The Structure of Health Status Among Hispanic, African American, and White Older Adults

Transcription

1 The Journals of Gerontology Series B 1997, Vol. 52B (Special Issue), Copyright 1997 by The Gemntological Society of America 4 The Structure of Health Status Among Hispanic, African American, and White Older Adults Timothy E. Stump, 1 Daniel O. Clark, 12 Robert J. Johnson, 3 and Fredric D. Wolinsky 4 'Indiana University School of Medicine. 2 Regenstrief Institute for Health Care, Indianapolis. 3 Kent State University, Kent, Ohio. "Saint Louis University Health Sciences Center. Activities of daily living (s), instrumental s, and disability markers have traditionally been the most common indicators of functional status. The study on Asset and Health Dynamics Among the Oldest Old (AHEAD) is used to replicate a five-dimensional measurement model composed of these observable indicators among the older adult self-respondents. The items available to measure upper body disability were found wanting, but the lower body disability, and the basic, household, and advanced constructs were confirmed. Analyses of the measurement model separately among subgroups of women, men, Hispanics, Mexican Americans, African Americans, and Whites found no meaningful differences. Two structural models linking the lower body disability, and the basic, household, and advanced constructs to perceived health and depression were also replicated among the older adult self-respondents, as well as separately among African Americans and among Whites. These models reaffirmed the dominant role of lower body disability on the everyday activities of older adults, and on their perceived health and depression. FOR more than three decades, activities of daily living (s) and disability markers have been the most common means used to assess functional health status among older adults (Kane & Kane, 1981, 1987; Ouslander, Osterweil, & Morley, 1991). Among the most cited and widely used measures, either in whole or in part, have been the index of (Katz, Ford, Moskowitz, Jackson, & Jaffee, 1963), the instrumental (I; Lawton & Brody, 1969), Nagi's (1976) disability index, and the Older Americans Research and Services Center instrument (OARS; Duke University Center for the Study of Aging and Human Development, 1978). Despite how central these measures are to gerontological and geriatric health services research, few studies have explicitly addressed the psychometric properties (i.e., the reliability, validity, and factorial structure) of these measures (Kane & Kane, 1981). Moreover, conflicting evidence has emerged from the few psychometric studies that have been conducted (Johnson & Wolinsky, 1993). For example, Spector, Katz, Murphy, and Fulton (1987) and Kempen and Suurmeijer (1990) have suggested that the and I items can be accommodated into a single, multi-item, hierarchical scale. In contrast, Garrad and Bennett (1971), Fillenbaum and Smyer (1981), and Charlton, Patrick, and Peach (1983) have argued that for conceptual and empirical reasons, the and I items need to be modeled separately. Still other studies, such as the research of Lazaridis, Rudberg, Furner, and Cassel (1994), have considered just one measure at a time (in this case, the ) and have raised questions about the accuracy of the hypothesized hierarchical structure within even that one scale. Taking a somewhat different approach, Wolinsky and colleagues have focused on the structure of health status among older adults by examining the items contained in or derived from the, I, and disability scales. Exploratory factor analysis (EFA) techniques used to routinely verify the psychometric properties of these three multi-item scales in the Longitudinal Study of Aging (LSOA) revealed that the items on eating, using the telephone, and managing money would not load with the other /I items (Wolinsky & Johnson, 1991). Those three items did, however, form a coherent, reliable factor of their own, which was labeled an advanced (or cognitive). After removing those three items from the /I pool, a forced two-factor EFA solution resulted in coherent, reliable factors reflecting separate and I dimensions (Charlton et al., 1983; Fillenbaum & Smyer, 1981; Garrad & Bennett, 1971), which were labeled as basic (or self-care) 49

2 50 STUMP ETAL. and household (or home-care) s, respectively. The routine EFA of the disability items also resulted in two coherent, reliable factors, which were labeled as lower body and upper body disabilities. Criterion validity of these five functional status scales was established by their differential ability to predict various health and health behavior outcomes in the LSOA. Replication of the three-dimensional structure of the /I items was obtained from EFA analyses of data from a randomized controlled trial of 668 consecutively hospitalized veterans aged 45 years or older (Fitzgerald, Smith, Martin, Freedman, & Wolinsky, 1993). Those analyses also supported the notion that the advanced was more cognitive in nature than the basic or household s, inasmuch as only the advanced was significantly related to scores on the Short Portable Mental Status Questionnaire (SPMSQ; Pfeiffer, 1975). Based on these findings and building on the work of Nagi (1976) and others, the Structure of Heaith Status Model shown in Figure 1 (and the measurement model it implies) was hypothesized (Johnson & Wolinsky, 1993). This model was assessed in the LSOA with confirmatory factor analysis (CFA) techniques. The specified causal sequence begins with sociodemographic characteristics and disease history; proceeds to disability (represented by the lower and upper body disability constructs); progresses to functional limitations (represented by the basic, household, and advanced constructs); and concludes with perceived health. With only a few minor modifications (i.e., dropping the sitting item from the upper body disability construct, dropping the walking item from the basic construct, and dropping the heavy housework item from the household construct), the CFA of the LSOA data supported if not confirmed the structural model (i.e., no plausible alternative model fit the data appreciably better). Subsequent CFA analyses of the four race-by-gender subsamples of the LSOA (Johnson & Wolinsky, 1994) and the National Long Term Care Survey (NLTCS; Clark, Stump, & Wolinsky, 1997) provided further support for the measurement and structural models. The NLTCS analyses also supported the notion that the advanced construct is more cognitive in nature (based again on its substantially higher correlations with the SPMSQ). The purpose of this article is to assess the Structure of Health Status Model shown in Figure 1 (and the measurement model it implies) using data on the 6,661 selfrespondents to the survey on Asset and Health Dynamics (AHEAD) who were aged 70 years or more. There are six reasons why this is the most stringent test to date of the Structure of Health Status Model, and five of these involve differences in the wording and availability of the observed indicators between the LSOA and the AHEAD survey (Wiener, Hanley, Clark, & Van Nostrand, 1990). First, the stems (i.e., lead-ins) for some of the items tapping various dimensions of health status in AHEAD are rather different from those used in the LSOA. For example, in AHEAD (as in the NLTCS) the stem for the basic items is "Does anyone ever help you...," and the stem for the household and advanced items is "Are you able to... without help," whereas in the LSOA the stem for all /I items is "Because of a health or physical problem, do you Sociodemographic Characteristics Disease Markers Perceived Health Figure 1. The Structure of Health Status Model. Specified, but not shown, are (a) the effects of sociodemographic characteristics and disease markers on each construct; (b) the correlated errors between lower and upper body disability constructs; and (c) the correlated errors between basic, household, and advanced constructs. have any difficulty...." Second, the activities addressed by the AHEAD and LSOA items are not all the same. For instance, the AHEAD study includes items on taking medications (an advanced item) and picking up a dime or driving a car (potential upper body disability items), none of which were in the LSOA; AHEAD does not include several items that were in the LSOA, including difficulties with light and heavy housework (household items); difficulties in stooping, kneeling, or crouching (a lower body disability item); or difficulties with reaching out, reaching over one's head, or grasping objects (upper body disability items). A third difference is that apparently comparable items included in AHEAD are worded differently from the way they were in the LSOA. Some of these differences are dramatic. For example, in the AHEAD study the eating item specifically directs the respondent's attention to mechanical limitations by focusing on "cutting up your food." This focus reflects considerably on basic s, whereas in the LSOA the eating item was not so constrained, leaving it as a more clearly cognitive function (see Wolinsky & Johnson, 1991). Thus, in AHEAD the eating item is likely to produce a substantial cross-loading on the basic construct. An example of a more modest difference is that in AHEAD, the meal preparation item is altered by focusing explicitly on "hot" meals, which places a somewhat more demanding task requirement on the respondent (i.e., competence to use the stove). This suggests that the meal preparation item might result in modest cross-loadings as well, especially on the advanced construct. The fourth difference is that only two items each are available in AHEAD for the household and upper body disability constructs. Moreover, the latter items (i.e., picking up a dime and driving a car) are conceptually complex and suggest the likelihood of several substantial cross-loadings and/or correlated errors. This begs the question of whether an adequate measure of the upper body disability construct can be obtained in the AHEAD study. Fifth, in AHEAD, only one perceived health item the traditional, global self-assessment is available to tap this construct, whereas the LSOA contained three such items. This reliance on a single-item indicator is offset by examining a second structural model that substitutes depression

3 STRUCTURE OF HEALTH STATUS 51 (measured by multi-item subscales for malaise and negative affect) as the health outcome. If support for the Structure of Health Status Model shown in Figure 1 (and the measurement model it implies) is found despite these five differences in the wording and availability of the observed indicators, then the internal validity of that model will be greatly enhanced. Finally, AHEAD contains the largest national probability samples of Hispanic and African American older adults currently available. This facilitates a much needed (Gibson, 1991a, 1991b, 1994; Johnson & Wolinsky, 1994) examination of the effects of ethnicity on the structure of health status as an exogenous (or additive) factor. It also facilitates an examination of the interaction effects associated with ethnicity by replicating the measurement model separately by ethnic group, and the structural model separately for African Americans and Whites (there are not enough cases to support a meaningful, separate analysis among Hispanics). If the models are supported in those separate analyses, then their external validity will be greatly enhanced. METHOD Data The AHEAD study comprises a nationally representative sample of 7,443 community-dwelling adults aged 70 years or older at the time of their baseline interviews in For conceptual and pragmatic reasons, the analyses focus on the 6,661 self-respondents. Conceptually, although proxy reports of s and disabilities are reasonably reliable and valid, the validity of proxy reports involving attitudes, beliefs, and feeling states is, at best, questionable (Rakowski, Mor, & Hiris, 1991). Pragmatically, only the self-respondents in AHEAD were asked the depression items. Thus, if we are to avoid the effects of compositional change in the comparison of the two structural models (i.e., one with perceived health as the outcome, and the other with depression as the outcome), only self-respondents can be included when modeling perceived health. The rate of proxy-respondents is low overall (10.5%) but varies across ethnic groups, being 18.6% among Hispanics, 13.1% among African Americans, and 9.3% among Whites. Therefore, to assess the effect of excluding the proxy-respondents from the measurement model, we have replicated those analyses among all 7,443 respondents. To further eliminate compositional change between the assessments of the two structural models, only the 6,470 respondents for whom complete data for both models exist are included in the analyses (75 respondents had missing data on ethnicity, 4 respondents had missing data on perceived health, and 112 respondents had missing data on one or more of the depression items). Although AHEAD oversampled Hispanics, African Americans, and Floridians, the unweighted data are reported here for two reasons. First, additional analyses of the 6,470 self-respondents (not shown) employing the complex sampling and post-stratification weights yielded equivalent results. Second, weighting the data makes it difficult to estimate the models separately among the oversampled groups. Measurement Table 1 contains the exact wording of the stem questions and the percentage with limitations by ethnic group on the 16 observed indicators used to measure the basic, household, and advanced s as well as the lower and upper body disability latent constructs. All 16 items have been dichotomized. A score of one (1) indicates either (a) receiving help with, (b) not being able to do, (c) having difficulty with, or (d) avoiding (because of health problems) the activity in question. A score of 0 reflects the ability to do the activity without help or difficulty. Because the items on walking several blocks, climbing up one flight of stairs, and pulling or pushing large objects were not asked of those who reported needing help in getting across a room, those items are scored as Is in this circumstance. Perceived health was measured using the traditional self-assessment question that asks respondents, "Would you say your health is excellent (10.7%; coded 5), very good (24.0%; coded 4), good (31.6%; coded 3), fair (22.5%; coded 2), or poor (11.2%; coded 1)?" Depression was measured using five items taken from or based on the Center for Epidemiological Studies-Depression Scale (CES-D; Radloff, 1977). EFA and CFA analyses of these items indicate that they tap the malaise and negative affect dimensions of depression. Those items were coded 1 if present and 0 if absent. Malaise items were feeling depressed (20.6%), feeling sad (19.8%), and feeling lonely (20.9%). Negative affect items were not feeling happy (12.3%) and not enjoying life (7.3%). CFA analyses (available on request) indicate that the two-factor measurement model fits the data well, and better than any alternative model. Thirteen sociodemographic and disease covariates are included in the structural models. The six sociodemographic characteristics are female gender (62.2%), age (years beyond age 70; M = 7.3), years of education (M = 10.9), and a set of three dummy variables reflecting ethnicity. These variables include markers for being Mexican American (3.0%), other Hispanic (2.2%), or African American (13.4%). Being White (81.4%) was the omitted or reference category. Although the numbers of Mexican Americans and other Hispanics are small, their separation in the pooled analysis permits the partial reduction of the heterogeneity that otherwise exists. Seven disease markers were included. Five are dichotomous indicators of whether the respondent had ever had cancer (13.8%), chronic lung disease (such as bronchitis or emphysema; 10.9%), heart disease (such as a heart attack, coronary heart disease, angina, or coronary heart failure; 31.1%), a stroke or transient ischemic attack (9.2%), or psychiatric problems (10.6%). The other two disease markers are dichotomous indicators of whether the respondent had had diabetes (13.0%) or arthritis or rheumatism (26.1%) during the past year. Analysis and Estimation Two phases to the analyses follow. The first phase assesses the fit of the measurement model. This assessment is done initially among the entire sample of 6,470 selfrespondents with complete data, and subsequently among the separate samples of 4,023 women and 2,447 men, and

4 52 STUMP ETAL. Table 1. Exact Wording of, and the Percent With Limitations on the 16 Observed Indicators Used To Measure the Basic, Household, and Advanced Latent Constructs, as Well as the Lower and Upper Body Disability Latent Constructs, Among the 6,470 Self-respondents Aged 70 Years or Older With Complete Data Items Does anyone ever help you dress, including putting on shoes and socks? Does anyone ever help you bathe or shower? Does anyone ever help you get in and out of bed? Does anyone ever help you use the toilet, including getting up and down? Are you able to prepare hot meals without help? Are you able to shop for groceries without help? Advanced Does anyone ever help you eat, such as cutting up your food? Do you manage money such as paying your bills and keeping track of expenses, without anyone's help? Are you able to make telephone calls without help? Are you able to take medications without help? Lower Body Limitation Do you have any difficulty walking several blocks? Do you have any difficulty climbing one flight of stairs without resting? Do you have any difficulty pulling or pushing large objects like a living room chair? Do you have any difficulty lifting or carrying weights over 10 pounds, like a heavy bag of groceries? Upper Body Limitation Do you have any difficulty picking up a dime from a table? Are you able to drive? of 338 Hispanics, 193 Mexican Americans, 864 African Americans, and 5,268 Whites. The second phase of the analysis examines two structural models. In one, the outcome is perceived health, and in the other, the outcome is depression. These analyses are also done initially among the entire sample of self-respondents with complete data; they include a dichotomous marker for gender, as well as a set of dummy variables reflecting ethnicity (i.e., being Mexican American, other Hispanic, or African American, vs being White). Next, these structural models are repeated among the separate samples of African Americans and Whites (there are not enough cases to support a meaningful separate analysis among Hispanics), with the ethnicity markers removed. The measurement and structural models are estimated using LISREL VIII (Joreskog & Sorbom, 1995). Because all items in the measurement models are dichotomous, the estimation method should rely on the gold standard of polychoric correlations and weighted least squares procedures (Joreskog, 1990). This minimizes the artificial attenuation of the standardized factor loadings and the bias that would otherwise result if maximum likelihood estimates were derived from product-moment correlations that are based on these dichotomous items (Bollen, 1989; Muthen, 1978, 1984, 1993). At the same time, however, reliance on this gold standard imposes a number of constraints. Principal All Self-Respondents N = 6, Percent With Limitations Hispanics n = African Americans n = Whites n = 5, among them is the very large sample size requirement for estimating the weight matrix, which increases quadratically when the number of observed variables approaches 20 (Bentler & Newcomb, 1992). This is problematic even with the large sample size in AHEAD, because the initial number of observed variables in the Structure of Health Status Model is 30 for perceived health and 34 for depression. Accordingly, although the gold standard can and is used to assess all of the measurement models, as well as the structure of health status models in the absence of the exogenous variables, the effects of the exogenous variables themselves can only be estimated with a less constraining method. That method involves using weighted least squares procedures on an asymptotic covariance structure derived from a product-moment correlation matrix (Joreskog & Sorbom, 1995). This innovative mixed analysis approach is "backed up" by comparing the estimates obtained from the gold standard method with those obtained from the less constraining method (in the absence of the exogenous variables) as an added safeguard (Bentler & Newcomb, 1992). Consistent with standard practice (Bentler & Bonett, 1980; Bollen & Long, 1993; Tanaka, 1993), overall goodness of fit is assessed using multiple measures. These include the adjusted goodness of fit index (AGFI), the chisquare to degrees of freedom ratio (x 2 /df), the root mean squared residual (RMR), and the root mean square error of

5 STRUCTURE OF HEALTH STATUS 53 approximation (RMSEA). A good fit overall is usually defined by an AGFI greater than.90, a x 2 /df less than 5 and preferably less than 3, an RMR less than.05, and an RMSEA less than.05 (Bollen, 1989). Of these measures, the RMSEA (a measure of the average discrepancy [between the estimated and observed correlations] per degree of freedom) is most appropriate to these analyses because it (a) is insensitive to both sample size and the scale of the dependent variable, (b) is sensitive to the number of parameters in the model (which minimizes the incentive to overparameterize), and (c) provides a more realistic test of close fit (rather than the exact fit interpretation of chi-square tests; see Browne & Cudeck, 1993). Individual items are evaluated on the basis of their standardized factor loadings, the /-values associated with those loadings, and the modification indices obtained for both cross-loadings and correlated error terms not specified in the original model. This evaluation is done, however, with the proviso that any parameter added to the model must be substantively meaningful (Bollen, 1989; Bollen & Long, 1993; Joreskog & Sbrbom, RESULTS Table 2 contains the standardized factor loadings obtained from three CFA measurement models using the gold standard of polychoric correlations and weighted least squares procedures among the 6,470 self-respondents with complete data. Also shown are the values for the AGFI, the X 2 /df ratio, the RMR, and the RMSEA. The first model consists of the 16 items and their hypothesized placement on the five latent constructs. As indicated, the standardized factor loadings are all large (i.e.,.60 or greater), and the goodness-of-fit measures generally indicate a good fit overall (i.e., the AGFI is greater than.90, the x 2 /df is less than 5, the RMSEA is less than.05, and the RMR is only somewhat greater than.05). The f-values (not shown) associated with the standardized factor loadings and the modification indices (also not shown) associated with the unspecified cross-loadings, however, suggest problems with the items about eating, picking up a dime, and driving a car. Specifically, the f-value for the eating item is considerably smaller than that for the other items loading on the advanced construct, and the modification indices for its cross-loadings on all of the other constructs are substantially greater, especially on the basic construct. The f-values for the picking up a dime and driving a car items are substantially lower than those for any of the other items on their respective constructs. Moreover, the modification indices for the cross-loadings of the picking up a dime and driving a car items are substantial, and greater on the basic and household and lower body disability constructs than their f-values on the upper body disability construct. Thus, and as expected, the eating, picking up a dime, and driving a car items do not work well in the AHEAD study, and they should be dropped from the measurement model. This finding results in the second model shown in Table 2. It consists of the 13 remaining items and 4 latent constructs. All four of the goodness-of-fit measures now indicate a good fitting model. Again, however, an examination of the modification indices reveals one additional relationship that warrants further attention. It involves the correlated error between pulling or pushing large objects, and carrying 10-pound items, which has a modification index of 83. This correlated error is plausible and has a straightforward substantive interpretation. Both these tasks have an element of upper body mobility and strength. In the absence of an independent upper body disability construct, freely estimating this correlated error is essential. This is done in the third model shown in the table, which represents a substantial improvement to an already good-fitting measurement model. Moreover, all of the remaining modification indices are less than 20, and the overwhelming majority are less than 10. Thus, the third measurement model fits the data extremely well, both globally and locally. Equivalent results (not shown) were obtained when the proxy-respondents were included. Nonetheless, as an added safeguard a number of alternative models were estimated. These all included the 13 items shown in the third model of Table 2, as well as the correlated error term. Where the models differed was on the number of constructs, and the alignment of the items on the constructs. Among these alternatives were a single-con- Table 2. Standardized Factor Loadings Estimated Using Polychoric Correlations and Weighted Least Squares Procedures in LISREL VIII for Measurement Models With Five Latent Constructs (Model 1), With Four Latent Constructs (Model 2), and With Four Latent Constructs and One Correlated Error Term (Model 3), Using the Unweighted Sample of 6,470 Self-respondents Over Age 70 With Complete Data" Item Dressing Bathing or showering Getting in and out of bed Using the toilet Preparing hot meals Shopping for groceries Advanced Eating Managing money Making telephone calls Taking medications Lower Body Limitation Walking several blocks Climbing one flight of stairs Pulling or pushing large objects Carrying 10 lbs. Upper Body Limitation Picking up a dime Driving a car Model Fit Statistics AGFI XVdf RMR RMSEA Model Model Model "Correlated error term involves pushing large objects and carrying 10- pound items.

6 54 STUMP ETAL. struct model (reflecting complete unidimensionality), a two-construct model (consisting of the advanced and a residual construct), and two three-construct models (one consisting of the advanced and household constructs along with a combined construct of basic and lower body disability items, and one consisting of the advanced and basic constructs along with a combined construct of household and lower body disability items). None of these alternatives came close to approximating the fit of the third model shown in Table 2. Indeed, the x 2 /df ratios and RMRs exceeded the thresholds for good fitting models, and the RMSEAs were all at least 50% larger. Table 3 contains the standardized factor loadings and goodness-of-fit measures using the gold standard of polychoric correlations and weighted least squares procedures for the 13-item, four-construct, one correlated error measurement model (i.e., the third model in Table 2) separately among the 4,023 women, 2,447 men, 338 Hispanic, 193 Mexican American, 864 African American, and 5,268 White self-respondents. Four points are worth noting here. First, the standardized factor loadings are all large and roughly equivalent across the gender and ethnicity groups. Second, the measurement model fits the data very well in each gender and ethnicity group. That is, the x 2 /df rat io and the AGFI always indicate a good fit. So, too, does the RMSEA, except among Mexican Americans, where it is marginally not a close fit (i.e.,.059). Indeed, only the RMR consistently falls above the desired level among the minority groups. Third, there are no blatant modification indices (not shown) for either the unspecified crossloadings or correlated errors. Fourth, there are, however, three standardized factor loadings marginally greater than. This most likely results from the small sample size among the minority groups coupled with the reliance on the polychoric correlation matrix and weighted least squares estimation procedure. Known as "Heywood cases" (Bollen, 1989), a standard solution to this problem is to fix the marginally negative error variances for the items involved to zero, and reestimate the models. When this constraint is invoked (not shown), the results are equivalent to those shown in the table. Therefore, the fit of the measurement model is extremely robust across all of the gender and ethnicity groups. The top panel of Table 4 contains the results obtained from the analysis of the Structure of Health Status Model when perceived health is the outcome using the gold standard of polychoric correlations with weighted least squares procedures in the absence of the exogenous variables among the 6,470 self-respondents with complete data. Shown are the standardized coefficients, asterisks indicating their p- values, and the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA. The bottom panel of Table 4 contains the same information obtained when depression is the outcome of the Structure of Health Status Model. Four points stand out in the top panel of the table. First, the lower body disability construct has highly statistically significant effects on all three of the constructs. Moreover, the magnitude of those effects is quite large on the basic (.93) and household (.89) constructs, although it Table 3. Standardized Factor Loadings Estimated Using Polychoric Correlations and Weighted Least Squares Procedures in LISREL VIII for Measurement Models With Four Latent Constructs and One Correlated Error Term, by Gender and Ethnicity, for Self-respondents Over Age 70 With Complete Data 0 Item Dressing Bathing or showering Getting in and out of bed Using the toilet Preparing hot meals Shopping for groceries Advanced Managing money Making telephone calls Taking medications Lower Body Limitation Walking several blocks Climbing one flight of stairs Pulling or pushing large objects Carrying 10 lbs. Model Fit Statistics AGFI XVdf RMR RMSEA Women n = 4, Men n = 2, Hispanic n = Mexican American n= African American n = White n = 5, 'Correlated error term involves pushing large objects and carrying 10-pound items.

7 STRUCTURE OF HEALTH STATUS 55 Table 4. Standardized Coefficients Estimated Using Polychoric Correlations and Weighted Least Squares Procedures in LISREL VIII for the Structure of Health Status Models in Which Perceived Health (top panel) and Depression (bottom panel) are the Outcomes in the Absence of the Exogenous Variables Among the 6,470 Self-respondents With Complete Data" Independent Concept Perceived Health Model Advanced Lower body Model Fit Statistics AGF1 = XVdf=2.59 RMR =.038 RMSEA =.O16 Depression Model Advanced Lower body Model Fit Statistics AGF1= X 2 /df = 2.31 RMR =.043 RMSEA =.O14 Perceived Health -.69*** Malaise.28* Negative Affect * Note: = not applicable. "Includes a correlated error term involving the pushing large objects and carrying 10-pound items. */><.05; **p< ;***/><.001. is somewhat less on the advanced construct (.62). Second, none of the constructs have statistically significant effects on perceived health net of the lower body construct. Third, the lower body disability construct does have a highly statistically significant effect on perceived health, and that effect is also quite large (-.69). Fourth, the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA all indicate a good fit of the model to the data, as do the R 2 levels for the basic (. 78), household (.71), advanced (.34), and perceived health (.40) constructs. Consistent with our expectations and previous reports (Johnson & Wolinsky, 1993, 1994), these results confirm that lower body disability plays the most important role in how older adults function on a daily basis, and how they perceive their health. Equivalent results (not shown) were obtained when the model was estimated separately among Whites. When the model was estimated separately among African Americans, equivalent results were also obtained, with one exception: The effect of the lower body disability construct on perceived health was not statistically significant. The bottom panel of Table 4 also can be summarized with four points. First, the lower body disability construct once again has highly significant effects on each of the constructs, and these are all quite large (range =.64 to.94). Second, the only construct to have a significant effect on depression is the basic. It has a modest effect (.28) on the malaise subscale, but no meaningful effect on the negative affect subscale. Third, and in contrast, the lower body construct has a statistically significant, modest 93*** 94*** Household 89***.89*** Advanced.62***.64*** effect (.25) on the negative affect subscale, but no meaningful effect on the malaise subscale. Fourth, the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA once again all indicate a good fit of the model to the data, although the R 2 levels for the malaise (.22) and negative affect (.22) constructs are not as robust as they were for the perceived health construct. Equivalent results (not shown) were obtained when the model was estimated separately among African Americans and among Whites. Using the less constraining approach that involves weighted least squares procedures, along with an asymptotic covariance matrix derived from product-moment correlations (not shown), yields very similar results. Indeed, the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA all indicate good fits for both the perceived health and depression models, and only three meaningful differences were observed among the parameter estimates. One involved the effect of the advanced construct on perceived health. As shown in the upper panel of Table 4, the gold standard estimate of that effect is not statistically significant (t = -.19) and is virtually zero (). In contrast, the less constraining approach estimated the effect of the advanced construct on perceived health as modest (-.18) and statistically significant (t = -2.48). A second difference involved the effect of the lower body disability construct on malaise. As shown in the lower panel of Table 4, the gold standard estimate of that effect is modest (.16) but not statistically significant (t = 1.62). The estimated effect of the lower body disability construct on malaise from the less constraining approach, however, while still modest

8 56 STUMP ETAL. Independent Concept (.26), is highly statistically significant (t = 8.59). The third difference involved the effect of the lower body disability construct on perceived health when estimated separately among African Americans. Using the less constraining approach, this effect is statistically significant. Given the overwhelming consistency between the gold standard and the less constraining approach, researchers seem justified in proceeding to estimate the effects of the exogenous variables using the latter method, although the effect of the advanced construct on perceived health (especially) and the effect of the lower body disability construct on malaise will be interpreted with caution. Table 5 contains the results obtained from the analysis of the Structure of Health Status Model when perceived health is the outcome using the less constraining approach involving weighted least squares procedures and an asymptotic covariance matrix derived from product-moment correlations among the 6,470 self-respondents with complete data. The table shows the standardized coefficients with asterisks indicating their p-values, and the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA. Because of AHEAD's large sample size and the potential for the less constraining method to underestimate standard errors, only coefficients significant at the.001 level or beyond will be considered meaningful. There are four major patterns in Table 5; equivalent patterns (not shown) were found when the model was estimated separately among African Americans and among Whites. Two of those major patterns are easily summarized. First, the overall fit of the model for perceived health remains solid after introducing the exogenous variables. This is reflected in the stable, stellar values for the AGFI, the X 2 /df ratio, the RMR, and the RMSEA. Moreover, the introduction of the exogenous variables increases the R 2 for perceived health to.41 (from.33), and increases the R 2 for the advanced construct to.29 (from.19). In addition, the exogenous variables explain 33% of the variance in the lower body disability construct (which had been considered exogenous in Table 4). Second, the effect of the lower body disability construct on perceived health, as well as the noneffects of the constructs on perceived health, is equivalent to those described previously for the upper panel of Table 4, regardless of the estimation method used, and need not be reiterated. A third major pattern in Table 5 involves the effects of the sociodemographic characteristics. These are generally consistent with our expectations and previous reports (Johnson & Wolinsky, 1993, 1994). Women report more lower body disability than men (but less basic and advanced difficulty) and better perceived health (cf. Verbrugge, 1985). Older adults have more lower body disability, and more basic, household, and advanced difficulty, although their perceived health is no different from that of younger adults (cf. Ferraro, 1980). Better educated respondents have less lower body disability and advanced difficulty, along with better perceived health (cf. Guralnik, Land, Blazer, Fillenbaum, & Branch, 1993). Among the ethnicity comparisons, the only meaningful difference is that African Americans perceive their health to be poorer than that of their White counterparts (cf. Ferraro, Table 5. Standardized Coefficients Estimated Using Weighted Least Squares Procedures and an Asymptotic Covariance Matrix Derived From Product-Moment Correlations in LISREL VIII for the Structure of Health Status Model With Perceived Health as the Outcome Among the 6,470 Self-respondents With Complete Data" Advanced Lower body Female Years from age 70 Education African American Mexican American Non-Mexican American Hispanic Diabetes Cancer Lung disease Heart disease Stroke or TIA Psychiatric problems Arthritis Perceived Health *** *.15*** - - ** -.10*** *** -.16*** -.05*** _ Basic.68*** -.05***.05***.03**.03** - _03*** Household.68*** 12*** -.03** *.07*** -.04** Advanced.21*** _ -.08*** - 03*** ** ** _03*** Lower Body.16***.13*** -.08*** * 13***.13*** 11***.05***.21*** Model Fit Statistics AGFI =.98 X7df = 4.87 RMR = 8 RMSEA = 4 "Includes a correlated error term involving the pushing large objects and carrying 10-pound items. *p <.05; **p < ; ***/? <.001.

9 STRUCTURE OF HEALTH STATUS ), which is somewhat surprising (cf. Andrews, 1989; Council on Scientific Affairs, 1991). The fourth major pattern in Table 5 involves the effects of the disease markers. With two exceptions, these are also generally consistent with our expectations and previous reports (Johnson &Wolinsky, 1993, 1994). That is, when the disease markers have meaningful effects, those effects always result in poorer functional status or poorer perceived health. More specifically, stroke has deleterious effects on all five health status constructs. Diabetes has deleterious effects on all of the health status constructs except the basic s. Chronic lung disease has deleterious effects on the basic s, lower body disability, and perceived health constructs. Heart disease, cancer, and arthritis each have deleterious effects on the advanced s and perceived health constructs. Psychiatric problems have deleterious effects only oh the lower body disability construct. The two unanticipated findings are (a) the protective effect of chronic lung disease on the basic construct, and (b) the protective effect of arthritis on the advanced construct. These findings reflect artificial overcompensations that result from the substantially stronger deleterious effects of chronic lung disease and arthritis on the lower body disability construct. This was verified when the lower body construct was removed from the model (not shown), and the protective effects disappeared. Table 6 contains the standardized coefficients, asterisks indicating their/7-values, and the values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA obtained from the analysis of the Structure of Health Status Model when depression is the outcome among the 6,470 self-respondents with complete data. Once again, because of AHEAD's large sample size and the potential for using weighted least squares procedures and an asymptotic covariance matrix derived from product-moment correlations to underestimate standard errors, only coefficients significant at the.001 level or beyond will be considered meaningful. As with Table 5, there are four major patterns in Table 6, and the first two are easily summarized; equivalent patterns (not shown) were found when the model was estimated separately among African Americans, and among Whites. The first is that the overall fit of the model for depression remains solid after introducing the exogenous variables. This pattern is reflected in the stable, stellar values for the AGFI, the x 2 /df ratio, the RMR, and the RMSEA. Moreover, the introduction of the exogenous variables increases the R 2 for the malaise construct to.19 (from.14), and increases the R 2 for the negative affect construct to.13 (from.11). Second, the effect of the lower body disability construct on the malaise and negative affect constructs, as well as the effect of the basic construct on the malaise construct, is equivalent to those described previously for the lower panel of Table 4 when the less constraining method was used, and does not need to be reiterated. A third major pattern in Table 6 involves the sociodemographic characteristics. With two exceptions, their effects on the lower body disability and constructs are the same as those shown in Table 5, and need not be reiterated. The two new effects are a modestly worse level for African Americans (compared to Whites) on the basic con- Table 6. Standardized Coefficients Estimated Using Weighted Least Squares Procedures and an Asymptotic Covariance Matrix Derived From Product-Moment Correlations in LISREL VIII for the Structure of Health Status Model With Depression as the Outcome Among the 6,470 Self-respondents With Complete Data" Independent Concept Advanced Lower body Female Years from age 70 Education African American Mexican American Non-Mexican American Hispanic Diabetes Cancer Lung disease Heart disease Stroke or TIA Psychiatric problems Arthritis Malaise [2***.06 19*** _ 09***.04** -.03** 09*** Negative Affect ***.09*** 69*** -.05***.05*** *.03** - _.03** Household.68*** 12*** -.05***.03* *. * * - _ Advanced 2i*** 09*** - 03*** *** ** 05*** _ Lower Body.16*** 14*** _09*** - 03***.13*** 12*** 1 ]*** 20*** Model Fit Statistics AGFI =.98 X 2 /df = 4.50 RMR = 7 RMSEA =3 "Includes a correlated error term involving the pushing large objects and carrying 10-pound items. *p <.05; **p < ; ***p <.001.

10 58 STUMP ETAL. struct, and a modestly better level on the household construct for those African Americans with more education. Both of these effects were shown in Table 5, but they fell marginally short of the meaningfulness criterion. Three of the sociodemographic characteristics have meaningful effects on depression. As expected (cf. Mirowsky & Ross, 1989), women report more malaise than do men, people with more education report less malaise, and non-mexican American Hispanics report more malaise and negative affect than do Whites. Somewhat unexpectedly, age and being either African American or Mexican American are unrelated to depression (cf. Mirowsky & Ross, 1989, 1992). The fourth major pattern in Table 6 involves the disease markers. With three exceptions, their effects on the lower body disability and constructs are the same as those shown in Table 5, and need not be reiterated. The three new findings are (1) the protective effect of chronic lung disease on the advanced construct, (2) the protective effect of arthritis on the household construct, and (3) the absence of a meaningful deleterious effect of stroke on the basic construct. Once again, the two new protective effects reflect artificial overcompensations that result from the substantially stronger deleterious effects of chronic lung disease and arthritis on the lower body disability construct; this finding was verified when the lower body construct was removed from the model (not shown), and the protective effects disappeared. The deleterious effect of stroke on the basic construct that had just met the meaningfulness criterion in Table 5 now just misses. Interestingly, the only disease marker that has a meaningful effect on depression is having a history of psychiatric problems. Individuals with psychiatric problems report more malaise and negative affect. DISCUSSION In this article we have assessed the Structure of Health Status Model shown in Figure 1, as well as the measurement model it implies on the unweighted sample of 6,470 self-respondents to the AHEAD survey who were aged 70 years or more and had complete data. Because of the dichotomous nature and the large number of the observed variables in the models, parameter estimation relied on a mixed approach involving two methods contained in LISREL VIII (Joreskog & Sorbom, 1995). One method was the gold standard of polychoric correlations and weighted least squares procedures (Joreskog, 1990). The other was the less constraining method of an asymptotic matrix derived from product-moment correlations and weighted least squares procedures (Bentler & Newcomb, 1992). Both methods produced equivalent results, enhancing our confidence in the findings. As shown in Table 2, the measurement model was supported, if not confirmed, with only two relatively minor modifications. One involved dropping both of the items available to tap the upper body disability construct (as well as the construct itself), and dropping the eating item from the advanced construct. As expected, the two upper body items were simply inadequate (i.e., nondiscreet) measures of the construct, and AHEAD contains no alternatives. Similarly, and as expected, the eating item also was not discreet, producing substantial cross-loadings on all of the other constructs. The other minor modification involved freely estimating the correlated error term that involves two lower body disability items: pushing large objects and carrying 10 pounds, both of which have an element of upper body mobility and strength. Furthermore, estimating the measurement model separately among women, men, Hispanic, Mexican American, African American, and White self-respondents revealed equivalent results, as shown in Table 3. Thus, despite substantial differences in the wording and availability of observed indicators in the AHEAD survey (compared to the LSOA or the NLTCS), the multidimensional structure of the basic, household, advanced, and lower body disability constructs has also been found to be rather robust across a variety of gender and ethnic groupings. This suggests that future studies using /I and/or disability markers need to disaggregate those measures to reflect the multidimensional structure of health status shown in Figure 1. Continued reliance on anything less, especially on single (i.e., gross) aggregations of /I and/or disability markers, is simply unrealistic and inappropriate. Similarly, the Structure of Health Status Model has also been supported, if not confirmed, with one exception. The most striking support is the resounding reaffirmation of lower body disability as the most important factor in how older adults function on a daily basis, how they perceive their health, and whether they are depressed. Indeed, the standardized coefficients for lower body disability on the basic, household, and advanced constructs, as well as on perceived health and depression, are always the largest. That is, within columns of Tables 5 and 6, the row entries (effect sizes) for the lower body construct are always larger than those for any other variable. What these findings failed to support about the Structure of Health Status Model was its assumption of saturated effects. As shown in Figure 1, the model assumed that each of the antecedents of perceived health or depression had direct effects on perceived health or depression. This is not the case. Indeed, none of the constructs (i.e., basic, household, or advanced) have meaningful direct effects on perceived health. And of the constructs, only the basic has a meaningful direct effect on depression, and then only for malaise (and just when the less constraining estimation method is used). In contrast, some of the sociodemographic characteristics and disease markers, and the lower body disability construct, always have meaningful direct effects on perceived health, malaise, and negative affect. It is important to note that the absence of meaningful direct effects of the constructs on perceived health and depression is not an artifact of erroneously placing the lower body disability construct as their antecedent in the Structure of Health Status Model. Indeed, strong theoretical precedents exist to substantiate that placement (Johnson & Wolinsky, 1993, 1994). And using longitudinal data from the multisite Established Populations for the Epidemiologic Study of the Elderly (EPESE), Guralnik, Ferrucci, Simonsick, Salive, and Wallace (1995) have demonstrated that physical performance tests of essentially lower body abilities predict the onset of /I difficulties over time.

11 STRUCTURE OF HEALTH STATUS 59 Nonetheless, as an added safeguard, the models shown in Tables 5 and 6 were reestimated (not shown) without the lower body disability construct. Although the basic and household constructs do pick up some of the joint variance previously attributed to the lower body disability construct, the R 2 for perceived health (for example) drops from.41 to.33, and the R 2 s for the basic, household, and advanced constructs themselves drop to.18 (from.48),.21 (from.45), and.19 (from.29), respectively. Thus, at the same time that the basic, household, and advanced constructs do not bring any unique explained variance to modeling perceived health or depression, the lower body disability construct does, and it also brings considerable unique explained variance to modeling the basic, household, and advanced constructs. This suggests that the long-standing emphasis on / I measures in gerontological and geriatric health services research on community-dwelling older adults is misplaced, at least when the outcome is perceived health or depression, and most likely when the outcome involves any other measure of health-related quality of life (see Wilson & Cleary, 1995). Therefore, interventions designed to improve health-related quality of life among communitydwelling older adults should focus on improving and maintaining lower body function rather than addressing environmental adaptations that target specific and I difficulties. Perhaps the strongest intervention possibilities lie with increasing physical activity, which has been repeatedly shown in both randomized controlled trials and observational epidemiologic studies to maintain and improve lower body function among older adults (see Wolinsky, Stump, & Clark, 1995). Although recommendations to increase physical activity have become routine (cf. Pate et al., 1995), the results reported here clarify that a principal target of such efforts should be lower body function. ACKNOWLEDGMENTS This work was supported by grants to Dr. Wolinsky from the National Institutes of Health (R37 AG-09692) and the AARP Andrus Foundation. The opinions expressed here are those of the authors and do not necessarily reflect those of the funding agencies, academic, or governmental institutions involved. Portions of this research were presented at the Annual Meeting of The Gerontological Society of America, Los Angeles, November Address correspondence to Dr. Fredric Wolinsky, Health Sciences Center, Saint Louis University, 3663 Lindell Boulevard, St. Louis, MO wolinsky@wpogate.slu.edu REFERENCES Andrews, J. (1989). Poverty and poor health among elderly Hispanic Americans. A report of the Commonwealth Fund Commission on elderly people living alone. New York: The Commonwealth Fund. Bentler, P. M., & Bonett, D. G. (1980). Significance tests and goodness of fit in the analysis of covariance structures. Psychological Bulletin, 88, Bentler, P. M., & Newcomb, N. D. (1992). Linear structural equation modeling with nonnormal continuous variables. Application: Relations among social support, drug use, and health in young adults. In J. H. Dwyer, M. H. Feinleib, P. Lippert, & H. Hoffmeister (Eds.), Statistical models for longitudinal studies in health (pp ). New York: Oxford University Press. Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley. Bollen, K. A., & Long, J. S. (1993). Introduction. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation models. Newbury Park, CA: Sage. Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fit. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation models (pp ). Newbury Park, CA: Sage. Charlton, J. R. H., Patrick, D. L., & Peach, H. (1983). The use of multivariate measures of disability in health surveys. Journal of Epidemiology and Community Health, 37, Clark, D. O., Stump, T. E., & Wolinsky, F. D. (1997). A race and gender specific replication of dimensions of disability and functional status. Journal of Aging and Health, 9, Council on Scientific Affairs (of the American Medical Association). (1991). Hispanic health in the United States. Journal of the American Medical Association, 265, Duke University Center for the Study of Aging and Human Development. (1978). Multidimensional functional assessment: The OARS methodology. Durham, NC: Duke University Press. Ferraro, K. F. (1980). Self-ratings of health among the old and old-old. Journal of Health and Social Behavior, 21, Ferraro, K. F. (1993). Are African-American older adults health pessimistic? Journal of Health and Social Behavior, 34, Fillenbaum, G. G., & Smyer, M. A. (1981). The development, validity, and reliability of the OARS multidimensional functional assessment questionnaire. Journal of Gerontology, 36, Fitzgerald, J. F., Smith, D. M., Martin, D. K., Freedman, J. A., & Wolinsky, F. D. (1993). Replication of the multidimensionality of activities of daily living. Journal of Gerontology: Social Sciences, 48, S28-S31. Garrad, J., & Bennett A. E. (1971). A validated interview schedule for use in population surveys of chronic disability. British Journal of Preventive and Social Medicine, 25, Gibson, R. C. (1991a). Race and the self-reported health of elderly persons. Journal of Gerontology. Social Sciences, 46, S235-S242. Gibson, R. C. (1991b). Age by race differences in the health and functioning of elderly persons. Journal of Aging and Health, 3, Gibson, R. C. (1994). The age-by-race gap in health and mortality in the older population: A social science research agenda. The Gerontologist, 34 t Guralnik, J. M., Ferrucci, L., Simonsick, E. M., Salive, M. E., & Wallace, R. B. (1995). Lower-extremity function in persons over the age of 70 years as a predictor of subsequent disability. New England Journal of Medicine, 332, Guralnik, J. M., Land, K. C, Blazer, D., Fillenbaum, G. C, & Branch, L. G. (1993). Educational status and active life expectancy among older African-Americans and Whites. New England Journal of Medicine, 329, Johnson, R. J., & Wolinsky, F. D. (1993). The structure of health status among older adults: Disease, disability, functional limitation, and perceived health. Journal of Health and Social Behavior, 34, Johnson, R. J., & Wolinsky, F. D. (1994). Gender, race, and health: The structure of health status among older adults. The Gerontologist, 34, Joreskog, K. G. (1990). New developments in LISREL: Analysis of ordinal variables using polychoric correlations and weighted least squares. Quality and Quantity, 24, Joreskog, K. G., & Sorbom, D. (1995). LISREL VIII: Analysis of linear structural relationships by maximum likelihood and least squares methods. Mooresville, IN: Scientific Software, Inc. Kane, R. A., & Kane, R. L. (1981). Assessing the elderly: A practical guide to measurement. Lexington, MA: Lexington Books. Kane, R. A., & Kane, R. L. (1987). Long term care: Principles, programs, and policies. New York: Springer. Katz, S., Ford, A. B., Moskowitz, R. W, Jackson, B. A., & Jaffee, M. W. (1963). Studies of illness in the aged. The index of : A standardized measure of biological and psychosocial function. Journal of the American Medical Association, 185, Kempen, G. I. J. M., & Suurmeijer, T. P. B. M. (1990). The development of a hierarchical polychotomous -I scale for noninstitutionalized elders. The Gerontologist, 30, Lawton, M. P., & Brody, E. M. (1969). Assessment of older people: Self-monitoring and instrumental activities of daily living. The Gerontologist, 9,

12 60 STUMP ETAL. Lazaridis, E. N., Rudberg, M. A., Furner, S. E., & Cassel, C. K. (1994). Do activities of daily living have a hierarchical structure? An analysis using the Longitudinal Study of Aging. Journal of Gerontology: Medical Sciences, 49, M47-M51. Mirowsky, J., & Ross, C. E. (1989). Social causes of psychological distress. New York: Aldine de Gruyter. Mirowsky, J., & Ross, C. E. (1992). Age and depression. Journal of Health and Social Behavior, 33, Muthen, B. O. (1978). Contributions to factor analysis of dichotomous variables. Psychometrika, 43, Muthen, B. O. (1984). A general structural equation model with dichotomous, ordered categorical, and continuous latent variable indicators. Psychometrika, 49, Muthen, B. O. (1993). Goodness of fit with categorical and other nonnormal variables. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation models (pp ). Newbury Park, CA: Sage. Nagi, S. Z. (1976). An epidemiology of disability among adults in the United States. Milbank Memorial Fund Quarterly, 54, Ouslander, J., Osterweil, D., & Morley, J. (1991). Medical care in the nursing home. New York: McGraw-Hill. Pate, R. R., Pratt, M., Blair, S. N., Haskell, W. L., Macera, C. A., Bouchard, C, Buchner, D., Ettinger, W., Heath, G. W., King, A. C, Kriska, A., Leon, A. S., Marcus, B. R, Morris, J., Paffenberger, R. S., Patrick, K., Pollock, M. L., Rippe, J. M., Sallis, J., & Wilmore, J. H. (1995). Physical activity and public health. A recommendation from the Centers for Disease Control and the American College of Sports Medicine. Journal of the American Medical Association, 273, Pfeiffer, E. (1975). A short portable mental status questionnaire for the assessment of organic brain deficit in elderly patients. Journal of the American Geriatrics Society, 10, Radloff, L. S. (1977). The CES-D scale: A self-report depression scale for research in the general population. Applied Psychological Measurement, 1, Rakowski, W., Mor, V., & Hiris, J. (1991). The association of self-rated health with two-year mortality in a sample of well elderly. Journal of Aging and Health, 3, Spector, W. D., Katz, S., Murphy, J. B., & Fulton, J. P. (1987). The hierarchical relationships between activities of daily living and instrumental activities of daily living. Journal of Chronic Disease, 40, Tanaka, J. S. (1993). Multifaceted conceptions of fit in structural equation models. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation models (pp ). Newbury Park, CA: Sage. Verbrugge, L. M. (1985). Gender and health: An update on hypotheses and evidence. Journal of Health and Social Behavior, 24, Wiener, J. M., Hanley, R. J., Clark, R., & Van Nostrand, J. F. (1990). Measuring the activities of daily living: Comparisons across national surveys. Journal of Gerontology: Social Sciences, 45, S229-S237. Wilson, I. B., & Cleary, P.D. (1995). Linking clinical variables with health-related quality of life. Journal of the American Medical Association, 273, Wolinsky, F. D., & Johnson, R. J. (1991). The use of health services by older adults. Journal of Gerontology: Social Sciences, 46, S345-S357. Wolinsky, F. D., Stump, T. E., & Clark, D. O. (1995). Antecedents and consequences of physical activity and exercise among older adults. The Gerontologist, 35,