| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
RESEARCH ARTICLE |
a Department of Sociology and Gerontology Program, Purdue University, West Lafayette, Indiana
Kenneth F. Ferraro, Stone Hall, Purdue University, West Lafayette, IN 47907-1365 E-mail: ferraro{at}purdue.edu.
 |
Abstract |
|---|
 TOP Abstract Self-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
Methods. This study reexamines the prognostic value of self-ratings of health on mortality with data from 20 years of the National Health and Nutrition Examination Survey–I Epidemiologic Follow-up Study (N = 6,833). Special attention is given to differences between White and African Americans.
Results. Results indicate that event history models of mortality with self-rated health treated as a time-dependent covariate are superior to those treating it as a baseline predictor only—the latter are likely to underestimate the effect. Moreover, self-ratings of health predict mortality for African Americans only when treated as a time-dependent covariate.
Discussion. The results suggest that self-ratings of health are sensitive to declines in physical health, especially those associated with terminal drop. The analysis also demonstrates the importance of using dynamic models for studying the link between self-rated health and mortality if data from multiple observation points are available.
Idler and Benyamini 1997
thoughtfully reviewed the burgeoning literature examining the link between self-rated (or self-assessed) health and mortality. They noted that interest in the subject was spurred by research on the validity of lay reports of health information as well as the influence of psychosocial factors on health and longevity. They identified more than 45 studies of the topic (
Benyamini and Idler 1999), and the findings are, in their words, "impressively consistent." Self-rated health is an independent predictor of mortality in most studies, despite numerous control variables, and the link appears especially strong for men. Idler and Benyamini issued a call for a new generation of studies on the topic to examine more carefully four major interpretations for why self-rated health may predict mortality.
The present study focuses on one of the four interpretations identified, namely that "self-rated health is a dynamic evaluation, judging trajectory and not only current level of health" (
Idler and Benyamini 1997, p. 29). There is some evidence in the literature to support this interpretation, but much of the extant literature relies on baseline measures of self-rated health used to predict mortality at follow-up. Although the modeling of mortality has been exemplary in many of these studies, there are very few studies that examine how change in self-rated health is related to mortality. This would appear to be a logical next step for studies of self-rated health and mortality, especially to determine the plausibility of the mechanism noted above (i.e., self-rated health is a dynamic evaluation). The present study uses data from a multiwave longitudinal national sample to compare static and dynamic models of the relationship between self-rated health and mortality. Extending findings on racial differences in the validity of lay information, special attention is also given to determining whether the predictive validity of self-rated health on mortality differs for White and African Americans.
|
Self-Rated Health as a Dynamic Evaluation |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
). Although the standard question for self-rated health does not ask the person to account explicitly for changes over time, it appears that respondents consider health trajectories in rendering their ratings. It has been shown in several studies that change in morbidity and disability is related to change in health ratings (e.g.,
Ferraro, Farmer, and Wybraniec 1997), thereby suggesting that self-rated health reflects, in part, the overall pattern of one's health trajectory.
As noted above, most of the literature on the relationship between self-rated health and mortality examines how baseline health ratings influence mortality (e.g.,
Borawski, Kinney, and Kahana 1996;
Bosworth, Siegler, et al. 1999;
Franks, Gold, and Clancy 1996;
Hays, Schoenfeld, Blazer, and Gold 1996;
Idler and Angel 1990;
Idler and Kasl 1991;
Idler, Kasl, and Lemke 1990;
Kaplan and Camacho 1983;
Mossey and Shapiro 1982;
Schoenfeld, Malmrose, Blazer, Gold, and Seeman 1994;
Wolinsky and Johnson 1992;
Wolinsky and Tierney 1998). To study more carefully the dynamic evaluation thesis, however, one needs repeated measures of health status, including self-rated health, to determine if change in them is linked to mortality. A small number of studies make use of any repeated measures, and these studies fall into one of two groups. First, a few studies have examined change in various health status measures, except self-rated health, as predictors of mortality.
Thomas, Kelman, Kennedy, Ahn, and Yang 1992 examined changes in activities of daily living (ADLs) and health conditions, but baseline self-rated health remained significant after accounting for such changes. By contrast, Wolinsky and colleagues found that accounting for declines in basic ADLs and lower body function rendered baseline self-rated health nonsignificant in predicting mortality among respondents of the Longitudinal Study on Aging (
Wolinsky, Callahan, Fitzgerald, and Johnson 1993;
Wolinsky, Callahan, and Johnson 1994).
Second, we identified only two studies that examined whether change in self-rated health was associated with mortality risk. These studies required a minimum of two interviews querying self-rated health followed by a mortality tracing.
Svardsudd and Tibblin 1990 studied Swedish men with such a design (i.e., a 7-year re-interview and an 8-year subsequent mortality follow-up). They found that declines in self-rated health were associated with higher mortality.
Strawbridge and Wallhagen 1999 examined survival over 28 years in the 4-wave Alameda County, California, study and reported that self-rated health remained a significant predictor of mortality even after accounting for the most recent information on health conditions.
Taken together, these two groups of studies point to a need for further research on how change in self-rated health may be related to mortality. They supply the most solid evidence to date regarding the dynamic evaluation thesis, but can be logically extended in many ways, two of which are considered in this research: (1) method of analysis and (2) consideration of minority differences.
|
Modeling Mortality with Multiwave Data |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
;
Cox 1972). The universe of models implementing time-dependent covariates is large, but we identified only one article that reported such an analysis (
Strawbridge and Wallhagen 1999). Failure to account for changes in self-rated health in previous research may lead to underestimates of the relationship between self-rated health and mortality. Consider the scenario for a 4-wave panel study in which participants rated their health as good at the beginning of the survey, fair at wave 2, poor at wave 3, and died before the final wave. (Indeed, eight respondents to the National Health and Nutrition Examination Survey–I Epidemiologic Follow-up Study (NHEFS) fit this exact scenario with scores of 3, 2, and 1 on self-rated health over the first three waves.) With baseline predictors only, these respondents would be assumed to be in good health throughout the survey until death. The use of time-dependent covariates would enable one to account for changes over the survey period. In order to highlight the influence of different types of analyses with multiwave data, we compared results from models that do not incorporate time-dependent covariates with those that do.
Models that reflect the timing of changes in self-rated health are important to consider when testing the dynamic evaluation thesis. If mortality is perceived as a temporally proximal outcome of a health decline, one would expect that health ratings would reflect such concern. In essence, persons may believe that the pace of their health decline is raising the risk of mortality, or even that they are dying. Gerontologists have long expressed interest in the concept of terminal drop (sometimes referred to as terminal decline or terminal change).
Kleemeier 1962 introduced the concept of terminal drop by studying the relationship between test performance and survival. He observed declines in test scores of men on several occasions over the course of 12 years; however, what was most striking was that the decline was much greater for those who died after one of the data collection periods as compared to those who survived during that same interval. He observed that declines in intellectual performance may be detected "several years prior to the death of the person" (p. 293, emphasis added).
Terminal drop indicates that there is a determinant chain of functional changes that are due to a death process (
Berg 1987;
Botwinick 1977;
Palmore and Cleveland 1976;
Riegel and Riegel 1972;
Siegler 1974;
White and Cunningham 1988). If self-rated health reflects a dynamic evaluation of health, then it should decline sharply in the period prior to death, and more dynamic models of self-rated health may uncover this phenomenon. Of course, how early such changes may be observed is an empirical question. In the literature examining terminal decline in cognitive functioning, a few studies report declines as early as 7 years or more prior to death (
Bosworth, Schaie, and Willis 1999;
Johansson and Berg 1989). The bulk of the studies, however, examine time windows of 1 to 5 years and show declines in selected features of cognitive functioning (
Jarvik and Blum 1971;
Johansson and Zarit 1997;
Lieberman 1965;
Lieberman and Coplan 1970;
Small and Backman 1997;
White and Cunningham 1988;
Wilkie and Eisdorfer 1974). Our interest in the dynamic evaluation hypothesis is focused on whether people might identify preclinical symptoms suggesting health problems. Therefore, the relevant time window is much longer than what people consider in cases of "failure to thrive" or "dwindling" in clinical settings (
Egbert 1996). Also, key to our concerns is a recognition that the person is, to some degree, aware of the change in their health.
|
Self-Rated Health Among Minority Populations |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
conclude their review of studies of self-rated health and mortality by offering four suggestions for the "next stage of studies." Two of their recommendations focus on studying "special populations" and "cognitive/cultural processes." These are superb suggestions, but we would like to add to the list a greater consideration of minority health and aging. Ethnic minority group members can be considered a special population, and cultural factors are certainly a part of ethnicity. Nevertheless, very little attention has been given to ethnic or racial differences in the predictive validity of self-ratings of health on mortality. The literature includes a number of studies of populations with limited ethnic or racial variation, either in the United States (
Borawski et al. 1996) or other countries (
Shahtahmasebi, Davies, and Wenger 1992); some of this is probably due to the composition of the population. Nevertheless, there are numerous studies of the United States or other diverse societies that do not give systematic consideration to ethnic or racial differences (
Deeg et al. 1989;
Kaplan, Barell, and Lusky 1988;
Kaplan and Camacho 1983;
LaRue, Bank, Jarvik, and Hetland 1979).
Given the substantial health inequality in American society by racial and ethnic groupings, one would expect more interest in the topic. Dozens of studies of morbidity, disability, and health ratings across the life course reveal that African and Hispanic Americans have poorer health (
American Medical Association 1991;
Ferraro and Farmer 1996;
Hummer 1996). Many members of ethnic minorities have less frequent or delayed contacts with physicians (
Himmelstein and Woolhandler 1995;
Wolinsky et al. 1989), suggesting that they may also be less likely to have a definitive diagnosis about conditions that bother them (and thus present to a physician at a more advanced illness stage). Indeed, some research shows that African Americans exaggerate their subjective probability of survival (
Hurd and McGarry 1995). Thus, one wonders if self-ratings of health have less predictive validity on mortality for minority group members. In a study comparing White and African Americans,
Ferraro and Farmer 1999 found that self-reports of morbidity were equally predictive of mortality for both groups, but whether the same holds true for self-rated health remains to be seen. Many studies do not even include a control for racial or ethnic variation (e.g.,
Thomas et al. 1992); of those that do, none have examined whether change in self-rated health is predictive of mortality. In short, we are unaware of any studies that compare the predictive utility of self-ratings of health on mortality across U.S. ethnic groups. The present analysis is designed as a first step to fill this gap by examining the predictive validity of self-rated health over time on the mortality of White and African American respondents. The basic research questions addressed are twofold:
|
Methods |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
;
National Center for Health Statistics 1987). Parts of this analysis focus on differences in outcomes based on race. At the baseline, respondents were asked to specify their race (White, "Negro," or Other) only if the answer could not be visually determined by the interviewer. Due to the small number of other racial and ethnic groups (i.e., 41 Asian Americans), further subsample analyses with these groups were not possible. The present study analyzes 5,955 White and 878 Black respondents at the baseline survey who were followed across time. Response rates among those who received the detailed components were very high (86% in 1982–1984, 87% in 1987, and 85% in 1992). The number of accumulated deaths over the approximate 20-year period is 1,938, or 28% of the baseline sample. Each survey is referred to as a wave (W); the first wave is considered W1 (or baseline).
Measures
The measurement of self-rated health in NHEFS was a single question asked at each wave: "In general, is your health excellent, very good, good, fair, or poor?" Excellent health is scored 5 and poor health is scored 1. Although some researchers have found it useful to examine contrasts between specific categories of self-rated health (or aggregates of those categories), the linear form of self-rated health was used here for two principal reasons. First, we desired to avoid the coarseness involved by collapsing the five responses into fewer categories. Second, treating self-rated health as a time-dependent covariate is far more parsimonious when the original form is retained.
In order to assure that the relationship between self-rated health and mortality is not spurious, most investigators have included a broad array of independent variables including morbidity and status characteristics and resources. Morbidity is measured with two variables, both counts of diagnosed health conditions. The National Health and Nutrition Examination Survey–I (NHANES–I) differs from many other surveys by asking if the respondent had been diagnosed with a particular health condition rather than relying on respondent opinion. Respondents were asked, "Has a doctor ever told you that you have ... hypertension or high blood pressure?" Each condition was first coded as a binary variable (1 equals condition is present). Next, separate sums of serious and chronic nonserious conditions were created (
Ferraro and Farmer 1999;
Small and Backman 1997). Serious conditions, those that are life threatening, include cancer, diabetes, heart failure (attack or trouble), hypertension, and stroke. Chronic nonserious conditions asked at all waves include arthritis, cataracts, hip fracture, and kidney trouble.
Several measures of health behavior were included in the analysis. First, respondents were asked two questions regarding their amount of daily physical activity: "In things you do for recreation, for example, sports, hiking, dancing, etc., do you get much exercise, moderate exercise, little, or no exercise?" and "In your usual day, aside from recreation, are you physically very active, moderately active, or quite inactive?" Subjects who answered "little or no exercise" on the first question and "quite inactive" on the second were considered quite inactive and coded 1 for restricted activity; all others were coded 0. Second, survey respondents were asked if they had a regular physician for their health needs (1 = yes, 0 = no). Finally, current smokers and past smokers of cigarettes, cigars, or pipes were separated out in binary variables, 1 indicating yes and 0 indicating no.
Indicators of socioeconomic status include income (ranges from 1 to 12, with 12 equal to $25,000) and education (ranges from 0 to 7, with 7 equal to graduate school). Those with private medical insurance and recipients of Medicaid are indicated by binary variables (1 equals yes).
The measurement of the other covariates is fairly straightforward. Female, Black, lives alone, widowed, and lives in a rural community are all dichotomous where 1 equals the name of the variable. Respondents with a body mass index (BMI) equal to or greater than 30 are considered obese and coded 1 (all others are coded 0). Given the nonlinear relationship between BMI and many health outcomes, underweight respondents—BMI less than 18.5—were also identified and coded 1 (all others are coded 0). In order to account for the possible influence of psychological distress on both self-rated health and mortality, a self-rating of "strain, stress, or pressure in the last month" was also included (
Farmer and Ferraro 1997). Ages ranged from 25 to 74 at baseline. Several additional variables were considered in preliminary analyses, but deleted from subsequent consideration after they did not manifest a significant relationship with mortality in multivariate models (e.g., South, overweight [BMI 
25 and < 30]). Unfortunately, measures of social support or social networks are conspicuously missing in NHEFS. Item-missing data were modest (i.e., <5%) and were handled via specific-mean imputation by age (<45, 45–64, 65+), sex, and race categories (
Little and Schenker 1995).
Vital status was determined at the follow-up surveys for all traced respondents and confirmed by death certificates. Brief interviews were conducted with proxies of the deceased respondents. Date of death was obtained for 1,935 decedents, so continuous-time event history models were applied. (Only all-cause mortality is considered; disease-specific cause of death, from death certificates, is not used.) For additional analyses related to identifying the time window for the dynamic evaluation thesis, four categories of vital status were created based on the timing of death relative to each survey: died within 12 months of the survey; died between 13 and 24 months of the survey; died between 25 months after the survey and the time of the next survey; and survived. This process was repeated for waves 2 and 3. For analyses for each wave, binary variables differentiating the categories were tested against a reference group of those who survived to the next survey wave. Mortality data are not available past the fourth wave.
Analytic Plan
Although our preferred analytic strategy is to apply time-dependent covariates in proportional hazards models, the analyses were completed with additional strategies to compare findings. First, the analysis began with event history models specifying baseline self-rated health as predictive of mortality—parallel to the common approach in the recent literature (e.g.,
Bernard et al. 1997). Second, we specified several types of models with time-dependent covariates (
Allison 1995;
Fisher and Lin 1999).
There are many different ways to implement time-dependent covariates, but the simplest uses the most recent value available. If no changes are reported, baseline information is used. If change is reported, the value reported at the most recent observation is used. A second type of model uses the most recent values but scales them by time of observation (
Allison 1995, pp. 140–147). In essence, a time stamp is placed on the most recent value. Although no information on initial values is used for those persons who changed their rating, the time of the most recent value is incorporated in the analysis. Knowing when the most recent value was observed is a critical piece of information when there are three or more data collection points.
A third type of time-dependent covariate analysis builds upon the preceding approach but incorporates change in the covariate. Both the initial value and all transitions in the covariate are used to estimate the partial-likelihood function. This approach may be described as time-indexed change because three pieces of information regarding change in the covariate are parsimoniously incorporated into the partial-likelihood estimation (
Allison 1995, pp. 150–152). Time-indexed change captures the magnitude (i.e., unit change in self-rated health), direction (i.e., increase or decrease), and timing of the transition(s). If the dynamic evaluation thesis is correct, self-rated health should remain a significant predictor in these models. Indeed, one may expect that its association with mortality will be stronger.
We estimated selected analyses with all three approaches and compared them to the results predicated on baseline values only. All of the models with time-dependent covariates yielded the same substantive conclusions, but the magnitude of effects varied slightly. Whereas model fit was superior for the time-indexed change analysis, those results will be presented in the following section and compared to those from models using baseline information only.
As noted above, the time between survey waves is fairly long, leading to the possibility of coarseness in the modeling of mortality. Especially for the first follow-up, self-ratings of health may have declined during the 10 years, but this was not measured among respondents who died. The likely effect of this would be an underestimate of the association of self-rated health on mortality. In order to test for potential coarseness, we estimated what
Allison 1995 refers to as ad hoc models of time-dependent covariates. We incorporated a time stamp to estimate changes in self-rated health using the baseline characteristics and follow-up information when available. Two strategies for these regression imputations were used: (a) imputing integer values of self-rated health, and (b) imputing noninteger values. Both methods yielded results that were very similar to our findings, suggesting that potential bias due to the coarseness of measurement intervals is minimal.
To summarize, our analysis proceeded in three stages. First, we estimated the models described here to compare the results. Second, we used the time-indexed change models to incorporate more than one time-dependent covariate. Consistent with our interest in minority aging, these models were estimated on the total and race-specific subsamples. Finally, we examined the timing of declines in self-rated health on mortality with retrospective analyses. Self-rated health at each of the first three waves of the study was regressed on vital status after the respective survey. The timing of death was used to delineate categories of deceased subjects anticipating that self-rated health would be lower for those dying shortly after the survey.
|
Results |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
|
The next set of equations (last 3 columns) presents the results from the analysis where self-rated health is treated as a time-dependent covariate. For the total sample, self-rated health is again significant, but the effect is much stronger when treated as a time-dependent covariate, thereby providing evidence for the dynamic evaluation thesis. The pattern of significant findings is almost identical to what was observed in the first equation, where only baseline self-rated health was considered, and the effects of most variables are similar. One notable difference between the first and fourth equations is the effect of distress. Although distress predicted mortality in the first equation, treating self-rated health as a time-dependent covariate rendered the effect due to distress nonsignificant. This is an important observation because many of the previous studies reporting a link between distress or depression and mortality may have been reflecting concern about unmeasured declining health.
In the equation for the African American subsample, treating self-rated health as a time-dependent covariate leads to different results from what was seen in the earlier model. Self-rated health, including change in it, is strongly linked to mortality risk, although such a conclusion could not be derived from the baseline predictor model. This means that African Americans, just as their White counterparts, adjust their health ratings in such a way that is predictive of mortality. While the effects of many of the covariates on mortality are similar to those in the baseline model, the magnitude of effects for several variables changes. Many of the effects are reduced slightly, but some are actually greater once accounting for self-rated health as a time-dependent covariate (probably reflecting suppression). For example, the effect due to serious illness is larger in the model with self-rated health treated as a time-dependent covariate. As anticipated, the results for the White sample are quite similar to the results for the total sample. Given the differences observed across the Black and White subsamples, a number of interactions with race were tested (i.e., age, self-reported health), but none were significant.
The results thus far reveal substantial support for the idea that self-rated health is a dynamic evaluation; it reflects trajectories at least as they pertain to mortality risk. This is an important finding, but it is possible that incident morbidity during the study could be at least partly responsible for the greater mortality risk. While
Fisher and Lin 1999 assert that "the use of time-dependent covariates offers exciting opportunities for exploring associations and potentially causal mechanisms" (p. 146), they also argue that these analyses must be approached with caution. In particular, they note that there is the "possibility of too much modeling and overfitting of a data set" (p. 152). With two measures for morbidity (serious and chronic, nonserious), we followed their advice and treated one morbidity variable as time dependent at a time. (We also tested for several other significant time-dependent covariates—one at a time—but found none.) The results, presented in Table 2 , compare models with and without self-rated health as time dependent, but where morbidity is time dependent throughout.
|
The race-specific results in Table 2 show again that self-rated health is a significant predictor of mortality for African Americans only when it is treated as a time-dependent covariate. Neither of the morbidity variables is significant for the Black subsample in Table 2 . For the White subsample, mortality risk is shaped by self-rated health in both models (used as a baseline or time-dependent predictor), but the magnitude of the relationship is reduced compared to Table 1 results. In short, incident morbidity during the study was partly responsible for the greater mortality risk associated with changing self-rated health. Nevertheless, for both White and African Americans, the link between changing self-rated health and mortality persists after controlling for changing morbidity.
Given the differences in the number of cases in the White and Black subsamples, supplementary analyses were undertaken to assure that the differences between the subsamples are not simply the result of the larger subsample of White respondents. First, the equation for the White subsample was reestimated on a random sample of 814 respondents in order to match the frequency of the Black subsample. The second set of analyses involved taking a subsample of White respondents so that the number of deceased persons equaled those deceased in the Black subsample. Given the lower mortality of White respondents, the first approach—matching the number of Black cases—yields fewer deaths in the White subsample. Therefore, we also reestimated the models on a slightly larger sample of White respondents (n = 1,179) so that the number of deaths (outcome variable) was equal across Black and White respondents. This set of analyses confirmed all of the significant relationships reported with the total White sample. Although the effects of some of the covariates were slightly different in the two approaches, the findings and conclusions regarding self-rated health and well-known risk factors (e.g., serious illness, female, age) were consistent across both of these sets of estimates and the results presented.
The final step in our analysis is designed to clarify the changes in self-rated health that occur as a result of approaching death and to provide some insight into the timing of the dynamic evaluation thesis. Self-rated health at the first three waves of the study was regressed on four categories of vital status based on the timing of death relative to each survey wave: died within 12 months, died between 13 and 24 months, died between 25 months and the time of the next survey, and survivors. The logic of these analyses is to exploit the longitudinal data by "retroactively" considering how proximity to death may have influenced self-rated health. Fig. 1 displays the mean levels of self-rated health for each of the four vital status categories at each wave. It appears that there is a terminal drop phenomenon on self-rated health for those respondents who died within 1 year of the survey. The effect is strongest for the most proximate deaths (within 1 year after being interviewed), but remains substantial for all of the decedents between survey waves. People adjust their health ratings to reflect the health problems experienced in the last few years of their lives.
|
|
|
Discussion |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
First, we sought to systematically examine
Idler and Benyamini 1997 thesis that self-rated health is a dynamic evaluation of health: Does self-rated health reflect a person's health trajectory? While it seems reasonable to assert that self-rated health reflects a dynamic evaluation of health status, there have been very few studies that systematically examine such a thesis. There have been dozens of studies that examined whether baseline self-rated health is related to mortality, and one that examined whether the expectation of declining health was independently related to mortality (
Wolinsky and Tierney 1998). A few studies have also examined how self-rated health affects the probabilistic assessment of life expectancy (
Hurd, McFadden, and Gan 1998;
Hurd and McGarry 1995). We are aware of only two studies, however, that use repeated measures of self-rated health to examine how change in the ratings is related to mortality (
Strawbridge and Wallhagen 1999;
Svardsudd and Tibblin 1990). Unlike those studies, the present research used a nationally representative sample while treating self-rated health as a time-dependent covariate. It also tested explicitly for differences in the effect of self-rated health on mortality for Whites and African Americans. We hope to see more studies that will examine these relationships with different samples and different time lags.
The findings from the NHEFS offer some of the most systematic evidence to date that self-rated health reflects health trajectories, especially the steep declines associated with death after the last survey measurement. Not only was baseline self-rated health predictive of mortality for NHEFS respondents, but models of change in self-rated health showed a stronger relationship with mortality risk. One conclusion from this study, therefore, is that self-rated health is probably a stronger predictor of mortality than has been reported in many previous studies. Many of the previous studies that relied on baseline self-rated health to predict mortality had fairly long lag times (e.g., 28 years,
Deeg et al. 1989). The fact that baseline self-rated health is an independent predictor of mortality with such long lag times is an important finding, because longer lag times work against finding such an effect (or underestimating such an effect). Accounting for more recent health ratings among the NHEFS survivors shows that the association between self-rated health and mortality is even stronger than estimates derived using baseline self-rated health alone. Moreover, estimating mortality with both morbidity and self-rated health treated as time-dependent covariates accounts not only for the more recent ratings, but also the transitions in health ratings from wave to wave among those who survive.
Another way to consider whether self-rated health is a dynamic evaluation is whether terminal drop reflects itself on self-rated health. That is, if self-rated health reflects a person's health trajectory, one would expect that temporal proximity to death would be associated with the steepest declines in self-rated health. In essence, evidence for steep declines in self-rated health associated with mortality risk would suggest that people are, to some degree, aware of terminal drop and adjust their health ratings accordingly.
Evidence from the NHEFS shows that declines in self-rated health precede most deaths within a 1-year follow-up period. The models presented in Table 3 for W2 and W3 are residualized change analyses, and in both equations persons who died within 12 months after the previous interview showed a decline in self-rated health. It is clear from these analyses that people adjust their health ratings to reflect their mortality risk, and that change in self-rated health associated with terminal drop is not solely a function of changing morbidity. We conclude that
Idler and Benyamini 1997 were quite correct in suggesting from previous research that self-ratings of health "reflect a dynamic, rather than static, perspective on health" (p. 29). While their suggestion was based on a thoughtful reading of a literature that was dominated by fairly static models, the findings from the present investigation based on more dynamic models offer some of the most compelling evidence that this is the case.
The second major contribution of the present research is to give greater attention to minority health ratings in the prediction of mortality (and exhort others to do likewise in future research). In the most simple of terms, one wonders whether self-rated health "works" as a predictor of mortality for minority groups. Many minority groups, especially African Americans studied in detail here, are health disadvantaged and limited in various ways to quality medical care (
Himmelstein and Woolhandler 1995;
Wolinsky et al. 1989). Thus, the question of whether African American self-rated health operates as a predictor has important policy implications. Do African Americans interpret and respond to a question about their self-rated health in the same way that White Americans do? Do Black persons understand that they are approaching their death? Given the concerns about early detection of disease, are they aware of their mortality risk due to incident disease?
The answers to these questions will differ depending upon the mode of analysis. Findings from the NHEFS that rely on baseline self-rated health show that self-rated health does not predict mortality risk. By contrast, models incorporating self-rated health as a time-dependent covariate show that they do. Three conclusions are inescapable. First, one would reach absolutely contradictory results depending upon model specification. Second, models incorporating time-dependent covariates are superior inasmuch as they account for transitions in self-rated health among survivors. Third, African Americans are quite perceptive of their health declines, including the rapid declines associated with terminal drop. Moreover, even adding morbidity as a time-dependent covariate did not negate the effect of self-rated health as a predictor of mortality. Self-ratings of health work just as well for African Americans as they do for Whites if the models incorporate change in health ratings.
We believe that this study is only a first step in what
Idler and Benyamini 1997 call "the next stage of studies." There are several lines of research that we believe merit further and more detailed examination. First, the present investigation has systematically examined one minority group only. Whether the findings extend to Hispanic, Native, or Asian Americans is not known. Second, we agree with
Idler and Benyamini 1997 call for pursuing qualitative approaches to the subject of health ratings and mortality. Several of these studies have shed important light on the topic (e.g.,
Krause and Jay 1994), and we believe that ethnographic analyses of the health trajectory thesis could be an important contribution. Third, we disagree with
Idler and Benyamini 1997 assertion that "researchers are fast approaching, if not already reaching, the limits of what secondary analysis of these large, longitudinal data sets can tell us about the relationship between self-ratings of health and mortality. There is a point when replication becomes redundancy" (p. 31).
While we agree that isomorphic replications of previous studies may not yield much in the way of scientific or health policy value, we disagree that secondary analysis is necessarily "approaching something of an impasse in understanding" (p. 34). Rather, we assert that the rich multiwave data sets that include repeated measures of self-rated health (and other health variables) open up a new vista of studies that can incorporate more dynamic models of self-rated health and mortality. Those types of analyses may help us to better understand
Idler and Benyamini 1997 assertion that self-rated health involves the assessment of one's health trajectory and not only the current level of health. There are other issues, such as a better understanding of minority health ratings, that can be gleaned from such quantitative analyses of large data sets. A number of such large prospective studies are underway, and we hope that investigations comparing various ethnic or racial groups will ensue.
Whichever direction the literature on self-rated health and mortality flows, the findings from the present analysis provide systematic evidence for the assertion that health ratings reflect a dynamic perspective on health. People adjust their ratings of health in a way that is associated with mortality risk, and this phenomenon is clearly observed in the year prior to death reflective of terminal drop. The changes in health ratings can be observed for both White and African Americans only when dynamic models of self-rated health and mortality are estimated.
|
Acknowledgments |
|---|
Received for publication February 22, 2000. Accepted for publication December 19, 2000.
|
References |
|---|
|
TOPAbstractSelf-Rated Health as a...Modeling Mortality with...Self-Rated Health Among Minority...MethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  J. B. Dowd and A. Zajacova Does the predictive power of self-rated health for subsequent mortality risk vary by socioeconomic status in the US? Int. J. Epidemiol., December 1, 2007; 36(6): 1214 - 1221. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 L. Winter, M. P. Lawton, C. A. Langston, K. Ruckdeschel, and R. Sando Symptoms, Affects, and Self-Rated Health: Evidence for a Subjective Trajectory of Health J Aging Health, June 1, 2007; 19(3): 453 - 469. [Abstract] [PDF]
|
|
|
|
 |
|
 S. A. Robert and E. Ruel Racial segregation and health disparities between black and white older adults. J. Gerontol. B. Psychol. Sci. Soc. Sci., July 1, 2006; 61(4): S203 - S211. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. L. Angel and R. J. Angel Minority Group Status and Healthful Aging: Social Structure Still Matters Am J Public Health, July 1, 2006; 96(7): 1152 - 1159. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. Ayis, R. Gooberman-Hill, A. Bowling, and S. Ebrahim Predicting catastrophic decline in mobility among older people Age Ageing, July 1, 2006; 35(4): 382 - 387. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. M. Irving and K. F. Ferraro Reports of Abusive Experiences During Childhood and Adult Health Ratings: Personal Control as a Pathway? J Aging Health, June 1, 2006; 18(3): 458 - 485. [Abstract] [PDF]
|
|
|
|
 |
|
 S. D. Barger Do Psychological Characteristics Explain Socioeconomic Stratification of Self-rated Health? J Health Psychol, January 1, 2006; 11(1): 21 - 35. [Abstract] [PDF]
|
|
|
|
 |
|
 L. W. Li Predictors of ADL Disability Trajectories Among Low-Income Frail Elders in the Community Research on Aging, November 1, 2005; 27(6): 615 - 642. [Abstract] [PDF]
|
|
|
|
|
|
P. Svedberg, M. Gatz, P. Lichtenstein, S. Sandin, and N. L. Pedersen Self-Rated Health in a Longitudinal Perspective: A 9-Year Follow-Up Twin Study J. Gerontol. B. Psychol. Sci. Soc. Sci., November 1, 2005; 60(6): S331 - S340. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. F. Ferraro and J. A. Kelley-Moore A Half Century of Longitudinal Methods in Social Gerontology: Evidence of Change in the Journal J. Gerontol. B. Psychol. Sci. Soc. Sci., September 1, 2003; 58(5): S264 - 270. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 Y. Benyamini, T. Blumstein, A. Lusky, and B. Modan Gender Differences in the Self-Rated Health-Mortality Association: Is It Poor Self-Rated Health That Predicts Mortality or Excellent Self-Rated Health That Predicts Survival? Gerontologist, June 1, 2003; 43(3): 396 - 405. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
