| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
RESEARCH ARTICLE |
a School of Public Health, University of Michigan, Ann Arbor
b Institute of Gerontology, University of Michigan, Ann Arbor
c Institute for Social Research, University of Michigan, Ann Arbor
d Tokyo Metropolitan Institute of Gerontology, Japan
e University of Tokyo, Japan
Jersey Liang, Department of Health Management and Policy, The University of Michigan School of Public Health, 109 S. Observatory, Ann Arbor, MI 48109-2029 E-mail: jliang{at}umich.edu.
Decision Editor: Fredric D. Wolinsky, PhD
 |
Abstract |
|---|
 TOP Abstract SES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
Methods. Data came from a 5-wave panel study of a national probability sample of 2,200 elderly Japanese conducted between 1987 and 1999. Hazard rate models involving time-varying covariates were used to ascertain the direct and indirect effects of SES. In addition, interaction effects involving SES variables with age and gender were evaluated.
Results. In contrast to prior findings from the Western developed nations, there is an educational crossover effect on mortality among older men, in that, at advanced age, those with less education live longer than those with higher education. On the other hand, there is some evidence that educational differences in the risk of dying tend to converge in the 70–79 age group. More interestingly, there is a crossover in the effect of education among the 80 and older age group.
Discussion. The observation that educational crossover exists only among elderly men may be because of gender and SES differences in causes of death, morbidity, and health behavior. On the other hand, possible explanations for age differences in the educational crossover include selective survival and cohort effects.
THERE is a well-established inverse gradient between socioeconomic status (SES) and mortality (Robert and House 2000
; Rogers, Hummer, and Nam 2000). At the same time, it is now common to conceptualize SES, age, and gender as distinct dimensions of social stratification (Lengermann and Niebrugge 1996; Riley 1987). Consequently, there is a growing interest in the interaction between the effects of SES on mortality and those of other dimensions of social status. This research examines gender and age differences in the effects of SES on mortality in a national sample of older adults over a 12-year period in Japan.
|
SES and Gender Interaction Effects |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
; Robert and House 2000). This conclusion has been supported by data from the United States, Canada, France, Hungary, England and Wales, and the Nordic nations (Elo and Preston 1996; Nathanson 1990; Valkonen 1989). However, it is not supported by several other studies (Christenson and Johnson 1995; Kitagawa and Hauser 1973; Liang et al. 2000).
These rather diverse findings are because of at least two reasons. The first is the uncontrolled heterogeneity. In many studies, important intervening variables, such as baseline health status, are often not taken into account (Christenson and Johnson 1995; Elo and Preston 1996; Martelin 1994; McDonough, Williams, House, and Duncan 1999). Without controlling for important confounding variables, the linkage between gender differences in SES effects on mortality would be difficult to evaluate and interpret. Second, it has long been recognized that gender difference in life expectancy varies across age spectrum, time period, and place (Nathanson 1990; Ostlin, George, and Sen 2001). For instance, gender differences in mortality tend to be high in early adult years, but much lower at either end of the age spectrum. At the same time, there is a steady widening of sex mortality differentials in the developed nations over the course of the 20th century (Lopez 1983; Nathanson 1990). In addition, in countries where social discrimination against women is less pervasive, women tend to outlive men. In societies where great female deprivation exists, women's mortality rate is higher or equal to that of men (Hemstrom 1998; Lopez 1983).
The variations across age, time, and place suggest that social factors have a significant influence on gender differences in survival (Nathanson 1990; Ostlin et al. 2001). Existing studies of gender and SES interaction effects on health are based on data from adult populations in general (e.g., Koskinen and Martelin 1994). There is very limited understanding concerning these interaction effects within the elderly population. Furthermore, the vast majority of studies are based on data from Western nations. Gender disparities in socioeconomic well-being in Western countries may differ substantially from those in non-Western societies because gender role is often reinforced by the educational system, employment policy, and family policy in determining the allocation of resources (Brinton 1989). For instance, among developed countries, Japan is a persistent outlier in terms of women's status (Brinton 1989). In Japan, the male-female wage gap is greater in that salaries among full-time female employees are 63% of their male counterparts. At the same time, women are much more likely to be unpaid family workers in small family-run businesses or farms. Furthermore, women's representation in managerial rank is lower with less than 10% in 1997 (Ministry of Foreign Affairs of Japan 2001). Although high school seems to have become the minimum acceptable level of education for both sexes in Japan, men are far more likely to go on to 4-year universities than are women. The gap has narrowed considerably, but even in 1997, 26% of girls and 43% of boys went to universities (Ministry of Foreign Affairs of Japan 2001).
|
SES and Age Interaction Effects |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
). This is because of the greater exposure of lower SES persons to a wide range of psychosocial risk factors. Health eventually converges across different socioeconomic strata because people inevitably weaken and die in old age regardless of social class (Elo and Preston 1996; House, Kessler, and Herzog 1990). In contrast, Ross and Wu 1996 advocate a hypothesis of cumulative advantage, because they find no evidence of convergence in health among different education and income levels in old age.
Whereas prior studies contribute significantly to our understanding of health changes across the adult life course, two important issues remain unresolved. First, current studies in general treat the elderly population as a whole without differentiating young-old from older-old and the oldest-old. In particular, does SES exert the same effect at 60 as well as 80? Does the SES effect on health continue to increase in old age or converge? In addition, the number of elderly subjects included in some studies is quite modest (Ross and Wu 1996). Second, theories of convergence and cumulative advantage imply changing interindividual differences over the adult life span. This calls for repeated observations of the same individuals over a very extended period of time, if not an entire life span. However, current studies are based on either cross-sectional data or 2-wave short-term panel studies of the adult population (House et al. 1994; Ross and Wu 1996). These hypotheses have not been evaluated by prospective studies over an extended period of time.
Third, observations concerning age and SES interaction effects on health are almost exclusively derived from Western developed nations, particularly the United States. Their external validity needs to be carefully evaluated, because the linkages between a given disease and SES may vary across different societal and cultural contexts. For example, Western Europe, North America, and Japan differ significantly with reference to the incidence of coronary heart disease (CHD). In Western countries, typically there has been a sharp increase in CHD mortality in parallel with the rapid increase in prosperity after World War II, followed by sharp falls as economic development has continued. Despite industrialization and supposed Westernization, Japan has escaped the CHD epidemic and has had among the lowest rates of heart disease of any developed nation (Marmot and Davey Smith 1989; Marmot and Mustard 1994). Finally, breast cancer mortality used to be higher among higher SES women because of higher incidence in this group. However, now among White American women, this association has largely disappeared (Heck, Wagner, Schatzkins, Devesa, and Breen 1997).
|
Old Age Mortality in Japan |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
).
The Japanese have the highest life expectancy at birth (77 years for males and 83.6 years for females in 1995) in the world, and can live 3 to 6 years longer than Americans (72.7 years for males and 79.4 years for females in 1995; OECD 1998). In terms of the leading causes of death, Japan has significantly lower rates of death from cancer and circulatory system disease than the United States. What is so remarkable is that the high life expectancy in Japan was only accomplished during the last 4 decades (Evans, Barer, and Marmor 1994).
In terms of social stratification, Japan has three distinct characteristics. First, there is less income inequality in Japan than in other market economies. According to the standard deviation calculated from percentage share of income distribution by quintile, Japan ranks 14th (11.14), whereas the United States ranks 40th (14.37) in equality in income distribution among 85 nations (World Bank 2001). Second, there is a great emphasis on educational credentials and intense competition for higher education. However, family background (i.e., parental income, education, and occupation) exerts a greater effect on educational attainment in Japan than in the United States and Britain (Ishida 1993). Third, Japan has a large middle class whose members are highly homogeneous in their attitudes and lifestyles (Kosaka 1994).
The traditional support of elderly people by their children is still very strong in Japan today. In 1996, 52% of elderly Japanese coresided with their adult children (Brown et al. 2002), compared with only 12% in the United States (U.S. Bureau of the Census 1999). In addition, Japanese elders are much more likely to receive financial support from their children than older adults in the United States (Maeda and Shimizu 1992). However, the popularity of joint households in Japan has declined significantly, and there has been a shift toward other sources of support, including public pensions (Maeda and Shimizu 1992; Schulz, Borowski, and Crown 1991). On the other hand, far more elderly people work in Japan than in other industrialized nations. In 1988, 36% of older Japanese men and 16% of older Japanese women worked, whereas in the United States, the corresponding figures were 16% and 8% (Bass 1996). However, in both nations, the labor-force participation rates of the elderly population have been declining for the past several decades (Noguchi and Wise 1994; Ogawa 1998).
With regard to the socioeconomic inequalities in health, Japan also exhibits a distinct pattern. Whereas a strong socioeconomic gradient has been observed in both infant and adult mortality, mortality rates among female professional workers are as high as among service workers and higher than among clerical workers (Hasegawa 2001). Moreover, Cockerham, Hattori, and Yamori 2000 recently reported that Okinawans traditionally rank at the top in health and life expectancy and at the bottom in socioeconomic indicators. They suggested that the social gradient thesis does not apply in Japan, and lifestyles such as diet and social support are more important factors. Nevertheless, the aforementioned studies have been entirely based on macro data and have not focused on old age mortality. Although there have been numerous studies of old age mortality in Japan, very little attention is directed to the impact of SES and how it interacts with age and gender (Honma, Kagamimori, and Nruse 1998; Iwamoto et al. 1994; Nakanishi and Tatara 2000; Nakanishi et al. 1997).
|
Model Specifications |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
The importance of age and gender in affecting mortality is well substantiated in the literature. In general, death rates decline from infancy to early teen ages and increase thereafter, growing more rapidly as old age approaches (Rogers et al. 2000). At older ages, the rate of mortality increase slows with age (Horiuchi and Wilmoth 1998). On the other hand, health status and mortality vary significantly between the sexes. In the developed nations, men have higher mortality rates than women, but women have higher rates of morbidity and health care utilization. According to Verbrugge 1989, this paradox can be explained by the fact that women have high rates of acute illnesses and of most nonfatal chronic conditions, whereas men experience higher prevalence rates of the leading fatal conditions. In addition to age and gender, residential setting can influence mortality. As an indicator of urbanization, size of community may reflect disparities in health care services and social structural conditions pertinent to individual health (e.g., ambient hazards and social disorders; Hayward, Pienta, and McLaughlin 1997).
SES status may involve knowledge, resources, community standing, and power. In this model, SES is represented by multiple measures, including education, household income, and home ownership. Education represents the amount of human capital one has, whereas income and home ownership are indicators of the flow and stock of one's material resources. High SES individuals have the knowledge, resources, and social connections to avoid risk or to minimize the effects of disease, impairment, and disability (Link and Phelan 2000). In accordance with the current literature, an inverse relationship between SES and mortality is hypothesized (Elo and Preston 1996; Kaplan 1997).
On the other hand, social relationships have been recognized as risk factors of mortality and morbidity since the early 1970s. Prospective studies have consistently shown increased risk of death among persons with a low quantity, and sometimes low quality, of social relationships (House, Landis, and Umberson 1988). Social relationships can affect individuals by providing them with a more positive view of themselves and their abilities, such as mastery, control, and social competence that, in turn, may protect or better prepare individuals in coping with a health event (Antonucci 1990). Finally, health status at the baseline, including diseases, functional limitations, and self-rated poor health, is postulated to be associated with a greater risk of dying during the subsequent 3 years (Idler and Benyamini 1997; Rogers et al. 2000).
The predictors of mortality are interrelated among themselves. Specifically, SES varies by age, gender, and place of residence (Hayward et al. 1997; O'Rand 1996). In addition, social relations are correlated with demographic variables and SES. At the same time, demographic variables, SES, and social relations influence baseline health status (House et al. 1994). Accordingly, baseline health conditions are conceptualized as intervening variables, in that SES may influence the risk of dying directly and/or indirectly. The theoretical rationale underlying the proposed framework stems from a sociomedical perspective of health (House et al. 1994; Kaplan 1989; Rogers et al. 2000).
|
Methods |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
). A number of procedures were taken to minimize the rate of attrition over time. In particular, all respondents were presented a small gift at the conclusion of the interview. A newsletter describing the findings was mailed to all respondents before the next wave of survey took place. In addition to the proxy interviews, a significant effort was made to convert each individual who had completed the baseline interview, but refused to participate in the subsequent follow-up. Typically, those considered by the interviewers and their supervisors as likely to reverse their refusal were identified and reassigned to another interviewer, often more experienced. Moreover, all those who were unable to participate in one survey (i.e., subjects with proxy interviews and refusals) were contacted again at the subsequent follow-up. As a result, many of them were recovered and retained in later surveys. As an example, in the 1993 survey, 210 individuals who did not participate in 1990 were followed-up, and 135 of them were recovered as respondents in our study. According to Table 1 ,143 three-year episodes between 1987 and 1999 have been included in our analysis, whereas 530 episodes (or 6.9% of the total 7,704 observations) were excluded.
|
The differences between those who dropped out (alive or dead) and those who remained in the study (including self-respondents, those with proxy interviews, and those deceased) were also examined. In particular, a multivariate logistic regression analysis of the probability of dropping out was done for each 3-year interval, by using the baseline characteristics to predict the response status at the follow-up. According to our multivariate analysis, the nonrespondents did not differ significantly from those who remained in the study in terms of all covariates included in our models. A further comparison of the surviving respondents (self-respondents and those with proxy interviews) with the nonrespondents revealed no systematic differences either. These results suggest that our estimates of the effects of SES and other covariates on the risk of dying are unlikely to be biased. The results are not included here, but will be made available on request.
Measures
Deaths among the 2,200 respondents included in the baseline survey were verified periodically against official records. Between 1987 and 1999, 724 members (33%) of this cohort died. In particular, 3 months before each survey, residential cards were sent to various local bureaus of resident registration across Japan to ascertain whether the respondents were still living at the last known addresses. If a respondent had moved, he or she would be interviewed at the new location. In addition, we were able to learn whether a given respondent was alive or not. In Japan, by law, a death must be reported to the Resident Registration Bureau with a death certificate within 7 days. Should a respondent be found deceased, the date of death was recorded. In addition, interviewers who reported deaths that had occurred during the last 3 months further updated this information. These field reports were subsequently checked against the official records. Through this procedure, we were able to ascertain the survival status of all 2,200 members of the cohort regardless of whether they chose to remain in the study or not. Given that the Japanese system of resident and vital registration is internationally known for its high quality, we are quite confident of the validity and reliability of our data on survival status.
SES was assessed by home ownership, education, and income. As an indicator of asset, home ownership was defined as a dummy variable, with "owning your own home" coded as 1. Education was indexed to differentiate among those who completed primary school (6 years), junior high (9 years), and senior high or more (12 years or more). Specifically, with a reference category of 0–5 years of education (4.3% of the total sample), three dummy variables were constructed to reflect 6–8 years of education (56%), 9–11 years of education (28%), and 12 years or more (13%). With a reference category of fewer than 1.65 million yen (i.e., approximately US$15,000 assuming an average exchange rate of 110 yen per US dollar; 11% of total sample), four dummy variables were created to represent various levels of annual household income, including (a) 1.65 through <3.95 million yen ($15,000–$35,909; 27%), (b) 3.95 through <5.75 million yen ($35,909–$52,727; 24%), (c) 5.75 through <9.75 million yen ($52,727–$88,636; 30%), and (d) 9.75 million yen or greater ($88,636 or more; 9%). We chose these specific cut points in representing education and income for two reasons. First, they should represent the entire distribution well. Second, to the extent possible, cut points should represent meaningful social categories as in the case of education. Finally, those with the least education and those with the lowest income were used, respectively, as reference categories in our multivariate Cox regression analyses. One may be concerned with the fact that they represented rather small proportions of the total sample, thus resulting in greater errors in estimation. However, we have experimented with several different choices of the reference group, and the results have remained invariant.
Three demographic measures (i.e., gender, age, and urbanicity) were included to control for population heterogeneity. Gender was defined as a dummy variable, with "female" coded as 1. Age was measured in terms of actual years of age. Urbanicity was measured by a single 5-interval scale item indicating a range of the population size of the area in which the respondent resided. The intervals were recoded to reflect a midpoint representative population size for each of the five categories (.25 = 25,000; .75 = 75,000; 1.5 = 150,000; 6 = 600,000; and 20 = 2,000,000).
Several social relations measures were used, including marital status, work status, size of household, and emotional and instrumental support. Both marital and work status were defined as dummy variables with "currently married" and "currently working full or part time" coded as 1. Size of household reflected the actual number of individuals, including the respondent, who resided within that household. Emotional support received was assessed as a composite of two items concerning how often the respondent felt that someone "listened to them" and "made them feel cared for" (r = .747). Instrumental support received also comprised a linear composite of two items asking how often someone provided "help to you when you were sick" and "help to you when you needed financial assistance" (r = .412). For all support items, a 4-point scale was used with the following coding scheme (scoring in parentheses): very often (4), fairly often (3), once in a while (2), never (1). Higher scores reflect greater support received.
Physical health status was assessed by measures of morbidity, functional status, and self-rated health. Information on morbidity was derived from a checklist of conditions. An index of serious conditions (i.e., those thought to be fatal) was generated by including diabetes, heart disease, hypertension, and stroke, whereas the remaining conditions (e.g., arthritis/rheumatism, respiratory diseases, chronic back pain, etc.) were grouped to be chronic conditions (Ferraro and Farmer 1999). Next, an index of functional status was created using (a) four 4-point items of functional limitations (i.e., crouching, grasping, lifting, and reaching; alpha = .839) and (b) six 5-point items of activities of daily living (i.e., bathing, climbing stairs, walking a half mile, using the phone, shopping, and traveling by bus or boat; alpha = .935). Z-scores for these two measures (r = .831) were computed and then summed. Finally, self-rated health was assessed via three indicators: (a) a rating of physical health [coded: excellent (1), fairly good (2), average (3), not very good (4), and poor (5)], (b) health comparisons with other people one's age [better (1), about the same (2), and worse (3)], and (c) a report of overall satisfaction with one's health [coded: very satisfied (1), relatively well satisfied (2), can't say (3), not very satisfied (4), and not at all satisfied (5)]. Health compared with others was scaled to reflect a 5-point scale, and then the three items were summed (alpha = .796). All physical health measures were coded to reflect greater morbidity, impairment, or poor health.
Regarding mental health, depressive symptoms were represented by seven items drawn from the Center for Epidemiological Studies–Depression scale (CES-D; Radloff 1977). These items included: (a) appetite was poor, (b) sleep was restless, (c) could not get going, (d) everything I did was an effort, (e) felt depressed, (f) felt lonely, and (g) felt sad [coded: most of the time (3), sometimes (2), and rarely (1)]. All items were scored such that a higher score reflects higher levels of depression (alpha = .807). In addition, cognitive impairment at each survey wave was assessed using Pfeiffer 1975 Short Portable Mental Status Questionnaire. Specifically, a count of the number of incorrect responses across nine questions covering short- and long-term memory, orientation to surroundings, knowledge of current events, and the ability to perform mathematical tasks was obtained. Unanswered questions were counted as incorrect (Fillenbaum 1980). With a range of 0–9, a higher score reflects greater cognitive impairment. Table 2 includes the descriptive statistics for selected variables in each episode and in the pooled sample.
|
, pp. 219–223). Each episode involved linking the independent variables (IVs) for each respondent at a given wave with his or her survival status during the subsequent 3 years [i.e., wave 1 (1987) IVs with 1987–1990 survival information; wave 2 (1990) IVs with 1990–1993 survival information; wave 3 IVs (1993) with 1993–1996 survival information; and wave 4 IVs (1996) with 1996–1999 survival information]. Accordingly a surviving respondent in 1999 could conceivably have contributed up to four episodes. The pooled data began with the 2,200 episodes from the 1987 baseline survey. Excluding deaths and nonresponse, these 2,200 episodes were pooled with 1,823 episodes from 1990–1993, 1,705 episodes from 1993–1996, and 1,446 episodes from 1996–1999 resulting in 7,174 episodes available for analysis.
To capture the time-varying nature of the covariates, their values are updated during each interval. One may be concerned with our approach of using episodes as the unit of analysis. However, Allison 1982, Allison 1995(pp. 219–224) has shown that the creation of multiple observations is not an ad hoc method, and it follows directly from factoring the likelihood function for the data. In particular, the probability of dying over the 12-year period for a given individual can be expressed as a product of a series of probabilities during the four 3-year intervals. Each of these terms may be treated as though it came from a distinct and independent observation. Furthermore, Allison has provided several empirical illustrations in which the data were analyzed by using persons as well as episodes as the units of analysis. The results are virtually identical. When there are many ties and many time-dependent covariates, the episode-based approach is relatively easy to implement and computationally more efficient. For illustrations of this approach, please refer to Crimmins, Hayward, and Saito 1996, Hayward, Friedman, and Chen 1998, and Miller, Longino, Anderson, James, and Worley 1999.
As longitudinal data were available over a 12-year period, measures of the period of observation were derived in addition to age. In particular, with the period of 1987–1990 as the reference, three dummy variables were created to represent the 1990–1993, 1993–1996, and 1996–1999 periods. These measures may be viewed as representing changes over time that were not fully captured by measures of other covariates, particularly those at the individual level.
Missing Data
Given the large number of variables involved at each wave of the survey, a suitable strategy for handling missing data needed to be derived. Only three variables contained missing data in more than 20% of the episodes, whereas the remaining variables had missing values in less than 10% of the episodes. To minimize the loss of subjects from missing data, multiple imputation (MI) was undertaken with software developed by Schafer 1997. In MI, each missing value is represented by a set of m > 1 plausible values drawn from their predictive distribution. The variations among the m imputations reflect the uncertainty with which the missing values can be predicted from the observed ones (Rubin 1987). MI has been shown to be an efficient imputation procedure with a sound statistical basis. In the present research, a multivariate normal distribution involving all covariates was assumed to generate imputations for the missing values. This model creates a system of simultaneous regression equations in which each variable potentially depends on all other variables. Although such a model is at best approximately true, experience has repeatedly shown that MI tends to be quite forgiving for such departure. Three complete data sets were imputed and analyses were run on each of these three data sets. Estimates were averaged across the three imputations to generate a single point estimate. Standard errors were then calculated using a formula that combines the average of the squared errors of the estimates and the variance of the parameter estimates across the three samples (Schafer and Olsen 1998).
|
Results |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
Gross Effects
The gross effects of all covariates are listed in Table 3 (under the column of bivariate analyses). The gross effect is measured by using a given covariate as the only predictor of the risk of dying without controlling for any other variable. Two of the three SES variables show a significant gross effect on mortality. In comparison with those with less than 6 years of education, the mortality risk for those with higher education is only 43%–54% as large. Similarly, in contrast with those with the lowest household income (less than 1.65 million yen or roughly US$15,000 a year), the risk of dying is only 60%–78% as large for those with higher household incomes. Home ownership is associated with slightly higher mortality (eb = 1.012), but is not statistically significant.
|
Direct and Indirect Effects
Next, the effects of SES variables are evaluated by controlling various covariates hierarchically (Table 3 ). In particular, SES effects are examined by themselves first (Model 1) and then by controlling for age, gender, and community size (Model 2). Following this, they are evaluated by adding variables of social relations (Model 3), and then health status measures to the equation (Model 4). Finally, period differences are included (Model 5). By analyzing the stability and change in the relative risk ratios across the hierarchical regressions, one may gain some insights concerning the direct as well as indirect effects of the predictors of mortality.
As various covariates are brought into the equations, the relative risk ratios associated with age diminish somewhat but remain quite robust. All else being equal, an increase of 1 year in age leads to a marginal increase of 9% in mortality (Model 5). Because age is treated as a time-varying covariate, this effect measures a mixture of cross-sectional and longitudinal differences. In contrast, when all covariates are controlled, gender differences in mortality become even more accentuated in that female mortality decreases from 52% to only 41% of that of the male. This suggests that the age and gender differences in mortality are substantial, and they cannot be explained by intervening variables, including SES, social relations, and baseline health conditions. More importantly, the observed gender differences are often masked by population heterogeneity.
Given the multidimensional nature of SES, the effects of education, household income and home ownership differ somewhat from one another. In particular, higher levels of education and income are associated with lower mortality than the respective reference categories (i.e., 0–5 years of schooling and an annual income of 1.65 million yen or less), but there are no systematic differences among themselves (Models 1 and 2). This is consistent with the hypothesis of the health ceiling effects among those of higher SES (Robert and House 2000). When demographic characteristics and social relationships are controlled, the net effects of education diminish somewhat but remain statistically significant (Model 3 in Table 3 ). This suggests that educational differences in mortality are partially from heterogeneity in age and sex composition, as well as variations in social networks and social support. Educational differences in mortality are no longer significant when baseline health status and time of observation are included in the equation, indicating that education influences mortality risk indirectly through baseline health conditions (Models 4 and 5 in Table 3 ). The same pattern of changes may be applied to the effects of household income. In contrast, home ownership does not show any significant effect on the risk of dying throughout the hierarchical regression analyses.
Social relations, including marital status and work status, are important predictors of old age mortality among the Japanese. In particular, marital status exerts an independent effect (eb = .779) on mortality in addition to demographic and other measures of social relations (Model 3, Table 3 ). However, its effect fails to maintain statistical significance when baseline health conditions are included in the equation, suggesting that the effect of marital status is mediated through health variables (Model 4). In contrast, the effects of employment status on mortality persist even when baseline health conditions are taken into account (Models 4 and 5). This reflects that the effect of work status on mortality cannot be completely explained by variation in baseline health.
Baseline physical and mental health variables are potent predictors of the risk of dying during the following 3 years (Model 4 in Table 3 ). In particular, poor functional status (eb = 1.123), self-rated poor health (eb = 1.122), and cognitive impairment (eb = 1.133) are associated with higher mortality. The fact that the net effects of serious conditions, chronic conditions, and CES-D are not significant may be from the substantial correlations among the health measures. In particular, the correlations between CES-D and other physical and mental conditions range from .145 to .407.
A relatively unique aspect of this research is that data were collected over a period of 12 years (i.e., 1987–1990, 1990–1993, 1993–1996, and 1996–1999). Measures of these four intervals may be viewed as representing changes over time that are not fully accounted for by various covariates included in the equation (Model 5). According to Table 3 , the initially significant variation in mortality across time periods ceases to be statistically significant, when all other covariates were included. This indicates that the observed increase in mortality over time can be explained by the changes in population heterogeneity as measured by various other covariates in the model. According to a further analysis, this is largely because age is treated as a time-varying covariate in Model 5, and it is significantly correlated with measures of time periods.
To obtain further insights concerning the indirect influences of SES on mortality, linear regressions were used to predict physical and mental health status at the follow-up by using baseline demographics, SES, and social relationship variables (Table 4 ). As suggested by Table 4 , higher education predicts better functional status and less cognitive impairment. Household income is significantly associated with all three health status predictors of mortality. In particular, higher household income at the baseline is correlated with less functional impairment, less self-rated poor health, and less cognitive impairment at the follow-up. Home ownership is associated with less self-rated ill health, suggesting the effect of home ownership has an indirect effect on mortality. In addition to SES variables, there are statistically significant associations between health conditions at follow-up and other covariates, including demographic characteristics and social relationships at baseline (Table 4 ). For instance, those who are employed are likely to have better functional status, better self-rated health, and less cognitive impairment. Accordingly, being employed may also reduce the risk of dying through baseline health.
|
|
|
|
|
Discussion |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
Old age mortality in Japan is directly influenced by age, gender, employment, functional status, self-rated ill health, and cognitive impairment. In addition, SES, along with demographic variables and social relations, exerts indirect effects on the risk of dying. Despite our use of longitudinal data gathered over a 12-year period in Japan, these findings are consistent with those from previous studies in the West (see, e.g., Kaplan 1997; Martelin 1994). Thus, there is further evidence in support of the causal effects of SES on old age mortality. In addition, although there is a substantial acceleration in mortality between 1987 and 1999 for the cohort of 2,200 respondents, such a trend is no longer significant when individual attributes are controlled. This suggests that the initially observed long-term increase in mortality is largely due to the changes in individual characteristics over the 12-year period. This may offer a partial explanation why our findings are similar to those based on short-term follow-up.
These results are consistent with prior findings suggesting the diminishing SES effects on health. For instance, there are substantially diminishing returns of higher income to health, with decreasing and even nonexisting relationships between income and mortality (Backlund, Sorlie, and Johnson 1996; McDonough, Duncan, Williams, and House 1997) and morbidity (House et al. 1994; Mirowsky and Hu 1996) at higher levels of income. This partially reflects a health "ceiling effect" in that people in the upper SES strata maintain overall good health until late in life, leaving little opportunity for improvement in average health among these groups throughout much of adulthood (House et al. 1994; Robert and House 2000). This is consistent with cross-national data in that there is a clear diminishing return of health (i.e., life expectancy) with increasing income per capita (World Bank 1993, p. 34). On the other hand, because we have observed significant gender and age differences in the educational effects on mortality (Table 5 ), we have to interpret the main effects of education in conjunction with the interaction effects. This is because the main effects alone no longer adequately represent the educational effects on mortality. Instead, the focus has to be placed on the simple effects of education with respect to specific gender or age groups. Given the evidence that the effects of education on mortality depend on gender and age, the hypothesis of the SES gradient effects on health may be an oversimplification.
Among the most significant contributions of this research are its findings concerning the interaction between education and other stratification variables, such as gender and age. For women, educational differentials are maintained throughout the 12-year period. On the other hand, for men, there appears to be a crossover in the effect of education on the risk of dying. Based on data from the first Whitehall study, Marmot and Shipley 1996 suggest that SES differences in mortality persist beyond retirement age and in magnitude, increasing up to 89 years of age. However, their study differs significantly from the present one in several respects. First, their sample consisted of over 18,000 male British civil servants. Second, SES is measured in terms of employment grade and car ownership. Third, in their analysis, there is no control for psychosocial and health risk factors at the baseline. These differences make a direct comparison and the interpretation difficult. Further replications based on comparable longitudinal data from other populations are clearly required to further evaluate the validity of our findings.
Why is an educational crossover observed only among Japanese elderly men? Although we are not sure what the underlying causes are, a plausible explanation may lie in gender and SES differences in major causes of death, morbidity, and health behavior. According to Nathanson 1990, the gender-social class-mortality interaction may be because the improvement of survival is more likely to be experienced by members of higher SES, and this increasing SES effect is more pronounced for males than females. Parallel trends in heart disease have been observed in Great Britain. Marmot and McDowall 1986 reported a reversal of the social class gradient for heart disease among men but not women between 1931 and 1971. On the other hand, the incidence of myocardial infarction appeared to decline more rapidly in white-collar male workers than their blue-collar counterparts, whereas no consistent trends were observed among female employees (Pell and Fayerweather 1985). In addition, men of lower levels of education are substantially more likely to smoke than either better-educated men or women (Nathanson 1990).
This study is among the first to evaluate the interactive effects of education and age on health among the elderly population by using repeated observations over an extended period of time. According to the thesis of social stratification of aging and health, SES differences in health are small in early adulthood, greatest in middle and early old age, and relatively small again in late old age (House et al. 1994). This study significantly augments this thesis by showing that educational differentials in old age mortality not only will converge, but also eventually cross over. Among the young-old (i.e., 60–69), the educational gradient is maintained during the entire 12-year period. For the older-old (i.e., 70–79), educational differentials begin to converge, and after 10 years into the follow-up, there is a crossover. Among the oldest-old, those with higher education have an increasingly lower probability of survival over time (Fig. 2). On the other hand, our research provides no support for the hypothesis of cumulative advantage of education (Ross and Wu 1996).
We are not quite sure why there is an educational crossover among those 70 or over in Japan. Nonetheless, this observation is parallel to the racial crossover in mortality observed in prior studies. In the United States, whereas Black persons generally have higher mortality than White persons, a mortality crossover occurs around age 80. After that age, Black older persons survive longer than White older persons (Corti et al. 1999). The existence of a racial crossover makes the possibility of an educational crossover somewhat plausible.
In this regard, at least two hypotheses concerning educational crossover need to be entertained. The first line of reasoning involves the selective mortality among those with less education. That is, high-risk individuals in this group are much more likely to die young, leaving behind a rather hardy group of survivors. Those better educated are more likely to postpone the onset of diseases and disability until late in old age. In comparison with the less educated, a large proportion of those better educated remaining alive at later ages may have a higher burden of disease and unfavorable risk factor profiles. Alternatively, for the well educated, morbidity is much more compressed (Fries 1980). Both scenarios lead to a convergence first and eventually a crossover of educational differentials in mortality.
What are the methodological implications of selective survival? According to a recent review, it may affect the estimation of age-related trends in risk factor-outcome associations in three ways (Kaplan, Haan, and Wallace 1999). First, if those with a given risk factor experience a higher and earlier mortality, then the distribution of this risk factor among the survivors would be altered relative to the "original" distribution. To the extent that the amount of variation in the risk factor is reduced, this may decrease the strength of the association. Second, there may be unmeasured heterogeneity in the susceptibility to the risk factor. If over time the exposed group becomes increasingly composed of those less susceptible to the risk factor, then the association between the risk factor and the outcome will diminish. Within the context of the present study, those with higher SES and those with lower SES, but less susceptible to the unmeasured ill effects, are likely to be overrepresented. Finally, as illustrated earlier in the case of racial crossover in old age mortality, a particular disease may be postponed to later ages rather than being completely prevented. Given that our data are limited to those aged 60 and over, the hypothesis of selective survival before age 60 cannot be evaluated. Further research involving longitudinal data on adults aged 60 or over, as well as those under 60, is required for this purpose.
In addition to selective survival, the second hypothesis postulates a cohort effect. Those born before 1917 in Japan might have lived through periods when certain lifestyles and risk factors led to the narrowing and crossover of educational differences in mortality. However, this hypothesis can only be evaluated with the accumulation of data derived from long-term follow-up of multiple cohorts. Even when observable longitudinal, cross-sectional, and time-lag differences can be measured, the inference and correct interpretation of the underlying age, period, and cohort effects are only possible with certain strong assumptions (Holford 1991; Palmore 1978). Moreover, when age, period, and cohort effects can be separated, underlying causes need to be identified. These include biological and psychosocial changes, environmental and genetic shifts, and interaction of historical circumstances and specific cohorts.
To further our understanding of the linkages between SES and mortality, future research should focus on how and why SES is related to various risk factors, health conditions, and specific causes of death. As suggested by House and associates 1994, social stratification of aging and health is primarily from differential exposure to, and perhaps impact of, major psychosocial and biomedical risk factors across SES and age groups. This information is important for understanding the pathways responsible for mortality differentials. For instance, in the United States, a racial crossover is observed for coronary heart disease deaths only (Corti et al. 1999). On the other hand, Koskinen and Martelin 1994 suggest that sex differences in the effects of SES on mortality may be explained by the differences in the structure of the major causes of death. In particular, mortality with large SES inequalities (e.g., cardiovascular disease and lung cancer) is common among men, whereas causes of death for which SES differentiation is less or reversed (e.g., breast cancer), tend to be common among women. In 1990, malignant neoplasm, heart disease, and cerebrovascular disease were the most important causes of mortality in Japan. It would be interesting to find out whether the educational crossover observed in this research can be replicated with mortality from these causes.
Similar questions may be posed concerning the socioeconomic differentials in the incidence, duration, and trajectory of morbidity, impairment, functional limitation, and disability. For instance, do the better-educated survivors have increased impairment and functional limitations? Or should we expect that morbidity in old age be increasingly compressed among those well educated? For instance, according to a recent study of 10,932 decedents 50 years of age at the baseline interview in the United States, decedents with higher SES experience lower morbidity and disability and better quality of life even in their last years of life (Liao, McGee, Kaufman, Cao, and Cooper 1999). In addition, Wingard and Cohn 1990 show that sex differences in health vary remarkably by age, cause, and outcome (i.e., morbidity vs. mortality).
In this regard, more attention needs to be devoted to the description and explanation of the trajectories of changes in old age. In particular, what are the major patterns of changes in health and other domains, such as financial well-being and social relationships? How do the trajectories in financial well-being, social relationships, and health relate to one another? However, current knowledge concerning these trajectories is extremely limited because of the lack of high-quality longitudinal data. According to the current literature, we only know that there are a multitude of such trajectories, and some of them are likely to be nonlinear (Aldwin, Spiro, Levenson, and Cupertino 2001). Our approach of using time-varying covariates in predicting the risk of dying in old age represents an incremental step in this direction. On the other hand, there have been some attempts of modeling changes spanned beyond a single episode. For example, Anderson, James, Miller, Worley, and Longino 1998 examined functional transitions by pooling up to three 2-year intervals within each of the more than 5,000 respondents in the Longitudinal Study of Aging. To explore whether the 2-year transitions might also be dependent on the respondent's previous functional status, they made an effort to include functional status measured 4 years prior as a predictor in their analysis. This was a modest but important step in the right direction.
Finally, our research also demonstrates the importance of examining the intersection between societal conditions and the biological differences. At issue is whether the observed age and gender differences in mortality stem chiefly from innate factors or reflect social divisions as mediated by SES. In this respect, a life course perspective (Hagestad 1990) in conjunction with the framework of social stratification would be most useful. Recent research suggests that exposure structured by socioeconomic circumstances may accumulate and ultimately increase the risk of adult disease (Lynch 2001). However, how such an observation can be extended to old age is still not well understood. At the same time, given the current state of knowledge, it is no longer sufficient to assert merely that a socioeconomic gradient in old age mortality exists. Nor is it sufficient to continue the debate between proponents of the "compression of morbidity" (Fries 1980) and those of the "failure of success" (Gruenberg 1977). We need to learn more about the underlying mechanisms and to specify the circumstances under which these results may take place. Our findings suggest that education interacts with other ascribed statuses, such as gender and age, in affecting old age mortality.
|
Acknowledgments |
|---|
Harold Pollack, Kuang Yee Liang, and Mark Hayward provided extremely useful advice concerning analytical strategies. Rod Little contributed valuable insights regarding multiple imputation of missing data. We have also benefited from the excellent comments and advice given by John Lynch, Stephanie Robert, and George Kaplan.
Received for publication April 11, 2001. Accepted for publication January 7, 2002.
|
References |
|---|
|
TOPAbstractSES and Gender Interaction...SES and Age Interaction...Old Age Mortality in...Model SpecificationsMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  S. Ito, R. Takachi, M. Inoue, N. Kurahashi, M. Iwasaki, S. Sasazuki, H. Iso, Y. Tsubono, S. Tsugane, and for the JPHC Study Group Education in relation to incidence of and mortality from cancer and cardiovascular disease in Japan Eur J Public Health, October 1, 2008; 18(5): 466 - 472. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. Liang, B. A. Shaw, J. M. Bennett, N. Krause, E. Kobayashi, T. Fukaya, and Y. Sugihara Intertwining Courses of Functional Status and Subjective Health Among Older Japanese J. Gerontol. B. Psychol. Sci. Soc. Sci., September 1, 2007; 62(5): S340 - S348. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 T. Chandola, J. Ferrie, A. Sacker, and M. Marmot Social inequalities in self reported health in early old age: follow-up of prospective cohort study BMJ, May 12, 2007; 334(7601): 990 - 990. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 V. Bos, A. E Kunst, J. Garssen, and J. P Mackenbach Socioeconomic inequalities in mortality within ethnic groups in the Netherlands, 1995-2000 J Epidemiol Community Health, April 1, 2005; 59(4): 329 - 335. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
