Predictors of Attrition Among Rural Breast Cancer Survivors

Predictors of Attrition Among RuralBreast Cancer Survivors

Karen Meneses, Andres Azuero, Xiaogang Su, Rachel Benz, Patrick McNees

Correspondence to Karen Meneses

E-mail: [email protected]

Karen Meneses

Professor & Associate Dean for Research

School of Nursing

University of Alabama, Birmingham, AL

Andres Azuero

School of Nursing


Xiaogang Su

Department of Mathematical Sciences

University of Texas, El Paso, TX

Rachel Benz

School of Nursing


Patrick McNees

School of Health Professions


Abstract: Attrition can jeopardize both internal and external validity. The goal ofthis secondary analysis was to examine predictors of attrition using baseline data of432 participants in the Rural Breast Cancer Survivors study. Attrition predictorswere conceptualized based on demographic, social, cancer treatment, physicalhealth, and mental health characteristics. Baseline measures were selected usingthis conceptualization. Bivariate tests of association, discrete-time Cox regressionmodels and recursive partitioning techniques were used in analysis. Resultsshowed that 100 participants (23%) dropped out by Month 12. Non-linear tree anal-yses showed that poor mental health and lack of health insurance were significantpredictors of attrition. Findings contribute to future research efforts to reduceresearch attrition among rural underserved populations.� 2013 Wiley Periodicals, Inc.

Keywords: attrition; predictors of attrition; secondary analysis; underserved popu-lations; rural women; breast cancer survivors

Research in Nursing & Health, 2014, 37, 21–31

Accepted 5 November 2013

DOI: 10.1002/nur.21576

Published online 11 December 2013 in Wiley Online Library (wileyonlinelibrary.com).

Attrition is typically defined as the loss of a research partici-pant prior to study completion (Deeg, 2002). Attrition is anunderreported problem. Nearly 20 years ago, Ribisl et al.(1996) documented that few investigators even bothered toreport attrition or patterns of attrition. More than a decadelater in 2007, Robinson, Dennison, Wayman, Pronovost,and Needham (2007) noted the continued persistence ofthe problem.

Attrition can jeopardize validity (Clough-Gorr, Fink, &Silliman, 2008). Attrition can threaten internal validitybecause it is difficult to determine whether group differen-ces result from the intervention or dropout (Cook &Campbell, 1979). Attrition can jeopardize external validitybecause participants who completed a study may signifi-cantly differ from those who dropped out (Deeg, 2002;

Demark-Wahnefried, Bowen, Jabson, & Paskett, 2011;Ribisl, Walton, Mowbray, & Boostsmiller, 1996).

Validity threats are a major concern in longitudinalresearch with rural underserved breast cancer survivors.Breast cancer is the most common female cancer, with a5-year relative survival rate of 90% and the vast majority

becoming long-term survivors (American Cancer Society,2013). Longitudinal research in cancer survivorship helpsinvestigators to better understand the processes andchanges occurring among survivors over time, and to iden-tify natural time points for intervention. The NIH Office ofWomen's Health data indicate that roughly one third ofwomen in the United States live in rural areas (Whitten-Goldstein, 2003). Yet, rural residents typically do not partici-pate in research (Baquet, Commiskey, Daniel Mullins, &Mishra, 2006; Bolen et al., 2006; Paskett et al., 2002).Thus, improving our knowledge and understanding of studyparticipation among rural women with breast cancer whoenroll in a longitudinal research study by an examination ofdescriptors and predictors of attrition in this underservedpopulation is warranted.

The purpose of this study was to determine predictorsof attrition using baseline characteristics among 432 partic-ipants in the Rural Breast Cancer Survivors (RBCS) study.Using secondary analysis of the baseline data, the specificaims were to: (a) describe sociodemographic, cancer treat-ment, physical health, and mental health characteristics at

�C 2013 Wiley Periodicals, Inc.

baseline; (b) describe attrition among rural women partici-pating in the RBCS study; (c) determine whether baselinecharacteristics predict risk of attrition in the RBCS study.We conclude by discussing opportunities and challenges toreducing attrition among underserved rural women.

The Concept of Attrition

There is a gap in the literature concerning conceptualframeworks in which to explore attrition (Marcellus, 2004;Robinson et al., 2007). Thus, the authors conceptualizedattrition based on demographic, social, cancer treatment,

physical health, and mental health characteristics of thestudy population of rural breast cancer survivors.

Demographic Characteristics

Demographic variables predicted attrition in previous longi-tudinal studies (Chatfield, Brayne, & Matthews, 2005; deGraaf, van Dorsselaer, Tuithof, & Ten Have, 2013; Young,Powers, & Bell, 2006). Chatfield et al. (2005) conducted asystematic review of attrition occurring among the elderlyafter baseline in longitudinal studies. The review was limitedto epidemiologic, population-based studies with a minimumof 1,000 participants aged 65 years and older. The investi-gators found three significant factors predicting attritionfrom baseline data: advancing age, lower education, andcognitive impairment.

Haring et al. (2009) used data from a longitudinalpopulation-based study of health among more than 7,000residents of the Pomerania region of Germany. They foundthat older age, lower educational level and unemploymentwere significant demographic and social predictors of attri-

tion. In their conclusion, the authors suggested that whenrecruiting potentially underserved populations, special tech-niques should be considered to increase the efficiency ofresponse.

Rural residence influences study recruitment (Baquetet al., 2006; Paskett et al., 2002). Baquet and coworkersuncovered serious gaps in efforts to recruit rural residentsin clinical trials, and found rural residence as a signficantpredictor of study participation among rural residents inMaryland's Eastern and Western Shore. Paskett et al.(2002) developed an intervention program to improverecruitment of North Carolina rural participants in cancerclinical trials consisting of rapid tumor-reporting system,nurse facilitation, quarterly newsletter communication, andhealth education. Yet, they found that the intervention wasineffective in improving rural participation, and like Baquetet al., confirmed thorny problems in rural recruitment.

Social Characteristics

Positing that social characteristics may be associated with

risk of attrition, Young et al. (2006) investigated patterns of

attrition in the Australian Longitudinal Study on Women'sHealth (ALSWH). The participants comprised a nationallyrepresentative sample of young women (age 18–23), mid-dle age women (age 45–50), and older women (age 70–75). Recognizing the paucity of rural participation, the inves-tigators oversampled women living in rural and remote Aus-tralian regions. They found that social characteristics wererelated to attrition. Attrition occurred most strikingly amongyoung women (21%) compared with older women (2%).Younger and middle-age women who dropped out weremore likely to be single (i.e., separated, divorced, or wid-owed) and experiencing financial difficulty. Clough-Gorret al. (2008), in a longitudinal study of older breast cancersurvivors, also noted that lack of finances was associatedwith attrition.

Cancer Treatment Characteristics

There is a paucity of evidence in the literature on how can-cer treatment characteristics relate to attrition in behavioral,

longitudinal studies. Clough-Gorr et al. (2008) found thatadvanced stage of disease at diagnosis and hormonal ther-apy (i.e., tamoxifen) were associated with attrition.

Physical Health Characteristics

Physical health is associated with attrition. Clough-Gorret al. (2008) found that a larger number of comorbid condi-tions was associated with higher attrition due to death orloss to follow-up. Among those who died, poorer health andworse tumor characteristics were noted. Similarly, Van Beij-sterveldt et al. (2002) found that poor overall health pre-dicted attrition.

Mental Health Characteristics

Poor mental health is associated with attrition (Clough-Gorret al., 2008; Ribisl et al., 1996; Young et al., 2006). In aprospective study of cardiovascular prevention, Bambset al. (2013) found that participants with CES-D scoresindicative of clinical depression were more likely to be lostto follow-up. Stressful life events such as caregiving anduncontrollable events (i.e., natural disasters, family deaths)that influence mental health appeared to increase drop-out or withdrawal (Ribisl et al., 1996; Young et al., 2006).Depression was also identified by Lee, Hayes, McQuaid,and Borrelli (2010) as a predictor of attrition from a smok-

ing cessation program. Individual participants who hadlower depressed mood scores were more likely than thosewith higher levels to complete the program (Lee et al.,2010).

From the above evidence, a multi-dimensional con-cept of attrition predictors derived from the literature andcontaining demographic, social, cancer treatment, physicalhealth, and mental health characteristics was applied to thissecondary analysis of breast cancer survivor attrition.

Research in Nursing & Health

22 RESEARCH IN NURSING & HEALTH

Methods

Study Population and Sample

The present study is a secondary analysis of data from theRural Breast Cancer Survivors Study (RBCS), a population-based study of psychoeducational support interventionsamong rural Florida-dwelling women. A three-prongedapproach using state cancer registry data with both activeand passive recruitment techniques was used for recruit-ment. Eligibility criteria were: women diagnosed with Stage0–III breast cancer, living in rural areas of Florida, within thefirst 3 years of completion of primary breast cancer treat-ment, at least 21 years of age, and with telephone or cell-phone access. Women receiving anti-hormonal therapy(i.e., tamoxifen or aromatase inhibitor) or anti-HER2 treat-ment at the time of study initiation were also eligible.Women with metastatic disease at the time of study con-tact, men with a breast cancer diagnosis, and those withno telephone or cellphone access were excluded fromparticipation.

Rural eligibility was established based on one of twocriteria: (1) residence in 1 of 33 Florida rural counties desig-nated by Florida statute (Florida Department of Health,2007); or (2) residence in a rural pocket of one of 34 Floridaurban with an Index of Research Access (IRA) score equalto or greater than 4. This score can be roughly translatedinto a criterion that women had at least four times the diffi-culty in securing face-to-face post-treatment support thanwomen who lived within one mile of the nearest treatmentfacility. Details and calculations for the IRA score havebeen reported in detail elsewhere (Meneses, Azuero,Hassey, McNees, & Pisu, 2012). The IRA was particularlysuitable for recruitment of rural residents because it pro-vides a detailed determination of rurality that cannot beobtained by county residence alone.

Once enrolled, participants were randomized to either

the Experimental (early intervention) or Wait Control (lateintervention) groups. The CONSORT diagram is shown inFigure 1. A total of 641 (13.5%) rural breast cancer survi-vors responded with interest in study participation. Of these,432 (67.4%) subsequently enrolled in the RBCS and wererandomized to one of two study arms. Study participationlasted 12 months. The specific details of the processes ofstudy recruitment, enrollment, randomization, retention, andintervention are described in detail elsewhere (Meneses,under review; Meneses, Benz, Hassey, Yang, & McNees,2013).

Measures

Demographic characteristics. The Breast CancerSurvivor Sociodemographic and Treatment Survey contains12 demographic variables including age, race, ethnicity, edu-cation level, marital status, type of health insurance, workstatus, and number of people living in the home. This mea-sure has been used in prior studies by the authors.

Social characteristics. The Medical OutcomesStudy Social Support Index (MOS-SSS) was used toassess social characteristics (Sherbourne & Stewart, 1991).The MOS-SSS contains an overall functional social supportindex with scores range from 0 to 100. Higher scores indi-cate more social support. In the present study, the MOS-SSS overall functional social support index had a Cronbacha reliability of .96.

Cancer treatment characteristics. The Socio-demographic and Treatment Survey contains 12 cancertreatment items related to breast cancer treatment includingmonths since diagnosis and time since end of treatment,type of breast cancer surgery, radiation therapy, chemother-apy, hormonal therapy, and/or anti-HER2 therapy.

Physical health. The Medical Outcomes SurveyShort Form v1 (SF-36) is a commonly used measure ofphysical health and function (Ware & Sherbourne, 1992).The SF-36 has 36 items that aggregate into eight subscalesconsisting of 2–10 items, and two overall summary scoresthat aggregate the eight subscales. The two overall sum-mary scores are the Physical Component Summary score(PCS) and the Mental Component Summary score (MCS).Four health concepts (i.e., physical functioning, role-physi-cal, body pain, and general health form the aggregate PCSscore).

Mental health. Two measures of mental healthwere used in the present study. First, the SF-36 v1 previ-ously described contains four health concepts (i.e., vitality,social functioning, role-emotional, and mental health) that

form the aggregate MCS score. The SF-36 v1 norm scoresrange from 0 to 100, where higher scores indicate betterphysical health and mental health, respectively. A normscore of 50 indicates a level comparable to that of the aver-age adult in the U.S. population. In the present study, thePCS and the MCS summary scores had Cronbach a reli-ability of .93 and .90 respectively.

The second measure of mental health used was theCenter for Epidemiologic Studies Depression Scale (CES-D), a widely used short self-report scale to measure depres-sive symptoms in the general population (Radloff, 1977).Total score ranges from 0 to 60, where higher scores indi-cate higher depressive symptomatology. Scores �16 sug-gest clinically significant levels of psychological distress.In the RBCS study, the CES-D had Cronbach a reliabilityof .91.

Analytic Strategy

Attrition was defined as non-completion of the study byMonth 12. To determine potential predictors of attrition,baseline demographic, social, cancer treatment, physicalhealth, and mental health characteristics were tabulated.The time point for the last data collection was tabulated for

all participants and was used as the indicator of retention (ifcompleted the Month 12 battery of measures) or attrition (ifthe last completed battery of measures was at baseline or


23PREDICTORS OF CANCER SURVIVOR ATTRITION/MENESES ET AL.

Month 3). Among those who dropped out, the median lastdata collection time point was computed.

First, descriptive statistics for the variables of interestwere computed at each of three study time points (i.e.,baseline, median last data collection time point, and Month12). Second, using bivariate tests of association, baselineparticipant characteristics were tested for differencesbetween those retained and those who dropped out by themedian last data collection time point, and by the end of thestudy (Month 12), with the purpose of identifying individualpredictors that were consistently associated with attritionacross the study period.

Third, a time-to-event analysis was performed. Dis-crete-time Cox regression models were used to assess theassociation between the baseline characteristics, when

considered simultaneously, and the instantaneous risk orhazard of attrition (i.e., the probability that a participantdropped out from the study at a certain time point, giventhat the participant had been retained up to that timepoint; Allison, 2010). An initial model with all covariates(i.e., a saturated model) was fitted. Next, a reduced modelwas fitted using a step-wise procedure that minimizedthe Akaike Information Criterion (AIC). A likelihood ratiotest was conducted to obtain a formal test of differencein goodness-of-fit between the saturated and reducedmodels. Wald tests were used to assess the significanceof regression coefficients in the final model and toextract interpretations in terms of hazard ratios. The pro-portional hazards assumption was tested in the final modelby fitting time-by-covariate interactions and using a linear

Assessed for Eligibility (n=4939 G1 thru G17)

Excluded (n=4507)

Randomized (n=432)

Experimental(n=215)

ExperimentalMonth 3 (215-28=187)Month 6 (215-49=166)Month 9 (215-64=151)Month 12 (215-41=174)

ExperimentalAnalyzed (n=174)

Enro

llmen

tA

lloca

tion

Follo

w-u

pA

naly

sis

Wait Control(n=217)

Wait ControlMonth 3(217-43=174)Month 6 (217-47=170)Month 9 (217-101=116)Month 12 (217-59=158)

Wait ControlAnalyzed (n=158)

Declined=1100

No Response=2779

Interested but not enrolled=209

Returned Mail=419

FIGURE 1. RBCS CONSORT.



contrast to conduct an overall test for these interactionterms.

A second exploratory analysis was conducted to iden-tify rules that would allow a classification of participants atrisk for attrition. To this end, recursive partitioning techni-ques were used to construct classification tree models forattrition status using baseline characteristics as predictors(Everitt, 2009; Su, McNees, Meneses, & Johnson, 2011).Classification trees are useful alternative models for identi-fying relevant combinations of predictors that cannot beuncovered using traditional statistical models that assumelinearity of predictors. Unlike linear models, tree models arenot easily described using equations and therefore are typi-cally shown graphically.

All the above analyses were conducted using SASv9.3 statistical software, except for the tree analysis, whichwas conducted using the package Party in the R statisticalsoftware (Hothorn, Hornik, & Zeileiswere, 2006; R-statisti-cal-software-v2.15.00., 2008). Statistical significance washeld at the p< .05 level.

Results

Characteristics at Baseline

Table 1 (left column) lists demographic, cancer treatmentand some social characteristics of the 432 RBCS partici-pants at baseline study entry. Missing data were detectedin the household income variable, with 64 (14.8%) decliningto provide information on income. The sample was predom-inately Caucasian, with a mean age of 63.1 years, married/partnered, and educated at the trade school/some college

level. More than 30% had household incomes <$30,000.The majority of participants were retired, and more than94% had health insurance. The average time since diagno-sis to study entry was 25.6 months (SD¼ 7.9) with an aver-age time since completion of primary breast cancertreatment of 18.8 months (SD¼ 7.9). The majority (56.9%)had lumpectomy and radiation therapy for control of localdisease, and chemotherapy (57.9%) for control of regionalsystemic disease. More than 66% received hormonal block-ing agents to reduce the risk for recurrence.

Participants reported minor to moderate levels ofdepressive symptomatology (M¼ 10.5, SD¼ 10.2), butbelow clinically significant levels of depression (CES-D cutscore of 16). Overall social support was good at 79% of themaximum scores. The SF-36 PCS and MCS compositescores were slightly below the general U.S. population aver-age estimate.

A total of 100 RBCS participants (23%) dropped outfrom the study. Table 1 (middle and right columns) furtherdemonstrates the bivariate tests of association betweenbaseline characteristics and attrition by Month 3 and byMonth 12. Of the baseline characteristics including groupassignment, demographics, social, cancer treatment, physi-cal health, and mental health, the characteristics that were

consistently and significantly associated with attrition acrossthe study period included not having health insurance,higher depressive symptomatology, lower mental health,and younger age.

Table 2 shows the time-to-event analysis for the lastdata collection time point and includes the initial saturatedCox model and the reduced Cox model. In the initial satu-rated model, baseline characteristics significantly associ-ated with attrition hazard were full-time employment, nothaving health insurance, social support, and lower mentalhealth. However, in the final reduced model, five character-istics remained: full time employment, education at thetrade school/some college level, not having health insur-ance, lower mental health, and hormonal therapy were sig-nificantly associated with the attrition hazard. Participantswith these baseline characteristics had significantly higherattrition when compared to those without these characteris-tics. Per the AIC criterion, the reduced model provided abetter balance between number of predictors and modelaccuracy than the saturated model. Goodness of fit was notsignificantly different between the two models (p¼ .071).The proportional hazards assumption was not significantlyviolated by any of the predictors in the final model (linearcontrast x2[5]¼ 4.95, p¼ .422).

An initial exploratory classification tree model indi-cated that lower mental health at baseline was a character-istic strongly associated with the risk of attrition (Fig. 2,left panel). Approximately 49% of participants with SF-36MCS scores at or below 31.307 dropped out, compared

to 20.5% among those with MCS scores greater than31.307. Because computing norm scores for the SF-36 v1requires specialized scoring algorithms, using this scaleas a screening device to identify rural participants atrisk for attrition may not be practical. However, a strongassociation between the CES-D and the SF-36 MCSscores was observed (Pearson r¼�.77, p< .001). Thus, asecond exploratory classification tree model was fittedafter removing the SF-36 MCS scores from the pool ofpredictors.

This second model (Fig. 2, right panel) used baselineCES-D scores and health insurance status to identify ruralbreast cancer survivors at risk for attrition. Approximately35% of participants with higher levels of depressive symp-tomatology (i.e., CES-D scores >12) dropped out of thestudy. However, 17% of participants who reported lowerlevels of depressive symptomatology (i.e., CES-D scoresless than or equal to 12) and had health insurance droppedout, compared with 54% of participants reporting lower lev-els of depressive symptomatology who had no healthinsurance.

Discussion

Outcomes from this secondary analysis of RBCS data haveimplications for future work in attrition among rural under-served women and may inform the design of future post-



Table 1. Baseline Sociodemographic, Treatment, and Selected Psychometric Measures for RBCS Participants, Compared byAttrition Status at Month 3 and Month 12

Baseline Month 3 Month 12

N¼ 432

Retained

(n¼ 361)

Dropped

(n¼ 71)

pa,b

Retained

(n¼ 332)

Dropped

(n¼ 100)

pa,cn % n % n % n % n %

Group assignment .056 .0592

Wait control 217 50.2 174 48.2 43 60.6 158 47.6 59 59.0

Experimental 215 49.8 187 51.8 28 39.4 174 52.4 41 41.0

Race .5698 .4832

Caucasian 408 94.4 342 94.7 66 93.0 315 94.9 93 93.0

Other 24 5.6 19 5.3 5 7.0 17 5.1 7 7.0

Annual income .8444 .1377

�$30,000 131 30.3 107 29.6 24 33.8 94 28.3 37 37.0

$30,001–$50,000 80 18.5 66 18.3 14 19.7 58 17.5 22 22.0

>$50,000 157 36.4 133 36.9 24 33.8 127 38.2 30 30.0

No response 64 14.8 55 15.2 9 12.7 53 16.0 11 11.0

Marital status .0536 .0459

Married/partnered 315 72.9 270 74.8 45 63.4 250 75.3 65 65.0

Other 117 27.1 91 25.2 26 36.6 82 24.7 35 35.0

Employment .1249 .1420

Full-time 111 25.7 85 23.5 26 36.6 77 23.2 34 34.0

Part-time 62 14.3 54 15.0 8 11.3 47 14.2 15 15.0

Retired 196 45.4 170 47.1 26 36.6 159 47.9 37 37.0

Other 63 14.6 52 14.4 11 15.5 49 14.7 14 14.0

Education .4777 .0574

<High school 24 5.6 22 6.1 2 2.8 21 6.3 3 3.0

High school grad 97 22.4 81 22.4 16 22.5 73 22.0 24 24.0

Trade school/some college 149 34.5 119 33.0 30 42.2 104 31.3 45 45.0

College grad 104 24.1 90 24.9 14 19.7 86 25.9 18 18.0

Postgraduate 58 13.4 49 13.6 9 12.7 48 14.5 10 10.0

Health insurance .0102 .0118

Yes 408 94.4 346 95.8 62 87.3 319 96.1 89 89.0

No 24 5.6 15 4.2 9 12.7 13 3.9 11 11.0

Surgery .2629 .3310

Lumpectomy 246 56.9 208 57.6 38 53.5 192 57.8 54 54.0

Mastectomy 128 29.6 109 30.2 19 26.8 100 30.1 28 28.0

Bilateral mastectomy 58 13.5 44 12.2 14 19.7 40 12.1 18 18.0

Chemotherapy .4420 .9761

Yes 250 57.9 206 57.1 44 62.0 192 57.8 58 58.0

No 182 42.1 155 42.9 27 38.0 140 42.2 42 42.0

Radiation therapy .0526 .0695

Yes 304 70.4 261 72.3 43 60.6 241 75.6 63 63.0

No 128 26.9 100 27.7 28 39.4 91 27.4 37 37.0

Hormonal therapy .1470 .0666

Yes 288 66.7 246 68.1 42 59.1 229 31.0 59 59.0

No 144 33.3 115 31.9 29 40.8 103 69.0 41 41.0

M SD M SD M SD pa,b M SD M SD

Age 63.1 10.5 63.7 10.5 60.2 10.1 .0105 63.8 10.2 60.7 11.1 .0110

Mo’s since Dx 25.6 7.9 25.5 7.8 25.8 8.1 .7564 25.5 7.8 25.8 8.0 .7270

Mo’s since end of Tx 18.8 8.6 18.8 8.6 18.6 8.5 .7708 18.9 8.7 18.6 8.1 .7729

Depression CES-D 10.5 10.2 9.7 9.6 14.3 11.8 .0029 9.6 9.6 13.6 11.3 .0016

MOS overall social support 79.1 21.2 79.4 20.4 77.6 24.4 .5656 79.3 20.5 78.2 23.3 .6261

SF-36 PCS 45.0 10.6 45.3 10.3 42.9 11.6 .0801 45.5 10.1 43.1 11.7 .0730

SF-36 MCS 48.8 11.5 9.7 10.9 44.2 13.0 .0013 49.9 10.7 44.9 13.0 .0006

aChi-squared, Fisher’s exact, or t tests, as appropriate.bTest of difference in baseline characteristics between retained and dropped-out by Month 3.cTest of difference in baseline characteristics between retained and dropped-out at the end of the study (Month 12).



Table 2. Characteristics of RBCS Participants Associated With Attrition Hazard in Cox Time-to-Event Analysis

Covariate

Saturated Model Reduced Model

Hazard Ratio (95% CI) p Hazard Ratio (95% CI) p

Group assignment

Experimental 0.65 (0.42, 1.01) .0646 — —

Wait control Ref.

Agea 0.71 (0.5, 1.01) .0641 — —

Race

Caucasian 0.72 (0.29, 1.79) .5520 — —

Other Ref.

Annual income

�$30,000 1.36 (0.62, 3.01) .4468 — —

$30,001–$50,000 1.61 (0.71, 3.63) .2560 — —

>$50,000 0.96 (0.42, 2.15) .9549 — —

No response Ref.

Marital status

Married/partnered 0.64 (0.37, 1.13) .1372 — —

Other Ref.

Employment

Full-time 2.65 (1.23, 5.73) .0174 1.71 (1.1, 2.65) .0173

Part-time 2.20 (0.91, 5.35) .1038 — —

Retired 1.98 (0.83, 4.72) .1479 — —

Other Ref.

Education

<High school 0.54 (0.13, 2.21) .4298 — —

High school grad 1.30 (0.55, 3.05) .5632 — —

Trade school/some college 1.41 (0.64, 3.07) .4228 1.58 (1.03, 2.41) .0345

College grad 0.97 (0.42, 2.22) .9459 — —

Postgraduate Ref.

Health insurance

Yes 0.37 (0.16, 0.85) .0251 0.40 (0.2, 0.83) .0130

No Ref.

Months since diagnosisa 1.42 (0.97, 2.08) .0812 — —

Months since end of treatmenta 0.70 (0.47, 1.03) .0807 — —

Surgery

Lumpectomy 1.18 (0.53, 2.6) .7253 — —

Mastectomy 0.74 (0.36, 1.52) .3933 — —

Bilateral mastectomy Ref.

Chemotherapy

Yes 0.74 (0.43, 1.27) .3011 — —

No Ref.

Radiation therapy

Yes 0.53 (0.28, 0.99) .0642 — —

No Ref.

Hormonal therapy

Yes 0.69 (0.45, 1.08) .1253 0.65 (0.42, 0.99) .0433

No Ref.

Depression CES-Da 0.99 (0.7, 1.42) .9618 — —

MOS overall social supporta 1.33 (1.02, 1.74) .0468 — —

SF-36 PCSa 0.79 (0.62, 1) .0669 — —

SF-36 MCSa 0.68 (0.49, 0.96) .0387 0.73 (0.6, 0.89) .0018

Fit statistics

AIC 683.924 673.059

�2 log likelihood (model parameters)b 631.924 (26) 663.059 (5)

Notes. AIC, Akaike Information Criterion. Smaller values indicate a better trade-off between the accuracy of a model and the number of

predictors included in the model.aHazard ratio for a standard deviation increase.bUsed to test the difference in fit between models: x2(21)¼ 31.135, p¼ .0714.



treatment support service design. Findings from the time-to-event modeling indicated that rural breast cancer survivorsat risk for early attrition from a psychoeducational programhad poorer mental health (which was strongly associatedwith higher depressive symptomatology), lack of health

insurance, full time employment, education at trade school/some college level, and no hormonal therapy. Across thedifferential testing and modeling analyses, the two baselinecharacteristics that robustly predicted higher attrition werelower mental health and lack of health insurance.

Almost half of participants reporting poorer mentalhealth (Fig. 2, left panel) dropped out of the study. This sig-nificant predictor of attrition supports previous work by otherinvestigators (Bambs et al., 2013; Clough-Gorr et al., 2008;Ribisl et al., 1996; Young et al., 2006). Bambs, in particular,found that depressive symptomatology and lower number ofsocial networks were associated with attrition. To our knowl-edge, this is the first study to report lower mental health asa predictor of attrition specifically among rural residents.While Young et al. (2006) included rural residents in theirsample, they did not analyze rurality as a predictor of attri-tion. It may be that lower mental health increased the bur-den of daily activities and mental resources, leaving lesstime and availability to remain engaged and participate in alongitudinal research study.

Lack of health insurance was the second major pre-dictor of attrition in this study. Bambs et al. (2013), in theirprospective study of cardiovascular prevention, also foundthat lack of health insurance was a significant predictor ofattrition. Young et al. (2006) and Clough-Gorr et al. (2008)did not specifically identify a lack of health insurance

but did report that inadequate finances and financialdifficulties were associated with attrition. Costs for medicalfollow-up and routine cancer surveillance, such as mam-mography, blood studies, and DEXA scans, may requirecontinued employment to maintain health insurance cover-

age, which may take precedence over a rural breast cancersurvivor's ability or desire to remain in a research studyover time.

While depressive symptomatology and lack of healthinsurance were independent predictors of attrition, the com-bination of the two characteristics significantly increasedthe likelihood of attrition. Nearly three times the proportionof RBCS participants reporting higher depressive symptom-atology dropped out if they had no health insurance, com-pared with those with depressive symptomatology who hadhealth insurance. In a previous study of urban breast can-cer survivors, the combination of worry about health insur-ance and out-of-pocket costs was problematic (Meneseset al., 2012). A significant increase in insurance premiumswas reported by breast cancer survivors within the first yearof cancer treatment, and an increase in economic hardshipevents was significantly associated with poor quality of life(Meneses et al., 2012). A difficult reality for some cancersurvivors is that lack of health insurance added economicburden to out-of-pocket costs, particularly among thosewith lower income (Pisu, Azuero, Meneses, Burkhardt, &McNees, 2011).

Full-time employment, education at the trade school/some college level, and no hormonal therapy were associ-ated with attrition in the time-to-event models. Full-timeemployment and/or return to full-time work status after

FIGURE 2. Classification tree models for attrition proportion.



breast cancer treatment may take priority, leaving less avail-able time for continued study participation. Alternatively, thedemands of full-time employment may have spurred survi-vors to move on with their lives after treatment ended.

The RBCS showed that lower education predictedattrition among rurally underserved breast cancer survivors.Findings support previous work by Chatfield et al. (2005)and Haring et al. (2009), who also found that lower educa-tion predicted attrition in their longitudinal observationalstudies. Clough-Gorr et al. (2008) also identified that hor-monal therapy was associated with attrition, but receipt ofhormonal therapy was associated with advanced diseaseand higher likelihood of attrition.

Contrary to other longitudinal studies with the elderly,older age was not associated with attrition in the RBCSstudy. Our findings are consistent with those of Van Beij-sterveldt et al. (2002), who did not find higher attritionamong older participants. They attributed their discrepantfindings to having relatively healthy older adults who couldparticipate and to compliance-promoting measures to main-tain retention in their study.

Implications for Future Research

The challenges faced over the course of this study providedvaluable information concerning how to manage attrition infuture work with rural residents and in breast cancer survi-vorship research. Ribisl et al. (1996) categorized causes ofattrition into participant characteristics and study character-istics. Participant characteristics include demographic,social, physical, and mental health features that are specificto the study population (Ribisl et al., 1996). We recommendthat investigators start to document and report participantcharacteristics associated with attrition in their longitudinalobservational and interventional studies. Reporting attritiondata will aid in improving our conceptual understanding andcan further determine baseline characteristics that predictattrition. The predictors tested here represent a beginningmodel conceptualizing the participant characteristics lead-ing to attrition. Additional conceptual development is war-ranted through empirical studies of the relationship betweenattrition and recruitment, enrollment, and retention in longi-tudinal studies.

Better understanding of attrition patterns occurring inlongitudinal studies of those who are hard to reach, such asrural residents and other underserved populations, is vital.Planned recruitment and oversampling of participants withbaseline study characteristics that predict attrition, suchas higher depressive symptomatology and lack of healthinsurance as was found in the present study, may aid inves-tigators to manage attrition. Chatfield et al. (2005) also sug-gested proxy interviews to help reduce attrition.

Study characteristics including design, methodologic

strategies, and analytic schemes (Ribisl et al., 1996) arelargely under the control and oversight of the researcher,who can actively implement retention strategies to reduce

future attrition. Deeg (2002) suggested that attrition must bemonitored because reasons for attrition can change overtime in a longitudinal study, which may ultimately influenceestimates of change, and even in intervention studies attri-tion can reduce power. Fayter, McDaid, and Eastwood(2007) suggested that prospective monitoring of study char-acteristics to manage and maintain study retention mayhelp reduce attrition bias over time.

Strengths and Limitations

Strengths of this study include a population-based sampleof rural breast cancer survivors. Previous studies have notreported attrition specifically in rural residents. Although weused the common approach of measuring attrition only untilthe intervention endpoint, we recognize that in longitudinalstudies there are other important concerns and questions,such as the durability or sustainability of intervention effectsover time. Such analyses are beyond the scope of thepresent work and remain an important topic of futureinvestigation.

Conclusion

Attrition is an expected outcome in any longitudinal study.Gaining an understanding of demographic, social, treat-ment, physical health, mental health, and other characteris-tics of study participants enrolled in longitudinal studies canhelp identify those at risk of early attrition and dropout. Ourfindings showed that poorer mental health and lack ofhealth insurance together were significant predictors ofretention among rural breast cancer survivors. These defin-ing attrition characteristics should be considered whendesigning future post-treatment support services. Becausemore than 30% of women reside in rural areas, understand-ing and improving methods to retain rural breast cancer sur-vivors as participants in cancer control research arewarranted. The benefits could be realized not only in termsof richer data but support services that are more responsiveto women who exhibit characteristics of attrition risk.

References

Allison, P. (2010). Survival analysis using SAS: A practical guide.

(2nd ed.). Cary, NC: SAS Institute, Inc.

American Cancer Society (2013). Cancer facts & figures, 2013.

Retrieved from http://www.cancer.org/acs/groups/content/@epi-

demiologysurveilance/documents/document/acspc-036845.pdf.

Bambs, C. E., Kip, K. E., Mulukutla, S. R., Aiyer, A. N., Johnson, C.,

McDowell, L. A., … Reis, S. E. (2013). Sociodemographic, clini-

cal, and psychological factors associated with attrition in a pro-

spective study of cardiovascular prevention: The Heart Strategies

Concentrating on Risk Evaluation study. Annals of Epidemiology,

23, 328–333. doi: 10.1016/j.annepidem.2013.02.007



http://www.cancer.org/acs/groups/content/@epidemiologysurveilance/documents/document/acspc-036845.pdf

http://www.cancer.org/acs/groups/content/@epidemiologysurveilance/documents/document/acspc-036845.pdf

Baquet, C. R., Commiskey, P., Daniel Mullins, C., & Mishra, S. I.

(2006). Recruitment and participation in clinical trials: Socio-

demographic, rural/urban, and health care access predictors.

Cancer Detection and Prevention, 30, 24–33. doi: 10.1016/j.

cdp.2005.12.001

Bolen, S., Tilburt, J., Baffi, C., Gary, T. L., Powe, N., Howerton, M.,

… Bass, E. (2006). Defining “success” in recruitment of under-

represented populations to cancer clinical trials: Moving toward a

more consistent approach. Cancer, 106, 1197–1204. doi:

10.1002/cncr.21745

Chatfield, M. D., Brayne, C. E., & Matthews, F. E. (2005). A system-

atic literature review of attrition between waves in longitudinal

studies in the elderly shows a consistent pattern of dropout

between differing studies. Journal of Clinical Epidemiology, 58,

13–19. doi: 10.1016/j.jclinepi.2004.05.006

Clough-Gorr, K. M., Fink, A. K., & Silliman, R. A. (2008). Challenges

associated with longitudinal survivorship research: Attrition and a

novel approach of re-enrollment in a 6-year follow-up study of

older breast cancer survivors. Journal of Cancer Survivorship, 2,

95–103. doi: 10.1007/s11764-008-0049-y

Cook, T. D., & Campbell, D. T. (1979). The conduct of randomized

experiments. In T. D. Cook & D. T. Campbell (Eds.), Quasi-exper-

imentation: Design and analysis issues for field settings. (pp.

341–386). Boston, MA: Houghton-Mifflin.

de Graaf, R., van Dorsselaer, S., Tuithof, M., & Ten Have, M. (2013).

Sociodemographic and psychiatric predictors of attrition in a pro-

spective psychiatric epidemiological study among the general

population. Result of the Netherlands Mental Health Survey and

Incidence Study-2. Comprehensive Psychiatry, 54, 1131–1139.

doi: 10.1016/j.comppsych.2013.05.012

Deeg, D. J. H. (2002). Attrition in longitudinal population studies:

Does it affect the generalizability of the findings? An introduction

to the series. Journal of Clinical Epidemiology, 55, 213–215. doi:

10.1016/S0895-4356(01)00472-3, pii: S0895-4356(01)00472-3.

Demark-Wahnefried, W., Bowen, D. J., Jabson, J. M., & Paskett,

E. D. (2011). Scientific bias arising from sampling, selective

recruitment, and attrition: The case for improved reporting. Can-

cer Epidemiology, Biomarkers, & Prevention, 20, 415–418. doi:

10.1158/1055-9965.EPI-10-1169

Everitt, B. H. T. (2009). A handbook of statistical analyses using R.

(2nd ed.). Boca Raton, FL: Chapman & Hall/CRC.

Fayter, D., McDaid, C., & Eastwood, A. (2007). A systematic review

highlights threats to validity in studies of barriers to cancer trial

participation. Journal of Clinical Epidemiology, 60, 990–1001.

doi: 10.1016/j.jclinepi.2006.12.013

Florida Department of Health, (2007). Vital Statistics Annual and Pro-

visional Report: Population. Retrieved October 15, 2012 from

http://www.flpublichealth.com/VSBOOK/VSBOOK.aspx.

Haring, R., Alte, D., Volzke, H., Sauer, S., Wallaschofski, H., John,

U., & Schmidt, C. O. (2009). Extended recruitment efforts mini-

mize attrition but not necessarily bias. Journal of Clinical Epidemi-

ology, 62, 252–260. doi: 10.1016/j.jclinepi.2008.06.010

Hothorn, T., Hornik, K., & Zeileiswere, A. (2006). Unbiased recursive

partitioning: A conditional inference framework. Journal of

Computational and Graphical Statistics, 15, 651–674.

Lee, C. S., Hayes, R. B., McQuaid, E. L., & Borrelli, B. (2010). Pre-

dictors of retention in smoking cessation treatment among Latino

smokers in the Northeast United States. Health Education

Research, 25, 687–697. doi: 10.1093/her/cyq010

Marcellus, L. (2004). Are we missing anything? Pursuing research

on attrition. Canadian Journal of Nursing Research, 36(3),

82–98.

Meneses, K., Azuero, A., Hassey, L., McNees, P., & Pisu, M. (2012).

Does economic burden influence quality of life in breast cancer

survivors? Gynecologic Oncology, 124, 437–443. doi: 10.1016/j.

ygyno.2011.11.038

Meneses, K. D., McNees, P., Loerzel, V. W., Su, X., Zhang, Y., &

Hassey, L. A. (2007). Transition from treatment to survivorship:

Effects of a psychoeducational intervention on quality of life in

breast cancer survivors. Oncology Nursing Forum, 34, 1007–

1016. doi: 334345R3225M1702, pii: 10.1188/07.ONF;1; 1007-

1016.

Meneses, K. M., (under review). Recruiting and enrolling rural breast

cancer survivors in cancer control research.

Meneses, K. M., Benz, R. L., Hassey, L. A., Yang, Z. Q., & McNees,

P. (2013). Strategies to retain rural breast cancer survivors in lon-

gitudinal research. Applied Nursing Research, 26, 257–262. doi:

10.1016/j.apnr.2013.08.001 [Epub ahead of print].

Paskett, E. D., Cooper, M. R., Stark, N., Ricketts, T. C., Tropman, S.,

Hatzell, T., … Atkins, J., (2002). Clinical trial enrollment of rural

patients with cancer. Cancer Practice, 10, 28–35.

Pisu, M., Azuero, A., Meneses, K., Burkhardt, J., & McNees, P.

(2011). Out of pocket cost comparison between Caucasian and

minority breast cancer survivors in the Breast Cancer Education

Intervention (BCEI). Breast Cancer Research and Treatment,

127, 521–529. doi: 10.1007/s10549-010-1225-0

Radloff, I. S. (1977). The CES-D Scale: A self-report depression

scale for research in the general population. Applied Psychologi-

cal Mesurement, 1, 385–401.

Ribisl, K. M., Walton, M. A., Mowbray, C. T., Luke, D. A., Davidson,

W. S., & Bootsmiller, B. J. (1996). Minimizing participant attrition

in panel studies through the use of effective retention and track-

ing strategies: Review and recommendations. Evaluation and

Program Planning, 19, 1–25. doi: 10.1016/0149-7189(95)00037-2

Robinson, K. A., Dennison, C. R., Wayman, D. M., Pronovost, P. J.,

& Needham, D. M. (2007). Systematic review identifies number of

strategies important for retaining study participants. Journal

of Clinical Epidemiology, 60, 757–765. doi: 10.1016/j.jclinepi.

2006.11.023

R-statistical-software. (2008). R statistical software v2.15.00. Vienna,

Austria: R Foundation for Statistical Computing.

Sherbourne, C. D., & Stewart, A. L. (1991). The MOS social support

survey. Social Science & Medicine, 32, 705–714.

Su, X., Meneses, K., McNees, P., & Johnson, W. (2011). Interaction

trees: Exploring the differential effects of an intervention pro-

gramme for breast cancer survivors. Journal of the Royal Statisti-

cal Society: Series C (Applied Statistics), 60, 457–474. doi:

10.1111/j.1467-9876.2010.00754.x

Van Beijsterveldt, C. E., van Boxtel, M. P., Bosma, H., Houx, P. J.,

Buntinx, F., & Jolles, J. (2002). Predictors of attrition in a longitu-

dinal cognitive aging study: The Maastricht Aging Study (MAAS).

Journal of Clinical Epidemiology, 55, 216–223.

Ware, J. E., Jr., & Sherbourne, C. D. (1992). The MOS 36-item short-

form health survey (SF-36). I. Conceptual framework and item

selection. Medical Care, 30, 473–483.

Whitten-Goldstein, K. (2003). Rural populations: Challenges in

recruitment, retention, and relevance. In Science meets reality:



http://www.flpublichealth.com/VSBOOK/VSBOOK.aspx

Recruitment and retention of women in clinical studies and the

critical role of relevance. (pp. 121–123). Bethesda, MD: NIH

Office of Research on Women’s Health.

Young, A. F., Powers, J. R., & Bell, S. L. (2006). Attrition in longitudi-

nal studies: Who do you lose? Australian and New Zealand Jour-

nal of Public Health, 30, 353–361.

Acknowledgments

Funded by National Cancer Institute, RO1CA-120638-07. The Florida cancer incidence data used in this report were collectedby the Florida Cancer Data System under contract with the Department of Health (DOH). The views expressed herein are solelythose of the author(s) and do not necessarily reflect those of the contractor or DOH. The authors are indebted to our breastcancer survivor participants and also thank research team members Silvia Gisiger-Camata, Madeline Harris, Lauren Hassey,Ann Wooten, and Ziqin Yang.



Documents

Predictors of Attrition Among Rural Breast Cancer Survivors