Educational differentials in mortality: United States, 1979–1985

~ Pergamon 0277-9536(95)00062-3

Soc. Sci. Med. Vol. 42, No. 1, pp. 47-57, 1996 Copyright © 1996 Elsevier Science Ltd

Printed in Great Britain. All rights reserved 0277-9536/96 $15.00 + 0.00

EDUCATIONAL DIFFERENTIALS IN MORTALITY: UNITED STATES, 1979-85

IRMA T. ELO and SAMUEL H. PRESTON Population Studies Center, University of Pennsylvania, 3718 Locust Walk, Philadelphia, PA 19104, U.S.A.

Abstract--The paper examines educational differentials in adult mortality in the United States within a multivariate framework using data from the National Longitudinal Mortality Survey (NLMS). As a preliminary step we compare the magnitude of educational mortality differentials in the United States to those documented in Europe. At ages 35-54, the proportionate reductions in mortality for each one year increase in schooling are similar in the United States to those documented in Europe. The analyses further reveal significant educational differentials in U.S. mortality among both men and women in the early 1980s. Differentials are larger for men and for working ages than for women and persons age 65 and above. These differentials persist but are reduced in magnitude when controls for income, marital status and current place of residence are introduced.

Key words--inequalities in mortality, education, socioeconomic status, race

Investigations of social class differences in mortality have nearly as long a history as measures of mortality themselves. Much of the interest in such differences reflects social concern with the distribution, and not only with the mean level, of socially-valued outcomes such as health, longevity and income. In this paper, we focus on an individual's educational attainment as the primary marker of socioeconomic status. This measure has become the basic indicator used for mortality and health studies in both demography and epidemiology in the United States. Kitagawa and Hauser's [1] major study of socioeconomic mortality differentials in the United States in 1960 used educational attainment as its principal measure, displacing the occupational groupings that had been the focus of the classic studies of socioeconomic mortality differentials by the Registrar General in England and Wales [2].

There are two basic analytic reasons for preferring educational attainment to other common markers of social standing such as occupation or income. First, educational level can be determined for all individuals, whereas not everyone has an occupation or an income. The distinction has been especially important for women and, in the case of occupation, for persons above age 65. Second, one's educational attainment is essentially unaffected by health impairments that may emerge in adulthood. These impairments can be an important influence on one's occupation and level of income and are, of course, also predictive of excess mortality. This specter of 'reverse causation' besets any attempt to provide a causal interpretation of associations between healthiness and income or occupation (see also Refs [3,4]). Educational attainment is essentially established in early adulthood and remains stable thereafter.

47

Like occupation and income, educational attainment is a composite socioeconomic variable, in the sense that it reflects a number of influences on health status and mortality. Most directly, it affects potential earnings and thus access to material resources such as diet, housing and health care services that influence health. It is also related to cognitive abilities and informational resources that influence an individual's abilities to secure valued outcomes such as healthiness and a longer life. Winkleby et al. [5] show that adult health behaviors such as smoking and exercise are more closely associated with educational attainment than they are with income or occupation. Educational attainment also reflects features of one's family and community of origin and is related to an individual's personality traits.

This paper examines educational differentials in adult mortality in the United States, and uses a recently-released data set, the National Longitudinal Mortality Survey (NLMS). It begins with a discussion of the magnitude of educational mortality differentials in the United States in 1979-1985. It then examines these differentials along with a broader set of demographic and socioeconomic variables within a multivariate framework. Since this is the first presentation of multivariate results for the United States for many of the other characteristics considered, we also briefly discuss their influence on mortality.

DATA The data used in this study come from the National

Longitudinal Mortality Survey (NLMS), a mortality follow-up study funded and directed by the National Heart, Lung, and Blood Institute. The NLMS Public Use Sample, the source of our data, is based on five Current Population Surveys (CPS), conducted

48 Irma T. Elo and Samuel H. Preston

Table 1. Death rates (per 1000) by educational attainment, United States 1979-1985, males and females aged 25--64 and 65-89

Males Females Age and year of Death rate Mortality Death rate Mortality school completed (per 1000) ratio index (per 1000) ratio index

25-64 0-8 years 7.41 149 3.69 151 9-11 years 7.05 142 3.51 143 12 years 4.96 100 2.45 100 13-15 years 4.71 95 2.55 104 16+ years 3.33 67 2.06 84

65-89 0--7 years 59.29 111 32.77 110 8 years 61.91 116 30.05 101 9-11 years 59.20 111 29.21 98 12 years 53.42 100 29.66 100 13-15 years 49.04 92 25.89 87 16+ years 40.66 76 23.80 80

Source: Public use sample of the National Longitudinal Mortality Survey. Age-adjusted death rates were calculated from age-specific death rates within five-year age groups with the total U.S. population on 1 July 1983 as the standard.

between M a r c h 1979 and M a r c h 1981, and conta ins 637,324 individual records which have been linked to the Nat iona l Dea th Index (NDI) for the years 1979-1985.* This record linkage identified 22,649 deaths tha t had occurred within a five-year period following the date of the CPS interview to members of the five CPS cohor ts (for details on the l inkage procedures, see Ref. [6].t The N L M S is unquest ion- ably the largest and mos t reliable source o f data for investigating recent socioeconomic differentials in Amer ican morta l i ty [8].

The N L M S data on socioeconomic and demographic characterist ics come f rom the interview da ta f rom the CPS surveys, in which the interviewer

*The NLMS Public Use Data File is a subset of the larger NLMS database consisting of 12 census samples numbering about 1.3 million persons in the United States. Eleven of the 12 cohorts were taken from Current Population Surveys conducted during the period from March 1973 through March 1985 with one sample drawn from the 1980 Census of Population. Sample individuals were then matched to the NDI beginning in 1979, when the NDI was established, with plans to continue mortality follow-up through 1993 [7].

tThe five CPS surveys were conducted in March 1979, April, August and December 1980, and March 1981. For all sample members follow-up in days is included in the Public Use Data File and all individuals who were not linked to the NDI, and thus considered to be alive at the end of the follow-up, are given a follow-up period of 1827 days (five years). We should note, however, that the March 1981 CPS cohort was followed only through the end of 1985, or approximately four years and nine and a half months. We cannot distinguish which sample individuals belong to this cohort and must accept a five-year follow-up period for them as well. In addition, the lack of perfect ascertainment of death in the NDI results in some deaths being missed. Rogot et al. [7] (p. 2) suggest that "there is some ascertainment loss, of perhaps 5%, occurring in the matching process because of recording errors in the files being matched".

:]:Individuals for whom education or race was missing are excluded from the calculations. The estimates presented are based on unweighted cases; the use of weights would not alter our conclusions.

contacts , by personal interview or th rough the telephone, the mos t knowledgeable adul t member of the household who provides informat ion on all household members . All baseline socioeconomic and demographic characterist ics per ta in to the t ime of the CPS interview. We present results separately for the age ranges 25-64 and 65-89 at the initial interview, a division tha t separates working ages f rom ret i rement ages.

Educational mortality differentials

Table 1 presents age-sex educat ion specific dea th rates in the Uni ted States for the age intervals 25-64 and 65-89. Co lumns 3 and 5 convert these rates to index numbers with the morta l i ty level of high school graduates assigned a value 100. The dea th rates are directly age s tandardized to the dis t r ibut ion of the total U.S. popu la t ion on 1 July 1983 within five-year age groups [9].~ An individual 's educat ional a t ta inment (i.e. n u m b e r of years of school completed) is classified into five ~ategories in the age interval 25-64 and six categorieo in the age interval 65-89 to take account of the differences in the dis t r ibut ion of educat ional a t t a inment in the two age intervals. In part icular , we expanded the 0-8 category for older persons because many more of them did not go beyond 8 years of schooling. At the same time, our classifications dist inguish a m o n g impor t an t educat ional markers , such as comple t ion o f high school and four or more years of college.

Educat ional a t t a inment clearly bears a close relat ionship to adul t mortal i ty. In all cases, college graduates have lower morta l i ty than high school graduates. Wi th one exception, individuals who did not complete high school have higher morta l i ty t han high school graduates. The cont ras t between the two extreme groups is largest for working-age men; the rat io o f dea th rates in the extreme educat ional categories is 2.22 for working-age males compared to 1.79 for working age females. The rat io among older males (1.46) is also somewhat more p ronounced t han

Educational differentials in mortality 49

Table 2. Inequality coefficients (percentage reduction of mortality per one year increase in education) by sex and country, persons aged 35-54

Country and Period Males Females

Denmark, 1971-80 8.1 (5.8)" 3.8 (2.9)" Finland. 1971-80 9.2 6.1 England & Wales, 1971-81 (April) 7.4 (7.4)" 7.8 (7.2)' Hungary, 1978-81 8.2 2.2 Norway, 1971-80 8.6 5.8 Sweden, 1971-80 8.0 (7.7)- 4.8 (4.7)" United States, 1979-85 7.7 7.3

• Inequality coefficient in parentheses estimated on a sample that excludes cases for which education was not reported.

Source: Valkonen [13], Table 7.2 for all causes of death; calculations for the United States are based on the NLMS Public Use Sample.

among older females (1.3 8), al though in both cases the contrast is less severe than during working ages.* The narrowing of differentials as age advances is consistent with widely-observed patterns. Models of age-patterns of mortali ty constructed from populations at different levels of life expectancy at birth show that

*Our confidence in these estimates is increased by the fact that educational differentials in mortality based on the NLMS are highly consistent with recent estimates by Feldman et al. [10] for the period 1971-1984, which are also based on a longitudinal mortality follow-up, the National Health and Nutrition Survey Epidemiologic Follow-up Study. An alternative set of estimates by Pappas et al. [11], on the other hand, show somewhat larger educational differentials in mortality than those documented here in the age range 25-64, but the data used by Pappas et al. are subject to numerator-denominator bias (see Ref. [8]).

tTo estimate the 'inequality coefficient' for each of the countries included in his analyses, Valkonen [13] fitted the logarithm of the number of deaths as a linear function of education (a continuous variable) and the categorical variables for age and period (and their interactions when data for two periods were available) plus the logarithm of person years of exposure with the statistical package GLIM. The input to the model is numbers of deaths and exposures aggregated according to the values of the covariates (for further details see Valkonen [13]). He then calculated the inequality coefficient as follows: I = 100[1 - exp(fl)], where # is the estimated coefficient on education. We have replicated Valkonen's method- ology with data from the NLMS to compute an equivalent 'inequality coefficient' for the United States. In our estimates deaths and person-years lived are counted cohort-wise for three 5-year cohorts aged 35-49 at the date of the interview, making the data structure comparable to Valkonen's estimates for the Nordic countries (see Valkonen [13], pp. 143-144). To linearize the categorical education variable in the NLMS we assigned the midpoint of an educational category as follows (assignments are shown in parentheses): none, 0-4 years (2); 5-7 years (6); 8 years (8); 9-11 years (10); 12 years (12); 13-15 years (14); 16 years (16); and 17 + years (I 8). Individuals for whom information on education was missing are excluded from these calculations (education was missing for 770 cases or 0.7% of 103,530 individuals aged 35-49 at the time of the interview). Our sample for these calculations includes individuals of all races. An alternative coding of education, collapsing the three lowest levels of schooling (estimated number of years of education for the combined category is 6) and the two highest education categories (estimated number of years of education for the combined category is 17), would not alter our substantive conclusions; the inequality coefficient for males is 8.0 and 7.4 for females.

proport ionate differences in adult age-specific death rates typically diminish with age [12].

Mortali ty differentials by education in the United States appear similar to those in other industrialized countries. Valkonen [13] has carefully analyzed educational differentials in mortality for persons aged 35-54 in the Nordic countries and England and Wales for two periods 1971-75 and 1976-80 and in Hungary for the period 1978-81. These analyses are also based on data from linked census-death certificate files, except for Hungary where death certificate and census data were not linked. We have replicated Valkonen's statistical procedures on the N L M S data. Results are presented in Table 2. These comparisons rely on an ' inequality coefficient': the coefficient measuring the slope of the regression line of the logarithm of the death rates on years of education. t This measure shows the estimated proport ionate reduction in mortality for each one year increase in the level of schooling.

The results show a remarkable consistency in the effects of education on male mortality. The male coefficients range only from 7.4 to 9.2 across the countries with the U.S. coefficient (7.7) falling towards the lower end of this range (see also Kunst and Mackenbach [14]). The results reported in Table 2 show a much greater dispersion of coefficients for females (ranging from 2.2 to 7.8) and the coefficient for the U.S. (7.3) is higher than in any other country except England and Wales, where the estimates for females appear sensitive to the treatment of cases for which education was missing and are fitted using only two distinct levels of schooling. Male inequality in mortality by educational attainment exceeds female inequality in all countries except England and Wales. The male excess in coefficients is smaller in the United States than in any country except England and Wales. But in general, there is nothing in these results that make U.S. differentials measured by the N L M S appear anomalous. Since access to health care is probably more unequally distributed in the U.S. than in any of the other countries, the similarity of its educational effects on mortality to those found in Europe is indirect evidence that such access may not be a highly important determinant of mortality.

Multivariate results

We now turn to an examination of educational differentials in mortali ty in the United States within a multivariate framework. We distinguish among demographic and socioeconomic characteristics that are acquired at bir th--race, region of birth and year of birth (age)---and those that are evaluated at the date of the initial interview: family income, marital status, metropoli tan residence and number of household members. Al though occupation is frequently em- ployed in studies of socioeconomic differentials in mortality in Europe, we do not include occupation in

50 Irma T. EIo and Samuel H. Preston

these analyses for reasons noted above.* We do, however, include a measure of income because of its importance for understanding the role of economic circumstances in mortality. The difficulties posed by the possibility of reverse causation (i.e. joint dependence of the variables on unmeasured health status) are partially mitigated by the following:

(A) We exclude from our sample all persons who were reported as being unable to participate in the labor force due to their own long-term physical or mental illness. These individuals typically have below-average incomes and die at unusually rapid rates [15, 16]; their inclusion would bias estimates of the effect of income on mortality. This category does not, however, exclude individuals who are temporarily ill or disabled or who expect to return to work within six months [7] (pp. 13-14).

(B) Income pertains to family income rather than to personal income, so it is less responsive to health impairments of the individual whose survival is being tracked.i"

*In the age range 25-89, for example, 20.6% of the males and 47.2% of the females fall in the category 'occupation not reported, or never worked'.

tAnnual family income is obtained in response to the following question: Which category in this card represents the total combined income of all members of this family during the past 12 months? This includes money income from jobs, net income from business, farm or rent, pensions, dividends, interest, social security payments and any other money income received by members of this family who are 14 years of age or older? A flashcard is shown to the respondent who chooses from the shown categories. The above categorical response to family income was obtained in April, August and December 1980 CPSs. For members of these cohorts all members of a household were given the income of the family of the reference person. For March 1979 and March 1981 CPSs the actual dollar amount of income in previous calendar year for each family in a household was determined [7] (pp. 11-12).

:~We also examined a linear specification of the income covariate. We elected to include (In income) in our models because this specification explained somewhat more of the variance in Model 3 in all four age-sex groups than when a linear term was used.

lIIndividuals for whom Hispanic origin information was missing were coded as non-Hispanics.

§The weights included in the Public Use Data File give projected population totals for the noninstitutionalized population of the United States on 1 April 1980. A total of 13,296 persons in the entire NLMS Public Use File could not be assigned weights for various reasons [20]. In our male sample, at the age range of interest, weights were missing for 1904 cases and in the female sample for 468 cases. The impact of sample weights was studied by comparing weighted and unweighted estimates of coefficients. The two sets were similar and do not change the substantive interpretation of the results. Standard errors for the coefficients were adjusted to account for the fact that more than one person from the same household could be included in the sample. The adjustments were based on the Huber formula that is incorporated into the hlogit procedure in STATA [21, 22].

(C) Income is recorded for the year preceding the beginning of the five-year follow-up, so that it is not influenced by health impairments that develop subsequently.

Annual income, adjusted to 1980 dollars by the Consumer Price Index, is given in the following categories: < $5000, $5000-9999, $10,000-14,999, $15,000-19,999, $20,000-24,999, $25,000-49,999 and $50,000 + , unknown. In order to treat income as a single variable in our models, we assign the dollar amount of the midpoint of each of the income category, e.g. $2500, $7500 etc., and then include the natural log of income as a covariate in our models.:~ Since family size is directly related to family income as well as to the demands placed on that income, we also control for the number of household members.

We employ a three-category racial variable distinguishing between white, black and other races (American Indian or Eskimo, Asian or Pacific Islander and other nonwhite). Place of birth distinguishes among individuals born in the four U.S. census regions, Northeast, South, North Central and West, and those born outside of the United States, among whom we further differentiate between those who are of Hispanic-origin and those who are not.II In addition, we include marital status and residential type (not in a metropolitan area, metropolitan-area central-city residence and metropolitan-area non-central-city residence) as covariates.

We have no data on a wealth of variables that have been shown to affect death rates. Some of these are correlated with variables that are included in the model. For example, the prevalence of cigarette smoking varies inversely with educational at tainment [17]. Because the data set is limited to socioeconomic and demographic variables, we are not able to elucidate the set of biomedical and behavioral variables through which they operate to influence death rates. What we are doing amounts to asking whether those variables are more closely associated with educational at tainment than, say, with income.

Methods

To assess the relative effects of the demographic and socioeconomic variables of interest on male and female mortality, we estimate a logistic regression model that can be expressed as follows:

ln[sq.,/(1 - sqx)] = a + rA + #X (1)

where sqx is the probability of dying during the five-year follow-up period; a is the intercept representing the log odds of dying for individuals for whom the values of all dummy covariates equal zero; r is a coefficient on the age covariate, A, (entered in its linear form); and fl is the vector of coefficients of the other covariates, X. To estimate our models, we use the maximum likelihood estimation method under the logit procedure in STATA [18] and show results based on unweighted regressions.§ We should note that in

Educational differentials in mortality

Table 3. Coefficients of equations predicting the log odds of dying in a five-year period: United States, 1979-1985

Males Females Characteristic Age 25-64 Age 65-89 Age 25--64 Age 65-89

Age 0.0881"* 0.0891"* 0.0836** 0.0913"*

Race White . . . . Black 0.2715"* -0.1200 0.3733** 0.0273 Other 0.1366 -0.4081"* 0.2353 -0.6244**

Region qfb#th Northeast . . . . South -0.0350 -0.0743 -0.0171 -0.1061" North Central -0.0670 -0.0457 -0.0378 -0.1515"* West -0.1439" -0.2204** -0.0434 -0.1489" Outside U.S.-Hisp. -0.5083** ---0.9268** -0.9824** -0.4428** Outside U.S.-other -0.5588** -0.3012"* -0.3331"* -0.2081"*

Years ofschool comp&ted 0-7 years 0.0675 0.0611 8 years 0.1414'** 0.1255"* 0.2341'** 0.0114 9-11 years 0.1469"* 0.1195" 0.1349" -0.0122 12 years . . . . 13-15 years 0.0262 -0.0588 0.0479 -0.1224" 16+ years -0.2541"* -0.2200** -0.0891 -0.1832"*

Log ofincome -0.3136"* -0.1278"* -0.2096** -0.0421"

Household size -0.0179 0.0558** -0.0234 0.0586**

Marital slatus Married . . . . Widowed 0.3877** 0.1848"* 0.2246** 0.1486"* Divorced/separated 0.3474** 0.3611"* 0.1847"* 0.4085** Never married 0.3849** 0.1856"* 0.2906** 0.1223

Residence Not a metro area Centra City (CC) 0.1127** Metro area, non-CC 0.0700

Constant -4.6514"*

Log-likelihood - 15448.2 X 2 = 738.6 df= 22

Sample size 134,835

0.0529 0.1289"* 0.0942** 0.0191 0.1386"* 0.0463

-6.5427** -6.2057** -8.2402**

-13439.1 -11021.3 -14198.8 X 2 = 249.4 X 2 = 337.9 X 2 = 128.7 df= 23 df= 21 df= 23

25,270 147,457 35,231

"Coefficient pertains to educational level 0-8 years of schooling. --Reference category. **P-value _<0.01. *P-value _<0.05. Note: Coefficients for categories with missing values on a characteristic not shown. For females aged

25~o4. individuals with missing value for residence are included in the reference category. 7~: obtained by subtracting the deviance of the current model from a model that controls only for age.

51

ou r m ode l s we have no t con t ro l l ed for u n o b s e r v e d he te rogene i ty , i.e. for the fact tha t indiv iduals m a y be selected in to var ious s ta tuses based on u n o b s e r v e d

var iables tha t a re re la ted to mor ta l i ty . The only s tudy o f s o c i o e c o n o m i c mor ta l i ty different ia ls in the U n i t e d

Sta tes tha t has m a d e an explicit a l lowance for u n o b s e r v e d he te rogene i ty f o u n d tha t coeff icients on e d u c a t i o n (and income) were robus t for the inc lus ion o f u n o b s e r v e d he te rogene i ty [19].

*Age is treated as a continuous linear variable because tests for nonlinearities (i.e. adding a term in age squared) showed them to be insignificant, except for Models l and 2 for males in the age range 25--64 (the age square term does not affect the substantive interpretation of the results and is not included in the models). That the logit ofdeath rates or death probabilities is highly linear in age was first demonstrated by actuaries in the 1930s (e.g. [23]) and has been repeatedly reaffirmed (e.g. [24]).

F o r each o f our four age - s ex g roups , we es t imate a

sequence o f th ree equa t i ons tha t reflects the fact tha t charac te r i s t ics are acqu i red at d i f ferent s tages o f life. W e begin by inc luding on ly charac ter i s t ics acqu i red at bir th: race, reg ion o f bir th , a n d year o f b i r th (age), in s u b s e q u e n t d i scuss ion re fe r red to as M o d e l 1.* A t a s econd s tage we a d d educa t iona l a t t a inmen t , w h o s e value is typical ly es tab l i shed in ado lescence o r ear ly a d u l t h o o d ( M o d e l 2). T h e th i rd stage a d d s ' cu r r en t ' var iables eva lua ted at the da te o f initial interview: family income , mar i t a l s ta tus , m e t r o p o l i t a n res idence a n d h o u s e h o l d size ( M o d e l 3). T w o types o f s tat is t ical tests are c a r d e d out: t - tes ts for tes t ing the signif icance o f ind iv idua l coefficients (e.g. for the ne t effects be t ween each ca tegory a n d the re ference ca tegory in the case o f ca tegor ica l covar ia tes) , and g lobal tests o f significance, c o m p a r i n g fits o f the sequent ia l (nes ted)


Table 4. Effect of an additional year of schooling on the Log Odds of dying: United States, 1979-1985

Coefficients of years of school completed** % Reduction

Age group Model 2 to and sex Number of cases Model 2 Model 3 Model 3

Men 25-64 133,229 -0.0546 -0.0250 54 Women 25-64 147,281 - 0.0501 - 0.0311 38 Men 65-89 25,217 -0.0299 -0.0172 42 Women 65-89 35,177 -0.0207 -0.0162 22 Range 0.0339 0.0149

**All coefficients are significant at 0.001. Note: Sample excludes persons for whom education or income is missing and persons who are

unable to work for health reasons. Model 2 controls for age, region of birth and race. In Model 3, controls are added for family income, household size, marital status and current residence.

models [25]. Our study samples consist of 160,105 males and 182,688 females aged 25-89 at the initial interview, of whom 10,312 and 7809 respectively, had died during the five-year follow-up period.*

RESULTS AND DISCUSSION

Table 3 presents results from the final model (Model 3) for the four age-sex groups. Categories are created for what are usually a small number of cases (region of birth is an exception) with missing values on a characteristic (results not shown). As noted above, we have excluded from our samples individuals who are unable to work because o f long-term disability. The inclusion of this group would result in overestimation of income effects on mortality. For example, for males aged 25-64 at the initial interview, if those unable to work for health reasons are included in the sample, the coefficient of log income is -0 .3454 instead of -0 .3136 as shown in Table 3. In other words, the coefficient on income is biased upwards in absolute value by 10.1% when people who are unable to work for health reasons are included. Income effects are thus overestimated when income is allowed to serve as a partial proxy for one's health status. We begin the discussion of the multivariate results with educational attainment.

*In the age range of interest, race was missing for 99 males and 30 females and these cases are excluded from our sample. As noted earlier, we have also excluded individuals who were reported as being unable to participate in the labor force due to long-term disability. Persons of unknown age have been excluded from the Public Use Sample. We have also excluded cases for which income was missing (about 5.5% of cases at ages 25-89); our substantive conclusions are not sensitive to this choice. Individuals within the same household for whom income was missing are assigned the same family income code as other members of the same household (< 1% of all cases). For the treatment of sample individuals who had missing values for other characteristics of interest, see text.

tTbese changes in coefficients are statistically significant for all four age-sex groups. To test whether the changes are statistically significant between Model 2 and Model 3, we follow Clogg et al. [26] (pp. 61-64, 71) and approximate the variance estimates for a change in the coefficient by the bootstrap sample reuse method.

Educational attainment

Educational attainment clearly has very substantial effects on adult mortali ty in all four age-sex groups (Table 3). The pattern of differentials, but not their magnitude, is similar to the unadjusted effects o f education shown in Table 1. In all cases, college graduates tend to have lower mortality than high school graduates, and people who did not attend high school tend to have higher mortali ty with one exception. The mortali ty of those who attended but did not complete college does not differ significantly from that of high school graduates, except for females aged 65-89.

To summarize these comparisons more efficiently, we have re-run Models 2 and 3 after converting years of schooling to a continuous variable, with each schooling category placed at the mid-point of that category. This variable replaces the set of dummy variables for years of school completed shown in Table 3. Its coefficient shows the average effect of an additional year of schooling on the log odds of dying in a five-year period. Results are presented in Table 4. Model 2 results confirm that the effect of education is larger for working age persons than for those aged 65 + ; within each age group, it is larger for men than for women.

When Model 2 is compared to Model 3, the largest reduction in coefficient, 54%, occurs for working-age men, and the second largest reduction occurs for older men (42%).t In both age groups, reductions in education coefficients for women are smaller than for men, and for each sex, the reduction is larger for younger people than older individuals. After the introduction of 'current ' variables in Model 3, the effect of education is actually larger for women than for men in the working ages. As discussed earlier, a greater degree of educational inequality in mortality for men than for women has been observed in Europe than in the United States. Based upon the present analyses, it appears that this sex difference in the United States is primarily attributable to the greater association of men's schooling with family income, combined with substantial responsiveness of mortality to income. This interpretation is confirmed by the


addition one at the time of income and other characteristics to Model 2. By far the largest reduction in the education coefficient in all four age-sex groups occurs when income is added to Model 2 (results not shown).

The mortality of older persons shows less responsiveness to years of schooling than that of younger persons, particularly in Model 2, but this difference is reduced substantially for both men and women when current income, residence and other individual and household characteristics are controlled in Model 3. The range of coefficients for the four groups is 0.0339 in Model 2 but only 0.0149 in Model 3.

The proposition that the effects of education diminish as age advances is also confirmed by the introduction of an interaction term, the product of age and years of schooling (expressed as a continuous variable), into Models 2 and 3. The coefficients of the interaction terms are all positive and significant except among women aged 25--64, where it is positive but insignificant. That is, the negative effect of schooling on mortality diminishes with age. As noted earlier, this finding is consistent with a general tendency across populations for the proportionate differences in adult age-specific death rates to diminish with age. Nevertheless, that mortality of older persons continues to be significantly affected by years of schooling holds out the possibility of additional reductions in old age mortality as better educated cohorts replace the cohorts who now occupy the older age groups.

Finally, we should note that to the extent that one's educational attainment reflects a poor health environment in childhood that influences both adult health and levels of schooling (see e.g. Ref. [27]), the direct effects of education on mortality are overestimated. The lack of data on an individual's childhood environment in the NLMS prevents us from adequately controlling for possible confounding factors. The results from a study by Mare [28], showing that the excess mortality of men whose fathers were in low occupational groups is eliminated when the man's years of schooling are controlled, suggest that the bias may not be large. Nevertheless, we must recognize the possibility that some of the effects of education may reflect the unmeasured influence of childhood environment.

Age

All four demographic groups have an age coefficient (controlling race and region of birth) in the range of 0.091-0.095 (results not shown). When other characteristics are introduced, the age coefficient diminishes in each case to the range of 0.084--0.091 (Table 3). The actuarial and demographic literature is replete with efforts to treat the age pattern of mortality as some intrinsic biological entity on which various theories of aging can be tested. However, it is dearly

also subject to social influences. The fact that older cohorts are more poorly educated than younger cohorts in the U.S. (and other countries) has steepened the unadjusted age pattern of mortality and exaggerated the apparent mortality consequences of aging.

Race

Controlling for age and region of birth, black males aged 25-64 at initial interview have a log odds of dying that is 58% higher than that of white males. This difference translates into an odds ratio .of exp {0.584} = 1.79 (results not shown). Of the difference in the log odds, 0.094 or 16.1% is eliminated when educational attainment is controlled. A large majority of the racial difference in mortality of working age males remains unexplained by educational differences between the races. This racial difference persists when differences in family income, marital status, household size and current residence are taken into account (Table 3).

The black-white contrast is even slightly larger among middle-aged women than among middle-aged men, with an odds ratio of 1.84 (exp {0.612}) controlling for race and region of birth (results not shown). Again, the difference persists but is diminished when other characteristics are taken into account (Table 3). Among persons aged 65-89, the black excess is sharply diminished relative to the working ages. It is insignificant for males controlling for age and region of birth (results not shown). As additional characteristics are introduced, the black excess at ages 65-89 also becomes insignificant for females, and is reversed in favor of black males (Table 3).

These results are clearly related to differences in age patterns of mortality for blacks and whites. Vital statistics-based death rates show a cross-over between the races throughout the twentieth century in the age range of 73-85 [29-33]. Whether this cross-over is real or a product of age misreporting is uncertain. That there is a substantial degree of age misreporting among older blacks is indisputable [34]. In a 1960 match of census and death records, ages were systematically reported as being older on census forms than on death certificates for the same individual [35]. Errors were much larger for blacks than for whites. The present longitudinal design has the advantage of eliminating inconsistencies between the sources. However, the fact that census ages for blacks were in general reported as older in the matching study still allows the possibility of net age overstatement among older black persons, most of whom were not issued birth certificates. Such a tendency could produce a racial cross-over in death rates at older ages even i fa single source is used for age because the reported black death rates at a given age would actually pertain to what is, on average, a younger age group. Introduction of an interaction term between age and being black in Model 1 results in a significant negative coefficient, indicating a lower

54 Irma T. EIo and Samuel H. Preston

black slope, for males over age 64 and females at ages 25-64. The interaction term is not significant for older women and younger men, although it is also negative for these groups. Because the interaction terms do not affect the substantive interpretation of results, we do not present results that include them.

We also tested for interactions between being black and each of the other characteristics shown in Table 3 (the search for region-of-birth interactions was limited to persons born in the South). Over all four age-sex groups no coefficient of a racial interaction term was significant at 5%. This result is surprising because even random variation alone would lead us to expect to find 2-3 significant coefficients. It is noteworthy that central city residence for black males aged 25-64 does not appear to convey any significant additional risk in view of vivid portraits of mortality conditions in such circumstances that have been presented [36]. Although black males in central cities have some excess mortality at ages 25-64, that excess is not significantly greater than that for white males.

Preston and Taubman [4] review other studies of racial differences in adult mortality and conclude that the introduction of a standard set of socioeconomic variables typically eliminates most or all of the excess mortality of blacks. Our results suggest that the ability of these variables to account for racial differences will depend strongly on the ages considered. At higher ages, where the racial difference in death rates is diminished, social and economic variables should be expected to explain more of the (smaller) gap. Below age 65, however, our results suggest that racial differences remain important. Analyses like that of Rogers [37], who combines all ages above 24, risk missing the important interactions between race and age, and its implications for the role of socioeconomic attributes in accounting for racial differences in mortality.

Region of birth

Region of birth effects are substantial. Relative to those born in the Northeast, those born in the West have unusually low mortality, except among working- aged women where the difference is small and insignificant. With two exceptions, other domestic contrasts are small and insignificant (Table 3). Hispanics born outside the U.S. also have unusually low mortality, and the Hispanic advantage grows in each case when other variables are controlled (results not shown). Others who were born outside the U.S.--primarily non-Hispanic whites and Asians-- also have low mortality, particularly among males aged 25-64. It is possible that the advantage of non-natives reflects in part the difficulties of following them up in U.S. death records, since it is likely that some may have died abroad. Nevertheless, the unusually low mortality of both Hispanic and Asian-American and foreign-born adults has been noted in other sources that are not subject to follow-up losses [38--40] (p. 13).

Income

Coefficients of family income have the expected negative sign for all four age-sex groups. They are larger for persons aged 25--64 than for older persons and larger for men than for women in both age intervals. Family income has a particularly large effect on the mortality of working age males, as shown in Table 3. The coefficient of -0 .3136 indicates that a doubling of family income (which increases log of income by 0.693) reduces the log odds of dying by 21.7% [(0.3136)(0.693)(100)].

The age-sex pattern o.f income coefficients has several possible sources:

(1) Differences in access to health care. The greater sensitivity of mortality to income among younger persons may reflect the steeper fees that younger persons must pay for medical insurance or medical services. The vast majority of the 65 + population has heavily subsidized medical coverage through Medi- care. However, as we noted earlier, reductions in mortality differentials are typically observed as age advances and any specific explanations must recognize the generality of this pattern.

Health care access may also be related to sex differences in income coefficients. Among younger persons, subsidized medical insurance coverage is more readily available for poor women than poor men through the Medicaid program. Eligibility for Aid to Families with Dependent Children (AFDC) program, a welfare program for single-parent households, confers automatic eligibility for Medicaid. Others qualify through the Supplemental Security Income (SSI) program for the aged, blind and disabled, and states have some flexibility in extending coverage to other groups as well. But, in fiscal year 1983, 69.3% of Medicaid recipients qualified for Medicaid through AFDC. Of all recipients (including children), 64.1% were female. This percentage is likely to be even higher among the 21-64 year olds (who make up 31.5% of all Medicaid recipients), since the vast majority of recipients in this age range qualify for Medicaid through AFDC [41].

(2) Contamination of the income variable with health status. Although we have removed from the analysis persons who are unable to work for health reasons, it is possible that other people remain whose income is reduced by health impairments. If so, income is serving as a partial proxy for health status and income coefficients will be biased upwards. This bias will be greater for those groups most likely to be in the labor force, i.e. for working age persons and for males. This potential bias thus mimics the age--sex pattern of income coefficients that we have uncovered.

(3) It is also possible that men's mortality is more responsive than women's to family income because accretions to family income are of greater benefit to men, who have more control over family budgets in many households [42]. And income coefficients may be smaller for older persons than for younger persons

Educational differentials in mortality

because income is a better marker of economic status for younger people than for older people, many of whom have accumulated significant wealth that does not appear in income flows. Wealth has been shown to have an independent effect on mortality [43]. Unfortunately, there are no wealth measures in our data set.

Given a particular level of family income, mortality should be higher in larger households, where there are more claimants on that income. This expectation is not realized among working-age persons, although the coefficients of household size are insignificant (Table 3). However, at ages 65 + for both men and women the coefficient is significant and properly signed. Controlling household income, each additional person in the household increases an individual's mortality by approx. 6%. However, this variable may also be functioning in part as an indicator of health state at initial interview, just as income itself may be. It is possible that the households of older persons in poorer health are more likely to include members who are specifically present for care-giving purposes. Those elderly people who live alone may be in unusually good health. The possibility is less salient for persons 25--64, where the prevalence of chronic conditions that may evoke a change in household composition is much lower [44].

Thus, our efforts to model the mortality effects of income and of demands on that income have been only partially successful. It is true that six of the eight pertinent coefficients are significant and properly signed (Table 3). However, the age-sex pattern of coefficients suggests that these two variables may also be functioning in part as indicators of health status.

Marital status

Our results on the relation between marital status and mortality are remarkably consistent. They show that, for all four age-sex groups, currently married people have the lowest mortality. The 12 coefficients representing the three other marital status categories among the four age-sex groups range from 0.12 to 0.41, and all but one are statistically significant.

These results are consistent with a broad range of literature that demonstrates lower mortality for married persons in the United States and elsewhere (see the review in Ref. [45]). Some of this literature finds marital status differentials to be larger among males than among females (see especially Ref. [46]). Our results are consistent with this pattern: five of the six comparisons of male and female coefficients for a particular age and marital status category show that the male coefficient is larger. Two of the largest coefficients pertain to 'divorced and separated' males and females at ages 65-89. Further analysis (not shown) traces the very high mortality of this group primarily to exceptionally high death rates for 'separated' persons. Once again these results may reflect the confounding effects of health status, since

55

the separation of spouses may be a product of health impairments.

The lower mortality of married persons is likely to reflect both the protective effects of marriage and health-based selection mechanisms that sort individuals into the different marital categories. As an example of selection, persons with health impairments are less likely to ever marry [47-49]. Our research is not well-designed to test the relative importance of selection vs protection effects. However, we consider the effects of widowhood to be least affected by the influence of health selection because one's own characteristics should play the smallest role in rates of transition into (but not necessarily out of) the state of being widowed. Averaged across the four age-sex categories, the mean widowhood effects (0.236) are somewhat smaller than those of other non-married categories and are larger for men than for women. This finding supports the possibility that coefficients of the other marital status categories include a more pronounced selection effect.

Residence

In three of the four age-sex groups, living in a central city of a metropolitan area is associated with significantly higher mortality than living in a non-metropolitan area. Residence in a metropolitan area outside the central city is also associated with elevated mortality, although only one of the four coefficients is significant. The magnitudes are similar to the 5% excess mortality of metropolitan area residents found by Kitagawa and Hauser [1] (p. 120) for 1960, although they did not control any other characteristics. Since health care resources are disproportionately located in metropolitan areas [50], these results are not consistent with a view that the availability of health care facilities and personnel plays a major role in adult mortality. Of course, these factors may be offsetting some disadvantage of urban living that would be even more visible in their absence. Since we cannot identify the SMSA in which individuals live, we cannot attach information on locality-specific health resources that could be used to test their influence more directly.

SUMMARY

Our analyses reveal significant educational differentials in mortality among both men and women in the United States in the early 1980s, with the differentials being somewhat larger for men than for women and during working ages than at ages 65 and above. Our results further show that these differentials persist but are reduced in magnitude when controls for current income, residence and other household and individual-level characteristics are introduced. The largest reduction (54%) occurs for working age males. This pattern is consistent with the interpretation that the initial (Model 2) effects are largest for this group because educational attainment has the largest

SSM 42/I--E


impact on family income among them. The responsiveness of mortality to years of schooling at ages 65 and above gives support to the view that old age mortality will continue to decline in the U.S. in the coming decades as better educated cohorts replace those currently aged 65 + .

The proportionate reductions in mortality for each one year increase in schooling in the United States at ages 35-54 are comparable to those estimated for a number of European countries by Valkonen [13]. The main difference between the United States and Europe is that in the U.S. the responsiveness of mortality to years of schooling is quite similar for both men and women, while in a number of European countries male mortality appears more responsive than female mortality to educational attainment. The similarity of our findings to those of Valkonen's is of interest in view of differences in the organization of the health care sector in the U.S. and in Europe. It suggests that educational differentials may not be very responsive to the organization of health care services (see also Ref. [51]).

Our data do not allow us to examine the association between an individual's level of education and health related behaviors or other intervening variables except current income, residence and other household and individual-level socioeconomic characteristics. The results suggest that the association between income and education is most important; controlling for income has the largest attenuating effect on the education coefficients in three of the four age-sex groups, although more so among men than women and at younger than older ages. We must, however, be cautious in the interpretation of these results, because of our inability to control completely for an individual's health status at the initial interview.

Although the focus of this paper is educational attainment, we have also uncovered a large mortality differential between blacks and whites at ages 25-64. The large mortality penalty associated with being black has often been obscured by research designs that include experience at ages above 65, where recorded mortality rates for the races appear to converge (although the quality of data at these ages is highly suspect: [34]). At ages 25-64, we can 'account for' only 54% of the black disadvantage for males and 39% for females by the adverse distribution of blacks on other measured socioeconomic variables; and even this accounting begs the question of why blacks are so adversely distributed. The NLMS, by far the largest reliable data set available to study recent socioeconomic mortality differentials in the U.S., reveals a landscape with many predictable gradients and a large and unexplained racial disparity at ages 25-64.

Acknowledgements--This research is supported by grants from the National Institute of Aging, AGI0168-01 and P30HDI0379-16. We would like to thank Tim Cheney for programming assistance and Andrew Foster, Charles Hirschman, Paul Sorlie, Tapani Valkonen and two anonymous reviewers for helpful comments. An earlier

version of this paper was presented at the annual meeting of the Population Association of America, May 1994.

REFERENCES

1. Kitagawa E. M. and Hauser P. M. Differential Mortality in the United States: A Study in Socioeconomic Epidemiology. Harvard University Press, Cambridge, MA, 1973.

2. Benjamin B. Social and Economic Factors Affecting Mortality. Mouton, The Hague, 1965.

3. Liberatos P., Link B. G. and Kelsey J. L. The measurement of social class in epidemiology. Epidemiol. Rev. 10, 87, 1988.

4. Preston S. H. and Taubman P. Socioeconomic differences in adult mortality and health status. In The Demography of Aging (Edited by Martin L. and Preston S. H.), pp. 279-318. National Academy Press, Washington, DC, 1994.

5. Winkleby M. A., Jatulis D. E., Frank E. and Fortman S. P. Socioeconomic status and health: how education, income and occupation contribute to risk factors for cardiovascular disease. Am. J. Publ. Hlth 82, 816, 1992.

6. Rogot E., Sorlie O. and Johnson N. J. Probabilistic methods in matching census samples to the National Death Index. J. Chronic Dis. 39, 719, 1986.

7. Rogot E., Sorlie P. D., Johnson N. J. and Schmitt C. A Mortality Study of 1.3 Million Persons by Demographic, Social, and Economic Factors: 1979-1985 Follow-up. National Institutes of Health. NIH Publication No. 92-3297. Washington, DC, 1992.

8. Preston S. H. and EIo I. T. Are educational differentials in adult mortality increasing in the United States? Population Aging Research Center, Working Paper 95-1. Population Studies Center, University of Pennsylvania, 1994.

9. U.S. Bureau of the Census. U.S. Population Estimates, by Age, Sex, Race and Hispanic Origin: 1980-1991. Current Population Reports, P25-1095. Government Printing Office, Washington, DC, 1993.

10. Feldman J. J., Makuc D. M., Kleinman J. C. and Cornoni-Huntley J. National trends in educational differentials in mortality. Am. J. Epidemiol. 129, 919, 1989.

11. Pappas G., Queen S., Hadden W. and Fisher G. The increasing disparity on mortality between socioeconomic groups in the United States, 1960 and 1986. N. Engl. J. Med. 329, 103, 1993.

12. Coale A. J. and Demeny P. Regional Model Life Tables and Stable Populations. Academic Press, New York, 1983.

13. Valkonen T. Adult mortality and level of education: a comparison of six countries. In Health Inequalities in European Countries (Edited by Fox J.), pp. 142-162. Gower, Aldershot, 1989.

14. Kunst A. E. and Mackenbach J. P. The size of mortality differences associated with educational level in nine industrialized countries. Am. J. Publ. Hlth 84, 932, 1994.

15. Sorlie P. D. and Rogot E. Mortality by employment status in the National Longitudinal Mortality Study. Am. J. Epidemiol. 132, 983, 1990.

16. Rogot E., Sorlie P. D. and Johnson N. J. Life expectancy by employment status, income and education in the National Longitudinal Mortality Study. Publ. Hlth Reports 107, 457, 1992.

17. Pierce J. J., Fiore M. C., Novotny T. E., Hatziandress E. J. and Davis R. M. Trends in cigarette smoking in the United States. J. Am. Med. Assoc. 261, 56, 1989.

18. Computer Resource Center. Stata. Reference Manual. Release 3. Santa Monica, CA, 1992.

19. Behrman J. R., Sickles R. and Taubman P. Some Causes and Consequences of Death. University of Pennsylvania, PA, mim¢o, 1993.


20. National Institutes of Health. National Longitudinal Mortality Study. Public Use File Documentation, 1992.

21. Huber P. J. The behavior of maximum likelihood estimates under non-standard conditions. Proc. Fifth Berkeley Syrup. Mathematical Stat. Probability 1, 221, 1967.

22. White H. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica 48, 817, 1980.

23. Perks W. On some experiments in the graduation of mortality statistics. J. Inst. Actuaries 63, 12, 1932.

24. Himes C., Preston S. and Condran G. A relational model of mortality at advanced ages in low mortality countries. Pop. Stud. 48, 269, 1994.

25. Hosmer D. W. and Lemeshow S. Applied Logistic Regression. John Wiley and Sons, New York, 1989.

26. Clogg C. C., Petkova E. and Shihadeh E. S. Statistical methods for analyzing collapsibility in regression models. J. Educ. Stat. 17, 51, 1992.

27. EIo I. T. and Preston S. H. Effects of early life conditions on adult mortality: a review. Pop. Index 58, 186, 1992.

28. Mare R. D. Socioeconomic careers and different mortality among older men in the United States. In Measurement and Analysis of Mortality (Edited by Vallin J., D'Souza S. and Palloni A.), pp 362-387. Clarendon Press, Oxford, 1990.

29. U.S. Bureau of the Census. United States Life Tables and Actuarial Tables 1939-1941. Government Printing Office, Washington, DC, 1946.

30. National Center for Health Statistics. United States Life Tables 1949-51. Vital Statistics-Special Reports, Vol. 41(1). Government Printing Office, Washington, DC, 1954.

31. National Center for Health Statistics. Life Tables: 1959-61. United States Life Tables: 1959--61. Govern- ment Printing Office, Washington, DC, 1968.

32. National Center for Health Statistics. United States Life Tables. U.S. Decennial Life Tables for 1969-71. Government Printing Office, Washington, DC, 1975.

33. National Center for Health Statistics. United States Life Tables. U.S. Decennial Life Tables for 1979-81. Government Printing Office, Washington, DC, 1985.

34. Elo I. T. and Preston S. H. Estimating African American mortality from inaccurate data. Demography 31, 427, 1994.

35. National Center for Health Statistics. Comparability of Age on the Death Certificate and Matching Census Record, United States--May-August 1960. Vital and

Health Statistics, Series 2, No. 29. Public Health Service, Washington, DC, 1968.

36. McCord C. and Freeman H. P. Excess mortality in Harlem. N. Engl. J. Med. 322, 173, 1990.

37. Rogers R. G. Living and dying in the U.S.A.: sociodemographic determinants of death among blacks and whites. Demography 29, 287, 1992.

38. Kestenbaum B. Mortality by nativity. Demography 23, 87, 1986.

39. Rosenwaike I. Mortality differentials among persons born in Cuba, Mexico, and Puerto Rico residing in the United States, 1979-81. Am. J. Publ. Hlth 77, 603, 1987.

40. National Center for Health Statistics. Health United States: 1990. Public Health Service, Hyattsville, MD, 1991.

41. U.S. Department of Health and Human Services. Health Care Financing. Program Statistics. Medicare and Medicaid Data Book, 1986. Health Care Financing Administration. Office of Research and Demonstrations. HCFA Pub. No. 03247. Government Printing Office, Washington, DC, 1987.

42. Millman M. Warm Hearts and Cold Cash: The Intimate Dynamics of Families and Money. Free Press, New York, 1991.

43. Menchik P. L. Economic status as a determinant of mortality among black and white older men: does poverty kill? Pop. Stud. 47, 427, 1993.

44. National Center for Health Statistics. Educational differences in health status and health care. Vital and Health Statistics. Series 10, No. 179. Government Printing Office, Hyattsville, MD, 1991.

45. Goldman N. Marriage selection and mortality patterns: inferences and fallacies. Demography 30, 189, 1993.

46. Gove W. Sex, marital status, and mortality. Am. J. Sociol. 79, 45, 1973.

47. Higashi T., Ishihara O. and Wada T. Medicosocial aspect of people with epilepsy in Japan: a survey from standpoint of epilepsy center. Folia Psychiat. Neurol. Jap. 33, 399, 1979.

48. Imaizumi Y. Mortality rate of Hunington Disease in Japan: secular trends, marital status, and geographic variations. Jap. J. Human Genetics 34, 169, 1989.

49. Kiernan K. E. 1988. Who remains celibate? J. Biosoc. Sci. 20, 253, 1988.

50. U.S. Congress, Office of Technology Assessment. Health Care in Rural America. OTA-H-434. Government Printing Office, Washington, DC, 1990.

51. Adler N. E., Boyce T., Chesney M. A., Folkman S. and Syme L. Socioeconomic inequalities in health: no easy solution. J. Am. Med. Assoc. 269, 3140, 1993.

Educational differentials in mortality: United States, 1979–1985

Documents