Methodological Considerations when Studying Aging 495cardiovascular risk factors present during neuropsychological
testing. For an overall composite score involving all of the
neuropsychological tests, there was a 23% (odds ratio1.23)
increase in risk for poor cognitive functioning (performance at
or below the 25 percentile)per risk factorbeyond zero risk
factors. In a secondary analysis, a long-term risk factor scale
was derived from cardiovascular risk data collected over 18 to
20 years. For this scale, there was an estimated 31% increase
(odds ratio1.31) in risk for poor cognitive functioning per
risk factor. The largest increase in risk (39%) was associated
with learning and immediate memory.
Unfortunately, no studies have asked the question as to
whether cumulative risk, as determined by multiple risk factor
scale, is exacerbated or diminished with increasing age. More-
over, studies adjusting the impact of age on cognitive func-
tioning for risk factors have not used a multiple risk factor
scale re”ecting the long-term and cumulative impact of risk.
Such studies are needed. While the examples have been given
with data from cardiovascular disease and cognition, the
principles apply generally in aging and health psychology.
Epidemiologic Concerns
The need for more sophisticated models that take into ac-
count relations between risk factors, aging, and disease is
apparent. But testing these models constitutes a major chal-
lenge. Design problems created by the dynamic character of
both disease and aging are evident. An excellent and well-
written summary of these designs and issues may be found in
Hennekens et al. (1987) and Collins and Horn (1991).
Understanding these issues is particularly important to
two critical decisions in the design of health-aging studies:
(a) selection of exclusionary variables; and (b) identi“cation
of confounders, necessary to model speci“cation. These
problems are common to all research areas but acutely
important in health-aging research for two reasons: (a) the
coexistence of chronic diseases increases with advancing age
and diseases interact in complex ways, and (b) the duration
of exposure to risk factor and disease is correlated with age
(Kaplan et al., 1999). These correlations are particularly
problematic in covariance analyses involving the adjustment
of risk factors for the impact of comorbidities and in designs
in which disease effects are adjusted for age, or vice versa.
Covariance assumptions are often not met in circumstances
in which they are most needed. Pedhauser and Schmelkin•s
(1991) discussion of covariance issues and solutions is most
valuable. Hennekens et al.•s (1987) chapter on confounding
and bias in public health research is very useful. Sackett et al.
(1991) offers an excellent discussion of issues surrounding
subject selection.
Kaplan et al. (1999) provide a good review of the particu-
lar odd things about aging when considering epidemiologic
research. The review is very valuable with respect to design
decisions such as whom to exclude, what confounders are
important conceptually, how to handle comorbidity (by
exclusion or by statistical adjustment), and issues of subject
selection, sample bias, and survival. This discussion begins
with complex and mixed “ndings in studies where cardiovas-
cular disease risk factors are related to cardiovascular disease
outcomes. It is widely assumed that the association between
CVD risk factors and CVD events and outcomes grows
weaker with advancing age, but the literature does not support
this conclusion (Kaplan et al., 1999). A pattern of declining
associations between CVD risk factors and cardiovascular
disease has been reported in some large population studies
(Psaty et al., 1990; Whisnant, Wiebers, O•Fallon, Sicks, &
Frye, 1996). Increasing strength of associations (Benefante,
Reed, & Frank, 1992; Keil, Sutherland, Knapp, & Gazes,
1992) and mixed results have been observed in others.
Kaplan et al. (1999) point out very complex and dynamic
associations among the following variables: (a) CVD risk
factors, (b) comorbid conditions, (c) subject selection and
attrition, (d) mortality, (e) subclinical disease, (f) clinical dis-
ease detection, (g) treatment, (h) clinical events, (i) metabolic
and physiologic changes, and (j) •aging senescence.ŽThese 10
variables provide a practical checklist for data design and
analysis. They all change with age and interact with each other.
Each of these changes (or modi“cations) correlate with age,
aging, and the passage of time. They affect, and are affected by,
manifestation of diseases, accuracy of self-report and recall of
exposures to risk, selection bias, accuracy of measurement,
changing diagnostic and assay methods, attrition and differen-
tial rates of attrition from longitudinal studies, selective sur-
vival, and the validity of covariance analyses.
A particular problem is the need to use proxy variables
and data from informants other than the study participant.
Efforts to avoid this problem often lead to exclusion of some
subject populations (e.g., institutionalized individuals) and
result in their underrepresentation in study samples.
The problem of dropout in longitudinal designs has been
discussed in our review of the hypertension-cognitive func-
tioning literature. Especially problematic for longitudinal
studies of cognitive functioning is that persons who perform
more poorly at one time are less likely to return for repeat test-
ing than those who perform well (M. Elias & Robbins, 1991b).Longitudinal Analysis MethodsThe recognition of problems, such as selective attrition and
the need to estimate missing data, has moved sophisticated