Logo of nihpaAbout Author manuscriptsSubmit a manuscriptHHS Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
. Author manuscript; available in PMC 2014 Apr 10.
Published in final edited form as:
PMCID: PMC3982329
NIHMSID: NIHMS462464

Rising Educational Gradients in Mortality: The Role of Behavioral Risk Factors

David M Cutler,a Fabian Lange,corresponding authorb Ellen Meara,c Seth Richards-Shubik,d and Christopher J Ruhme
Author information ► Copyright and License information ►

Abstract

The long-standing inverse relationship between education and mortality strengthened substantially at the end of the 20th century. This paper examines the reasons for this increase. We show that behavioral risk factors are not of primary importance. Smoking declined more for the better educated, but not enough to explain the trend. Obesity rose at similar rates across education groups, and control of blood pressure and cholesterol increased fairly uniformly as well. Rather, our results show that the mortality returns to risk factors, and conditional on risk factors, the return to education, have grown over time.

Keywords: Health Inequality, Risk Factors, Education and Mortality, Smoking, Obesity

1. Introduction

Mortality in the United States has declined by much more among the better educated than among the less educated. Between 1960 and 1986, education-related differences in mortality grew 20 percent (Pappas et al. 1993). The gap widened further and more rapidly in the decade after 1990, when life expectancy of those attending college increased an additional 1.6 years with no change among those who did not go to college, yielding a 30 percent growth in life expectancy gaps by education (Meara, Richards, Cutler 2008). By 2000, college-educated 25-year olds could expect to live 7 years longer than their peers with less schooling. These patterns have thrust the issue of health disparities onto the political agenda. Reducing such disparities (by race and ethnicity as well as economic status), along with improving population health, are two major components of the Healthy People 2010 objectives (U.S. Department of Health and Human Services 2000).

Sources of the increase in these educational gradients remain poorly understood. Some studies show that advantaged individuals receive better and earlier medical care than their less advantaged counterparts (e.g. Rathore et al. 2000). Other analyses stress behavioral differences: the better educated are less likely to smoke, drink, or (at least among women) to be obese (Cutler and Lleras-Muney 2008). Still other research suggests the possibility that high status individuals are less exposed to unalleviated stress (Marmot 2006).

This paper analyzes the extent to which behavioral differences (smoking and obesity), and their immediate medical correlates (hypertension and high cholesterol), explain changes in educational mortality gradients occurring over the past three decades.1 Smoking and obesity are natural to examine because they are the two leading behavioral causes of death in the United States. Tobacco use is responsible for about 435,000 premature deaths annually (Mokdad et al. 2004) and obesity for between 100,000 and 400,000 early deaths per year (Flegal et al. 2005; Willett et al. 2005).2 Since the obese often develop high blood pressure (hypertension) and high cholesterol (hypercholesterolemia), and management of these risk factors is itself a behavioral issue, we further examine how disease management varies across education groups. Finally, we separately consider deaths due to cardiovascular disease (CVD, mostly heart disease) and cancer, since these represent major sources of premature mortality and are responsive to changes in modifiable risk factors.

Cutler, Glaeser, and Rosen (2009) show that behavioral risk factors play an important role in understanding changes in overall mortality trends. They estimated the contribution of demographics and risk factors (smoking, drinking, obesity, and blood pressure) to 10-year mortality risk during the last three decades of the 20th century and provide evidence that decreases in smoking and better control of hypertension contribute most to the substantial reductions in age-adjusted deaths rates, while increases in obesity raised mortality risk (also see Olshansky et al. 2005, and Stewart, Cutler, and Rosen 2009). We use similar data but, instead of focusing on the entire population, assess the extent to which differential changes in risk factors explain secular increases in education-related mortality gaps.

Our analysis reveals three primary findings. First, education-gradients in mortality are stronger for men than women but have increased over time for both sexes. Second, despite the importance of smoking, obesity, hypertension, and cholesterol as determinants of population health, differential changes in these risk factors do not explain the widening educational gap in death rates since the 1970s. Finally, the mortality returns to risk factors and the return to education, conditioning on them, have grown over time for reasons that are not yet understood. Thus, even if less educated populations were able to achieve risk factor profiles mirroring those with more education, widening mortality differentials would likely persist.

Three explanations seem likely to explain why the impacts of risk factors and education have increased over time. First, access to medical care may have become more important for detecting disease early and treating it appropriately, and the better educated have superior access to care. Second, the living environments (i.e. the exposure to environmental health risks) may have improved more over time for the better educated. Third, the management of chronic health problems may have become more sophisticated in ways that favor those with more schooling. Our data are not adequate to test these theories, which we leave to subsequent research.

Sections II and III present the data we analyze and descriptive trends in mortality, and the fourth section our empirical approach. The fifth section reports our mortality regressions, and the sixth section uses these to understand changes in the education-gradient in mortality. Section VII examines CVD and cancer deaths. The last section concludes.

2. Data and Key Variables

Our analysis utilizes data from various waves of the National Health and Nutrition Examination Surveys (NHANES) and multiple years of the National Health Interview Survey (NHIS). The NHANES and NHIS both provide samples that are nationally representative of the non-institutionalized U.S. population in the specified period, using stratified, multi-stage probability cluster designs. Information on subsequent mortality for selected surveys is used to estimate education-related differences in death rates. The remainder of this section describes the surveys and key explanatory variables.

The National Health and Nutrition Examination Survey

Baseline information for this study was obtained from NHANES I, covering the period 1971–1975. Subsequent descriptive data are from the NHANES II (1976–1980) and NHANES III (1998–1994), and the first six years (1999–2004) of the continuous NHANES survey (hereafter referred to as NHANES IV).3 The NHANES is particularly useful since it includes a physical exam component providing clinical measures of weight, height, blood pressure, and cholesterol. Our analysis is limited to non-Hispanic whites, since the earlier NHANES does not provide adequate size to estimate mortality among non-white or Hispanic populations. All results are age-adjusted and separated by gender, so that trends do not reflect changes in these characteristics.

The absence of persons institutionalized at baseline reduces the estimates of future mortality, since the institutionalized have higher death rates than the rest of the population. However, Meara, Richards and Cutler (2008) demonstrate that, conditional on surviving one year, the mortality rates for adults initially living in the community closely resemble those reported in published statistics for the entire population. Therefore, we limit the sample to individuals surviving at least one year from the baseline interview for all mortality models estimated below.

We further restrict the sample to those aged 25–74.4 This provides 5,848, 8,357, 4,720 and 4,777 respondents from NHANES I, II, III and IV respectively. NHANES I and NHANES III include mortality follow-up surveys, in which respondents are matched to National Death Index records providing information on the timing and cause of death. For NHANES I, the follow-up covers a period of 10 years (1981–85). For NHANES III we utilize public use data including deaths through 1998–2000, providing 10 years of information for around half of survey respondents.5 Relatively small sample sizes for the NHANES III follow-up reduce the precision of these estimates, so they are only used as a supplementary source of information on mortality differentials. The main data set employed during the later sample period is the NHIS, which is described below. In addition to total mortality, we examine deaths due to cardiovascular disease (CVD) and cancer, as defined in NHANES I using International Classification of Diseases, Ninth Revision (ICD-9) codes.6

The National Health Interview Surveys

We supplement NHANES data with data from the 1987–88, 1990–95, and 1997–2000 years of the National Health Interview Survey.7 The NHIS data have been linked to deaths through 2002, and we compute survival for a follow-up period of a decade, or through 2002 for individuals interviewed after 1992. The advantage of using the NHIS is its large sample size, which allows greater precision of the estimated mortality functions than in the NHANES (particularly NHANES III). The cost is that, unlike NHANES, all information is self-reported. Therefore, the NHIS lacks the clinical markers used to evaluate the role of hypertension or cholesterol control in NHANES. Also, self-report data underestimate obesity prevalence since height is over-reported and weight is understated (Kuczmarski, Kuczmarski, and Najjar 2001; Villanueva 2001).

As with NHANES, our mortality estimates are for persons surviving at least one year, to adjust for the exclusion of the institutionalized at baseline. Similarly, we restrict the sample to non-Hispanic whites aged 25–74 at baseline, leaving a final sample size of 207,985.8 Cause-of-death in the NHIS is classified using “U codes” that represent broad categories compiled from the underlying International Classification of Diseases-10th Revision codes. We use the latter to identify cancer and CVD mortality.9

When selecting data sets for our analysis, we traded off the precise measurement of features like BMI, cholesterol, and hypertension over long time periods in the NHANES surveys, with the larger sample size, but a smaller set of self-reported measures available in the NHIS. The magnitude of educational gaps in risk factors and mortality rates measured in later time periods were remarkably similar in the NHIS and the NHANES data, suggesting that the overall under-reporting of risk factors likely to occur in the NHIS does not occur differentially by education group, something we discuss further in section 6. We acknowledge that the biological measures available over time for our analysis are limited and imperfect, but these same measures of hypertension and cholesterol have been shown in previous research to respond to medical and behavioral intervention (Chobanian et al. 2003; Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults 2001).

Education

We divide the population into two education groups: those who completed more than 12 years of schooling (college attendees) and those with less education. These broad categories were chosen to minimize errors introduced by the known tendency to over-report the attainment of a high school degree (Sorlie and Johnson 1996). Secular changes in the composition of the education groups are potentially problematic, since the share of individuals attending college has grown dramatically over time: from 33 percent of men and 24 percent of women in 1971–75 to 60 and 61 percent in 1999–2004. We test for the importance of such compositional changes by estimating specifications where individuals on the margin between the low and high education group are reassigned so as to equalize the compositional shares across time periods, as detailed in section 6.

Behavioral Risk Factors

We focus on smoking, obesity, blood pressure and cholesterol as potentially important mortality risk factors. Current smokers are those who report smoking at the time of interview; former smokers are those who had previously smoked at least 100 cigarettes but do not report smoking at the interview date.10 In supplemental specifications, we also control for the number of cigarettes smoked per day (for current smokers) and the time since last use (for former smokers). These are measured with greater error than current/former/never status, and so are not included in the main models.

We distinguished five weight categories based on body mass index (BMI): weight in kilograms divided by height in meters squared. Height and weight are obtained from medical examinations employing standardized procedures and equipment in NHANES but from (less accurate) self-reports in the NHIS. Following national and international standards, we define “underweight”, “healthy weight”, “overweight”, “class 1 obese”, and “class 2 or class 3 obese” for BMI of: <18.5, 18.5 to <25; ≥ 25 to <30, ≥ 30 to <35 and ≥ 35 (World Health Organization 1997; National Heart, Lung and Blood Institute 1998). Persons with a BMI of 35 or more are sometimes referred to as “severely obese”. Although BMI is less accurate than laboratory measures of body composition, since it does not account for variations in muscle mass or in the distribution of body fat, it is a favored method of assessing excess weight because it is simple, rapid, and inexpensive to calculate.11

One reason excess body weight increases mortality is because obese individuals tend to have high rates of hypertension and hypercholesterolemia (Must et al. 1999; Cutler, Glaeser, and Rosen 2009), placing them at risk of serious cardiovascular events. The development of effective medications to control high blood pressure and cholesterol may have substantially reduced the risk of death from these conditions, raising the possibility that larger mortality gradients might reflect better success in ameliorating these health problems for high than low educated individuals. We investigate this by estimating models that added controls for blood pressure and cholesterol.

Blood pressure was divided into four groups, following the recommendations of the National Heart, Lung, and Blood Institute (2004): normal (systolic blood pressure (SBP) <120 mmHG and diastolic blood pressure (DBP) <80 mmHG); pre-hypertension (120≤ SBP<140 or 80≤ DBP<90); stage 1 hypertension (140≤ SBP<160 or 90≤ DBP<100); and stage 2 hypertension (SBP ≥ 160 or DBP ≥ 100).12 Total cholesterol levels were divided into three groups, based on guidelines from the National Heart, Lung, and Blood Institute (2002): normal (<200); borderline high cholesterol (200 to <240); and high cholesterol (≥ 240). We focus on total cholesterol rather than HDL and LDL because the latter were not available in NHANES I.

3. Descriptive Trends

5-year mortality and education

Table 1 displays 5-year age-adjusted death rates by sex and education group. Between the early 1970s and the 1990s, mortality rates declined, but with particularly large reductions for men and the college-educated. Among less educated males, the estimated 5-year death rate fell from 6.4% in NHANES I (1971–75) to 5.9% in the NHIS (1987–96), a drop of 0.5 percentage points; this compares to a 2.0 point reduction (from 5.9% to 3.9%) for college attendees. The death rates of non-college educated women increased 0.5 points (from 3.1 % to 3.6%) over the same period, while falling 1.0 percentage point (from 3.7% to 2.7%) among college attendees.13 Thus, survival gains were large in recent decades but mainly for the highly educated. Table 1 also illustrates that education gradients were stronger for males than for females in each time period except for NHANES I, when education gradients were larger for females than for males.

Table 1
Age-Adjusted 5-year Mortality Rates (percent) By Sex, Education and Time Period

College-educated women have slightly higher estimated mortality rates than their peers in NHANES I (3.7% versus 3.1%) and NHANES III (3.2% versus 2.9%). Given the data required to construct 5-year mortality rates, the sample sizes are relatively small (2,299 and 739 in NHANES I and 1,542 and 966 in NHANES III) and these differences are not statistically significant. Moreover, the finding that education conferred no mortality advantage to adult women in the 1970s echoes a similarly flat mortality gradient, especially for deaths due to cancer and cardiovascular disease, obtained in earlier analysis of the National Cancer Society cohorts from 1959–1971 (Preston and Elo 1995; Steenland, Henley, and Thun 2002). The lack of sample size in the NHANES III data prompted us to use the larger NHIS samples for these years. Although the small NHANES I samples preclude us from drawing firm conclusions regarding the nature of female mortality in the early 1970s, data from the NHIS unambiguously indicates that there were large education-mortality gradients for both sexes at the end of the 20th century. Table 2 confirms the well-known trend of rising education over time. For men, college attendance rose from 33 to 54 percent between NHANES I and the NHIS. Among women the growth was even larger: from 24 to 49 percent.14 In section 6, we discuss our efforts to ensure that our findings are not due to the resulting compositional changes in education groups.

Table 2
Completed Education by Gender and Data-Set (Percent)

Age-adjusted risk factors

Tables 3A and ​and3B3B show how risk factors related to smoking, obesity, hypertension and hypercholesterolemia differ across education groups and over time. The results are presented separately for males and females, and we compare changes from 1971–1975 (using NHANES I) to either 1987–2000 (from the NHIS) or 1999–2004 (using NHANES IV).

Table 3A
Summary Statistics for Age Adjusted Risk Factors, By Time Period, Males
Table 3B
Summary Statistics for Age Adjusted Risk Factors, By Time Period, Females

An inverse relationship between education and current smoking exists in all periods but has strengthened over time, particularly for women. Current smoking fell 8 percentage points between 1971–75 and 1999–2004 for less educated males, versus 15 points among college attendees. During the same period, smoking increased by one percentage point for non-college educated females but dropped 12 points for college attendees. These results, mirror findings previously obtained by Meara, Richards, and Cutler (2008) and raise the possibility that differential reductions in tobacco use contribute to increasing educational gaps in mortality.

The trends in obesity are somewhat sensitive to the choice of endpoints and data sets, with larger secular growth observed for NHANES IV than the NHIS.15 Table 3 nevertheless demonstrates a rapid growth in obesity for all groups but without strong educational differences. Between 1971–75 and 1999–2004, the fraction of less educated males who were obese (class I or higher) increased 17 percentage points (from 14% to 31%) and the proportion severely obese (class II or higher) rose 8 points (from 2% to 10%). The corresponding expansion in obesity and severe obesity for college educated men was 20 points (from 9% to 29%) and 8 points (from 1% to 9%). Among females, the 18 percentage point rise in obesity (from 11% to 29%) for college-educated women was only slightly less than the 20 point increase (from 17% to 37%) for those with less schooling; both groups had an 11 percentage point growth in severe obesity.16 Although obesity growth is smaller when using the NHIS rather than the NHANES IV, similar changes are observed for the two education groups, further indicating that obesity trends are unlikely to explain much of the growth in education gaps in mortality we focus on.

The bottom portion of tables 3A and ​and3B3B document success stories in the battle to reduce risk factors for cardiovascular disease, as measured by the reductions in high blood pressure and cholesterol. Hypertension (Stage I or II) dropped 29 (18) percentage points for less educated males (females) and 26 (17) points for college attendees. Corresponding reductions in severe (Stage II) hypertension were 14 (10) points for non-college educated males (females), and 13 (8) percentage points for college attendees. The prevalence of high cholesterol declined by between 10 and 13 percentage points for both sexes and education categories. The less educated are more likely to have high blood pressure or cholesterol at both points in time but decreases in the severe forms of these health risks tend to be slightly larger for them (except for high cholesterol among men), so that the secular changes are unlikely to explain the widening in education-mortality gradients.

4. Empirical Approach

We examine whether trends in risk factors contribute to widening education differentials in death rates, using an approach similar to Cutler, Glaeser, and Rosen’s (2009) investigation of the sources of overall gains in life expectancy. Our main empirical analysis involves three steps. First, we estimate mortality hazard rates as a function of behavioral risk factors and education, using information from a single data set (such as NHANES I or the NHIS). This describes the relationship between the control variables and mortality given other determinants (e.g. medical technologies) present at that point in time. The schooling coefficient indicates the educational differential for persons with given risk factors. Next, we use the parameter estimates generated in the first stage to simulate how mortality would change within an education group, given the distribution of risk factors present at a different time period. Finally, we calculate ratios of predicted death rates for less versus more educated individuals for each of the four combinations of population characteristics and mortality equations from the earlier (NHANES I, 1971–75) and later (NHIS 1987–2000) period.

For example, the actual mortality gradient during 1971–75 is calculated using NHANES I sample characteristics and mortality hazard rates estimated from this same data set. We then calculated the 1987–2000 education-mortality gradient predicted using the parameter estimates from the 1971–75 mortality risk model but applied to the data from 1987–2000 (and thus the 1987–2000 population characteristics). The difference between these two estimates shows the change in the gradient that is due to differences in population risk factors over the two time periods. Since educational and mortality trends vary dramatically between men and women, we estimate all models separately for subsamples stratified by sex.

The basic equation is a proportional hazard model of mortality risk taking the form:

mortality(X,college,age)s=exp(β′∗Xs+λ∗colleges+γ∗ages),
(1)

where the s subscript specifies the survey (NHANES I or the NHIS in the main models). College is an indicator of whether a respondent reported >12 years of education, age is a continuous variable expressed in single years between 25 and 74, and X contains the set of behavioral risk factors. This specification offers a natural fit for the log-linear relationship between mortality risk and age.17 The models are estimated using maximum likelihood and we report hazard ratios throughout.

In the basic specification, X includes two indicators of smoking status (“current” and “former” smokers, with those never smoking the reference group) and four indicators of weight (underweight, overweight, obese class 1, and obese class 2 or 3, with healthy weight as the reference group). We experimented with more complete or flexible models by including controls for hypertension and hypercholesterolemia, interactions of smoking and obesity with education, or both. These results are discussed briefly below. One caveat to the overall approach is that we assume there are no changes in year effects within a given survey, so that we average mortality effects of risk factors across these years.

As mentioned, counterfactual estimates of death rates were calculated using data from one survey but the mortality equation from another. For instance, we predict what the mortality risk would have been in the NHANES I period if the population contained the risk factors present in the NHIS (between 1987 and 2000) by applying the parameters estimated from NHANES I (denoted with the subscript 71–75) to the distribution of risk-factors in the NHIS (subscripted with 87-00). For the college educated, this is calculated as:

E^[mortality(X,college,age;θ1971-75)∣college=1,survey=1987-2000]=1N∑i∈87-00,college=1exp(β71-75′∗Xi+λ71-75∗collegei+γ71-75∗agei).
(2)

We can compare this to the mortality hazard directly estimated from NHANES I to see how changes in the behavioral risks affected predicted mortality between the two time periods. Alternatively, we can compare the mortality rate calculated from (2) to that obtained using both the NHIS mortality equation and risk factors, to indicate how much of the total secular trend is due to changes in the model coefficients – that is how much occurs for reasons other than changes in the risk factors. In the tables, we perform the estimates described above separately by education groups and calculate both actual and counterfactual educational gradients.

The decompositions based on equation (1) are closely related to the well-known Oaxaca-Blinder (OB) decomposition. In fact, equation (1) demonstrates that an OB decomposition on log mortality risk observed by individuals would deliver identical results (in large samples) as our procedure. However, since individual mortality risk is not observable and needs to be estimated, this avenue is not open to us. 18

Because of small sample sizes in NHANES III, our primary trend decomposition is for NHANES I versus the NHIS. However, we also estimated mortality models for the 1990s and simulated predicted mortality using the smaller NHANES III sample, as shown in the Appendix.

The standard errors reported below are more conservative (i.e. larger) than those usually obtained in regression decompositions because we accounted for both estimating error in the mortality model and sampling error for the distribution of risk factors in different survey periods (rather than just the former as is typical).19 Appendix A details the process used to derive the correct standard errors.

5. Mortality Hazard Rate Estimates

Table 4 displays estimates of the mortality models described in equation (1) for subsamples stratified by sex, controlling for education, smoking and obesity status. Since we are estimating hazard models, coefficients greater (less) than one indicate increased (decreased) risk of death.

Table 4
Mortality Model with Controls for Education, Smoking and Obesity

Two findings stand out. First, the college-educated have lower expected mortality rates than their less educated peers, and this differential has increased over time, controlling for smoking and body weight (Model B). Column 1 shows that the predicted hazard rates of highly educated NHANES I men are a statistically insignificant 12% below those of non-college attendees. For NHANES III and NHIS males, the corresponding differences are a statistically significant 26% and 25%. College-educated women have a (statistically insignificant) 9% higher hazard rate than their counterparts in NHANES I; for NHANES III and NHIS females, the predicted rates are 12% and 20% lower. The estimated college education effect is statistically significant in the NHIS.

Obesity and smoking do explain a portion of the educational differential in mortality rates at a point in time, as expected since both represent significant health risks. This can be seen by noting that the predicted risk reduction associated with a college education is always larger in specifications that do not control for smoking or body weight (Model A of the table) than when these are included (Model B). In models without controls, college attendance is associated with a risk reduction of 22%, 31% and 34% for males in the NHANES I, NHANES III and the NHIS, and a -6%, 24% and 25% reduction for corresponding females.

Second, current smoking is associated with a much larger increase in mortality rates than other risk factors, and these adverse effects appear to have risen over time. Averaging across specifications, the hazard rates of current smokers are around twice as high as for persons who have never smoked. Severe (class 2 or 3) obesity also confers a substantial, although not always statistically significant, elevation in mortality risk, raising the predicted hazard rates by 44% to 128%. Being underweight or mildly obese predicts increased death rates, although usually by statistically insignificant amounts. Overweight (but non-obese) persons sometimes have slightly reduced mortality risk. The mortality equations estimated using the NHIS resemble those from NHANES III, except that the former are more precise. For this reason, we focus on results using the NHIS below.

6. Education-Mortality Gradients

We next use the hazard rate estimates obtained above to examine the extent to which trends in education-mortality gradients result from changes in behavioral risk factors. Throughout, we present mortality ratios with college attendees as the reference group, so that ratios above one indicate higher mortality for the less educated. Table 5 shows observed ratios (in bold) on the main diagonal. For instance, less educated males were predicted to have 31% higher mortality hazard rates than male college attendees in 1971–75, and the gap rose to 52% in 1987–2000; the 22 percentage point increase is shown as the third element of the main diagonal. Moving across each row in Table 5, we compare the mortality ratio based on early (NHANES I) or later (NHIS) profiles of risk factors. In other words, we ask how the education gradient in mortality would differ, given the observed trends in smoking and obesity but assuming that the returns to these risk factors in our estimated mortality models remained constant.

Table 5
Observed and Predicted Mortality Ratios and Differences

Table 5 shows that changes in population risk factors play essentially no role in explaining the increase in mortality differentials. As mentioned, less educated NHANES I males have a 31% higher risk of death than their college-attending peers. The predicted gap using the same hazard rate equation but the population risk factors from the NHIS – 15 to 25 years later –is a slightly smaller 27%. The second row shows similar results, but based on parameters estimated using the NHIS data: changes in risk factors imply a (statistically insignificant) 4 percentage point smaller mortality differential in the more recent period. Patterns of smoking and obesity play a similarly small role in all of the remaining comparisons in Table 5, regardless of whether estimates use parameters based on the NHANES I or NHIS data.

By contrast, the mortality decrement for being less educated, holding risk factors constant, worsens with time. Continuing with the previous example, predicted mortality hazard rates of the less educated NHANES I male sample would be 57% higher than for those with more schooling, had parameters from the NHIS mortality equation applied, exceeding the actual (52%) differential.

The bottom panel of Table 5 shows analogous results for females. Secular changes in smoking and obesity, from the 1970s to the 1990s account for a 2 to 3 percentage point rise in the education-mortality ratio, out of an actual 42 point total increase (from 0.92 to 1.34). Thus, trends in these risk factors explain less than 10% of the growth in the education-mortality gradient, with more than 90% resulting from changes in the returns to behavioral risks and education.

These results might be an artifact of the blunt measures of smoking used. To this point, we have considered tobacco use only at the extensive margin, however, changes at the intensive margin might be important if college-educated tobacco users have cut the amount they smoke by more than their non-college counterparts.20 Similarly, there could be differences in the timing of last use among former smokers, which could matter because health risks are elevated for a substantial period of time after smoking cessation. To address these possibilities, Table 6 displays simulated mortality ratios based on models that control for smoking intensity (<5, 6–10, 11–19, 20, 21–39, > 39 cigarettes on smoking days) and the time since quitting (<1, 2–5, 6–10, 11–15, 16–20, and >20 years).

Table 6
Observed and Predicted Mortality Ratios and Differentials With Additional Controls for smoking intensity and quit timing

The results in Table 6 are virtually identical to those in Table 5 for males, indicating that little was lost by using the parsimonious smoking covariates. In contrast, differential trends in the intensity and timing of tobacco use do seem to explain some of the rise in the excess mortality of less educated females. Specifically, in these models, trends in risk behaviors account for 7 to 17 percentage points of the total 43 point (from 0.92 to 1.35) increase in the gradient for women, or 20% to 40% percent of the total change. However these estimates are not statistically significant and the majority of the secular trend (60% to 80%) does not result from changes in measured behavioral health risks. One should still interpret changes in the “return” to risk factors like smoking with caution, however. Given the dramatic drop in smoking rates in the population, the composition of smokers (and importantly the difference between current smokers and others) has likely changed in ways that could amplify the correlation between smoking and mortality, even if smoking’s relationship to mortality remained unchanged.

The composition of education groups

The proportion of males attending college rose over 60 percent (from 33% to 54%) and that of females more than doubled (from 24% to 49%) between NHANES I and the NHIS, making it likely that the composition of individuals within these education groups also changed (e.g. the non-college educated were likely to be lower in the ability distribution in the 1990s than the 1970s). To test whether our results were sensitive to such compositional changes, we first estimated the probability of attending college within each NHIS survey year conditional on the age of individuals, region of residence, marital status, and income. Based on these estimates, we computed an index of the propensity to attend college. Next, within each survey year, we ranked college attendees using this index and reassigned the bottom 34% (44%) of such males (females) to the “high school or less” education category. This balances the share of people who attend college and not in the NHIS and the NHANES I. Under the assumption that the estimated propensities indicate who attends college at the margin, this reassignment should roughly preserve the education composition over time.21 Third, we estimated the basic mortality models for the revised NHIS data, and simulated mortality as before.22

The results of this exercise, summarized in Table 7, are similar to the findings in Table 5, indicating that our main findings are not being driven by compositional changes of education subgroups. Specifically, in Table 7, changing risk factors are estimated to explain none of the widening educational gap in mortality for men and no more than one-eighth of the increase for women, again indicating that the widening education gap in death rates primarily reflects changes in the mortality function rather than in the distribution of risk factors.

Table 7
Observed and Predicted Mortality Ratios and Differentials With Constant Share of Individuals in High Education Group

As a final check, we replicated the results using NHANES III rather than NHIS data (see Appendix Table A.1). Although smaller sample sizes imply less precise estimates, changes in smoking and obesity continue to explain little of the trend in education-mortality gradients. The male mortality ratio increases 15 percentage points (from 1.31 to 1.46), between 1971–75 and 1988–94, with changes in smoking and obesity predicted to reduce the gap by 6 or 7 points. For women, trends in the risk factors account for 7 to 10 percentage points of a 38 point rise (from 0.92 to 1.30) in the gradient.

Using the NHANES III data, we also examined patterns for alternative mortality models that: 1) added interactions between smoking and obesity, and 2) included these interactions plus supplementary controls for high blood pressure and cholesterol. These alternatives (also summarized in Table A.1) had essentially no effect on the results, except that changes in risk factors explained even less of the rising education gaps of women: with the most comprehensive controls, changes in the risk factors were responsible for a 0 to 7 percentage point widening of educational mortality ratios, compared with a 34 point increase in the actual gradient; as before, trends in the risk factors have no explanatory power for men.

The evidence that tobacco use explains a large share of point-in-time differences in mortality across groups but not trends in the gradient might seem surprising, given that smoking has declined more over time for the highly educated. However, a more comprehensive analysis of smoking patterns, detailed on Appendix Tables A.2 and A.3, shows that the differences in educational trends in smoking are concentrated among adults who are younger than the ages at which deaths from cancer usually occur, with either no differential or much smaller differentials for older individuals. For instance, current smoking among people aged 60 and older fell more between 1971–75 and 1987–2000 for less educated males (8 percentage points) than for their college educated counterparts (1 point), and it rose only slightly more for non-college educated women (4 versus 2 percentage points). For the entire population, in contrast, current smoking fell much more for college-educated males (13 percentage points) than among less-educated males (7 points). The widening smoking differential at younger ages raises the possibility that changes in this risk behavior could contribute to future increases in education-related cancer mortality gradients, as these persons age, but they do not explain growth in the gradient in the period we examine.

Unobserved factors and education

Given strong secular trends in education, one may also worry that education picks up some unobserved risk factor for mortality that is changing over time. If this were the case, our results suggesting that the protective effect of a college education increased over time would simply reflect omitted variables bias. We consider this possibility in relation to other important observed risk factors. Smoking and obesity are by far the most important behavioral risk factors identified in the health literature. For our results to be an artifact of an omitted risk factor, this risk factor would have to significantly raise mortality risk and its gradient with education would have to increase substantially over the time-period examined here. We use our estimates from NHANES I and the NHIS reported in Columns 1 and 3 of Table 4 to ask how large an omitted risk factor would have to be to generate the decline in the relative risk associated with a college degree that we estimate, a change from 0.88 to 0.75 between NHANES I and NHIS.

Our thought experiment assumes that a relative mortality risk comparable to that of being a smoker (i.e. the coefficient on this factor equals that on smoking in mortality models), and it assumes that in the early period, NHANES I, it is uncorrelated with any other characteristics in the mortality model. We then compute how large the mean difference in this risk factor would have to be in the later period to explain the difference in the education gradient between NHANES I and NHIS observed in Table 4. We estimate that the increase in the incidence of such an omitted risk factor would have to increase by 18–26 percentage points more among less educated than among more educated males. 23 Among females, the incidence would have to increase among the less educated by about 25–60 percentage points more than among the more educated. This increase in the education gradient of the omitted risk factor would therefore have to be 4–10 times larger than the increase in the smoking gradient reported in table 3A and ​and3B.3B. We believe it unlikely that there is an omitted behavioral risk factor that both carries as significant a relative risk as does smoking and whose relative incidence changed by the amounts required to explain the widening of the SES-gradient.

Finally, due to concerns regarding differences in reporting bias across education groups, we explored how self-reported vs. measured BMI differed by education in the NHANES III. Although reporting bias affects both education groups, differences by education were small and insignificant. Thus, we believe that differential bias in reported risk factors would not have substantial effects on trends in mortality ratios.

7. Cardiovascular and Cancer Mortality

In this section, we separately analyze the gradients in deaths due to CVD and cancer, the two most important sources of mortality in the United States.24 Both are influenced by behavioral risks such as smoking and obesity. Ideally, we would examine even more detailed categories – such as cancers closely linked to smoking – but sample size limitations prohibit this. It is possible to precisely measure cause-specific mortality ratios in the NHIS data, but small sample size in the NHANES I makes it difficult to get precise estimates of the changes in the mortality ratios by education, even for broad categories of cancer and CVD mortality. This limitation should be kept in mind when considering the following results.

Cancer and CVD account for 70% to 77% of all deaths in our data, with slightly higher shares for women than men and in NHANES I than in the NHIS. In relative terms, cancer has become more important over time for males as death from this source increased modestly (from 1.5% to 1.6%) while CVD mortality has fallen rapidly (from 2.9% to 1.8%). A similar pattern was observed among females: 5-year death rates from cancer increased slightly from 1.2% to 1.3% and those for CVD declined from 1.4% to 1.0%.

These overall patterns conceal strong differences across education groups shown in Table 8, particularly for cancer. Five-year death rates from cancer rose from 1.7% to 1.9% (1.0% to 1.4%) for less educated men (women), while declining from 1.4% to 1.3% (1.4% to 1.2%) among college attendees. Conversely, CVD death rates fell for both sexes and education groups, although more sharply for men than women. Between the 1970s and the 1990s, changes in male cancer deaths account for 20% of the entire increase of 1.5% percentage points in the raw 5-year mortality gap by education. The unadjusted education disparity for cardiovascular mortality actually declined during this time period. For females, we observe that the mortality gaps have been increasing for all causes, although trends in cancer deaths account for two-fifths of the 1.5 point increase in total 5-year mortality differences by education.

Table 8
Age-Adjusted 5-year Mortality Rates By Years of Education and Cause of Death

Hazard rate estimates for deaths from cancer and cardiovascular disease are displayed in Table 9. Estimates for the early 1970s are imprecise, as discussed, making comparisons across time periods difficult. For males, the estimated education gradient in cancer deaths rose between NHANES I and NHIS: the relative mortality hazard associated a college education declined from 1.12 to 0.77, implying that the protective effect of schooling increased. We do not observe a similar decline in CVD mortality risk among male college attendees. The pattern by broad disease group is reversed for females: hazard rates for the college-educated relative to others remained unchanged for cancer deaths, but declined significantly for CVD mortality (from 0.99 to 0.76).

Table 9
Mortality Model for Malignant Neoplasms & Cardiovascular Disease

We next calculated counterfactual cancer and CVD mortality ratios, corresponding to those for total mortality in Table 5. For both cancer and cardiovascular disease, there is again little evidence that the changing risk profiles explain the widening education-mortality gradient (Table 10). This is true regardless of the whether the mortality model is estimated using NHANES I or the NHIS. Instead, the widening gradient results from the increase in relative risk of death associated with behaviors like smoking, and, conditional on these behaviors, a rising relative risk of low education for cancer mortality among males and for CVD mortality for females. Our most robust finding is that risk factors did not vary in ways that made CVD or cancer deaths change differentially over time by education. Instead, the changing return to schooling and to risk factors has favored the more educated.

Table 10
Observed and Predicted Mortality Ratios and Differentials For Malignant Neoplasms and Cardiovascular disease

Our analysis does not indicate why the impact of the risk factors changed to increase educational gaps in mortality. To garner some information on this, we estimated models for two risk factors for cardiovascular mortality, cholesterol and hypertension, controlling for education, smoking, obesity, and gender. Both are quantitatively important and useful to examine because prescription drugs to control cholesterol and blood pressure have been widely credited with mitigating the role of these risk factors in deaths due to cardiovascular disease. For brevity, these results are not shown, but they are available upon request.

Importantly, the advantage of education for cholesterol control and hypertension control did not change over time– schooling conferred a protective effect against high cholesterol that remained nearly identical in each NHANES sample, and rates of hypertension were also similar across education groups in each of the four NHANES samples.25 Taken together, the results suggest that the health risk of obesity, as measured by hypertension, fell significantly over time. However, there is no suggestion that access to medical care, or compliance with complex medical regimens, for example, contributed to educational disparities in these important risk factors for CVD deaths.

8. Discussion

This paper examines how education-related disparities in mortality rates changed between the early 1970s and the end of the twentieth century, and investigates the role of behavioral risk factors in explaining these trends. Consistent with previous research, we document that the educational gradients are steeper for men than women but have widened over time more for females than males.

Our most striking finding is that the widening educational gaps in death rates are not explained by secular changes in key behavioral risk factors. Although higher levels of mortality among the less educated are partially due to their greater rates of tobacco use, smoking trends explain little if any of the increase in relative mortality risk for the less educated over the last three decades. This result occurs because gaps in smoking rates are much smaller among older adults, who have the highest death rates. In our main estimates, the mortality differential between males without and with college rose 22 percentage points, whereas corresponding trends in smoking and obesity predict a 4 point decrease, ceteris paribus. For women, patterns of smoking and obesity explain approximately 3 points out of the 41 percentage point increase in the relative risk of death for the less educated. Risk factors play a more important role in some (but not all) alternative specifications but never explain more than a small fraction of the rising education-gradient in mortality.

Instead, it is the return to education (conditional on health behaviors) and changes in returns to behaviors that are important. For example, the detrimental impact of smoking has strengthened over time for both men and women, as have the negative consequences of severe obesity for females (although small NHANES I sample sizes limit the precision of these estimates).

There are several possible explanations for these findings. One is that the highly educated have better access to medical care, and thus achieve superior results due to higher productive efficiency as described by Grossman (1972) and Kenkel (1991). This might become more pronounced over time, as sophisticated treatments improve the survival of those receiving them26 and new technologies lead to earlier detection.27 Adherence to prescribed regimes may also have become both more important and more difficult over time, yielding larger gains in life expectancy for highly educated individuals who have better adherence rates.28

Cancer for males and CVD among females play important roles in the increasing education gradients observed for total mortality. In cancer, the trend reduction in death rates appears to be due to improvements in screening, earlier diagnosis, and treatment (Cutler 2008).29 However, the same does not hold for CVD, as evidenced by the similar increases over time in the control of high cholesterol and hypertension for less and more educated groups.

A second explanation is that environmental risks may have declined more over time for the highly educated. For instance, changes in the nature of employment may have led to larger reductions in job stress and work-related health hazards for professional than manual occupations.30 There could also be similar patterns for geographically-based risks, such as those due to pollution, particularly if housing has become more geographically segregated by education, or because advantaged groups change their behaviors to avoid pollution in ways that other groups do not (see for example, Moretti and Neidell 2011). Chay and Greenstone (2003) demonstrated that recession-induced factory slowdowns and closings had beneficial impacts on infant health (due to the associated reductions in certain air pollutants), especially among disadvantaged groups, measured based on black versus white infant deaths. Using plant closings interacted with close proximity to an industrial plant, Currie, Davis, and Walker document that differential location near sources of pollution (such as plants), can explain about 6 percent of existing gaps in birth weight between the most advantaged (white college educated) and least advantaged (black high school dropouts) mothers. Similarly, by comparing infant health before and after superfund cleanups, Currie, Greenstone, and Moretti (2011) show significant improvements in infant health following superfund clean-ups. Still, these factors are relatively small.

Understanding the mechanisms for the effects we identify is important for achieving national goals of reducing disparities in longevity and medical outcomes. Given their importance in cross-sectional data, there is some sense in which we expect the most important risk factors for mortality, smoking, obesity, hypertension, and high cholesterol to explain differential mortality trends over time. Yet our simple exercise decomposing differences in mortality over time show that they do not. Thus, our research highlights a fundamental puzzle in the literature on mortality and socioeconomic status. Our results do not imply that improvements in the health-related lifestyles of the less educated yield no benefits. To the contrary, reducing smoking, obesity, hypertension and high cholesterol would improve health. However, the results suggest that even the complete elimination of disparities in behavioral risks across education groups would be unlikely to substantially reduce education-related differentials in mortality.

Supplementary Material

01

Footnotes

1We are not the first to look at these issues. Evelyn M. Kitagawa and Philip M. Hauser (1968) identified educational differentials in age-specific mortality, present in 1960, and there has been substantial related recent research (e.g. Feldman et al. 1989; Duleep 1989; Pappas et al., 1993; Crimmins and Saito 2001; Steenland, Henley, and Thun 2002; Wong et al. 2002; Lin et al. 2003; Singh and Siahpush 2006; Meara, Richards, and Cutler 2008), most finding that the educational gaps have increased over time.

2The third leading behavioral cause of death, alcohol use, accounts for only around 22,000 deaths annually (Heron et al. 2009). Many of these are from automobile accidents involving persons younger than 25, who are excluded from our analysis.

3See: http://www.cdc.gov/nchs/nhanes.htm for further information on the NHANES.

4The upper age limit was 74 years old for NHANES I and II.

5The NHANES III public use mortality data include perturbations of data elements to protect confidentiality of respondents. These do not affect the estimates of demographic differences in death rates.

6The three-digit ICD-9 codes corresponding to cancer and cardiovascular disease are 140–240 and 390–460.

71989 and 1996 are excluded due to a lack of information on smoking. Information on the NHIS is available at: http://www.cdc.gov/nchs/nhis.htm.

8The age and race/ethnicity restrictions eliminate 155,763 respondents. Another 11,454 observations were deleted due to missing information (on education, race/ethnicity, weight, height, smoking status or future mortality from the National Death Index follow-up) and 1,598 were excluded because they died within a year of the baseline interview.

9The ICD-10 codes indicating cancer are C00-C97, and for cardiovascular disease the codes are I00-I99.

10The exact wording of the questions changed in the 1992 NHIS to measure individuals who smoked on some days. There is no reason to believe this change would differentially affect smoking rates by education.

11Some researchers prefer anthropometric measures such as waist circumference (Sönmez et al. 2003), waist-hip ratio (Dalton et al. 2003), or waist-height ratio (Cox and Whichelow 1996). Burkhauser and Cawley (2008) recommend the use of Bioelectrical Impedance Analysis (BIA).

12In cases where systolic and diastolic blood pressure put a person in different categories, the more severe category is used, as is standard.

13These patterns resemble recent estimates by Meara, Richards, and Cutler (2008).

14Corresponding secular increases in education are found in other national data sets. For instance, Current Population Survey data indicate that the fraction of white males (females) 25 and older with a high school degree or more rose from 54.0% to 84.8% (55.0% to 85.0%) between 1970 and 2000 (U.S. Census Bureau 2009; Table 222).

15This occurs because the self-report data in the NHIS lead to an understatement of BMI, as discussed, and since the NHANES IV covers a later period (1999–2004 rather than 1987–96) and so picks up the substantial growth in weight occurring at the very end the 20th century.

16These results are consistent with previous research indicating that SES-BMI differentials are constant or narrowing over time (Zhang and Wang 2004; Chang and Lauderdale 2005).

17This specification gives rise to a Gompertz distribution of time to death.

18An alternative approach would be to perform OB decompositions based on linear probability models, using death within some pre-specified time period as the outcome. We rely on hazard models instead, because these allow us to make full use of the information available on the date of death and they have the advantage of restricting predicted mortality rates within any time-period to lie between zero and one. Further, it is well-known that the relationship between age and mortality is log-linear. A log-linear specification in a linear probability model would preclude use of the standard OB decomposition. A linear probability model would further impose the restriction that the percentage point increase in mortality associated with any risk factor is identical across ages, whereas we prefer the assumption, implicit in equation (1), that the relative risk associated with behaviors does not change with age.

19We find that for our standard errors are 20–25% larger than those obtained without accounting for errors in the distribution of risk factors.

20This possibility is salient given the importance of tobacco-related causes of death shown in prior literature (Wong et al. 2002; Meara, Richards, and Cutler 2008).

21This does not exactly preserve the age distribution within education groups. The less educated group in the later period has a younger average age and the more educated an older average age after reassignment. By construction, the age difference before reassignment is zero. After reassignment, the more educated are on average about 3 months older than the less educated, compared with before. This would tend to narrow any gap in disparities. In an alternative procedure (that yields nearly identical results) we duplicated all observations with education equal to some college (13–15 years of education). Second, we re-weighted these observations to shift 24.89% of women and 20.72% of men from the high to low education group in the NHIS data. This approach – randomly splitting the data until shares were equal by duplicating observations and re-weighting them – produced qualitatively and quantitatively similar results.

22The ability of this procedure to correct for selection depends on the power of the regression used to predict education. The R-squared values for the regressions used to construct this index vary between 0.15 and 0.2 depending on the survey year.

23Details of this calculation are available from the authors upon request.

24Heart and cerebrovascular diseases, both of which are included in CVD, were the first and third most common causes of mortality in 2006, together accounting for 31.7 percent of deaths; cancer was the second most frequent cause, being responsible for 23.0 percent of deaths (Heron et al. 2009).

25Our analysis revealed a striking change in the effect of obesity on hypertension over time. In NHANES III and IV, the elevated risk of hypertension among overweight individuals observed in earlier surveys nearly disappeared. This pattern did not hold for cholesterol.

26There is some evidence on this point from Goldman and Smith (2002) based on better adherence to complex treatment regimens for HIV and diabetes, and Cutler and Lleras-Muney (2008), who document better adherence to cancer prevention and screening among better educated. There is evidence of substantial SES differences in the receipt of cardiac catheterization and coronary revascularization following heart attacks (Alter, et al. 1999; Rathore et al. 2000). Similar differences in access to expensive cancer diagnostic technology and treatments might be expected.

27For instance, Smith (2007) shows that the overall prevalence of undiagnosed diabetes fell dramatically between 1976–80 and 1998–2002 but with much larger reductions for college attendees than for the non-college educated.

28Lange (2010) shows that less educated individuals have relatively low rates of screening for various cancers, especially among those with significant risk factors for cancer. The same patterns are observable in self-assessed cancer risk, suggesting that education might play a role in making individuals aware of their risks and consequently in obtaining adequate treatment.

29Recent evidence suggests that breast and prostate cancer are detected at later stages and treated less aggressively, once detected, for low compared to high SES individuals and that these differences are associated with an inverse SES-mortality relationship (Bouchardy et al., 2006; Byers et al., 2008; Rapiti et al., forthcoming).

30Ickovics, et al. (1997) show that life stress, social isolation and depression were all inversely related to social class (as measured by education and occupation) among men hospitalized following heart attacks.

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

References

  • Allison DB, Fontaine KR, Manson JE, Stevens J, VanItallie TB. Annual Deaths Attributable to Obesity in the United States. JAMA. 1999;282(16):1530–1538. [PubMed]
  • Alter DA, Naylor D, Austin P, Tu JV. Effects of Socioeconomic Status on Access to Invasive Cardiac Procedures and on Mortality After Acute Myocardial Infarction. New England Journal of Medicine. 1999;341(18):1359–1367. [PubMed]
  • Bouchardy C, Verkooijen HM, Fioretta G. Social Class is an Important and Independent Prognostic Factor of Breast Cancer Mortality. International Journal of Cancer. 2006;119:1145–1151. [PubMed]
  • Burkhauser RV, Cawley J. Beyond BMI: The Value of More Accurate Measures of Fatness and Obesity in Social Science Research. Journal of Health Economics. 2008;27(2):519–529. [PubMed]
  • Byers TE, Wolf HJ, Bauer KR, Bolick-Aldrich S, Chen VW, Finch JL, et al. The Impact of Socioeconomic Status on Survival After Cancer in the United States. Cancer. 2008;113:582–91. [PubMed]
  • Chang VW, Lauderdale DS. Income Disparities in Body Mass Index and Obesity in the United States, 1971–2002. Archives of Internal Medicine. 2005;165(18):2122–2128. [PubMed]
  • Chay K, Greenstone M. The Impact of Air Pollution on Infant Mortality: Evidence from Geographic Variation in Pollution Shocks Induced by a Recession. Quarterly Journal of Economics. 2003;118(3)
  • Chobanian AV, Bakris GL, Black HR, Cushman WC, Green LA, Izzo JL, Jr, et al. National Heart, Lung, and Blood Institute Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure. The Seventh Report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure: the JNC 7 report. JAMA. 2003;289:2560–72. [PubMed]
  • Cox BD, Whichelow MJ. Ratio of Waist Circumference to Height is Better Predictor of Death than Body Mass Index. British Medical Journal. 1996;313(7070):1487–1488. [PMC free article] [PubMed]
  • Crimmins EM, Saito Y. Trends in Healthy Life Expectancy in the United States, 1970–1990: Gender, Racial, and Educational Differences. Social Science and Medicine. 2001;52:1629–1641. [PubMed]
  • Currie Janet, Davis Lucas, Walker Reed. Department of Economics, Columbia University Working Paper. 2011. Toxic Releases and Health at Birth: Evidence from Plant Closures.
  • Currie Janet, Greenstone Michael, Moretti Enrico. Superfund Cleanups and Infant Health. American Economic Review: Papers and Proceedings of the Annual Meeting of the American Economic Association.May, 2011.
  • Cutler DM. Are We Finally Winning the War on Cancer? Journal of Economic Perspectives. 2008;22(4):3–26. [PubMed]
  • Cutler DM, Lleras-Muney A. Education and Health: Evaluating Theories and Evidence. In: House James, Schoeni Robert, Kaplan George, Pollack Harold., editors. Making Americans Healthier: Social and Economic Policy as Health Policy. New York: Russell Sage Foundation; 2008.
  • Cutler DM, Glaeser EL, Rosen AB. Is the US Population Behaving Healthier? In: Brown Jeffrey R, Leibman Jeffrey, Wise David., editors. Social Security Policy in a Changing Environment. Cambridge, MA: National Bureau of Economic Research; 2009.
  • Dalton M, Cameron AJ, Zimmet PZ, Shaw JE, Jolley D, Dunstan DW, et al. Waist Circumference, Waist-Hip Ratio and Body Mass Index and Their Correlation with Cardiovascular Disease Risk Factors in Australian Adults. Journal of Internal Medicine. 2003;254(6):555–563. [PubMed]
  • Duleep HO. Measuring Socioeconomic Mortality Differentials Over Time. Demography. 1989;26(2):345–351. [PubMed]
  • Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults. Executive Summary of The Third Report of The National Cholesterol Education Program (NCEP) Expert Panel on Detection, Evaluation, And Treatment of High Blood Cholesterol In Adults (Adult Treatment Panel III) JAMA. 2001;285:2486–97. [PubMed]
  • Feldman JJ, Makuc DM, Kleinman JC, Cornoni-Huntley J. National Trends in Educational Differentials in Mortality. American Journal of Epidemiology. 1989;129(5):919–933. [PubMed]
  • Flegal KM, Graubard BI, Williamson DF, Gail MH. Excess Deaths Associated with Underweight, Overweight, and Obesity. JAMA. 2005;293:1861–1867. [PubMed]
  • Goldman DP, Smith JP. Can Patient Self-Management Help Explain the SES Health Gradient? Proceedings of the National Academy of Sciences of the United States of America. 2002;99(16):10929–10934. [PMC free article] [PubMed]
  • Grossman M. The Demand for Health: A Theoretical and Empirical Investigation. New York: Columbia University Press; 1972.
  • Heron M, Hoyert DL, Murphy SL, Xu J, Cohanek KD, Tehada-Vera B. National Vital Statistics Reports. 14. Vol. 57. Hyattsville, MD: National Center for Health Statistics; 2009. Deaths: Final Data for 2006.
  • Ickovics JR, Viscoli CM, Horwitz RI. Functional Recovery After Myocardial Infarction in Men: The Independent Effects of Social Class. Annuals of Internal Medicine. 1997;127:518–525. [PubMed]
  • Kenkel D. Health Behavior, Health Knowledge, and Schooling. Journal of Political Economy. 1991;99(21):287–305.
  • Kenkel D. Unpublished Manuscript. Vol. 2007. Ithaca, NY: 2007. Mar, [accessed on September 2, 2009]. The Evolution of the Schooling-Smoking Gradient. at http://paa2007.princeton.edu/download.aspx?submissionId=71558)
  • Kitagawa EM, Hauser PM. Differential Mortality in the United State by Cause of Death, 1960. Demography. 1968;5(1):318–333.
  • Kuczmarski MF, Kuczmarski RJ, Najjar M. Effects of Age on Validity of Self-Reported Height, Weight, and Body Mass Index: Findings from the Third Health and Nutrition Examination Survey, 1988–1994. Journal of the American Dietetic Association. 2001;101(1):28–34. [PubMed]
  • Lange F. The Role of Education in Complex Health Decisions: Evidence from Cancer Screening. Journal of Health Economics. 2010 forthcoming. [PubMed]
  • Lin CC, Rogot E, Johnson NJ, Sorlie PD, Arias E. A Further Study of Life Expectancy by Socioeconomic Factors in the National Longitudinal Mortality Study. Ethnicity and Disease. 2003;13(2):240–247. [PubMed]
  • Marmot MG. Status Syndrome: A Challenge to Medicine. JAMA. 2006;295:1304–1307. [PubMed]
  • Meara E, Richards S, Cutler DM. The Gap Gets Bigger: Changes In Mortality And Life Expectancy, By Education, 1981–2000. Health Affairs (Millwood) 2008;27:350–360. [PMC free article] [PubMed]
  • Mokdad AH, Marks JS, Stroup DF, Gerberdin JL. Actual causes of death in the United States. JAMA. 2000;291:1238–1245. [PubMed]
  • Moretti E, Neidell M. Pollution, Health, and Avoidance Behavior: Evidence from the Ports of Los Angeles. Journal of Human Resources. 2011;46(1):154–175.
  • Must A, Spadano J, Coakley EH, Field AE, et al. The disease burden associated with overweight and obesity. JAMA. 1999;282(16):1523–1529. [PubMed]
  • National Heart, Lung, and Blood Institute. The Seventh Report of the Joint Committee on Prevention, Detection, Evaluation and Treatment of High Blood Pressure. Washington D.C: US Government Printing Office; 2004.
  • National Heart, Lung, and Blood Institute. Third Report of the National Cholesterol Education Program (NCEP) Expert Panel on: Detection, Evaluation and Treatment of High Blood Cholesterol in Adults (Adult Treatment Panel III), Final Report. Washington D.C: US Government Printing Office; 2002. [PubMed]
  • National Heart, Lung, and Blood Institute. Clinical Guidelines on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults: The Evidence Report. Washington D.C: US Government Printing Office; 1998. [PubMed]
  • Olshansky SJ, Passaro DJ, Hershow RC, Layden J, et al. A Potential Decline in Life expectancy in the United States in the 21st Century. New England Journal of Medicine. 2005;352(11):1138–1145. [PubMed]
  • Pappas G, Queen S, Hadden W, Fisher G. The Increasing Disparity in Mortality between Socioeconomics Groups in the United States, 1960 and 1986. The New England Journal of Medicine. 1993;329:103–109. [PubMed]
  • Preston SH, Elo IT. Are Educational Differentials in Adult Mortality Increasing in the United States? Journal of Aging and Health. 1995;7:476–496. [PubMed]
  • Rapiti E, Fioretta F, Schaffar R, Neyroud-Caspar I, et al. Impact of Socioeconomic Status on Prostate Cancer Diagnosis, Treatment, and Prognosis. Cancer Forthcoming. [PubMed]
  • Rathore SS, Berger AK, Weinfurt KP, Feinleib M, et al. Race, Sex, and the Medical Treatment of Myocardial Infarction in the Elderly. Circulation. 2000;102:642–648. [PubMed]
  • Singh GK, Siahpush M. Widening Socioeconomic Inequalities in US Life Expectancy, 1980–2000. International Journal of Epidemiology. 2006;35(4):969–979. [PubMed]
  • Smith JP. National Bureau of Economic Research Working Paper No 12905. 2007. Diabetes and the Rise of the SES Health Gradient.
  • Sönmez K, Akcakoyun M, Akcay A, Demir D, Duran NE, et al. Which Method Should Be Used to Determine the Obesity, in Patients with Coronary Artery Disease? (Body Mass Index, Waist Circumference or Waist-hip Ratio. International Journal of Obesity. 2003;27(3):341–346. [PubMed]
  • Sorlie PD, Johnson NJ. Validity of Education Information on the Death Certificate. Epidemiology. 1996;7(4):437–439. [PubMed]
  • Steenland K, Henley J, Thun M. All-Cause and Cause-specific Death Rates by Educational Status for Two Million People in Two American Cancer Society Cohorts, 1959–1996. American Journal of Epidemiology. 2002;156(1):11–21. [PubMed]
  • Stewart S, Cutler DM, Rosen AB. Forecasting the Effects of Obesity and Smoking on U.S. Life Expectancy. New England Journal of Medicine. 2009;361(23):2252–2260. [PMC free article] [PubMed]
  • U.S. Census Bureau. Statistical Abstract of the United States: 2009. 128. Washington, D.C: U.S. Government Printing Office; 2009.
  • U.S. Department of Health and Human Services. Healthy People 2010: Understanding and Improving Health. 2. Washington, DC: U.S. Government Printing Office; 2000.
  • Villanueva EV. The Validity of Self-Reported Weight in US Adults: A Population Based Cross-Sectional Study. BMC Public Health. 2001;21(1) Epub. [PMC free article] [PubMed]
  • Willett WC, Hu FB, Colditz GA, Manson JE. Underweight, Overweight, Obesity, and Excess Deaths. JAMA. 2005;294:551. [PubMed]
  • Wong MD, Shapiro MF, Boscardin JW, Ettner SL. Contribution of Major Diseases to Disparities in Mortality. New England Journal of Medicine. 2002;347(20):1585–1592. [PubMed]
  • World Health Organization. Report of a WHO Consultation on Obesity. Geneva, Switzerland: World Health Organization; 1997. Obesity: Preventing and Managing the Global Epidemic. [PubMed]
  • Zhang Q, Wang Y. Trends in the Association between Obesity and Socioeconomic Status in U.S. Adults: 1971 to 2000. Obesity Research. 2004;12(10):1622–1632. [PubMed]

Formats:

Loading... Loading portlets...