A comparison of the relationships of education and income with mortality: the national longitudinal mortality study

Social Science & Medicine 49 (1999) 1373±1384

www.elsevier.com/locate/socscimed

A comparison of the relationships of education and income p with mortality: the national longitudinal mortality study Eric Backlund a,*, Paul D. Sorlie b, Norman J. Johnson a a US Bureau of the Census, Suitland, MD 20233, USA Epidemiology and Biometry Program, National Heart, Lung, and Blood Institute, National Institutes Of Health, Bethesda, MD, USA

b

Abstract A sample of over 400,000 men and women, ages 25±64, from the National Longitudinal Mortality Study (NLMS), a cohort study representative of the noninstitutionalized US population, was followed for mortality between the years of 1979 and 1989 in order to compare and contrast the functional forms of the relationships of education and income with mortality. Results from the study suggest that functional forms for both variables are nonlinear. Education is described signi®cantly better by a trichotomy (represented by less than a high school diploma, a high school diploma or greater but no college diploma, or a college diploma or greater) than by a simple linear function for both men ( p < 0.0001 for lack of ®t) and women ( p = 0.006 for lack of ®t). For describing the association between income and mortality, a two-sloped function, where the decrease in mortality associated with a US$1000 increase in income is much greater at incomes below US$22,500 than at incomes above US$22,500, ®ts signi®cantly better than a linear function for both men ( p < 0.0001 for lack of ®t) and women ( p = 0.0005 for lack of ®t). The di€erent shapes for the two functional forms imply that di€erences in mortality may primarily be a function of income at the low end of the socioeconomic continuum, but primarily a function of education at the high end. Published by Elsevier Science Ltd. Keywords: Educational status; Income; Mortality; United States

Introduction The inverse association between detrimental health outcomes and socioeconomic status (SES) has been well established by epidemiological research

The views expressed in this manuscript are attributable to the authors and do not necessarily re¯ect those of the US Census Bureau. * Corresponding author. Tel.: +1-301-457-4270; fax: +1301-457-3766. E-mail address: [email protected] (E. Backlund) p

0277-9536/99/$ - see front matter Published by Elsevier Science Ltd. PII: S 0 2 7 7 - 9 5 3 6 ( 9 9 ) 0 0 2 0 9 - 9

(Antonovsky, 1967; Kitagawa and Hauser, 1973; Townsend et al., 1988; Rogot et al., 1992; Feinstien, 1993). Most researchers view the relationship between socioeconomic status and health as arising from an association with intermediate risk factors (Rose and Marmot, 1981; Dutton and Levine, 1989; Haan et al., 1989; Adler et al., 1993) which produce a `generalized susceptibility' to ill-health (Syme and Berkman, 1976). Some hypothesized intermediate risk factors that may be associated with both social status and health are material standard of living, access to and knowledge of health care, exposure to environmental risk factors, and a wide variety of behavioral, social, and psycho-

1374

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

logical processes (Rose and Marmot, 1981; Dutton and Levine, 1989; Haan et al., 1989; Adler et al., 1993). However, the pathways through which the intermediate risk factors interact to produce the SES health gradient are poorly understood. One of the ®rst steps in understanding the biological and social pathways in the association of SES and health is understanding the socioeconomic di€erentials themselves in as much detail as possible, because any hypothesized pathway which explains these di€erentials must be consistent with their observed pattern. The gradient between socioeconomic status and health has been shown to exert itself at all points along the socioeconomic scale with each increase in socioeconomic status being accompanied by a corresponding increase in health (Rose and Marmot, 1981; Haan et al., 1989; Adler et al., 1993, 1994; Williams and Collins, 1995). Researchers have suggested that hypotheses that seek to explain the gradient need to take account of its persistence along the entire socioeconomic continuum (Haan et al., 1989; Adler et al., 1994; Williams and Collins, 1995). One strategy for further understanding of the socioeconomic health gradient is to consider di€erent alternative measures of SES jointly because each measure may represent di€erent aspects of social status and may be associated with di€erent intermediate risk factors. Two of the most important measures of SES are education and income. Education may represent acquired knowledge, skill in managing the social system to meet desired ends, is closely associated with lifestyles and behaviors (Matthews et al., 1989) and is often thought to be closely associated with the status component of the Weberian description of social class. Income represents material resources, potential access to di€erent lifestyles, a sense of security, and is often thought to be closely associated with the class or economic component of the Weberian description of social class. Thus, examination of the pattern of mortality di€erentials when income and education are considered jointly may provide new insight into the pathway through which the SES di€erentials are produced. A straightforward way of examining patterns of mortality di€erentials in relation to the joint e€ect of income and education would be to compare and contrast the shape or functional form of each of their relationships with mortality. Di€ering e€ects of income and education with respect to their functional forms would imply that mortality di€erentials between subgroups at di€erent points along the socioeconomic continuum may primarily be functions of either education or income alone, which in turn may suggest di€erent casual pathways exist at di€erent points along the continuum. Research on large data sets has shown that both education and income have substantial net associ-

ations, after adjustment for each other, with both mortality (Kitagawa and Hauser, 1973; Sorlie et al., 1995; Elo and Preston, 1996) and health status (House et al., 1990). Other research has shown that the relationship between income and health has a steep slope at lower income levels and a much more gradual slope at high income levels (Wilkinson, 1986; Duleep, 1995; Backlund et al., 1996). To our knowledge, the shape of the relationship between education and health has not been studied rigorously. This report uses data from the National Longitudinal Mortality Study (NLMS) to undertake a statistically rigorous examination of the functional form of the relationship between education and mortality. The study also uses NLMS data to determine the shape of the relationship between income and mortality for men and women between the ages of 25 and 64. A comparison functional form of the mortality gradient for education and income is then discussed. Methods The NLMS is a long term prospective study of mortality in the United States which is primarily used to study socioeconomic and demographic mortality di€erentials (Rogot et al., 1988, 1995). For the analyses in this manuscript, data from nine di€erent samples taken from Current Population Surveys (CPS, US Department of Commerce, 1978), conducted by the US Bureau of the Census between March 1979 and March 1985, are utilized. The CPS is a survey of households conducted by personal and telephone interview which has a response rate of close to 96%. The CPS does not include the institutionalized population and is more likely to miss the mobile, homeless, and people without telephones. From the identi®ed CPS samples, a subset of 189,414 men and 225,810 women, ages 25±64 with nonmissing values for all variables analyzed, were followed for subsequent mortality from the date of survey until 31 December 1989. Thus, the maximum length of follow-up ranges from 4.75 to 10.75 years depending on the date of the individual survey. Decedents were ascertained by matching the CPS sample to the National Death Index (NDI, National Center for Health Statistics, 1990). The NDI has been shown to be an e€ective and accurate means of ascertaining deaths using personal identi®ers (Wentworth et al., 1983; Stampfer et al., 1984; Williams et al., 1992). During the follow-up period, 7679 deaths were observed for women and 8458 were observed for men. The CPS elicited information on annual income at baseline from six di€erent sources: wages and salaries; nonfarm self-employment income; farm self-employment income; interest and dividends; social security

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

1375

Fig. 1. Graphical depictions of hypothetical discontinuous models with no slope, a single slope and multiple slopes for education, and a two slope model for income.

income; and public assistance income. Family income was then obtained by summing the six sources of income over all members of the family. The summed total was then adjusted to re¯ect 1980 dollars and assigned to one of seven categories: US$0±4999; 5000± 9999; 10,000±14,999; 15,000±19,999; 20,000±24,999; 25,000±49,999 or over 50,000. Years of education ranged from 0 to 19 with 19 representing two or more years of postgraduate training. A high school diploma is represented by 12 years of education while a college diploma is represented by 16. The CPS also collected information at baseline on several demographic variables which were used in this study as control variables. Age at baseline was categorized into ®ve year age groups. Race was categorized as white, black, or other. Household size was categorized as 1, 2, 3, 4, 5, 6, 7 or 8 or more persons per household. Marital status was categorized as married, widowed, divorced, separated, or never married. Employment status was categorized as employed (working for pay during the survey week), unemployed (not working during survey week but had looked for work in past four weeks), housework (occupied with

own housework during survey week), unable to work (due to long term physical or mental illness or disability), or other (includes retired, volunteer work, discouraged workers and any other persons not classi®ed elsewhere). Occupational status for men was classi®ed using Seigal scores (Miller, 1991) categorized by 5 percentile intervals. In all analyses presented in this manuscript, a Cox regression model (Kalb¯eisch and Prentice, 1980; Collet, 1994) was used to model the relationship between the log of the relative risk of mortality and SES. Results from Cox models are presented in terms of relative risks obtained by exponentiating the regression coecients. For the results presented, the baseline value is the group with the highest SES which, by de®nition, is assigned a relative risk of 1. For some models two sets of coecients are presented. One set of coecients is adjusted for only demographic variables (marital status, household size, employment status, age, and race) and the other set is additionally adjusted for three SES variables (education, income, and for men only, Seigal scores for occupational status). Income and education can also be viewed in a

1376

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

causal framework by comparing the education coecients adjusted for only the demographic variables with the income coecients adjusted for both SES and demographic variables. All models are sex speci®c. The shape of the relationship was determined by comparing the log-likelihood of a Cox model containing a linear function of either education or income as the independent variable to one containing an alternative nonlinear function as the independent variables. The statistic ÿ2(the log-likelihood of the linear modelÿthe log likelihood of the nonlinear alternative model) follows a w 2 distribution which has an associated p-value (Collet, 1994). A signi®cant p-value indicates that the alternative nonlinear model is signi®cantly better for describing the association between either education or income and mortality than the linear model. In the Cox model, it is the log of the relative risk rather than the actual relative risk that is related to the dependent variables in the model. Thus, the linear functions of the SES components are actually linearly related to the log of the relative risk. In the small range of observed risks, however, the curvature of the log function is small, therefore the relationship between relative risk and SES is nearly linear if a linear function is speci®ed in the Cox model. Some researchers assert that the association between education and health related variables is probably not linear because 12 and 16 years of education represent the attainment of signi®cant educational credentials (high school and college diplomas) (Liberatos et al., 1988; Krieger et al., 1993). Therefore, the e€ect on mortality of increasing education from 11 to 12 years or from 15 to 16 years is greater than a one year increase in education at other points along the educational gradient. Consequently, the relationship between education and mortality may be discontinuous at 12 and 16 years, and the relationship between education and mortality is broken into three segments represented by less than a high school education; a high school diploma but no college degree; or a college degree or higher. The segments are allowed to be discontinuous. Three alternative discontinuous models are examined: one where the slopes for each segment are zero; one where the slopes for each segment are equal but nonzero and one where the slopes for the di€erent segments are not equal to each other, i.e. there may be multiple slopes (Fig. 1). The three discontinuous models are hierarchical with increasing levels of complexity. Thus, the best ®tting model for education can be determined by comparing the log-likelihood for each of the four possible models in hierarchical fashion. When the slopes for each segment of education in the discontinuous model are all zero, the model is simply a trichotomy (see discontinuous Ð no slope,

Table 1 Sample size, number of deaths, and age-race adjusted mortality rates per 100,000 person years in the US population by education and income for men and women aged 25±64a Men

Women

N

D

R

N

Education 8 or less 9±11 12 13±15 16 Over 16

18,838 21,708 68,699 33,662 24,914 21,593

1635 1433 3027 1188 652 523

518 23,251 1676 354 501 30,082 1513 340 414 100,199 3011 249 399 37,058 879 253 294 22,197 376 207 260 13,023 224 187

Income US$0±US$4999 US$5000±US$9999 US$10,000±US$14,999 US$15,000±US$19,999 US$20,000±US$24,999 US$25,000±US$49,999 Over US$50,000

7621 16,504 27,443 29,997 31,936 62,684 13,229

583 1015 1443 1288 1287 2355 487

674 598 494 414 373 337 272

18,247 30,145 36,363 33,075 32,587 62,473 12,929

D

1311 1549 1347 968 816 1365 296

R

469 337 298 257 246 214 202

a N = sample size; D = number of deaths, R = mortality rate per 100,000 person years adjusted for age and race by the indirect method.

Fig. 1), and the only association education has with mortality is with the attainment of high school and college diplomas. When the slopes for each segment in the discontinuous model are equal but nonzero, education has a constant linear association with mortality, except when signi®cant credentials are attained. When a high school or college diploma is attained, there is an additional decrease in mortality, above and beyond the linear e€ect (see discontinuous Ð single slope, Fig. 1). When the discontinuous model allows for multiple slopes for each education segment, the attainment of signi®cant credentials is accompanied by both a discontinuity and a change in the slope of the relationship (see discontinuous Ð multiple slopes, Fig. 1). The design matrices for the three discontinuous models have been considered in standard regression textbooks (Neter et al., 1985) The education variable used in each model is continuous except that values from 0 to 8 are grouped together and assigned a value of 6 (the mean) because the sample size of the individual values is small. Income is often viewed as representing a proxy for the material environment and it might be suspected that an additional unit of income is less likely to be used to provide health necessities; such as proper housing, nutrition, and medical care, when income is high than when it is low. Thus, an alternative model of diminishing returns to income can be represented by a two slope model (see two slope model, Fig. 1) where

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384 Table 2 Relative risks by education and income for men and women of ages 25±64 adjusted for age, race, household size, marital status, employment status, occupation (for men only), and either income (for the education model) or education (for the income model) with 95% con®dence intervals Men RR Education (yr) 0±8 1.42 9±11 1.54 12 1.42 13±15 1.41 16 1.12 16+ 1.00 Income (US$1000) 0±4.999 1.57 5±9.999 1.38 10±14.999 1.31 15±19.999 1.14 20±24.999 1.08 25±49.999 1.06 Over 50 1.00

Women 95% CI

RR

95% CI

(1.24±1.57) (1.37±1.73) (1.28±1.58) (1.26±1.58) (0.99±1.26)

1.33 1.36 1.18 1.18 1.02 1.00

(1.15±1.53) (1.18±1.58) (1.02±1.35) (1.01±1.36) (0.87±1.21)

(1.38±1.80) (1.22±1.55) (1.17±1.46) (1.02±1.27) (0.96±1.20) (0.96±1.18)

1.44 1.33 1.28 1.17 1.09 1.08 1.00

(1.25±1.66) (1.16±1.52) (1.12±1.45) (1.03±1.34) (0.95±1.25) (0. 94±1.21)

income has one slope when it is below the median income category for the total US population (US$20,000±25,000) and another slope when it is above the median. The sensitivity of the model to using points other than the median for the change in slope will also be considered. The midpoint of each income category was used as the independent variable in the income model. A value of US$60,000 was used for the US$50,000 and over category, because it is near the median for the category. Graphical methods showed no serious violations of

1377

the assumption of proportional hazards, implicit in the Cox model, between any of the subgroups used in the analyses. The complex design of the CPS is not taken into account in these analyses because previous research on both the NLMS (Anderson et al., 1997) and a similar database (Feldman et al., 1989) suggests that neither the point estimates of the socioeconomic mortality gradient nor their standard errors are greatly a€ected by the complex nature of the survey. Results Table 1 shows the sample size, number of deaths, and death rates per 100,000 person years adjusted for age and race by the indirect method (Fliess, 1981) for various categories of education and income. An inverse socioeconomic gradient is observed for both measures. Table 2 displays the relative risk for categories of income and education for men and women adjusted for both SES, and demographic variables. For both men and women, income appears to show a strong inverse mortality gradient among people with low income, but the gradient appears to become much weaker as income increases (Table 2). The observed pattern of relative risks for educational categories suggests a nonlinear relationship (Table 2). Any of the three discontinuous functions discussed earlier would appear to be reasonable alternatives. Table 3 presents the degrees of freedom and p-values obtained from the log-likelihood statistic for hierarchical comparisons of pairs of discontinuous models for education. For both men and women, a comparison of the log-likelihoods in Table 3 shows the multiple slopes discontinuous model ®ts signi®cantly better than the linear model regardless of whether or not the model is adjusted for SES. Therefore, the linear model is

Table 3 p-values and degrees of freedom for a hierarchical comparison of alternative models for education for men and women Adjustment variables men demographic variables

women demographic+SES variables

Discontinuous, multiple slopes vs. linear p < 0.0001 p < 0.0001 df=4 df=4 Discontinuous, multiple slopes vs. discontinuous, no slope p = 0.08 p = 0.13 df=3 df=3 Discontinuous, single slope vs. discontinuous, no slope p = 0.19 p = 0.69 df=1 df=1

demographic variables

demographic+SES variables

p < 0.0001 df=4

p = 0.006 df=4

p = 0.84 df=3

p = 0.8 df=3

p = 0.9 df=1

p = 0.41 df=1

1378

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

Fig. 2. Predicted relative risks for education using the linear model (thick grey line), the discontinuous multiple slopes model (thin black line), the discontinuous no slope model (thicker black line) and categorical variables (square markers) for men and women ages 25±64. Models have been adjusted for income, occupational status (for men), marital status, household size, employment status, age, and race.

rejected in favor of one of the alternative discontinuous models. For both sexes, neither the single slope nor the multiple slopes discontinuous models ®t signi®cantly better than the no slope discontinuous model in either the SES-demographic variables adjusted model or the model adjusted only for demographic variables. Thus, the no slope model would be chosen, on the basis of parsimony, as the model providing the best ®t. However, because the multiple slope model is approaching signi®cance in men, it is important to realize that the nonsigni®cant improvement in the ®t does not imply rejection of the multiple slopes model, but rather a lack of statistical evidence to reject the no slope model as a less complicated null hypothesis. Fig. 2 shows the predicted relative risk of mortality, adjusted for both SES and demographic variables, at each level of education according to the linear model, the multiple slopes discontinuous model, the no slope discontinuous model and the categorized education variables. The linear model does a particularly poor job of predicting relative risk for men because the

observed relationship between education and mortality is not monotonic. For income, the two slope model, with a break at US$22,500, ®ts better than the linear model for both sexes in both the model adjusted for SES and demographic variables model ( p < 0.0001 for men; p = 0.0005 for women) and the model adjusted for demographic variables only ( p < 0.0001 for both sexes). The two slope model shows the decrease in mortality associated with an increase in income of a given dollar amount to be signi®cantly greater at low levels of income than at higher levels of income. The ®tted values according to the best ®tting alternative models for education and income are depicted graphically in Fig. 3. Models adjusted for SES and demographic variables and for demographic variables only are both presented. The ®tted data in Fig. 3 show that men who have not attained a high school degree have a greater risk of mortality than men who have 12±15 years of education (high school diploma or more but no college

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

1379

Fig. 3. Predicted relative risks using the best ®tting models for education, and income for men and women aged 25±64. The relative risks are predicted using two models: one which adjusts for other socioeconomic variables, as well as demographic variables (marital status, household size, employment status, age, and race), and another model which adjusts only for the demographic variables. For education, the chosen model is the discontinuous Ð no slope model in both the model adjusted for only demographic variables (solid line with square markers) and the model adjusted for SES and demographic variables (solid line). The chosen model for income is the two slope model in both the model which is adjusted for only demographic variables (solid line with square markers) and in the model adjusted for SES and demographic variables (solid line).

degree) in both the SES-demographic variables adjusted model (RR=1.06) and the model adjusted for only demographic variables (RR=1.15). A much larger drop in mortality is associated with the attainment of a college degree. The relative risk of mortality for men with 13±15 years of education as compared to those with a college diploma or greater is 1.33 when adjusted for SES and demographic variables and 1.48 when adjusted for only demographic variables. The relative risk of mortality at the bottom of the educational scale compared to the top is 1.70 when adjusted for only demographic variables and 1.40 when also adjusted for SES. In contrast to men, the decrease in mortality associated with the attainment of high school and college diplomas is about equal for women (Fig. 3). The total e€ect of education for

women from top to bottom is 1.32 in the SES-demographic variables adjusted model and 1.48 when only adjusted for demographic variables. At incomes of under US$22,500, the relative risk of mortality associated with a US$10,000 decrease in income is 1.23 for men and 1.18 for women when adjusted for only demographic variables and 1.21 for men and 1.15 for women when adjusted for SES and the demographic variables. At incomes of over US$22,500, the relative risk associated with a US$10,000 decrease in income is only 1.07 for men and 1.04 for women when adjusted for only demographic variables and 1.02 for men and 1.03 for women when additionally adjusted for SES. The relative risk for an individual with an income of US$22,500 as compared to an individual with an

1380

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

income of US$60,000 is 1.27 for men and 1.17 for women when adjusted for only demographic variables and 1.07 for men and 1.11 for women when adjusted for SES in addition (Fig. 3). The relative risk of an individual with an income of US$2500 relative to one with an income of US$22,500 is 1.51 for men and 1.40 for women when adjusted for only demographic variables and 1.46 for men and 1.32 for women in the SES-demographic variables adjusted model (Fig. 3). The total top to bottom relative risk for incomes between US$60,000 and 2500 is 1.92 for men and 1.64 for women when adjusted for only demographic variables and 1.56 for men and 1.47 for women in the SES-demographic variables adjusted model (Fig. 3). Because income and education have di€erent functional forms, each might be a relatively more important predictor of mortality di€erentials among di€erent socioeconomic subgroups of the US population. Two large divergent subgroups will be used to exemplify. Consider two groups with congruent status for education and income, where the lower status group consists of respondents with incomes below US$22,500, and high school educations or less, while the upper status group consists of respondents with incomes above US$22,500, and greater than high school educations. The lower status group represents 28 and 49.9% of the population for men and women, respectively. The upper status group represents 26 and 15.3% of the population for men and women, respectively. For men in the lower status group, the relative risk associated with having an income of US$2500 vs. 22,500 is 1.46, which is signi®cantly greater than the relative risk of 1.06 associated with not having attained of a high school diploma (using the fully adjusted model for both education and income). In contrast, for the upper status group, the relative risk associated with the not having attained a college diploma, 1.33, is signi®cantly greater than the 1.07 relative risk associated with having an income of US$22,500 vs. 60,000. The e€ect of education and income can be considered in a casual framework by comparing the education coecients which are adjusted for only the demographic variables to the income coecients which adjusted for both demographic variables and SES. Viewed in a causal framework, the income e€ect is still much greater than the e€ect of earning a high school diploma in the lower status group and the e€ect of having a college degree is still greater than the e€ect of income in the upper status group. For women in the lower status group the relative risk associated with having an income of US$2500 vs. 22,500 is 1.32 and a relative risk of 1.15 is associated with the attainment of a high school diploma. For the higher status group a relative risk of 1.11 is associated with having an income of US$22,500 vs. 60,000 and a relative risk of 1.16 is associated with not having

attained of a college diploma. The di€erence between the two relative risks is not signi®cant. Thus, in contrast to men, the e€ect of income and education appears to be more equal for women in the high status group. As for men, the general conclusion reached when the variables are viewed in a causal framework is the same as when they are not. The within subgroup analysis above does not consider interactions between education and income. If strong interactions are present, the model would need to be modi®ed to include them. A test for interactions did not reveal any of consequence. However, the power to detect signi®cant e€ects for some interaction terms is low. Therefore, further research on interactions should be conducted, but would require very large data sets. Discussion The shape of the relationship between education and mortality appears to be well described by a trichotomy among people with less than high school educations, a high school diploma but no college degree, and those with a college degree or greater. The nonlinear shape is contrary to other data sources which hypothesize the shape to be more linear (Kitagawa and Hauser, 1973; Kuntz and Mackenbach, 1994). Some of these di€erences may be due to temporal (Kitagawa and Hauser, 1973) or international di€erences (Kuntz and Mackenbach, 1994) or because education was not adjusted for the same set of demographic control variables. Four possible explanations for the trichotomous relationship between education and mortality will be suggested. Firstly, the attainment of high school and college degrees may serve as markers for selective factors, which are innate at birth or established in early childhood or adolescence, which are related to both the ability to attain educational credentials and to health. Possible selective factors may be intelligence or innate ability, health in early childhood, and personality traits such as the ability to defer grati®cation. Secondly, the attainment of educational credentials may serve as a marker for acquired knowledge which may be used to secure health related outcomes. Thirdly, a person's relative status within society may be related to the attainment of a high school or college degree (Faia, 1981). Relative status in society may be related to a wide range of societal bene®ts such as job security, stimulating work, and sense of control and self-worth. Relative societal position, as marked by educational credentials, may also be an important factor in the acceptance or non-acceptance of certain lifestyles and behaviors. Therefore, socioeconomic di€erences in lifestyle and behaviors, which are

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

thought to be more strongly related to education than to other SES variables (Winkleby et al., 1990, 1992), may be delineated by educational credentials. Finally, the social networks of individuals may be partially determined along the lines of educational credentials. Social networks formed among people with similar educational credentials may provide a pool of common health related information and coping strategies which may be drawn upon by its members to secure positive health related outcomes. The shape of the mortality gradients for education and income appears to be generally similar for men and women although the size of the e€ects may di€er. One exception is that the risk of mortality for nonhigh school graduates, when compared to those with 12 to 15 years of education, is larger for women than for men (1.15 vs. 1.06), but the relative risk of those with 12 to 15 years of education compared to those with a college diploma or greater is larger for men (1.16 vs. 1.30). The result is similar to one observed in an analysis of death certi®cates from Michigan (Christenson and Johnson, 1995). Several factors may be responsible for the di€erence in the relationship between education and mortality for men and women. One possibility is that residual confounding (Kaufman et al., 1997) of education with income is greater for women than for men. For example, women comprise a larger percentage of the long term poor than men and recent studies suggest that the attainment of a high school diploma may be an important factor in escaping long term poverty for women (Bassuk et al., 1996). Because these permanently poor women cannot be identi®ed through income at baseline, their e€ect on mortality may be manifested through low levels of education. A second possibility is that mortality in women has been shown to be more closely related to the social status of their husbands than to their own social status (Arber, 1987), and education describes the social status only of the women themselves while family income describes the social status of both the women and their husbands. A third possibility is that returns in terms of income and status, from obtaining a college degree may, be smaller for women than for men (Krieger et al., 1993). A ®nal possibility is that even though obtaining a college degree may provide returns, in terms of increased income and status for women, this e€ect might be counterbalanced by other stressors; such as maintaining both a demanding job and an acceptable family relationship or sexism on the job and the glass ceiling (Krieger et al., 1993). For men, the NLMS data suggests that the relative net contribution of education and income to socioeconomic mortality di€erentials is di€erent at di€erent points along the socioeconomic mortality gradient. Therefore, any group of intermediate risk factors that

1381

is responsible for the entire gradient would likely be strongly related to income in the lower portion of the gradient and strongly related to the attainment of a college degree in the upper portion. For example, if a group of behaviors and lifestyles were hypothesized to be responsible for the gradient, the case would be strengthened if the group had a strong net relationship with income in the lower portion of the gradient and strong net relationship with the attainment of a college degree in the upper portion. The fact that two di€erent components of SES are primarily responsible for the gradient at di€erent points along its scale may also suggest di€erent causal mechanisms exist in di€erent portions of the gradient. Income is closely associated with the class or economic component of the Weberian concept of social class and represents ®nancial security and the opportunity to ful®ll material desires. Some material limitations; such as limited access to health care, poor nutrition, exposure to environmental dangers (e.g. deteriorating housing and hazardous chemicals) and hazardous social environments may a€ect health directly (Dutton and Levine, 1989; Haan et al., 1989; Adler et al., 1993). Indirectly, lack of income produces physical and social constraints which e€ect the psychological well being and social networks of the individual (Williams, 1990; Wilkinson, 1996). Psycho-social constraints may subsequently e€ect the individual's health attitudes and behaviors (Williams, 1990). Income also e€ects the individual's perceived rank in society, which, again, may have psycho-social e€ects which ultimately a€ect behavior (Wilkinson, 1989, 1996). In contrast, education is often thought to represent acquired knowledge, the ability and opportunity to relate to a complex environment, and skill in using the social system to meet desired ends. Education is also thought to be closely associated with the status component of the Weberian concept of social class. Like income, education also a€ects an individual's psychosocial environment, social networks and perceived status. The NLMS data may suggest that, for both men and women, the economic component of social status may be the aspect of social status which is most indicative of health up to a certain critical point. In men, however after the critical point is reached, mortality is more strongly associated with acquired knowledge, societal status and other bene®ts related to the attainment of a college degree than with further increases in economic bene®ts. The NLMS data may also suggest that a minimal material standard is necessary to maximize individual health. The analyses presented su€er from several limitations. Inferences which can be made about causal pathways from multivariate regression models are limited. The formulation of multistage models is necessary

1382

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

to fully describe interrelationships between SES, intermediate variables and mortality (Graber, 1989; Haan et al., 1989). Also, the relationship between SES and intermediate risk factors is poorly understood. The manner in which education and income are related to even their most basic attributes, such as di€erential material conditions and di€erential cognitive skills, is not well understood and constantly changing. For example, educational attainment is becoming increasingly homogeneous with each successive birth cohort. As a high school diploma becomes the societal norm, it may be possible that income serves as a better marker for cognitive skills than education, especially at the low end of the educational scale. The e€ect of income may be overstated relative to education, because the temporality of its association with mortality cannot be unequivocally determined, a phenomenon known as reverse causality (Fox et al., 1985; Kitagawa and Hauser, 1973; Backlund et al., 1996; Elo and Preston, 1996). In this study, health related selection out of the labor force, which some researchers believe is responsible for a large portion of the reverse causal e€ect (Fox et al., 1985), has been controlled for by adjusting for employment status. However, some e€ect of reverse causality may remain because reductions in income from reduced work hours and selection into lower paying jobs due to illness (Fox et al., 1985) cannot be controlled for. In contrast, since income is only available in the year prior to baseline, its e€ect may be understated relative to that of education because the income variable does not properly re¯ect the instability of income over time (Liberatos et al., 1988). Particularly troublesome is the inability of a single measure of income to separate the transitory poor from the permanently poor. Recent studies suggest that long term poverty is an important predictor of mortality (McDonough et al., 1997). Measures of wealth, which are not available on the NLMS data set, have also been shown to be important predictors of mortality independent of income (Menchek, 1993). Since income and wealth both represent material aspects of social status, the relationship between material well-being and mortality may be understated in this study. Two problems mentioned in the comparisons of e€ects of men and women, residual confounding (Kaufman et al., 1997), and that family level and individual level SES variables may have di€erent meanings (Krieger et al., 1993), may also cause general interpretive problems. Also, income has not been adjusted to consider poverty levels for di€erent family sizes as some researchers have suggested (Krieger et al., 1993). Determining the functional form of a statistical relationship requires a very large amount of data, therefore as large as the NLMS is, it is limited by its `small' size. The number of deaths in the NLMS is too limited

to estimate the functional forms for either income or education separately for blacks or other races with any con®dence. A further weakness of the study is that it has not dealt with the older population. Previous research, from the NLMS, has shown that the shape of the relationship between income and mortality for the elderly appears to follow the same shape as that of the younger population, although its magnitude is smaller (Backlund et al., 1996). However, the exact shape of the relationship between education and mortality for the elderly remains a topic of future research. Therefore, the results presented in the manuscript only apply to premature mortality occurring between the ages of 25 and 64. The functional forms presented here should be regarded as approximate. For example, it can be stated with con®dence that the slope is attenuated in the upper portion of the income gradient, but the point where the change in slope occurs and the nature of this change are uncertain. The change in slope is almost certainly not as abrupt as the model suggests. Nonetheless, the two slope model for income has been shown to be robust to moderate changes in the point of the slope change and has also been shown to ®t better than several mathematical functions of income (Backlund et al., 1996). However, the presented comparisons made within the low and high status groups may be sensitive to the point of the slope change used for income. A sensitivity analysis was preformed by moving the point of the slope change for income to either US$17,500 or 37,500. For men, both changes resulted in the same general conclusion that income is a more important predictor of mortality in the low status group while the attainment of a college diploma is a more important predictor of mortality in the high status group. For women, moving the point of the slope change to US$17,500 did not change the conclusions presented in this study. However, moving the slope change to US$37,500 resulted in education (the attainment of college a degree) being more important than income in the high status group instead of income and education being of about equal importance. The conclusion reached for the low status group did not change. The true functional forms of the mortality gradient for education and income may well be more complex than the simple alternatives model presented in this study. However, because mortality among the working age population in NLMS is limited and because of the limited number of categories for income, functional forms cannot explored in their full complexity. Therefore, generalized additive models (Hastie and Tibshirani, 1990), which examine the functional form of a statistical relationship without appealing to a predetermined alternative hypothesis, have not been uti-

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

lized. The authors certainly applaud e€orts to explore the functional form of the relationship between SES and health using generalized additive models (Kaufman et al., 1998) and fully realize these e€orts may reveal complexities of the functional forms that may eventually supplant the simpler alternatives that have been hypothesized here. Thus, the functional forms presented should not necessarily be interpreted as the `correct' functional form, but rather as an alternative functional form for which there is signi®cant statistical evidence to prefer over a simpler functional form. This report has suggested some hypotheses concerning the nature of the socioeconomic mortality gradient based on the comparative functional forms of education and income. Further research on how di€erent types intermediate variables are jointly related to education and income may help to further elucidate the interpretation of the comparative functional forms. The functional forms presented in this manuscript may also need to be re®ned as more mortality data becomes available. Research also needs to be conducted on di€erences in functional forms among more speci®c age groups and di€erent racial and ethnic groups. Hopefully, a better understanding of the functional forms of the relationship between di€erent components of socioeconomic status and mortality combined with greater understanding of how the di€erent components act as markers for di€erent groups of intermediate variables, will provide further insight into the causes of the socioeconomic mortality gradient.

Acknowledgements The authors would like to thank Charles Alexander and Jay Kaufman for their valuable comments.

References Adler, N.E., Boyce, T., Chesney, M.A., Folkman, S., Syme, L., 1993. Socioeconomic inequalities in health: no easy solution. Journal of the American Medical Association 269 (24), 3140±3145. Adler, N.E., Boyce, T., Chesney, M.A., Cohen, S., Folkman, S., Kahn, R.L., Syme, S.L., 1994. Socioeconomic status and health: the challenge of the gradient. American Psychologist 49 (1), 15±24. Anderson, R.T., Sorlie, P.D., Backlund, E., Johnson, N.J., Kaplan, G., 1997. Individual and area-based indicators of socioeconomic status and mortality: the national longitudinal mortality study. Epidemiology 8, 42±47. Antonovsky, A., 1967. Social class, life expectancy, and overall mortality. Milbank Quarterly 45, 31±73. Arber, S., 1987. Social class, non-employment, and chronic ill-

1383

ness: continuing the inequalities in health debate. British Medical Journal 294, 1069±1073. Backlund, E., Sorlie, P.D., Johnson, N.J., 1996. The shape of the relationship between income and mortality in the United States: evidence from the national longitudinal mortality study. Annals of Epidemiology 6 (1), 12±20. Bassuk, E.L., Browne, A., Buckner, J.C., 1996. Single mothers and welfare. Scienti®c American 275 (4), 60±67. Christenson, B.A., Johnson, N.E., 1995. Educational inequality in adult mortality: an assessment with death certi®cate data from Michigan. Demography 32 (2), 215±229. Collet, D., 1994. In: Modeling Survival Data in Medical Research. Chapman & Hall, London, pp. 53±106. Duleep, H.O., 1995. Mortality and income inequality among economically developed countries. Social Security Bulletin 58 (2), 34±50. Dutton, D.B., Levine, S., 1989. Socioeconomic status and health: overview, methodological critique, and reformulation. In: Bunker, J.P., Gomby, D.S., Kehrer, B.H. (Eds.), Pathways to Health. The Henry Kaiser Family Foundation, Menlo Park, CA, pp. 29±69. Elo, I.T., Preston, S.H., 1996. Educational di€erentials in mortality: United States, 1979±85. Social Science & Medicine 42 (1), 47±57. Faia, M.A., 1981. Selection by certi®cate: a neglected variable in strati®cation research. American Journal of Sociology 86, 1093±1111. Feinstien, J.S., 1993. The relationship between socioeconomic status and health: a review of the literature. Milbank Quarterly 71 (2), 279±322. Feldman, J.J., Makuc, D.M., Kleinman, J.C., CornoniHuntley, J., 1989. National trends in educational di€erentials in mortality. American Journal of Epidemiology 129 (5), 919±933. Fliess, J.L., 1981. In: Statistical Methods for Rates and Proportions, 2nd ed. John Wiley and Sons, New York, pp. 237±255. Fox, A.J., Goldblatt, P.O., Jones, D.R., 1985. Social class mortality di€erentials: artefact, selection, or life circumstances. Journal of Epidemiology and Community Health 39 (1), 1±8. Graber, A.L., 1989. Pursuing the links between socioeconomic factors and health: critique, policy, implications, and directions for future research. In: Bunker, J.P., Gomby, D.S., Kehrer, B.H. (Eds.), Pathways to Health. The Henry Kaiser Family Foundation, Menlo Park, CA, pp. 271±315. Haan, M.N., Kaplan, G.A., Syme, S.L., 1989. Socioeconomic status and health: old observations and new thoughts. In: Bunker, J.P., Gomby, D.S., Kehrer, B.H. (Eds.), Pathways to Health. The Henry Kaiser Family Foundation, Menlo Park, CA, pp. 76±135. Hastie, T.J., Tibshirani, R.J., 1990. Generalized Additive Models. Chapman & Hall, London. House, J.S., Kessler, R.C., Herzog, A.R., 1990. Age, socioeconomic status, and health. Milbank Quarterly 68, 383±411. Kalb¯eisch, J.D., Prentice, R.L., 1980. In: The Statistical Analysis of Failure Time Data. John Wiley and Sons, New York, pp. 70±142. Kaufman, J.S., Cooper, R.S., McGee, D.L., 1997. Socioeconomic status and health in blacks and whites: the

1384

E. Backlund et al. / Social Science & Medicine 49 (1999) 1373±1384

problem of residual confounding and the resiliency of race. Epidemiology 8, 621±628. Kaufman, J.S., Long, A.E., Liao, Y., Cooper, R.S., McGee, D.L., 1998. The relationship between income and mortality in US blacks and whites: the national health interview survey (1986±1990). Epidemiology 9 (2), 147±155. Kitagawa, E.M., Hauser, P.M., 1973. Di€erential Mortality in the United States. Harvard University Press, Cambridge, MA. Krieger, N., Rowley, D.L., Herman, A.A., Avery, B., Phillips, M.T., 1993. Racism, sexism, and social class: implications for studies of health disease and well-being. American Journal of Preventive Medicine 9, 82±122. Kuntz, A.E., Mackenbach, J.P., 1994. The size of mortality di€erences associated with educational level in nine industrialized countries. American Journal of Public Health 84, 932±937. Liberatos, P., Link, B.G., Kelsey, J.L., 1988. The measurement of social class in epidemiology. Epidemiological Reviews 10, 87±121. Matthews, K.A., Kelsey, S.F., Meilahn, E.N., Lewis, K.H., Wing, R.R., 1989. Educational attainment and behavioral and biologic risk factors and coronary heart disease in middle-aged women. American Journal of Epidemiology 129, 1132±1144. McDonough, P., Duncan, G.J., Williams, D., House, J., 1997. Income dynamics and adult mortality in the United States, 1979 through 1989. American Journal of Public Health 87, 1476±1483. Menchek, P., 1993. Economic status as a determinant of mortality among older black and white men: does poverty kill? Population Studies 47, 427±436. Miller, D., 1991. In: Handbook of Research Design and Social Measurement. Sage Publications, Beverly Hills, pp. 323±365. National Center for Health Statistics, 1990. User's manual the National Death Index. VSOHHS, Public Health Service, Hyattsville, MD, DHOWS Publication No. (PHS) 901148. Neter, J., Wasserman, W., Kutner, H.K., 1985. Applied Linear Statistical Models, 2nd ed. Irwin, Homewood, IL, pp. 343±351. Rogot, E., Sorlie, P.D., Johnson, N.J., Glover, C.S., Treasure, D., 1988. A mortality study of one million persons by demographic, social, and economic factors: 1979± 1981 folow-up. NIH publication 88-2896. Bethesda. Rogot, E., Sorlie, P.D., Johnson, N.J., Schmitt, C., 1992. A mortality study of 1.3 million persons by demographic, social, and economic factors: 1979±1985 folow-up. NIH publication 92-3297. Bethesda.

Rose, G., Marmot, M.G., 1981. Social class and coronary heart disease. British Heart Journal 45, 13±19. Sorlie, P.D., Backlund, E., Kehler, J., 1995. Mortality by economic, demographic, and social characteristics; the national longitudinal mortality study. American Journal of Public Health 85, 949±956. Stampfer, M.J., Willet, W.C., Speizer, F.E., 1984. Test of the National Death Index. American Journal of Epidemiology 119, 837±839. Syme, S.L., Berkman, L.F., 1976. Social class, susceptibility and sickness. American Journal of Epidemiology 104 (1), 1±8. Townsend, P., Davidson, N., Whitehead, M., 1988. Inequalities in Health. Penguin, London. US Bureau of the Census, 1978. The current population survey: design and methodology. In: Technical Paper, 40. US Department of Commerce, Washington, DC. Wentworth, D.N., Neaton, J.D., Rasmussen, W.L., 1983. An evaluation of the social security administration master bene®ciary record ®le and the national death index in the ascertainment of vital status. American Journal of Public Health 73, 1270±1274. Wilkinson, R.G., 1986. Income and mortality. In: Wilkinson, R.G. (Ed.), Class and Health: Research and Longitudinal Data. Tavistock, London, pp. 88±114. Wilkinson, R.G., 1989. Class mortality di€erentials, income distribution and trends in poverty 1921±1981. Journal of Social Policy 18 (3), 307±335. Wilkinson, R.G., 1996. Unhealthy Societies: The A‚ictions of Inequality. Routledge, London. Winkleby, M.A., Fortmann, S., Barret, D., 1990. Social class disparities in risk factors for disease: eight year prevalence patterns by level of education. Preventive Medicine 19, 1± 12. Winkleby, M.A., Jatulis, D.E., Frank, E., Fortmann, S., 1992. Socioeconomic status and health: how education, income, and occupation contribute to risk factors for cardiovascular disease. American Journal of Public Health 82, 816± 820. Williams, B.C., Demitrack, L.B., Fries, B.E., 1992. The accuracy of the national death index when personal identi®ers other than social security number are used. American Journal of Public Health 82, 1145±1147. Williams, D.R., 1990. Socioeconomic di€erentials in health: a review and redirection. Social Psychology Quarterly 53 (2), 81±99. Williams, D.R., Collins, C., 1995. US socioeconomic and racial di€erences in health. Annual Review of Sociology 21, 349±386.