases and Terminal Illness

Health Psychology: Psychological Adjustmentto Chronic DiseaseAnnette L. Stanton, Tracey A. Revenson, and Howard Tennen � � � � � � � � � � � � � � � � � � � � � � � � � � � 565

Contents ix

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.


Research Methodology

Mediation AnalysisDavid P. MacKinnon, Amanda J. Fairchild, and Matthew S. Fritz � � � � � � � � � � � � � � � � � � � � � 593

Analysis of Nonlinear Patterns of Change with Random CoefficientModelsRobert Cudeck and Jeffrey R. Harring � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 615

Indexes

Cumulative Index of Contributing Authors, Volumes 48–58 � � � � � � � � � � � � � � � � � � � � � � � � � � � 639

Cumulative Index of Chapter Titles, Volumes 48–58 � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 644

Errata

An online log of corrections to Annual Review of Psychology chapters (if any, 1997 to thepresent) may be found at http://psych.annualreviews.org/errata.shtml

x Contents

Ann

u. R

ev. P

sych

ol. 2

007.

58:5

93-6

14. D

ownl

oade

d fr

om a

rjou

rnal

s.an

nual

revi

ews.

org

by V

IVA

on

02/1

8/09

. For

per

sona

l use

onl

y.

mediation analysis - university of british columbiaschaller/528readings/mackinnon... · ·...

Documents