This Article Abstract Full Text (PDF) Services Related articles in Journals of Gerontology Series A: Biological Sciences and Medical Sciences Download to citation manager PubMed PubMed Citation

How Are Biomarkers Related to Physical and Mental Well-Being?

¹ Office of Population Research, Princeton University, New Jersey.
² Center for Population and Health, Georgetown University, Washington, District of Columbia.
³ Bureau of Health Promotion, Center for Population and Survey Research, Department of Health, Taichung, Taiwan.

Address correspondence to Christopher L. Seplaki, PhD, Office of Population Research, 263 Wallace Hall, Princeton University, Princeton, NJ 08544. E-mail: cseplaki{at}princeton.edu

	Abstract

Top Abstract Methods Results Discussion References

 
We investigate how biological markers of individual responses to stressful experiences are associated with profiles of physical and mental functioning in a national sample of middle-aged and elderly Taiwanese. Data come from a population-based sample of middle-aged and elderly Taiwanese in 2000. The data combine rich biological measures with self-reported information on physical and mental health. Grade of membership methods are used to summarize functional status, and multinomial logit models provide information on the association between biological measures and function. The analysis identifies significant associations between biomarkers of stressful experience and profiles of physical and mental functioning. The estimates reveal the potential importance for health of both low and high values of biological parameters. The findings point to directions for future research regarding development of aggregate measures of cumulative dysregulation across multiple physiological systems.

The experience of stressful stimuli precipitates a complex and counterbalancing set of hormonal responses in the sympathetic nervous system (SNS) and hypothalamic-pituitary-adrenal (HPA) axis. Chronic stimulation of these responses may cause imbalances in their functioning, as well as dysregulation in cardiovascular functioning [see McEwen (5), Brunner and Marmot (6), and McEwen and Goodman (7)]. Reflecting the "fight-or-flight" aspect of the stress response, epinephrine and norepinephrine are released in the SNS to stimulate the cardiovascular system and blood flow. The HPA axis response to stress is characterized by cortisol secretion, which has numerous physiological effects, including glucose metabolism and immune response. Dehydroepiandrosterone sulfate (DHEA-S) is believed to be an antagonist to cortisol, although its precise mechanisms are not well understood (8). Also among the hormonal responses to stress are the growth hormone insulin-like growth factor-1 (IGF-1) and immune system factors such as interleukin-6 (IL-6). IGF-1 is important for muscle growth (9), while IL-6 is involved with the inflammation response (10) and has been shown to be elevated in those experiencing chronic stress (11). Cardiovascular parameters that are affected by the stress response include cardiovascular disease risk factors and aspects of insulin resistance syndrome, or "syndrome X" (12–14), such as blood pressure, cholesterol level and composition, blood sugar levels, and visceral fat deposition.

Persistently high or low levels (i.e., levels outside of normal operating ranges) of many of these factors may be associated with various poor health outcomes. For example, chronically elevated levels of cortisol may result in hypertension, abdominal obesity, and memory impairment, while low levels of cortisol may be associated with excessive immune responses (15). Other examples include hypotension and hypertension, hypoglycemia and hyperglycemia, and hypocholesterolemia and hypercholesterolemia and their sequelae. Alternatively, for some factors, only one tail appears to be related to risk; e.g., only low levels of DHEA-S appear to be associated with worse mental and physical health (16–18), while only a high ratio of total to high-density lipoprotein (HDL) cholesterol level is typically associated with cardiovascular disease risk.

Theories of allostatic load (5,19,20) describe how the experience of stressful events precipitates a complex set of hormonal and physiological responses across the neuroendocrine, sympathetic nervous, immune, and cardiovascular systems, which, when repeatedly or chronically stimulated, can result in dysregulation of these responses and poor health outcomes. Expositions by McEwen and Seeman (21) and McEwen (22) define the hormonal elements of this cascade (i.e., HPA axis, SNS, and immune system parameters) as "primary mediators," which, in turn, become manifest in a collection of intermediary cardiovascular disease risk factors defined as secondary outcomes (which include the symptoms of insulin resistance syndrome). The ultimate disease end-points of this cascade are termed "tertiary outcomes."

The theory and physiological processes that define the health effects of allostatic load are, by definition, the aggregate consequences of mechanisms spanning several physiological systems. Thus, empirical support for the theory does not derive from the existence of significant associations between outcomes and individual biomarkers, but instead between outcomes and measures of the combined influence of multiple physiological parameters (23). Accordingly, composite indices of allostatic load have been developed and implemented in population-based studies (e.g., 23,24). The original and most frequently-used index of allostatic load (23) consists of a summation of the number of times an individual falls into a (single tail) "highest risk" quartile for each of 10 biomarkers affiliated with the stress response, i.e., beyond a high or, in some cases, low threshold defining values outside of "normal" operating ranges. Alternatively, a recent reformulation of the index (24) uses a weighted linear combination of the same biomarkers measured on a continuous scale, where the weights are optimized to predict downstream health outcomes.

Current research that examines the associations between health end-points and either individual biomarkers of the stress response or related aggregate measures of physiological risk suffers from several limitations. First, the individual associations between many of these biomarkers and health end-points are not well documented in representative, population-based samples. Most existing work using such biological information examines relatively small or selected samples (e.g., clinical or healthy samples) that are not representative of a general population. Second, most existing studies consider associations with end-points that define a single condition (e.g., mortality, physical or cognitive performance) rather than associations with more realistic states of comorbid functional impairment. Third, only linear relationships are typically investigated, leaving unknown the potential association of both high and low biomarker values with risk. Fourth, studies have paid little attention to identifying what cut-points should be used to define the extreme values, i.e., what are outside "normal" ranges. Lastly, the relative importance of the various biomarkers forming aggregate indexes is generally unexplored, and current measures assume either that all components count equally or weight them using theoretically endogenous tertiary outcome information.

In this analysis, we investigate how biological indicators of cumulative physiological dysregulation associated with the stress response are individually associated with profiles of physical and mental functioning in a national sample of middle-aged and elderly Taiwanese. These profiles are derived from a statistical procedure known as Grade of Membership (GOM), which synthesizes the available information on physical limitations, depressive symptoms, and cognitive functioning. Our specific aims are to: 1) use GOM models to summarize the burden of impairment in this population, and 2) use multinomial logit regression models to investigate the associations between the resulting profiles of functioning and a broad range of biomarkers. We analyze data representing a number of improvements over existing sources and address several methodological limitations of prior analyses. This research is likely to inform the development of improved measures of cumulative physiological dysregulation.

	METHODS

Top Abstract Methods Results Discussion References

 
Data
This analysis uses data on biological markers and self-reports of physical and mental functioning from the 2000 Social Environment and Biomarkers of Aging Study (SEBAS). Respondents for SEBAS were drawn from the Taiwan Survey of Health and Living Status, a longitudinal study of a nationally representative probability sample that began in 1989 with persons aged 60 years and older, including the institutionalized population. A new sample of middle-aged persons (aged 50 to 66 years) was included in 1996. Respondents for SEBAS were drawn from a random subsample of 1713 survivors selected from the 1999 combined middle-aged and older cohorts. Those aged 71 years and older in 2000 and persons in urban areas were over-sampled. The survey procedures were approved by the institutional review boards at Princeton University, Georgetown University, and the Bureau of Health Promotion, Department of Health, Taiwan, and conformed to the principles embodied in the Declaration of Helsinki.

SEBAS consists of two parts: an in-home interview (n = 1497, a 92% response rate among survivors) and a medical exam (n = 1023, 68% of those interviewed). The in-home interview collected information on self-reports of physical limitations, depressive symptoms, social and demographic variables, and cognitive performance measures. Each participant in the medical exam was asked to fast overnight and collect a 12-hour overnight urine sample (to provide integrated measures of neuroendocrine function). The following morning, a medical exam was conducted in a nearby hospital, where a physician or nurse drew a blood sample from the participant and took blood pressure and anthropometric measurements. Compliance by those completing the medical exam was extremely high: All but 10 individuals followed the urine protocol, provided a sufficient volume of blood suitable for analysis, and completed the medical exam.

Among the original 1497 respondents to the initial interview, 111 (7%) were ineligible for the exam because of a health condition. Among the remaining 363 (24%) who declined to participate in the exam, the principal reasons included: 1) the respondent felt that he or she was healthy and did not need an exam; 2) the exam was too much trouble; 3) the respondent just had a health exam; and 4) the respondent had no free time or was out of town during the several-day period during which exams were offered. A comparison of the characteristics of nonparticipants and participants in the medical exam reveals that, although persons older than age 70 were less likely to participate than younger respondents, sex and various measures of socioeconomic status were not significantly related to participation. Persons who received the medical exam reported the same average self-assessed health status (5-point scale) as those who did not. These results suggest that, in the presence of controls for age, estimates derived from clinical information are unlikely to be seriously biased (25).

There were 43 individuals for whom profiles of physical and mental functioning could not be calculated because of missing data; they are excluded from this analysis. Unweighted comparisons between the included and excluded groups reveal that, although the excluded group is older (p =.02) and is comprised of more women relative to the analysis sample (p =.006), IL-6 is the only biomarker with a significant difference (Mann-Whitney, p =.041) between those included and excluded from the analysis. A few individuals for whom functional profile scores were calculated have missing values for some of the biomarkers, resulting in final sample sizes between 976 and 980, depending on the specific biomarker (as noted in Tables 3 and 4). The mean age of the sample (n = 980) is 68.8, 44% are over the age of 70, and 41% of the sample are women.

View this table: [in this window] [in a new window]   Table 3. Multinomial Logit Estimates for Regression of Each Biomarker on GOM Profile Scores, Primary Mediators (10% and 90% Cut-Points).

View this table: [in this window] [in a new window]   Table 4. Multinomial Logit Estimates for Regression of Each Biomarker on GOM Profile Scores, Secondary Outcomes (10% and 90% Cut-Points).

Measures
Measures of cortisol, epinephrine, norepinephrine, and dopamine (a precursor to the synthesis of norepinephrine) were derived from the 12-hour urine specimens. Results for cortisol, epinephrine, and norepinephrine are reported as "micrograms per gram creatinine" in order to adjust for body size (see Table 1). The fasting blood specimens provided measures of DHEA-S, IGF-1, IL-6, total cholesterol, the ratio of total to HDL cholesterol, triglycerides, fasting glucose, and glycosylated hemoglobin. Systolic and diastolic blood pressures were calculated as the average of two seated blood pressure readings (taken about 1 minute apart, using a mercury sphygmomanometer). Anthropometric information used for the BMI (weight in kilograms divided by height in meters squared) and for the waist/hip ratio calculation was gathered during the hospital exam.

In these analyses, biomarker values are divided into three discrete categories indicating those in the lowest 10% of nonmissing values of the full sample of 1023 respondents, the middle 10%–90%, and above 90%. Table 1 gives the cutoff values defining each category for each biomarker. For IL-6 and epinephrine, a large number (approximately 33% and 20%, respectively) of readings fell below the sensitivity of the assays: 0.1 pg/mL for IL-6 and 2 µg/L for epinephrine. Thus, the categories for these measures are as follows: below assay sensitivity, above assay sensitivity and below the 90th percentile, and above the 90th percentile. We explored the robustness of the results to the percentile cut-off levels by dividing observations for each biomarker into a second set of three categories based on alternative criteria: those below the 25th percentile (which included observations below assay sensitivity for IL-6 and epinephrine), those between the 25th and 75th percentiles, and those above the 75th percentile. This second characterization based on quartiles was designed to parallel existing formulations of allostatic load in the literature (23), although, unlike current measures of allostatic load, it permitted us to examine associations at both extremes.

Measures of functioning include 6 activities of daily living (ADLs), 7 mobility impairments, and items from indices of cognitive performance and depressive symptoms. The ADL items were composed of difficulty with bathing, dressing or undressing, eating, getting out of bed, moving around the house, and using the toilet. Mobility impairments include difficulty squatting, walking up 2–3 flights of stairs, lifting or carrying 11–12 kg, working around the house, walking 200–300 meters, standing continuously for 15 minutes, and running 20–30 meters.

Depressive symptoms are measured using questions from a 10-item version of the original 20-item Center for Epidemiologic Studies Depression Scale (CES-D) (28) that has been used in Chinese populations (29,30). Shortened forms of the CES-D, including a 10-item Chinese version in a sample of elderly Chinese (31), have been shown to perform well (32,33). CES-D items used here include reports (in the past week) of no interest in eating, sleeping poorly, being in a terrible mood, feeling lonely, people not being nice, feeling anguished, having no energy to do things, feeling joyful (reverse-coded), that doing anything is exhausting, and life is going well (reverse-coded). Measures of cognitive function include 12 items from 3 tests: the modified Short Portable Mental Status Questionnaire (34), the modified Rey Auditory Verbal Learning test (35), and a modification of the Digits Backward test (36) (see Table A3 in Appendix for a list of items).

Following standard notation (4,37), for any pure-type k (k = 1...K) _kjl denotes the probability that an individual whose functional status exactly matches that of pure-type k gives response l to variable j. Also, for each individual i, a K-dimensional vector g_i = (g_i1,...,g_iK) is defined so that and 0 g_ik 1. The individual GOM score elements g_ik represent the similarity of individual i to pure-type k. The GOM model assumes that, conditional on g_i, the probability p_ijl that person i gives response l to variable j is equal to . The _kjl and g_i are estimated iteratively by maximizing a conditional likelihood function based on p_ijl, beginning from a set of initial values for the g_i and the _kjl [for details see Manton and colleagues (37), and additional treatments by Berkman and colleagues (4) or Wachter (40)]. All GOM analyses in this paper are performed using software developed by Charpentier (41).

The number of pure-types K is set in advance of the estimation. In exploratory analyses, models can be estimated by choosing successively higher values of K. However, while an increase in K tends to improve the fit of the model, the pure types and mixes among them generally become more difficult to interpret. Thus, following work by Berkman and colleagues (4) and Singer (42), we chose the number of pure types on the basis of a priori hypotheses and model parsimony. In particular, comparable prior studies have found that between 4 and 6 pure types provide reasonable summaries of health and functional status in elderly populations (2–4). Our analyses showed that, with the SEBAS data, 5 pure types provided a balance between pure-type detail and interpretability.

Subsets of characteristics that distinguish one pure type from another, and thus form the basis for verbal descriptions of the pure types (below), are defined using criteria developed by Singer (42) and Berkman and colleagues (4), which compare each _kjl to the corresponding marginal frequency. Specifically, we define a particular response as a distinguishing characteristic for a pure-type profile if its estimated response probability (shown in Appendix Tables A1, A2, and A3) is at least twice the marginal frequency of that response in the overall sample. For relatively prevalent characteristics (assumed here to be responses with marginal probabilities greater than 0.4), the response is considered to be distinguishing if the estimated pure-type probability is at least 35% greater than the marginal frequency.

GOM models can be sensitive to initial parameter values, particularly those for the _kjl, because of the difficulty of identifying a global maximum on the complex likelihood surface. We address this issue by choosing initial values for the _kjl that reflect highly idealized and simplified combinations of functional impairment that were informed both by exploratory analyses (using random values) and existing studies. This strategy of using "informed" [(37): p. 72] initial values increases confidence in the optimized solution, relative to one derived from purely random or other numerical criteria. Initial values for the individual GOM scores were set to be uniform across all individuals and equal to one divided by the number of pure types.

The second stage of our analysis uses multinomial logit models to examine the individual associations between the GOM scores of functioning and each biomarker. Each model includes four GOM scores, along with controls for age and sex. The middle category (e.g., 10%–90%) serves as the reference group for the specific biomarker, and the GOM score for profile I (described below as a fully functioning pure type) is the reference GOM score. Thus, each model predicts the probability that an individual has a biomarker score in each distributional category (e.g., below the 10th percentile, 10%–90%, or above the 90th percentile).

Although the specification of the biomarkers as outcomes and the functioning scores as explanatory variables may appear to be a reverse implementation of our conceptual model, the two are functionally equivalent in the sense that both sets of variables—functioning and biomarkers—are measured at the same time (i.e., the survey period in 2000). We prefer the current specification because modeling the GOM scores as the outcome would require a separate equation for each profile score and for each biomarker, multiplying the already considerable number of equations under consideration by a factor of K.

Following the estimation of each model, a ² test for equality of the distributions of the GOM profile scores in that model is calculated. For biomarkers that demonstrate significant (p <.10) joint associations with the GOM profile scores, pair-wise tests for the equality of GOM profile score distributions are subsequently performed. For each pair of profiles A and B, the null hypothesis for these tests is that profile scores A and B have the same distribution of the given biomarker. Nesting the pair-wise tests within a joint test of significance in this way provides control over our overall error rate. The use of a Bonferroni multiple-test procedure (43) was considered but rejected because the approximation is often too conservative (44).

To illustrate the magnitude of the relationships in cases of significant joint associations, predicted probabilities of having a biomarker value in each outcome category are calculated for each of the five pure-type function profiles. For each biomarker, these simulated probabilities are obtained by: 1) sequentially assigning all individuals to the first GOM profile, followed by the second profile, and so on, while leaving age and sex at their observed values, and 2) in each case, averaging the resulting predicted probabilities across individuals. All summary analyses, multinomial logit estimation, tests, and predictions were done using Stata 7.0 (Stata Corp., College Station, TX) (45).

	RESULTS

Top Abstract Methods Results Discussion References

 
Grade of Membership Classification of Functioning
Estimation of the five pure-type GOM model yields a set of profiles with the following distinguishing characteristics (the full set of estimated pure-type response probabilities is given in Appendix Tables A1, A2, and A3, with those satisfying the criteria for distinguishing characteristics in boldface). The names given below in italics for each of the profiles are oversimplifications that are intended as shorthand descriptions for ease of reference:

• Pure-Type I, Fully Functioning: Pure-type I has no physical or mental impairments.
  
• Pure-Type II, Cognitively Impaired: Pure-type II has no physical or depressive impairments (apart from some difficulty running 20–30 meters). However, this pure type has significant cognitive deficits, indicated by the substantial number of incorrect cognitive evaluation responses that distinguish it from other pure types.
  
• Pure-Type III, Moderately Depressed: Pure-type III has no physical impairment but displays moderate depressive symptoms (and has some difficulty with the 10-item recall).
  
• Pure-Type IV, Limited Mobility: Pure-type IV has mobility impairments, but no ADL or cognitive impairments. This type is also distinguished by the presence of sleeping problems and a lack of interest in eating, although each of these is reported as rare.
  
• Pure-Type V, Substantially Impaired: Pure-type V displays severe mobility and ADL impairments, severe depressive symptoms, and some cognitive impairment.
  

The distribution of functional impairment in the sample across these five pure-type profiles is indicated by the distribution of GOM scores estimated for each individual from the GOM model (Table 2). Consistent with Berkman and colleagues (4), we define individuals with a GOM score of 0.9 or greater as representing a single pure type. Given this definition, 26.7% of males and 14% of females are described solely by the fully functioning profile, indicating a substantial difference by sex in the proportion that can be described as having no functional impairments. Very few individuals, male or female, are described solely by any one of the remaining profiles.

Multinomial Logit Model Estimates of Associations Between Biological Measurements and Functioning
The results for each multinomial logit model are given in Tables 3 (primary mediators) and 4 (secondary outcomes), based on the 10% and 90% cut-points. A p value less than 0.1 (suggesting rejection of the null hypothesis that all five of the GOM profile scores have the same distribution of the given biomarker) was obtained for the following primary mediators—dopamine (p .104), cortisol, DHEA-S, and IL-6—and the following secondary outcomes—ratio of total to HDL cholesterol, triglycerides, fasting glucose, BMI, and waist/hip ratio.

Conditional on p < 0.1, we next examined pair-wise tests for the equivalence of a given biomarker distribution between each pair of profiles (shown in Table 5 as a box with four columns; pair-wise test results not shown in any particular order). The appearance of two xs in any one (or more) of the columns represents a significant difference between the two GOM profiles indicated by the corresponding x values. For example, Table 4 demonstrates a significant joint association (p .09) between the five profiles and the ratio of total to HDL cholesterol. Table 5 presents the subsequent tests for the equivalence of the distribution of total/HDL cholesterol between each pair of profiles. Results show that significant differences in the ratio of total to HDL cholesterol were found between individuals characterized by the substantially impaired profile and those characterized by each of the following profiles: fully functioning, moderately depressed, and limited mobility. The predicted probabilities that accompany the pair-wise test results in Table 5 indicate the nature of these differences. For example, individuals denoted by the substantially impaired profile are more likely than those in the three other profiles to have high values (e.g., 28% vs 10%, 6%, and 5%, respectively) of total/HDL cholesterol. Table A4 in the Appendix provides the full set of ² statistics and p values for all pair-wise comparisons.

View this table: [in this window] [in a new window]   Table A4. Joint and Pair-Wise 2 Test Results (10th and 90th Percentile Cut-Points).

The secondary outcomes demonstrate similar associations with the profiles of functioning. Individuals denoted by the substantially impaired profile are significantly more likely than those represented by the fully functional profile to be in the highest category of the distribution for every secondary outcome, except the waist/hip ratio. Results also demonstrate that persons described by the substantially impaired profile are distinct from: 1) those characterized by the moderately depressed profile with regard to the distributions of the total/HDL cholesterol ratio, fasting glucose, and BMI; and 2) those represented by the cognitively impaired profile with regard to triglyceride levels and fasting glucose. In each of these cases, individuals with scores indicating increasing similarity to the substantially impaired profile are more likely to have values in the highest category of the relevant biomarker.

Sensitivity results (Tables A5, A6, and A7 in the Appendix) based on alternative cut-point values at the 25th and 75th percentiles for each of the biomarkers display similar joint associations as under the original 10th and 90th percentile cut-points for several of the primary mediators (dopamine, DHEA-S, and IL-6) and secondary outcomes (total cholesterol/HDL ratio, fasting glucose, BMI, and waist/hip ratio) (Table A7). However, three additional biomarkers now reveal joint (and pair-wise) significant associations with functioning (diastolic blood pressure, total cholesterol, and IGF-1), while two that were significant under the 10th/90th percentile criteria (triglycerides and cortisol) fail to demonstrate significant joint associations with functioning under the 25th/75th percentile criteria.

View this table: [in this window] [in a new window]   Table A5. Multinomial Logit Estimates, Primary Mediators (25th and 75th Percentile Cut-Points).

View this table: [in this window] [in a new window]   Table A6. Multinomial Logit Estimates, Secondary Outcomes (25th and 75th Percentile Cut-Points).

View this table: [in this window] [in a new window]   Table A7. Joint and Pair-Wise 2 Test Results (25th and 75th Percentile Cut-Points).

	DISCUSSION

Top Abstract Methods Results Discussion References

 
This analysis provides a summary of the functional status of a national sample of older Taiwanese, which captures the complex mixture of cognitive, psychological, and physical impairments that coexist at older ages. The resulting profiles, in turn, permit us to identify biomarkers that are associated with functional status in this population as well as to examine the nature of the ensuing relationships.

There are several limitations of this study. First, these analyses are cross-sectional. We do not attempt to estimate or infer the direction of the association between the biomarkers and functioning, and recognize that effects in both directions are likely to operate. For example, whereas various indicators of syndrome X are known to affect subsequent physical health, other associations are less clear cut. Specifically, although it is tempting to suggest, as much of the scientific and popular literature does, that low levels of DHEA-S lead to poor health and a short life span, it is also plausible that certain illnesses reduce the levels of DHEA-S (16,46). Second, the distributions of biomarkers in this analysis are divided intro three categories designed to distinguish "high" and "low" values from those within "normal" ranges. The choice of these cut-points is arbitrary, but the empirical approach we have taken in considering two sets of cut-point criteria is reasonable in the absence of clinically defined, or otherwise substantively meaningful, criteria. Although population distributions and risk cut-points are not known for many of the biomarkers considered here, clinical criteria for high cut-point values do exist for several markers (e.g., blood pressure, total cholesterol and total/HDL cholesterol ratio, triglycerides, fasting glucose, glycosylated hemoglobin, and BMI). Results not shown here that incorporate clinical criteria for high values are generally consistent with the findings presented in this article.

Despite these limitations, our results offer several insights into the nature of functional impairment among older Taiwanese. Specifically, our analysis shows that, although a large proportion of the population was relatively unimpaired functionally in 2000, a substantial proportion of physically unimpaired persons, particularly men, displayed modest levels of depressive symptoms. Overall, women comprised a more heterogeneous mix of physical, psychological, and cognitive impairments than men.

Our findings also reveal large and statistically significant associations between profiles of functioning and physiological parameters of the stress response. In particular, the analysis uncovered numerous distinctions between the substantially impaired and fully functional profiles in their associations with primary mediators and secondary outcomes. These patterns suggest that the association between multiple-system dysregulation and health is observable across several points in the theoretical cascade of physiological events represented by primary mediators, secondary, and tertiary outcomes. An important implication of our results for future measurement strategies derives from the significance of both high and low values for several of the biomarkers and the lack of robustness of some relationships to the choice of cut-points. Thus, current formulations of allostatic load, which account for only high (or in a few cases, only low) values may fail to capture some underlying physiological relationships. Our findings also highlight the important role that biological information drawn from population-based surveys can play in future research on the connections between the social environment and health.

	Acknowledgments

 
This research has been supported by the Demography and Epidemiology Unit of the Behavioral and Social Research Program of the National Institute of Aging, grant numbers R01AG16790 and R01AG16661, and by the National Institute of Child Health and Human Development, grant number 5P30HD32030.

We thank Bruce McEwen for his guidance through the literature on the physiology of stress, Germán Rodríguez for statistical advice, and Dana Glei for helpful comments on earlier versions of the manuscript. We acknowledge helpful comments from participants in the 2003 Population Association of America meeting in Minneapolis, Minnesota, in which an early draft of this article was presented, the session discussant, Dr. David Lindstrom, and two anonymous referees.

	Footnotes

 
Decision Editor: James R. Smith, PhD

Received July 18, 2003

Accepted December 4, 2003

	References

Top Abstract Methods Results Discussion References

 

1. Weinstein M, Willis R. Stretching social surveys to include bioindicators: possibilities for the health and retirement study, experience from the Taiwan Study of the Elderly. In: Finch C, Vaupel J, Kinsella K, eds. Cells and Surveys: Should Biological Measures be Included in Social Science Research? Washington, DC: National Research Council, National Academy Press; 2000:250–275.
2. Manton KG, Myers GC, Andrews GR. Morbidity and disability patterns in four developing nations: their implications for social and economic integration of the elderly. J Cross-Cultural Gerontol.. 1987;2:115-129.
3. Lamb VL. A cross-national study of quality of life factors associated with patterns of elderly disablement. Soc Sci Med.. 1996;42:363-377.
4. Berkman LF, Singer BH, Manton K. Black/white differences in health status and mortality among the elderly. Demography.. 1989;26:661-678.[Medline]
5. McEwen BS, Stellar E. Stress and the individual. Mechanisms leading to disease. Arch Intern Med.. 1993;153:2093-2101.[Abstract/Free Full Text]
6. Brunner E, Marmot M. Social Organization, Stress, and Health. In: Marmot M, Wilkinson RG, eds. Social Determinants of Health. New York: Oxford University Press; 1999:17–43.
7. McEwen BS, Goodman HM, eds. Handbook of Physiology Section 7: The Endocrine System. Volume IV: Coping with the Environment: Neural and Endocrine Mechanisms. New York: Oxford University Press; 2001.
8. Kalimi M, Shafagoj Y, Loria R, Padgett D, Regelson W. Anti-glucocorticoid effects of dehydroepiandrosterone (DHEA). Mol Cell Biochem.. 1994;131:99-104.[Medline]
9. Borst SE, Lowenthal DT. Role of IGF-I in muscular atrophy of aging. Endocrine.. 1997;7:61-63.[Medline]
10. Arkins S, Johnson RW, Minshall C, Dantzer R, Kelley KW. Immunophysiology: the interaction of hormones, lymphohemopoietic cytokines and the neuroimmune axis. In: McEwen BS, Goodman HM, eds. Handbook of Physiology Section 7: The Endocrine System, Volume IV: Coping with the Environment: Neural and Endocrine Mechanisms. New York: Oxford University Press; 2001:469–495.
11. Kiecolt-Glaser JK, Preacher KJ, MacCallum RC, Atkinson C, Malarkey WB, Glaser R. Chronic stress and age-related increases in the proinflammatory cytokine IL-6. Proc Natl Acad Sci U S A.. 2003;100:9090-9095.[Abstract/Free Full Text]
12. Meigs JB. Epidemiology of the insulin resistance syndrome. Curr Diab Rep.. 2003;3:73-79.[Medline]
13. Meigs JB. Invited commentary: insulin resistance syndrome? Syndrome X? Multiple metabolic syndrome? A syndrome at all? Factor analysis reveals patterns in the fabric of correlated metabolic risk factors. Am J Epidemiol.. 2000;152:908-911 discussion 912.[Abstract/Free Full Text]
14. Reaven GM. Banting lecture 1988. Role of insulin resistance in human disease. Diabetes.. 1988;37:1595-1607.[Abstract]
15. Sapolsky RM, Romero LM, Munck AU. How do glucocorticoids influence stress responses? Integrating permissive, suppressive, stimulatory, and preparative actions. Endocr Rev.. 2000;21:55-89.[Abstract/Free Full Text]
16. Berr C, Lafont S, Debuire B, Dartigues JF, Baulieu EE. Relationships of dehydroepiandrosterone sulfate in the elderly with functional, psychological, and mental status, and short-term mortality: a French community-based study. Proc Natl Acad Sci U S A.. 1996;93:13410-13415.[Abstract/Free Full Text]
17. Mazat L, Lafont S, Berr C, et al. Prospective measurements of dehydroepiandrosterone sulfate in a cohort of elderly subjects: relationship to gender, subjective health, smoking habits, and 10-year mortality. Proc Natl Acad Sci U S A.. 2001;98:8145-8150.[Abstract/Free Full Text]
18. Svec F, Lopez A. Antiglucocorticoid actions of dehydroepiandrosterone and low concentrations in Alzheimer's disease. Lancet.. 1989;2:1335-1336.
19. McEwen BS. Stress, adaptation, and disease. Allostasis and allostatic load. Ann N Y Acad Sci.. 1998;840:33-44.[Medline]
20. McEwen BS. Protective and damaging effects of stress mediators. N Engl J Med.. 1998;338:171-179.[Free Full Text]
21. McEwen BS, Seeman T. Protective and damaging effects of mediators of stress. Elaborating and testing the concepts of allostasis and allostatic load. Ann N Y Acad Sci.. 1999;896:30-47.[Medline]
22. McEwen BS. Sex, stress and the hippocampus: allostasis, allostatic load and the aging process. Neurobiol Aging.. 2002;23:921-939.[Medline]
23. Seeman TE, Singer BH, Rowe JW, Horwitz RI, McEwen BS. Price of adaptation—allostatic load and its health consequences. MacArthur studies of successful aging. Arch Intern Med.. Oct 27 1997;157:(19): B2259-2268.
24. Karlamangla AS, Singer BH, McEwen BS, Rowe JW, Seeman TE. Allostatic load as a predictor of functional decline. MacArthur studies of successful aging. J Clin Epidemiol.. 2002;55:696-710.[Medline]
25. Goldman N, Lin I-F, Weinstein M, Lin Y-H. Evaluating the quality of self-reports of hypertension and diabetes. J Clin Epidemiol.. 2003;56:148-154.[Medline]
26. Schnorpfeil P, Noll A, Schulze R, Ehlert U, Frey K, Fischer JE. Allostatic load and work conditions. Soc Sci Med.. 2003;57:647-656.
27. Goldman N, Weinstein M, Chang M-C, et al. 2000 Social Environment and Biomarkers of Aging Study in Taiwan (SEBAS). No. 2003-08. Located at: Office of Population Research Working Paper Series, Princeton University, Princeton, NJ (available for download at http://www.opr.princeton.edu/papers/); 2003.
28. Radloff LS. The CES-D Scale: A self-report depression scale for research in the general population. Appl Psychol Meas.. 1977;1:149-166.
29. Krause N, Liang J. Cross-cultural variations in depressive symptoms in later life. Int Psychogeriatr.. 1992;4:(Suppl 2): B185-202.
30. Ofstedal MB, Zimmer ZS, Lin HS. A comparison of correlates of cognitive functioning in older persons in Taiwan and the United States. J Gerontol Soc Sci.. 1999;54B:S291-S301.
31. Boey KW. Cross-validation of a short form of the CES-D in Chinese elderly. Int J Geriatr Psychiatry.. 1999;14:608-617.[Medline]
32. Shrout PE, Yager TJ. Reliability and validity of screening scales: effect of reducing scale length. J Clin Epidemiol.. 1989;42:69-78.[Medline]
33. Kohout FJ, Berkman LF, Evans DA, Cornoni-Huntley J. Two shorter forms of the CES-D (Center for Epidemiological Studies Depression) depression symptoms index. J Aging Health.. 1993;5:179-193.[Abstract/Free Full Text]
34. Pfeiffer E. A short portable mental status questionnaire for the assessment of organic brain deficit in elderly patients. J Am Geriatr Soc.. 1975;23:433-441.[Medline]
35. Lezak MD. Neuropsychological Assessment. 2nd ed. New York: Oxford University Press; 1983.
36. Wechsler D. WAIS-R Manual. New York: Psychological Corporation; 1981.
37. Manton KG, Woodbury MA, Tolley HD. Statistical Applications Using Fuzzy Sets. New York: John Wiley; 1994.
38. Manton KG, Cornelius ES, Woodbury MA. Nursing home residents: a multivariate analysis of their medical, behavioral, psychosocial, and service use characteristics. J Gerontol Med Sci.. 1995;50A:M242-M251.
39. Manton KG, Liu K, Cornelius ES. An analysis of the heterogeneity of U.S. nursing home patients. J Gerontol B Psychol Sci Soc Sci.. 1985;40:34-46.
40. Wachter KW. Grade of membership models in low dimensions. Stat Papers.. 1999;40:439-457.
41. Charpentier P. GOM3, Version 3.4 (with modifications). Freeware. Available at: http://lib.stat.cmu.edu/DOS/general/.index.html. Accessed December 2002.
42. Singer B. Grade of membership representations: concepts and problems. In: Andersen TW, Athreya KB, Iglehart DL, eds. Probability, Statistics, and Mathematics: Papers in Honor of Samuel Karlin. New York: Academic Press, Inc.; 1989:317–334.
43. Bland JM, Altman DG. Multiple significance tests: the Bonferroni method. BMJ.. 1995;310:170.[Free Full Text]
44. Perneger TV. What's wrong with Bonferroni adjustments. BMJ.. 1998;316:1236-1238.[Free Full Text]
45. Stata Statistical Software [computer program]. Version 7.0. College Station, TX: Stata Corportation; 2001.
46. Yaffe K, Ettinger B, Pressman A, et al. Neuropsychiatric function and dehydroepiandrosterone sulfate in elderly women: a prospective study. Biol Psychiatry.. 1998;43:694-700.[Medline]

Related articles in Journals of Gerontology Series A: Biological Sciences and Medical Sciences:

Erratum

Journals of Gerontology Series A: Biological Sciences and Medical Sciences 2006 61: 209. [Full Text]  

This Article Abstract Full Text (PDF) Services Related articles in Journals of Gerontology Series A: Biological Sciences and Medical Sciences Download to citation manager PubMed PubMed Citation