Journal of Gerontology: MEDICAL SCIENCES
1998. Vol. 53A. No. I. M39-M46
Copyright 1998 by The Gerontological Society of America
Random Versus Volunteer Selection
for a Community-Based Study
Mary Ganguli,12 Mary E. Lytle,2 Maureen D. Reynolds,1 and Hiroko H. Dodge2
'Division of Geriatrics and Neuropsychiatry, Department of Psychiatry, School of Medicine,
and department of Epidemiology, Graduate School of Public Health, University of Pittsburgh, Pittsburgh, Pennsylvania.
Background. Selection methods vary greatly in ease and cost-effectiveness. The effects of selection factors associated with subjects' recruitment into studies can introduce bias and seriously limit the generalizability of results.
Methods. For an epidemiologic study, we recruited an age-stratified random sample of 1,422 community-dwelling
individuals aged 65+ years from the voter registration lists in a rural area of southwestern Pennsylvania. The first 1,366
of these were accrued through intensive recruitment efforts; the last 56 of them responded to a single mailing. To increase sample size for future risk factor analyses, we also recruited by direct advertisement a sample of 259 volunteers
from the same area. The three groups were compared on selected baseline characteristics and subsequent mortality.
Results. The two subgroups of the random sample were not significantly different on any of the variables we examined. Compared to the random sample, in cross-sectional analyses, volunteers were significantly more likely to be
women, more educated, and less likely to have used several health and human services. Volunteers also had higher cognitive test scores and Instrumental Activities of Daily Living (IADL) ability. Over 6-8 years (10,861 person-years) of
follow-up, volunteers had significantly lower mortality rates than randomly selected subjects.
Conclusions. Health-related studies with populations composed partly or entirely of volunteers should take potential
volunteer bias into account when analyzing and interpreting data.
demographic projections for the twenty-first
CURRENT
century anticipate a dramatic increase in the population over age 65 (1). Planning for the service needs for a
growing elderly population, recommendations for specific
interventions, as well as the development of etiopathogenic
hypotheses about aging-related disorders, will be based on
results from descriptive and experimental studies of the elderly. Such studies employ a variety of sampling and recruitment methods, but empirical data are largely lacking
on how these methods affect the interpretation of study results. Selection methods vary greatly in ease and costeffectiveness, yet the effects of selection factors associated
with subjects' inclusion in studies can seriously limit the
generalizability of results.
Random or otherwise representative samples are considerably more difficult and expensive to recruit than are volunteer samples. To the extent that volunteers differ from
nonvolunteers, however, parameters estimated from volunteer samples may be seriously biased, particularly in survey
research (2). We conducted a survey of the elderly members of a rural community as part of an effort to establish a
population-based dementia registry (3). The current article
presents data from the three randomly selected and volunteer subgroups of the sample who participated in the survey. We predicted that there would be significant differences among the random and volunteer groups, postulating
that these differences would help indicate the nature of volunteer bias. Our particular concerns were with variables
relevant to dementia, such as demographics, cognitive and
functional ability, and the use of health services.
METHODS
Sampling and Recruitment
The survey was conducted in the mid-Monongahela
Valley of southwestern Pennsylvania, an economically depressed rural area approximately 25 miles south of Pittsburgh. This population is largely blue-collar and of European descent. The project was named the Monongahela
Valley Independent Elders Survey (MoVIES). Entry criteria were: (a) age 65+ years at the time of study entry; (b)
being community-dwelling (i.e., not in a nursing home) at
the time of study entry; (c) fluency in English; and (d) at
least a grade 6 education — the last two requirements
being related to interpretability of neuropsychologic (cognitive) test data.
General information about the study and its objectives
was provided to the community through newspaper articles,
flyers, talk radio, information booths at local fairs, and presentations to various religious and secular community
groups. A series of 1:13 age-stratified (65-74, 75+) random
subsamples was drawn from a master sampling frame based
on the voter registration lists for the study area. Because
these communities have very stable populations, the electoral rolls were believed to be the most complete and least
biased sampling frame for the area, as they included subjects who had voted not just in the most recent election but
in past elections as well. Selected subjects were mailed
letters informing them that they had been randomly chosen
to participate in the study, asking them to mail back a form
indicating the most convenient times for them to be con-
M39
M40
GANGULIETAL
tacted, and informing them that they would receive a follow-up telephone call. Subjects who responded agreeing to
participate were scheduled for a home visit, during which
the study was explained in greater detail and informed consent was obtained. Subjects who could not be contacted by
letter or telephone were visited directly ("door knocking"),
so that an effort could be made to recruit them in person. In
some cases, several calls and/or "door knocking" visits
were made before contact could be made and consent to
participate was either obtained or denied. In some cases,
positive word-of-mouth publicity from those who had entered the study was helpful in recruiting individuals who
had initially been reluctant to participate. When it was felt
that maximal recruitment had been reached with a particular random subsample, another random subsample was
drawn from the master list, this process being repeated until
an adequate total sample size was attained. Thus, recruitment of the random sample was a labor- and time-intensive
process. Further details have been reported previously (3).
Using the methods described above, a total age-stratified
random sample of 1,366 individuals was recruited, and designated for this article as the random sample (RS).
When it was determined that an adequate random sample
had been accrued, recruitment was officially closed. However, recruiting letters had already been mailed out for the
next 1:13 random subsample. Each individual in this final
random subsample received only this single letter inviting
participation in the study; no follow-up recruiting letters,
calls, or visits were made. Fifty-six subjects, designated as
random volunteers (RV) for the purposes of this article, responded spontaneously and entered the study. We have no
information about the individuals who were sent, but did
not respond to, the final mailing. A general recruiting drive
was then publicized in the local media, asking for volunteers. The project was then deluged with calls and letters
from persons who had not been randomly selected but
wanted to participate. A total of 259 such individuals who
met study entry criteria were recruited into the study before
time ran out and recruitment had to be closed. They are
designated as direct volunteers (DV) for this article.
We were concerned that randomly selected individuals
who agreed to participate after receiving a single letter
might resemble the volunteers more than the rest of the random sample. Choosing to err on the side of caution, we initially demarcated the RS, RV, and DV as distinct subgroups
of the Mo VIES cohort. Our intention was to pool the RV
and DV groups with the RS group for risk factor determination but not for prevalence or incidence estimation.
Data Collection
Subjects were interviewed at their own residences unless
they expressed a preference to be interviewed at the project
offices. After obtaining informed consent, trained interviewers administered a battery of cognitive tests (lasting
approximately 25 min in an intact subject), which have
been described in detail elsewhere (3,4) and which included
a general mental status test, the Mini Mental State Examination (MMSE) (5). Subjects were also asked to provide information about a variety of characteristics, including demographics, living arrangements, functional (Instrumental
Activities of Daily Living, IADL) ability (6), use of health
and human services, and use of prescription drugs (7).
The MoVIES survey has since become a prospective
cohort study, as part of which the project has followed all
surviving and consenting participants through visits, telephone calls, and letters. Mortality is tracked through these
means as well as through daily scanning of obituaries in
six local newspapers. Dates, places, and listed causes of
deaths are eventually confirmed through abstraction of
death certificates.
Statistical Methods
Cross-sectional analyses. — Random sample (RS) and
direct volunteer (DV) subjects were compared to identify
differences, if any, between subjects randomly sampled
from the voter registration lists (RS) and subjects who volunteered in response to advertisements (DV).
Random sample (RS) and random volunteer (RV) groups
were compared to identify differences, if any, between randomly sampled subjects who agreed to participate with
minimal recruitment efforts (RV), and the randomly sampled subjects in the more intensively recruited prevalence/
incidence cohort (RS).
Between-group differences in proportions with selected
characteristics at study entry were examined univariately
between RS and RV samples, and between RS and DV
samples. The variables examined were as follows: demographics: age, sex, education, race; social characteristics:
living arrangements (living alone, living at home with
others, group living situations), marital status, previous
occupation; health service utilization: usual source of
health care, hospitalization during the past 6 months, personal care or nursing home use in the past, use of professional health care at home in the past year, current use of
social services, number of current prescription drugs; functional abilities: IADL total score (6), possible score ranging
from 11 to 45 (dichotomized as no deficit vs any deficit, or
11 vs 12+, and also treated as a continuous variable);
whether still driving a motor vehicle; cognitive functioning:
MMSE scores, possible scores ranging from 0 to 30,
(dichotomized as 0 to 23 and 24+, also treated as a continuous variable).
The chi-square test for association, or Fisher's exact test
if the expected values for any cell was less than 5, was used
to assess statistical significance defined as p < .05. For variables that were continuous in nature (age, IADL, and
MMSE scores), we calculated mean, standard deviation
(SD), and median. Pairwise between-group comparisons of
continuous variables were performed using the nonparametric Wilcoxon rank sum tests, as these continuous variables had highly skewed distributions.
Because the large number of statistical tests performed in
this exploratory study increases the Type I error rate for
each individual test, the reader may choose to interpret p
values using a different criterion. The establishment of an
appropriate value for reporting statistical significance is
difficult, if not impossible, because many unreported exploratory comparisons were performed while examining
these data. However, based on the number of comparisons
RANDOM VS VOLUNTEER SELECTION
reported in this article, a Bonferroni adjustment would
yield a criterion of p = .0014.
Longitudinal analysis. — Given the similarity on all
variables between RS and RV groups (see below), they
were combined for longitudinal comparisons between random (RS + RV) and volunteer (DV) groups.
To examine differences in mortality between random and
volunteer groups, first, we examined age-specific and overall mortality rates per 100 person-years for each group.
Second, we used the proportional hazards model (8), for
which time was measured from the date of study entry to
his or her death. If subjects were still alive, or had dropped
out of the survey, time from the first contact until the dates
when the subjects were last observed were included as censored cases. Control variables included in the model were
age (continuous variable), gender, and education (at least
high school versus less than high school), because these
variables are known to be significant predictors of mortality
and (as reported below) the above cross-sectional analyses
showed that DV were more likely than RS to be younger,
female, and to have higher educational levels. A second
proportional hazards model was also fitted including interaction terms between DV*age, DV*gender, and DV*education, in addition to participant status (random vs volunteer).
A third model, similar to the first, used all-cause attrition
(i.e., not restricted to mortality) as the outcome variable.
RESULTS
For the cross-sectional analyses, three groups were compared: random sample (RS) (n = 1,366), random volunteer
sample (RV) (n = 56), and direct volunteer sample (DV)
(n = 259). Tables 1-4 show the proportions of each group
with various selected characteristics, and the results of tests
for significant differences in proportions between (a) RS
and RV groups and (b) RS and DV groups. The tables show
p values <.05, and also indicate p values that are significant
after adjustment for multiple comparisons at/? < .0014.
M41
Demographic characteristics (age, gender, education,
race) of the three groups are shown in Table 1. There were no
statistically significant differences in proportions between RS
and RV samples. Compared to RS subjects, DV subjects
were significantly more likely to be in a younger age category; age distribution was also significantly (p = .002,
Wilcoxon rank sum test) different between RS and DV (not
shown in the table). RS, RV, and DV subjects had mean (SD)
ages of 73.1 (6.0), 73.9 (5.8), and 71.8 (5.3) years, respectively. DV were more likely to be female and better educated.
Social characteristics (marital status, living arrangements, previous occupation) of the three groups are shown
in Table 2. These variables also showed no significant differences between RS and RV. Male DV were significantly
more likely than male RS to have previously held professional/technical positions; occupational class differences
among women were not detectable, probably because the
majority had been (and still were) housewives.
Cognitive (MMSE) (5), functional (IADL) (6), and driving status are shown in Table 3. While there were no significant differences between RS and RV group, DV were significantly more likely than RS to have MMSE scores above 24.
However, mean (SD) MMSE scores in the RS, RV, and DV
groups were 27.1 (3.0), 27.3 (2.1), and 27.5 (2.6); MMSE
distributions were not significantly different (Wilcoxon rank
sum test) between RV and DV groups (p = .318).
Functional (IADL) scores, when categorized, were not
significantly different between RS and RV groups, while
the DV group had significantly better IADL scores than the
RS group. The distribution of IADL scores was also significantly different (p = .016, Wilcoxon test) between DV and
RS groups, even though mean (SD) IADL scores were
close among RS, RV, and DV groups: 12.5 (4.5), 12.0 (3.0),
and 11.5 (1.7). With respect to automobile driving, a much
larger proportion of women than men reported never having
driven. However, among both men and women who did
previously drive, significantly greater proportions of RS
than DV had quit driving before study entry.
Table 1. Demographic Characteristics: Percentage of Subjects in Each Group (column %)
Random
Sample
(RS)
(n= 1,366)
Random
Volunteers
(RV)
(n = 56)
Direct
Volunteers
(DV)
(n = 259)
Age
<75
75-84
85+
64.1
31.0
4.9
58.9
32.1
8.9
Gender
Men
Women
45.4
54.6
Education
< High school grad
£ High school grad
Race
Caucasian
Other
Subject Characteristics
(/V= 1,681)
RS vs RV
P-
RS vs DV
P=
72.6
25.5
1.9
0.375
0.011*
42.9
57.1
25.5
74.5
0.709
0.001**
45.8
54.2
39.3
60.7
30.9
69.1
0.335
0.001**
96.6
3.4
100.0
98.8
1.2
0.256
0.057
0.0
*p < .05; **p < .0014 (Bonferroni adjustment for multiple comparisons).
M42
GANGULIETAL
Table 2. Social Characteristics: Percentage of Subjects in Each Group (column %)
Subject Characteristics
(W = 1,681)
Marital status (n = 1,681)
Married
Other
Living arrangements (n = 1,679)'
Alone
With others at home
Group living situation
Previous occupation (n = 1,678)'
Men (n = 708)
Professional/technical
Other
Women (n = 970)
Professional/technical
Other (includes housewife)
Random
Sample
(RS)
(n= 1,366)
Random
Volunteers
(RV)
(n = 56)
Direct
Volunteers
(DV)
(n = 259)
57.0
43.0
62.5
37.5
59.1
40.9
0.417
0.542
31.0
68.1
31.7
68.3
0.0
0.792
0.398
0.9
33.9
66.1
0.0
0.204
0.012*
18.8
81.2
29.2
70.8
31.8
68.2
0.753
0.281
14.5
85.5
12.5
87.5
17.6
82.4
RS vs RV
P=
RS vs DV
P=
'Analyses based on n < 1,681 on variables with missing or incomplete data.
*p < .05; **p < .0014 (Bonferonni adjustment for multiple comparisons).
Table 3. Cogitive and Instrumental Functioning: Percentage of Subjects in Each Group (column %)
Random
Sample
(RS)
(n= 1,366)
Random
Volunteers
(RV)
(n = 56)
Direct
Volunteers
(DV)
(n = 259)
MMSE score2 (n= 1,671)'
<23
24+
8.2
91.8
5.5
94.6
3.1
96.9
IADL score3 (n= 1,674)'
11
12+
78.3
21.7
80.4
19.6
83.8
16.2
2.9
12.2
84.9
4.2
4.2
91.7
0.0
3.0
97.0
29.6
21.1
49.4
25.0
21.9
53.1
21.8
13.5
64.8
Subject Characteristics
(N= 1,681)
Driving automobile (n = 1,673)'
Men (n = 707)
Never drove
Quit driving
Still driving
Women (n = 966)
Never drove
Quite driving
Still driving
RS vs RV
P=
RS vs DV
P=
0.618
0.004*
0.713
0.046*
0.376
0.020*
0.855
0.001**
'Analyses based on n < 1,681 on variables with missing or incomplete data.
Mini Mental State exam (5).
'Instrumental Activities of Daily Living (6).
*p < .05; **p < .0014 (Bonferroni adjustment for multiple comparisons).
:
Table 4 shows the utilization of health and human
services and the regular use of prescription medications
among the three groups. Again, there were no statistically
significant differences between RS and RV groups. DV subjects were significantly less likely than RS subjects to report no usual source of health care (no identified primary
care physician), to have been hospitalized in the previous 6
months, to have ever been in a personal care home or nursing home, to have had home health care in the previous
year, and to be currently using social services. There was,,
however, no significant difference with respect to number
of regularly used prescription drugs.
Longitudinal Analyses
For the mortality analyses, two groups were compared:
random (RS + RV) and volunteer (DV). Between the beginning of the survey (April 1987) and July 1996 (10,861 total
person-years), we observed 531 total deaths: 495 random
subjects (481 RS and 14 RV) and 36 volunteers. Table 5
shows age-specific and overall mortality rates per 100
person-years in each group. Overall mortality rates were
significantly lower (p < .001) among the volunteers than
among randomly selected subjects.
In a Cox proportional hazards model, risk of dying (risk
ratio 0.484, p = .0001) was significantly lower among vol-
RANDOM VS VOLUNTEER SELECTION
M43
Table 4. Recent Use of Health Services and Current Use of Prescription Drugs: Percentage of Subjects in Each Group (column %)
Random
Sample
(RS)
(«= 1,366)
Random
Volunteers
(RV)
(n = 56)
Direct
Volunteers
(DV)
(n =259)
Usual source of health care (n = 1,677)'
No doctor
All other
5.4
94.6
5.4
94.6
2.3
97.7
Hospitalization past 6 months (n = 1,677)'
Yes
No
11.2
88.8
14.3
85.7
5.0
95.0
Previous (ever) stay in nursing or personal
care home(n = 1,672)'
Yes
No
1.6
98.4
1.8
98.2
0.0
100.0
Home health care previous year(/j = 1,677)'
Yes
No
5.2
94.8
7.1
92.9
2.3
97.7
Current social services (n = 1,675)'
Yes
No
3.2
96.8
3.6
96.4
0.8
99.2
Current prescription drug use(/i = 1,681)
0
1-4
5+
29.4
59.2
11.4
26.8
60.7
12.5
31.3
60.6
8.1
Subject Characteristics
(N= 1,681)'
RS vs RV
P=
RS vs DV
P=
0.980
0.034*
0.480
0.002*
0.923
0.037*
0.527
0.045*
0.863
0.033*
0.900
0.297*
'Analyses based on n < 1,681 on variables with missing or incomplete data.
*p < .05; **p < .0014 (Bonferroni adjustment for multiple comparisons).
Table 5. Total Mortality Rates per 100 Person-Years (95% CI)
Age
Random Subjects
(RS + RV)
Mortality Rate
(95% CI)
Volunteers
(DV)
Mortality Rate
(95% CI)
65-74
75-79
80-84
85+
2.9 (2.4, 3.5)
4.5 (3.8, 5.4)
7.8 (6.5, 9.3)
15.9(13.4, 18.9)
1.3(0.6,2.3)
3.0(1.6,5.0)
1.6(0.4,4.0)
6.0(0.0, 12.5)
Overall
5.4 (5.0, 5.9)
2.1 (1.5,2.9)
p Value
p < .001
/?<.O4
p<.00l
p<.00\
p<M\
unteer than among random subjects. Of the control variables, factors predicting mortality were greater age (for
each year, risk ratio 1.1, p - .0001) and male gender (risk
ratio 1.68, p = .0001), but not education. When the model
was rerun including interactions between DV status and
age, gender, and education, no interactions were significant;
that is, the effects of age, gender, and education on mortality were similar among volunteer and random samples.
We also examined a proportional hazards model in which
the outcome variable was dropouts for any reason (not restricted to death). This added a further 169 dropouts to the
531 deaths, bringing the total to 700 events. Results were
similar to those of the mortality model: significant predictors of dropout were age (risk ratio 1.1, p = .0001) and male
gender (risk ratio 1.56, p = .0001); volunteer status reduced
the risk of dropout (risk ratio = 0.553, p = .0001).
DISCUSSION
Summary of Findings
In a cohort drawn from a rural community, we have compared two subgroups who were selected by random sampling, with greater and lesser ease of recruitment, and one
subgroup that was composed entirely of volunteers. Our
major finding from the univariate analyses is that, compared to the random sample, the volunteers were significantly younger, better educated, more likely to be women,
with higher functional and perhaps cognitive ability, and
lesser use of health and human services. (We note that after
adjusting for the multiple comparisons reported here, only
the differences in gender and education met the more stringent test of significance.) Several of these differences, as
well as the significantly greater survival rates we found
among our volunteers, suggest that, in general, those who
volunteer for research are a healthier group than the general
population from which they come. As would be expected,
those who volunteered in the first place were less likely to
drop out of the study for any reason.
Volunteer "Bias"
All consenting research participants are "volunteers" to a
greater or lesser extent, but may differ in the degree of bias
they introduce into a study. In their classic volume, Rosenthai and Rosnow stated (2, p. 1): "There is a growing suspicion among behavioral researchers that those human subjects who find their way into the role of research subject
may not be entirely representative of humans in general."
M44
GANGULIETAL
They summarized the results of several studies showing
that research volunteers tend to be younger and female,
with higher educational and occupational attainment. Two
decades later, many cohort and case-control studies still
have to resort to recruiting volunteers because of the sometimes prohibitive cost of other strategies. We are not aware
of other published attempts to compare random samples
with volunteer samples. Our findings regarding volunteer
bias parallel many, but not all, of those studies examining
response bias by comparing elderly study responders with
nonresponders. Norton et al. (9) found initial nonresponders (who later agreed to participate) to have lesser education and lower cognitive (MMSE) scores. Hebert et al. (10)
found nonresponders to be more cognitively impaired,
more disabled, and to have a higher 1-year mortality rate
than responders. Koval et al. (11) found nonresponders to
be older, more likely to be women, more likely to be married, to have had less education, and to be less likely to
have had a white-collar job. Launer et al. (12) found nonresponders to be more likely to be unmarried, with lesser
education and lower cognitive status, and to report higher
rates of physical and mental illness. The similarities among
characteristics of responders (as compared to nonresponders) in the literature and volunteers (as compared to randomly selected subjects) in our study intuitively suggests a
continuum of responsiveness from outright volunteers to
"draftees" to refusers. An interesting contrast is provided
by Ives et al. (13), who identified three groups initially selected for a trial of preventive health services: participants,
refusers, and individuals who could not be reached. In that
study, refusers appeared to be the healthiest, followed by
participants, followed by unreachables.
In experimental studies, including intervention trials, the
selection bias inherent in volunteer- or referral-based samples is typically countered by randomization and matching
to controls, who for the most part are also volunteers. Confounding, unlike bias, cannot be countered as readily. Poon
et al. (14) pointed out that confounding variables are not
randomly distributed among cases and controls, nor can
they be randomized. Nonvolunteer controls, such as hospital controls, are subject to Berkson's bias. The growing
recognition of the importance of random sampling and
community-based studies lies in their perceived ability to
overcome selection bias. However, it is well known that
research response rates in the elderly decrease with age
(15), and recruitment of a random sample can be difficult,
expensive, and of low yield. Thus, many well-known studies (16,17) of the elderly have relied on volunteer samples.
Potential Bias in Dementia Studies
The ongoing Mo VIES study is focused primarily on
prevalence, incidence, risk factors, and outcome in dementia. In our sample, on the cognitive functioning measure
(MMSE), significant differences between random and volunteer groups were found when the scores were dichotomized as 0-23 and 24+ (the conventional clinical cutpoint),
but not when MMSE was treated as a continuous variable
so that its overall distribution could be compared between
the groups. The mean and median MMSE scores were almost identical between groups, but the skewness of the
MMSE score distributions made it difficult simply to compare means. The discrepancy in results between continuous
and categorical treatment of the MMSE score is most likely
a function of dichotomous classification (above or below a
semiarbitrary screening cutpoint). Thus, although volunteer
subjects may not be less impaired in a general sense, they
would be less likely than randomly selected subjects to fall
below this predetermined cutpoint. The consequences
would depend on the specific purpose of cognitive screening in a given study; for example, in the Mo VIES study,
volunteers would have been less likely than random sample
subjects to be selected for the detailed clinical assessment
designed to determine whether or not they were prevalent
cases of dementia. Thus, the MoVIES study could not have
included the volunteer sample in analyses involving characteristics at the time of recruitment (e.g., for prevalence estimation), because of the expected bias; for example, individuals with cognitive impairment are less likely to be able or
willing to volunteer.
We recruited the volunteer sample only to enlarge the
overall cohort for future analyses of risk factors for conditions that were absent at the time of recruitment. Thus, in a
combined cohort of random plus volunteer subjects who
were not demented at study entry, we could prospectively
determine risk factors for (future) incident dementia. However, concern is still raised by the fact that several of the
characteristics (lesser age, greater education, higher cognitive and functional ability, lower mortality) of volunteers
are known to be inversely associated with the presence of
dementia (15,18-21). Thus, by including volunteers in our
sample for risk factor analyses, we might have increased
power at the expense of introducing a systematic bias.
Rosenthal and Rosnow (2) observed that researchers often
suspect bias but are unable to quantify it. One solution for
our cohort is to pool the volunteer and random subsamples
for risk factor analyses, but to stratify the analyses by subjects' participant type (volunteer vs random status) so as to
account for potential bias. Another possibility is to include
participant status in the model and look for interactions between participant status and suspected risk factors; this approach has been demonstrated in the mortality analyses reported in this article.
Much of what is known about risk/protective factors for
Alzheimer's and other dementias is based on retrospective
case-control studies of patients referred (or volunteering)
for dementia research — for example, in Alzheimer centers
and memory disorder clinics at major urban centers. For example, at the Mayo Clinic in Rochester, Minnesota (22),
patients with Alzheimer's disease who lived locally and received their primary care at the Mayo Clinic had dementia
of later onset, were less educated, more often lived alone,
and were more often institutionalized than were those who
were referred to the clinic from farther away specifically for
the evaluation of dementia. Alzheimer's disease patients referred to the University of Washington Alzheimer Center
were significantly younger, more educated, more severely
impaired, had less comorbidity, and even had different gene
frequencies than Alzheimer's disease patients identified
within the HMO-based dementia registry at the same center
(19,23). Thus, studies that are based entirely on volunteers
RANDOM VS VOLUNTEER SELECTION
might find systematically higher or lower rates of dementia,
and/or of risk factors for dementia, and their results may
not be as generalizable to the community at large as those
based on representative samples. These types of bias pose
more of a concern for cross-sectional studies in which
prevalent cases, that is, subjects demented at study entry,
must be detected, than for prospective studies, in which
dementia-free subjects are recruited and incident cases of
dementia are identified over time. For results of risk factor
analyses to be biased by the inclusion of volunteers, the relationship between the outcome (e.g., mortality or incident
dementia) and the risk factor would have to be different between random and volunteer subgroups. The data presented
here show that this is not the case with respect to mortality
in the Mo VIES cohort.
Cost-Effectiveness
The cost-effectiveness of random selection vs volunteer
recruitment must be considered depending on the objectives
of the given study. For example, the scientific costs of inadequate sample sizes must be weighed against the diminishing returns of open-ended recruitment; Silagy et al. (24)
suggest that the pressure felt by investigators to obtain high
response rates may lead to the use of a variety of low-yield
recruiting measures, which can substantially increase the
cost of the study without proportionately increasing its
value. Based on experience in the Duke Longitudinal Study,
Maddox (16) concluded that for clinical studies and
prospective studies, representative samples would be ideal
but impractical. Schleser et al. (25) observed that, although
successful in meeting the requirements of individual studies, the absolute cost of incentive programs prohibits largescale recruitment and may be effective in attracting only a
certain subset of the population. Describing attempts to recruit volunteers as normal controls, Shtasel et al. (26) reported that, of 1,670 persons responding to newspaper advertisements for "healthy people ages 18-45," only 9% met
inclusion criteria and were willing to participate.
A secondary finding of our study, which should be tempered by the recognition of the small sample size on which it
is based, is that the RV subsample, those who were randomly
selected but immediately agreed to participate, was not significantly different from the larger random sample on any of
the baseline characteristics we measured. The fact that no
statistically significant differences were detected between the
RS and RV samples may be construed as reflecting largely
lack of power resulting from the relatively small size of the
RV sample. However, it is notable that the percentages of
each characteristic in Tables 1-4 are extremely close between
the RS and RV samples, suggesting that the groups are indeed similar in all these aspects. This finding has led to our
decision to pool the two random samples for future incidence
as well as risk factor analyses. A potential methodologic implication is, however, worth discussing, given the effort required to obtain "maximum" response rates in random samples, and the potential for diminishing returns as recruitment
proceeds. Our data suggest that a sample similar in characteristics to the intensively recruited random sample can be recruited by enrolling the most willing members of a random
subsample. Hypothetically, perhaps an entire cohort of the
M45
desired size can be accrued, with minimal selection bias as
well as minimal cost, by drawing a series of random subsamples, recruiting those who respond immediately, and moving
on to the next random subsample. This hypothesis would
need to be tested with a larger "random volunteer" sample
than that which we recruited. While to some extent this is de
facto what occurs in most studies, we have attempted a systematic post hoc examination of its effects. Although some
degree of bias is inevitable with such an approach, this strategy may represent a cost-effective middle road between conventional random sampling and volunteer selection.
ACKNOWLEDGMENTS
The authors gratefully acknowledge the contributions of Jill Bennett,
Rebecca Veschio, Faith Gallioto, Norma Maatta, and Diane Beley for recruitment and data collection; Mary Marcini for mortality tracking and
on-site clerical services; Deborah Echement, Wilma Furlong, and Meribeth Riccio for data management; Eric C. Seaberg for assistance with statistical analyses; and Dr. Steven Belle for overall supervision of data management and analysis. The cooperation of the Mon Valley Community
Health Center is also appreciated.
The work reported here was supported in part by grants AG00312,
AG06872, and AG07562 from the National Institute on Aging, U.S. Department of Health and Human Services.
Address correspondence to Dr. Mary Ganguli, WPIC, 3811 O'Hara
Street, Pittsburgh, PA 15213-2593. E-mail: [email protected]
REFERENCES
1. Kinsella K. Demographic imperative. Cancer Control 1995; 7-10.
2. Rosenthal R, Rosnow RL. The volunteer subject. New York: John
Wiley, 1975.
3. Ganguli M, Belle S, Ratcliff G, et al. Sensitivity and specificity for dementia of population-based criteria for cognitive impairment: the
MoVIES project. J Gerontol 1993;48:M152-61.
4. Ganguli M, Ratcliff G, Huff FJ, et al. Effects of age, gender, and education on cognitive tests in a rural elderly community sample: norms
from the Monongahela Valley Independent Elders Survey. Neuroepidemiology 1991;10:42-52.
5. Folstein MF, Folstein SE, McHugh PR. Mini-Mental State: a practical
method for grading the cognitive state of patients for the clinician. J
Psychiat Res 1975; 12:189-98.
6. Lawton MP, Brody EM. Assessment of older people: self-maintaining
and instrumental activities of daily living. Gerontologist 1969;9:
179-86.
7. Ganguli M, Seaberg E, Belle S, Fischer L, Kuller LH. Cognitive impairment and the use of health services in an elderly rural population:
the MoVIES project. J Am Geriatr Soc 1993 ;41:1065-70.
8. Cox DR. Regression models and life tables. J Roy Stat Soc 1972;34:
187-220.
9. Norton MC, Breitner JCS, Welsh KA, Wyse BW. Characteristics of
nonresponders in a community survey of the elderly. J Am Geriatr
Soc 1994;42:1252-6.
10. Hebert R, Bravo G, Korner-Bitensky N, Voyer L. Refusal and information bias associated with postal questionnaires and face-to-face interviews in very elderly subjects. J Clin Epidemiol 1996;49:373—81.
11. Koval JJ, Ecclestone NA, Paterson DH, Brown B, Cunningham DA,
Rechnitzer PA. Response rates in a survey of physical capacity among
older persons. J Gerontol 1992;47:S140-7.
12. Launer LJ, Wind AW, Deeg DJH. Nonresponse pattern and bias in a
community-based cross-sectional study of cognitive functioning among
the elderly. Am J Epidemiol 1994;139:803-12.
13. Ives DG, Traven ND, Kuller LH, Schulz R. Selection bias and nonresponse to health promotion in older adults. Epidemiology 1994;
5:456-61.
14. Poon LW, Krauss IK, Bowles NL. On subject selection in cognitive
aging research. Exp Aging Res 1984; 10:43-9.
15. Herzog AR, Rodgers WL. Age and response rates to interview sample
surveys. J Gerontol 1988;43:S200-5.
M46
GANGULIETAL
16. Maddox GL. Goals and methods. Selected methodological issues.
In: Palmore E., ed. Reports from the Duke Longitudinal Study,
1955-1969. Durham, NC: Duke University Press, 1970; 18-27.
17. Katzman R, Aronson M, Fuld P, et al. Development of dementing illnesses in an 80-year-old volunteer cohort. Ann Neurol 1989;25:317—24.
18. Perls TT, Morris JN, Ooi WL, Lipsitz LA. The relationship between
age, gender and cognitive performance in the very old: the effect of
selective survival. J Am Geriatr Soc 1993;41:1193—201.
19. Tsuang D, Kukull W, Sheppard L, et al. Impact of sample selection on
APOE e4 allele frequency: a comparison of two Alzheimer's disease
samples. J Am Geriatr Soc 1996;44:704-7.
20. Williams DE, Vitiello MV, Ries RK, Bokan J, Prinz PN. Successful
recruitment of elderly community-dwelling subjects for Alzheimer
disease research. J Gerontol 1988;43:M69-74.
21. Norris FH. Characteristics of older nonrespondents over five waves of
a panel study. J Gerontol 1985;40:627-36.
22. Kokmen E, Ozsarfati Y, Beard CM, O'Brien PC, Rocca WA. Impact of
referral bias on clinical and epidemiological studies of Alzheimer's
disease. J Clin Epidemiol 1996;49:79-83.
23. Barnhart RL, Belle G, Edland SD, et al. Geographically overlapping
Alzheimer's disease registries: comparisons and implications. J Geriatr Psychiat Neurol 1995;8:203-8.
24. Silagy CA, Campion K, McNeil JJ, Worsam B, Donnan GA, Tonkin
AM. Comparison of recruitment strategies for a large-scale clinical
trial in the elderly. J Clin Epidemiol 1991 ;44:1105-14.
25. Schleser R, West RL, Boatwright LK. A comparison of recruiting
strategies for increasing older adults' initial entry and compliance in a
memory training program. Int J Aging Hum Dev 1986-87,24:55-65.
26. Shtasel DL, Gur RE, Mozley D, et al. Volunteers for biomedical research. Recruitment and screening of normal controls. Arch Gen Psychiat 1991;48:1022-5.
Received January 30, 1997
Accepted May 23, 1997
DIRECTOR, KUNIN-LUNENFELD
CLINICAL RESEARCH UNIT
(Associate to Full Professor Level)
Reporting to the Vice President, Research, the Director
will be responsible for the scientific leadership and management of a newly-developed and expanding unit which
focuses on clinical research designed to enhance the quality of care and life of the elderly.
Research in the Unit functions at the interface between
basic and applied science and clinical epidemiology, and involves both the evaluation of the validity or current treatment programs and the development of new management
and treatment procedures. The Unit includes fully-funded
scientists, postdoctoral fellows, and technical support staff.
This unit also includes the Katz Centre for Gerontological
Social Work, a joint Centre with the University of Toronto
and the Faculty of Social Work.
Baycrest Centre for Geriatric Care is a leader in the care
of the elderly, and is fully affiliated with the University of
Toronto as a teaching and research organization. Its unique
campus includes a 300-bed geriatric hospital, 372-bed
home for the aged, a 204-unit seniors' supported apartment residence, a 14-unit group home, and a range of ambulatory and community outreach programs. The Rotman
Research Institute is also an integral part of the Research
Division at the Centre.
The Director will be expected to provide a vision for the
future development of the Unit, scientific leadership, promotion of training for postdoctoral fellows and graduate
students, and a commitment to excellence.
The successful candidate will be a PhD and MD with
proven leadership abilities and a strong background in clinical research with a specific focus on the elderly, including a
record of grant awards and publications. Research
strengths could include clinical trials, theoretically driven
clinical research, evaluation, and/or clinical epidemiology.
An Assistant Director will provide support and compliment
the Director's skills. The Director will be eligible for a status-only appointment at the University of Toronto at the Associate or full Professor level. The salary offered will be
commensurate with that offered by the University.
Baycrest and the University of Toronto encourage applications from qualified women and men, members of visible
minorities, aboriginal peoples, and persons with disabilities. In accordance with Canadian immigration requirements, this advertisement is directed firstly to Canadian citizens and permanent residents.
Applicants should submit a C.V., together with the
names of three references, by March 31,1998, to:
Baycrest Centre
for Geriatric Care
Vice-President, Human Resources
Baycrest Centre for Geriatric Care
3560 Bathurst Street
Toronto, Ontario M6A 2E1
Canada
Fax:(416)785-2490