## Abstract

Objectives. The vast majority of studies on socioeconomic status (SES) and old age mortality are based on data derived from developed nations. This research examined the SES differentials in old age mortality in China, a developing nation.

Methods. Hazard rate models in conjunction with ordinary least squares and logistic regression analyses were used to ascertain the gross, direct, indirect, and interaction effects of SES on mortality during a 3-year period in a probability sample of 2,943 persons aged 60 years or older in Wuhan, China.

Results. Education, household economic well being, and urban-rural residence showed statistically significant gross effects on old age mortality. Education influenced mortality directly and indirectly. Household economic well being and urbanicity exerted indirect effects on mortality through mediating variables such as stress, social relations, and baseline health status. The mechanism through which education affected mortality differed between men and women, but SES differentials in mortality did not interact with age.

Discussion. SES differentials in old age mortality may be extended to a developing nation such as China. The observed gender by SES interaction effect on old age mortality has important implications for intervention. In particular, improving education among women in underdeveloped areas must remain a high priority, for policy makers in efforts to extend the life expectancy of women.

## Model Specifications

Fig. 1 explicates the hypothesized linkages among SES, stress and social relations, health, and mortality. The model consists of three major components. First, direct linkages are hypothesized between SES and mortality. Second, financial strain, social relations (i.e., social networks and social support), and health status are conceptualized as mediating the effects of SES on mortality. Third, age and gender are postulated as antecedent variables.

In Fig. 1, urban residence, education, and luxury household items are included as measures of SES. Their inclusion reflects our conceptualization of SES as a multidimensional construct and the rather unique societal conditions in China. In particular, education, household economic well being, and urban or rural residence represents SES at the individual, household, and area levels (Feinstein 1993; Krieger and Fee 1994). As an indicator of SES, education has several advantages. It is causally prior to income and occupation and is stable throughout life after young adulthood. More importantly, it is a universal indicator of SES to all adults, whereas occupation is specific to those who are employed (Kitagawa and Hauser 1973; Ross and Wu 1996). Income was not used as a measure of SES, because many respondents were no longer working. Moreover, urban standards of living are much higher than the income data would suggest because of the large subsidies in housing, education, and health care (World Bank 1997b). In rural areas economic exchange often does not involve currency. Consequently, instead of income, we used a count of the presence of 12 household luxury items (e.g., television, refrigerator, and car or truck) to assess household economic well being. Finally, given the substantial disparities between urban and rural areas in income and other dimensions of welfare, urban residence was employed as an indicator for SES.

Within the proposed framework, several research issues may be examined. First, the nature of the SES effects on mortality is often not well defined. The gross effects of SES on mortality are those found without controlling for other confounding variables. In multivariate analyses, the effect of SES on the risk of dying often represents the direct effect, net of the impacts of other covariates. Furthermore, SES may influence mortality indirectly through various mediating variables such as social relationships and baseline health status. There has been a fair amount of research on the gross and direct effects of SES on mortality, whereas little attention has been directed to its indirect effects.

## Methods

### Sample and Data

Data for this research came from the 1991 Survey of Health and Living Conditions of the Aged in Wuhan, China, including the city proper and its surrounding suburbs and rural areas. Wuhan is the capital of Hubei province, which is part of the middle and lower Yangtze River basin in south central China. Wuhan is known as an agricultural center and for its major industries, including iron and steel works, shipbuilding and machinery, and textile and chemical manufacturing. The fifth largest city in China, Wuhan had a population of 4 million in its urban and suburban districts in 1990, and some 3 million people lived in the rural counties. Eight percent of the residents were aged 60 years or older in 1991.

The survey involved a three-stage probability sample of individuals aged 60 or older. Using the 1990 Chinese Census as the sampling frame and stratifying by administrative areas including seven urban districts, two suburban districts, and four rural counties, we selected 3,543 eligible respondents. Interviews were conducted in the respondents' homes and generally lasted between 1 and 2 h. With a response rate of 83%, interviews were completed with 2,943 individuals, including 178 proxy respondents.

### Measures

The survival status for each respondent was followed during a 3-year period, from 1991 to 1994. Of the 2,765 respondents who completed the baseline interview in 1991, 391 individuals (14%) died during the 3 following years. The time of death was recorded by the interviewer on the basis of information given by the appropriate proxy respondent (i.e., next of kin or neighbor). We verified deaths by checking against the Household Registration records, because vital events (i.e., birth, death, and marriage) are by law required to be reported to the local administrative units (i.e., neighborhood resident committees in the city and village offices in the countryside) and recorded in the Household Registration. In general, the quality of Household Registration information is better in urban areas than in rural areas. The increasing migration from rural to urban areas that has occurred since the economic reforms were enacted may also influence the quality of registration. However, reporting of old age mortality is less affected because elderly people are less mobile.

Three measures were used to represent SES. Urban residence was a dummy variable, with urban household registration coded as 1 and rural household registration coded as 0. Educational attainment was indexed by the total number of years in school. Household economic well being was measured by a count of the presence of 12 household luxury items, including telephone, motorcycle, black and white television, color television, refrigerator, washing machine, videocassette recorder, stereo, air conditioner, piano, bicycle, and car or truck.

Two demographic characteristics, age and gender, were included in the analyses. Age was measured in terms of the actual years of age at the time of survey. Gender was a dummy variable, with female coded as 1.

With reference to stress in old age, we focused on stressors related to finance. Three items of current financial difficulty (i.e., not having enough pocket money, dissatisfaction with financial situation, and how finances compare to like-aged others), were transformed into z scores and then summed to create a scale of financial strain

. Marital status, size of household, and current work status were used as measures of social networks. Marital and working status were treated as dummy variables, with currently married and currently working both coded as 1. Size of household was defined as the number of persons living in the household, including the respondent. Measures of social support included emotional as well as instrumental support items. These items were derived on the basis of current conceptualization and a series of factor analyses (Liang, et al. 1992). Emotional support consisted of a composite of two 5-point items assessing (a) the amount of love and caring the respondent can expect from his or her significant others and (b) the willingness of significant others to listen to the personal problems and inner feelings of the respondent. We summed the two emotional support items
$$(r\ =\ .415)$$
to form a composite. Instrumental support comprised two 5-point items indicating the amount of sick care and financial assistance the respondent can expect from his or her significant others. We summed the two instrumental support items
$$(r\ =\ .465$$
) to form a composite. Higher scores for both composites represent more support received.

were computed and then summed. Finally, self-rated health was assessed as a composite of the z scores from four indicators including an overall rating of physical health, health comparisons to other people one's age and to one's health a year prior, and a report of overall satisfaction with one's health
$$(\mathrm{{\alpha}}\ =\ .759)$$
.

### Data Management

To maximize the response rate, we administered an abbreviated version of the survey questionnaire to a proxy respondent when the selected respondent could not complete the full interview for a variety of reasons, including physical illness, hearing problems, or mental problems. A total of 178 proxy respondents were identified at the onset of or during the interview. Proxy responses differed substantially from the self-responses (Table 1 ). In particular, the death rate among those interviewed by proxies was more than three times as high as that among the self-respondents (43.1% vs 12.4%). Furthermore, self-respondents were likely to be younger, male, better educated, married, working, and living in a larger household and to have better functional status. As a result, it is important to assess the bias caused by the exclusion of proxy interviews in the analysis.

### Data Analysis

We used Cox regression analyses to ascertain the direct and indirect effects of SES on mortality, and we used ordinary least squares regression and logistic regression to assess the interrelationships among the covariates. The bulk of the analysis was done on the data collected from the self-respondents. To assess the potential bias caused by the exclusion of proxy responses, findings from data on self-respondents were contrasted with the results involving both proxy and self-respondents.

## Results

### Gross Effects

The gross effects of all variables as the only predictor of the risk of mortality without controlling for any other variable are listed in Table 2 , under the column of bivariate regressions. SES variables all show a significant gross effect on mortality. To illuminate the effects of SES on the risk of dying, one may examine the antilogarithm of the unstandardized regression coefficients (eb), often referred to as hazard ratios or relative risk ratios (Table 2 ). In particular, an older person residing in an urban area was 27% less likely

$$(\mathrm{e}^{\mathrm{b}}\ =\ .729)$$
to die than someone who lived in a nonurban area. Likewise, each additional reported year of education was associated with a 5% reduction in risk of mortality
$$(\mathrm{e}^{\mathrm{b}}\ =\ .947)$$
, and the possession of each additional luxury item resulted in an associated 7% decrease in risk of dying
$$(\mathrm{e}^{\mathrm{b}}\ =\ .930)$$
.

As expected, age and sex also exhibited statistically significant gross effects, with increased age

$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.100)$$
related to a higher risk of mortality and being female
$$(\mathrm{e}^{\mathrm{b}}\ =\ .709)$$
showing a reduction in the risk of mortality. Consistent with current literature, financial strain was associated with an increased gross risk of dying
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.072)$$
, and all social relations measures except size of household were significantly related to a decreased risk of mortality. Specifically, being married
$$(\mathrm{e}^{\mathrm{b}}\ =\ .627)$$
, currently working
$$(\mathrm{e}^{\mathrm{b}}\ =\ .552)$$
, and higher levels of emotional support
$$(\mathrm{e}^{\mathrm{b}}\ =\ .887)$$
and instrumental support
$$(\mathrm{e}^{\mathrm{b}}\ =\ .945)$$
tended to lower the risk of subsequent mortality. For health status, greater numbers of serious health conditions
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.371)$$
, more functional limitations
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.358$$
), and self-rated poor health
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.188)$$
were all predictors of higher mortality. However, chronic conditions did not show a significant impact on mortality.

### Direct and Indirect Effects

The effects of the various predictors of mortality were subsequently examined within each conceptual block so that the correlations among variables within the same block were controlled (Table 2 ). Results from the demographic block containing age and sex were basically the same. With regard to the block containing the SES variables, when urbanicity, education, and luxury items were all entered, only the effect of education

$$(\mathrm{e}^{\mathrm{b}}\ =\ .959)$$
remained significant. This may be due to the substantial correlations between urban residence and luxury household items
$$(r\ =\ .500)$$
. As for the effects of the block of stress and social relations variables on mortality, the results were somewhat similar to the findings of the bivariate analysis. However, financial strain and instrumental support were no longer significant because of their correlations with the other social relations variables. For instance, there was a significant correlation between emotional and instrumental support
$$(r\ =\ .451)$$
. In addition, financial strain was significantly correlated with emotional support
$$(r\ =\ {-}.298)$$
and instrumental support
$$(r\ =\ {-}.268)$$
. When we used the entire block of health status variables to predict mortality, functional status and self-rated ill health remained significant, whereas serious health conditions no longer showed a significant effect, and chronic conditions were associated with a reduced risk of dying
$$(\mathrm{e}^{\mathrm{b}}\ =\ .917).$$

Finally, the effects of SES and other determinants on mortality were analyzed with Cox regression analysis hierarchically (Table 3 ). Specifically, measures of demographics and SES were entered as predictors of mortality (Table 3 , columns 2 and 3). Subsequently, we evaluated the effects of SES by controlling for the effects of intervening variables, including financial strain and social relations first (Table 3 , columns 4 and 5) and then adding the block of health items to the equation (Table 3 , columns 6 and 7). 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 (Table 2 and Table 3 ).

As various covariates were brought into the equations, the relative risk ratios associated with age and gender diminished somewhat but remained quite robust. This suggests that age and gender differences in mortality were substantial, and they could not be explained by intervening variables including SES, stress, social relations, and baseline health conditions.

Given the multidimensional nature of SES, the effects of education differed somewhat from those of urban residence. When age and gender differences on mortality were controlled, the net effects of education remained statistically significant, indicating that educational differences in mortality were not due to population heterogeneity in age and sex composition. Moreover, education had a direct net effect on old age mortality

$$(\mathrm{e}^{\mathrm{b}}\ =\ .958)$$
, although educational differences in old age mortality were partially mediated by stress, social relationships, and baseline health conditions.

On the other hand, demographic characteristics, financial strain, and social relationships appeared to exert suppressive effects on the association between urbanicity and mortality. That is, the relative risk ratio associated with urban residence was not statistically significant when education and luxury household items were controlled (Table 2 ). However, urban-rural differences in old age mortality became statistically significant at the .05 level and somewhat greater when age, gender, stress and social relations were controlled (Table 3 ). Nevertheless, the net effect of urbanicity on mortality could be explained by the differences in baseline health status such as functional status and self-rated ill health.

Across all models, household luxury items did not exert a statistically significant net effect on the risk of dying. This was largely due to the fact the possession of luxury household items was substantially correlated with education

$$(r\ =\ .294)$$
, urbanicity
$$(r\ =\ .500)$$
, and household size
$$(r\ =\ .428)$$
in that 46% of its variance was explained by these three variables.

When demographic and SES variables were controlled, employment continued to show a protective effect, whereas the initially observed effects of marital status and emotional support on mortality were no longer statistically significant. In the full model containing demographic characteristics, SES, stress, social relations, and baseline health status, serious health conditions

$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.203)$$
, functional status
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.115)$$
, and self-rated health
$$(\mathrm{e}^{\mathrm{b}}\ =\ 1.135)$$
all showed significant direct net effects on the risk of dying (Table 3 ). However, within the same equation, financial strain and social relations measures did not show any direct net effects on mortality. This suggests that the gross effects of financial strain and social relations variables were largely due to the population heterogeneity in age, gender, and SES. Furthermore, the direct net effect of work status on mortality was explained by differences in baseline health status.

To obtain further insights concerning the indirect effects of SES on mortality, we examined the relationships between SES and intervening variables such as stress and social relations and health (Table 4 ). According to our model, SES variables indirectly impacted the risk of dying via two mechanisms. First, urbanicity, education, and luxury items impacted mortality indirectly via their effect on serious health conditions, functional status, and self-rated health. Specifically, residing in an urban area and having higher levels of education reduced functional limitations, whereas residing in an urban area, having higher levels of education, and possessing more luxury items reduced self-rated poor health, in turn decreasing the risk of mortality. Second, SES indirectly affected the risk of dying via its impact on stress and social relations and their subsequent impact on health status. For instance, higher levels of education were related to increased emotional social support

$$(b\ =\ .052)$$
, which in turn was associated with improvements in functional status limitations and self-rated poor health.

The findings described thus far are all presented as unstandardized regression coefficients. Because of the different metrics associated with various variables, the relative magnitudes of their direct effects are difficult to discern. Consequently, we computed standardized regression coefficients by multiplying the unstandardized coefficients by the standard deviation of the covariate (see Selvin 1991). Among the major determinants of old age mortality, age

and gender
$$(B\ =\ {-}.454,\ \mathrm{e}^{\mathrm{B}}\ =\ .635)$$
had the largest direct effects, followed by self-rated poor health
$$(B\ =\ .371,\ \mathrm{e}^{\mathrm{B}}\ =\ 1.449)$$
. The effect of education
$$(B\ =\ {-}.170,\ \mathrm{e}^{\mathrm{B}}\ =\ .844)$$
was similar in magnitude to those of serious diseases
$$(B\ =\ .136,\ \mathrm{e}^{\mathrm{B}}\ =\ 1.145)$$
and functional status
$$(B\ =\ .188,\ \mathrm{e}^{\mathrm{B}}\ =\ 1.207)$$
.

### Interaction Effects

To explore the SES differences in old age mortality with respect to age and gender strata, we formed six interaction terms between SES variables (i.e., urban residence, education, and household luxury items) and strata variables such as age and gender. As shown in column 4 of Table 3 , none of the interaction terms involving age and SES variables were statistically significant at the .05 level. Thus, the hypothesis of cumulative advantage of education across age (Ross and Wu 1996) was not supported by data in this study.

On the other hand, there was a statistically significant interaction effect involving gender and education (

in Table 4 ). To obtain additional information, we performed further analyses on the male and female subsamples separately (results not shown). First, the average education for older men was more than three times that for elderly women (i.e., 4.163 years vs 1.226 years). Second, the gross effect of education on old age mortality among women
$$(\mathrm{e}^{\mathrm{b}}\ =\ .782)$$
was more than three times as large as that among men
$$(\mathrm{e}^{\mathrm{b}}\ =\ .940)$$
. Each additional year of formal schooling reduced the relative mortality risk by 6% among men and by 22% among women. Third, when all the antecedent and intervening variables were controlled, the marginal effect of 1 additional year of education remained very strong
$$(\mathrm{e}^{\mathrm{b}}\ =\ .843)$$
, whereas among older men the net direct effect of education was not significant at the .05 level. Among older men, the effect of education was completely mediated by financial strain and social relations (results not shown). Among elderly women, the effect of education remained quite robust even when stress, social relations, and baseline health status were controlled.

Another way to highlight the education by gender interaction effect on mortality among older people in China is to compare the survival functions for different gender and education subgroups. Statistically significant effects on the hazard rates may not necessarily translate into meaningful differentials in survival, depending on the magnitude of baseline mortality rate (Teachman and Hayward 1993). On the basis of the full model (Table 3 , column 3), Fig. 2 displays the proportions of survivors during a period of 36 months at four levels of education for both men and women. The four levels of educational attainment included illiterate, 2.5 years (i.e., mean education for the sample), 6 years (i.e., primary school), and 12 years (i.e., high school). The gender differences in the effects of education on survival were quite apparent in that educational differentials were much greater among women than men.

### Replication With Proxy Interviews

Bias due to the exclusion of proxy respondents was a concern in the current study. To assess such bias, we replicated the Cox regression analyses by including 2,765 self-respondents as well as 178 proxy respondents and by using a subset of the covariates without several measures based on self-report. Specifically, measures of current financial strain, emotional and instrumental support, and self-rated health were not included. In addition, a dummy variable indicating whether a respondent was a self-respondent was included. The results were replicated in that being older, male, and less educated and having more functional status limitations were associated with a greater risk of dying, which was revealed in an analysis of the sample consisting of only self-respondents. Finally, being a self-respondent at the baseline was significantly associated with a lower risk of dying during the subsequent 3 years. This is very likely a reflection of the better health status enjoyed by the self-respondents in general.

## Discussion

A major contribution of the present research is that it provides new information concerning the linkages between SES and old age mortality in a developing nation such as China. Our results show that urban residence, education, and household luxury items all have statistically significant gross effects on mortality. These convergent findings provide some evidence that the socioeconomic gradient in mortality observed in the developed nations can be generalized to developing nations. Furthermore, various measures of SES have distinct direct and indirect effects on old age mortality. When all antecedent and mediating variables are controlled, education has a direct net effect on mortality, whereas urbanicity and luxury household items no longer show any direct net effect. In addition, SES can indirectly impact the risk of dying via stress, social relations, and health status.

Consequently, investigators need to be sensitive to the multidimensional nature of SES. Whenever possible, multiple measures of SES should be employed. More important, attention should be given to the gross, direct, and indirect effects of SES on old age mortality. Given the current state of knowledge, it is no longer sufficient to assert merely that a socioeconomic gradient in old age mortality exists. Researchers need to learn more about the underlying mechanisms and to specify the conditions and circumstances under which SES may make a difference in mortality. In this regard, analyzing data from the Third World countries would be particularly useful in supplementing observations made in the developed nations.

One of our major purposes in this research was to explore the differential SES effects on mortality in relation to age and gender. Several recent studies have offered evidence for the hypothesis of cumulative advantage of education on health (e.g., House et al. 1994; Ross and Wu 1996). However, these studies focused on health measures such as functional limitations and self-rated health among adults in the United States. It would be instructive to determine whether similar evidence may be obtained with reference to old age mortality in a developing nation. No interaction effects involving age and SES measures were found in the present analysis, raising some questions about the generalizability of prior findings on old age mortality.

However, this study constituted a somewhat limited test of the initial hypothesis of cumulative advantage of education on health, because our data were derived entirely from persons aged 60 years or older. The educational levels of the survivors aged 60 or older might be quite different from those of the cohort at earlier ages. According to the theory of cumulative advantage of education on health, SES differentials in health would be smaller at younger ages than at older ages. This would imply that the educational levels of the survivors are not substantially different from those of the cohort at younger ages. However, it is difficult to evaluate this hypothesis because of the lack of data on those aged less than 60 years. Furthermore, the length for the period of mortality follow-up may not be sufficient to detect the hypothesized interaction effect. Ideally, researchers should analyze mortality across the life course by following one or more cohorts over an extended period of time. The selection effect and the possibility of a time varying effect of SES and other covariates on mortality could be examined. Our conclusions can apply only to a sample of Chinese people who survived to 60 years of age. It goes without saying that our findings need to be replicated in other older populations, particularly those in the developing nations.

There are several possible reasons why education exerts such differential effects between older Chinese women and their male counterparts. The difference may be due to the fact that the effect of education on survival diminishes as one becomes more educated. In China, 23% of the older men and 76% of the older women were illiterate in 1991. Furthermore, the average education for older men in our sample was more than three times that for elderly women (i.e., 4.163 years vs 1.226 years). Given the very low average education and its skewed distribution among elderly Chinese women, the marginal educational benefits in survival are likely to be quite substantial. Because older Chinese men have a much higher average education, the marginal effect of education on survival is still significant but more moderate.

Although this study represents an important step toward the understanding of old age mortality in developing nations, more research is needed. Given the multidimensional nature of SES, one may wish to examine the interaction effects involving different SES dimensions. For instance, there is some evidence that the SES differential in mortality is less strong in rural areas compared with urban areas in the United States (Hayward, McLaughlin, & Pienta, 1997). This is a very interesting research question, but it cannot be adequately addressed within the context of the current study. Further research replicating such findings in developing nations is definitely warranted.

Finally, it is conceivable that reciprocal relationships may exist among stress, social support, and health status within our proposed framework (Fig. 1). In particular, major health changes may be conceptualized as stress. Sickness may reduce social networks and elicit more social support (Pruchno, Michaels, and Potashnik 1990). Because stress, social support, and health status are specified as intervening variables between SES and old age mortality, the existence of these reciprocal linkages should not significantly affect the assessment of the effects of SES.

Table 1.

Differences Between Self- and Proxy Respondents on Variables in the Analyses

 Variable Self-Respondents (n = 2,765) Proxy Respondents (n = 178) M SD M SD Survival (months) 33.696*** 6.741 29.073 10.815 Mortality (dead = 1) 14.1%*** 39.3% Age (years) 68.760*** 6.098 73.101 8.300 Sex (female = 1)a 54.4%** 65.2% Urbanicity (urban = 1)a 55.3% 56.7% Education (years)b 2.569** 3.949 1.733 3.345 Luxury items 2.592 1.914 2.556 1.878 Marital status (married = 1)a 61.5%*** 46.6% Work status (currently works = 1)a 32.3%*** 9.6% Size of household 3.778* 2.026 4.174 2.175 Number of serious (life-threatening) health conditions .501*** .730 .803 1.014 Number of chronic (non-life-threatening) health conditions 2.765 1.857 2.781 1.923 Functional status Activities of daily living limitations 1.177*** 1.720 3.598 2.340 Functional limitations 9.28*** 1.206 2.169 1.571
 Variable Self-Respondents (n = 2,765) Proxy Respondents (n = 178) M SD M SD Survival (months) 33.696*** 6.741 29.073 10.815 Mortality (dead = 1) 14.1%*** 39.3% Age (years) 68.760*** 6.098 73.101 8.300 Sex (female = 1)a 54.4%** 65.2% Urbanicity (urban = 1)a 55.3% 56.7% Education (years)b 2.569** 3.949 1.733 3.345 Luxury items 2.592 1.914 2.556 1.878 Marital status (married = 1)a 61.5%*** 46.6% Work status (currently works = 1)a 32.3%*** 9.6% Size of household 3.778* 2.026 4.174 2.175 Number of serious (life-threatening) health conditions .501*** .730 .803 1.014 Number of chronic (non-life-threatening) health conditions 2.765 1.857 2.781 1.923 Functional status Activities of daily living limitations 1.177*** 1.720 3.598 2.340 Functional limitations 9.28*** 1.206 2.169 1.571

c 8 cases missing.

a

Chi-square test was used instead of t test.

b

2 cases missing.

*

p < .05;

**

p < .01;

***

p < .001 (2-tailed test).

Table 2.

Unstandardized Estimates From Cox's Regression Block Analyses of Old Age Mortality in Wuhan (N = 2,745)

 Independent Variable Bivariate Demographics (df = 2) Socioeconomic Status (df = 3) Stress and Social Relations (df = 6) Health (df = 3) B Exp (B) B Exp (B) B Exp (B) B Exp (B) B Exp (B) Age .095*** 1.100 .102*** 1.108 Sex (female = 1) −.343*** .709 −.548*** .578 Urbanicity (urban = 1) −.316** .729 −.165 .848 Education −.054*** .947 −.042* .959 Luxury items −.073** .930 −.027 .972 Financial strain .069*** 1.072 .031 1.032 Marital status (married = 1) −.467*** .627 −.313** .731 Work status (currently works = 1) −.595*** .552 −.502*** .606 Size of household .015 1.015 .031 1.032 Emotional support −.120*** .887 −.086** .917 Instrumental support −.056* .945 −.015 .985 Number of serious health conditions .315*** 1.371 .079 1.083 Number of chronic health conditions .017 1.017 −.087** .917 Functional status .306*** 1.358 .231*** 1.260 Self-rated poor health .172*** 1.188 .098*** 1.103 −2 LL without covariates 6073.330*** 6073.330*** 6073.330*** 6073.330*** −2 LL with covariates 5891.578*** 6054.887*** 6012.303*** 5899.777*** Model chi-square 181.752*** 18.443*** 61.026*** 173.553***
 Independent Variable Bivariate Demographics (df = 2) Socioeconomic Status (df = 3) Stress and Social Relations (df = 6) Health (df = 3) B Exp (B) B Exp (B) B Exp (B) B Exp (B) B Exp (B) Age .095*** 1.100 .102*** 1.108 Sex (female = 1) −.343*** .709 −.548*** .578 Urbanicity (urban = 1) −.316** .729 −.165 .848 Education −.054*** .947 −.042* .959 Luxury items −.073** .930 −.027 .972 Financial strain .069*** 1.072 .031 1.032 Marital status (married = 1) −.467*** .627 −.313** .731 Work status (currently works = 1) −.595*** .552 −.502*** .606 Size of household .015 1.015 .031 1.032 Emotional support −.120*** .887 −.086** .917 Instrumental support −.056* .945 −.015 .985 Number of serious health conditions .315*** 1.371 .079 1.083 Number of chronic health conditions .017 1.017 −.087** .917 Functional status .306*** 1.358 .231*** 1.260 Self-rated poor health .172*** 1.188 .098*** 1.103 −2 LL without covariates 6073.330*** 6073.330*** 6073.330*** 6073.330*** −2 LL with covariates 5891.578*** 6054.887*** 6012.303*** 5899.777*** Model chi-square 181.752*** 18.443*** 61.026*** 173.553***
*

p < .05;

**

p < .01;

***

p < .001 (2-tailed test).

Table 3.

Unstandardized Estimates From Cox's Regression Hierarchical Analyses of Old Age Mortality in Wuhan (N = 2,745)

 Independent Variable Demographics and SES (df = 5) Demographics, SES, and Stress and Social Relations (df = 11) Demographics, SES, Stress and Social Relations, and Health (df = 15) Demographics, SES, Stress and Social Relations, Health, and Interactions (df = 21) B Exp (B) B Exp (B) B Exp (B) B Exp (B) Age .099*** 1.104 .086*** 1.090 .079*** 1.082 .074*** 1.077 Sex (female = 1) −.714*** .490 −.806*** .447 −.911*** .402 −.865*** .421 Urbanicity (urban = 1) −.233* .792 −.333* .717 −.216 .806 −.857 .424 Education −.059** .943 −.054** .948 −.043* .958 −.000 1.000 Luxury items .019 1.019 .031 1.031 .048 1.049 .047 1.048 Financial strain .018 1.018 −.033 .968 −.032 .968 Marital status (married = 1) −.049 .952 −.044 .957 −.035 .965 Work status (currently works = 1) −.453** .636 −.121 .886 −.128 .880 Size of household .008 1.008 .003 1.003 .002 1.002 Emotional support −.047 .954 −.023 .977 −.021 .979 Instrumental support −.028 .973 −.021 .980 −.021 .979 Number of serious health conditions .185** 1.203 .185** 1.204 Number of chronic health conditions −.047 .954 −.042 .959 Functional status .109** 1.115 .109** 1.115 Self-rated poor health .127*** 1.135 .127*** 1.135 Age × Education −.000 1.000 Gender × Education −.141* .869 Age × Luxury items −.000 1.000 Gender × Luxury items .058 1.059 Age × Urbanicity .010 1.010 Gender × Urbanicity −.122 .885 −2 LL without covariates 6073.330*** 6073.330*** 6073.330*** 6073.330*** −2 LL with covariates 5871.952*** 5849.803*** 5755.968*** 5748.465** Model chi-square 201.378*** 223.527*** 317.362*** 324.865***
 Independent Variable Demographics and SES (df = 5) Demographics, SES, and Stress and Social Relations (df = 11) Demographics, SES, Stress and Social Relations, and Health (df = 15) Demographics, SES, Stress and Social Relations, Health, and Interactions (df = 21) B Exp (B) B Exp (B) B Exp (B) B Exp (B) Age .099*** 1.104 .086*** 1.090 .079*** 1.082 .074*** 1.077 Sex (female = 1) −.714*** .490 −.806*** .447 −.911*** .402 −.865*** .421 Urbanicity (urban = 1) −.233* .792 −.333* .717 −.216 .806 −.857 .424 Education −.059** .943 −.054** .948 −.043* .958 −.000 1.000 Luxury items .019 1.019 .031 1.031 .048 1.049 .047 1.048 Financial strain .018 1.018 −.033 .968 −.032 .968 Marital status (married = 1) −.049 .952 −.044 .957 −.035 .965 Work status (currently works = 1) −.453** .636 −.121 .886 −.128 .880 Size of household .008 1.008 .003 1.003 .002 1.002 Emotional support −.047 .954 −.023 .977 −.021 .979 Instrumental support −.028 .973 −.021 .980 −.021 .979 Number of serious health conditions .185** 1.203 .185** 1.204 Number of chronic health conditions −.047 .954 −.042 .959 Functional status .109** 1.115 .109** 1.115 Self-rated poor health .127*** 1.135 .127*** 1.135 Age × Education −.000 1.000 Gender × Education −.141* .869 Age × Luxury items −.000 1.000 Gender × Luxury items .058 1.059 Age × Urbanicity .010 1.010 Gender × Urbanicity −.122 .885 −2 LL without covariates 6073.330*** 6073.330*** 6073.330*** 6073.330*** −2 LL with covariates 5871.952*** 5849.803*** 5755.968*** 5748.465** Model chi-square 201.378*** 223.527*** 317.362*** 324.865***

Note: SES = socioeconomic status.

*

p < .05;

**

p < .01;

***

p < .001 (2-tailed test).

Table 4.

Unstandardized Estimates From Regression Analyses of Functional Status and Self-Rated Health in Wuhan (N = 2,745)

 Independent Variable Financial Strain Marital Statusa Work Statusa Emotional Support Instrumental Support Serious Health Conditions Functional Status Self-Rated Poor Health Age .017* −.132*** −.145*** −.039*** .003 −.000 .094*** .001 Sex (female = 1) .361*** −1.156*** −1.463*** .117 .451*** .125*** .466*** .401*** Urbanicity (urban = 1) −.687*** .355*** −1.215*** −.174* −.668*** .105** −.176* −.773*** Education .009 .053*** −.003 .052*** −.007 .009* −.025** −.033* Luxury items −.308*** .033 −.070* .138*** .270*** −.001 −.018 −.131*** Financial strain .017** .090*** .298*** Marital status (married = 1) .077* −.024 .197 Work status (currently works = 1) −.213*** −.694*** −1.205*** Size of household −.004 .025 .041 Emotional support −.001 −.055*** −.133*** Instrumental support −.009 −.020 −.086***
 Independent Variable Financial Strain Marital Statusa Work Statusa Emotional Support Instrumental Support Serious Health Conditions Functional Status Self-Rated Poor Health Age .017* −.132*** −.145*** −.039*** .003 −.000 .094*** .001 Sex (female = 1) .361*** −1.156*** −1.463*** .117 .451*** .125*** .466*** .401*** Urbanicity (urban = 1) −.687*** .355*** −1.215*** −.174* −.668*** .105** −.176* −.773*** Education .009 .053*** −.003 .052*** −.007 .009* −.025** −.033* Luxury items −.308*** .033 −.070* .138*** .270*** −.001 −.018 −.131*** Financial strain .017** .090*** .298*** Marital status (married = 1) .077* −.024 .197 Work status (currently works = 1) −.213*** −.694*** −1.205*** Size of household −.004 .025 .041 Emotional support −.001 −.055*** −.133*** Instrumental support −.009 −.020 −.086***
a

Logistic regression was used instead of OLS.

*

p < .05 ;

**

p < .01;

***

p < .001 (2-tailed test).

Figure 1.

A conceptual model of socioeconomic differentials in mortality.

Figure 1.

A conceptual model of socioeconomic differentials in mortality.

Figure 2.

Educational differentials in mortality by gender.

Figure 2.

Educational differentials in mortality by gender.

This research was supported by Grants R01-AG08094, T32 AG00114, and T32 AG00134 from the National Institute on Aging. Carol Weissman, Paula Lantz, and Xian Liu contributed many useful comments to this article. The assistance of Cathy Fegan is gratefully acknowledged. Additional tables involving correlations among the independent variables and the analyses including proxy respondents are available from the first author upon request.

## References

Bian Y., Logan J. R.,
1996
. Market transition and the persistence of power: The changing stratification system in urban China.
American Sociological Review

61:
739
-758.
Branch L., Ku L.,
1989
. Transition probabilities to dependency, institutionalization, and death among the elderly over a decade.
Journal of Aging and Health

1:
370
-408.
Christenson B. A., Johnson N. E.,
1995
. Educational inequality in adult mortality: An assessment with death certificate data from Michigan.
Demography

32:
215
-229.
Cook T. D., Campbell D. T.,
1979
.
Quasi-experimentation
Rand McNally, Chicago.
Dorey F. J., Little R. J. A., Schenker N.,
1993
. Multiple imputation for threshold-crossing data with interval censoring.
Statistics in Medicine

12:
1589
-1603.
Elo I. T., Preston S. H.,
1996
. Educational differentials in mortality: United States, 1979-85.
Social Science & Medicine

42:
47
-57.
Feinstein J. S.,
1993
. The relationship between socioeconomic status and health: A review of the literature.
Milbank Quarterly

71:
279
-322.
Feldman J. J., Makuc D. M., Kleinman J. C., Cornoni-Huntley J.,
1989
. National trends in educational differences in mortality.
American Journal of Epidemiology

129:
919
-933.
Ferraro K. F., Farmer M. M.,
1999
. Utility of health data from social surveys: Is there a gold standard for measuring morbidity?.
American Sociological Review

64:
303
-315.
George L. K.,
1996
. Social factors and illness. Binstock R., George L. K., , ed.
Handbook of aging and the social sciences
4th ed.
229
-252. Academic Press, New York.
Grigsby J. S., Olshansky S. J.,
1989
. The demographic components of population aging in China.
Journal of Cross-Cultural Gerontology

4:
307
-334.
Gu S., Guo J.,
1989
.
A comprehensive study of the urban elderly
Wuhan University Press, Wuhan, China.
Gu S., Liang J.,
1994
. Old age support in China.
Almanac of China's population 1994

118
-125. Chinese Academy of Social Sciences, Population Research Institute, Beijing, China.
Gu S., Wang B.,
1989
.
International comparison of the health and social life of the elderly
Wuhan University Press, Wuhan, China.
Hagsted G. O.,
1990
. Social perspectives on the life course. Binstock R. H., George L. K., , ed.
Handbook of aging and the social sciences
3rd ed.
151
-168. Academic Press, New York.
Hanushek E. A., Jackson J. E.,
1977
.
Statistical methods for social sciences
Academic Press, New York.
Hayward M. D., Crimmins E. M., Saito Y.,
1998
. Cause of death and active life expectancy in the older population of the United States.
Journal of Aging and Health

10:
192
-213.
Hayward M. D., McLauglin D. K., Pienta A. M.,
1997
. Inequality in men's mortality: The socioeconomic status gradient and geographic context.
Journal of Health and Social Behavior

38:
313
-330.
Ho S. C.,
1991
. Health and social predictors of mortality in an elderly Chinese cohort.
American Journal of Epidemiology

133:
907
-921.
House J. S., Lepkowski J. M., Kinney A. M., Mero R. P., Kessler R. C., Herzog A. R.,
1994
. The social stratification of aging and health.
Journal of Health and Social Behavior

35:
213
-234.
Hsiao W. C., Liu Y.,
1996
. Economic reform and health lessons from China.
New England Journal of Medicine

335:
430
-432.
Hummer R. A., Rogers R. G., Eberstein I. W.,
1998
. Sociodemographic differentials in adult mortality: A review of analytic approaches.
Population and Development Review

24:
553
-578.
Hunt K., Annandale E.,
1999
. Relocating gender and morbidity: Examining women's and men's health in contemporary Western societies.
Social Science & Medicine

48:
1
-5.
Idler E. L., Benyamini Y.,
1997
. Self-rated health and mortality: A review of twenty-seven community studies.
Journal of Health and Social Behavior

38:
21
-37.
Kaplan G. A.,
1997
. Behavioral, social, and socioenvironmental factors adding years to life and life to years. Hickey T., Speers M.A., Prohaska T.R., , ed.
Public health and aging

37
-52. Johns Hospkins University Press, Baltimore.
Kaplan H. B.,
1989
. Health, disease, and the social structure. Freeman H. E., Levine S., , ed.
Handbook of medical sociology
4th ed.
46
-68. Prentice-Hall, Englewood Cliffs, NJ.
Kitagawa E. M., Hauser P. M.,
1973
.
Differential mortality in the United States: A study in socioeconomic epidemiology
Harvard University Press, Cambridge, MA.
Knudsen C., McNown R.,
1993
. Changing causes of death and the sex differential in the U.S.: Recent trends and projections.
Population Research and Policy Review

12:
27
-42.
Koskinen S., Martelin T.,
1994
. Why are socioeconomic mortality differences smaller among women than among men?.
Social Science & Medicine

38:
1385
-1396.
Krieger N., Fee E.,
1994
. Social class: The missing link in U.S. health data.
International Journal of Health Services

24:
25
-44.
Land K. C., Guralnik J. M., Blazer D. G.,
1994
. Estimating increment-decrement life tables with multiple covariates from panel data: The case of active life expectancy.
Demography

31:
297
-319.
Lengermann P. M., Niebrugge J.,
1996
. Contemporary feminist theory. Ritzer G., , ed.
Sociological theory

436
-486. McGraw-Hill, New York.
Liang J., Gu S., Krause N.,
1992
. Social support among the aged in Wuhan, China.
Asia-Pacific Population Journal

7:
33
-62.
Liu X., Hermalin A. I., Chuang Y. L.,
1998
. The effect of education on mortality among older Taiwanese and its pathways.
Journal of Gerontology: Social Sciences

53B:
S71
-S82.
Martelin T.,
1994
. Mortality by indicators of socioeconomic status among the Finnish elderly.
Social Science & Medicine

38:
1257
-1278.
Martikainen P.,
1995
. Socioeconomic mortality differentials in men and women according to own and spouse's characteristics in Finland.
Sociology of Health and Illness

17:
353
-375.
McDonough P., Williams D. R., House J. S., Duncan G. J.,
1999
. Gender and socioeconomic gradient in mortality.
Journal of Health and Social Behavior

40:
17
-31.
Nathanson C. A.,
1977
. Sex, illness, and medical care: A review of data, theory, and method.
Social Science & Medicine

11:
13
-25.
O'Rand A. M.,
1996
. The cumulative stratification of the life course. Binstock R. H., George L. K., , ed.
Handbook of aging and the social sciences
4th ed.
188
-207. Academic Press, New York.
Pruchno R. A., Michaels J. E., Potashnik S. L.,
1990
. Predictors of institutionalization among Alzheimer disease victims with caregiving spouses.
Journal of Gerontology: Social Sciences

45:
S259
-S266.
Riley M. W.,
1987
. On the significance of age in sociology.
American Sociological Review

52:
1
-14.
Ross C., Wu C.,
1996
. Education, age, and the cumulative advantage in health.
Journal of Health and Social Behavior

37:
104
-120.
Schoenfeld D., Malmrose L., Blazer D., Gold D., Seeman T.,
1994
. Self-rated health and mortality in the highly functioning elderly—A close look at healthy individuals: McArthur field study of successful aging.
Journal of Gerontology: Medical Sciences

49:
M109
-M115.
Selvin S.,
1991
.
Statistical analysis of epidemiologic data
Oxford University Press, New York.
Sorlie P. D., Backlund E., Keller J. B.,
1995
. U.S. mortality by economic, demographic, and social characteristics: The National Longitudinal Mortality Study.
American Journal of Public Health

85:
949
-956.
Sugisawa H., Liang J., Liu X.,
1994
. Social network, social support, and mortality among older people in Japan.
Journal of Gerontology: Social Sciences

49:
S3
-S13.
Teachman J. D., Hayward M. D.,
1993
. Interpreting hazard rate models.
Sociological Methods and Research

21:
340
-371.
U.S. Bureau of the Census
1999
.
International data base

U.S. Department of Commerce, Bureau of Economic Analysis
1995
. Table 2. Total and per capita disposable income by state and region, 1989–94.
Survey of Current Business

75:
41
Whyte M. K.,
1995
.
City versus countryside in China's development
Australian National University, Canberra, Australia.
World Bank
1992
.
China: Long-term issues and options in the health transition
Author, Washington, DC.
World Bank
1997
.
Financing health care: Issues and options for China
Author, Washington, DC.
World Bank
1997
.
Sharing rising incomes
Author, Washington, DC.
Yu E. S., Kean Y. M., Slymen D. J., Liu W. T., Zhang M., Katzman R.,
1998
. Self-perceived health and 5-year mortality risks among the elderly in Shanghai, China.
American Journal of Epidemiology

147:
880
-890.
Yu M. Y., Sarri R.,
1997
. Women's health and gender inequality in China.
Social Science & Medicine

45:
1885
-1898.