A Community-Based Heart DiseaseIntervention: Predictors of Change

Marilyn A. Winkleby, PhD, JuneA. Flora, PhD, and Helena C. Kraemer, PhD

IntroductionThe Stanford Five-City Project is a

long-term field trial designed to testwhether a comprehensive program ofcommunity organization and health educa-tion produces favorable changes in cardio-vascular disease risk.' Changes in overall

M: risk factors have been reported2 andW4~k: indicate a beneficial educational effect,tV especially with regard to blood pressure3

K and smoking.4K.'R,'~ In this paper, we present prospective

data on respondents from the two treat-ment cities. We identify those who showedfavorable risk factor changes and com-pare their baseline sociodemographic,psychosocial, and physiologic characteris-

g- tics with those of respondents who showedunfavorable changes. This procedure al-lows for the identification of subgroupswho are most likely to respond to a

|YQ5!:^ community intervention like the Five-CityProject, as well as subgroups who deservespecial attention because of their lack ofpositive change.

Because intervention programs in-creasingly involve the targeting of mul-tiple risk factors"15'6 and because thepnmary goal of community cardiovasculardisease intervention studies is overallimprovement in the community's health,we used a summary measure that incorpo-rates multiple cardiovascular disease riskfactors as our indicator of change. Using asignal detection model designed to de-velop optimal sequential rules,7 we identi-fied baseline health-related variables thatbest divide the sample into subgroups onthe basis of the probability of positivechange in risk factor status.

Our selected variables reflect thetheoretical bases of the Project's educa-tion program: social learning theory,8'9 thecommunication-persuasion model,10 thesocial marketing model,i" and the theory

of reasoned action.12 We evaluate knowl-edge and attitudes about cardiovasculardisease that often accompany behaviorchange13; perceived self-efficacy aboutdiet and exercise, which has been shownto be important in predicting behaviorchange14; perceived cardiovascular dis-ease risk, which has been shown to beimportant in understanding how individu-als change15; and health media use, whichhas been shown to be related to exposureto the Five-City Project campaign.'6

MethodsThe Five-City Project drew subjects

from two treatment and two control citiesin Northern California, ranging in popula-tion size from 35 000 to 145 000 residents(a fifth city was monitored for morbidityand mortality, but not for risk factorchange). To assess change in risk factors,independent cross-sectional surveys ofrandomly selected households and re-peated surveys of a cohort were con-ducted. All persons aged 12 through 74years were eligible to participate and wereinvited to attend survey centers located ineach community. Detailed descriptions ofthe study design and methodology havebeen published previously. ii7, 18

This analysis focuses on adults aged25 through 74 years living in the twotreatment cities who participated in atleast the baseline cohort survey (1979/80)and the final cohort survey (1984/85).

The authors are with the Stanford Center forResearch in Disease Prevention, StanfordUniversity School of Medicine, Palo Alto,Calif.

Requests for reprints should be sent toMarilyn A. Winkleby, PhD, Stanford Centerfor Research in Disease Prevention, StanfordUniversity School of Medicine, 1000 Welch Rd,Palo Alto, CA 94304-1885.

This paper was accepted August 3, 1993.

American Journal of Public Health 767:. No ''

Wimldeby et al.

When these subjects also participated ineither or both of the interim cohortsurveys (1980/81 and 1982/83), interimdata were also used. Eighty-seven percentof those who participated in the baselineand final surveys also participated in bothof the interim surveys.

The Education InterventionThe 6-year education intervention

targeted all residents in the treatmentcommunities and involved a multiple riskfactor strategy delivered through multipleeducational channels.'9 Mass media (tele-vision and radio) programs formed amajor proportion of the intervention.Print media (newspaper health columnsand pamphlets) were delivered throughdirect mail, at work sites, and throughmedical care givers.20 Special emphasiswas given to designing materials for thosewith low literacy levels and disseminatingprograms to low-income Spanish-speak-ing populations, particularly via radio.

Outcome VariableAs an overall measure of change in

cardiovascular disease risk factors, wechose change in a composite risk factorfunction (based on the FraminghamStudy) that provides an estimate of all-cause mortality risk per 1000 persons in 10years.2' The risk function, which is sex-specific, is a combination of cigarettes perday, total plasma cholesterol, mean dia-stolic blood pressure, and pulse (menonly). Although we recognize that anindividual's response may be specific toone component of the risk score (e.g.,smoking) but not to another component(e.g., cholesterol), our objective was toselect an outcome indicative of overallimprovement in health. Furthermore, asingle outcome variable avoids problemsof collinearity and multiple-hypothesistesting. It thus provides a more powerfulmethod of assessing the simultaneouseffects of multiple risk factors than doanalytic methods that become impracticalwhen separate variables are used.

Information on variables used in therisk assessment was obtained by experi-enced nurses and laboratory technicians.Respondents were asked if they hadsmoked any cigarettes in the past week. Ifno cigarettes were smoked, respondentsreceived a code of zero; if one or morecigarettes were smoked, they received acode reflecting the number of cigarettessmoked per day. Although no biochemicalmeasures were used to validate the num-ber of cigarettes smoked per day, previousanalyses have shown a high degree of

accuracy between self-reported and bio-chemical measures for overall smokingstatus.4'22 Total plasma cholesterol wasderived from nonfasting venous samplesand analyzed by standard methods estab-lished by the Lipid Research ClinicProgram.23 Three blood pressure measure-ments were taken on the right arm with asemiautomatic recorder (SphygmetricsSR-2 automatic blood pressure recorder)that minimizes observer bias.17 The meanof the second and third diastolic readingswere used for analyses. Resting pulse wasdetermined by direct palpation for 1minute.

We calculated respondents' compos-ite risk factor functions at each surveytime from the values of their risk factors atthat time. (The mean baseline risk scorewas 3.5 [SD = 1.2].) Age was held con-stant for each individual at his or herbaseline age, allowing for the examinationof change in risk due to factors targeted bythe intervention. After calculating arespondent's risk function values at eachsurvey, we employed a random regressionmodel to fit a unique slope for eachrespondent.24 On the basis of the value oftheir individual slopes, individuals weredivided into two groups, one representingpositive changers (those with a positiveslope, indicating a general improvementin risk score over the course of thesurveys) and one representing negativechangers (those with a negative slope,indicating a worsening of risk score).

Predictor VariablesPredictor variables used in our analy-

sis include six baseline sociodemographicvariables (age, education, marital status,ethnicity, sex, city of residence) and fivebaseline psychosocial variables (cardiovas-cular disease health knowledge, healthattitudes, self-efficacy, perceived risk, andhealth media use).

Information on sociodemographicvariables and psychosocial variables wasobtained from self-reported question-naire data. Age and education werecollected as continuous variables andmarital status (currently married or notcurrently married) and ethnicity (White,Hispanic) were collected as dichotomousvariables. The 2% of the sample who wereAsian and the 2% who were AfricanAmerican were excluded from this analy-sis. Education was selected as the mea-sure of socioeconomic status becauseeducation is available for all individualsregardless of employment status, has highreliability and validity, is generally stableafter early adulthood,2526 and has recently

been shown to have a stronger associationwith cardiovascular disease risk factorsthan income or occupation.27'28

Nine psychosocial predictor vari-ables were initially selected for analyses-these were reduced to five by means of aprincipal components analysis to reducecollinearity. Health knowledge was mea-sured on a summative scale of 17 itemsabout cardiovascular disease risk factors(Cronbach's a = .66). Health attitudeswere measured on a summative scaleencompassing four dimensions-8 physi-cal activity items, 7 weight items, 10nutrition items, and 10 stress manage-ment items (a = .83). Self-efficacy wasbased on 13 diet and exercise items(a = .79), and perceived cardiovasculardisease risk was measured by one item(scale of 1 to 7). Health media use, whichassesses the frequency with which healthinformation is obtained from newspapers,magazines, and pamphlets sent in themail, was based on four items and wasstandardized to a mean of 0 and astandard deviation of 1 (a = .70). Higherscores reflect higher health knowledge,more positive health attitudes, higherself-efficacy, greater perceived risk, andhigher health media use.

Statistical ModelA multiple logistic regression model

including interaction terms is a commonanalytic method for identifying character-istics of respondents who show favorablerisk factor changes compared with thosewho do not. We used this method initially,but we found many significant higher-order interaction terms, making resultsdifficult to interpret. In addition, orderingindividual subjects on a scale, as the riskscore does, is not as useful for planningand implementing prevention programsas is the identification of specific sub-groups. To overcome these problems, weused a signal detection model that pro-vides more interpretable results whenthere are a large number of interactions.7This method, which is a form of recursivepartitioning, uses baseline characteristicsto define distinct subgroups of respon-dents, which are mutually exclusive andmaximally discriminated from each other,on the basis of probability of positivechange in risk function score.

The baseline sociodemographic andpsychosocial predictor variables were en-tered into the signal detection analysisalong with minimum and maximum valuesand interval cutpoints (further informa-tion is available from the authors uponrequest). The signal detection algorithm

768 American Journal of Public Health May 1994, Vol. 84, No. 5

Heart Disease Prevention

then examined each variable and itspossible cutpoints and selected a variableand cutpoint on the basis of a combinedoptimal measure of sensitivity and specific-ity with regard to the outcome measure(cardiovascular disease risk factor change).

After choosing and splitting on thefirst optimally efficient variable, the signaldetection program separately searchedeach subgroup or "branch" of the firstsplit for the next most efficient variableand cutpoint (Figure 1). This procedurewas repeated separately in each subgroupwith all the remaining predictor variables.The procedure ended when (1) therewere too few subjects in a subgroup forfurther analysis, (2) no further significantdiscriminating variable at (P < .05) couldbe found, or (3) no further predictorvariables remained.

ResultsThe cohort treatment sample was

composed of 221 women and 190 menaged 25 through 74 years who participatedin at least the baseline and final surveys.Of those who participated in the firstsurvey, 70.0% returned for the secondsurvey, 58.7% for the third survey, and54.0% for the final survey.

The baseline sample consisted pre-dominantly of White adults (92%) whohad completed at least high school (86%)and who were married (71%). Sixty-ninepercent (n = 284) of the study populationshowed a positive change in their compos-ite risk factor scores over the 6-year studyperiod.

Figure 1 shows the development ofthe optimally efficient algorithm for iden-tifying distinct groups on the basis ofpositive change in cardiovascular diseaserisk factor score. The first optimallyefficient variable (P < .001) that distin-guished positive changers from negativechangers was age (<55 years vs >55years). In the older age group (group 1),83% of the participants showed positivechange. No other variable provided signifi-cant discrimination within this group.

Among younger participants (<55years of age), educational attainmentfurther discriminated respondents. Forthose with higher educational attainment( > 11 years), the next significant split wasfound for self-efficacy. This variable subdi-vided the younger subgroup with highereducational attainment into two smallergroups. In the group with lower self-efficacy scores (group 2), 70% showedpositive change; in the group with higherscores (group 3), 55% showed positive

TOTAL SAMPLE69 % CHANGERS

AGEp<.OOl

Group 1 YOUNGER AGEOLDER AGE (. 55 YEARS)(>55 YEARS)

83% CHANGERS31% OF SAMPLE EDUCATION

p<.Ol

HIGHER EDUCATION(>11 YEARS) LOWI

(42fonr

SELF-EFFICACY SCOREp<.05

Group 2 Group 3LOWER HIGHEL

SELF-EFFICACY SELF-EFFIC(<7 SCORE) (.7 SCORI

70% CHANGERS 55% CHANG45% OF SAMPLE 17% OF SAM

Group 4ER EDUCATION.11 YEARS)% CHANGERSXo OF SAMPLE

zACYE)sERSIPLE

FIGURE 1-Signal detection analysis: development of the optimally efficientalgorithm for identifying distinct groups on the basis of positivechange In cardiovascular disease risk factor score.

change. The last subgroup of youngerparticipants was discriminated on thebasis of lower educational attainment( < 11 years). This group (group 4) had thelowest proportion of positive changers(42%). After the identification of thesefour subgroups, no further significantpredictors were found and the signaldetection test development stopped.

As in all correlational analyses, thefactors selected identify respondents atdifferent risk levels but are not necessarilythemselves causal factors. Thus, to gainfurther understanding of the nature of thefour subgroups, we examined the profileof each (Table 1). Group 1, the oldest agegroup and the most likely to show apositive overall risk function change, wasdistinguished from the other subgroups byhigher cholesterol and blood pressurelevels and higher prevalence of hyperten-sion at baseline. Group 1 also had thehighest perceived risk and highest healthmedia use scores. Group 2, which had

lower self-efficacy scores, had a higherproportion of Hispanics, a higher propor-tion of smokers and heavy smokers,higher rates of hypertension, and higherlevels of cholesterol than did group 3.Members of group 3, which had thehighest self-efficacy scores, were theyoungest participants, the most highlyeducated, the most likely to be single, andthe most likely to be male. They also hadthe lowest perceived risk and the mostpositive risk factor profile. Group 4, thegroup with the lowest proportion ofpositive changers, had the highest propor-tion of women, the highest proportion ofHispanics, and the lowest levels of healthknowledge.

DiscussionIn this paper we provide a prospec-

tive examination of factors associated withchanges in overall cardiovascular diseaserisk factor scores in individuals living in

American Journal of Public Health 769May 1994, Vol. 84, No. 5

XYIr

Winkleby et al.

TABLE 1 Baseline Characteristics of Four Change Groups: Men and Women Aged 25 through 74 Years, Stanford Five-CityProject, 1979 through 1985

Group 2: Group 3:Younger, More Younger, More Group 4:

Group 1: Educated, Lower Educated, Higher Younger, LessOlder Adults Self-Efficacy Self-Efficacy Educated pa

No. in group 127 184 69 31Positive changers, % 83 70 55 42Sociodemographic variablesMean age, y (SD) 63.3 (5.0)b 39.4 (9.3)b 36.0 (8.1)b 40.4 (10.0)b .001Meaneducation, y(SD) 13.1 (3.6) 14.3 (2.3)b 15.1 (2.4)b 8.1 (2.6)b .001Married, % 70.1 75.5 53.6 80.6 .004Hispanic, % 4.7 7.6 1.5 38.7 .001Women, % 55.1 55.4 40.6 67.7 .06

Physiologic variablesSmokers, % 22.2 32.1 26.1 32.3 .27Mean no. cigarettes/d (smokers) 16.5 (10.2) 22.2 (12.9) 12.2 (8.5) 17.7 (9.8) .02

(SD)Mean diastolic blood pressure, mm 81.9 (11.1) 77.6 (10.0) 76.8 (10.7) 77.5 (10.8) .01Hg (SD)

Hypertensive, % 57.6 23.9 15.9 23.3 .001Mean plasma cholesterol, mg/dL 224.5 (37.8) 200.3 (38.7) 192.9 (38.3) 199.7 (41.7) .001

(SD)Mean pulse, beats/min (SD) 69.9 (12.3) 69.7 (9.6) 69.0 (8.7) 70.2 (8.8) .94

Psychosocial variablesMean cardiovascular disease knowl- 6.9 (3.0) 6.9 (2.7) 6.9 (2.9) 4.6 (2.3) .001edge score (SD)

Mean health attitudes score (SD) 4.2 (0.7) 3.8 (0.6) 4.3 (0.7) 3.7 (0.7) .001Mean self-efficacy score (SD) 6.0 (1.5) 5.4 (1.1)b 7.8 (0.6)b 4.8 (1.5) .001Mean perceived risk score (SD) 4.2 (1.6) 3.7 (1.3) 3.3 (1.4) 3.9 (1.4) .001Mean health media use score (SD)C 0.2 (0.7) -0.2 (0.7) -0.1 (0.8) -0.1 (0.8) .001

aBased on analysis of variance for continuous variables and chi-square tests for categorical variables.bNumbers that segmented the sample into subgroups in the signal detection model.cStandardized to a mean of 0 and a standard deviation of 1.

the two Stanford Five-City Project treat-ment cities that received the 6-yeareducation program. Our objective is toprovide researchers who are planningcardiovascular disease community inter-ventions with information about sub-groups that are the most likely (orunlikely) to make positive changes inresponse to intervention programs, espe-cially those involving media interventions.Although other studies have examinedthe effectiveness of demographic andpsychosocial variables in segmenting popu-lations,29'30 few have examined the rela-tionship between segmentation variablesand risk factor changes over time.

Four distinct subgroups of individu-als were identified on the basis of changesin their cardiovascular disease risk factorscores. Group 1 (31% of the sample)consisted of older adults who initially hadthe highest mean blood pressure levels,double the rate of hypertensives of othergroups, and the highest mean cholesterollevels. This group had the most positivemotivation and information-seeking hab-

its, as indicated by its high perceived riskand health media use scores. Given theirpsychosocial and physiologic profiles, it isnot surprising that members of this groupwere the most likely to change; therefore,some portion of the change may reflect aregression to the mean phenomenon.Their perception of risk was accuratelyhigh, their cardiovascular disease knowl-edge was adequate, and their healthmedia use was high (providing a relativelyinformation-rich environment via printmaterials). Moreover, their self-efficacyscores show a confidence about makingbehavioral changes needed to lower risk.The demographic and psychosocial pro-file of members of this subgroup, com-bined with their health media use, makethem receptive to public health cam-

paigns that disseminate risk reductioninformation.

Members of groups 2 and 3, whoconstituted 62% of the sample, were

identified by younger age and highereducational attainment. Group 2 (with70% changers) was distinguished from

group 3 (with 55% changers) by self-efficacy score (5.4 and 7.8, respectively).Group 2 also had a higher smoking rate

(32%, vs 26% in group 3), a higherpercentage of hypertensives (24%, vs 16%in group 3), and a higher mean cholesterollevel (200.3 mg/dL in group 2 and 192.9mg/dL in group 3). The higher percent-age of positive changers in group 2 than ingroup 3 is most likely explained by group2's greater cardiovascular disease risk at

baseline, which gives a greater possibilityfor change.

At first glance, the fact that group 3had a higher level of self-efficacy and a

lower proportion of positive changersseems counter to the large empiricalliterature on self-efficacy and behaviorchange.9'4 On closer examination, how-ever, we see that group 3 was the youngestof all groups, the most educated, and themost likely to be single, non-Hispanic, andmale. This group was also the healthiest interms of blood pressure and cholesterol.Thus, while its members' self-efficacy andhealth attitudes were highest, their low

770 American Journal of Public Health May 1994, Vol. 84, No. 5

levels of perceived risk may indicate thatthey believe they could change but do notsee a need to do so. It is also possible thatmembers of this group have already madechanges and do not see the need forfurther change. The effort to reach thosewho are younger, more educated, andhealthier, but less likely to seek healthinformation via the media, may be mostsuccessful if prevention is stressed viamainstream formats such as newspaperarticles rather than specialized healthmedia such as doctors' columns. Further-more, strategies that incorporate short-term outcomes and use a social influencesframework may be most appealing to thisaudience.

Like group 1, group 4 is one thatcould be labeled "high risk," but unlikegroup 1, group 4 had a low rate of change(42%). Its members were the least edu-cated of all groups (mean of 8.1 years ofschooling). Although predominantlyWhite, almost 40% were Hispanic. Over65% were women. Members of this grouphad the highest smoking rates, the lowestlevels of cardiovascular disease knowl-edge, the lowest health attitude scores,and the lowest self-efficacy scores. Al-though it constitutes only 8% of thesample, this group is at high risk for healthproblems because of its lower socioeco-nomic status27'31 and poorer health knowl-edge and attitudes. Although persons ofall educational levels are beginning toshow substantial declines in cardiovascu-lar disease risk factors,32 those with loweducational attainment may be unlikely torespond to health education campaignsthat are based on a high level ofcardiovas-cular disease knowledge and the initiationof complex actions. It is possible thatmembers of this subgroup were notprovided with the requisite skills to per-form many of the behaviors promoted bythe campaign and that they lived inenvironments where resources were scarceand where norms did not support positivechanges in health. It is clear that moreresearch is needed on how to provideeffective health education to those who fitthe profile of this subgroup. For example,community interventionists need to gain abetter understanding of how poverty andliteracy levels, normative beliefs, andcommunication styles affect behaviorchange. Furthermore, interventionistsneed to increase their understanding ofthe social environments in which adultsfrom lower educational levels reside.33-37

Mayl1994, Vol. 84,No.5

Strengths and LimitationsA major strength of the Five-City

Project is the opportunity it offers toconduct cohort analyses on large numbersof men and women for whom multiplesociodemographic, psychosocial, andphysiologic measurements are available.In addition, participants in the Project aremore likely to represent the communitythan are clinical or volunteer samples,thus enhancing the study's generalizabil-ity. Because of the study's location andthe type of intervention, the results aremost relevant to California communitieswith campaigns similar to the Five-CityProject.

As noted, only 54.0% of those whoparticipated in the baseline survey(1979/80) also participated in the finalcohort survey (1984/85). Although themajority of dropouts were out-migrantsrather than refusers, dropouts were signifi-cantly different from nondropouts on anumber of sociodemographic variables.For example, dropouts were significantlyyounger (mean age 42.2 vs 46.6 years) andless educated (mean years of education12.5 vs 13.4) than nondropouts.

Another potential limitation of thestudy is the possible bias that could arise ifthose with more motivation to changetheir cardiovascular disease risk factorswere more likely to continue participatingin the study. We assessed this potentialsource of error by comparing respondentswho participated in at least the baselineand final cohort surveys with those whodropped out after the baseline survey. Wefound no significant differences betweendropouts and nondropouts on a summaryindex of nine items assessing confidencein changing dietary habits (P = .15) or ona single item assessing intention to quitsmoking (P = .71, smokers only).

Public Health ImplicationsThis study has three important public

health intervention implications. First,this research reinforces the work of othersin pointing out the need for campaigndesigners to segment audiences into morehomogeneous and identifiable subgroupsto enhance the likelihood that eachsubgroup will be reached effectively. Sec-ond, it points out the need to usesociodemographic, psychosocial, andphysiologic variables when segmentingaudiences. Third, the distinct makeup ofthe four population subgroups illustratesthat communitywide health educationcampaigns need to develop specific inter-ventions that target different age, socioeco-

Heart Disease Prevendon

nomic, and cultural subgroups to enhancethe likelihood that all will be reached. fa

AcknowledgmentsThis research was supported by Public HealthService grant lRO1-HL-21906 from the Na-tional Heart, Lung, and Blood Institute to DrJohn W. Farquhar.

The authors gratefully acknowledge theassistance of Drs Stephen P. Fortmann andErica Frank for valuable comments on themanuscript; Beverly Rockhill, Darius Jatulis,and Jennifer Lin for statistical advice; PatriciaCruz for the preparation of tables; and DonnaBalderston, Christine Brereton, Doreen Chap-man, Mary Collins, Kathryn Kandler, JenneferSantee, Sarjeet Singh, and Sammy Van Gundyfor their valuable participation in all phases ofrecruitment and data collection.

References1. Farquhar JW, Fortmann SP, Maccoby N,

et al. The Stanford Five-City Project:design and methods. Am J Epidemiol.1985;122:323-334.

2. Farquhar JW, Fortmann SP, Flora JA, etal. Effects of communitywide education oncardiovascular disease risk factors: theStanford Five-City Project. JAMA. 1990;264:359-365.

3. Fortmann SP, Winkdeby MA, Flora JA,Haskell WL, Taylor CB. Effect of long-term community health education on bloodpressure and hypertension control: theStanford Five-City Project. Am J Epide-miol. 1990;132:629-646.

4. Fortmann SP, Taylor CB, Flora JA, JatulisDE. Changes in adult cigarette smokingprevalence after 5 years of communityhealth education: the Stanford Five-CityProject.AmJEpidemiol. 1993;137:82-96.

5. Blackburn H, Luepker RV, Kline FG, et al.The Minnesota Heart Health Program: aresearch and demonstration project incardiovascular disease prevention. In:Matarazzo JD, Miller NE, Weiss SM, eds.Behavwral Health: A Handbook of HealthEnhancement and Disease Prevention. NewYork, NY: John Wiley & Sons; 1984:1171-1178.

6. Carleton RA, Lasater TM, Assaf AR,Lefebvre RC, McKinlay SM. The Paw-tucket Heart Health Program: an experi-ment in population-based disease preven-tion. R IMedJ. 1987;70:533-538.

7. Kraemer HC. Evaluating Medical Tests:Objective and Quantitative Guidelines. New-bury Park, Calif: Sage Publications; 1992.

8. Bandura A. Self:efficacy mechanisms inhuman agency. Am PsychoL 1982;37:122-147.

9. Bandura A. Social Foundations of ThoughtandAction. Englewood Cliffs, NJ: Prentice-Hall; 1986.

10. McGuire W. Attitudes and attitude change.In: Lindzey G, Aronson E, eds. TheHandbook of Social Psychology. 3rd ed.Hillsdale, NJ: Lawrence Erlbaum; 1985.

11. Lefebvre RC, Flora JA. Social marketingand public health intervention. HealthEduic Q. 1988;15:299-315.

12. Fishbein M, Ajzen I. Belief Attitude, Inten-tion and Behavior. Menlo Park, Calif:Addison Wesley; 1975.

American Journal of Public Health 771

Wrnkleby et al.

13. Roser C, Flora JA, Chaffee SH, FarquharJW. Using research to predict learningfrom a PR campaign. Public Relations Rev.1990;Summer:61-77.

14. Maibach E, Flora JA, Nass C. Changes inself-efficacy and health behavior in re-sponse to a minimal contact communityhealth campaign. Health Commun. 1991;3:1-15.

15. Chaffee SH, Roser C. Involvement and theconsistency of knowledge, attitudes andbehaviors. Commun Res. 1986;13:373-399.

16. Jatulis D, Flora JA, Winkleby MA, Fort-mann SP. Who remembers heart diseaseprograms? Presented at the 119th An-nual Meeting of the American PublicHealth Association; November 12, 1991;Atlanta, Ga.

17. Fortmann SP, Marcuson R, Bitter PH,Haskell WI. A comparison of the Sphygmet-rics SR-2 automatic blood pressure re-corder to the mercury sphygmomanometerin population studies. Am J Epidemiol.1981;114:836-844.

18. Fortmann SP, Rogers T, Haskell WL,Solomon DS, Vranizan K, Farquhar JW.Indirect measures of cigarette use: expiredair carbon monoxide vs. plasma thiocya-nate. Prev Med. 1984:13;127-135.

19. Flora JA, Maccoby N, Farquhar JW.Communication campaigns to prevent car-diovascular disease: the Stanford commu-nity studies. In: Rice RE, Atkin CK, eds.Public Communication Campaigns. New-bury Park, Calif: Sage Publications; 1989:233-252.

20. Chaffee SH, Roser C, Flora JA, FarquharJW. A functional interpretation of a healthcampaign's effects. Health Commun. Inpress.

21. Truett J, Cornfield J, Kannel W. A multi-variate analysis of the risk of coronaryheart disease in Framingham. J Chron Dis.1967:20;511-524.

22. Winkleby MA, Fortmann SP, Rockhill B.Cigarette smoking trends in adolescentsand young adults: the Stanford Five-CityProject. PrevMed 1993;22:325-334.

23. Lipid Research Clinics Manual of Labora-tory Operations. Vol 1. Lipid and LipoproteinAnalysis. Washington, DC: US Dept ofHealth, Education, and Welfare; 1974.

24. Gibbons RD, Hedeker D, Waternaux C, etal. Some conceptual and statistical issues inanalysis of longitudinal psychiatric data.Arch Gen Psychiatry. 1993;50:739-750.

25. Mueller CW, Parcel TL. Measures ofsocioeconomic status: alternatives and rec-ommendations. Child Dev. 1981;52:13-30.

26. Liberatos P, Link BG, Kelsey JL. Themeasurement of social class in epidemiol-ogy. Epidemiol Rev. 1988;10:87-121.

27. Winkleby MA, Fortmann SP, Barrett DC.Social class disparities in risk factors fordisease: eight-year prevalence patterns bylevel of education. Prev Med. 1990;19:1-12.

28. Winkleby MA, Jatulis DE, Frank E, Fort-mann SP. Socioeconomic status and health:how education, income, and occupationcontribute to cardiovascular disease.Am JPublic Health. 1992;82:816-820.

29. Slater MD, Flora JA. Health lifestyles:

audience segmentation analysis for publichealth interventions. Health Educ Q. 1991;Summer:221-233.

30. Snyder LB, Rouse RA. Target the audi-ence for AIDS messages by actual andperceived risk. AIDS Educ Prev. 1992;4:160-172.

31. Marmot MG, McDowall ME. Mortalitydecline and widening social inequalities.Lancet. 1986;ii:274-276.

32. Winkleby MA, Fortmann SP, Rockhill B.Trends in cardiovascular disease risk fac-tors by educational level: the StanfordFive-City Project. Prev Med. 1992;21:592-601.

33. Rogers EM, Shoemaker FF. Communica-tion of Innovations: A Cross-Cultural Ap-proach. 2nd ed. New York, NY: MacMil-lan; 1971.

34. Wallack L, Winkleby MA. Primary preven-tion: a new look at basic concepts. Soc SciMed. 1987;25:923-930.

35. Bunker JP, Gomby DS. Preface. In: Bun-ker JP, Gomby DS, Kehrer BH, eds.Pathways to Health: The Role of SocialFactors. Menlo Park, Calif: Henry J. KaiserFoundation; 1989:15.

36. Dervin B. Audience as listener and learner,teacher and confidante: the sense-makingapproach. In: Rice RE, Atkin CK, eds.Public Communication Campaigns. New-bury Park, Calif: Sage Publications; 1989:67-86.

37. Freimuth VS, Mettger W. Is there ahard-to-reach audience? Public Health Rep.1990;105:232-238.

772 American Journal of Public Health May 1994, Vol. 84, No. 5