.l L ..:.
I 477111.37

I In? ‘ Z‘
1 ,

£13.. . .. 3
”5:53.: at:
5.2., .2“.

.
v

. 2.5.9..

~.T...r.

~ ‘ . 1.; n 1‘
,n2.- 1...:

.v. ,..H.1.N..x

 

 

 
 

4 .1 .l r
.. $I , ff... . t 5 . , . ., - .
_ . c .i . ., a. 1...; 5:. . , . av; i $.15? ‘ . .... I . , .. ‘ . . ‘ . . -
:1 . uf . . :7 ‘ ‘. . . 1-: 3.. . .7 . ‘ v . , '1 01nd. i
. . . a ‘ ‘ .. . u y . w 3.... .uuahb‘

 

W . 4 .

 

WSYS

-' If 1
10/ '1

This is to certify that the

dissertation entitled

Research Design and Subject Characteristics
Predicting Sample Attrition in a Panel Survey
of Older Families with Cancer: An Ecological
Model and Empirical Tests

presented by

Donna Edwards Neumark

has been accepted towards fulfillment ‘
of the requirements for

 

-Ph,D. mgmmhLEamil¥_and Child Ecology

Miafl '

Major professor

 

[mm June 14. 2002
54%02-

MS U is an Affirmative Action/ Equal Opportunity Institution 0-12771

 

 

LIBRARY

Michigan State

University

 

 

PLACE IN RETURN Box to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

59318 % 9&4

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 c:/CIRC/DateOue.p65-p.15

 

RESEARCH DESIGN AND SUBJECT CHARACTERISTICS PREDICTING SAMPLE
AT'I'RITION IN A PANEL SURVEY OF OLDER FAMILIES WITH CANCER:
AN ECOLOGICAL MODEL AND EMPIRICAL TESTS
By

Donna Edwards Neumark

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Department of Family and Child Ecology

2002

 

ABSTRACT

RESEARCH DESIGN AND SUBJECT CHARACTERISTICS REDICTING SAMPLE
ATTRITION IN A PANEL SURVEY OF OLDER FAMILIES WITH CANCER:
AN ECOLOGICAL MODEL AND EMPIRICAL TESTS
By

Donna Edwards Neumark

A common problem in the conduct of health care research with families is sample
attrition. Researchers describe racial, cultural, gender, functional, psychological, and
age-related variables associated with higher attrition in surveys, cancer prevention trials,
and intervention research. Others suggest that the research design has substantial effects
on subjects’ attrition and retention over time. Staring strategies is imperative, however,
comprehensive frameworks which encompass subject, societal, and research design
characteristics have not been systematically applied to panel health care surveys.

This dissertation achieves two purposes. The first is the development of the
Ecological Model of Sample Attrition (EMSA). The model conceptualizes the
interactions of subject characteristics, research design features, and the societal, family,
and health care enviromnents that influence whether subjects participate or do not
participate in family health care panel surveys, and whether participating subjects drop
out over time. The second purpose is to test the ecological model. Statistical techniques
are used to examine the relationships between research design and subject predictors on
the outcome of attrition from a sample of older individuals with cancer and their families.
Data are from the Family Care Study, a survey of elderly cancer patients and family

caregivers, interviewed by telephone 4 times in the first year after cancer diagnosis.

Two secondary analyses are discussed. In the first, demographic and cancer
variables and features of the study protocol are used as predictors in a multinomial logit
regression model to Compare those who declined participation (n=748), consenters who
dropped out before the first interview (n=208), and consenters who participated in the
interview (n=992). Among subject characteristics, age and cancer diagnosis played major
roles in whether consent was obtained. Race did not influence consent but increased odds
of early drop out. Study design features, such as if caregivers participated, recruitment
stafl roles, and whether recruiters were paid, strongly affected the probability of subjects
declining participation or dropping out prior to the interview. In the second analysis,
similar subject and research design variables, as well as data fiom patient interviews, are
included in pooled time-series cross-section models to compare those who died (n=133)
or those who dropped out (n=168), to those who remained on the study for one year
(n=704). The analysis helped distinguish that when attrition was due to reasons other than
death, research design characteristics were more likely to be associated with the attrition.
Thus, the effects of the research design characteristics diminished, mainly in the later
waves of the study, when odds of dropping out for reasons other tlmn death decreased.

Sample attrition is inevitable in studies about families with cancer. Although
having little control over health or age-related reasons contributing to attrition,
researchers can design studies flexible to the needs of ill participants, families, and
project personnel, particularly during accrual and the early phases of data collection.
Studies investigating sample attrition are needed. Predicting who is prone to decline
participation or to drop out, and why, may help researchers target efforts and reduce the

extent to which study-related factors contribute to attrition.

Copyright by
DONNA EDWARDS NEUMARK

2002

Dedicated to the women and the dogs at Fort Funston

ACKNOWLEDGEMENTS

My dissertation topic was not typical for the department of Family and Child
Ecology, yet the flexible program and the latitude of my committee allowed me to study
what I set out to do. My advisor and dissertation chair, Dr. Barbara Ames, helped me
expand my perspectives especially about family systems. I thank her for pushing me to
inject my own voice in my writing, and for all the editorial suggestions, especially during
cnmch times. The direction of my dissertation took a theoretical turn during the third
year thanks to Dr. Norm Bobbitt. Although the process was sometimes exasperating as I
struggled to articulate the model, I am pleased with the outcome. I appreciate the time of
my other committee members, Dr. Robert Grifl‘ore and Dr. Marsha Carolan in PCB, and
Dr. Barbara Given and Dr. Manfred Stommel in the College of Nursing. A special thanks
to Dr. Stommel for sticking with the attrition project since its inception, and for helping
with the statistical analyses. It was a pleasure to work with him and to write with him

The data for this dissertation are from the Family Care Study at Michigan State
University (NINR/NCI Grant #2R01 NR/CA l995—03A3). I thank Dr. Barbara Given and
Dr. Bill Given for sharing their expertise in managing a large research study. Although
my doctorate is in Family and Child Ecology, and my work reflects a family and
ecological perspective, my original flame of reference is oncology nursing. Through my
years at MSU, I valued many discussions with Dr. Barbara Given about cancer nursing
issues. Moreover, I am grateful I had the chance to learn and work with Dr. Ruth
McCorkle at the University of Pennsylvania, before I came to MSU. Dr. McCorkle had
highexpectations ofher students, herresearchstafi‘, and ofherself, whileatthe same
time she wholeheartedly supported my situation as a working mother with two small

children. In my last assignment for Dr. McCorkle, I wrote a professional philosophy

paper. In April, 1993, I wrote, “I recognize that the ideal is never the reality. We work
within systems of individuals, families, communities, and governments which ofien .
impose constraints and require negotiation and compromise.” In reading this passage, I
realize that my journey from advanced nursing practice in oncology, to gaining an
understanding of ecosystems and family systems, has come full circle. In recent years,
through my research and through personal experiences, I have learned even more how
cancer affects the family system. Thus I am so grateful that my mother, Esther Edwards,
was able to come to Michigan to see me get my PhD. I also amthankful for my dad’s
encouragement. He often asks about my work, and willingly agreed to be the subject of
my PCB 801 paper and ecomaps.

My graduate work was eased by generous financial support. I am gratefiil to
Michigan State University for the University Distinguished Fellowship, and the
dissertation completion fellowship, and to the Department of Family and Child Ecology
for research and teaching assistantships. I enjoyed my work in West Fee with Dr. Esther
Onaga and Helen Irvine. I also appreciate funding fi'om the Oncology Nursing
Foundation, the Nurses’ Educational Fund, and particularly the American Cancer Society
(DSCN 98-331-01; DSCN 98-331-03).

My experiences as a graduate student would have been greatly diminished if I did
not meet Barb Gormley in the first semester of classes. I am so glad our life paths are
keeping us near each other for now. I also welcomed the support fi'om other people
going through graduate school, especially Julie Haddow, Jessica Kovan, Annie Farrell,
and Susan Hoppough Figuring out how to balance everything was a common endeavor.
Finally, I want to thank my husband David, and my children, Noey and Eitan. You guys

keep me laughing. You are the center of my life.

TABLE OF CONTENTS page I

LIST OF TABLES ......................................................................
LIST OF FIGURES .....................................................................

INTRODUCTION ......................................................................

Background and Scope of Problem .............................................
Purposes and Research Questions ..............................................
Overview of Ecological Model of Sample Attrition .........................
Relevance to Family Health Care Research ...................................
Clarifying the Language of Sample Attrition ..............................
Methodological Considerations in Family Research ......................
Support for an Ecological Perspective of Inquiry .........................

Scope of Dissertation and Definition of Terms ...............................
A Blending of Paradigms .........................................................

EXISTING FINDINGS AND MODELS OF SAMPLE ATTRITION ..........
Consequences of Attrition in Family Studies .................................
Phases of Sample Attrition ......................................................
Methods to Assess and Report Attrition .......................................
Existing Models Related to Sample Attrition .................................

Theories of Survey Response Behavior .....................................
Access and Recruitment in Health Care Studies ...........................
Participation in Clinical Trials ................................................
Summary ...........................................................................

CONCEPTUALIZING AN ECOLOGICAL MODEL .............................
Overview of Model ...............................................................
Ecological Characteristics of the Model .......................................
Time in the Model ................................................................

Components of the Model .......................................................

Subject Characteristics and Research Design Characteristics Affecting
Sample Attrition ..................................................................

Subject Characteristics ......................................................
Research Design Characteristics ..........................................

Pages

14-29
14
1 8
20

24
24

24
25
28

29-60
29
3 1
33
34

35
36

4O

TABLE OF CONTENTS page 2 Pages

The Environments ................................................................ 48
Social Environment (SE) ................................................... 49
Health Care Environment (HCE) .......................................... 51
Family Environment (FE) .................................................. 54

Propositions fiom the Model .................................................... 58

Support for an Ecological Perspective of Inquiry ............................ 59

METHODS .............................................................................. 61-69

Ex Post Facto Case Study of Sample Attrition ............................... 61

Research Questions ............................................................... 62

Describing the Family Care Study .............................................. 63

Attrition Analyses ................................................................. 69

EARLY ATTRITION ANALYSIS: NONPARTICIPATION AND EARLY

DROPOUT ............................................................................... 7048
Research Question ................................................................ 7O
Defining the Target Sample ..................................................... 70
Variables for Early Attrition Analysis .......................................... 71

Variables Pertaining to Subject Characteristics ............................ 71
Variables Pertaining to Research Design Characteristics ................. 72
Statistical Analysis of Early Attrition .......................................... 75
Results of Early Attrition Analysis ............................................. 76
Brief Discussion ................................................................... 78

ATTRITION OVER TIME: DROPOUT OVER FOUR WAVES OF THE 79-98

FAMILY CARE STUDY. ............................................................
Defining the Target Sample ...................................................... 79
Statistical Analysis of Attrition Over Time .................................... 80
Variables for Attrition Over Time Analysis ................................... 82

Variables Pertaining to Subject Characteristics ............................ 83
Variables Pertaining to Research Design Characteristics ................. 86
Results of Attrition Over Time Analysis ....................................... 89
Bivariate Analysis. ............................................................ 90
Time-Series Cross-Section Analysis of Attrition Over Time ............ 91

TABLE OF CONTENTS page 3 Pages

Comparing Attrition Outcomes: Attrition for Other Reasons Versus

Death Attrition. ................................................................... 94
Attrition for Reasons Other than Death ..................................... 96
Attrition Due to Death ......................................................... 97

Brief Discussion .................................................................. 97

CONCLUSIONS AND IMPLICATIONS .......................................... 99417

Patterns and Predictors of Attrition in the Family Care Study .............. 99

Subject and Research Design Characteristics and Sample Attrition in

the PCS ............................................................................ 100

Data fiom the Field Notes ....................................................... 103

Limitations of the Study ......................................................... 109

Implications of this Dissertation ................................................ 11o
Enhancing Understanding of Sample Attrition ............................. 110
The Language of Sample Attrition .......................................... 111
Incorporating a Family Perspective ......................................... 1 13
Ecological Characteristics of Sample Attrition ............................ 114

Conclusions ........................................................................ 115

APPENDIX A: Tables ................................................................. 118-133
APPENDIX B: Figures ............................................................... 134-139
REFERENCES ........................................................................... 140-151

Table 1

Table 2

Table 3

Table 4

Table 5

Table 6

Table 7
Table 8

Table 9

Table 10

Table 11

LIST OF TABLES

Early Attrition Analysis: Variables Available for Target
Sample of 1948 Cases ...................................................

Multinomial Logistic Regression Comparing Participants to
Early Dropouts to Nonconsenters .....................................

Attrition Over Time in the Family Care Study: Overall Status
of Cases ...................................................................

Bivariate Analyses of Time-Independent Subject
Characteristics and Attrition Over Time .............................

Bivariate Analyses of Time-Dependent Subject Characteristics
and Attrition Over Time ................................................

Bivariate Analyses of Research Design Characteristics
and Attrition Over Time ................................................

TSCS Models of All Attrition in the Family Care Study ..........
Interaction Model: Primary Cancer Diagnosis x Wave Time. .

Attrition Over Time: Comparing Attrition Outcomes for Target
Sample of 1005 Cases ...................................................

TSCS Models of Attrition Due to Causes Other Than Death. .

TSCS Models of Death Attrition in the Family Care Study .......

Pages

119

120

121

122-123

124

125
126-127

128

129
130-131

132-133

Figure 1
Figure 2
Figure 3
Figure 4

Figures

LIST OF FIGURES

Ecological Model of Sample Attrition in Family Health Care Panel
Studies ...........................................................................

Subject Characteristics and Research Design Characteristics
Predicting Attrition in Family Health Care Panel Studies ...............

The Environments .............................................................

Operationalization of the Independent Variables in the Family Care
Study in Relationship to the Theoretical Components of EMSA .......

Cancer Diagnosis * WaveTime Interaction Model .......................

Pages

135

136

137

138

139

INTRODUCTION
Background and Scope of Problem

Longitudinal research designs are necessary in family health care studies when
investigators are concerned with long-term outcomes of an illness or its treatment on
individuals and families. Panel surveys in particular allow for observation of the natural
history of the process being studied. They help identify and delineate factors which my
predict or mediate outcomes, and are useful for the analysis of change over time
(Johnson, 1988; Miller & Wright, 1995; Ribisl et al., 1996; Ryan & Hayman, 1996).
However, despite these advantages, methodological problems ranging from sample
recruitment and data collection, to the analysis of repeated measures and the
generalizability of the results are prevalent in studies that depend on repeat data
collection on the same subjects at regular intervals over time. A particularly important
challenge is sample attrition (Arean & Gallagher-Thompson, 1996; Coen, Patrick, &
Shem, 1996; Given, Keilrnan, Collins & Given, 1990; Motzer, Moseley & Lewis,l997;
Zapka , Chasan-Taber, Bigelow & Hurley, 1994). Although attrition rates are widely
variable, levels of 15% to 40% are not uncommon, and participant loss as high as 70%
has been reported (Moser, Dracup, & Doering, 2000). The success of panel research
depends in part on subjects consenting to participate and making an ongoing commitment
to remain in the study (Clinton et al., 1986; Shanna, Tobin, & Brant, 1989).

Anecdotal examples of efforts to reduce sample attrition and improve access,
recruitment, and retention in panel studies of families are prevalent. Often, these reports
highlight the obstacles and suggest strategies to enroll and retain families based on

subject characteristics. Researchers recount racial, cultural, linguistic, socioeconomic,

gender, functional, psychological and age-related attributes that are often associated with
lower participation rates and higher attrition. Barriers to participation persist particularly
among minority populations, older persons, and those with lower education and lower
socioeconomic status (Martin, 1995; Swanson & Ward, 1995). Likewise, subjects with
poor social support and weaker social networks are more likely to drop out over time
(Norris, 1987; Shaw, Cronan, & Christie, 1994). Typically, the ability to maintain high
response rates is reduced with greater fimctional impairment of subjects, greater intensity
of health service needs, poorer physical health, and poorer self-reported health, all of
which grow worse with age (Corder & Manton, 1991; Thompson, Heller & Rody, 1994).
In cancer studies, an unavoidable source of attrition is the morbidity and mortality related
to the illness itself (Davies, Reimer, Brown & Martens, 1995; Mock, Hill, Dienemann,
Grimm, & Shivnan, 1996; Rinck, van den Bos, de Haes, Schadé, & Veenhof, 1997).

In addition to the subject characteristics, research design features may have
efl‘ects on sample attrition. Subjects are influenced by how much time is involved, the
predominant mode of data collection, the number of family members required, the ease of
participation, and interactions with study personnel (Brehrn, 1993; Groves, Cialdini, &
Couper, 1992; Mihelic & Crimmins, 1997). In health care studies, varying levels of
access and recruitment support at clinical sites, or competition with other research studies
in a particular setting, challenge investigators to overcome recruitment barriers (Eaves,
1999; Neurnark, Stommel, Given, & Given, 1998; Motzer et al., 1997; Ryan & Hayman,
1996).

Bischofl‘ and Sprenkle (1993) identify the main drawback in the literature as

inconsistent operational definitions of attrition, which limits the comparability of results

across studies. For example, although refusal and recruitment rates play important roles
in evaluating data, most family health studies overlook or ignore these issues (Holder et
al., 1998). In an analysis of published empirical studies examining the impact of adult
cancer diagnosis on family functioning, Cooley and Moriarty (1997) reported that just
over halfof the studies state response rates, but only 38% of the studies include refirsal
rates. Similarly, in a review of published reports of longitudinal investigations of
caregiving in families with cancer, the majority (86%) report on response rates and
attrition over time. However, only 4 of the 14 reviewed articles (29%) included
comments about nonparticipants or gave information about how many people were
identified as eligible to participate compared to how many actually initiated the first
phase of data collection (Neumark, 2000).

Another salient issue is that multiple terms are used by researchers and
methodologists to communicate about sample attrition. For example, subjects who begin
a panel study but do not complete it may be called attritors, dropouts, or nonrespondents;
others describe refusal, nonresponse, or noncompliance (Ribisl et al., 1996).
Alternatively, the occurrence of attrition is captured indirectly through the description of
related phenomena such as subject access, recruitment, participation, retention, or study
completion. In this dissertation, sample attrition is the general term referring to all
subjects who were identified, found to be eligible, but who do not complete the final
observation in a panel survey, for whatever reason. Therefore, sample attrition includes
specific subsets of the eligible and accessrble sample, including individuals who initially
refuse participation, who fail to reply to interview requests at any time, who are lost to

follow up, or who have died (Ribisl et a1).

Purposes and Research Questions

There are two overall purposes of this dissertation. The first is to develop the
Ecological Model of Sample Attrition (EMSA) in Family Health Care Panel Studies
(Figure 1). The model proposes that subjects and researchers are situated in the family
environment, the health care environment, and the social environment; simultaneously,
the research design and subject characteristics contribute independemly and
interdependently to influence whether subjects participate or do not participate in a panel
survey (early attrition), and if they do participate, what may influence dropout over time
(attrition over time).

The second purpose of the dissertation is to utilize the Ecological Model of
Sample Attrition as a conceptual guide in order to operationalize and to armlyze
statistically the relationships between research design and subject characteristics
(independent variables) on the outcome of sample attrition fiom a target sample of older
individuals with cancer and their families Data are fiom the Family Care Study (FCS), a
descriptive panel survey of elderly cancer patients, with or without family caregivers,
who were interviewed by telephone four times in the first year following cancer diagnosis
(Given & Given, 1991-1996).

Several research questions emerge from the Ecological Model. Specifically, this
dissertation seeks to answer the following: 1) Do research design or subject
characteristics predict who was more likely to refuse participation or to drop out prior to
the initiation of a panel survey of elderly cancer patients and family caregivers?; 2) Do
research design or subject characteristics predict who was more likely to drop out over

time fi'om a panel survey of elderly cancer patients and Emily caregivers?; and 3) Do

difl°erent characteristics predict who drops out due to death versus who drops out for
reasons other than death?

Two separate statistical analyses were completed to answer these questions. The
first analysis explored the simultaneous effects of demographic and cancer
characteristics, as well as features of the study protocol, on early attrition (question 1). A
multinomial logit regression model compared subjects eligible for the FCS who did not
consent to participate (nonparticipants), to consenters who dropped out prior to initial
data collection (early dropouts), to consenters who participated in the intake interview
(participants). The second analysis compared those who initiated the study but dropped
out or died, to those who remained on the study for one year following cancer diagnosis
(questions 2 and 3). Sociodemographic, psychosocial, and cancer-related predictor
variables representing subject characteristics, as well as features of the research design
were included in pooled time-series cross-section models to reveal patterns of attrition
over time in this sample of elderly cancer patients and their Emily caregivers.

Overview of Ecological Model of Sample Attrition

Based in systems thinking, an ecological perspective highlights the
interrelationships among and the connectedness of organisms and their contexts (Bristor,
1990; Gfifiore & Phenice, 2001; Wideen, Mayer-Smith, & Moon, 1998). It is
particularly useful when the problem under study is complex in order to capture the
multidimensional character of a phenomenon. An ecological model is applicable to the
study of sample attrition in the attempt to untangle the complexities of this phenomena.

The Ecological Model of Sample Attrition in Family Health Care Panel Studies

(Figure 1) depicts an organizational structure made up of interdependent, interrelated

parts. EMSA is a holistic, non-linear representation of the dynamic interactions and the
connectedness between subjects, researchers, and their environments; and sample
attrition is conceptualized as multileveled and subject to numerous influences over time,
both within and outside of the research study. The ecological perspective is represented
by the shape of the model and the nature of the relationships between the concepts and
the variables. The nested model theorizes a multilayered situation incorporating the fact
that individuals and Ernilies and researchers operate simultaneously in multiple contexts
(Benjamin & Hollings, 1997).

The important features of the model are the three environments, the subject
characteristics, the research design clmracteristics, the outcome of attrition, and the notion
of tinre. The model proposes that subject characteristics and the research design are
situated in the Emily environment, the health care environment, and the social
environment. Simultaneously, the research design and subject characteristics
mdependently and interdependently influence whether subjects participate or do not
participate in a panel survey (early attrition), and if they do participate, what may
influence dropout over time (attrition over time). The model in Figure l is dissected into
Figure 2 and Figure 3. Figure 2 illustrates the subject and research design components of
EMSA and Figure 3 illustrates the components of the environments that interact to
influence whether attrition from the target sample occurs. These figures, although
presented separately, should be viewed concurrently as they represent the multiple layers
underlying subject attrition.

A key aspect of EMSA is its dynamic nature. Although diflicult to capture

graphically, this attribute allows the elements of the model to change in shape and size

according to their relative importance over time (Hancock, 1993). Inherent in this idea is

that the relationships are complex and multidirectional. This is depicted in the model's

shape, its multiple layers, the overlapping and intersecting environments, the dashed lines

representing permeable boundaries, and the interaction arrow between the subject

characteristics and the research design. These features are expanded in Chapter 3.
Relevance to Family Health Care Research

Clarifying the Language of Sample Attrition

This dissertation seeks to specify and clarify terminology with respect to sample
attrition in all phases of Emily health care panel research. In order to compare results
across studies, it is imperative to reduce variability in the methods used for calculating
attrition rates, and to be consistent in descrrlring and reporting sample attrition in panel
studies (Bischofi‘ & Sprenkle, 1993). Typically, when researchers report on sample
attrition, they compare subject characteristics of the sample at intake to those of the
sample of participants who remain in the study until the last observation of interest. It is
less typical to extend the analysis bacde to the original target sample (Carter, Elward,
Mahngren, Martin, & Carson, 1991; Kamey et al., 1995; Thompson, Heller, & Rody,
1994).

In the model developed for this dissertation, “target sample” is defined as all
potential subjects who were found to be eligible for a study, and who were approached to
participate. Sample attrition occurs either when potential subjects did not participate, or
when subjects failed to complete a study alter initiating data collection (Brehrn, 1993;
Fogg & Gross, 2000; Given et al., 1990). In particular, sample attrition tends to be large

in the beginning (Mihelic & Crimmins, 1997). Thus, an important consideration in an

attrition analysis is whether people who consent to participate are inherently different
fi'om those who do not. Unfortunately, it is the nonconsenting subjects about whom little
is known because they did not participate in any extensive data collection efl'orts (Brehm,
1993). Nonetheless, a thorough understanding of sample attrition in panel studies must
begin the all potential subjects who were found to be eligible for a study and who were
approached to participate. (Brehnr; Groves et al.,1992; Mihelic & Crirmnins).
Methodological Considerations in Family Research

There are discrepancies in the literature about whether the sample unit of
observation is related to attrition, and specifically whether or not having multiple Emily
members involved as subjects is beneficial (Bonvicini, 1998; Eaves, 1999; Killian &
Newton, 1993; Ryan & Hayman, 1996; Thornton, Freedman, & Camburn, 1992).
Furthermore, studies of families are complicated because of the complex nature of the
relationships within Emilies. Attrition may be influenced by Emily relationships, Emily
composition, and Emily dynamics (Bray, Maxwell, & Cole, 1995; Holder et al., 1998).
Yet existing models related to health behavior typically focus on the individual (Padula,
1997). The interaction with the Emily that is central to understanding sample attrition in
studies of Emilies is missing.

Ultimately, the purpose of the research should guide the investigator in
determining whether data fiom individual Emily members rather tlmn data from multiple
sources are sufiicient. Family researchers must decide which parts or aspects of the
Emily need to be evaluated and need to determine the sampling unit, recruitment and
retention strategies,as well as appropriate analytical techniques (Bray, 1995; Eaves,

1999; Maguire, 1999). In addition to incorporating a family perspective into the study of

sample attrition, the'ecological model can be applied to studies involving either
individual subjects or multiple Emily members in panel data collection.
Support for an Ecological Perspective

Researchers agree that inadequate recruitment and drop out over time affects the
implementation, outcomes, and the conclusions generated from panel research on
individuals and Emilies. Clearly, sharing strategies to prevent and reduce sample
attrition in similar studies is essential; however, this universal problem is seldom placed
in a conceptual framework. Developing such a fi'amework Ecilitates collection of data on
variables known to afl‘ect sample attrition, and may encourage Emily researchers to
design studies in light of anticipated attrition (Goodman & Blum, 1996). Tested theories
of nonresponse are found in research about public opinion polls and consumer marketing
surveys (Brehm, 1993; Evangelista, Albaum, & Poon, 1999; Groves et al., 1992);
however, comprehensive fiameworks which encompass subject, societal, and research
design characteristics have not been applied to Emily health care studies.

Some models have been used to explain various types of health behavior, and a
few have been empirically tested to explain why people would or would not use health
services or participate in prevention, screening, or intervention programs. While these
models have merit, they are inadequate in their application to understanding sample
attrition in descriptive panel survey research of Emilies with chronic illness. First, the
models are most often applied to predict preventive health behavior and adherence with
medical regimens. The emphasis on “compliance” and “adherence” implies that the
impetus to participate depends solely on subject behavior, but seems to neglect that both

subject characteristics and research design may simultaneously influence these outcomes.

Second, the models typically focus on the individual. The interaction with the Emily that
is central to understanding participation in panel studies of families is missing. Finally,
although some theoretical and operational evidence for this dissertation emerges fiom
reports of successful strategies for recruitment and retention in health promotion and
intervention studies, many issues associated with participation in clinical trials and other
studies with experimental designs are difi‘erent from those with descriptive designs.

Progress in research surrounding sample attrition warrants an ecological
perspective of inquiry. No existing models sufliciently capture the complex, dynamic,
and interactive nature of sample attrition in longitudinal Emily health care studies.
Creating the Ecological Model of Sample Attrition enhances understanding of this
ubiquitous problem, and provides conceptual guidance for multivariate analysis to predict
early attrition and attrition over time in panel survey research of individuals with chronic
illness and their Emilies.

Scope of Dissertation and Definition of Terms

The Ecological Model of Sample Attrition in Family Health Care Panel Studies
conceptualized in this dissertation pertains to panel surveys of older individuals with
cancer and their families. A survey is an observational research technique, typically used
for descriptive, explanatory, and exploratory purposes (Babbie, 1998), and a panel study
is an inquiry that depends on repeat data collection on the sample subjects over time
(Ribisl et al., 1996). EMSA addresses sample attrition of older adults fi’om studies that
access and recruit subjects within the health care system. The model is applicable to
studies in which the subject unit involves either individuals or two or more Emily

members in panel data collection, and in which the focus of the study is on health care

10

issues, particularly related to outcomes of chronic illness or its treatment on individuals
and Emilies.

The target sample is a useful starting point for empirical exploration of the
phenomenon of subject attrition since sample attrition is a subset of the overall target
sample. To firrther clarify, early attrition is defined as the nonparticipation that occurs
between the verification of eligibility of potential subjects and the initiation of panel data
collection. Nonparticipation includes nonconsenters and early dropouts. Nonconsenters
are defined as eligible subjects, who are approached to participate in a panel study, but
who decline participation. Early dropouts refer to eligible subjects who initially give
consent but then decline participation, who die, or who are lost to follow-up prior to
initiation of the intake interview. Attrition overtime, is defined as dropout, for whatever
reason, fiom the sample who initiated data collection. It includes subjects who dropped
out due to death, as well as those who leave the study for any other reason other than
death. In Figure l and Figure 2, the shapes representing target sample and sample
attrition widen to capture the notion of the growing sample size and the accumulation of
subjects who do not participate or who drop out over the course of a panel study.

Theoretical and operational evidence for the model was gathered fi'om a wide
array of literature and is presented in the next two chapters of the dissertation. Chapter 2
examines existing models related to sample attrition, reviews the consequences and
phases of attrition in panel studies, and highlights methods of assessing and reporting
attrition. Chapter 3 reviews causes of sample attrition and further explicates the
ecological components of EMSA which were introduced above. Then the chapter

expands the discussion of subject characteristics and research design characteristics

11

associated with sample attrition, and finally, the three environments are described.
Chapters 4 through 6 in the dissertation present the methods and results of the analyses of
early attrition and attrition over time in the Family Care Study. Finally, although the
statistical analyses are separate, the discussion and implications of the study are
combined in the concluding chapter. The tables and the figures are all found in the
appendices.
A Blending of Paradigms

This dissertation is an ex post facto case study, based on the Family Care Study
(Given & Given, 1991-1996). The major rationale for a case study research strategy is
when the investigation must cover both a particular phenomenon and the context within
which the phenomenon is occurring, either because (a) the context is hypothesized to
contain important explanatory variables about the phenomenon or (b) the boundaries
between the phenomenon and context are not clearly evident (Yin, 1994). The primary
characteristic of an ex post Ecto study is that the variables are studied as they occur in a
natural setting, because they have already occurred (Babbie, 1998). Therefore the
variables cannot be manipulated, and the analysis is limited to data collected during
sample accrual and recruitment, or from data collection via interviews, questionnaires,
and medical record audits. One common concern of case studies is that they provide little
basis for scientific generalization. Yet, according to Yin (1994), the case study allows an
investigation to retain the holistic and meaningful characteristics of real-life events and is
suitable to pose questions such as what are the relationships and the possible effects

among the variables.

12

A case study can contribute to our knowledge of individual, organizational, social,
and political phenomena, particularly'when the phenomena is complex. It is appropriate
to include quantitative evidence in a case study, and to rely on multiple sources of
evidence, with data needing to converge. Furthermore, a case study benefits from
development of theoretical propositions to guide data collection and analysis and is usefirl
to expand and generalize these propositions. This research strategy is appropriate when
trying to capture relationships and the possible efi‘ects among the variables, and not just
wanting to describe a situation (Creswell, 1994; Yin, 1994). This case study achieves
these goals.

Both interpretive and empirical, this dissertation requires a blending of paradigms
and modes of inquiry. At the conceptual level, the inquiry involves the development of
the Ecological Model of Sample Attrition in order to illuminate the multidimensional and
interactive Ectors associated with sample attrition in descriptive panel survey research of
Emilies with chronic illness, particularly cancer in older adult Emily members. As
recruitment coordinator of the F CS (1994-1997), I was closely involved in ongoing field
procedures associated with recruitment, retention, and attrition of subjects fiom over 60
clinical sites around Michigan and Indiana (N eumark, Given, & Given, 1997; Neumark et
al., 1998). Clearly, the empirical testing and predictive nature of the models are suitable
to a quantitative paradigm. Yet, the Family Care Study was not designed as a study to
evahrate sample attrition. Furthermore, the emergence and inductive evolution of the
ecological model during the past several years, as well as my role as a researcher within
the research study, are within a qualitative paradigm (Creswell, 1994). Thus, reliability

and validity in this study cross paradigms since the variables are based upon my ability to

13

EXISTING FINDINGS AND MODELS OF SAMPLE ATTRTION

Evidence for this dissertation was gathered fiom literature focusing on research
methodology, survey response behavior models, health care recruitment models,
anecdotal recruitment and retention analyses, Emily and human ecology, Emily health
models, and an ecological health communication model. This chapter reviews the
consequences, the pluses, and the methods of describing and reporting sample attrition in
Emily panel studies. It also summarizes existing models and fiameworks which are
related to sample attrition.

Consequences of A ttrition in Family Panel Studies

It is generally recognized that differential selection into a sample and dropout
over time complicates data analysis and may contribute to bias in the results, and these
issues are commonly addressed in the design, implementation, and analysis of panel
studies. Methodologists highlight techniques of assessing and/or correcting for biases
due to attrition and describe statistical corrections when subjects or data are missing
(Goodman & Blum, 1996; Miller & Wright, 1997). In order to reveal evidence of bias,
investigators must examine and describe the patterns, degree, and causes of sample
attrition (Ribisl et al., 1996). Norris (1985, 1987) reported there was much greater loss of
sample representativeness over the course of a gerontological study when attrition
occurred because of poor health or when the participants could not be located to be
interviewed, in comparison to attrition due to nrore benign reasons, such as simple
unwillingness to participate.

Ribisl and others (1996) differentiate between attrition that occurs randomly and

that which is not randomly distributed. If subjects who are lost from the sample are not

14

 

comparable to those who remain (i.e. not random), the interpretation of study findings is
affected. There is greater potential for bias if the Ectors contributing to attrition result in
systematic differences such that persons with certain characteristics are difl‘erentially
selected out of the sample. For example, if attrition is related to morbidity in studies of
older Emilies with members with chronic illness, participating subjects systematically
present a healthier picture than what prevails in the underlying population (Norris, 1985).

On the other hand, sample attrition does not necessarily result in bias. In
evaluating a survey of community-dwelling elderly, Mihelic and Crirnmins (1997) found
that attrition did not significantly bias the outcomes of interest. Furthermore, in health
care panel studies, attrition is often due to subject death, which is a “legitimate” outcome
(Mihelic & Crirnrnins). Nonsurvivors do not affect panel representativeness in the same
way as those who drop out for other reasons, because if death was the only cause of
attrition in a representative sample, the sample would continue to be representative of the
sm'viving cohort (Norris, 1985). Thus, if attrition is solely due to disease—related
mortality, then the panel attrition does, indeed, reflect the natural history of the disease as
it unfolds in the target population, and bias is not evident.

In the context of panel surveys, bias due to sample attrition is often discussed in
terms of both external and internal validity. If, for example, a researcher starts with a
probability sample at intake, any nonrandom loss of subjects at subsequent data
collection waves will compromise external validity. Thus some researchers focus the
discussion of potential bias on how representative the obtained study sample is of the
defined target population (Braver & Bay, 1992; Corder & Manton, 1991; Groves et al.,

1992; Miller & Wright, 1995: Motzer et al., 1997). In an article addressing issues of

15

difference relative to sample selection, participation, and attrition in studies with elderly
Afi'ican American subjects, Dennis and Neese (2000) identified the final sample as both a
percentage of the projected sample and a percentage of the total possrhle population
available to be sampled.

One concern regarding family measurement deals with the question of sample
selection as a source of systematic bias and error that restricts generalizability of research
findings (Thomas, 1987). Ideally, the operational and practical considerations of
selecting the sample and establishing eligibility criteria will not diminish the
generalizability of the findings, thus compromising the external validity of the study.
However, while researchers attempt to recruit representative samples, in reality, those
who do participate my be so unlike those who did not that generalization of findings is
significantly diminished. Furthermore, eligibility depends on the sample universe for the
study, the criteria for eligibility, the goals of the study, and whether the subject has the
ability to answer the questions (Brehm, 1993). For example, some may be excluded if
they do not speak English or if they are not healthy enough to participate in an interview.
Finally, populations of Emilies experiencing similar health concerns may not in reality
be comparable because of differences in health care providers, regional differences in
access to health care or in health care practices, provider attitudes towards research, and
Emily willingness to participate in research (Thomas).

Researchers agree that loss of participants over time any introduce bias and
complicate statistical analysis of repeated measures; however, it is less often
acknowledged that the problem of sampling bias already starts with the initial

nonparticipation of eligible subjects (early attrition). Comparing the characteristics of

16

eligible individuals who refuse to participate with those who consent enables researchers
to gain information about the external validity of a study, yet the consequences of refusal
and the potential resulting biased data is seldom studied. Hinds and others (1995)
highlighted the efi‘ect of refusals in clinical research studies. The researchers
acknowledged that interpreting data and extrapolating from study findings depended on
the characteristics of the participants, the nonparticipants, and the dropouts, who they
defined as those who began but did not finish the study (Hinds, Quargnenti, & Madison,
1995).

In any case, concern with how representative a sample is may not always be
pertinent. Given the nature of subject recruitment in health care settings, probability
samples are rare to begin with, and generalizability is often an elusive target (Patrick,
Pruchno, & Rose, 1998). Typically, researchers use sampling techniques that are
appropriate to research questions and methodologies (Karney et al., 1995). Furthermore,
many of the issues in determining the choice of a particular sample are pragrmtic and
stem fiom the fact that the researcher functions within a broader social order (Sjoberg &
Nett, 1968). Convenience and access to subjects or data, and resources such as time,
money, and personnel rmy also influence the representativeness of a sample.

Recruitment into panel studies is premised on research designs that specify
criteria for entry and designate goals for numbers of subjects for each criterion or
combination of criteria, and it may not be possible to obtain the number of subjects
actually eligible for recruitment. When a specific clinical population is the focus of a

study, and inclusion criteria are used to assist in sampling, those who are both eligible

l7

and accessible to the researcher make up the target sample (Patrick et a1, 1998; Porter,
1999).

Still, even with nonprobablity samples, there is good reason to be attentive to bias.
Accordingly, often researchers emphasize internal validity and the extent to which
nonparticipation and dropout affect the coefficients that describe the nature of the
relationships among variables in the data (Braver & Bay, 1992, Johnson, 1995; Miller &
Wright, 1995; Norris, 1987; Thompson et al., 1994). As the extent of attrition increases
over time, the resulting sample of those who remain in the study may become
progressively less representative of the original sample (Hayslip, McCoy-Roberts, &
Pavur, 1998). While many investigators compare demographic characteristics, it is less
comrrron to compare retained and lost subjects with respect to the measured outcomes
under investigation (Janus & Goldberg, 1997). It is especially important for researchers
to interpret results carefully when sample attrition is related to the phenomenon being
studied.

Phases of Sample Attrition

Each study phase may make a unique contribution to sample attrition (Brehm,
1993; Hooks et al., 1988; Norris, 1985; Zapka et al., 1994). Sample attrition occurs when
potential subjects do not participate, or when subjects Eil to complete a study after
initiating data collection (Brehrn; Fogg & Gross, 2000; Given et al., 1990). Thus sample
attrition starts with the nonparticipation that occurs in the earliest phases of a panel study,
reflecting subjects who are lost to follow-up, or the weeding out of those who are not
interested (Mihelic & Crimmins, 1997). Others face greater obstacles to participation,

such as health problems or Emily members resistant to enrollment in a research study

18

(Motzer et al., 1997). In this dissertation, the target sample, defined as all potential
subjects who are found to be eligible and who are approached to participate in a panel
study, is the starting point for exploration of the phenomenon of subject attrition.

There is additional opportunity for sample attrition between the time of giving
consent and actual participation in the intake interview. In a discussion of participant
recruitment in research among vulnerable Emilies, Demi & Warren (1995) descn'bed the
phenomenon of subjects agreeing to participate in studies when approached, due to their
deference to authority figures, or sometimes, due to their strong need to talk to someone.
They found that although some people did not overtly refuse to participate, they were
subsequently unavailable for follow-up. In this dissertation, these subjects are designated
as early dropouts, referring to eligible subjects who initially give consent but then decline
participation, who die, or who are lost to follow-up prior to initiation of the intake
interview.

The final phase of sample attrition in panel studies is that which occurs after the
initiation of data collection. Dropout refers to the decision by some individuals to leave a
study once they have completed some part of the research (Fogg & Gross, 2000). In this
dissertation, attrition over time, is defined as dropout, for whatever reason, from the
sample who initiated data collection. It includes subjects who dropped out due to death,
as well as those who leave the study for any other reason other than death.

Most assessments of sample attrition in panel studies focused on the subjects who
drop out after the initial intake interview. In an analysis of published empirical studies
examining the impact of adult cancer diagnosis on Emily functioning, Cooley and

Moriarty (1997) reported that just over halfof the studies stated response rates, but only

19

38% of the studies included refusal rates. Similarly, in a review of published reports of
longitudinal investigations of caregiving in Emilies with cancer, the majority (86%)
reported on response rates and attrition over time. Minimally, numbers and percentages
compared the initial sample size to the sample that was available for analysis. Many
offered general reasons about why people dropped out, and a few gave details of how the
missing data were accounted for in statistical amlysis. However, only 4 of the 14
reviewed articles (29%) included comments about nonparticipants or gave information
about how many people were identified compared to how many actually initiated the first
phase of data collection. Two publications presented no data about attrition, although the
researchers emphasized that the samples were small convenience samples and that
cautious conclusions were warranted (Neumark, 2000).
Methods to Assess and Report Attrition

Ribisl and others (1996) reviewed methods of describing and reporting attrition in
panel studies and found multiple terms to describe individuals who begin a study but do
not complete it, and great variability in the reported rates of participation in longitudinal
studies. Their excellent review article ofl‘ered multiple suggestions for describing,
tracking, and reporting attrition in panel studies. Another problem is that there are
inconsistencies regarding definitions of attrition and whether there are differences
between those who never begin a study and those who do and then drop out.

Some researchers studied the refusal component of nonresponse in household
marketing surveys and public opinion polls, and emphasized the importance of
recognizing the observed influences of sociodemographic and survey design Ectors, with

the less observable impact of the psychological components of the relatively brief

20

interaction between interviewer and participant (Groves et al., 1992). In addition to
improving consensus in operationalizing attrition, Bischofl and Sprenkle (1993)
suggested studies are needed to examine the interaction between research and subjects in
its relationship to dropout over time.

Some researchers differentiated between item nonresponse in which there are
missing data for individual survey items, and unit nonresponse which occurs when a
given research participant is not included in the follow-up sample, thereby yielding
missing data on all variables. In an analysis of attrition from a longitudinal study of
Emilies with chronically ill children, researchers analyzed both loss of participation
when subjects dropped out of the study, and level of participation, when there were
missing data although the Emily remained in the study (Janus & Goldberg, 1997). In a
description of retention strategies in a study of patients and spouses during early recovery
fiom heart bypass surgery, researchers decided that some data were better than none, and
designated ‘partial participants” in order to compare differences in characteristics of
subjects who did and did not complete certain parts of the study (Killian & Newton,
1990). However, others purport that in the context of nonparticipation and attrition in
panel studies, failure to complete a follow-up measurement is more salient than item
nonresponse (Ribisl et al., 1996).

Some studies that evaluated attrition depended solely on bivariate comparisons to
determine if there were significant differences with respect to predictors of attrition.
Researchers did a secondary analysis to determine characteristics differentiating those
who completed fi'om those who dropped out of a longitudinal multicenter clinical trial of

caretakers of infants at risk for cardiopulmonary arrest (Moser et al., 2000). The analysis

21

compared baseline sociodemographic, emotional, psychosocial, and infant characteristics
of those who dropped from the trial with those who completed the trial. Using chi-square
analysis or independent t-tests, the results showed that emotional and psychosocial
characteristics were important in predicting attrition in this sample.

The next Step to assess sample attrition is to progress to multivariate techniques
such as regression analysis, or logistic regression, to compare groups of predictor
variables covering different types of hypothesized effects (Mihelic & Crimmins, 1997).
There are conflicting findings regarding who is most likely to refuse participation or to
drop out of longitudinal surveys (Mihelic & Crimmins; Shaw, Cronan, & Christie, 1994;
Thompson et al., 1994). Mihelic and Crimmins attributed the ambiguities to the lack of
multivariate analyses and inadequate consideration of confounding characteristics that
may account for differential attrition. They criticized work that neglects to account for
individual characteristics that my change over time or ignores the contribution of
unobserved characteristics (Mihelic & Crirmnins). In a statistical model of attrition, these
researchers found that controlling certain characteristics (i.e. demographic variables) in a
series of logistic regression models allowed them to rmke more accurate predictions
about who was more likely to be lost to follow-up fi'om a longitudinal survey of
community dwelling elders.

Predicting attrition in health studies has been dificult and inconsistent although
Shaw and colleagues (1994) find that psychosocial variables were important in predicting
attrition among older subjects in health research. In a comparative study of Emilies with
or without chronically ill children, researchers distinguished levels of participatiOn and

differentiated between the Emilies who participated fully, partially, or not at all. Using t-

22

tests and chi-square tests, and a series of stepwise regression analyses, the three groups
were compared with respect to both demographic and study measure indicators (Janus &
Goldberg, 1997). Additionally, the researchers classified reasons for loss of participation
as unknown, situational, or personal. Situational attrition was due to objective, external
conditions preventing participation, such as change of residence, lack of sufficient
English, or the subject's death. Personal attrition was considered voluntary, due to
internal reasons such as lack of interest, being too busy, or other Emily problems.

In a study to identify Ectors predicting program attrition among participants in
HIV risk reduction trials, the researchers found that most attrition occurred between the
baseline assessment and the start of the intervention, which they called early dropout
(DiFranceisco et al., 1998). Initially, univariate comparisons were done between those
who completed the program and early dropouts. Then the relative contributions of the
predictors to a subject’s likelihood of Eiling to complete the program were determined
by stepwise multiple logistic regression. Similarly, in an analysis of attrition in a Emily
study with children, researchers used a risk model to assume that there were several
Ectors, rmny of which included small efl‘ects, but that these Ectors accmnulated to
increase the likelihood of attrition (Kazdin, Stolar, & Marciano, 1995). They argued that
multivariate techniques must be used to investigate attrition because univariate models
are simplistic and do not consider the simultaneous influence that variables have on both

each other and the phenomenon under investigation.

23

 

Models Related to Sample Attrition

Existing explanatory models related to sample attrition approach the phenomenon
indirectly by addressing outcomes related to subject access, participation, response,
recruitment, or retention.
Theories of Survey Response Behavior

Some theories of survey response behavior have been specified and empirically
studied (Brehm, 1993; Evangelista et al., 1999; Groves et al., 1992). In research
involving public opinion polls and household marketing surveys, Groves and others
(I 992, 1998) developed and tested a theoretical framework to describe multiple Ectors
that influence survey participation. The “multilevel conceptual fiamewor ” (Groves and
Couper, 1998) consisted of societal level Ectors, attributes of the survey design,
clmracteristics of the sample person, attributes and behavior of the interviewer, the
respondent-interviewer interaction, and the social environment in which the request for
survey participation is rmde. Although many of the concepts included in this model of
survey response belmvior are applicable to sample attrition in longitudinal health care
studies, the model falls short in its application to social research studies which typically
use purposive sampling with restrictive eligibility criteria, as compared with surveys
using random population sampling (Patrick et al., 1998).
Access and Recruitment in Health Care Studies

Similar theoretical models occasionally have been applied to participation in
longitudinal Emily health care studies; however, empirical testing is limited (Hautrnan &
Bomar, 1995; Holder et al., 1998; Lengacher et al., 2001). Hooks and others (1988)

described a model for recruiting participants in a study of healthy Emilies assessed for

24

dietary and exercise habits. This model included four temporal phases whereby the
researcher establishes trust within the community, establishes contacts with Emilies,
encourages Emilies to commit to the project and maintains the conneCtion between
Emilies and the research project. Although theoretical contributions of social learning,
social support, and community development were cited, the theoretical underpinnings of
the model were not expanded (Hooks et al). Hautrnan and Bomar (1995) proposed an
interactional model of research recruitment and retention, particularly with application to
people fi'om diverse ethnic and. cultural backgrounds. The key figures in their model
were community(ies), site(s), participant(s), and researcher(s). Intended primarily for use
in qualitative research where researchers-participant interactions are inherent, support for
this model in panel survey studies is limited. However, irrespective of study
methodology, researchers and participants are involved in a mutual relationship
(Hautman & Bomar).
Participation in Clinical Trials

Recently researchers developed a model that identified barriers, issues, and
strategies related to participation in clinical trials (Lengacher et al., 2001). The authors
identified key Ectors, including study design, participant factors, issues related to ethnic
diversity, the informed consent process, and physician Ectors, and suggested that the
model be used prior to and during enrollment of subjects into clinical trials. However, the
model has yet to be validated empirically. Additional models have been used to explain
various types of health behavior, and several have been applied to explain why people
would or would not use health services, or participate in prevention, screening, or

intervention programs (Manfredi, Lacey, Wamecke & Balch, 1997; McKenzie & Jurs,

25

1993; Padula, 1997; Spoth & Redmond, 1994). These models are rooted in a variety of
conceptual frameworks including social learning theory (Perry, Baranowski, & Parcel,
1990), models of interpersonal health behavior (Lewis, 1990), marketing theory
(McKenzie & Jurs; Novelli, 1990), and health behavior models (DiFranceisco et al.,
1998; McKenzie & Jurs; Seigley, 1998).

While existing models related to health care studies have merit, they are
inadequate in their application to understanding sample attrition in descriptive panel
survey research of Emilies with chronic illness. First, the models are most often applied
to predict preventive health behavior and adherence with medical regimens (Patrick et al.,
1998; Seigley, 1998; Spoth & Redmond, 1994). The emphasis on “compliance” and
“adherence” implies that the impetus to participate depends solely on subject behavior
(Wuest, 1993), but seems to neglect tint both subject characteristics and research design
may simultaneously influence these outcomes. Interestingly, more recent literature
criticizes the stereotyping of nonparticipants as “noncompliant,” “unreliable,” and
“unwilling” (Underwood, 2001); and the focus in the clinical trials literature has shifted
to improving “participation” (Lengacher et al., 2001; Underwood & Alexander, 2001).

Second, although some evidence for the ecological model developed in this
dissertation emerged fi'om reports of successful strategies for recruitment and retention in
health promotion and intervention studies, many issues applicable to clinical trials and
other studies with experimental designs are different fi'om those with descriptive designs.
Specifically, clinical trials typically are conducted to establish a relationship between
treatment and outcomes, and usually involves placement of subjects into experimental

groups on a random basis (Lengacher et al., 2001 ). On the other hand, a survey is an

26

observational research technique, typically used for descriptive, explanatory, or
exploratory purposes, and does not involve randomization of subjects (Babbie, 1998.)

Third, in existing models, demographic and personal Ectors are explored in
relationship to health behaviors whereas environmental Ectors are excluded (Seigley,
1998). Inasmuch, such models do not fully capture the multidimensional characteristics
that any contribute to sample attrition, particularly in longitudinal studies of Emilies
with chronic illness. There are a few exceptions. One is a study of health behaviors of
community-dwelling older adults in which Seigley (1998) emphasized the importance of
understanding individuals in the context of their unique environment. Conceptualizing
health belmviors as a function of the continuous reciprocal interaction between personal
and environmental Ectors, she found that the interactions of social support and functional
health, and the interaction of social support and self-esteem significantly increased the
explained variance in health behavior for older residents of resource-poor environments.
Powell (1988) also incorporated an ecological approach in identifying program design
issues regarding the recruitment and participation of individuals in a community-based
parent education and support program. By presenting a retrospective case study of the
interplay between client characteristics and the design of a community-based program,
Powell portrayed the program as a dynamic fluid organization and depicted the
participants as active individuals who engage in the program in different ways. While
Powell recognized that all aspects of program-environment relations, such as linkages
with funding agencies, affect program design, he did not explore that aspect.

Finally, although individuals are profoundly influenced by Emily relationships

and dynamics, existing models typically focus on the individual (Padula, 1997). Some

27

studies do focus on families in which an elderly member is ill, but there is a great deal
more in the literature on Emilies with pediatric chronic illness. Furthermore, while
reports about studies of individuals are numerous, there are relatively few reports
describing strategies for the recruitment and retention of families in longitudinal research
(Ryan & Hayman, 1996). Thus, incorporating 3 Emily perspective into the study of
sample attrition is imperative.
Summary

A wide array of literature deals with the consequences, the correlates, and the
phases of nonparticipation, early dropout, and attrition over time. Yet, a serious limitation
in Emily health care research is that information relating to sample attrition emerges
from anecdotal examination of completed research studies, rather than studies that are
designed to specifically evaluate dropout (Bischofl‘ & Sprenkle, 1993). Furthermore,
researchers do not usually ask people why they do not participate or why they dropped
out, but instead focus on a variety of characteristics of either the subjects or of the study,

examined in a retrospective fashion.

28

CONCEPTUALIZING AN ECOLOGICAL MODEL
' Overview of Model

A primary goal of this dissertation is the development of an Ecological Model of
Sample Attrition in Family Health Care Panel Studies (EMSA), shown in Figure 1.

Based in systems thinking, an ecological perspective highlights the interrelationships
among and the connectedness of organisms and their contexts (Bristor, 1990; Grifl'ore &
Phenice, 2001; Wideen et al., 1998). It is particularly useful when the problem under
study is complex, and it captures the multidimensional character of a phenomenon.

The ecological model depicts an organizational structure made up of
interdependent, interrelated parts. It is a holistic, non-linear representation of the dynamic
interactions and the connectedness between subjects, researchers, and their environments.
The nested model theorizes a multilayered situation incorporating the Ect tint individuals
and Emilies and researchers operate simultaneously in multiple contexts (Benjamin &
Hollings, 1997). Sample attrition is conceptualized as multileveled and subject to
numerous influences, both within and outside of the research study.

The important features of the model are the three enviromnents, the subject
characteristics, the research design, target sample, the outcomes of attrition, and the
notion of time. The model proposes that subject characteristics and the research design
are situated in the Emily environment, the health care environment, and the social
enviromnent. Simultaneously, the research design and subject characteristics
independently and interdependently influence whether subjects participate or do not
participate in a panel survey (early attrition), and if they do participate, what may

influence dropout over time (attrition over time). The model in Figure 1 is dissected into

29

Figure 2 and Figure 3. Figure 2 illustrates the subject and research design components of
EMSA and Figure 3 illustrates the components of the environments that interact to
influence whether attrition Item the target sample occurs. These figures, although
presented separately, should be viewed concurrently as they represent the multiple layers
underlying subject attrition.

A key aspect of this model is its dynamic nature. Although difficult to capture
graphically, this attribute allows the elements of the model to change in shape and Size
according to their relative importance over time (Hancock, 1993). Inherent in this idea is
that the relationships are complex and multidirectional. This is depicted by the shape of
the model, the multiple layers, the overlapping and intersecting environments, the dashed
lines representing permeable boundaries, and the shapes and colors of the arrows.

In the model, “target sample” is defined as all potential subjects who are found to
be eligible and are approached to participate in a panel study. Attrition, then, is a subset
of the target sample. Attrition occurs when subjects do not participate, for any reason
(Brehm, 1993). It also occurs when subjects fail to complete a study after initiating the
study (Fogg & Gross, 2000; Given et al., 1990). The shapes in the model widen to
capture the notion of the growing target sample size and the accumulation of subjects
who do not participate, who drop out, or who die over the course of a panel study.

Currently, the Ecological Model of Sample Attrition in Family Health Care Panel
Studies pertains to panel surveys of older individuals with cancer and their families.
EMSA addresses sample attrition of older adults fiom studies that access and recruit
subjects within the health care system The model is applicable to studies in which the

subject unit involves either individuals or multiple Emily members in panel data

30

collection, and in which the focus of the study is on health care issues, particularly related
to outcomes of chronic illness or its treatment on individuals and Emilies.

In order to facilitate understanding of the Ecological Model, the first part of this
chapter expands the ecological features of the model introduced earlier. The next section
reviews the subject characteristics and research design characteristics (Figure 2), and
finally, the three environments (Figure 3) are described. The nature of this project does
not allow separation of the literature review from the narrative descriptions of the
model’s features. Theoretical and operational evidence for the model was gathered fi'orn
the categories of literature listed at the beginning of Chapter 2

Ecological Characteristics of the Model

A key concept underlying this model is its ecological characteristics. An
ecological perspective is nonlinear, perceiving interaction between elements of the
ecosystem and acknowledging feedback mechanisms (Bristor, 1990). From a systems
perspective, the components of the model create an organizational structure made up of
interdependent, interrelated parts. The ecosystem of a research subject has an important
role in the determination of attrition in a research study for which that individual is
eligible. As illustrated in Figure 3, subjects (individuals and/or Emily members) are
situated in families, in the health care system, and in the social environment, and each of
these environments interacts to shape one another. The research study is also situated in
an ecosystem incorporating the researcher, the health care environment, the Emily
environment, and the social environment.

The notion of adaptation is another ecological feature of the model. Adaptation is

the ongoing process by which an organism changes its structure in order to accommodate

31

a new condition in the environment (Bristor, 1990). Subjects adapt to requirements for
participation; for example, the allocation of the resource of time often changes to
accommodate participation in a study, since usually a significant time commitment is
required to complete interviews or questionnaires several times over the course of a panel
investigation. The research design also has an adaptive quality when a study protocol is
accommodated to be flexible to the unique attributes of the environments. For example,
when field procedures and resources are modified in response to attrition, increased
efforts may be targeted to those who are most prone to nonparticipation and dropout over
time (Neumark et al., 1998).

Another important concept of the model is that it is holistic, and the variables are
interdependent, not exclusive (Witte et al., 1996). Witte and colleagues offered an
ecological theory of health communication to capture how individual behaviors are
influenced across multiple levels of communication at several time points. In a
multilayered view, they described health messages to which individuals are exposed at
multiple levels, through multiple channels (Brashers & Babrow, 1996). The strength of
the ecological perspective is the emphasis on interdependencies among variables and the
holistic patterns that emerge fi'om those interdependencies.

An ecological perspective is particularly usefirl when the problem under study is
complicated in order to capture the multidimensional character of a phenomenon. An
ecological approach links methodology to the social and cultural conditions surrounding
the research, the researcher, and the research subjects, and highlights the interrelations
and connectedness of organisms and their environments (Bond & Pyle, 1998; Wideen et

al., 1998). An assumption of systems theory is the interdependence of the individual

32

Emily member, the Emily, and the environments. Inherent in this assumption is the idea
that the relationships are complex and multidirectional. Thus the environments in EMSA
are characterized as interdependent in that the elements making up the system interact
and exert influence on one another such that changes in one component may have
radiating effects on other components (Bond & Pyle).
Time in the Model

In the model, the time dimension figures prominently. An ecological perspective
is useful to capture multiple changes over time in environmental as well as personal
factors (Seigley, 1998). Groves and Couper (1998) highlight the temporal dimensions of
survey participation by descrrhing the process of contacting a sample household,
identifying persons in the household, choosing an informant, and seeking their
participation in the survey. Brehrn (1993) describes survey participation as a process
with three stages. Brehm’s model differentiates between organizational or administrative
problems that influence accessibility, and behavioral problems that determine whether an
individual will comply with the request to participate in a survey. Thus two important
notions fiom the survey response models are that different factors affect whether an
individual moves from one step to the next, and that nonparticipation and dropout can
occur at any point in the process. These notions are represented by the black arrows in
the figures, as well as the varying sizes of the red and blue arrows.

Given the multiple stages of contact with subjects in a panel study, it may turn out
that different phases of the study are associated with variable patterns of attrition (Brehm,
1993; Hooks et al., 1988; Norris, 1985; Zapka et al., 1994). While subject characteristics

and research design contribute more equally in the process of initial recruitment in the

33

study, the model hypothesizes that the subject characteristics will be stronger predictors
of attrition over the longitudinal time period of involvement in a panel survey. For
example, as interviewers gain Skill and confidence in the research protocol, interviewer
influence on study participation might diminish (Groves et al, 1992). The strength of this
relationship is indicated by the varying thickness of the arrows.

The notion of time is also captured by the trajectory of the chronic illness. Jacobs
(1993) described illness phases and Emily response which may affect attrition in a
research study. In the acute phase, during the initial period of diagnosis, the family must
incorporate accurate information about the disease and mobilize family resources to deal
with ensuing medical demands. The time of diagnosis of a serious illness constitutes a
significant Emily crisis, and Emily members must delegate their time and financial and
emotional resources to dealing aggressively with it (Jacobs). It is precisely during this
time period that subjects may be approached for participation in a time-consuming
longitudinal survey. Over time, in the chronic phase of the illness, after initial treatments
have controlled the disease to the extent possible, family members may have found ways
to routinize and contain illness management (Jacobs). The illness may no longer
dominate family life, and ongoing participation in a study may not be so unmanageable.
Ontheotherhand, themorbidityandmortalityassociated with cancerandcancer
treatment may be an unavoidable source of attrition over time (Davies et al., 1995; Mock
et al., 1996, Rinck et al., 1997).

Components of the Model
The essence of EMSA is to capture the dynamic interactions among the three

environments that underlie both subject and research design characteristics, and

34

consequently the outcomes of early attrition and attrition over time. Evidence for the
model was gathered from a wide array of literature including research methodology,
survey response behavior models, health care recruitment models, anecdotal recruitment
and retention analyses, Emily and human ecology, Emily halth models, and an
ecological health communication model. The first section reviews subject characteristics
and research design characteristics that affect attrition in Emily health care panel surveys
(See Figure 2). The next part describes the three environments (See Figure 3) and
expands on the ecological features of the model.
Subject Characteristics and Research Design Characteristics
A fleeting Sample Attrition
Multiple interacting factors, related to both research design and subject
characteristics, afl‘ect whether an individual consents to participate or to remain in a panel
study to the end (Brehm, 1993; Groves et aL, 1992). Potential participants (target
sample) are not likely to systematically review and incorporate all the available
information in a decision, but are more likely to base decisions to participate or to refirse
to participate on one or two highly prominent considerations (Groves et al). In the same
way, this ecological model classifies the independent variables in the broad categories of
subject characteristics and research design. Rather than trying to indicate specific
directional relationships among the variables, the model intends to capture the range of
possible influences on the outcome of sample attrition fiom Emily health care studies.
The inherent relationships are complex and multidirectional.
In the Ecological Model of Sample Attrition, the subject characteristics refer to

the distinctive qualities of a person who is studied in order to collect data These include

35

demographic qualities, such as age, race and gender, socioeconomic status, personality
traits, attitudes, physical and psychological capacities, residence, and social support. In
the context of Emily health care research, subject characteristics refer to the qualities of a
person that determine eligibility in a particular panel study. Research design is defined as
all the processes a researcher undertakes in the planning and implementation of Emily
health care panel surveys, including sampling procedures, access procedures, recruitment
procedures, methods of measurement and data collection, data analysis, and data
dissemination, hiring and/or management of research personnel, and the maintenance of
the research protocol.

While researchers have little control over health-related reasons that contribute to
subjects’ reluctance to participate or to their subsequent attrition, researchers do have
control over study design and data collection techniques (Arean & Gallagher-Thompson,
1996; Eaves, 1999; Motzer et al., 1997; Neumark et al., 1998; Sullivan, Rumptz,
Campbell, Eby, & Davidson, 1996). Predicting who is prone to decline participation or to
dropout, and why, may help researchers target recruitment and retention efforts,
particularly to reduce the extent to which study-related Ectors contribute to attrition.
Subject Characteristics

Although complications are inevitable when multiple factors associated with
attrition are considered simultaneously, certain variables have emerged consistently as
correlates of attrition in health research. Older participants in panel studies are believed
to be more likely to refirse participation or to drop out (Carter et al., 1991; Mihelic &
Crimmins, 1997). Not surprisingly, the ability to obtain and maintain high participation

rates is negatively affected by decreases in physical functioning, greater intensity of

36

health service needs, poorer physical health, and worse self-reported health, all of which
grow worse with age (Corder & Manton, 1991; Mihelic & Crimmins; Norris, 1985).
Corder and Manton reviewed longitudinal national health surveys and considered special
problems in measuring the health of the elderly, particularly the oldest-old. Difliculty
with hearing and speaking, cognition, and physical stamina may be more prevalent.

The health and firnctional status of the elderly population is complex.
Furthermore, any person may have some combination of the characteristics that are
attributed to higher attrition rates. Thus one would expect older subjects to be less willing
or able to participate in research over time (Corder & Manton, 1991; Mihelic &
Crimmins, 1997; Thompson et al., 1994). Surprisingly, however, some researchers have
found that, once older adults agree to participate in a research protocol, they are less
likely to drop out (Arean & Gallagher-Thompson, 1996; Carter et al., 1991; Thompson et
al). One major reason for this may be greater availability of fiee time among the elderly
in retirement age, and the desire for social contact may more than outweigh problems
such as difliculty with hearing and speaking, cognition, and physical stamina (Carter et
a1; Corder & Manton). In fact, older adults are often good study participants, willing to
comply with study protocols and to provide high-quality responses to surveys compared
with other age groups (Carter et al). Shaw and colleagues (1994) reported an interesting
finding that self-selection of subjects involved the loss of subjects at both the high and
low levels of physical and psychosocial functioning.

Even so, in panel studies of Emilies experiencing chronic illnesses, an
unavoidable source of sample attrition is the subject’s physical mortality, which is related

to the illness itself. Unfortunately, researchers have little control over attrition fi'om these

37

sources in any study of an elderly cohort (Corder & Manton, 1991; Norris, 1985; Sharma
et aL, 1989), but especially in a longitudinal study of cancer patients. For study
participants in advanced stages of cancer, attrition can be a major problem (Davies et al.,
1995; Rinck et al; 1997). For instance, the one-year survival rates for patients with the
four primary diagnoses relevant to this study are as follows: breast 97.0%, colorectal
83.1%, lung 38.1%, and prostate 97.9% (National Cancer Institute, 1994).

Although conclusions about demographic characteristics as predictors of attrition
are often fi'om bivariate analysis of attrition by one variable after another, there is
consensus that lower socioeconomic status, whether measured in terms of lower
household incomes, fewer years of education, or renting rather than owning one’s home
is associated with higher attrition in older samples (Corder & Manton, 1991; Shaw et
al.,l994; Thompson et al., 1994). Demi and Warren (1995) address methodological
issues when race is treated as a dichotomous variable whereby there is a tendency to
assume that results obtained are manifestations of racial differences, rather than due to
economic, social, educational, and cultural differences.

Social support is related to study participation; study participants with weaker
social networks, who are unmarried, or who live alone are more likely to drop out over
time (Norris, 1987; Shaw et al, 1994). Geographic region of residence, particularly
among subjects living in rural areas, may be an impediment to research participation
(Eaves, 1999). In addition, residential relocation, which often occurs among the elderly
because of the death of a spouse or because they can no longer live independently, also
leads to loss of elderly panel subjects in follow—up due to the difficulty in locating moved

sample members (Mihelic & Crinrrnins, 1997).

38

Demi and Warren (1995) described specific methodological and ethical issues
associated with studies of vulnerable Emilies, such as those with lower socioeconomic
status or minorities. Motivation to participate may be limited among these families;
sometimes potential participants are reluctant to overtly refuse to participate but
subsequently are unavailable for follow-up. Again, participant retention may be a greater
problem in vulnerable Emilies with a low income, due to change in residence or lack of
telephones. Conversely, in a paper describing characteristics of participants that
promoted retention, Killien and Newton (1990) indicated that geographic stability was an
important Ector.

Personality characteristics and other psychological concepts are often offered as
explanations about why people participate or do not participate in research studies
(Groves et al., 1992). For example, inflexibility about daily life patterns and preference
for routines may account for attrition in studies of older adults (Thompson et al., 1994).
Others have a tendency towards altruism or feel they have an obligation to participate as
cooperative and responsible individuals (Evangelista et al., 1999). Furthermore, a
person’s decision to participate or refirse may be related to basic attitudes, such as
suspicion of strangers, or fear of breach of privacy or confidentiality (Souder, 1992).
Other reasons may include attitudes towards surveys in general, previous experience with
research, or the perceived costs and benefits associated with participation (Brehm, 1993;
Evangelista et al; Groves, 1989;Ki11ien & Newton, 1990; Shaw et al., 1994; Zapka et al.,
1994). For example, participants cite altruistic reasons to participate, as well as the
opportunitytotalkwithsorrreoneabout theirconcemsasreasonsto enrollandremainin

a panel study (Killien & Newton; Souder). Bonvincini (1998) highlights an important

39

benefit of participation that may enhance retention in a study. Often subjects appreciate
the opportunity to speak about themselves, but they do not realize this benefit until after
they have completed an interview.

Maintaining contact with the same individuals results in participants perceiving an
established personal relationship with an interviewer, may translate into a commitment to
the study over time. In a longitudinal study of patients and spouses during early recovery
fiom heart surgery, researchers described the interactions between the researcher and the
participants and found that consistent interaction over time was an important strategy in
reducing dropout (Killien & Newton, 1990). Perceived importance of the topic under
study, who is conducting the study, the nature of the questions, and the proposed
beneficiaries of study results may influence initial participation (Groves, 1989).
However, their effects on continued participation in panel studies is less certain. Groves
and colleagues (1992) describe social psychological concepts associated with
participation in a study. For example, subjects may be more willing to comply if the
request to participate comes from someone with properly constituted authority, such as
when physicians endorse a study. Whereas demographic data on nonrespondents rmy be
accessible, albeit in limited form, it is more difiicult to gather information about attitudes
and inherent personality traits.

Research Design Characteristics

In addition to the characteristics of the study participants, as captured through

sociodemographic and psychological variables, there are research design characteristics

that my have substantial effects on whether an individual or Emily agrees to participate

40

or remains on a panel study (Eaves, 1999; Groves et al., 1992; Hooks et al., 1988;
Neumark et al., 1998; Zapka et al., 1994).

Perceived importance of the topic under study, who is conducting the study, the
nature of the questions, and the proposed beneficiaries of study results may influence
initial participation (Groves, 1989). However, their effects on continued participation in
panel studies is less certain Groves et a1. (1992) describe social psychological concepts
associated with compliance. For example, subjects may be more willing to comply if the
request to participate comes fiom someone with properly constituted authority, such as
when physicians endorse a study. Also, characteristics of the health care service
environment, such as varying levels of access and recruitment support at hospitals and
clinical sites, or competition with other research studies in a particular setting, challenge
investigators to overcome recruitment and retention barriers (Arean & Gallagher-
Thompson, 1996; Motzer et al., 1997).

Sample selection can be affected by decisions about research questions,
hypotheses, instrumentation, and demographics of the target group. Subjects may be less
likely to agree to the research protocol when there are language barriers or cultural beliefs
that differ fiom what is expected and were not considered in the planning of the research
(Dennis & Neese, 2000). In an amlysis of sampling problems encountered in a study of
Afi'ican American elders’ long-term care choices, Groger and colleagues (1999)
commented that the wording of recruitment letters may have excluded potential
participants because the researcher’s definition of caregiving may have been different
hour that of the potential participants. Furthermore, the researchers raised an interesting

point in that they may have excluded potential participants by using the term “Afiican

41

American” that offended some elders who would have preferred to be called “colored”
(Groger, Mayberry, & Straker, 1999). '

Multiple factors come into play when making a decision to participate. When the
request occurs amid external distraction, or when it is too complex or too quickly
presented to be easily comprehended, it my result in poorer participation (Groves et al.,
1992). Subjects’ decisions are likely to be influenced by the burden the sm'vey design
imposes on respondents, including how much time is involved, the ease of participation,
and the length ofthe individual interview. Other study procedures, like the predominant
mode of data collection, also may contribute to the potential for sample attrition. Mailed
sm'veys, self-response questionnaires, in-person interviews, and telephone interviews all
have features that result in differential participation rates, although little is known about
the effect of their combined use (Brehm, 1993; Groves, 1989; Mihelic & Crimmins,
1997).

Interactions with study personnel could be associated with sample attrition
(Brehm, I993; Eaves, 1999; Harrington & Nicoteri, 1999; Hooks et al., 1988; Zapka et
al., 1994). Subject’s behavior can be influenced by both explicit and unintended cues
provided by the research personnel, including verbal and nonverbal behaviors, gender,
age, race, ethnicity, experience level, and interpersonal style (Bonvicini, 1998). A critical
element in a theory of survey attrition is to understand the interaction between the subject
and the research personnel and the behaviors, attitudes, and expectations each brings to
the interaction (Groves et al., 1992). The appearance, skills, and demeanor of the research

personnel can all affect the decision to join or remain on a study (Brehm).

42

Although routines for recruitment of potential subjects are standardized to protect
the scientific integrity of the research protocol, in the individual interaction between the
participant and the researcher, standardization may not really occur. So while research
personnel may be trained with a script for recruitment and interviewing, some suggest
that researchers should be trained in a variety of techniques and given guidelines for
choosing strategies about which combination to use (Groves et aL 1992). In Powell’s
(1988) ecological description of program recruitment, a strong relationship was found
between the type of recruitment technique used and the type of subject agreeing to
participate. Although issues associated with program participation difler from
participation in a descriptive research study, the emphasis on understanding the
interaction between subjects and researchers is important.

Ultimately, the context in which a request for participation takes place, modes and
manners of initial contact, persuasion categories, and the skills of the research stafi‘ may
all be relevant to survey participation (Corder & Manton, 1991; Eaves, 1999; Groves et
al., 1992; Souder, 1992). Ryan and Hayrnan (1996) identified the quality of the first
contact with subjects as vital to gaining entry into the family system A recruiter’s prior
experience can affect the skill and confidence with which the sample person is
approached with the survey request. Experienced personnel often adapt their approach to
the sample person by looking for cues about the attributes of the person and focusing on
those attributes that may be relevant to engaging that person in the study (Groves et al;
Martin, 1992; Souder, 1992). Thus, over the length of the study, there may be a
diminished effect on attrition as the researcher gains skill and confidence in approaching

subjects.

43

Sometimes researchers address early attrition via scrutiny of recruitment
procedures. Grant and DePew (1999) suggested that studies in which research personnel
assume full responsibility for recruitment of potential research participants will have
more recruitment success than studies having to rely on intermediaries. In a study by
Holder and others (1998), recruitment success occurred in clinical sites that had site
coordinators who were members of both the research team and the health care team
Understanding the routine of clinical visits, preoperative and postoperative protocols, and
follow-up procedures can assist researchers to make contact with potential participants. In
this way, recruitment may be more successful when subjects are not overwhelmed by
other health care providers or screening tests (Killien & Newton, 1990). Motzer and
others (1997) also identified some recruitment procedures including using intermediaries
to identify potential participants and to make initial contact with them; the intermediaries
received a nominal honorarium for subjects who eventually consented to participate in
the study. Eaves (1999) found that using intermediaries on the recruitment staff with
whom Emilies were already familiar helped to decrease the mistrust of the research,
which thus decreased early attrition. Furthermore, accessibility and recruitment are
enhanced when qualified and committed intermediaries are chosen (Cartwright &
Lirnandri, 1997; Motzer et al).

Spoth and Redmond (1994) empirically investigate the effectiveness of two
recruitment strategies for a Emily-focused substance abuse prevention study of rural
Emilies and emphasized that predictors of participation vary and depend on both
program and participant characteristics. An interesting point is that there were

discrepancies between professional and participant perceptions of Ectors that would

maximize participation. They concluded that participation is not a simple matter since it
depends on a combination of attitudes, Emily-role related time demands and other
Ectors. Other researchers also have described inconsistencies in the perceptions of
professionals and Emily members in terms of types of procedures to maximize retention
of Emilies in early intervention programs for special-needs young children (Saylor,
Elksnin, Farah, & Pope, 1990).

A hindrance to participation may be the repetitive, time-consuming nature of
longitudinal data collection (Killien & Newton, 1990). Eaves (1999) also found that a
major barrier to recruitment and retention was the longitudinal nature of the study. Spoth
and Redmond (1995) compared two different recruitment strategies and found higher
accrual rates in Emilies agreeing to a single pretest and subsequently being asked to
participate in a longer study, as compared to those who were informed initially of the
longitudinal nature of the study. On the other hand, retention was higher in the group that
had firll knowledge of the study time commitment at the time of consent. Clinton and
others (1986) also found that retention was enhanced when researchers emphasized the
long-term cormnitment during initial contacts with subjects about the study.

Certain design characteristics may especially impede participation of an older
sample (Corder & Manton, 1991; Mihelic & Crimmins, 1997). Lack of willingness to
participate in telephone interviews may contribute to lower participation rates for older
adults, particularly the older old (Carter et al., 1991; Corder & Manton). In one national
survey on medical care utilization and expenditures, the elderly population did not
respond well to telephone interviews, either because they lacked telephones, were hard-

of-hearing, or because long telephone conversations were burdensome. Others found that

45

men in particular might find long telephone surveys distasteful (Mihelic & Crimmins).
Pruchno and Hayden (2000) compared the costs and data quality in alternative modes of
data collection in a sample of older women and found higher attrition and more missing
data among those participating in the study via self-administered questionnaires as
compared with either in-person or telephone interviews.

Survey attributes may also contribute to attrition. The elderly cohort rmy be
suspicious of research and of questions asking personal information. Corder and Manton
(1991) described item nonresponse rates in a national health interview survey on
noninstitutionalized older people and found the highest nonresponse rate (~10%) is for
income. Other studies had missing data on sensitive program participation, income items,
mental health and emotional problems (Corder & Manton; Thompson et al., 1994).

Bussell and colleagues (1995) explored whether Emily research has an effect on
the participants and concluded that the research experience is not inconsequential for
participants. A pertinent example is choosing the home as a physical setting for research
This setting influences the interactions between researchers and subjects by creating a
more intimate encounter. Conducting research in this setting may enhance the
willingness of participants to provide persoml information.

There are discrepancies in the literature about whether the sample unit of
observation is related to attrition, and specifically whether or not having multiple Emily
members involved as subjects is beneficial. Some suggest that when a study requires
multiple Emily members, obtaining consent and maintaining the sample is difficult
(Eaves, 1999; Ryan & Hayman, 1996; Thornton et al., 1982), particularly in studies of

vulnerable or minority Emilies. Resistance on the part of any one member may render

46

Emily participation impossible (Hooks et al., 1988). Others described high recruitment
and retention rates when spouses or other family members also participated (Bonvicini,
1998; Killien & Newton, 1990). Either way, collecting information from multiple Emily
members is challenging, and the challenge extends to analyzing such data.

Subject retention may be affected by lag time between sample identification and
interviewing. In a longitudinal study of expectant Ethers, researchers found that the
quicker the follow-up between recruitment and the date of the first data collection was an
important recruitment Ector that influenced long-term retention. The quicker the follow-
up, the more likely subjects were to view project staff as serious and efficient about the ’
study (Clinton et al., 1986).

Maintaining contact with the same individuals results in participants perceiving an
established personal relationship with an interviewer, which may translate into a
commitment to the study over time. In a longitudinal study of patients and spouses
during early recovery fiom heart surgery, researchers described the interactions between
the researcher and the participants and found that consistent interaction over time
translated to reduced dropout time (Killien & Newton, 1990). Others found that the
continuity of research personnel with the subjects increased the likelihood of establishing
trust and mutual respect, which enhanced the quality and quantity of data collection
(Clinton et al., 1986; Given et al., 1991; Ryan & Hayman, 1996). Furthermore, with the
longitudinal nature of contact between researcher and subjects, over time, the researcher
may be perceived as a fiiend, and the Emily might disclose more intimate information

than in other circumstances.

47

Systematic retention procedures such as follow-up reminders, acknowledging
Emilies for participation, personalized greetings, small gifts, and thank you notes also
may reduce attrition (Killien and Newton, 1990; Motzer et al., 1997; Ryan & Hayman,
1996). Sullivan and colleagues (1996) found that in-between interview contacts was
important in maintaining retention in a longitudinal study of a high-risk population.
Finally, efions to assure subjects of the importance, value, and contribution of their
participation can enhance retention of subjects over time (Given et al, 1990).

The Environments

An ecological approach links methodology to the social and cultural conditions
surrounding the research, the researcher, and the researcher subjects, and highlights the
interrelations and connectedness of organisms and their environments (Bond & Pyle,
1998; Wideen et al., 1998). Thus the environments in EMSA are characterized as
interdependent in that the elements making up the system interact and exert influence on
one another such that changes in one component may have radiating effects on other
components (Bond & Pyle).

What is described next is essentially a literature review which serves to articulate
the features of the three environments. As illustrated in Figure 3, subjects (individuals
and/or Emily members) are situated in Emilies, in the health care system, and in the
greater social environment, and each of these environments interacts to shape one
another. The research study is also situated in an ecosystem incorporating the researcher,

the health care environment, the Emily environment, and the social environment.

48

Social Environment (SE)

In the model, social environment does not refer to physical setting; rather, it refers
to the social mics, roles, customs, and policies that affect people’s relationship to
research studies (Bond & Pyle, 1998). The location of the research design in the social
environment is evident as well. Namely, many of the issues in determining the choice of
a particular sample are pragmatic and stem from the fact that the researcher functions
within the context of a broader social order (Sjoberg & Nett, 1968).

The social environment includes social institutions and social interactions (Witte
et al., 1996). Groves and others (1992) describe global characteristics that have an effect
on survey participation. These Ectors determine the context within which the request for
participation takes place and constrain both the researchers and the respondents. Some
Ectors tint may underlie survey participation include the legitirmcy of social institutions,
for example, and the perceived legitimacy of surveys. Endorsement fi'om local or
national groups may add legitirmcy to the research, making participation attractive to
potential subjects (Diekmann & Smith, 1989). From the researcher’s perspective, a
barrier to recruitment and retention of participants can emerge from lack of familiarity
with the state and local agencies, or the communities in which subjects are located
(Eaves, 1999).

Bussell and colleagues (1995) describe the interpersonal interactions that occur
between the researcher and the subject whereby research participation occurs as an
agreement between the investigator and the subject to engage in a specific form of social
interaction. The research context incorporates the social interaction that occurs between

research personnel and subjects, shaped by both parties’ motivations, expectations, and

49

perceptions. Furthermore, other personal characteristics, including sex, age, race,
experience level, and interpersonal style can affect the responses of research subjects. In
studies of Emilies, additional levels of complexity are introduced, layered upon the social
interaction between the researcher and the subject. Thus a critical element in the model
is to capture the interaction between the subject and the interviewer, and the behaviors,
attitudes and expectations each brings to the interactions (Groves et al., 1992).

The importance of social and cultural context is widely recognized in the
recruitment and retention literature, and articles that focus on access, recruitment and
retention of multicultural and diverse populations are prevalent. Wariness about
researchers may be common among groups who have been disenfranchised by the
dominant culture, exploited, or abused by researchers (Hautrnan & Bomar, 1995). In a
study of Afiican American Emilies with a chronically, physically ill member, Holder and
others (1998) asserted that Afiican American communities share unique beliefs and
experiences that have implications for their recruitment into research. More specifically,
Patrick, Pruchno, and Rose (1998) identify social barriers to participation and retention in
social research samples. For example, they attribute attitudes of fear and suspicion of
research to the historical events surrounding the role of research in the Afiican American
community. In Ect, references to the legacy of distrust resulting from unethical medical
research such as the Tuskegee Syphilis Experiment is nearly universal in literature
addressing recruitment of this population. In spite of this common historical context,
researchers must recognize heterogeneity within groups which precludes assumptions

about group practices and home, including generational differences, lifestyles, values,

50

social class, economic status, or education. (Dennis & Neese, 2000). Subtle differences of
groups cannot always be seen or known by the researcher.

Researchers often describe differences between researchers and research subjects
that may impede interactions that occur in the undertaking of a study. Demi and Warren
(1995) describe the potential for racism and classism that may occur when investigators
are middle-class, well—educated Caucasians studying people from the lower class, poorly
educated, and minorities. Although researchers may adopt strategies rooted in genuine
concern for the research participants, they may have beliefs, attitudes, and experiences
that impede interactions with subjects who are culturally, racially, and socioeconomically
different (Demi & Warren).

To summarize, the social environment is a component of the ecosystem in which
both research subjects and the researchers are situated. The social enviromnent is
conceptualized as the social Ectors affecting subjects’ and researchers’ relationships to
research studies, including social rules, roles and customs, social institutions, social
interactions, culture and ethnicity, historical context, race, age, and the legitimacy of
social institutions.

Health Care Environment (HCE)

Boundaries and accessibility are key components of the health care environment.
Investigator accessibility depends on institutional approval and recruitment support at
hospitals, physicians’ omces, and other health care Ecilities. Without institutional
endorsement and access, recruitment is impossible (Diekmann & Smith, 1989).
Strategies for gaining and mintaining access vary based on type ofsetting, the research,

cultural norms, and oficial policies in place regarding access to a particular setting

51

(Hautman & Bomar, 1995). Impedirnents to access include studies that do not have
direct benefits to the institution, or an unwillingness to divert stafl‘ time to participate in
recruitment efforts. In particular, researchers Ece barriers to access subjects in health
care agencies in which the researcher has no established reputation (Cronenwett, 1986;
Motzer et al., 1997; Nokes & Dolan, 1992). Nokes and Dolan descrrhe the problem in
conducting research when the investigator is not affiliated with a health care agency
because each agency poses certain requirements that must be satisfied before the research
can be initiated and before the researchers can gain access to potential subjects.
Furthermore, the rigidity of the institutional environment, as well as bureaucratically
organized institutional regulations, may impede the identification and hiring of the best
staff available to work on a study (Coen, Patrick, & Shem 1996).

While research studies are designed to allow the potential subjects to decide
whether they wish to participate in a particular study, an important issue regarding
recruitment is access to subjects. The accessrhility of an individual influences whether
research personnel are able to initiate contact with the person (Brehm, 1993). Potential
subjects may not find out about studies if investigators are denied access to their names
by physicians, by hospital research committees, or by other gatekeepers of practice
settings where subjects would be found (Cronenwett, 1986; Feetharn, 1991). Inflexible
boundaries exist as the healthcare team may protect the ill patients and their families. On
the other hand, gaining cooperation fi'om professiomls, particularly physicians, may
improve subject recruitment (Grant & DePew, 1999). These concepts are widely

supported in the recruitment literature.

52

Another feature of the health care environment that underlies issues of sample
attrition is competition with other research studies. Gaining access tohurnan subjects
needed for a research study is affected by simultaneous requests to access relatively small
populations of patients (Cronenwett, 1986; Motzer et al., 1997). Potential subjects are
often not approached for participation when investigators are denied access due to
institutional priorities about which studies take precedence to accrue patients.

Community-based research involves interaction between the researcher and the
practitioners’ ofices that can provide access to eligible subjects for health care research
with special patient populations (Carey, Kinsinger, Keyserling, & Harris, 1996). This
collaboration can affect access, sampling, and recruitment of subjects, particularly if
research is time consuming or if physician enthusiasm is low. Access and recruitment of
subjects has been hindered by clEnges in health care delivery, including downsizing of
services, and shifting care from hospital to ambulatory settings (Motzer et al., 1997).
Investigators must interface with multiple agencies to access potential participants.
Within the health care environment, this includes institutional review boards,
administrators, tumor registries, and medical and nursing staffs.

In EMSA, the subject is also situated within the health care environment. Many
Ectors influence the ability to obtain samples in research of Emilies and health. For
example, there is variability of health resources available to potential sample members
which may be due to differences in health care providers or regional or geographic
disparities in access to health care or in health care practices (Feetharn, 1991). Medically

underserved areas rmy be less likely to support research endeavors. Thus, equity of

53

access to health care or use of health care services can affect access to diverse
populations (Dennis & Neese, 2000; Thomas, 1987).

To summarize, health care environment is a component of the ecosystem in which
both research subjects and the researchers are situated. The health care enviromnent
refers to the health care system Ectors affecting subjects’ and researchers’ relationships
to research studies. It is characterized by boundaries which afiect accessibility to research
subjects, researcher afliliation with the health care agency, competition with other
studies, and regioml or geographic disparities in health care.

Family Environment (FE)

The Emily environment incorporates Emily configurations, norms, roles, and
relationships that may influence nonparticipation and dropout over time. Although
subjects may share their own unique experiences, individual experiences are influenced
by Emily relationships, family composition, and Emily dynamics (Holder et al., 1998;
Padula, 1996). Studies of Emilies are complicated because of the complex nature of the
relationships within Emilies (Bray, Maxwell, & Cole, 1995). Furthermore, researchers
must recognize the uniqueness of each Emily system and their unique patterns of
interaction (Ryan & Hayman, 1996).

Family relations are embedded within cultural and ethnic contexts that may define
and influence the processes of certain groups of families (Bray, 1995). Context also can
affect the sampling ofmultiple Emily members and the collection ofEmily data. An
individual may respond differently to requests to participate, depending on whether or not
other Emily members are present (Feetham, 1991). In particular, Emily members ofthe

elderly often become involved in the decision to participate in research studies and

54

express moral, ethical, personal and financial concerns when potential subjects are
approached (Williams, 1992). Williams describes the involvement of Emily members in
decisions of whether an elderly Emily member should participate in a research study, and
indicates that many Emily members advise elderly relatives not to sign papers before
getting their approval. Although some people may be submissive to requests to
particMe, in the clinmte of increased health care consumer awareness, subjects and
Emilies may readily question any requests for volunteering (Martin, 1995).

Researchers need to be attentive to the Ect that Emilies my have spokespersons
who present the family to the outside world; these spokespersons may serve as key
infornEnts, but they also my act as gatekeepers (Daly, 1992; Groger et al., 1999).
Bonvicini (1998) finds that a Emily gatekeeper, or one who has considerable influence
within the Emily, is paramount to successfirl recruitment and retention. Family members
may shelter subjects and present a barrier to contacting respondents because of suspicion
of the interview. Researchers suggest that overcoming distrust of researchers via
establishment of strong and positive relationships with Emily members contributes to
successful recruitment and retention of older research participants (Arean & Gallagher-
Thompson, 1996).

The impact of chronic illness on Emilies varies tremendously. Both intrafamilial
dimensions as well as areas in which Emilies interact with outside systems can influence
a Emily’s response to chronic illness (Jacobs, 1993). Families may be especially
protective of ill family members, fearing that partaking in a study rmy expose their
Emilymembertounnecessarystressorharm, orthatmedicalcaremaybeafl‘ccted (Coen

et al., 1996; Holder et al., 1998; McNeely & Clements, 1994). In a study of Afiican

55

American Emilies with a chronically, physically ill member, Holder and colleagues
(1998) examined reasons Emilies provided for refusing participation. Interestingly, when
comparing recruited to nonrecruited families on some simple measures of illness severity,
there were no differences in the groups, suggesting that the patient’s or Emily’s
subjective feelings about the illness may be important to recruitment.

Boundaries are an important feature of the Emily environment. One particularly
salient challenge in working with families in research is to gain access to the private
activities of Emily members. Families differ with regard to ways boundaries are defined
and the processes by which Emily boundaries are presented and maintained. Boundary
regulation involves the management of space and privacy among individuals in the
Emily, between generations, and between the family and the outside world. In the
presence of chronic illness, these boundaries can be disrupted (Jacobs, 1993). Family
boundaries firnction to demarcate insiders fi'om outsiders and vary with respect to their
permeability (Daly, 1992). Although Daly writes that qualitative research, in comparison
with more remote methods of data collection, allows for the construction of relationships
with the participants, this similar sort of interaction occurs between members of the
research team and Emily members also in the context of survey research.

Recruitment and retention of Emilies for longitudinal research may be impeded
by reluctance to participate, family changes and crises, and situational stressors (Ryan &
Hayman, 1996). External Emily circumstances and concurrent life events, unrelated to
the research study, may be a primary reason for attrition in Emily studies. For example,
Emilies with weak social networks may feel subjectively overwhelmed. In a study of

Afiican American families, some refused to participate because of chronic illness or

56

substance abuse in other family members, or marital conflict (Holder et al., 1998).
Additionally, individual and family system changes over time. C6hanges in Emily
composition due to health, illness, or situational transitions may induce changes in Emily
roles and functions, thus hampering ongoing participation in research (Ryan & Hayman,
1996).

One conceptualization of the family is that Emily is the environment for the
individual Emily members. F eetham ( 1991) extends this definition to recognize that the
Emily- isamcdiator between the individual Emily members and the environment.
Recognizingthe Emily as the environment supports the assumption of interdependence
of individual Emily members with the Emily, and the interdependence of the individual
and Emily within the health care environment. Family systems theory assumes the
interdependence of the individual Emily member, the family, and the environment
(F eetham). Inherent in this assumption is the idea that the relationships are complex and
multidirectional. Data analytic techniques that test simple linear relationships are not
consistent with these theories.

To summarize, Emily environment is a component of the ecosystem in which
both research subjects and the researchers are situated. More specifically, Emily
environment is conceptualized as family factors affecting attrition. It is influenced by
family relationships and roles, family dynamics, decision-making processes, Emily
gatekeepers, and the permeability of Emily boundaries. The Emily environment interacts
with the health care environment in relationship to family response to illness, and/or the
illness trajectory. It is further influenced by external family circumstances and concurrent

life events.

57

Propositions from the Model

The following propositions are statements to describe the relationships between

the concepts and variables in the Ecological Model of Sample Attrition. The propositions

are directed towards the outcome of sample attrition in panel surveys of Emilies in which

an older member has cancer.

1.

A researcher must have access to a potential subject in order for the request for
participation to occur.

If a researcher has access to a potential subject, and there is a social interaction
that occurs between the researcher and the subject, it may result in the subject
deciding not to participate in the research study.

Sociodemo graphic subject characteristics (independent variables) like age,
gender, and race rmy afl‘ect whether subjects are more likely to refuse
participation or to drop out (dependent variables).

The age of a subject (IV) may be related to whether the subject is more likely to
refuse participation or to drop out (DV).

The race of a subject (IV) may be related to whether the subject is more likely to
refuse participation or to drop out (DV).

The gender of a subject (IV) may be related to whether the subject is more likely
to refuse participation or to drop out (DV).

The presence of a cancer diagnosis in a subject (IV) may be related to whether the

subject is more likely to refuse participation or to drop out (DV).

58

8. Research design characteristics (independent variables) may influence whether
subjects are more likely to refirse participation or to drop out (dependent
variables).

9. Whether a family member participates along with a subject (IV) may influence
whether subjects are more likely to refuse participation or to drop out (DV).

10. Recruitment procedures (IV) may influence whether a subject is more likely to
refuse participation or to drop out (DV).

Support for An Ecological Perspective of Inquiry

Progress in research surrounding attrition in family health care panel research
requires an ecological perspective. Currently, no model exists to capture the dynamic and
interactive nature of sample attrition in Emily heath care studies. It is assumed for the
purposes of this research that an ecological model is a useful way to depict sample
attrition to enhance understanding of this complex phenomenon in Emily health care
studies. Benjamin and Hollings (1997) raise the issue of theoretical parsimony, the idea
that the best models are those involving the fewest assumptions and factors. They suggest
tlmt a thrifty modeling strategy is appropriate when considering simple phenomena,
where relationships are linked in a linear manner. However, in trying to conceptualize
and analyze the complex processes and patterns of relationships underlying attrition in
Emily health care panel studies, conceptual parsimony is counterintuitive. Rather, the
ecological model is multilayered and holistic. Furthermore, the model is dynamic. As
awareness of different levels of complexity become apparent, new properties and insights

into the phenomena emerge (Wideen et al., 1998). Therefore, although EMSA continues

59

to evolve, in its current form, it provides the conceptual foundations to operatiomlize and

empirically test the model .

6O

METHODS
Ex Post F acto Case Study of Sample Attrition

The first purpose of this dissertation is to develop and elucidate the Ecological
Model of Sample Attrition (EMSA) in Family Health Care Studies (Chapter 3). The
model illustrates the complexity and interaction of subject, researcher, and the social,
Emily, and health care Ectors which my be associated with attrition in panel research
on Emilies with a member experiencing chronic illness. The second purpose is to utilize
the model as a conceptual guide in order to operationalize and to statistically analyze the
relationships between research design and subject characteristics on the outcome of
attrition from a target sample of older individuals with cancer and their Emilies.

As described in the introductory chapter, the dissertation is an ex post Ecto case
study, based on the Family Care Study (Given & Given, 1991-1996). The primary
characteristic of an ex post Ecto study is that the variables are studied as they occur in a
mtural setting, because they lmve already occurred (Babbie, 1998). Therefore the
variables cannot be manipulated, and the analysis is limited to data collected during
sample accrual and recruitment, or from data collection via interviews, questionnaires,
and medical record audits.

It is appropriate to include quantitative evidence in a case study, and to rely on
multiple sources of evidence, with data needing to converge (Yin, 1994). Furthermore, a
case study benefits from development of theoretical propositions to guide data collection
and analysis and is usefirl to expand and generalize these propositions. This research
strategy is appropriate when trying to capture relationships and the possible effects

among the variables, and not just wanting to describe a situation (Creswell, 1994; Yin).

61

Although the Family Care Study was not designed as a study of sample attrition, it is
hypothesized that the context of this descriptive panel survey contains important
explanatory variables about the phenomenon of sample attrition.

The empirical portion of the dissertation is based on the Family Care Study
(FCS), a longitudinal comparative survey of elderly cancer patients, with or without
Emily caregivers, who were observed for one year following initial cancer diagnosis
(Given & Given, 1991-1996). The goal of the FCS was to describe how the diagnosis of
cancer and treatment alter the daily lives of elderly patients and their Emilies. The survey
probed the time and effort that Emilies devoted to care during cancer treatment, and
gathered details about the costs of cancer to patients and their Emilies. Finally, the
survey asked participants to identify health agencies and community resources most
beneficial in assisting families with care.

As recruitment coordinator of the Family Care Study, I was closely involved in
ongoing field procedures associated with access, recruitment, retention, and attrition of
elderly subjects fiom over 60 clinical sites around Michigan and Indiana (Neumark et al.,
1997; Neumark et al., 1998). My role as a researcher within the context of the case study
enables this inquiry to retain the holistic and meaningful characteristics of real-er events
and is suitable to pose questions such as what are the relationships and the possible
efl’ects among the variables (Yin, 1994).

Research Questions

The inquiry warrants the formation of several research questions to specify what

the study will attenrpt to explain about sample attrition. First, with regard to early

attrition, the research question is: Do research design or subject characteristics predict

62

who was more likely to refirse participation or to drop out prior to the initiation of a panel
survey of elderly cancer patients and Emily caregivers?

With regard to attrition over time, research questions emerge fi'om the Ecological
Model of Sample Attrition: Do research design or subject characteristics predict who was
more likely to drop out over time from a panel survey of elderly cancer patients and
Emily caregivers? Do different characteristics predict who drops out due to death versus
who drops out for reasons other than death?

Describing the Family Care Study

Eligibility criteria for the FCS included the following: (a) a cancer patient had to
be 65 years old or older; (b) was newly diagnosed (not recurrent) with one of four
primary diagnoses: breast, colorectal, lung, or prostate cancer; and (c) was recruited
within six weeks of initial surgery or within two weeks of the start of chemotherapy or
radiation therapy.

The original accrual goal for the FCS was to enroll 1235 patients, in order to
ensure that approximately 800 patients (200 of each diagnosis) would survive to the one
year observation. In order to gain access to as many subjects as possible and to achieve
the target sample, subjects were recruited over a three-year period fiom over 30 locations
in Michigan and Indiana, including in-hospital surgical units or oncology units,
comprehensive cancer treatment centers, radiation therapy units, outpatient
chemotherapy clinics, and outpatient surgeon offices. The Family Care Study had
recruitment procedures that resulted in selective case identification, rather than complete
case identification so accurate attainment of the number of subjects eligible for

recruitment or comprehensive documentation of patient refusal and attrition was not

63

possible. Since subjects were recruited from hospitals and multiple outpatient settings, it
was not possrble to identifi the numbers of patients within a specific age or gender
category who had a particular cancer diagnosis. There was no central location where this
information was stored, and most of the institutional review boards did not allow access
to all of this information. Thus the population of eligible cases was not known with
conrplete certainty (C. Given, personal communication, 1995). However, even under
these less than perfect conditions, it was essential that the recruiters were able to track
and to separate the subjects who chose to participate fiom those who refused
participation.

In the Family Care Study, community-based recruiters were responsible for
determining eligibility, establishing contact with the potential subject, explaining the
study, and obtaining informed consent. Each location designated who would work as
recruiters, and every agency stipulated different mechanisms through which recruiters
were allowed to determine to approach potential patients and their caregivers to ascertain
consent or refirsal to participate in the PCS. There was a lot of variability in the ease of
identifying and accessing potential subjects. Study personnel identified patients via
clinical conferences, pathology reports, surgery schedules, chart review, or contact in
their employment setting. Once eligibility was determined, contact was made during
hospitalization, during a clinic visit, or at a time convenient to both the recruiter and the
patient.

Potential participants were informed tlmt the study included four telephone
interviews, four parallel self-administered questionnaires, and reviews of medical

records and Medicare claim files to examine treatments and physical, emotional, and

financial outcomes over the first year following a new cancer diagnosis. F ollow-up by
telephone and obtaining consents by mail or at a follow-up visit was an option if within
the time frame of eligibility. Altogether, between February, 1994 and May, 1997, 67
different people served as recruiters and enrolled at least one case into the study.
Successful enrollment required verification of case eligibility and the receipt of informed
consent at the research office at the university.

Family Care Study subjects signed consents to participate in health care settings
and then were contacted for the first full-length interview some time after written consent
was received. There were four waves of participation in the Family Care Study. The first
contact between the interviewer and the subject occurred after consent was received at
the research oflice. On average, contact with subjects was established within 11 days
after the receipt of the written consent. For 83 cases (6.9%) the interviewer contact
occurred on the same day as the receipt of consent, and for more than half of the cases,
this contact was accomplished within the first 9 days after the consent form was received.
However, for 17 cases (1.4%) it took interviewers more than a month to establish contact.
The second wave interview was approximately 6-8 weeks after the first; the timing was
established by an algorithm in the research design in order to link the initial adjuvant
cancer therapy with functional and symptomalogic milestones experienced by patients
and Emilies (Given & Given, 1991 -1996).Waves 3 and 4 were scheduled at 6 months
and 12 months. At each wave of data collection, individual participants were interviewed
by telephone by the research stafl‘ at the university. Following completion of an
interview, participants were asked to complete a self-administered questionnaire and to

return the questionnaire to the research office at the University.

65

The Family Care Study recruitment protocol permitted patients to enroll without a
family caregiver, if there was not a caregiver available or if a caregiver was not willing to
participate. Additionally, there were a few instances in which caregivers only were
invited to participate. In the FCS, the cases were counted depending on whether consent
forms were received fiom patients only, dyads of cancer patients and their designated
Emily caregivers, or caregivers only. Furthermore, a case that was initially designated as
a patient-caregiver dyad remained assigned to this case type for the duration of the study,
even if the patient or the caregiver later withdrew fi'om participation.

Although counted as a single case, patients and caregiver dyads were not
interviewed together. Several interview tools were available in order to acquire as much
data as possible, depending on whether caregivers were participating with a patient.
Certain portions of the interview could be answered by either Emily member, but some
questions, such as patient symptom experience, were only answered if a patient
completed the interview. The rationale for admitting “caregiver only” cases was that
many of the questions exploring financial and service use issues could be answered by
Emily members, and patients who were distressed at the time of their initial diagnosis
might later decide to participate more fully in the study.

During the three years of data collection, 30 different individuals worked as
interviewers on the Family Care Study. Primarily graduate students in nursing, mdicine,
or epidemiology, the interviewers underwent extensive training and were monitored
periodically to assure adherence to the research protocol. During the second year of data
collection, the telephone interview was programmed into laptop computers and the

interviewers entered data directly as they conducted the interview. Interviewers received

66

assignments on a weekly basis, and were responsible to contact the subjects, schedule the
interviews and conduct the interviews within a specified time fiame according to the.
research protocol. The questionnaire and addressed, stanrped envelopes were sent to
patients and caregivers following completion of the telephone interview, although in a
few cases, questionnaires were completed although interviews were not. Interviewers
were required to document dificulties in contacting the subjects, as well as information
about subject dropout or death, and field notes were entered into the database used to
track subject retention and attrition overtime.

The tracking database was developed as a management system to assist in the
conduct of the Family Care Study. The key goals of the system were to follow sampling
and accrual data by each recruitment site, to verify eligibility, and to keep track of contact
and data collection at each wave. The tracking system was developed to monitor and
evaluate the research process as it was being undertaken and to allow for new or revised
approaches as necessary. It allowed the researchers to monitor if attrition was occurring
and helped make certain that there was adherence to the research protocol with regard to
the timing of the panel waves.

In a complicated design like the Family Care Study, a management system was
necessary to monitor accrual, data collection processes, and the changes in the status of
patients over time. The management system was not without flaws; yet, it was developed
to handle a panel survey with multiple observations, multiple forms of data collection,
many research personnel, and many subjects. It allowed the generation of reports to

assess where attrition was occurring, whether it was related to specific recruiters,

67

interviewers, or types of patients, and to evaluate strategies for decreasing the rate of
attrition (C.Given, personal communication, January, 1995).

The management system was developed using computers and database
technology. Once a case was entered into the system, the programmed algorithm
automatically determined the timing of subsequent waves. The tracking system allowed
the researchers to anticipate stafling needs and allowed interviewers to plan their
schedules by producing a list of interviews that needed to be done within a two week time
period. Additionally the system generated mports that Ecilitated day~to-day nmnagement
of the Family Care Study, such as newly assigned interviews, and the status of interview
and questionnaire schedules. Additionally, the intervals for observations were an integral
part of the research design. Interviewers were required to schedule interviews within
designated time parameters and were expected to follow up on late interviews and
questionnaires. Thus, the management tracking system helped the researchers monitor
adherence to the overall research protocol and the time intervals, and allowed
interviewers to anticipate upcoming caseloads and to plan accordingly.

Retention enhancing procedures in the Family Care Study included sending
participants small gifts and personal notes after each wave of data collection, and sending
subjects birthday cards. Also, interviewers ascertained consent from participants to
remain on the study at each contact. Finally, in response to attrition anticipated between
the 6-month and 12-month interview, a "3.5" wave was instituted which consisted of a
telephone call midway between wave 3 and wave 4 in order to check in and to remind
participants of the upcoming 4"1 interview. It was often at this contact that interviewers

learned that subjects had died.

68

Attrition Analyses

The next two chapters of the dissertation describe the two separate empirical
analyses that were conducted to evaluate attrition in the Family Care Study. The first
analysis (Neumark, Stommel, Given, & Given, 2001) explored the simultaneous effects
of subject and research design characteristics on early attrition Demographic and cancer
characteristics, as well as features of the study protocol, were used as predictors in a
multinomial logit regression model to enable a three-way comparison between those who
declined participation (nonparticipants, n=748), consenters who dropped out prior to
participating in the first interview (early dropouts, n=208), and consenters who
participated in the first Ell-length interview (initial participants, n=992).

In the second empirical analysis, several time-series cross-section statistical
models explored the combined data fiom all waves of the panel survey in order to
estirmte the determinants of attrition over time. Time-independent and time-dependent
sociodemographic, psychosocial, and cancer-related predictor variables representing
subject characteristics were included, in addition to characteristics of the research design
The statistical models helped identify the characteristics that differentiated subjects who
remained from those lost to the study, and also explored if there were differences
between attrition due to death and attrition due to other causes.

Although the empirical analyses are presented separately, the discussion of the
discussion and the irrrplications of the findings are incorporated in the concluding chapter

ofthe dissertation.

69

EARLY ATTRITION ANALYSIS: NONPARTICIPATION AND EARLY DROPOUT
Research Question .

With regard to early attrition in the Family Care Study, the following research
question emerges from the Ecological Model of Sample Attrition: Do research design or
subject characteristics predict who was more likely to refirse participation or to drop out
prior to the initiation of a panel survey of elderly cancer patients and Emily caregivers?
This research question focuses on the period between the subject’s identification by the
recruitment staff and telephone contact for the wave 1 intake interview. Emphasizing
nonparticipation, the analysis distinguishes between nonconsenters, consenters who
dropped out prior to data collection (early dropouts), and consenters who participated in
the first full-length interview (Wave 1 participants).

The inforrrration about sorrre subjects who were eligible but did not consent to
participate or who dropped out prior to an interview is quite scant. However, the
demographic and illness-related characteristics provided by the recruitment staff, as well
as observation of the research design, yielded valuable information. A model was
constructed to evaluate the relative importance of several Ectors in predicting who is
likely to refuse consent or to drop out of the study prior to the actual implementation of
the survey protocol (Neumark, 2001).

Defining the Target Sample

The starting point for the early attrition analysis is the target sample, defined as
the 1948 elderly cancer patients identified in the various community agencies as eligible
to participate in the Family Care Study, and who were approached for recruitment

purposes only. Signed consent forms were obtained in the community settings, and then

70

subjects were contacted by members of the interviewing stafl‘ for entry into the study.
Thus, subjects could be lost to the study because they did not consent in the first place, or
they did not follow through with the data collection efforts despite their prelinrinary
consent to do so. Of the 1948 cancer patients approached in the community-based
agencies, 748 (38.4%) were nonconsenters, and 1200 were consenters. Of the 1200
consenters, 208 dropped out prior to initiating the Wave 1 interview, thus leaving 992
cases who participated in data collection. These 992 cases represent 50.8% of the target
sample of 1948 patients.
Variables for Early Attrition Analysis

Data for this analysis were obtained from a computerized database constructed to
track patient recruitment, enrollment, and refusals. The completeness of inforrmtion
obtained, especially for the 748 nonparticipants, differed fiom measure to measure.
Nevertheless, for eight variables, data were suficiently complete to conduct a
multivariate comparison of the three groups of nonconsenters, early dropouts, and initial
participants. Table 1 details the variables available for the target sample of the 1948
eligible subjects identified and approached to participate in the Family Care Study.
Variables Pertaining to Subject Characteristics

Because this analysis is concerned with nonconsenters and early dropouts,
information on subject characteristics was limited. However, some data were available
for most cases in the target sanrple of 1948. Subject characteristics on which enough
data were available to be included in a multivariate analysis included primary cancer
diagnosis (available for 99.6% of the target sample), patient sex (99.5%), patient age

(98.9%), and patient race (93.9%).

71

Over half (53. 1%) of the target sample were males, the mean age was 73.0 years
(range 65-98), and the majority (92%) were not a minority race. The target sample was
almost equally divided among the four cancer diagnoses, with about a quarter of the
patients having breast or prostate cancer, 22.5% had colon cancer, and 28.7% had lung
cancer. While pertinent variables, such as staging of the cancer and whether or not they
were Medicaid recipients, were available for sorrre of the target sample, data were too
incomplete for nonconsenters to be included. Thus, these variables were not included in
the three-way comparisons of nonconsenters, early dropouts, and Wave 1 participants.
Variables Pertaining to Research Design Characteristics

Case Type (Sample Unit of Observation). Case type refers to the unit of data
collection, which may be an individual subject or an aggregation of two or nrore subjects
fi'om which information is collected (Babbie, 1998). Including case type in the early
attrition analysis allows comparison of cases in which a caregiver participated to those in
which only a patient participated. Information on case type was available for 96.4% of
the target sample, based on whether or not a caregiver was involved in the refirsal to
participate. In the early attrition analysis, the few caregiver only cases were grouped with
the dyad cases to consider the comparison between patients with a caregiver and patients
without a caregiver. More than halfof the target sample (56.6%) were patients with
caregivers; the remainder (43.4%) were patient only cases, meaning there was no
caregiver identified or a caregiver was not interested in participating in the study.

Recruiter Role in Health Care Environment. The FCS had three categories of
recruiters. (1) Some recruiters were hired specifically as FCS study personnel. Most often

nurses, these recruiters received their salary and benefits directly from the study grant,

72

and worked 10-20 hours per week coordinating FCS recruitment in their institutions,
seeking subjects in both inpatient and outpatient locations. (2) Other recruiters were
members of a research stafl, often with responsibility for obtaining consents fiom
patients to participate in clinical trials and other studies at the agency. These recruiters
worked variable numbers of hours each week for the FCS. (3) Finally, some recruiters
were staff nurses or advanced practice nurses who identified potential participants during
their working hours, most often on their assigned units. Often, the institution designated
nurses in several units to recruit subjects; some worked exclusively in specialized
oncology settings, while others did not.

Of the 67 recruiters, 10 (14.9%) were employed as FCS recruiters, 19 (30.6%)
were part of the research staff in their respective institutions, and 38 (56.7%) were staff
nurses. Combined, the FCS recruiters and the staff nurses were responsible for
identifying and approaching 80% of the target sample (see Table 1).

Recruiter Payment. Although many recruiters were paid directly by the FCS,
others received no remuneration for recruitment efforts. Recruiters in hospitals and
cancer care ofiices became recruiters in a variety of ways, and each recruitnrent setting
determined if and how reimbursement was offered for enrollment of participants. The
Family Care Study has a variety of methods to reimburse agencies or individuals for
recruiting patients. In some cases, hospitals and clinics (not individuals) received
payment for each case identified as participants, and were paid $10.00 when cases
enrolled in the study, and $7.00 when eligible participants refused to enroll. In other
cases, individuals received a direct payment for recruitment. Some recruiters, primarily

those who were members of a research department, submitted time cards for the precise

73

number of hours worked as recruiters, and finally, others received a pre-determined ‘/4
time to V2 time salary.

In order to evaluate what sorts of Ectors may contribute to differential attrition,
the recruiters were initially classified according to the mechanism through which
recruiters and hospitals were reimbursed for acquisition of cases to the study. The
payment plan for recruitment was entered as an independent variable in sonre preliminary
testing of the early attrition model. However, trying to characterize such strategies for
reimbursement for recruitment was not useful because strategies were so diverse, agency-
specific, and were modified throughout the course of the study. Eventually, recruiter
payment was simplified to a dichotomous variable whereby recruiters either were paid for
recruitment efi‘orts or were not paid directly. Of the 67 recruiters, 35 (52.2%) did not
receive any reimbursement tied specifically to the FCS. Ultimately, the recruiters who
were not paid identified nearly 70% of the target sample. (See Table 1).

Study Phase. In the Family Care Study, four distinct phases emerged during the
three years of subject recruitment. The first phase involved the start-up period in which
the recruitment protocol was initiated and refined and new recruitrrrent locations were
being added. This period lasted for six months fiom February to August, 1994, during
which 162 cases (8.3%) making up the target sample were identified. The second phase, a
period of full enrollment effort with established routinized procedures, lasted fi'om
September, 1994 to August, 1995 and added 40% of the target sample. At that time,
recruitment ceased for six weeks when annual review of the F CS coincided with federal
policy for stricter assurance of compliance in the documentation of informed consent in

federally funded research involving human subjects (Mark, Geddes, Salyer, & Smith,

74

1999). Specifically, recruitment efforts were temporarily suspended in order to obtain
Single Project Assurance [SPA] fi'om all collaborating institutions without Multiple
Project Assurance [MPA], as required by the Office for Protection from Research Risks
(National Institutes of Health, Office for Protection from Research Risks, 1991). When
recruitment started up again six weeks later, in October of 1995, subject accrual lost
momentum, especially at agencies where recruiters’ time was diverted to other work
responsibilities. This resumption phase lasted fiom October, 1995 to August, 1996 and
during that period, only 33% of the target sample was identified. The final phase, during
which 18.5% of the target sample was found, was fi'om September, 1996 to May, 1997.
During this period, recruitment efforts were winding down and recruiters were elirnimted
as accrual reached an adequate sample. While these recruitment periods are idiosyncratic
to this particular study, similar phases occur in other studies, and therefore this variable
was included to examine if the different phases have any effects on the success in
obtaining and retaining subjects.
Statistical Analysis of Early Attrition

In order to examine how research design and subject characteristics affect the loss
of subjects starting with the identification of eligibility, all eight variables shown in Table
1 were employed as predictors in a multinomial logit regression model (Agresti, 1990;
Long, 1997). Multinomial logit regression represents an extension of the more Emiliar
binary logistic regression (Hosmer & Lemeshow, 1989) involving comparisons of all
possible pairs of outcomes. Specifically, instead of employing separate logistic regression
models to conrpare nonconsenters to early dropouts, early dropouts to participants and

nonconsenters to participants, these comparisons are contrasts in the multinomial model.

75

Thus, in addition to obtaining adjusted odds ratios that take into account the simultaneous
influence of all predictors, this analysis produces overall goodness-of-fit indicators and
multivariate significance tests that indicate whether coefficients associated with a
particular independent variable have a simultaneous effect on the outcome across all
predictor equations. For example, a multinomial p-value of less than .05 for a subject’s
age would indicate that age has a significant effect on whether or not a subject is going to
be a nonconsenter, early dropout, or participant. The specific odd-ratios and their
associated significance levels then indicate if and to what degree age affects each two-
way contrast (dropouts-participants, nonconsenters-participants, nonconsenters-
dropouts). The analysis was conducted using the mlogit procedure of the STATA
statistical software package (StatCorp, 1999).
Results of Early Attrition Analysis

Table 2 presents the results from the multinomial logit regression model. To ease
interpretation, the table displays three columns of the adjusted odds-ratios (ORs),
comparing all two-way comparisons of outcomes. (The third column, comparing
nonconsenters to dropouts, is redundant as it represents the ratio of the other two
columns.) As the multinomial p-values in the right-hand column indicate, with the sole
exception of patient sex, every independent variable is a significant predictor of
nonparticipation. Among the subject characteristics, patient age affects nonconsent. For
every 10 years, the odds ofnot consenting rise by 29.3% (1.026'0 = 1.293, p 5 .009).
However, age does not predict dropout rates between consent and actual participation in

the initial interview (p 5 .774).

76

The reverse pattern emerges for race. While non-black minorities do not have a
different consent rate compared to white patients (p 5 .634), they have nrore than five
times greater odds of dropping out between consent and the first interview (OR=5.33, p 5
.002). Being of Afiican-American race does not increase either the risk of being a
nonconsenter or early dropout compared to whites (the reference category for race).
Finally, the patient’s primary diagnosis emerges as a strong predictor of both nonconsent
and of early dropout. Colorectal cancer patients have attrition odds roughly twice as large
(2.353 for dropping out and 1.952 for not consenting) as those of prostate cancer patients
(the reference category in the dummy coding scheme). Likewise lung cancer patients
also have greater odds of dropping out (2.317) and not consenting (1.411).

Table 2 also reveals that research design characteristics have significant effects on
attrition. Based on the Wald statistic, case type emerges as the most powerful predictor of
both consent and participation. For patients with a Emily caregiver, the odds of
participating in the first interview double, both when compared to the nonconsenters
(1/.504=l.98, p 5 .001) and to dropouts after consent (1/.419=2.39, p 5 .001). Among the
recruiter roles, staff nurses show the greatest success in obtaining subject consents. They
outdo the study’s own recruiters (OR=.467, p 5 .001), as well as agency research stafl”
(.528=.467/.885, p 5 .001). Despite higher recruitment success, patients enrolled by staff
nurses have no greater odds of dropping out after consenting. Overall predicted
probabilities of initial participation, adjusted for all variables in the equation, are 58.2%
for staff RNs, 48.5% for FCS recruiters and 47.7% for research personnel. Paid recruiters
are noticeably nrore successful in getting subjects to consent to participate: the odds of

not consenting versus participating are only halfthose (.557, p 5 .001) among unpaid

77

recruiters. However, consenting patients of paid recruiters have 66% higher odds of
dropping out after consent (OR = 1.659, p 5 .022), leading to a neutral overall effect on
initiating study participation (.924 = 1.659 x .557). Finally, the ORs associated with the
study phases suggest a monotone increase in nonconsenters, but not dropouts, as the
study progressed over the three years.

Brief Discussion

An important consideration in an attrition analysis is whether people who
participate are different from those who do not. In this multivariate analysis, early
attrition was examined by focusing on both nonconsenters and consenters who dropped
out prior to initial data collection. This retrospective analysis was restricted to the details
included on the recruitment eligibility forms. Additional pertinent demographic and
social variables, such as marital status and living arrangements, and nrore complete
design variables, such as actual costs incurred and number of recruitment tries, would
strengthen the model.

Findings suggest that both subject and research design characteristics afl'ect the
likelihood of early attrition fiom the Family Care Study prior to the actual
implementation of the survey protocol. Age and cancer diagnosis played roles in
whether consent was obtained, whereas race did not aflea consent but raised odds of
drop out after consent. Powerful evidence emerged that design features, such as if a
caregiver participated, recruitment roles, and payment to recruiters, affected the
probability of subjects not participating or dropping out before being interviewed. Further

discussion and implications of these findings are included in Chapter 7.

78

ATTRITION OVER TIME:
DROPOUT OVER THE FOUR WAVES OF THE FAMILY CARE STUDY

The next goal of the dissertation is to extend the analysis of attrition to the entire
length of the F anrily Care Study to compare those who remain in the study for one year
following cancer diagnosis to those who withdraw. With regard to attrition over time, the
following two research questions are asked: 1) Do research design or subject
characteristics predict who is more likely to drop out over time, for whatever reason,
fi'om a panel survey of elderly cancer patients and family caregivers? and 2) Do difi‘erent
characteristics predict attrition due to death versus attrition for reasons other than death?

Defining the Target Sample

In order to evaluate attrition over time, only those who participated in the Wave 1
data collection are included. As described in the early attrition analysis, FCS subjects
signed consent forms to participate in community settings, and when consent forms were
received at the university, the consenters were contacted by telephone for the first
interview. Of the 1200 consenting cases, 195 (16.3%) dropped out prior to initiating the
first interview (early dropouts). Thus, the cases who participated in the intake interview
(n=1005) constitute the target sample for the empirical analysis of predictors of attrition
over time in the Family Care Study. It is worthwhile to bear in mind that, although
counted as a single case, and treated as a unit in this analysis, patients and caregiver
dyads were not interviewed together, so overall, the 1005 cases include 1640 individuals
(981 elderly cancer patients and 659 Emily caregivers).

[Please note that due to some discrepancies between the computerized attrition

tracking database and the interview and questionnaire database, which was not used in

79

the early attrition analysis, the number of cases who initiated the intake interview, as
. described in the earlier analysis, is slightly diflemnt in this analysis of attrition over
time.]

Statistical Analysis of A ttrition Over Time

The very concept of attrition implies longitudinal follow-up. In the Family Care
Study, attrition is defined in terms of study participation status at three discrete follow-up
times: a case’s dropping out of the study at waves 2, 3 and 4 of the panel study. Given
the panel nature of the data, the statistical methods employed must take account of the
fact that repeated observations on the same subjects, taken at different waves, are not
independent of each other. Thus, statistical models for cross-sectional analyses are not
appropriate, since they would treat each observation, whether fi'om a different case or
from a different wave, as an independent observation. Instead, in the current analysis,
pooled time-series cross-section models (TSCS models) are employed to estimate
attrition efl‘ects (Diggle, Liang & Zeger, 1994).

These models offer three important advantages. (1) In pooling data across all of
the waves rather than analyzing each wave separately, the effects of time-dependent or
time-independent variables can be estimated more precisely, assuming that these effects
are constant across the three observations. (2) This approach allows the estimation of the
effect of time on attrition, by capturing variation between subjects as well as variations
within subjects fiom wave to wave. Specifically, the pooled times series cross-section
model will be used to identify both case and research design characteristics that predict,
which cases remain or are lost to the study. In addition, the model allows for the

possibility that the same predictor variable measured at different waves can have different

80

effects on attrition in the following wave. (3) Third, pooled time-series cross-section
models can easily lEndle variation in the number of waves, a feature that is essential for
the analysis of attrition patterns. This means that subjects, who contributed data to some
waves but not to other waves, can be included, and the tirrre interval between the data
collection waves does not have to be equal (Johnson, 1995). Thus the analysis can take
advantage of all available information in the data.

While pooling observations increases the effective sample size, and hence the
precision of the estimates, if unobservables (captured in the residual) are correlated
across observations on the same individual, the standard errors fi'om the pooled data will
be understated, which means that the significance ofthe results will be overstated. To
address this problem, a random effects estimator is used (Diggle et al., 1994). This
estimator incorporates a subject-specific component of the residual that is common to
observations on the same subject across waves, and therefore yields correct standard
errors.

The panel analysis is preceded by some descriptive procedures to look at bivariate
relationships between attrition and its various predictors at each wave of data collection.
Statistical significance for continuous variables are estinrated by independent t-tests for
equality of means; whereas, the significance of proportional differences between those on
study and those ofl‘ study is determined by chi-square analyses for the categorical
variables. Although bivariate models do not consider the more complex multivariate
relationships between attrition and its predictors, they are useful in revealing attrition
patterns in the Family Care Study data, providing a sense of how these patterns are

associated with subject and research design characteristics over time, and identifying the

81

relationships which need to be explored more thoroughly in the multivariate models. For
example, in the bivariate analysis, minority status and attrition are strongly correlated;
however because income and race are also strongly correlated, it is important to discern
whether minority status has an effect on attrition independent of income.
Variables for Attrition Over Time Analysis

The dependent variable in the first part of this analysis is attrition over time,
referring to dropout, for whatever reason, fi'om the target sample of cases for which data
collection was initiated in the Family Care Study (n=1005). Time-independent and time-
dependent sociodemographic, psychosocial, and cancer-related predictor variables
representing subject characteristics are included as independent variables in the analysis
of attrition over time. Conceptually emerging fi'om the ecological model, subject
characteristics include demographic and socioeconomic variables, Emily structure,
physical and psychosocial functioning, and cancer-related variables. Unfortunately,
because the number of caregivers participating is fewer than the number of cases
available for analysis, variables pertaining to caregivers are not included in this analysis.
Research design predictors capture the study protocol features of the ecological model,
including sampling unit, data source, and wave time. Figure 4 contains the
operationalization of the potential independent variables available in the Family Care
Study and their relationship to the theoretical components as described by the Ecological
Model of Sample Attrition. It includes variables fiom the Family Care Study that were
considered for the analyses of sample attrition. The figure includes some variables that
were part of the databases; however, sonre data were too incomplete or too unreliable to

be included in the final multivariate statistical models.

82

The next section describes the measures and items that were evaluated as
predictors of attrition over time. For some of the subject characteristics, there is
descriptive informtion, in order to help the reader understand the makeup of the target
sample, but these variables were ultimately not included in the multivariate models. Some
of the design variables, which are theorized as being related to attrition, such as
consistency of interviewer and protocol adherence, were not available for analysis in the
pooled data set. However, they were evaluated in a cross-sectional manner and are
discussed below.

The data used came from the interview and questionnaire responses, the medical
record audits (staging of cancer information), as well as the database used to track subject
retention and attrition over time. In the Family Care Study, there were up to four
observations (= waves) for each case, at each of which cases could remain on study or
drop out. However, since the goal is to predict attrition status, which can only be
ascertained at the wave following the wave at which subject and design characteristics
are gathered, the analysis, in effect, includes within-subject data from up to three discrete
time intervals. Each time interval comprises independent variables, obtained at waves 1,
2 or 3 coupled with the attrition status (dependent variable) ascertained at the relevant
succeeding wave (fiom waves 2,3, or 4).

Variables Pertaining to Subject Characteristics

Demographic and Socioeconomic Characteristics. Demographic variables
include patient sex, patient age, patient race, patient education, marital status, household
income, and whether the patient lives alone or with others. The target sample (n=1005)

was comprised of 551 men (54.8%) and 454 women (45.2%), primarily Caucasian

83

(92.3%), with a mean age of 72.6 years (SD=5.3 years). The majority of the sample who
participated in initial data collection were married (66.4%) or widowed (25.4%), and
77% lived with their spouses or other people. A small number of patients (8.2%) reported
less than a high school education, and about 40% had attended or graduated from college.
Most patients were retired (78.2%), and the median household income of the target
sample was $22,500.00 (mean income $27,982.00, range 4K-100K, SD=$17,776).

Cancer-Related Variables. Information on cancer diagnosis was obtained at the
time of patient enrollment into the Family Care Study. Of the 1005 patients in the target
sample, 283 (28.2%) had lung cancer, 279 (27.8%) had prostate cancer, 255 (25.4%) had
breast cancer, and 188 (18.7%) had cancer of the colon. Medical records of these patients
revealed nearly 30% of the sample had advanced cancer (stage 3 and stage 4), whereas
61.6% had cancer in an earlier stage. (Cancer stage information is not available for 9.9%
of the sample). Around the time of enrollment into the study, about 60% of the patients
had had surgery for cancer, 18% were undergoing chemotherapy, and 46.5% were having
chemotherapy. (Some patients had multiple treatment modalities.) Additionally, over
85% of the subjects reported at least one nredical condition besides the cancer diagnosis,
and nearly half indicated three or more comorbidities.

Symptom Experience. The symptom scale (Given & Given, 1991 -l 996) sought
information about 31 symptoms associated with cancer or with cancer treatment. Patients
were asked if they have experienced each symptom in the past two weeks (yes/no). If
“yes,” they were asked how severe this symptom was (mild, moderate, severe). The
symptom inforrmtion fiom the Family Care Study was aggregated into two variables: a

simple count of the symptoms reported by each patient and a summated severity score

84

which weights each reported symptom based on whether it was perceived to be ‘mild’
(=1), ‘moderate’ (=2) or ‘severe’ (=3). Questions about symptoms were answered only by
patients, so there are no symptom data if a patient did not complete the interview portion
of the protocol. At the first wave of data collection, the mean symptom count was 8.14
(range 0-27, SD = 4.61); and the mean symptom severity score was 12.86 (range 0-72,
SD = 9.30). As could be expected, these two measures are highly correlated (r = .935 at
Wave 1, r = .932 at Wave 2, and r = .946 at wave 3).

Physical and Emotional Functioning. Measures of physical and emotional
functioning are from the Medical Outcomes Study Short Form, SF-36, which was
incorporated in the patient telephone interview in the Family Care Study. The SF-36 is a
health survey for adults with chronic conditions and varying socioeconomic states which
was originally constructed to provide a comprehensive assessment of physical and mental
components of health status (Ware & Sherboume, 1992). The instrument, which has well
established psychometric properties, consists of 36 items representing eight generic
health concepts including physical firnctioning, role disability due to physical health,
bodily pain, general health perceptions, vitality, social firnctioning, role disability due to
emotional health problems, and general mental health (McHomey, 1996; McHomey,
Ware, & Raczek, 1993). The eight subscales were computed in the F CS data set fiom the
36 items using a scoring manual provided by the Medical Outcomes Study. Higher scores
on the subscales (potentially ranging from 0 to 100) indicate better functioning.

Table 5 contains the scores for each of the subscales across the waves of the
Family Care Study. At the initial wave of data collection, the lowest mean scores were

for the vitality subscale and the (reversed) role restrictions due to physical health

85

subscale. By contrast, the highest mean scores were observed on the (reversed scored)
role restrictions due to emotional health problems and general mental health subscales.
Variables Pertaining to Research Design Characteristics

Case Type (Sample Unit of Observation). As described in the early attrition
analysis, the Family Care Study recruitment protocol permitted patients to enroll, with or
without a Emily caregiver, and the case type was based on who signed the consent forms:
cancer patients only, dyads of cancer patients and their designated Emily caregivers, or
caregivers only. Including case type in the attrition over time analysis allows comparison
of cases in which a caregiver participated to those in which only a patient participated. In
this analysis, the few ‘caregiver only’ cases are grouped with the dyad cases to consider
the comparison between patients who participated with a caregiver and patients who
participated without a caregiver. In the target sample (n=1005), 659 cases (65.6%) were
patients with caregivers participating, and 346 (34.4%) were patient only cases, meaning
there was no caregiver identified or a caregiver was not interested in participating in the
study.

Data source/Method of Data Collection. In the management of the protocol, in
order to maintain a longitudinal sample, the investigators decided that a case could be
moved on to the next wave if any portion of either the patient or caregiver interview or
questionnaire were completed in the previous wave and the participant(s) agreed to be
contacted again. As a research design characteristic in this analysis, two variables were
utilized to investigate source of data (e.g., telephone interview or questionnaire), which
were fully or partially completed at the previous wave, had an effect on dropout at the

next wave. The first variable indicated whether a telephone interview (either a patient

86

interview or a caregiver interview) was done at the previous wave (yes/no); the second
variable summarized if any self-administered questionnaire (either a patient questionnaire
or a caregiver questionnaire) was completed and returned at the previous wave (yes/no).
Among cases who remained on study through Wave 2, 99.5% participated in a wave one
patient or caregiver interview and only 4 cases (0.5%) did not complete the interview
portion of the design. On the other hand, anrong the cases participating in a Wave 1
interview, only 94.9% returned at least one of the mailed questionnaires; thus, 45 cases
(5.1%) did not rmil back either the patient or the caregiver self-administered
questionnaire.

Completeness of data collection/Item nonresponse. The F CS interview and
questionnaire compiled multiple measures to assess the physical, emotional, and financial
outcomes of a cancer diagnosis on older individuals and Emily caregivers. Several
versions of the interview and questionnaire instruments were available in order to
accormnodate the different configurations of case types, and because certain portions of
the interview (such as household financial information) could be obtained fiom either a
patient or a caregiver.

Typically, in order to assure standardization of the research protocol, interviewers
are trained to ask questions in the specified sequence, and to follow appropriate skip
patterns; however, the FCS design allowed interviewers to complete a shortened version
in order to accommodate participants, who could not or would not complete the entire
tool. For example, in order to achieve the goals of determining costs associated with
cancer in this population, the investigators requested that interviewers try to gather data

pertaining to these issues fiom all participants, even if other portions of the interview

87

were omitted. In the full length interview, due to the sensitive nature of these items,
financial questions were asked towards the end. A good part of the data was obtained
from either the patient or the caregiver (e. g., sociodemographic inforrmtion, service
utilization), but other items were asked exclusively of patients (e. g., symptom
experience), or of participating caregivers only (e.g., caregiver symptom assistance).
Unfortunately, inforrmtion on who answered the specific questions was not captured at
the time the data were entered into the computer, nor is it clear whether data are missing
because a shortened interview was done or because the participant simply did not want to
answer certain questions.

In order to examine the effect of completeness of data collection in a previous
wave on attrition at a subsequent wave, the degree of response to the various interview
sections was captured through specially created “non-response variables”. Since
caregiver data are not available for patient-only cases, and since a lot of data from the
self-administered questionnaires were missing, only the interview sections that are part of
the patient interview were evaluated. The variables were operationalized by looking at
the degree of response (no response, partial response, complete response) to the interview
sections. Complete response indicates that all items in that section were answered. For
example, at the first interview, the majority of patients (87.4%) fully answered
sociodemographic items, 2.1% partially responded, and 10.4% had no response to this
section of the interview.

Consistency of interviewer. The review of the literature indicated that consistency
of interviewer over time may be an important Ector related to attrition in samples of

older individuals. .To explore whether this may be associated with attrition in the Family

88

Care Study, this variable was created fiom the tracking database where interviewers’
initials were entered at each wave of data collection. Unfortunately, because of how the
management tracking database was deve10ped, in addition to the complicated research
design, information about interviewers is often missing or inconsistent, and this variable
could not be included in the pooled data. However, information about interviewers was
complete on 930 subjects in the target sample. Of these, 78.4% had the sanre interviewer
contact them by telephone at Wave 1 and Wave 2, and 22% had different interviewers.
Interviewer consistency was significant only at the third wave of data collection when
having more than one interviewer across waves of participation was associated with less
attrition.

Wave Time. The longitudinal nature of the panel survey design and the notion of
time in the ecological model is represented by the wave time variable. As mentioned
above, although there are four observations times or interview waves in the F arnily Care
Study, the analysis starts with all cases who participated in wave one and evaluates which
of the selected variables, measured at waves 1,2 and 3, are predictors of attrition at the
subsequent waves (2,3 and 4). Although there may be missing data fiom the sample of
cases who were still on at the start of the firm] interview, these Ectors are not useful in
predicting attrition over time.

Results of the Attrition Over Time Analyses

The overall attrition status of cases over time is found in Table 3. The data
clearly show that the proportions of dropouts decreased at each successive wave, a
finding tlnt appears even more pronounced, when one considers that the average time

interval between successive waves increased, from 6 weeks between waves one and two,

89

over three months between waves two and three, to 6 months between wave three and
four. At Wave 2, 128 cases, representing 12.7% of the target sample dropped out; at
Wave 3, the dropout rate was 10.5% of the remaining sample (92 cases out of 877); and
at Wave 4, attrition accounted for 10.3% of the remaining sample (81 cases out of 785).
Bivariate Analysis

Table 4 and Table 5 display the results of the bivariate analyses of subject
characteristics and attrition over time, referring to dropout, for whatever reason, from the
target sample of cases who initiated data collection in the Family Care Study (n=1005).
The variables shown in Table 4 are the time-independent subject characteristics.
Although in principle patient marital status and patient living an'angements may vary
over the course of the study, except for 3-4 cases, these characteristics remained stable
among sample cases and, thus, are not treated as time-dependent covariates. The
variables in Table 5 are the time-dependent measures that could, and often did, change at
each wave of data collection.

As Table 4 shows, cancer diagnosis and cancer stage were significantly associated
with attrition over time. It is not surprising that lung cancer patients had the highest
dropout rates at all waves, with only 42% of the original 283 cases left at Wave 4. Some
demographic and socioeconomic indicators were associated with higher attrition,
although the specific variables differ across the waves. Those who dropped out were
slightly older at Wave 2 and had lower incomes at Wave 3. Overall, men were more
likely to drop out of the study at Wave 3 and Wave 4; but marital status, and whether the

patient lived with someone were not significant at any wave. At Wave 2 and Wave 3, the

90

patients who were minority and less educated had higher rates of dropping out, although
neither Ector was significant at Wave 4.

Without exceptions, the time-dependent variables displayed on Table 5, are
significant predictors of dropout at waves 3 and 4, but not wave 2. Greater reported
numbers of symptoms and symptom severity, and lower mean scores on the SF-36,
indicating worse emotional and physical firnctioning at the wave prior to attrition are
consistent predictors of dropping out.

Table 6 displays the results of the bivariate analyses relating research design
characteristics to attrition over time. Except for “case type,” which was designated at the
time of participant(s) consent, the research design predictors may vary over time.
Significant differences in case type were found only at Wave 3, where patients with a
caregiver participating tended to drop out more than the patient-only cases. The sources
of data collection at a previous wave were usefirl to distinguish between cases remaining
on the study and dropouts. In particular, not returning a mailed questionnaire was
associated with greater subsequent attrition, although the completion of a telephone
interview did not predict attrition at the next wave.

Time-Series Cross-Section Analysis of A ttrition Over Time

Results fiom three random-efl‘ects pooled time-series analyses are displayed in
Table 7. The first model is the full model that includes most of the subject and research
design variables that emerged as significant at one or more wave in the bivariate analysis.
However, some of the variables, such as consistency of interviewer, were not available
for the pooled dataset; others were eliminated because of substantial multi-collinearity

among the predictors. In particular, while all of the time-dependent variables fi'om the

91

SF—36 subscales were significant in differentiating subsequent dropouts fi'om subsequent
. participants in the bivariate analyses described above, this information is, to a large
degree, redundant, given the high correlations among the SF-36 subscales. Thus, only
the general health, physical functioning and emotional functioning subscales were
retained in the multivariate model. Finally, the first two models contain some common
demographic and socioeconomic indicators, such as age, race, marital status and social
support (living alone), that arose fi'om the review of literature and are conceptualized in
EMSA, even if they are not necessarily significant in the bivariate analysis.

In the second and third models displayed in Table 7, several predictor variables
have been dropped, in an effort to arrive at a more parsimonious specification. In
addition, one of the predictor variables in Model 1, the failure to return the self-
administered questionnaire, may not be considered a fully independent predictor, but
rather already a part of the outcome. Thus, the strong association between attrition and
Eilure to return the questionnaire in Model 1 (OR=.107, p S .001) may be more a
reflection of dropping out behavior than an exogenous predictor of dropping out. Model 2
retains variables only if they meet one of two criteria: they either display a significant
effect on the odds of dropping out or they have been identified in the literature as
important determinants of attrition. Model 3 takes this one step further and simply drops
all variables that do not have a significant efi‘ect on the odds ratio, even if a role for them
has been suggested in past work. While time as a feature of the research design is
removed, additional models, looking at interaction effects with time (not on the table)

were tested and are described below.

92

Although the fill] results of the attrition over time analysis are provided in Table
7, the discussion of significant effects focuses primarily on the final, most parsimonious
model (Model 3). As seen in the last 2 columns of the table, demographic and
socioeconomic variables with significant effects on attrition included age, income, and
widowhood status. People with lower incomes and older people were more likely to drop
out. For every 10 years of additional age, the odds of attrition over time increased by
35.7% (1.03110 = 1.357, p S .000). Similarly, the odds of attrition are 10.5% lower for
each $10,000 increment in income (.98910 = .395, p s .0042). Although the joint
significance tests across the multiple categories of marital status is not significant at the
.05 level, surprisingly, widowed persons in the sample had relatively lower odds of
attrition than married people.

The patient’s cancer diagnosis and stage of cancer were extremely strong
predictors of attrition over time, for both the individual categories as well as the multiple
category predictors as indicated by the joint significance tests. The odds of attrition were
highest for those with lung cancer, followed by colon cancer, prostate cancer, and breast
cancer. The patients with early stage disease had about halfthe odds of dropping out,
compared to those with late stage disease (OR=0.451, p _<_ .000).

Another illness characteristic, the time-dependent variable measuring perceptions
of general health, was robust in predicting attrition in this sample. Because the units of
the subscale are arbitrary (a 0—100 scale), it is best to interpret the estimated odd ratio in
light of a one standard deviation change in the variable. (This is equivalent to reporting
the coefficient of the subscale score divided by its standard deviation). Since the standard

deviation of this variable is 23, the estimated odds ratio implies that a subject with a score

93

that is one standard deviation higher (=better self-rated health) than another subject’s has
.67 lower odds of dropping out (.983 23 = 0.674 ).

The last set of variables in the initial model is research design predictors including
case type, wave time, and return of the questionnaire. As discussed above, the last of
these predictors was dropped from Models 2 and 3. In addition, time (in Model 2) and
case type (in Model 3) were not revealed as significant predictors of attrition over time.

Finally, in order to gain a clearer picture of the differential effects of time on
attrition for patients with different primary cancer diagnoses, a statistical model was
tested which included the interactions between wave time and cancer diagnosis, in
addition to the Mode12 variables in Table 7. The results of the interaction model are
displayed on Table 8. In this model, whose interaction eflects are graphically depicted in
Figure 5, the odds of attrition for those with lung cancer increases with each successive
wave, while remaining essentially stable or declining for the other cancer diagnoses.
Particularly for breast cancer patients, the odds of attrition continued to decline over time,
while, for prostate cancer patients there was a slight increase in odds of attrition at wave
4. It is also notable that in the model incorporating the interactions between primary
cancer diagnosis and wave time, the research design variables, time and case type, were
once again significant predictors of attrition. The odds of dropping out were 1.4 higher
for patient only cases compared with cases in which a caregiver also participated.

Comparing Attrition Outcomes: Attrition For Other Reasons Versus Death Attrition

The dependent variable in all of the above analyses is attrition over time, referring

to dropout, for whatever reason, fiom the target sample of cases who participated in the

wave one data collection of the Family Care Study (n=1005). However, the second

94

research question refers to differences in predictors of attrition due to death as compared
with attrition due to all other causes. .

Table 9 compares the specific attrition outcomes. Over the course of the study,
fiom the target sample of 1005 cases, 133 (5% of the target sample) died while they were
still participating in the study, and 168 (6.3%) dropped out for reasons other than death.
At the second wave, those who dropped out due to death represent a much smaller
proportion (2.4%) of the sample at that wave, than the proportions who dropped fi'om the
study due to death at either Wave 3 (6.3%) or at Wave 4 (6.9%). Conversely, attrition due
to causes other than death, diminishes in magnitude fi'om Wave 2 to Wave 4. At Wave 2,
81.3% of those who dropped out did so for reasons other than death. This decreases to
40% of the dropouts at Wave 3 and 33% of the dropouts at Wave 4.

Thus, in order to garner a clearer picture of predictors of attrition in this sample of
older families with cancer, the analysis has been supplemented by time-series cross-
sectional models that differentiate among the two types of attritions: death or attrition for
other reasons. Subjects were counted in the category “attrition due to dea ”, if a case
participated in the study at a given wave and the patient’s documented date of death
occurred prior to the next attempted contact by the interviewing stafl‘. Even though
patients may have died later, if cases dropped fi'om the study when the patient was still
alive they were counted as dropouts “for reasons other than death”.

Of those who died after being on the study at the first wave (n=24), 79.2% had
lung cancer, and 16.7% had colon cancer. The majority were men (62.5%), in dyad cases
(79.2%), and all were Caucasian. Of those who dropped out for other reasons after Wave

1 (n=104), 54.8% were male, 16% were minorities, and 53.9% were in dyad cases.

95

Although more lung cancer patients withdrew than patients with any other diagnosis,
attrition statistics for diagnosis were Eirly similar to the distribution of the target sample.

Attrition after the second wave included 37 cases who dropped out for reasons
other tlmn death, but 55 who died in that time period. Again, the patients who died were
primarily nrale (69.1%), with lung cancer (85.5%) and in dyad cases (89.1%). Only
85.4% were Caucasian, and 11% had early stage cancer. Fifty-four nrore patients died
between Wave 3 and Wave 4, the first time there were deaths among those with breast
cancer.
Attrition for Reasons Other than Death

In order to examine predictors of attrition over time for reasons other than death,
Table 10 shows the results fi'om a pooled time-series analysis that excluded the 133
patients, who died while still on study. Thus, the dependent variable in this analysis
refers to the odds of dropping out for reasons other than death versus remaining on study.
While all of the intermediary steps are not specified here, the steps to achieve a more
parsimonious model were similar to the procedures described previously. The resulting
Model 2 indicates that other than cancer diagnosis (where only the joint test is significant
at the .05 level), subject characteristics are not predictive of dropout for reasons other
than death. Instead, it is the research design variables, including time and case type,
which have significant effects on attrition. Patients who participated ill the study, but did
not have a caregiver participating had dropout odds nearly twice as large as cases that
included a participating caregiver (OR = 1.89, p S .002.) The wave time variable specifies
that the odds of dropping out due to reasons other than death decrease in likelihood over

time. Odds of dropout at Wave 3 are less than half(.49) the odds of dropout at Wave 2,

96

and they continue to decrease at Wave 4. The implications of these findings are
addressed in Chapter 7.
Attrition due to Death

Finally, Table 11 displays three models, parallel to those in Table 7, to explore
predictors of attrition over time due to death while on study. This analysis again used the
pooled data, but this time, excluded the 168 subjects, who dropped out over time but did
not die during the period they were participating on the studyfor other reasons. Here,
the results in Model 3 reveal subject characteristics, particularly cancer diagnosis and
cancer stage are significantly related to attrition over time. Other characteristics, like
increasing age also approach significance, particularly when controlling for sex and race.
Two research design characteristics, wave time and case type, remain significant
predictors of attrition due to death The odds of attrition due to death are over ten times
greater at Wave 4, and still five times greater at Wave 3, when compared to wave 2. Case
type also has a significant effect on the outcome, although now, attrition due to death is
halfas likely in the patient only cases as compared with dyads (OR= .456, p _<_ .031).

Brief Discussion

This analysis confirmed that both subject characteristics and research design are
important predictors of attrition over time in a panel survey of older people with cancer
and Emily caregivers. It is impossible in this retrospective analysis to evaluate all the
Ectors involved; nonetheless, there is evidence that multiple interacting Ectors operate to
influence who drops out and why. Given that lung cancer patients typically have worse

prognoses and higher staging at diagnosis conrparcd to the other cancers, it was not

97

surprising that those with lung cancer were most likely to drop out of the study, whether
due to death or due to other reasons. '

The Ecological Model proposes that sample attrition is multileveled and subject to
numerous influences, both within and outside of the research study. Furthermore,
different patterns emerged over time. In particular, research design contrrhuted to
patterns of attrition in the earlier waves, especially in cases of attrition for reasons other
than death. On the other hand, subject characteristics emerged as more likely indicators
of attrition as the study progressed. Time is an important Ector in working with a sample
of older people, whether because of disease, increasing age, or increasing functional or
psychological impairments that may affect ongoing participation in a panel study.

Further discussion of these findings are found in the concluding chapter.

98

CONCLUSIONS AND IMPLICATIONS

This dissertation aims to enhance understanding of the complexities associated
with sample attrition in panel surveys of older individuals with chronic illness and their
Emilies. This objective is achieved in several ways. First, this dissertation specifies and
clarifies the terminology with respect to sample attrition these studies. Second, the
dissertation addresses conceptual and methodological issues associated with sample
attrition in health care research with Emilies. Third, this project incorporates an
ecological perspective by establishing and expanding the Ecological Model of Sample
Attrition. Finally, the multivariate analyses provide empirical evidence that both research
design and subject characteristics are useful in predicting who is more likely to refuse
participation, dropout prior to the initiation of data collection, or drop out over time from
a panel sm'vey of elderly cancer patients and Emily caregivers. This chapter discusses
the results, implications, and limitations of the statistical analyses that were presented in
Chapters 5 and 6. Quantitative and qualitative examples from the retrospective case
study are provided in order to illustrate the nuances and multiple layers of the ecological
model, the Emily characteristics, and the use of consistent language. Additionally, this
chapter contains suggestions for further research on the complex and inevitable
phenomenon of sample attrition in Emily health care panel studies.

Patterns and Predictors of A ttrition in the Family Care Study

As hypothesized, the findings revealed tlurt both subject characteristics and the
research design were influential in contributing to early attrition and attrition over time in
the Family Care Study. Although the Family Care Study had less than the projected

sample size of 800 surviving to the fourth wave of data collection, the data clearly show

99

that the proportion of dropouts decreased at each successive wave (Table 9). This finding
is noteworthy when one considers that the average time interval between successive
waves increased over the year of the study: there were 6 weeks between waves one and
two, over three months between waves two and three, and 6 months between wave three
and four. In the FCS, once cases made it past the second wave of data collection, attrition
was much less frequent, other than when it was due to death.

Specifically, the analysis helped distinguish that when attrition was due to reasons
other than death, research design characteristics were nrore likely to be associated with
attrition over time (Table 10). Accordingly, the effects of the research design
characteristics diminished, particularly in the later waves of the study, when the odds of
dropping out due to reasons other than death decreased (Table 11). Clearly, in panel
studies of older patients with serious illnesses like cancer, an unavoidable source of
attrition is the patients’ physical morbidity and mortality. While having little control over
age or health-related reasons that contribute to subjects’ reluctance to participate or to
their subsequent attrition, researchers do have control over study design and data
collection techniques (Arean & Gallagher-Thompson, 1996; Eaves, 1999; Motzer et al.,
1997; Sullivan et al., 1996). Predicting who is prone to decline participation or to drop
out, and why, may help researchers target recruitment and retention efforts, particularly
to reduce the extent to which study-related Ectors contribute to attrition.

Subject and Research Design Characteristics and Sample Attrition in the FCS

The cancer diagnosis was a strong predictor of both early attrition and attrition

over time. Given that lung and colorectal cancers typically have worse prognoses and

higher staging at diagnosis compared to breast and prostate cancers, it was not surprising

100

that patients with lung or colorectal cancers were twice as likely to drop out before
initiating a time-demanding panel study. Over time, the odds of attrition were lowest for
those with breast cancer, followed by prostate cancer, and colon cancer. As shown in the
diagnosis-time interaction models, at each wave of the study, those with lung cancer were
most likely to drop out, and these dropout rates accelerated through all phases of the
study (Table 8 and Figure 5). Furthermore, symptoms related to cancer or cancer
treatment can impede participation. Some of the common symptoms reported were mouth
sores, diarrhea, nausea, weakness, shortness of breath and fatigue, all of which may make
it difficult or impossrhle to complete lengthy telephone interviews or questionnaires.

In a panel study of elderly people with cancer, attrition due to death is an expected
outcome. In 1994, when subject accrual began for the Family Care Study, the one-year
survival rates for patients with the four primary diagnoses relevant to this study were as
follows: breast 97.0%, colorectal 83.1%, lung 38.1%, and prostate 97.9% (National
Cancer Institute, 1994). Additionally, although not differentiated in the early attrition
analysis, 13 patients died between the time of consent and when they were contacted for
an interview. While Mihelic and Crimmins (1997) assert that attrition due to patient
death is considered a "legitimate" outcome since it reflects the natural history of the
disease, the fact that it diminishes sample size is still important in the planning and
implementation of panel studies.

In addition to illness, the results confirmed that older age played a major role in
the likelihood of obtaining consent and maintaining the FCS sample. Older age affected
nonparticipation and remained an important predictor of attrition over time. Although in

the bivariate analysis age effects were found only at Wave 2 (Table 6), when controlling

101

for other subject characteristics, including sex and minority race, old age remained a
consistent predictor of higher odds of attrition (Table 7 Model 3). The patterns were
different for early dropout, which was not associated with older age. Finally, the
literature indicates weaker social networks are associated with higher attrition, yet
proxies for weak social networks such as living alone or not being married did not predict
attrition.

The notion of time and the dynamic nature of sample attrition that came from the
review of the literature and are conceptualized in EMSA, were tested by inclusion of
time-dependent measures in the statistical analyses. The results suggest that especially
variables measured one wave before the attrition outcome surface as significant
predictors in differentiating between those remaining on study and those dropping out
after Wave 2. On average, those who dropped out of the study at Waves 3 and 4 had
more severe symptoms and worse emotional and physical functioning at the previous
wave of data collection than those who did not drop out at the next wave (Table 6).
Similar significant differences in mean scores were not prevalent at Wave 2. This
evidence supports the conceptual (as well as intuitive) notion that subject characteristics
remain strong indicators of sample attrition over time.

The analyses of early attrition yielded powerful evidence that, in addition to
subject variables, research design characteristics independently influenced sample
attrition. A predominant question in research involving Emilies is whether Emily
support hinders or enhances participation in a panel study. In the Family Care Study,
having a caregiver involved was a robust indicator of study enrollment and initial

participation. This result substantiates the importance of the family environment and

102

confirms the importance of mobilizing a social support network for recruitment (Hooks
et al., 1988; Eaves, 1999). On the other hand, while having a caregiver available
generally enhanced participation, occasionally it was the caregiver who refirsed to
consent on behalf of the Emily member with cancer.

The interaction of the subject (and the Emily) with the health care environment is
a relevant issue related to sample attrition since the patients are typically undergoing
treatment, often requiring daily visits to cancer care Ecilities. In fact, the Family Care
Study had interview questions about service use that probed the time and effort Emilies
devoted to undergoing conventional as well as alternative treatments for cancer. Results
indicated that patients had multiple visits with primary care physicians, oncologists, and
radiation specialists, and many required multiple hospitalizations due to their cancer
(F CS, 1996).

The following observations fi'om the FCS illustrate the application of the
Ecological Model to reveal some of the different and multiple Ectors that may contribute
to sample attrition in all phases of a panel survey. For example, minority race (subject
characteristic, social environment, family environment) is often singled out as a predictor
of attrition in panel surveys. However, this characteristic did not have an effect on
attrition over time in the FCS, particularly after controlling for income. According to the
bivariate analysis, minority patients were more likely than white patients to drop out over
time, but due to the small number of minorities remaining in the sample compared with
the number of whites (research design, social environment, health care environment),

this effect disappeared entirely in the multivariate models.

103

An interesting finding fi'om the early attrition analysis was that there were
extremely large odds of dropping out after consent among non-black minorities (Table 2).
It is possible that the social environment and the family environment influenced this
phenomenon. Perhaps there were language barriers or cultural beliefs that increased the
likelihood of the non-black minorities being early dropouts. Ahnost all were Hispanic
patients from Saginaw, Michigan; two were Native Americans. While this finding was
based on a small sub-sample, it warrants further research about mechanisms of early
attrition, particularly among minorities. Studies with other target samples are needed to
test and evaluate research design features to reduce nonparticipation, early dropout, and
attrition over time. In a discussion of participant recruitment in research among
vulnerable Emilies, Demi & Warren (1995) described the phenomenon of subjects
agreeing to participate in studies when approached, due to their deference to authority
figures, or sometimes, due to their strong need to talk to someone. They found that
although some people did not overtly refuse to participate, they were subsequently
unavailable for fo llow-up.

Another Ector contributing to the early dropout may be the recruiter who
recruited the majority of subjects from Saginaw. The recruiter, a nurse, was hired as an
F CS recruiter and was paid a half-time salary to contact multiple clinical sites to access
subjects, verify eligibility, and obtain consent (or refusal) (research design, recruiter
role, recruiter payment, health care environment, /social environment). She was a “ ”
recruiter, enthusiastic, and thorough, and she enrolled a large number of subjects after
joining the staff nearly the sanre week as the recruitment coordinator (September, 1994).

The interaction between the FCS and the health care environment was Evorable since

104

Saginaw Cooperative Hospitals had established ties to the Michigan State researchers.
Yet the recruiter Eced other challenges (health care environment, social environment,
research design). As a hired F CS recruiter, she was responsible for finding subjects in
several hospitals, treatment centers, and physicians’ omces around the town. Saginaw
provided health care for a lot of elderly people with cancer who lived in the “thumb” of
Michigan (social environment), and who commuted daily for radiation treatments (health
care environment). Others came to Saginaw for diagnosis but were treated in home
communities not available to our researchers. The efforts to verify eligibility and to
access these patients took careful planning, coordimtion, and cooperation with the
community-based agencies.

Operational characteristics such as recruitment personnel and payment to
recruiters were important predictors of early attrition in the Family Care Study. These
mechanisms support an important characteristic of the Ecological Model, namely the
interaction and interdependence between the ecosystem of the research study and the
health care environment. Similar to Motzer and colleagues (1997), the F CS encountered
gatekeepers at every entrance. Willingness of physicians and institutions to provide
access to patients or to allow office staff a role in recruitment varied widely, particularly
when there was no previous relationship between the researchers and the agency. To
overcome such barriers, significant resources were committed to subject recruitment. An
important strategy was the employment of people dedicated solely to identifying and
recruiting study participants.

In the Family Care Study, FCS recruiters had higher recruitment rates, but

unfortunately, these recruits also tended to have higher rates of dropout and

105

nonparticipation than those recruited by stafi nurses. Among those who dropped out
early were the 10 minority patients from Saginaw described above. Contrary to
expectations, the analysis showed that staff nurses, and not FCS recruiters, were able to
obtain the highest participation rates. In hindsight it appears that the relative success of
the staff nurses may be due to several Ectors. First, unlike the agency research personnel,
staff nurses did not have to juggle competing demands for enrollment in other studies.
Also, staff nurses typically volunteered as recruiters, while many research staff were
“assigned” to the FCS. More importantly, difi'erent fiom both the FCS recruiters and
research personnel, staff nurses worked in the specific recruitment locations, giving them
superior access to patients and their Emilies (health care environment, family
environment). In particular, staff nurses in radiation oncology settings had higher consent
and lower refusal rates than recruiters in chemotherapy clinics or inpatient settings. Since
patients visited radiation clinics on multiple, consecutive days, rapport with patients and
Emilies was established and recruitment did not have to occur simultaneously with
postoperative teaching or with the initiation of chemotherapy. Finally, it is possible that
subjects were more willing to participate when recruited by a nurse who assisted them in
the clinical setting, instead of when approached by a stranger.

Aside fi‘om the recruiter roles, an interesting pattern emerged regarding
compensation of recruiters. Somewhat surprising, direct payment to recruiters did not
improve participation rates among the eligible patients. In Ect, the results suggest that
financial incentives may have lead some recruiters to urge patients to sign up, even if

they represented poor prospects of retention. Whereas subject incentives are fiequently

106

discussed in the literature, payments to recruiters presents an interesting area of further
research.

There may be distinct phases of survey participation in which different factors
interact to predict whether subjects agree to participate or not (early attrition), and if
subjects consent to participate, what determines whether they move to the next stage of
the study or drop out (attrition over time). From a systems perspective, the components
of the Ecological Model are interdependent and interrelated. Researchers must expect that
during long-term enrollment periods changes in the health care environment or social
environment rmy impinge on successful subject accrual and consequent retention. The
adaptive qualities of the model allow the structure to accommodate new conditions in the
environment (Bristor, 1990).

Recruitment may be especially slow during the initial phase of a study as
investigators are establishing contacts, honing procedures, and training personnel
(Diekmann & Smith, 1989). Although literature that describes recruitment and retention
successes discuss the interaction between research personnel and subjects, few research
reports of panel studies evaluate the effect of the study procedure as a contributor to
attrition, other than noting the problems associated with a small convenience sample or
using data collection tools that were difficult for participants to complete (Neumark,
2000)

Identifying research design and subject characteristics affecting the likelihood of
nonparticipation and attrition during all phases of an investigation is essential. In the
Family Care Study, the phase of the study affected early attrition, with refirsal rates

steadily increasing over time. Despite adapting and accommodating the protocol to the

107

unique attributes of each community agency, ultimately it proved harder and harder to
enroll eligible subjects, particularly after obtaining sufficient numbers of some diagnoses.
These design features characterize the dynamic interactions occurring between the
environments of the ecological model that is pertinent to the outcome of sample attrition.
During the three years of sanrple accrual, cancer care services underwent rapid
restructuring, including increased acuity of hospitalized patients, shorter hospital stays,
and the expansion of outpatient care. While challenges were anticipated in undertaking a
panel design with elderly cancer patients, systemic changes in cancer care delivery, with
fewer patients in the communities than projected and less support in the agencies than
promised, was difficult, if not impossible to anticipate (health care environment, social
environment). Although some issues were specific to the FCS, the ebbs and flows of
subject accrual and attrition are common to all studies.
Data from the Field Notes

In addition to the variables examined in the quantitative analysis, a review of
interviewer and recruiter field notes yielded a list of motives for attrition in the FCS
(Neumark et al., 1997; Neumark et al., 1998). While the notes incomplete, and often no
reasons were offered particularly for nonconsenters, about half of the early attrition that
was documented was attributed to cancer-related reasons; for example, patients reported
being too sick, too busy with treatments, overwhelmed by the diagnosis, or disliked
talking about it. Interestingly, some people, when contacted for their second interview,
felt they were no longer eligible to participate since the surgery " got the cancer." Finally,
sometimes people dropped out because they were feeling better, and no longer wanted to

be bothered.

108

According to the field notes, design-related reasons for early attrition included a
few patients who did not recall signing the consent, and some who could not be contacted
based on the information provided by the recruitment staff. Others declined or dropped
out due to poor vision, limiting ability to complete the questionnaire; or poor hearing,
limiting ability to complete the telephone interview. Some simply did not like talking on
the telephone. Other research design features that were mentioned as reasons for not
continuing participation included the length of the interview, the repetitiveness of the
questions, and the personal nature of the instruments.

Limitations of the Study

Although this dissertation helps untangle the phenomenon of sample attrition in
panel survey studies of families in which an older member has cancer, this ex post Ecto
case study has limitations. Because the sociodemographic background variables cannot
be manipulated and research design features were not systematically varied and randomly
assigned to different participants, the analysis was limited to observatioml data collected
during sample accrual and recruitment, or from data that were collected via the telephone
interviews, mailed questionnaires, and medical records reviews.

Some of the research design concepts of EMSA did not always match data that
were available fi'om the Family Care Study. While the management tracking system was
a valuable source of data for this study, the flaws of the database were limiting at times.
For example, consistency of interviewer, protocol adherence, and degree of nonresponse
to interview items, may all be related to attrition over time, yet often these data were
missing or inconsistent. Also, other interviewer attributes, and the interaction between

interviewers and subjects that is an important proposition of the Ecological Model, are

109

impossible to evaluate in retrospective research that was not designed to study sample
attrition.

Implications of this Dissertation
Enhancing Understanding of Sample Attrition

The results of this dissertation will be important for those involved in Emily
health care research since obtaining reliable results depends in part on acquiring and
retaining the sample for the duration of the project. Predicting attrition in health studies
has been diflicult, and reports are often conflicting. Some of the discrepant findings may
be due to the lack of multivariate analyses, inadequate consideration of confounding
characteristics that may account for differential rates of attrition, or not taking advantage
of the availability of panel data (Mihelic & Crimmins, 1997). Although interpretation of
findings is complicated when multiple predictors and multiple time periods are
considered simultaneously, this complexity mirrors the reality of sample attrition in
Emily health care panel studies.

A strength of panel studies is the ability to describe change over time in
individuals and families. Yet, although panel designs are Eirly common in Emily health
care research, single cross sectional analyses are far more prevalent than those using
pooled data (Neumark, 2000). The empirical examination of sample attrition warrants
multivariate statistical procedures using pooled data. In the analyses for this dissertation,
statistical models explored the combined data fi'om all waves of the panel survey in order
to estimate the determinants of attrition over time. The multivariate analyses provided
evidence that both research design and subject characteristics were useful to predict who

would be more likely to refuse participation, to dropout prior to the initiation of data

110

collection, or to drop out over time from a panel survey of elderly cancer patients and
Emily caregivers.
The Language of Sample Attrition

Sampling issues abound in research involving Emilies and health Ideally,
researchers and funding agencies want accurate information about population statistics,
but it may not be possible to obtain the number of subjects actually eligible for
recruitment. In reality, many health care environment Ectors influence the ability to
access and obtain these sanrples. Researchers are constrained when a specific clinical
population is the focus of a study, and inclusion criteria are used to assist in sampling.
Thus, although researchers make predictions about potential numbers of subjects, based
on hospital records and tumor registries, in reality, true “population” sampling does not
occur. Rather, the subjects for the studies are drawn fi'om the accessible sample. In this
dissertation, the term “target sample” was derived to represent the accessible sample and
to capture the assumption that the studies to which EMSA is applicable do not obtain
probability samples or use random population sampling techniques. Operationally,
"target sample" is defined as the accessible subjects who are found to be eligible for a
specific family heath care study, and who are approached to ascertain consent to
participate.

There are many anecdotal (and valuable) examples of studies which describe
successfirl and unsuccessful access, recruitment and retention efforts targeted to specific
“target samples.” (e. g., women, minority, elderly, mentally challenged, vulnerable).
Optirml conditions for sampling special populations may not meet the realities. The

literature revealed that community-based recruitment strategies are more successful than

111

health care system based strategies, yet the practical necessity to recruit patients when
they are first identified with cancer requires hospitalization (Peterson, 1998). Among
other things, obtaining the target sample depends on access to subjects or data, and
resources such as time, money, and research personnel.

A pertinent illustration is that researchers are responsible for recruiting minority
and underserved populations, particularly for studies with federal funding, and yearly
reports are required to detail these efi'orts (social environment, health care environment,
research design). For example, in the FCS, special efforts were applied, such as using
minority recruiters in Flint, Michigan (Neumark et al., 1998; Peterson, 1998). In spite of
targeted efforts to enhance minority recruitment, the realities of the social and health care
environments of the communities in which recruitment for the F CS occurred were
reflected in the resulting obtained sample which was primarily Caucasian, mrried, and
with at least a high school education.

Consistent terminology is imperative in order to communicate findings about
sample attrition. In particular, it is appropriate to use the accessible target sample as a
starting point to evaluate sample attrition. Although this terminology is distinctive, it
captures the notion that refirsal and recruitment rates play important roles in interpreting
data and evaluating sarrrpling bias. Yet, systematic analysis of refirsal or early drop out is
limited. Unfortunately, it is the nonconsenting subjects about whom little is known
because they did not participate in any extensive data collection efforts (Brehm, 1993).
Nonetheless, a thorough understanding of sample attrition in panel studies must begin the
all potential subjects who were found to be eligible for a study and must consider whether

people who consent to participate are inherently different fi'om those who do not.

112

Incorporating a family perspective

Special conceptual and operational issues arise in research with Emilies.
According to Feetham (1991), it is important to identify the conceptualization of family
guiding a study, to explain how the conceptualization is derived from a theoretical
fiarnework, and to elucidate how the conceptualization influences the empirical aspects
of the study. In a review of studies of people with cancer and Emily caregivers
(Neumark, 2000), definitions of “Emily caregiver” ranged fiom general (designated as
primary caregiver for their patients) to specific (spouse or male partner lives with
patient). In one study, sample attrition occurred due to divorce since the research design
stipulatedthepartnersneeded to bemarriedto eachother. Inthisway,thesubject
characteristics, research design, and Emily environment were interdependent.

One conceptualization of the Emily is that the Emily is the environnrent for
individual Emily members. F eetharn (1991) extends this definition to recognize that the
Emily is the mediator between the individual members and the environment.
Recognizing the Emily as the environment supports the assumption of interdependence
ofthe individual and Emily within the health care enviromnent conceptualized in EMSA.
In the Emily environment, the model conceptualizes that the process of sample attrition
my be influenced by Emily roles and relationships, Emily boundaries, decision-making,
and concurrent life events.

Researchers must decide which parts or aspects of the Emily need to be evaluated
and must determine the appropriate unit of analysis (Bray, 1995; Cooley & Moriarty,
1997 ; Maguire, 1999). However, a controversy persists in the literature. Some believe

that information fi‘om individuals yielding conclusions about relationships in Emilies

113

reflects the perspective of the individual and not the entire family system (Bray; Thomas,
1987). Others suggest that a dyadic approach to studying relationships requires that the
dyad be considered as the unit of analysis throughout all phases of the research process,
fi'om research design to analysis. In reality, this may not occur. For example, although a
review of longitudinal caregiving studies included only samples made up of dyads, only
one of fourteen studies . explicitly specified that the dyad was the unit of analysis. More
typically, data were collected separately, and researchers compared individual data and
reported discrepancies or similarities in responses (Neumark, 2000).

The intent of this dissertation was not to settle this dispute. Ultimately, the
purpose of the research should guide the investigator in determining whether data fi'om
individual Emily members rather than data fi'om multiple sources is suflicient. Thus the
ecological model of sample attrition can be applied to studies involving either individual
or multiple Emily members in panel data collection.

Ecological Characteristics of Sample Attrition

This project established and expanded the Ecological Model of Sample Attrition.
Similar conceptual models which encompass subject, societal, Ernilial, organimtional,
and research design characteristics have not been applied to Emily health care studies.
This researcher’s field experiences with recruitment and retention and clinical practice as
an oncology nurse, as well as the extensive literature review, supplemented the
conceptual guidance provided by the Ecological Model to evaluate the subject
characteristics and research design characteristics predicting sample attrition in the

Family Care Study.

114

Mihelic and Crimmins (1997) criticize work that neglects to account for
individual characteristics that rmy change over time or ignores the contribution of
unobserved characteristics. Conceptually, the holistic ecological perspective overcomes
these criticisms by acknowledging that sample attrition is multileveled and subject to
numerous and simultaneous influences over time, both within and outside of the research
study. EMSA enhances understanding of sample attrition in Emily health care studies.
The Emily environnrent, the social environment, and the health care environment are
systems that independently and interdependently affect the subjects and the researchers.
These relationships are complex, dynamic, and multidirectional.

To capture the assumption of system theory that relationships under study are
complex and multidirectional, appropriate data analytic techniques must be utilized
Exploring simple bivariate relationships is not suflicient (F eetham, 1991; Witte ct al.,
1996). It is sometimes difficult to translate statistical methods to conceptual conclusions,
and future interdisciplinary research is needed to capture the essence of an ecological
model. Witte and colleagues (1996) offer suggestions of ways to operationalize and
analyze health communication processes at the ecological level. Their models are based
in chaos theory. Further exploration of similar applications to the ecological model of
sample attrition is possible.

Conclusions

Focusing attention on the complexities of attrition in a panel survey of elderly
cancer patients and Emily caregivers highlights some important considerations. Multiple
interacting Ectors, related to both study procedures and subject characteristics affect

whether an individual consents to participate or to remain in a panel study to the end.

115

Research studies designed to investigate sample access, recruitment, retention, and
attrition are nwded. Because Emily health care studies are multiEceted, and often very
specific to the population and community under study, research on recruitment, retention,
and attrition may be difficult to design and may require creative means for Ending.
Nonetheless, overcoming individual, societal, family, and organizational barriers to
participation in Emily health care studies is imperative.

Conceptualization of the complexities and interaction of the multiple Ectors
which may influence sample attrition in descriptive studies of Emilies with cancer is a
start. Using a model to predict who is more likely to drop out, when, and why, enables
researchers to implement various strategies to target individuals and Emilies at highest
risk for attrition. The ecological perspective is dynamic and holistic (Wideen et al.,
1998). In the future, the model may be expanded to describe and analyze sample attrition
in clinical trials, particularly Emily cancer prevention and intervention studies with
experimental designs. Thus the Ecological Model will continue to evolve as awareness of
different levels of complexity become apparent, and new properties and insights about
sample attrition emerge.

In making design decisions, researchers must strike a balance between the study
purposes, the characteristics of the sample, field procedures, and available resources to
develop systems to track those most susceptible to attrition. Although they have little
control over health or age-related reasons for attrition, or health care institutional barriers,
researchers do have control over recruitment and retention procedures and can establish
mechanisms to be flexible to the needs of ill participants, Emilies, and project personnel.

The ecological model can be applied to develop and refine tracking systems to monitor

116

sample accrual, recruitment, retention, and attrition in Emily health care panel studies,
- while being responsive to the needs of the researchers, project personnel, ill participants
and Emilies. Implications of such strategies extend to issues of accessing Emilies for

assessment, provision, and evaluation of care as they live with cancer.

117

APPENDIX A

Tables

118

Table 1

Early Attrition Analysis: Variables Available for Target Sample of 1948 Cases

 

Subject Characteristics an % Eligiiiccltleggicgsn 'n %
Patient Sex 1938 Case Type 1877
Female 908 46.9 Patient & Caregiver 1062 56.6
Male 1030 53.1 Patient Only 815 43.4
Patient Ageb 1927 Recruiter Role
Patient Race 1829 FCS Recruiters 771 39.6
Non-Hispanic White 1682 92.0 Research Staff 389 20.0
Afiican American 129 7.1 Stafl‘ Nurses 786 40.4
Other minority 18 1.0 Recruiter Payment
Cancer Diagnosis 1940 Paid by FCS 592 30.4
Breast 483 24.9 Not paid by FCS 1354 69.6
Colon/Rectum 436 22.5 Study Phase
Lung 557 28.7 Start up 162 8.3
Prostate 464 23.9 Full effort 783 40.3
Resumption 640 32.9
Winding down 360 18.5

 

“The “n” represents the data available for each variable. The completeness of information
obtainable, especially for nonconsenters, differs from measure to measure.
I’Mean age of target sample = 73.0 years; range 65-98 (SD. = 5.8).

119

Table 2

Multinomial Logistic Regression Comparing Participants“ to Early Dropouts” to
Nonconsentersc

 

 

 

 

Dropouts Nonconsenters Nonconsenters M ulti-
vs Partrcrpants vs Partrcrpants vs Dropouts no mial
figigem OR va’lile 0R vaplire OR value p- value

Patient Age (years) 1.004 .774 1.02 .009 1.022 .150 .031

Patient Sex (F vs M) .947 .797 .947 .726 1.000 .998 .928

Patient Race (White)R 006s .754s .016S .010S
Afiican American 1.352 .319 1.13 .570 .837 .574 .584
Other 5.330 .002 .719 .634 .135 .005 .002

Cancer Diagnosis (Prostate)R .003S .004S .234S .0019
Breast 1.828 .060 1.44 .098 .791 .493 .080
Colon/Rectum 2.353 .002 1.95 .000 .830 .525 .000
Lung 2.317 .001 1.41 .040 .609 .063 .001

(Cs/5551261, only) .419 .000 .504 .000 1.201 .283 .000

Recruiter role (FCS recruiter) .581s .000S .0005 .000S
Research Staff 1.101 .678 .885 .442 .803 .365 .599
Staff Nurse 1.245 .298 .467 .000 .375 .000 .000

Recruiter Payment

(Paid by FCS vs Not 1.659 .022 .557 .000 .336 .000 .000

Paid)

Study Phase (Strut-up)R .282S .000: .0008 .000S
Full Effort .771 .394 .456 .000 7.077 .000 .000
Resunrption 1.069 .845 7.33 .000 6.910 .000 .000
Winding Down .763 .432 9.28 .000 12.16 .000 .000

 

Note. Number of cases with complete variables, an=992; b n=202; °n=534.
Note. OR= adjusted odds ratio.

Likelihood Ratio )8: 193.66; degrees of fieedom: 28; p<.001

R Reference Category (comparison group).

S Significance tests associated with multiple category predictors.

120

Table 3

Attrition Over Time in the Family Care Study

 

 

 

Overall Status of Cases
At Wave 1 At Wave 2 At Wave 3 . At Wave 4
N % N % N % N %
On
Study 1005 83.8 877 87.3 785 89.5 704 89.7
CE 195 16.3 128 12.7 92 10.5 81 10.3
Study
Total 1200 1005 877 785

 

121

Table 4 (page 1 of 2)

Bivariate Analyses of Time-Independent Subject Characteristics and

 

 

 

Attrition Over Time
vita" Start Wave 2 Start Wave 3 Start Wave 4
ave 1
gage: On Off On on On on
3"“) e n=877 n=128 n=785 n=92 n=704 n=81
n—1005
Subject N N N N N N N
Characteristics (%) (%) (%) (%) (%) (%) (%)
Patient sex p S .729 p S .031 p S .006
M 1 551 479 72 419 60 364 54
ac (54.8) (54.6) (56.2) (53.4) (65.2) (51.7) (67.9)
Female 454 398 56 366 32 340 26
(45.2) (45.4) (43.8) (46.6) (34.8) (48.3) (32.1)
Patient age p 5 .072 p S .677 p _<_ .801
Mean age, yrs 72.60 72.48 73.39 72.51 72.27 72.53 72.37
(range) (65-98) (65-98) (65-94) (65-98) (65-84) (65-98) (65-91)
Patient racea p S .028 p 5 .001 p S .407
White 927 815 1 12 740 75 662 78
(92.3) (93.0) (87.5) (94.3) (82.4) (94.0) (96.3)
. lit 77 61 16 45 16 42 3
Mm" y (7.7) (7.0) (12.5) (5.7) (17.6) (6.0) (3.7)
Patient educationb p 5 .057 p 5 .001 p _<. .484
< High 73 65 8 54 1 1 49 5
School (8.2) (8.2) (8.4) (7.5) (15.7) (7.5) (6.8)
High 474 413 61 368 45 326 42
School (53.4) (52.1) (64.2) (50.9) (64.3) (50.2) (57.5)
Colle 6+ 341 315 26 301 14 275 26
g (38.4) (39.7) (27.4) (41.6) (20.0) (42.3) (35.6)
MMarital StatusC p S .599 p S .218 p S .363
Married 649 574 75 516 58 458 58
(66.4) (66.7) (64.1) (66.4) (69.0) (65.4) (75.3)
Never 23 21 2 21 0 l9 2
married (2.4) (2.4) (1.7) (2.7) (2.7) (2.6)
Divorced/ 5 8 48 10 46 2 43 3
Separated (5.9) (5.6) (8.6) (5.9) (2.4) (6.1) (3.9)
Widowed 248 218 30 194 24 180 14
(25.4) (25.3) (25.6) (25.0) (28.6 (25.7) (18.1)

 

122

Table 4 (page 2 of 2)

Bivariate Analyses of Time-Independent Subject Characteristics and

 

 

 

 

Attrition Over Time
Sta“ Start Wave 2 Start Wave 3 Start Wave 4
Wave 1
Snug: On on On 03“ On on
“3&5 n=877 n=128 n=785 n=92 n=704 n=81
Subject N N N N N N N
Characteristics (%) (%) (%) (%) (%) (%) (%)
Patient Lives Aloned p 5 .559 p 5 .301 p 5 .109
Lives with 694 617 77 559 58 497 62
other(s) (77.0) (76.7) (79.4) (76.3) (81.7) (75.4) (83.8)
Lives alone 207 187 20 174 13 162 12
(23.0) (23.3) (20.6) (23.7) (18.3) (24.6) (16.2)
”Household lncome p 5 .173 p 5 .003 p 5 .196
Mean 27,982 28, 262 25,742 28,869 22,640 29,155 26,322
mcome ($)
Freq. Of 935 831 104 750 81 674 76
response (11)
Cancer Diagnosis p 5 .001 p 5 .001 p 5 .001
L 283 228 55 164 64 119 45
ung (28.2) (26.0) (43.0) (21 .0) (69.6) (16.9) (54.3)
Colon 188 167 21 155 12 137 18
(18.7) (19.0) (16.4) (19.8) (13.0) (19.5) (22.2)
Breast 255 231 24 224 7 220 4
(25.4) (26.3) (18.8) (28.5) (7.6) (31.2) (4.9)
Prostate 279 251 28 242 9 228 14
(27.8) (28.6) (21.8) (30.7) (9.8) (32.4) (17.3)
Cancer Stage p 5 .001 p 5 .001 p 5 .001
Late Stage 619 562 57 537 25 509 28
(3-4) (61.6) (64.1) (44.5) (68.4) (27.2) (72.3) (34.6)
Early Stage 286 240 46 189 51 147 44
(0-2) (28.5) (27.3) (35.9) (24.1) (55.4) (20.1) (54.3)
No staging 100 75 25 59 16 50 9
info (9.9) (8.6) (19.5) (7.5) (17.4) (7.1) (11.1)

 

123

Table 5

Bivariate Analyses of Time-Dependent Subject Characteristics and Attrition Over Time

 

 

 

Start Wave 2 Start Wave 3 Start Wave 4

On 011” On Off On Off
Suhj ect Characteristics n=877 n=128 n=785 n=92 n=704 n=81
Symptom Count p 5 .305 p 5 .001 p 5 .001
““13“” “mm at 800 97 703 63 642 61
prevrous wave
Mean at previous wave 8.09 8.60 6.93 9.41 5.94 8.84
Symptom Severity p 5 .105 p 5 .001 p S .001
Freql’ency ““51”” at 800 97 703 63 642 61
prevrous wave
Mean at previous wave 12.69 14.31 10.96 15.80 8.91 15.50
Physical Functioning” p 5 .073 p 5 .001 p 5 .001
le’ency “’8sz at 794 96 741 85 678 77
prevrous wave
Mean at previous wave 63.34 57.76 72.64 49.85 73.34 48.78
Bodily Pain" p S .683 p 5.001 p 5 .001
Freqf’ency “response at 789 95 737 85 680 77
prevrous wave
Mean at previous wave 67.10 65.88 76.93 62.96 79.26 63.21
Physical Role lmpact* p 5 .590 p 5 .001 p 5 .001
“6‘1me “”51”“ at 789 92 739 84 681 77
prevrous wave
Mean at previous wave 40.56 42.93 55.68 28.87 65.36 34.52
Emotional Role lmpact* p 5 .901 p 5 .004 p 5 .001
Freql‘ency 0““an at 789 92 739 84 681 74
prevrous wave
Mean at previous wave 80.69 81.16 85.70 75.40 90.75 75.00
General Health Perceptions* p 5 .043 p 5 .001 p 5 .001
““13”” 0“ “31’0““ at 786 92 737 85 675 77
prevrous wave
Mean at previous wave 68.48 63.63 69.7 51.95 70.60 46.11
Vitality" p 5 .213 p 5.001 p 5.001
Frequncy Of’esl’Ons" at 769 91 691 64 644 61
prevrous wave
Mean at previous wave 51.81 48.46 57.36 43.36 60.0 41.39
Social Functioning“ p 5 .693 p 5 .001 p 5 .001
Freql'emy 0f “390““ at 792 93 741 83 682 76
prevrous wave
Mean at previous wave 70.11 71.37 81.90 60.67 86.95 57.73

 

Note. The (*) items are the scale items from the Medical Outcomes Study Short-Form 36
(SF-36) described in the methodology section.

124

Table 6

Bivariate Analyses of Research Design Characteristics and Attrition Over Time

 

 

Wave] Wave 2 Wave 3 Wave 4
gang: On 013‘ On 08‘ On 011“
11$) 5 n=877 n=128 n=785 n=92 n=704 n=8l
Research Design N N N N N N N
Characteristics (%) (%) (%) 4%) (%) (%) (%)
Case Type p 5 .075 p 5 .023 p 5 .135
Patient with 659 584 75 513 71 454 59
Caregiver (65.6) (66.6) (58.6) (63.4) (77.2) (64.5) (72.8)
. 346 293 53 272 21 250 22
Pam 0“” (34.4) (33.4) (41.4) (34.6) (22.8) (35.5) (27.5)
Data Source
Any Telephone Interview p S '129 p 5 .324 p 5 .693
figg‘itefggeifeggus 873 126 751 90 691 80
wave p (99.5) (98.4) (95.7) (97.8) (98.2) (98.7)
No interview done at 4 2 34 2 13 1
previous wave (0.5) (1.6) (4.3) (2.2) (1.8) (1.2)
Data Source
Any Mailed Questionnaire p S '001 p S ”003 p S '00]
P and/or CG questionnaire 832 65 722 76 687 70
returned at previous wave (94.9) (50.8) (92.0) (82.6) (97.6) (86.4)
No interview done at 45 63 63 16 17 11
previous wave (5.1) (49.2) (8.0) (17.4) (2.4) (13.6)
Consistency of Interviewer p 5 .479 p 5 .001 p 5 .350
Same interviewer at Wave 663 66
l and Wave 2 (78.1) (81.5)
Different interview at 186 15
Wave 1 and Wave 2 (21.9) (18.5)
Same interviewer across 263 48 240 23
all waves of interviews (33.4) (53.3) (34.0) (28.8)
isszsrsersimm“ 42 .7
(66.6) (46.7) (66.0) (71.2)

interviews

 

125

 

Table 7 (page 1 of 2)

TSCS Models of All Attrition in the Family Care Study

 

Model 1‘ Model 2b Model 3c
Independent Variables OR p-value OR p-value OR p—value

 

 

 

 

Subject Characteristics

 

Patient sex (M vs F) .801 .338 .749 .183

Patient age (years) 1.04 .032 1.04 .023 1.03 .041
Patient race

(White vs minority) .830 .612 .986 .966
Education J
(High school”) '330
College .688 .565
Less than HS .774 .386
No info .505 .439
Marital Status 1 J j
. .107 .086 .074
(MamedR)
Never married .688 .565 .722 .599 .954 .931
Divorced! Separated .627 .230 .608 .1 75 .629 .1 85
Widowed .547 .017 .583 .017 .606 .013
Live alone?
(No vs Yes) 1.00 .978
Income
7
($10003) .991 .153 .990 .088 .989 .04..
Diagnosis (LungR) .0005 .0005 .000S
Colon .436 .001 .443 .001 .404 .000
Breast .282 .000 .334 .000 .294 .000
Prostate .298 .000 .317 .000 .375 .000
Cancer Stave 4 s S
(Late stage?) .00.) .001 .000
Early stage (0-2) .515 .001 .496 .000 .451 .000
No staging info .631 .217 .871 .325 .831 .475

 

Note. Number of observations with complete variables, ‘n=2201; I’n=2210; cn=2319.
RReference category (comparison group).

’ Joint significance tests for multiple category predictors; no info category excluded.
8 Joint significance tests for multiple category predictors; all categories included.

126

Table 7 (page 2 of 2)

TSCS Models of All Attrition in the Family Care Study

 

 

 

 

 

 

Model 1‘ Model 2b Model 3c

Independent Variables OR p-value OR p-value OR p-value
Symptom Count .935 .185
Symptom Severity 1.04 .125 1.01 .198
6690”“. "63'“ .987 .004 .988 .003 .983 .000
perceptions
Physical functioning 1.00 .627
Emotional functioning 1.00 .978
Research Design Characteristics
Time (olrat w2") .150S .513 5

Ofl‘ at w3 .958 .838 .871 .477

Ofiat w4 1.41 .105 1.10 .602
83261135561, with CG) 1.33 .126 1.43 .041 1.23 .217
gigsvtfgggwone .107 .000

 

Note. Number of observations with complete variables, I"n=2201; bn=2210; cn=2319.
RReference category (comparison group).
1 Joint significance tests for multiple category predictors; no info category excluded.
S Joint significance tests for multiple category predictors; all categories included.

127

Table 8

Interaction Model: Prinmry Cancer Diagnosis x Wave Time

 

 

wiggiem OR p-value
Patient sex (M vs F) .732 .154
Patient age (years) 1.039 .019
White vs Minority Race 1.001 .998
Marital Status Married“) .079‘
Never married .686 .547
Divorced/ Separated .605 .1 73
Widowed .574 .015
Income ($10005) .990 .086
Diagnosis (LungR) .432S
Colon .657 .273
Breast 1.090 .820
Prostate .670 .241
Cancer Stage (Late stageR) .001s
Early stage (0-2) .480 .000
No staging info .691 .301
Symptom Severity 1.014 .125
Health Perceptions .989 .006
Time (ofl‘at w2“) .1114s
OE at w3 1.746 .045
Off at w4 2.322 .005
83:13 1;: P with CG) 1.421 .046
Diagnosis‘WaveTime .011S
Colon*wave3 .467 .176
Colon*wave4 .574 .305
Breast*wave3 . 140 .002
Breast*wave4 .080 .000
Prostate‘wave3 .255 .011
Prostate‘wave4 .294 .015

 

I"Reference category (comparison group).

JJoint significance tests for multiple category predictors; no info category excluded.
S Joint significance tests for multiple category predictors; all categories included.

128

Table 9

Attrition Over Time: Comparing Attrition Outcomes for Target Sample
of 1005 Cases

 

 

 

 

 

At Wave 2 At Wave 3 At Wave 4
N % N % N %
On Study 877 87.3 785 89.5 704 89.7
Ofl‘ Study
Attrition NOT
due to Death 104 10.3 37 4.2 27 3.4
Attrition Due
to Death 24 2.4 55 6.3 54 6.9
Total 1005 100.0 877 100.0 785 100.0

 

129

Table 10 (page 1 of 2)

TSCS Models of Attrition Due to Causes Other Than Death

 

 

 

 

 

Full Model 11‘ Model 2b
Independent Variables OR p-value OR p-value
Subject Characteristics
Patient sex (M vs F) .701 .312
Patient age (years) 1.02 .369
(we...)
1313i]; t3161:1001") '908’
College .919 .744
Less than HS .871 .726
No info
Marital Status (MarriedR) .822J
Never married 1.20 .801
Divorced/Separated l . 1 9 .823
Widowed .854 .822
$53322) .547 .131 .988 .858
Income ($10008) .990 .196
Diagnosis (LungR) .297S .025S
Colon .821 .570 .603 .165
Breast .607 .237 .442 .077
Prostate .542 .084 .590 .066
552.2223 .0.
Early stage (0-2) 1.06 .831
No staging info .584 .371

 

Note. Number of observations with complete variables, an=2201; I’n=2222.
RReference category (comparison group).

JJoint significance tests for multiple category predictors; no info category
excluded.

S Joint significance tests for multiple category predictors; all categories included.

130

Table 10 (page 2 of 2)

TSCS Models of Attrition Due to Causes Other Than Death

 

 

 

 

 

 

 

Full Model 1‘ Model 2b
Independent Variables OR p-value OR p-value
Subject Characteristics
Symptom Count .979 .772
Symptom Severity 1.00 .949
General health perceptions .992 .153 .993 .136
Physical functioning 1.00 .663
Emotional functioning 1.01 .217
Research Design Characteristics
Time (olfat w2") .002 5 .001 5
Off at W3 .395 .001 .490 .003
Off at W4 .547 .039 .434 .001
($261355, with CG) 2.37 .000 1.89 .002
$353331? Done .061 .000

 

Note. Number of observations with complete variables, an=2201; l’n=2222.

RReference category (comparison group).
JJoint significance tests for multiple category predictors; no info category

excluded.

8 Joint significance tests for multiple category predictors; all categories included.

131

Table 11 (page 1 of 2)

TSCS Models of Death Attrition in the Family Care Study

 

Full Model 1a Model 2b Model 3c
Independent Variables OR p-value OR p-value OR p-value

 

 

 

 

Subject Characteristics

 

Patient sex (M vs F) .709 .352 .685 .228
Patient age (years) 1.08 .032 1.06 .042 1.06 .068
Patient race (White VS .450 .331 .469 .293
mlnorlty)
l(EICIiiugCliitslglilool") '150 J
College .557 .126
Less than HS .344 .111
No info .685 .059
12:23:23.1?“ .774’
Never married
Divorced/Separated .508 .1 11
Widowed .674 .250
i413: 3:032; 1.07 .694
Income ($10008) .991 .440
Diagnosis (LungR) .005 S .002 S .000 S
Colon .203 .010 .222 .002 .219 .002
Breast .025 .005 .028 .002 .022 .001
Prostate .092 .002 .097 .000 .112 .000
Efgtiegtiéi 3’ 005 S .002 S 000 5
Early stage (0-2) .161 .001 6.64 .000 .155 .001
No staging info .555 .305 3.32 .050 .547 .001

 

Note. Number of observations with complete variables, an=2201; bn=2203; cn=2203.
RReference category (comparison group).

1 Joint significance tests for multiple category predictors; no info category excluded.
8 Joint significance tests for multiple category predictors; all categories included.

132

Table 11 (page 2 of 2)

TSCS Models of Death Attrition in the Family Care Study

 

Full Model 1° Model 2" Model 3°

 

 

 

Independent Variables OR p-value OR p-value OR p-value

 

Subject Characteristics

 

 

 

Symptom Count .924 .353

Symptom Severity 1.06 .139 1.03 .108 1.02 .144

gigglié‘sjm‘ .978 .029 .976 .003 .976 .003

Physical functioning 1.00 .859

Emotional functioning .994 .172 .993 .105 .993 .111

Research Design Characteristics

Time (oaat M") .022 S .013 S .002 S
Offat W3 5.13 .008 5.42 .001 5.28 .001
Ofiat W4 11.3 .007 11.1 .000 10.7 .001

83:6 “31:1, with CG) .498 .101 .458 .032 .456 .031

gills/53(6):)naire Done .706 .582

 

Note. Number of observations with complete variables, an=2096;"n=2210; cn=2203.
RReference category (comparison group).

J Joint significance tests for multiple category predictors; no info category excluded.
S Joint significance tests for multiple category predictors; all categories included.

133

APPENDIX B

Figures

134

Seaman ._§e_>_eauc

 

 

 

 

       

“ED 038% m z
I‘ §
085. I EoEco:>=m
game
8:53. ' f
oaeam nomae§ .. . ...... /§\ ........
cmfiom spammem I . ,K

mozmtouomano 80.35 I .....

\

     

      

\

/ .\\ \

 

EoEcoem>cm
Emu 5:63

 

 

 

 
    

 

EoEcocScm
860m

 

 

 

seam sea 28 £36m Ease 5 SEE 295m do 382 338m ._ 2:5

135

mucogohEm ‘

a

8:32. '
23% Egg
:w_moQ nonmemom I
85:28.86 Sufism a .x I /\
a

\\a

 

l
‘

 

 

  

 

was 59888 CO m=§m
532385 mo 553550
858005 coccfiom
855;? .80on
€052: 550200 8.5
mozmcgomnmco $32.38:—
ES 299%

8:68er EoEzaoom
$53005 mmuoo<

 

 

 

 

 

852mm: cinnamooo

559292” 9 EoEHEEoU

cofiamsm .Em_E:mimoe2_z<

$8: 3:285m

2:62am memEtoaasm Boom
m:_:oco:£ _mo_wo_ono>mm

mcwcomwoce 323i

oEooE .cocmozvulmgfim 28088063
52% .88 .uwmimoEaSmoEoQ

 

 

seam seam gem Esau a eoEaZ usages
momfiuouofiuno :wmmom 53.30% 93 agar—88.520 Sufism .N 0.53m

136

 

 

 

323505

0:50:50: 5:3 3:033:
355% 35:35:—
5:33.083:

 

805308385 5:50.: 50:50:00
305: .5 0.3::

$05: 8 8:88: 5::3»
50:08:85.: 33:3“:

0:35:30: 52:3“—

wE3E 53:00: 5:50»
3:858 2:50 9.3 3:330
850:»: 3:31.:

005: :3 5:35:33: 52:3:

 

 

m0_3_0::om
5:883: 5:3 £0380: ’
5338 038:0 3535590 1 l,

 

 

 

* ::0E:o:_>:m 030 530:: _

 

082. A17

5:0380M

seesaw a

::0E:o:_>:m 350m

0::0E:o::>:m 0.2. .m 05mm”:

           

:

50:50:35: 353m

137

 

05:03: 050500038
>52:

50:0?

5:53.: .5380: .5 33>
60:50

30:55:: :3 3:330 .338
20:03:05: 35cm

25350 .85: £03: 350m

 

 

 

Figure 4. Operationalization of the Independent Variables in the Family Care Study
in Relationship to the Theoretical Components of EMSA

 

 

Subject Characteristics of
Ecological Model ofSample Attrition
o F CS variable available for analysis
0 F CS variable unavailable for analysis

Demographics
0 patient sex
0 patient age
0 patient race
Socioeconomic status
0 family income
0 patient employment“
0 patient education
0 medicaid status
Physical functioning
0 cancer diagnosis
0 cancer stage
0 cancer-related symptom count"
0 cancer-related symptom severity“
0 patient comorbidities
o hospitalizations*
0 cancer treatment (surgery,
chemotherapy, radiation)
0 physical functioning (SF-36)*
0 physical role impact (SF-36)*
0 general health perceptions (SF-36)*
Psychological functioning
0 emotional role impact (SF—36)*
o vitality (SF-36)*
0 social fimctioning (SF-36)*
0 mental health (SF -36)*
o depression“
Social support/flimily structure
0 marital statue“
0 does patient live alone?*
Geographic residence
0 urban versus rural
Personality traits
Attitudes
Commitment to participation

 

Research Design Characteristics of
Ecological Model of Sample Attrition
0 FCS variable available for analysis
0 FCS variable unavailable for analysis

Access procedures

Recruitment procedures
Skills of research staff
0 recruiter role
0 recruiter payment
0 skill of recruiter (time on study)
0 recruitment location (in hospital,
community MD, chemotherapy
clinic, radiation clinic)
Sample unit
0 case type (with or without family)
Interviewer characteristics
Skills of research staff
0 consistency of interviewer
0 skills of interviewer (time on study)
Data collection methods
0 data source——telephone interview
Odatascurce self ’ "‘ ‘
questionnaire
Protocol adherence
0 interview done
a questionnaire returned
0 wave contact time
0 time between consent and initial
contact
Retention procedures
0 small gifls
o reminder phone calls
0 birthday cards
0 telephone contact at wave 3.5

 

 

 

 

 

HEP 02;?

 

 

a. as? m 33 N 0%?
_ _ _ C
Di
- 2
e853: - _
398:: Jon
58+ .
0:5 ill - m _
- N
fi m.”

3:32 5:088:— 05? 0203 _._ £00335 50:30 .m 05mE

 

SPPO “(fl-”W

139

 

REFERENCES

140

REFERENCES

Agresti, A. (1990). Categorical Data mysis. New York, NY: Wiley.

Areén, P. A. & Gallagher-Thompson, D. (1996). Issues and recommendations for
the recruitment and retention of older ethnic minority adults into clinical research. Journal
of Counsel_t_ng' and Clini_cal Psychology, fl, 875-880.

Babbie, E. (1998). The Practice of Social Resear_ch. Belmont, CA: Wadsworth.

Benjamin, M., & Hollings, A. (1997). Student satisfaction: Test of an ecological
model. Journal of College Student Development, 3; 213-228.

Bischofi‘, R J., & Sprenkle, D. H. (1993). Dropping out of marriage and family
therapy: A critical review of research. Palm Process, _3_2, 353-375.

Blumenthal, D., Sung, J., Coates, R, Williams, J., & Lifl‘, J. (1995). Recruitment
and retention of subjects for a longitudinal cancer prevention study in an inner-city Black
community. Health Services Research, 3_0, 197-204.

Bond, M.A., & Pyle, J .L. (1998). The ecology of diversity in organizational
settings: Lessons from a case study. Human Relations, 51, 589-623.

Bonvicini, K. A. (1998). The art of recruitment: The foundation of family and
linkage studies of psychiatric illness. Fa_rn_1ly' Process, 32, 153-165.

Brashers, D.E., & Babrow, AS. (1996). Theorizing communication and health. .
Communication Studies 41, 243-251.

 

Braver, S. L, & Bay, R. C. (1992). Assessing and compensating for self-selection
bias (non-representativeness) of the family research sample. Journal of Mm’ e and the
Family, 54, 925-939.

Bray, J. H. (1995). Family assessment: Current issues in evaluating families. family
Relations, fl, 469-478.

Bray J.H., Maxwell, S.E., & Cole, D. (1995). Multivariate statistics for family
psychology research. Journgl of Fa_m_ily Psychology, _9_, 144-160.

Brehrn, J. (1993). Why do people participate in surveys? In J. Brehrn (Ed), '_I_'h_e
Phantom Re ndents: inion Surve and Political Re resentation (pp 39-71). Ann
Arbor, MI: University of Michigan.

Bristor, M.W. (1990). Individuals families and environments. Dubuque, IA:
Kendall/Hunt.

 

141

 

Bussell, D. A, Matsey, KC., Reiss, D., & Hetherington, M. (1995). Debriefing the
family: Is research an intervention? Family Process. 3, 145-160.

Carp. F. M. (1989). Maximizing data quality in community studies of older people.
In M.P. Lawton, & A. R. Herzog (Eds), Smcial Research Methods for Gerontology (pp.
93-122). Amityville, NY: Baywood.

Carey, T.S., Kinsinger, L, Keyserling, T., & Harris, R. (1996). Research in the
community: Recruiting and retaining practices. Journal of Community Health, 21, 315-32

Carter, W. B., Elward, K., Malmgren, J., Martin, M.L., & Larson, E. (1991).
Participation of older adults in health programs and research: A critical review of the
literature. The Gerontologist, 31, 584-592.

Cartwright, J., & Limandri, B. (1997). The challenge of multiple roles in the
qualitative clinician researcher-participant client relationship. Qflh’tative Hflth Research,
1, 223-235.

Clinton, J., Beck, R., Radjenovic, D., Taylor, L., Westlake, S., & Wilson, SE.
(1986). Time-series designs in clinical nursing research: Human issues. Nursmg' Research,
3_5, 188-191.

Coen, A. 5., Patrick, D. C., & Shem, D.L. (1996). Minimizing attrition in
longitudinal studies of special populations: An integrated management approach.
Evaluation and Program Planning, fl, 309-319.

Cooley, M., & Moriarty, H. J. (1997). An analysis of empirical studies examining
the impact of cancer diagnosis and treatment of an adult on family functioning. Journal of

Family Nursmg' , 3, 318-347.

Corder, L. S., & Manton, KG. (1991). National surveys and the health and
functioning of the elderly: The efl‘ects of design and content. Journal of the American
Statistical Association, 86, 513-525.

Creswell, J.W. (1994). Research design: Qualitative and quantitative approaches.
Thousand Oaks, CA: Sage.

Cronenwett, LR. (1986). Access to research subjects. Journal of Nursmg'
Administration, 16, 8-9.

Daly, K (1992). The fit between qualitative research and characteristics of

families. In J. F. Gilgun, K Daly, & G. Handel (Eds), Qu_al_i_tative Methods in Family
Research (pp. 3-11). Newbury Park, CA: Sage.

142

Davies, B., Reirner, J. C., Brown, R, & Martens, N. (1995). Challenges of
conducting research in palliative care. Omega, 3_1_, 263-273.

Demi, A. S., & Warren, N. A. (1995). Issues in conducting research with
vulnerable families. Western Journal of Nursmg' Research, l_7, 188-202.

Denham, S. (1999a). Part 1: The definition and practice of family health. Journal of
Fm’fl Nursing, 3, 133-159.

Dennis, B.P., & Neese, J. B. (2000). Recruitment and retention of African
American elders into community-based research: Lessons learned. Archives of ngchiatric
Nursing, 311, 3-11.

Diekrnann, J.M., & Smith, J.M. (1989). Strategies for accessment and recruitment
of subjects for nursing research. Western Journal of Nursmg’ Research, _1_l, 418-430.

DiFrenceisco, W., Kelly, J.A., Sikkema, K.J., Somlai, A.M., Murphy, D.A., &
Stevenson, L.Y. (1998). Differences between completers and early dropouts from 2 HIV
intervention trials: A health belief approach to understanding prevention program attrition.
American Journal of Public Health, _8_8_, 1068-1074.

Diggle, P.J., Liang, K.Y., & Zeger, S.L. (1994). Oxford Statistical Science Seiries:
Vol. 13. Analysis of lorgitudinal data Oxford, UK: Oxford University Press.

Eaves, Y. D. (1999). Family recruitment issues and strategies: Caregiving in rural
African Americans. Nursm’ Research, fl, 183-187.

Evangelista, F., Albaum, G., & Poon, P. (1999). An empirical test of alternative
theories of survey response behavior. Journal of the Market Research Socieg, 4_1_, 227-
244.

Farnily Care Study (1996). Family Home Care For Cancer—A Community-Based
Model. Year IV Executive Summary. Michigan State University.

Feetham, S.L. (1991). Conceptual and methodological issues in research of

families. In A.L.Whall & J. Fawcett (Eds), Fa_m_rly' theory development in nursing: State of
the science and art (pp. 55-68). Philadelphia: Davis.

Fogg, L., & Gross, D. (2000). Threats to validity in randomized clinical trials.
Research in Nursmg’ and Health, 33, 79-87.

Given, B. A., & Given, C.W. (1991-1996). Farm home-care for cancer: A
community-based model. National Institute of Nursing Research (NINR) and National

Cancer Institute (NCI), Grant #2R01 NR/CA l995-O3A3.

Given, B.A., Keilman, L.J., Collins, C., & Given, C. W. (1990). Strategies to

143

minimize attrition in longitudinal studies. Nursing Research 32, 184-186.

Goodman, J.S., & Blum, T.C. (1996). Assessing the non-random sampling effects
of subject attrition in longitudinal research. Journal of Management, _23, 627-652.

Grant, J.S., & DePew, DD. (1999). Recruiting and retaining research participants
for a clinical intervention study. Journal of Neuroscience Nursing, 31, 357-361.

Grifi‘ore, RJ., & Phenice, LA. (2001). The mlage of human ecology: A
general gstems pgpsmtive. Dubuque, IA: Kendall/Hunt.

Groger, L., Mayberry, P. S., & Straker, J. K (1999). What we didn’t learn
because of who would not talk to us. Qualitative Health Research, _9_, 829-835.

Groves, R M. (1989). Surveyerrors_ and survey costs. New York: Wiley.

Groves, R M. (1991). Measurement error across the disciplines. In P. P. Biemer,
R. M. Groves, L. E. Lyberg, N. A. Mathiowetz, & S. Sudman (Eds), Measurement errogs
in survefi, (pp. 1-25). New York: Wiley.

Groves, R M., Cialdini, R B., & Couper, M. P. (1992). Understanding the
decision to participate in a survey. Public Opinion May, 36, 475-495.

Groves, R M., & Couper, M. P. (1998). A conceptual framework for survey
participation Nonre_sp_qnse in Household Interview Surveys (pp.25-46). New York: Wiley.

Hancock, T. (1993). Health, human development and the community ecosystem:
three ecological models. Haalth Promotion International, _8_, 41-47

Harrington, P., & Nicoteri, J. (1999). The human element in data collection: How
to avoid the pitfalls. Journal of the American Academy 10f Nurse Practitioners, 11, 161-3.

Hautman, M. A., & Bomar, P. (1995). Interactional model for recruiting ethnically
diverse research participants. The Journal of Muflcultuaal Nursm' & Health, 1(4), 8-27.

Hayslip, B., McCoy-Roberts, L., & Pavur, R. (1998). Selective attrition efl'ects in
bereavement research: A three-year longitudinal analysis. OMEGA, 33, 21 -35.

 

Herzog, A.R., & Kulka, R A. (1989). Telephone and mail surveys with older
populations: A methodological overview. In M.P. Lawton, & A. R Herzog (Eds),

Smcial Research Methods for Gerontology (pp. 63-89). Amityville, NY: Baywood.

 

Herzog, A. R & Rodgers, W. L. (1988). Age and response rates to interview
sample surveys. Journal of Gerontolcgy, Q, 8200-8205.

144

Herzog, A. R, Rodgers, W. L., & Kulka, R A. (1983). Interviewing older adults:
A comparison of telephone and face-to-face modalities. Public Opinion Quarterly. 51],
405-418.

Heymann, S. J. (1995). Patients in research: Not just subjects, but partners.
Science. 269. 797-800.

Hinds, P.S., Quargnenti, A., & Madison, J. (1995). Refusal to participate in clinical
nursing research. Western Jourrgl of Nursing Research, 11, 232-236.

Holder, B., Tumer-Musa, J., Kimmel, P.L., Alleyne, S., Kobrin, S., Sirnmens, S.,
Cruz 1., & Reiss, D. (1998). Engagement of African American families in research on
chronic ilhress: A multisystem recruitment approach. me' Process, 32, 127-151.

Hooks, P. C., Tsong, Y., Baranowski, T., Henske, J. C., Nader, P. R, & Levin, J.
S. (1988). Recruitment strategies for multiethnic family and community health research.

Pam Comm] Hglfla, 11, 48-59.

Hosmer, D.W., & Lemeshow, S. (1990). Applied Logistic Reggssion. New York:
Wiley.

Jackson, J .S. (1989). Methodological issues in survey research on older minority
adults. In M.P. Lawton, & A. R Herzog (Eds), S cial Research Metho for
Gerontology (pp. 137-161). Amityville, NY: Baywood.

Jacobs, J. (1993). Family therapy in the context of chronic medical illness. In A.
Stoudemire & B.S. Fogel (Eds), Psychiatric Care of th_e Medical PatiLnt (pp.l9-30). New
York: Oxford University Press.

Janus, M., & Goldberg, S. (1997). Factors influencing family participation in a
longitudinal study: Comparison of pediatric and healthy samples. Journal of Pediatric
chhology, 32, 245-262.

Johnson, DR (1988). Panel analysis in family studies, Journal of Mm’ e and the
Fj1n_ily, fl, 949-955.

Johnson, D. R. (1995). Alternative methods for the quantitative analysis of panel
data in family research: Pooled time-series models. Joumal of Mm‘ e and the Farrah,
3‘1, 1065-1077.

Kamey, B.R., Davila, J., Cohan, C.L., Sullivan, KT., Johnson, M.D., & Bradbury,
TN. (1995). An empirical investigation of sampling strategies in marital research. Journal

of M_a_rg_ag° e and the Far—nth , 51, 909-920.
Kazdin, A.E., Stolar, M.J., & Marciano, PL. (1995). Risk factors for dropping out

145

of treatment among White and Black families. Journal of Family Psychology, _9_, 402-417.

Killien, M., & Newton, K (1990). Longitudinal research: The challenge of
maintaining continued involvement of participants. Western Jouma1 of Nursirg Researah,
112, 689-692.

Lengacher, C.A., Gonzalez, L., Giuliano, R, Bennett, M.P., Cox, C. E., &
Reintgen, D. S. (2001). The process of clinical trials: A model for successful clinical trial
participation Oncology Nursing Forum, E, 1115-1120.

Lebowitz, B.D. (1989). Scientific change and longitudinal research subjects,
methods, and environments. In M.P. Lawton, & A. R Herzog (Eds), Smial Research
Methpds for Gerontology (pp. 309-317). Amityville, NY: Baywood.

Leventhal, H., Leventhal, B.A., & Nguyen, T.V. (1985). Reactions of families to
illness: Theoretical models and perspectives. In DC. Turk & RD. Kems (Eds), Health,
Illness and Families (pp. 108-145). New York: Wiley.

 

Lewis, F .M. (1990). Perspectives on models of interpersonal health behavior. In K

Glanz., F.M. Lewis, & Rimer, B.K. (Eds), Health behavior and health fiucation: Theory,
research, and practice (pp. 242-251). San Francisco: Josey-Bass.

Lillie-Blanton, M., & Hoffman, S. C. (1995). Conducting an assessment of health
needs and resources in a racial/ethnic minority community. Health Services Remh, 30,
225-236.

Long, J. S. (1997). Regggssion Mode1s for Categorical and Limited erehdent
Variables. Thousand Oaks, CA: Sage.

 

Maguire, M.C. (1999). Treating the dyad as the unit of analysis: A primer on three
analytic approaches. Journal of Marriage and the Family, 61, 213-223.

Manfredi, C., Lacey, L., Warnecke, R, & Balch, R (1997). Method efi‘ects in
survey and focus group findings: Understanding smoking cessation in low-SES Afiican-
American women. _Hgalth Education & Behavior, _2_4_, 786-800.

Mark, B. A., Geddes, N., Salyer, J., & Smith, C. S. (1999). Managing the
informed consent process in large multicenter projects. Nursing Research, 53, 114-117.

Martin, P. A. (1995). Recruitment of research subjects. Applifl Nursmg' Raggeh,
3, 50-54.

McHomey, CA. (1996). Measuring and monitoring general health status in elderly

persons: Practical and methodological issues in using the SF-36 health survey. _Tha
Gerontologist, 36, 571-583.

146

McHomey, C.A., Ware, J.E., & Raczek, A.E. (1993). The MOS 36-item short-
forrn health survey (SF-36): II. Psychometric and clinical tests of validity in measuring
physical and mental health constructs. Medical Care, 31, 247-263.

McKenzie, J. F., & Jurs, J. L. (1993). le’ g, m lementing, and evaluatihg'
health promotion programs. New York: McMillan.

McKenzie, M., Tulsky, J.P., Long, H.L., Chesney, M., & Moss, A. (1999).
Tracking and follow-up of marginalized populations: A review. Joumi of Health Care for
aha Poor and Underserved. 1_O_, 409-429.

McNeely, E. A., & Clements, S. D. (1994). Recruitment and retention of the older
adult into research studies. Jouanal of Neuroscience nursrhg' , 26, 57-61.5

Mihelic, A. H. & Crimmins, E. M. (1997). Loss to follow-up in a sample of
Americans 70 years of age and older: The LSOA 1984-1990. Journal of Gerontology,
3233, 837-848.

Miller, R B. & Wright, D. W. (1995). Detecting and correcting attrition bias in
longitudinal family research. Joumal of Mm’ e and the Family, fl, 921-929.

Mock, V., Hill, M., Dienemann, J., Grimm, P., & Shivnan, J. (1996). Challenges to
behavioral research in oncology. Cancer Practice: A Muhidiscgp' linary Journal of Cancer
Care, 5, 267-273.

Morgan, S. P., & Teachman, JD. (1988). Logistic regression: Description,
examples, and comparisons. Journal of mg“ e and the Family, fl, 929-936.

Moser, D.K, Dracup, K, & Doering, L.V. (2000). Factors differentiating
dropouts fiom completers in a longitudinal, multicenter clinical trial. Nursithesearah,
_42, 109-116.

Motzer, S. A., Moseley, J. R., & Lewis, F. M. (1997). Recruitment and retention
of families in clinical trials with longitudinal designs. Western Journal of Nursing
Research, fl, 314-333.

Napoles-Springer, A.M., Grumbach, K, Alexander, M., Moreno-John, G., Forte,
D., Rangel-Lugo, M., Perez-Stable, El (2000). Clinical research with older Afi'ican
Americans and Latinos: Perspectives fiom the community. Research on Agmg' , 33, 668-
691.

National Cancer Institute. (1994). SEER Cancer Statistics Review. (NIH
Publication No. 94-2789). Washington, DC: US. Government Printing Office.

147

National Institutes of Health, Ofiice for Protection from Research Risks (1991).
Title 45 Code of Federal Regulations Part 46: Protection of Human Subjects (OPRR
Reports 45 CFR 46).

NeunErk, DE. (2000). A review of longitudinal investigations of caregiving in
families with cancer: Gathering evidence for an ecological model of nonparticipation and
attrition. Unpublished manuscript.

Neumark, D. E., Given, B., & Given, C. W. (1997). Recruitment and retention of
elderly cancer patients in a community-based Emily caregiving study. Poster session
presented at the 22"" Annual Congress of the Oncology Nursing Society, New Orleans,
LA.

Neumark, D. E., Stommel, M., Given, C. W., & Given, B. (2001). Research
Design and Subject Characteristics Predicting Nonparticipation in a Panel Survey of
Families with Cancer. Nursmg' Research, _5_Q, 363-368.

Neumark, D. E., Stommel, M., Given, C. W., & Given, B. (1998). Attrition of
elderly cancer patients in a longitudinal survey--Who drops out, why, and what can we do
about it? Poster session presented at the 23"1 Annual Congress of the Oncology Nursing
Society, Columbus, OH.

Nokes, K M, & Dolan, M. S. (1992). Experiences of nurse-researchers in gaining
access to subjects for clinical nursing research. Journal of Professional Nursmg' , 3, 115-
119.

Norris, F. H. (1985). Characteristics of older nonrespondents over five waves of a
panel study. Journal of Gerontology, £9, 627-636.

Norris, F. H. (1987). Effects of attrition on relationships between variables in
surveys of older adults. Journal of Gerontology, 43, 597-605.

Novelli, W.D. (1990). Applying social marketing to health promotion and disease
prevention. . In K. Glanz., F.M. Lewis, & Rimer, B.K. (Eds), Health behavior and health
education: Theory, research, and practice (pp. 342-369). San Francisco: Josey-Bass.

Padula, C. (1996). Older couples’ decision making on health issues. Western
Journal of Nursihg_Research. 3, 675-680.

 

Padula, C. A (1997). Predictors of participation in health promotion activities by
elderly couples. Journal of Family Nursmg' , 3, 88-106.

Patrick, J.H., Pruchno. R. A., & Rose, MS. (1998). Recruiting research
participants: A comparison of the costs and effectiveness of five recruitment strategies.
The Gerontologist, 33, 295-302.

148

Perry, C.L., Baranowski, T., & Parcel, GS. (1990). How individuals,
environments, and health behavior interact: Social learning theory. In K. Glanz., F.M.

Lewis, & Rimer, B.K (Eds), Health behavior and health education: Theory, research, and
practice (pp. 161-186). San Francisco: Josey-Bass.

Peterson, E. (1998). Improving Recruitment and Retention of African Americans
into Clinical Trials. Cultural Diversity Research Program. Student Research Abstrgaa
1997 and 1996. Michigan State University: College of Human Medicine.

Porter, E. J. (1999). Defining the eligible, accessible population for a
phenomenological study. Western Journal of Nursmg' Research, 21, 796-804. .1

Powell, D. R (1988). Client characteristics and the design of community-based

intervention programs. In A.R Pence (Ed), Ecological research with children and famili§§
From Concepts to Methodology (pp. 122-142). New York: Teachers College.

 

Pruchno, RA., & Hayden, J.M. (2000). Interview modality: Efl‘ects on costs and
data quality in a sample of older women. Journal of Agmg' and Health, fl, 3-24.

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 tracking strategies: Review and recommendations. Evaluation

W 12. 1-25.

Rinck, G. C., van den Bos, G., de Haes, H., Schadé, E., & Veenhof, Cees. (1997).
Methodological issues in efl‘ectiveness research on palliative cancer care: A systematic
review. JouM of Clinical Oncology, Q, 1697-1707.

Ryan, E. A, & Hayman, L. L. (1996). The role of the Emily coordinator in
longitudinal research: Strategies to recruit and retain Emilies. Journal of Family Nursing,
2, 325-329.

Sabatelli, R M., & Bartle, S. E. (1995). Survey approaches to the assessment of
Emily functioning: Conceptual, operational, and analytical issues. Journal of Marriage and
the FM’ , 32, 1025-1039.

Saylor, C.F., Elksnin, N., Farah, B.A., & Pope, J. A. (1990). Depends on who you
ask: What maximizes participation of Emilies in early intervention programs. Journal of
Pediatric Psychology, Q, 557-569.

Seigley, L. (1998). The efi‘ects of personal and environmental factors on health
behaviors of older adults. Nursmg' Connections, 11(4), 47-58.

Sharrm, S.K., Tobin, J.D., & Brant, LJ. (1989). Attrition in the Baltimore

149

longitudinal study of aging during the first twenty years. . In M.P. Lawton, & A. R.
Herzog (Eds), Spacial Resegch Methods for Gerontology (ch. 9). Amityville, NY:
Baywood.

Shaw, W. S., Cronan, T. A., & Christie, M. D. (1994). Predictors of attrition in
health intervention research among older subjects with osteoarthritis. Health Pachology,
13, 421-431.

Sjoberg, G., & Nett, R (1968). A Met_h_odology for Soaial Reseaach New York:
Harper & Row.

Souder, J. E. (1992). The consumer approach to recruitment of elder subjects.
Nursing Research, 41, 314-316.

Spoth, R, & Redmond, C. (1994). Effective recruitment of parents into Ernily-
focused prevention research: A comparison of two strategies. Pachology and thh, 2,
353-370.

StatCorp. (1999). Stata statistical software: Release 6.0. College Station, TX:
Stata Corporation. -

Stoep, A.V., Williams, M., Jo nes, R, Green, L., & Trupin, E. (1999). Families as
full research partners: What’s in it for us? Journal of Behavioral Health Services &
Research, 26, 329-344.

 

Sullivan, C. M., Rumptz, M. H., Campbell, R, Eby, K K, Davidson, W. S.
(1996). Retaining participants in longitudinal community research: A comprehensive
protocol. Journal of Applied Behflioral Scieng, 32, 262-276.

Swanson, G.M., & Ward, AJ. (1995). Recruiting minorities into clinical trials:
Toward a participant-fi'iendly system. Journal of the National Cancer Institute, fl,1747-
1759.

Thomas, RB. (1987). Methodological issues and problems in Emily health care
research. Journal of Mm’ e and the me , Q, 65-70

Thompson, M. G., Heller, K, & Rody, C. A. (1994). Recruitment challenges in
studying late-life depression: Do community samples adequately represent depressed older

adults? Paychology and Agm , 1, 121-125.

Thornton, A., Freedman, D. S., & Camburn, D. (1982). Obtaining respondent
cooperation in family panel studies. Sociological Methods & Research. 11(1), 33-51.

Underwood, S. (2001). Minorities, women, and clinical cancer research: The
charge, promise, and challenge. Annals of Epidemiology, 1Q(Suppl 8), S3-812.

150

Underwood, 8., & Alexander, G.A. (Eds), (2001). Participation of minorities and
women in clinical cancer research. Anhals of Epidemiolcgy, _1_Q(Suppl. 8).

Underwood, S., Sander, E., & Davis, M. (1993). Determinants of participation in
state-of-the-art cancer prevention, early detecting/screening, and treatment trials among
Afi'ican-Americans. Cancer Nursing, 16, 25-33.

Ware, J.E., & Sherboume, CD. (1992). The MOS 36-item short-form health
survey (SF-36): 1. Conceptual fi'amework and item selection. Medical Care, _3_()_, 473-483.

Wideen, M., Mayer-Smith, J., & Moon, B. (1998). A critical analysis of the 2*
research on learning to teach: Making the case for an ecological perspective on inquiry.
Revaiew of Educational Research _63, 130-178.

 

Williams, S. G. (1992). Research considerations: Family opinions about elderly
relatives in research Journal of Gerontological Nursing, 13(12), 3-8.

Witte, K, Meyer, 6., Bidol, H., Casey, M.K, Kopfinan, J., Maduschke, K,
Marshall, A., Morrison, K, Ribisl, KM., & Robbins, S. (1996). Bringing order to chaos:
Communication and health. Communication Studies, 41, 229-242.

Worth, A., & Tierney, A. J. (1993). Conducting research interviews with elderly
people by telephone. Journal of Advanced Nursmg' , 13, 1077-1084.

Wuest, J. (1993). Removing the shackles: A feminist critique of noncompliance.
Nursing Outlook, fl, 217-24.

Yin, R. K (1993). Applications of case study reseaach. Thousand Oaks, CA: Sage

Zapka, J. G., Chasan-Taber, L., Bigelow, C., & Hurley, T. (1994).
Methodological issues for health-related surveys of multicultural older women.
Evaluation & The Health Profes_sions. Ll, 485-500.

Zimmer, A. W., Calkins, E., Hadley, E., Ostfeld, A.M., Kaye, J. M., Kay, D.

(1985). Conducting clinical research in geriatric populations. Annals of Internal Medicine,
m, 276-283.

151