Mediation Analyses: Applications in Nutrition Research and Reading the Literature

This article appeared in a journal published by Elsevier. The attachedcopy is furnished to the author for internal non-commercial researchand education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling orlicensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of thearticle (e.g. in Word or Tex form) to their personal website orinstitutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies areencouraged to visit:

http://www.elsevier.com/copyright

http://www.elsevier.com/copyright

Author's personal copy

RESEARCH

Research and Practice Innovations

Mediation Analyses: Applications in NutritionResearch and Reading the LiteratureCHONDRA M. LOCKWOOD, PhD, MA; CAROL A. DEFRANCESCO, MA, RD; DIANE L. ELLIOT, MD;SHIRLEY A. A. BERESFORD, PhD, MSc, MA; DEBORAH J. TOOBERT, PhD, MA

ABSTRACTMediation analysis is a newer statistical tool that is be-coming more prominent in nutrition research. Its useprovides insight into the relationship among variables ina potential causal chain. For intervention studies, it candefine the influence of different programmatic compo-nents and, in doing so, allows investigators to identifyand refine a program’s critical aspects. We present anoverview of mediation analysis, compare mediators withother variables (confounders, moderators, and covari-ates), and illustrate how mediation analysis permits in-terpretation of the change process. A framework is out-lined for the critical appraisal of articles purporting touse mediation analysis. The framework’s utility is dem-onstrated by searching the nutrition literature and iden-tifying articles citing mediation cross referenced with theterms “nutrition,” “diet,” “food,” and “obesity.” Seventy-two articles were identified that involved human subjectsand behavior outcomes, and almost half mentioned me-diation without tests to define its presence. Tabulation ofthe 40 articles appropriately assessing mediation demon-strates an increase in these techniques’ appearance andthe breadth of nutrition topics addressed. Mediationanalysis is an important new statistical tool. Familiaritywith its methodology and a framework for assessing ar-ticles will allow readers to critically appraise the litera-ture and make informed independent evaluations ofworks using these techniques.J Am Diet Assoc. 2010;110:753-762.

The word mediate comes from the Latin mediare,which means to place in the middle, and mediationrefers to a facilitator resolving a dispute between two

parties. During the past 10 years, mediation has increas-ingly been used in another context related to study de-sign, data analysis, and interpretation of findings (1-3).The term is well used in that setting as mediation anal-ysis assesses events between two variables or between anintervention and its outcomes.

Researchers from many fields, including dietetics, havestressed moving beyond evaluating only study efficacy todeconstruct how programs achieve their results (4-8).Mediation analysis can determine the influence of eachlink in a hypothetical chain of events and define thecontribution of different program components. It providesan explicit check on an intervention’s theoretical under-pinnings and whether the proposed change process wasachieved. Importantly, mediation analysis helps re-searchers modify, improve, and make interventions morecost-effective by identifying and refining their criticalcomponents.

Mediation analysis is a newer statistical technique.The Journal of the American Dietetic Association hasbeen a vehicle of information about statistical concepts(9) and the application of evidence-based assessments(10). In that tradition, this article describes mediationanalyses and contrasts mediators with others types ofvariables, such as a confounders, covariates, and moder-ators. Basic statistical methods to assess mediation arereviewed and an analytic framework proposed to guidereading articles purporting to use mediation analysis.The concepts are illustrated by nutrition-related exam-ples and the framework is applied to recent publicationsin the field. The objective is to enhance readers’ ability tocritically interpret the literature when investigators usethese newer statistical tools.

METHODSOvid MEDLINE and PubMed were searched using “me-diation” combined with “nutrition,” “diet,” “obesity,” and“food” for the dates January 2006 through May 2008,expanded by reviewing the indexes of prominent nutri-tion journals for relevant works. Titles and abstracts ofcitations, and articles involving human beings and usingphysical activity, nutrition, or other related behaviors asmeasures were accessed. For these works, at least twoauthors (C.M.L., C.A.D., or D.L.E.) read the article andapplied the proposed evaluation framework. When dis-agreements occurred, a third author read the work, andthe consensus scoring was used.

C. M. Lockwood is a statistical consultant in Portland,OR. C. A. DeFrancesco is a senior research associateand D. L. Elliot is a professor of medicine, Division ofHealth Promotion and Sports Medicine, Department ofMedicine, Oregon Health and Science University, Port-land. S. A. A. Beresford is a professor of epidemiology,Department of Epidemiology, University of Washington,Seattle. D. J. Toobert is a senior research scientist, OregonResearch Institute, Eugene.

Address correspondence to: Diane L. Elliot, MD, Divi-sion of Health Promotion and Sports Medicine, OregonHealth and Science University, 3181 SW Sam JacksonPark Rd, Portland, OR 97239-3098. E-mail: [email protected]

Manuscript accepted: September 1, 2009.Copyright © 2010 by the American Dietetic

Association.0002-8223/10/11005-0009$36.00/0doi: 10.1016/j.jada.2010.02.005

© 2010 by the American Dietetic Association Journal of the AMERICAN DIETETIC ASSOCIATION 753


MEDIATION AND RELATED MODELSMediation Uses Regression AnalysisIn a simple cross-sectional observational study, differentstatistical methods can be used to compare groups. Forexample, if one wanted to understand how tooth decayvaries by sweetened beverage intake, individuals could bepartitioned into those who do and do not drink sweetenedbeverages. The number of cavities for each group could becalculated and their means directly compared with anindependent t test to determine the probability or P valuethat the observed difference could have occurred bychance.

Regression analysis is a different approach to answerthe same question. A predictor or independent variable(X) is defined—which in this example would be drinkingsweetened beverages—and it is related to the dependentvariable (Y)—in this case, number of cavities. This pro-vides a regression coefficient, which is a measure of therelationship between beverage status and cavity number.This design is depicted in panel A of Figure 1, with theeffect of X on Y labeled c. Each approach is equally valid,and the conclusions and P values from the two analyseswould be the same (11).

For certain questions, regression analysis has advan-tages over a t test. With a t test, a single variable must beused to split samples into groups, and in general, only onevariable can be examined at a time. Regression analysisallows variables to be evaluated as continuous predictors(eg, number of sweetened beverages consumed each day),and more than one predictor variable can be added intothe analysis.

The use of regression analysis implies directionality, asX is defined as the predictor or independent variable andY is the outcome or dependent variable. Although exist-ing understanding and common sense might suggest thatdrinking sweetened beverages leads to more cavitiesrather than the reverse, one cannot determine causationwhen data are gathered from one time point. No statisti-cal test proves causality. Cause and effect only can beestablished from a prospective randomized experiment,which assesses the population before and following anintervention. Regression analysis can be used where dataare gathered over time and panel A of Figure 1 also couldrepresent such a trial, where X represents a randomizedintervention. Assuming the control and experimentalcondition were comparable at baseline, intervention sta-tus could be used to predict number of cavities followingthe intervention.

Mediation Adds Another Variable into the Regression SequenceRegression analysis provides the ability to introduceother variables, including those that may have mediatedthe path between X and the outcome Y. Mediating vari-ables also may be referred to as process variables, surro-gate endpoints, or proximal outcomes; each term relatesto intervening parameters that come between the predic-tor variable (or initiation of an intervention) and the finaloutcome of interest (1-3,12). A basic mediation model isshown in panel B of Figure 1. In this case, a secondregression path is assessed that includes a potentiallymediating variable, M. The independent predictor or in-tervention (X) is presumed to affect the mediator (M), andthat path (a) is the action theory. In turn, M affects theoutcome or dependent variable (Y), and the latter medi-ation path (b) is the conceptual theory (13). Althoughsecond in the sequence, when designing an intervention,the conceptual theory is the initial step, as investigatorsdecide what variables will relate to the outcomes of in-terest. Those become the purported mediating variablesand, once identified, researchers decide on actions to af-fect those parameters.

Returning to the prospective tooth decay example, ifthe intervention was removing sweetened carbonatedbeverage machines from schools, then the proposed me-diator might be number of sweetened carbonated bever-ages consumed each day. A reduction in that mediatingvariable would be predicted to reduce tooth decay. TheX¡Y path is recalculated after statistically removing thecomponent of that relationship accounted by the a and bpaths, yielding c=. If that mediating variable completelyaccounts for the change in Y, then that c= path goes tozero. In other words, if sweetened carbonated beveragesconsumed (X) explained all the variability in prevalenceof tooth decay (Y), then c= would go to zero. In most casesa single mediator will not account for all of the effect, andthe c= direct effect still will be present.

Other Variables: Confounders, Covariates, and ModeratorsMediation is only one of several relations that may bepresent when a third variable is included, and other typesare confounders, covariates, and moderators (Figure 2).The distinction among these is in their relationship to

Figure 1. Simple regression analysis and analysis with a third medi-ating variable added to the model.

754 May 2010 Volume 110 Number 5


other variables in the model. A variable that is a covari-ate in one study can be a moderator in another. Under-standing these relationships is important when interpret-ing findings.

A confounder is a variable that relates to both X and Y,but is not in the causal pathway (panel A of Figure 2).Confounders are alternative explanations for the ob-served relationship of X¡Y (11,14,15). An example mightbe represented by the epidemiologic observation that fre-quent urination was related to losing weight, and a re-gression analysis might indicate urination frequency pre-dicted weight loss. It might be tempting to interpret this

as a cause-and-effect connection. Instead, the apparentrelationship between X and Y is due to a confoundingvariable, hyperglycemia, which accounted for both obser-vations. Confounding variables are particularly impor-tant to consider when interpreting observational studiesbecause methods to control for confounders in prospectiverandomized trials, such as enrollment restrictions, ran-domization, and matching, are not always present inobservational studies (14).

A covariate is a variable that was not changed by theintervention, does not alter the X¡Y relationship, andimproves prediction of the outcome (panel B of Figure 2).Covariates often are a parameter measured in the popu-lation being studied, such as age, sex, or socioeconomicstatus; including covariates in the analysis will explainadditional variability in the dependent outcome variable.For example, investigators conducting a study of fooddiaries as a means to increase fruit and vegetable con-sumption find that those who kept food diaries ate morefruits and vegetables. When the influence of sex wasassessed, women were found to eat more fruits and veg-etables in both the intervention and control groups. Ad-justing for that covariate will reduce the variability infruit and vegetable consumption, allowing for more pre-cise estimation of the effect of food diaries on changingfruit and vegetable consumption.

Moderators are a third type of variable that also mustbe factored into conclusions. Unlike covariates, modera-tors, also called interaction effects, change the relation-ship of X¡Y (16,17) (panel C of Figure 2). For continuousmoderators, the relationship varies across the range ofthat variable. As with covariates, moderators frequentlyare features of the study group and not affected by theintervention. In a study examining the relationship be-tween calcium intake and bone density among premeno-pausal women, estrogen status might be a moderator.Calcium intake’s affect on bone density would be differentfor women experiencing hypoestrogenic amenorrhea thanfor women with normal estrogen levels. Moderators ex-plain differential effects and specify for whom and whena treatment will be effective. Once identified, they becomeimportant variables to consider when enrolling subjectsand randomizing them to study conditions.

MEDIATION IN STUDY DESIGN AND INTERPRETATIONMediation analysis can be critical when interpretingstudy findings. As a simplified example of its utility,suppose researchers were designing a weight loss study.They observed that individuals with obesity eat largerportion sizes (18) and also are less physically active (19).Those variables and relationships form the conceptual the-ory (M¡Y). Then, investigators must decide how to imple-ment an intervention to affect those mediators, which willdefine the action theory (X¡M). Accordingly, they may de-sign a program that teaches appropriate portion size andprovides an incentive plan for increased physical activity. Intheir methods, investigators must use means to sequen-tially measure participants’ portion size and physical activ-ity level, along with body weight. In addition, potentialconfounders, covariates, and moderator variables need to beconsidered as the subject group is recruited, enrolled, andrandomized to study conditions.

Figure 3 presents a matrix of potential study findings.

XIndependent or Predictor

Variable

YDependent or Outcome

Variable

Confounder


Variable


Variable


Variable


Variable

Covariate

Moderator

Figure 2. Third variable models and relationships to independent anddependent variables: A confounder (panel A), a covariate (panel B), anda mediator (panel C).

May 2010 ● Journal of the AMERICAN DIETETIC ASSOCIATION 755


For each row, the intervention group lost more weightthan the control condition. However, the mediation out-comes and the result’s implications differ. Concludingthat the intervention achieved its objectives and is readyfor replication and wider dissemination would hold onlyfor Study 1. It represents the ideal situation, with theintervention successfully affecting both targeted media-tors, each of which are related to the outcome. In thisexample, both the conceptual and action theory weresupported, and portion size and physical activity mediatethe intervention effect.

In the second situation, the intervention successfullyreduced portion size, and portion size was related to bodyweight. Although physical activity was related to bodyweight, the intervention failed to affect that variable, andonly the conceptual theory was supported. With thosefindings, researchers must critically examine their phys-ical activity intervention and decide whether tostrengthen or discontinue that component to focus re-sources on the successful portion size aspect.

Study 3 findings indicate that the intervention changedboth targeted mediators, but only physical activity wasrelated to body weight. Thus, the conceptual theory re-lating portion size to body weight was not supported. Thiscould mean that the conceptual theory is faulty. Perhapsit was based on cross-sectional observations, with bothportion size and weight relating to some other confound-ing variable? With these study findings, investigatorsmay wish to omit the intervention’s portion size compo-nent. Alternatively, if prior evidence for the portion sizeconceptual theory was robust, researchers may need toreassess their methodology for indexing that variable orincorporate another potential variable, such as time, into

the analysis. Possibly a longer interval was required toestablish the relationship between portion size and bodyweight? In either situation, the mediation analyses find-ings should lead to better understanding and inform sub-sequent study designs.

The final situation is where the intervention affectedboth potential mediators, supporting the action theory,but neither was related to the outcome. Following theintervention, participants were eating smaller portionsand were more active, but those variables did not predictthe change in weight. The implication is that other pa-rameters altered by the intervention, such as providercontact or social support, also were changed, and thoseother parameters mediated the outcome. More likely thanthis extreme case would be that portion size and physicalactivity accounted for only a small amount of the medi-ating effect, leaving a large direct effect. Rather thanincorrectly concluding that altering physical activity andportion size are ideal means to achieve weight loss, theinvestigators have the opportunity to identify other vari-ables not included in the original mediation model thatmay have been affected to account for the positive inter-vention effects.

Despite the differences in mediation analyses out-comes, the examples in Figure 3 all share a significantintervention effect. Although early guidelines for media-tion analyses required an overall program effect (20),more recent research has shown that is not necessary(3,21), and mediation analysis can be performed evenwhen an intervention does not achieve significantchanges in the study outcome. The examples in Figure 3also involve an intervention with only two manipulations,and most programs involve multiple components. When

Study

Action theory:Intervention (X)3Mediator (M)(Is intervention related to mediator?)

Conceptual theory:Mediator (M)3Outcome (Y)(Is mediator related to outcome?) Interpretation and action

1 Intervention changed portion size:yes

Portion size related to body weight:yes

Conceptual and action theory supportedfor both mediators. Disseminateeffective intervention.

Intervention changed physical activity:yes

Physical activity related to weight:yes


Portion size related to body weight:yes

Portion size is a mediator. Conceptualtheory supported for physical activitybut exercise intervention not effective.Redesign or omit physical activityintervention component.

Intervention changed physical activity:no



Portion size related to body weight:no

Physical activity is mediator. Interventionchanged portion size but becauseconceptual theory not supported, itwas not a mediator. Assess theoryand measurements and consideromitting that component.

Intervention changed physical activity:yes



Portion size related to body weight:no

Intervention changed targeted variables,but those changes were not relatedto the outcome. Program effectsoccurred through other mechanisms.

Figure 3. Exploration of potential program effects and interpretation through mediation analysis.



several mediators are present, they can be examinedseparately or simultaneously. For this example, in addi-tion to physical activity and portion size, a weight lossstudy might attempt to reduce television viewing (22) andfood energy density (23). Imagine Figure 3 expanded toinclude additional intervention components, where thecombination of potential mediating outcomes increasesgeometrically. Interpreting study findings and knowingwhich parts of the intervention to drop, alter, and en-hance become impossible without analyzing the media-tion of each of its components.

Similarly, these examples have a limited number oftime points, and longitudinal data with repeated mea-sures of variables provide additional rich information forthe investigation of mediation. More advanced tech-niques, such as structural equation and latent growthcurve modeling, allow examination of mediation chainsacross multiple waves of data. However, those more com-plicated models are extensions of the basic mediationanalysis (3).

Mediation analysis generates evidence for how a pro-gram achieved its effects and provides a check on whetherthe program produced a change in variables that it wasdesigned to change. Second, its results may suggest thatcertain program components need to be strengthened.Third, program effects on mediating variables in the ab-sence of effects on outcome measures suggest that thetargeted constructs were not critical in changing out-comes or that measurement of those parameters wasfaulty. Thus, identification of mediating variables pro-vides information on the change process and allowsstreamlining programs by focusing on their effective com-ponents.

APPLYING MEDIATION CONCEPTS WHEN READING THELITERATUREProponents of critical literature review have suggestedsteps that allow appropriate interpretation of researchpublications involving prognosis, diagnostic tests, ther-apy, and economic analyses (24). A four-step structurecan be applied to assess articles using mediation, ex-plained in greater detail below:

1. Is mediation properly assessed?2. Are theoretical underpinnings clearly stated and sup-

ported by prior research?3. Is it a single time point observation? If more than one,

are variables measured in the correct temporal se-quence?

4. Is it a prospective randomized controlled interventionstudy?

At a minimum, a study purporting to show mediation(or lack of mediation) needs to conduct a statistical test ofmediation (Step 1). Although this may seem obvious, theterm mediation often is used when no statistical media-tion was performed (25).

There are three major approaches to statistical media-tion analysis: causal steps, difference in coefficients, andproduct of coefficients (25). The most widely used tech-nique is the causal steps approach outlined in the classicwork of Baron and Kenny (20) and Judd and Kenny(26,27). Four steps are involved: establish an overall ef-

fect between X and Y (the c path), establish an affect of Xon M (the a path), establish an effect of M on Y aftercontrolling for X (the b path), and establish a reductionfrom the total effect (c) to the direct effect (c=), aftercontrolling for M. A simpler, but equally valid, causalsteps test is to require that both paths a and b be signif-icant (24). A modified version of the Baron and Kennyapproach that applies specifically to randomized trials iscalled the MacArthur model (12,28). Keywords to look forin identifying a study that uses this method are “causalsteps” or citations by Baron and Kenny (20), Judd andKenny (26,27), or MacKinnon (25).

The second and least used mediation method involvescalculating the primary difference in the total effect anddirect effect (c-c=). That value can then be divided by astandard error and tested for significance. Keywords forthis statistical technique would include references toFreedman and Schatzkin (29) or Clogg (30). The thirdapproach is the primary product of coefficients method,which calculates the mediated effect as a�b. This methodis commonly used by statistical software packages, andthe methods or results description frequently will includea reference to Sobel (31,32). Although these differenttests for mediation vary in their type I error rates andstatistical power, identifying that investigators used anyone of the three methods will establish that statisticalmediation was assessed.

Steps 2 through 4 strengthen the conclusion that me-diation occurred. A strong foundation in theory and priorresearch strengthens mediation claims (Step 2). Thismeans that the authors describe both the conceptual the-ory for the purported mediators and the action theory fortheir intervention’s ability to affect the mediating vari-ables by citing prior research that supports those rela-tionships.

Although mediation analysis can be conducted oncross-sectional observations, a stronger case is made withlongitudinal data, when variables that come earlier in thecausal model are measured before those that come later(Step 3). Adding a randomized design to longitudinal datacollection makes causal mediation claims stronger byeliminating many possible confounders (Step 4) (33).

Using the Framework: An ExampleThe four-step analytic framework can be applied to anearly work utilizing mediation published in the Journalof the American Dietetic Association by Fisher and col-leagues (34). The investigators assessed a cohort of moth-ers and their infants. The variables of interest were du-ration of breast-feeding (X), maternal control of feeding(M), and energy intake of the toddlers (Y). Maternal con-trol of feeding was a self-reported measure of child-feed-ing beliefs and behaviors.

Is Mediation Properly Assessed? Fisher and colleagues (34)displayed the results of their regression analyses usingthe causal steps method. They determined an overalleffect (the c path) of breastfeeding history on toddlerenergy intake. They also established the a path (breast-feeding history affects maternal control of feeding) andthe b path (maternal control affects toddler energy in-take). They defined the c= path by statistically removingthe paths through the mediating variable. They cited



Author(s)(reference no.) Subjects

Predictor/independentvariable

Dependent/outcomevariable Purported mediator(s)

Observational studies with data gathered at a single time pointAnnus and

colleagues (35)US collegiate

womenFamily experiences

(teasing, modeling)Disordered eating Expectancies for life

improvement from thinness

Bere and colleagues(36)

Norwegianadolescents

Sex Fruit and vegetable intake Accessibility, preferences, andself-efficacy

Beydoun andcolleagues (37)

Cross-section USadults

Nutrition knowledge andbeliefs

Diet quality Away from home foodexperiences

Beydoun and Wang(38)

Cross-section USadults

Income and education Diet quality Perceived barriers food pricesand benefits of quality diet

Beydoun andcolleagues (39)

National Healthand NutritionExamination data

Ethnicity/race Obesity and indexmetabolic syndrome

Dairy food group intake

Brown andcolleagues (40)

Adults with mentalillness

Mental performancecomposite score

Grocery shopping skills Grocery shopping knowledge

Brug and colleagues(41)

European schoolchildren

Sex Fruit and vegetable intake Accessibility, modeling, andpreferences

Caperchione andcolleagues (42)

Australian adults Body mass index Intentions for physicalactivity

Attitudes, behavior control, andsubjective norms

Cerin and colleagues(43)

Australian adults Education, income,household size

Physical activity Individual and environmentalvariables

Decaluwe andcolleagues (44)

Flemish obeseadolescents

Parent characteristicsand behaviors

Psychological problems Parenting style (inconsistentdiscipline)

Hanson and Chen(45)

US high schoolstudents

Income, parenteducation andoccupation

Body mass index Sedentary behaviors

Hesketh andcolleagues (46)

Australian families Maternal education Children’s televisionviewing

21 aspects family televisionenvironment

Jago and colleagues(47)

Houston boy scouts Distance to food storesand restaurants

Fruit and vegetable intake Food preferences andavailability in home

Janicke andcolleagues (48)

US overweightadolescents

Peer victimization Quality of life Depressive symptoms

Klepp and colleagues(49)

11-year-olds fromnine Europeancountries

Exposure foodcommercials ontelevision

Fruit and vegetable intake Attitudes and preferences aboutfruits and vegetables

Luyckx andcolleagues (50)

Dutch adults withtype 1 diabetes

Identification diabetesrelated problems

Depression Adaptive and maladaptivecoping

Proper andcolleagues (51)

Working Australianadults

Education and income Body mass index Sitting on weekdays, weekends,and in leisure time

Mai and colleagues(52)

8- to 10-year-oldCanadian youth

Milk consumption Asthma Being overweight

Mond and colleagues(53)

Australian women Psychological functioningand quality of life

Obesity Weight/shape concerns andbinge-eating

Sacco andcolleagues (54)

US adults with type2 diabetes

Behavioral adherence,self-care, body massindex

Depressions Symptoms diabetes and self-efficacy

(continued)

Figure 4. Survey of mediation analysis in the nutrition literature.



Baron and Kenny (20) as further evidence of their use ofthe causal steps method.Are Theoretical Underpinnings Clearly Stated and Supported byPrior Research? The authors reported previous researchrelating child-feeding strategies to food intake (concep-tual theory).

Although they do not present prior evidence of a linkbetween breastfeeding and maternal control of feeding,they establish a clear basis for the hypothesis based onprior research.Are Variables Measured in the Correct Temporal Sequence? Thestudy was observational and longitudinal. It began afterthe child’s first year, when participants were grouped intothose who had breastfed for 12 months and those who didnot. Six months later the outcome measure was assessed.Thus, the independent variable preceded measurement ofthe outcome, but the description of methods is less clearas to whether maternal control of feeding was measuredbefore or concurrent with the dietary records used tocalculate energy intake.Is It a Randomized Intervention Study? The study was a lon-gitudinal observational study and did not have random-ization or an intervention. Therefore, one must be con-cerned about confounders and recognize that making firmconclusions about causal direction is problematic. Wom-en’s existing attitudes about child feeding may have in-fluenced both their decision to breastfeed and their child’slater energy intake. Applying the proposed mediationframework allows readers to interpret the mediation jar-gon and thoughtfully consider the authors’ conclusions.

Mediation in Recent Nutrition LiteratureWe identified 157 articles using the search terms media-tion crossed with keywords “nutrition,” “diet,” “food,” and“obesity” in the title, abstract or text for the 29 monthsbeginning in January 2006. Following the exclusion ofreviews and works that involved non–human beings orexclusively biochemical measures, 72 manuscripts werereviewed. Thirty-two of the 72 did not use statisticalmediation and failed Step 1, including nine with media-tion in the articles’ title.

The 40 articles presented in Figure 4 and Figure 5 useda statistical test of mediation (35-74). Each also includeda discussion of the theoretical underpinnings for theirpurported mediators’ effects (Step 2). The articles aregrouped according to the third and forth framework com-ponents. The majority (Figure 4) were observational stud-ies with data collected at a single or more than one timepoint. Figure 5 presents the 11 articles that reportedmediation findings for controlled intervention trials.

The works listed demonstrate that mediation analysishas been used with many types of nutrition study and isincreasing in frequency (six in 2006, 17 in 2007, and 17 inJanuary through May 2008). However, critical reading isrequired, as almost half of the 72 articles in this inclusivereview discussed mediation without performing tests toestablish its presence.

The growing appreciation of mediation’s utility also issuggested by recent topic reviews. Ventura and Birch (75)reviewed 67 articles assessing the association among par-enting characteristics, child eating habits, and child

Author(s)(reference no.) Subjects

Predictor/independentvariable


Sawatzky andcolleagues (55)

Elderly Canadians Chronic medicalconditions

Quality of life andfunctional abilities

Leisure-time activities

Shin and Shin (56) Korean childrengrades 5 and 6

Obesity Self-esteem Body dissatisfaction

Wansink andcolleagues (57)

French and UScollege students

Country of origin Body mass index Internal and external cues ofmeal cessation

Woo and colleagues(58)

US adolescents Race and parentaleducation

Obesity Breastfeeding history

Zeller and colleagues(59)

Obese USadolescents

Child obesity status Family dynamics andfunctioning

Maternal distress

Observational studies with data gathered at more than one time pointFranko and

colleagues (60)US adolescents Family meals Adolescent health

(smoking, stress anddisordered eating)

Family cohesion and copingskills

Ornelas andcolleagues (61)

US adolescents Perceived parentalinfluences

Vigorous physical activity6 y later

Self esteem and depression atbaseline

Walker andcolleagues (62)

Jamaican children Early childhood stunting Psychological functioning Cognitive ability

Wansink andcolleagues (63)

US secretaries Proximity and coveringcandy dish

Candy intake Self-perceived intake candy

Figure 4. Continued



weight. They found that no study met the criteria neededto test a full mediation model, and that despite increasedpublications on these topics, the “evidence for the influ-ence of parenting and feeding practices on children’s eat-ing and weight status is limited.” The authors noted theimportance of including mediation analysis in subse-quent studies. Authors of a recent review of studies ex-amining the psychosocial predictors of fruit and vegetableintake (76) reached similar conclusions about the impor-tance of stronger experimental designs and including me-diation analysis in the assessment. Thus, the field ap-pears poised for expanded work using these statisticaltechniques, and readers can anticipate more publicationswith mediation methodology.

CONCLUSIONSMediation analysis is a newer statistical tool that canestablish the existence of causal relationships amongvariables. It allows researchers to move beyond simply

establishing an intervention’s efficacy to determiningwhich aspects of an intervention are contributing tochange. By establishing the temporal sequence amongvariables, mediation analysis helps describe behaviors,and for interventions influencing those behaviors, it de-fines means for their modification and improvement. Theappearance of nutrition-related articles using thesemethods is increasing. However, as with most methodol-ogies, it has its own terminology, advantages, and limi-tations. Familiarity with the classes of mediation testsand a few key parameters will allow readers to makeinformed evaluations of articles using these methods.

STATEMENT OF POTENTIAL CONFLICT OF INTEREST:No potential conflict of interest was reported by the au-thors.

FUNDING/SUPPORT: This research was supportedby the National Institutes of Health grants no.

Author(s)(reference no.) Subjects Intervention


Barerra andcolleagues (64)

Postmenopausal USwomen with type 2diabetes

Mediterranean LifestyleProgram

Fat consumption,physical activity, andglycemic control

Social support variables

Barerra andcolleagues (65)

Postmenopausal USwomen with type 2diabetes

Mediterranean LifestyleProgram

Fat consumption,physical activity, andglycemic control

Social support variables

Burke andcolleagues (66)

Overweight hypertensiveAustralian adults

Weight loss dietaryintervention

Saturated fat intake andphysical activity

Self-efficacy, barriers, beliefs,and social support

Campbell andcolleagues (67)

Pooled data of 5 UScommunityinterventions

Population-based strategiesto increase fruit andvegetable consumption

Fruit and vegetableintake

Knowledge and self-efficacy

Doerksen andEstabrooks (68)

Participants in exerciseintervention

Messaging newsletterabout fruit and vegetableintake

Fruit and vegetableintake

Social cognitive messagesnewsletter

Elliot andcolleagues (69)

Career US firefighters Team-based andmotivational interviewing

Fruit and vegetableintake and quality oflife

Knowledge, beliefs, and socialsupport

Epstein andcolleagues (70)

Overweight US children Increasing fruit, vegetable,and low-fat dairyconsumption vs reducinghigh-energy-dense foods

Body mass index change Parent concern for childweight

Fuemmeler andcolleagues (71)

African-American USchurchgoers

Body and Soul intervention Fruit and vegetableintake

Social support, self-efficacy,and autonomous motivation

Haerens andcolleagues (72)

Flemish adolescent girls School-based and parentintervention

Fat intake Psychosocial determinants

Scholz andcolleagues (73)

German adults incardiac rehab

Self-managementinstructions

Depression Perceived achievement goalsand level physical activity

Stice andcolleagues (74)

US adolescent girls Dissonance or healthyweight intervention

Body dissatisfaction,affect, and eatingbehaviors

Thin internal ideal, healthfuleating and physical activity

Figure 5. Survey of mediation analysis in randomized intervention trials in the nutrition literature.



R01CA105835, R01CA105774, R01HL077120, and inpart by PHS M01 RR00334.

References1. MacKinnon DP, Lockwood CM. Advances in statistical methods for

substance abuse prevention research. Prevention Sci. 2003;4:155-171.2. MacKinnon DP, Fairchild AJ, Fritz MS. Mediation analysis. Annu

Rev Psychol. 2007;58:593-614.3. MacKinnon DP. Introduction to Statistical Mediation Analysis. Mah-

wah, NJ: Erlbaum; 2007.4. Shadish WR. Meta-analysis and the exploration of causal mediating

processes: A primer of examples, methods, and issues. Psychol Meth-ods. 1996;1:47-65.

5. Baranowski T, Anderson C, Carmack C. Mediating variable frame-work in physical activity interventions: How are we doing? How mightwe do better? Am J Prev Med. 1998;15:266-297.

6. Kraemer HC, Wilson T, Fairburn CG, Agras WS. Mediators andmoderators of treatment effects in randomized clinical trials. ArchGen Psychiatry. 2002;59:877-883.

7. Baranowski T. Advances in basic behavioral research will make themost important contributions to effective dietary change programs intime. J Am Diet Assoc. 2006;106:808-811.

8. Kristal AR, Glanz K, Tilley BC, Li S. Mediating factors in dietarychange: Understanding the impact of a worksite nutrition interven-tion. Health Educ Behav. 2000;27:112-125.

9. Boushey C, Harris J, Bruemmer B, Archer SL, Van Horn L. Publish-ing nutrition research: A review of study design, statistical analyses,and other key elements of manuscript preparation, Part 1. J Am DietAssoc. 2006;106:89-96.

10. Gray GE, Gray LK. Evidence-based medicine: Applications in dieteticpractice. J Am Dietetic Assoc. 2006;102:1263-1272.

11. Keppel G, Wilkens TD. Design and Analysis: A Researcher’s Hand-book, 4th ed. Englewood Cliff, NJ: Prentice-Hall; 2004.

12. Kraemer HC, Stice E, Kazdin A, Kupfer D. How do risk factors worktogether to produce and outcome? Mediators, moderators, indepen-dent, overlapping, and proxy risk factors. Am J Psychiatry. 2001;158:848-856.

13. Chen HT. Theory-Driven Evaluations. Newbury Park, CA: Sage;1990.

14. Grimes DA, Schultz KF. Bias and causal associations in observationalresearch. Lancet. 2002;359:248-252.

15. Jager KJ, Zoccali C, MacLeod A, Dekker FW. Confounding: What it isand how to deal with it. Kidney Int. 2008;73:256-260.

16. Aiken LS, West SG. Multiple Regression: Testing and InterpretingInteractions. Newbury Park, CA: Sage; 1991.

17. Kraemer HC, Frank E, Kupfer DJ. Moderators of treatment out-comes: Clinical, research, and policy importance. JAMA 2006;296:1286-1289.

18. Rolls BJ, Morris EL, Roe LS. Portion size of food affects energy intakein normal-weight and overweight men and women. Am J Clin Nutr.2002;76:1207-1213.

19. Rippe JM, Hess S. The role of physical activity in the prevention andmanagement of obesity. J Am Diet Assoc. 1998;98(suppl):S31-S38.

20. Baron RM, Kenny DA. The moderator–mediator variable distinctionin social psychological research: Conceptual, strategic, and statisticalconsiderations. J Personality Social Psychol. 1986;51:1173-1182.

21. Holland PW. Causal inference, path analysis and recursive structuralequation models. Sociol Methodol. 1988;18:449-484.

22. Gable S, Chang Y, Krull JL. Television watching and frequency offamily meals are predictive of overweight onset and persistence in anational sample of school-aged children. J Am Diet Assoc. 2007;107:53-61.

23. Rolls BJ, Drewnowski A, Ledikwe JH. Changing the energy density ofthe diet as a strategy for weight management. J Am Diet Assoc.2005;105(suppl 1):S98-S103.

24. Sackett DL, Haynes RB, Guyatt GH, Tugwell P. Clinical Epidemiol-ogy: A Basic Science for Clinical Medicine. Boston, MA: Little, Brownand Company, 1991.

25. MacKinnon DP, Lockwood CM, Hoffman JM, West SG, Sheets V. Acomparison of methods to test mediation and other intervening vari-able effects. Psychol Methods. 2002;7:83-104.

26. Judd CM, Kenny DA. Estimating the Effects of Social Interventions.Cambridge, UK: Cambridge University Press; 1981.

27. Judd CM, Kenny DA. Process analysis: Estimating mediation intreatment evaluations. Eval Rev. 1981;5:602-619.

28. Kraemer HC, Kiernan M, Essex M, Kupfer DJ. How and why criteriadefining moderators and mediators differ between the Baron and

Kenny and MacArthur approaches. Health Psychol. 2008;27(suppl):S101-S108.

29. Freedman LS, Schatzkin A. Sample size for studying intermediateendpoints within intervention trials or observational studies. Am JEpidemiol. 1992;136:1148-1159.

30. Clogg CC, Petkova E, Shihadeh ES. Statistical methods for analyzingcollapsibility in regression models. J Ed Stat. 1992;17:51-74.

31. Sobel ME. Asymptotic confidence intervals for indirect effects instructural equation models. Sociol Methodol. 1982;13:290-312.

32. Sobel ME. Some new results on indirect effects and their standarderrors in covariance structure models. Sociol Methodol. 1986;16:159-186.

33. MacKinnon DP, Krull JL, Lockwood CM. Equivalence of the media-tion, confounding, and suppression effect. Prev Sci. 2000;1:173-181.

34. Fisher JO, Birch LL, Smiciklas-Wright H, Picciano MF. Breast-feed-ing through the first year predicts maternal control in feeding andsubsequent toddler energy intakes. J Am Diet Assoc. 2000;100:641-646.

35. Annus AM, Smith GT, Fischer S, Hendricks M, Williams SF. Associ-ations among family-of-origin food-related experiences, expectancies,and disordered eating. Int J Eat Disord. 2006;40:179-186.

36. Bere E, Brug J, Klepp K-I. Why do boys eat less fruit and vegetablesthan girls? Public Health Nutr. 2007;11:321-325.

37. Beydoun MA, Powell LM, Wang Y. Reduced away-from-home foodexpenditures and better nutrition knowledge and belief can improvequality of dietary intake among US adults. Public Health Nutr. 2008;22:1-13.

38. Beydoun MA, Wang Y. How do socio-economic status, perceived eco-nomic barriers, and nutritional benefits affect quality of dietary in-take among US adults? Eur J Clin Nutr. 2008;62:303-313.

39. Beydoun MA, Gary TL, Caballero BH, Lawrence RS, Cheskin LJ,Wang Y. Ethnic differences in dairy and related nutrient consump-tion among US adults and their association with obesity, centralobesity, and the metabolic syndrome. Am J Clin Nutr. 2008;87:1914-1925.

40. Brown CE, Rempfer MV, Hamera E, Bothwell R. Knowledge of gro-cery shopping skills as a mediator of cognition and performance.Psychiatr Serv. 2006;57:573-575.

41. Brug J, Tak NI, te Velde SJ, Bere E, de Bourdeaudhuij I. Tastepreferences, liking and other factors related to fruit and vegetableintakes among schoolchildren: Results from observational studies.Br J Nutr. 2008;99(suppl 1):S7-S14.

42. Caperchione CM, Duncn MJ, Mummery K, Steele R, Schofield G.Mediating relationship between body mass index and the direct mea-sures of the Theory of Planned Behaviour on physical activity inten-tion. Psych Health Med. 2008;12:168-179.

43. Cerin E, Leslie E. How socio-economic status contributes to partici-pation in leisure-time physical activity. Soc Sci Med. 2008;66:2596-2609.

44. Decaluwe V, Braet C, Moens E, Van Vlierberghe L. The association ofparental characteristics and psychological problems in obese young-sters. Int J Obesity. 2006;30:1766-1774.

45. Hanson MD, Chen E. Socioeconomic status, race, and body massindex: The mediating role of physical activity and sedentary behaviorsduring adolescence. J Pediatr Psychol. 2007;32:250-259.

46. Hesketh K, Ball K, Crawford D, Campbell K, Salmon J. Mediators ofthe relationship between maternal education and children’s TV view-ing. Am J Prev Med. 2007;33:41-47.

47. Jago R, Baranowski T, Baranowski JC, Cullen KW, Thompson D.Distance to food stores and adolescent male fruit and vegetableconsumption: Mediation effects. Int J Behav Nutr Phys Act. 2007;4:35.

48. Janicke DM, Mariel KK, Ingerski LM, Novoa W, Lowery KW, SallinenBJ, Silverstein JH. Impact of psychosocial factors on quality of life inoverweight youth. Obesity. 2007;15:1799-1807.

49. Klepp K-I, Wind M, de Bourdeaudhuij I, Rodrigo CP, Due P, BjellandM, Brug J. Television viewing and exposure to food-related commer-cials among European school children, associations with fruit andvegetable intake: a cross sectional study. Int J Behav Nutr Phys Act.2007;4:46.

50. Luyckx K, Seiffg-Krenke I, Schwartz SJ, Gossens L, Weets I, Hen-drieckx C, Groven C. Identity development, coping, and adjustment inemerging adults with a chronic illness: The sample case of type 1diabetes. J Adolesc Health. 2008;43:451-458.

51. Proper KI, Cerin E, Brown WJ, Owen N. Sitting time and socio-economic differences in overweight and obesity. Int J Obesity. 2007;31:169-176.

52. Mai X-M, Becker AB, Sellers EAC, Liem JJ, Kozyrskyj AL. Infrequent



milk consumption plus being overweight may have great risk forasthma in girls. Allergy. 2007;62:1295-1301.

53. Mond JM, Rogers B, Hay PJ, Darby A. Obesity and impairment inpsychosocial functioning in women: The mediating role of eating dis-order features. Obesity. 2007;15:2769-2779.

54. Sacco WP, Wells KJ, Friedman A, Matthew R, Perez S, Vaughan CA.Adherence, body mass index, and depression in adults with type 2diabetes: The mediational role of diabetes symptoms and self-efficacy.Health Psychol. 2007;2:693-700.

55. Sawatzky R, Liu-Ambrose T, Miller WC, Marra CA. Physical activityas a mediator of the impact of chronic conditions on quality of life inolder adults. Health Qual Life Outcomes. 2007;5:68.

56. Shin NY, Shin MS. Body dissatisfaction, self-esteem, and depressionin obese Korean children. J Pediatr. 2008;152:502-506.

57. Wansink B, Payne CR, Chandon P. Internal and external cues of mealcessation: The French paradox redux? Obesity. 2007;15:2920-2924.

58. Woo JG, Dolan LM, Morrow AL, Geraghty SR, Goodman E. Breast-feeding helps explain racial and socioeconomic status disparities inadolescent obesity. Pediatrics. 2008;121:e458-e465.

59. Zeller MH, Reiter-Purtill J, Modi AC. Controlled study of criticalparent and family factors in the obesigenic environment. Obesity.2007;15:126-136.

60. Franko DL, Thompson D, Affenito SG, Barton BA, Streigel-Moore RH.What mediates the relationship between family meals and adolescenthealth issues? Health Psychol. 2008;27(suppl 2):S109-S117.

61. Ornelas IJ, Perreira KM, Ayala GX. Parental influences on adolescentphysical activity: A longitudinal study. Int J Behav Nutr Phys Act.2007;4:3.

62. Walker SP, Chang SM, Powell CA, Simonoff E, Grantham-McGregorSM. Early childhood stunting is associated with poor psychologicalfunctioning in late adolescence and effects are reduced by psychoso-cial stimulation. J Nutr. 2007;137:2464-2469.

63. Wansink B, Painter JE, Lee Y-K. The office candy dish: proximity’sinfluence on estimated and actual consumption. Int J Obesity. 2006;30:871-875.

64. Barrera M Jr, Toobert DJ, Angell KL, Glasgow RE, MacKinnon DP.Social support and social-ecological resources as mediators of lifestyleintervention effects for type 2 diabetes. J Health Psychol. 2006;11:483-495.

65. Barrera M Jr, Strycker LA, MacKinnon DP, Toobert DJ. Social-ecologi-

cal resources as mediatiors of two-year diet and physical activityoutcomes in type 2 diabetes patients. Health Psychol. 2008;27(suppl):S118-S125.

66. Burke V, Beilin LJ, Cutt HE, Mansour J, Mori TA. Moderators andmediators of behavior change in a lifestyle program for treated hy-pertensives: A randomized controlled trial (ADAPT). Health EducRes. 2008;23:583-591.

67. Campbell MK, McLerran D, Turner-McGrievy G. Mediation of adultfruit and vegetable consumption in the national 5 A Day for BetterHealth community studies. Ann Behav Med. 2008;35:49-60.

68. Doerksen SE, Estabrooks PA. Brief fruit and vegetable messagesintegrated within a community physical activity program successfullychanged behavior. Int J Behav Nutr Phys Act. 2008;4:12.

69. Elliot DL, Goldberg L, Kuehl KS, Moe EL, Breger RK, Pickering MA.The PHLAME (Promoting Healthy Lifestyles: Alternative Models’Effects) Firefighter Study: Outcomes of two models of behaviorchange. J Occup Environ Med. 2007;49:204-213.

70. Epstein LH, Paluch RA, Beecher MD, Roemmich JN. Increasinghealthy eating vs reducing high energy-dense foods to treat pediatricobesity. Obesity. 2008;16:318-326.

71. Fuemmeler BF, Masse LC, Yaroch AL, Resnicow K, Campbell MK,Carr C, Wang T, Williams A. Psychosocial mediation of fruit andvegetable consumption in the body and soul effectiveness trial. HealthPsychol. 2006;25:474-483.

72. Haerens L, Cerin E, Deforce B, Maes L, De Bourdeaudhuij I. Explain-ing the effects of a 1-year intervention promoting a low fat diet inadolescent girls: A mediation analysis. Int J Behav Nutr Phys Act.2008;4:12.

73. Scholz U, Knoll N, Sniehotta FF, Schwarzer R. Physical activity anddepressive symptoms in cardiac rehabilitation: Long-term effects of aself-management intervention. Soc Sci Med. 2006;62:3109-3120.

74. Stice E, Presnell K, Gau J, Shaw H. Testing mediators of interven-tion effects in randomized controlled trials: An evaluation of twoeating disorder prevention programs. J Consult Clin Psychol. 2007;75:20-32.

75. Ventura AK, Birch LL. Does parenting affect children’s eating andweight status? Int J Behav Nutr Phys Act. 2008;5:15.

76. Shaikh AR, Yaroch AL, Nebeling L, Yeh M-C, Resnicow K. Psychos-ocial predictors of fruit and vegetable consumption in adults: A reviewof the literature. Am J Prev Med. 2008;34:535-543.


Mediation Analyses: Applications in Nutrition Research and Reading the Literature

Documents