Education inequality in mortality: The age and gender specific mediating effects of cigarette smoking

Abstract

A debate within the mortality literature centers around the impact of health behaviors on the prospects of disadvantaged groups. Meanwhile, a growing body of work illustrates the social processes that shape changes in smoking levels by socioeconomic status (SES), especially educational attainment. These literatures are merged by examining the mediating effects of cigarette smoking on education gaps in U.S. adult mortality by age and gender. Findings reveal that cigarette smoking is an important mediator of the education-mortality gap for all males and for younger females. In particular, education-mortality gaps for young men narrow considerably when cigarette smoking is accounted for, while older women experienced no reduction in the education-mortality gap with controls for smoking. These results are consistent with diffusion arguments that describe SES differences in smoking adoption by age and gender and provide strong evidence that smoking is an important differentiator of mortality risks by education.

Two well-known findings suggest that the strong, inverse, graded relationship between education and mortality (Jemal et al. 2008; Lauderdale 2001; Pappas et al. 1993; Warren and Hernandez 2007) might stem in good part from cigarette smoking: smoking greatly increases the risk of premature adult mortality (e.g., Rogers et al. 2005) and smoking has become increasingly concentrated in low education groups (e.g., Pampel 2005). Smoking cigarettes contributes to about 20% of all deaths in the United States and is considered the single most preventable contributor to premature mortality (Himes 2010). Recognizing these findings, several studies have attempted to identify the role of lifestyle behaviors, smoking in particular, as mediators of the education and mortality relationship. But these studies and the claims made about the importance of smoking relative to other factors have generated some dispute.

On one side, Lantz et al. (1998) reject claims that elevated mortality risks among disadvantaged groups come primarily from their higher prevalence of risky behaviors. In comparing mortality risks across education groups over a 7.5-year period, they find that the odds ratio for the lowest to the highest education group falls by only 14% with controls for smoking, drinking, sedentary lifestyle, and relative body weight. Similarly, a study of British civil servants over a 25-year period shows that smoking and other coronary risk factors, such as high cholesterol, account for 27% of the inverse social gradient in coronary heart disease (Marmot 2006). In both cases, the methods involve comparison of the effects of education and other status characteristics on mortality risk, first without and then with controls for smoking and other health behaviors.

On the other side, studies using different methods find much stronger effects. Using indirect estimation techniques to attribute mortality to either smoking-related causes or other causes among men aged 36–69 in four nations, Jha et al. (2006) calculate that smoking accounts for nearly half of the excess mortality among the lowest socioeconomic stratum in each nation. Balia and Jones (2008) decompose a Gini coefficient measure of inequality in mortality and find that with controls for lifestyles, the contribution of education to inequality in mortality drops by up to 79%. Based on these results, smoking and other health behaviors seem to predominate in accounting for socioeconomic-based health disparities.

These differences in results have obvious theoretical and policy implications. Although advocates of the first view do not dismiss the importance of smoking by any means, they emphasize that it is only a relatively small component in accounting for education disparities in mortality risk. This conclusion answers those who might claim that disparities exist only because of freely chosen, harmful lifestyles. Thus, beyond smoking, underlying socioeconomic conditions do most to harm the health of disadvantaged groups through other mechanisms. In Marmot’s (2006: 341) words, claims that socioeconomic disparities stem largely from smoking might lead health officials to “forget social conditions, neighbourhood deprivation, employment conditions, early childhood and subsequent adult disease.” Link and Phelan’s (1995) fundamental cause perspective makes much the same point in criticizing policies that focus largely on behavioral risk factors. Even if differences in smoking and health behaviors disappeared, the relationship between education and health would change little—other sources of disparities would grow in importance and replace previous ones (Link and Phelan 1996).

While acknowledging the importance of structural factors, those advocating the second view highlight the policy value of tobacco interventions in reducing health disparities. Public health officials have accordingly included health behaviors in goals to reduce health disparities (Department of Health and Human Services 2008; Fagan et al. 2004). This approach need not neglect underlying social conditions because, as nearly all scholars would agree, social conditions shape the motivations, opportunities, and resources to smoke or not smoke. Among factors contributing to educational differentials in smoking are varied levels of stress, social capital (such as social ties with networks of smokers or non-smokers), exposure to ads and marketing, risks of dying from non-smoking causes, and efficacy in overcoming the obstacles to non-smoking. Even so, the enormous harm of smoking suggests that directly intervening to reduce the habit among disadvantaged groups can have important benefits for such groups, even without changing social conditions.

But both sets of arguments may miss a key point in understanding the influence of smoking on mortality disparities: the mediating impact may vary importantly across age and gender groups. Education groups have adopted smoking at different rates over the course of the last century, which would affect the size of disparities in smoking across age groups and their potential influence on mortality disparities. Such dispersion around the average may have important implications. Growth of disparities in smoking among younger relative to older people foretells greater health disparities in the future, whereas the opposite pattern suggests the influence of smoking will wane. Similarly, given different patterns of male and female smoking over past decades, the disparities may be larger for one gender than the other and affect future sex differences in mortality.

This paper thus takes a contingent approach to examining how smoking mediates the relationship between education and mortality.¹ It draws out theoretical predictions on the nature and direction of the varied influence of smoking on education differences in mortality across age and gender groups, and it then uses National Health Interview Survey (NHIS) Linked Mortality File (LMF) data to test the hypotheses. Similar to Lantz et al. (1998), the methods measure the mediating impact of smoking by comparing educational differences in mortality risk without and with controls for smoking. Although the effects of other components of socioeconomic status (SES) are considered and other health behaviors examined, education and smoking prove most important for mortality disparities and are the primary focus of the analysis.

A Diffusion Model of Disparities in Smoking and Mortality

Epidemiologists note that population changes in smoking take a form analogous to an epidemic that spreads from relatively small parts of a population to other parts and then eventually recedes (Lopez 1995; Lopez, Collishaw, and Piha 1994). In the early stages, smoking emerges first among high SES groups. During the middle stages of the process, smoking diffuses to the rest of the population but begins to decline first among high SES groups. In the later stages of the epidemic, smoking falls among all groups, but disparities widen as the decline occurs faster among high than among low SES groups.

To illustrate, in 1966 the prevalence of smoking among individuals aged 20 years and over in the United States differed little by education: 36.5% for those with less than a high school degree and 33.7% for college graduates. By 1987, the smoking prevalence rate remained about the same for individuals with less than a high school degree, at 35.7%, but fell by half to 16.3% for college graduates (Department of Health and Human Services 1989). And by 2007, the graded relationship between smoking and education had strengthened considerably: the percentage of adults aged 18 and above who were current smokers was 33.3 for those with 9–11 years of schooling, 23.7 for those with a high school diploma, 20.9 for those with some college but not an undergraduate degree, 11.4 for those with an undergraduate degree, and just 6.2 for those with a graduate degree (National Center for Health Statistics [NCHS] 2008).

Two closely-related mechanisms may account for the tendency of the epidemic to begin with high SES groups, diffuse to lower SES groups, and recede most among high SES groups. The first involves the diffusion of innovations through population networks of communication and imitation, whereas the second involves diverging incentives for healthy behavior. Both mechanisms imply a changing rather than an invariant relationship between SES and smoking and explain the current tendency for smoking to become concentrated among low SES groups.

First, a large literature on diffusion of innovations recognizes the tendency for high SES persons to most quickly adopt new ideas and behaviors (Rogers 2003; Strang and Soule 1998). The diffusion of the use of manufactured cigarettes, both a technological and cultural innovation (Griswold 1994), follows such a status-based pattern (Ferrence 1989). Given their greater openness to innovations and access to resources, high SES groups led, ironically, by physicians (Lopez, Collishaw, and Piha 1994), begin smoking earlier than the general public. Smoking spreads first within high SES networks, but later patterns of imitation lead to diffusion of the practice and normative change across classes and down the status hierarchy (Pampel 2005). As smoking diffuses to lower SES groups, however, new concerns about health emerge among higher SES groups, who are among the first to reject smoking (again led by physicians, and hastened by the negative publicity about the harm of smoking). The later diffusion of smoking to lower SES groups and the adoption of innovative health-promoting behaviors by higher SES groups serve to concentrate smoking among lower SES groups (Link 2008). But this does not occur until cigarette use has passed the early stages of diffusion and spread throughout the population (Ferrence 1989). Education plays a key role in this diffusion process. Highly educated persons adopt innovations most quickly (Rogers 2003), including the latest health-promoting technologies and behaviors (Glied and Lleras-Muney 2008; Link et al. 1998; Waldron 1991).

Second, changes in the balance of affordability and health costs of smoking for SES groups underlie the diffusion process (Cutler and Glaeser 2006; Pampel 2007). The early stages of cigarette diffusion tend to occur at lower levels of national income—early in the 20th century in the United States, for example. In this economic context, high SES groups can best afford the price of cigarettes. As national income grows, cost-sensitive groups can better afford manufactured cigarettes and their stimulating and addictive properties. As affordability and smoking rise among low SES groups, high SES groups respond to another factor. Their growing longevity accentuates the health costs of smoking so that the costs come to outweigh the benefits. For lower SES groups, however, the risks of premature mortality from non-smoking sources may limit the harm of smoking and the pace of the decline (Lawlor et al. 2003). Although couched in economic terms, the arguments relate to education differences. Smoking studies find that educational attainment heightens people’s motivations to use resources that help them quit, adopt effective therapies, and respond to antismoking campaign efforts (Cummings and Hyland 2005; Honjo et al. 2006; Niederdeppe et al. 2008).

Gender also plays a role in the diffusion process (Pampel 2002). Although a similar pattern of change occurs for women and men, the process of cigarette diffusion among women typically lags a decade or two behind men (Lopez 1995). Because men adopt cigarettes before women, the earliest stage of the epidemic shows a rising gap between men and women. In the middle stage, smoking among men levels off while it rises more quickly among women (particularly young and high status women), and the gap stops growing. In later stages, smoking continues to decline less slowly among women than men, and the gap narrows further. The difference in timing of adoption and diffusion of smoking produces varied education disparities among men and women. With women adopting smoking later, current education disparities tend to appear less strongly than for men, and among older women in particular, high rather than low SES groups tend to smoke more.

These status-based processes of change in cigarette smoking should produce diverse mortality disparities across age groups and genders (Ferrence 1989). Because smoking begins by adulthood for the vast majority of those adopting the addictive habit, the attitudes and behaviors at the time of a cohort’s adolescence will shape later patterns of smoking (Preston and Wang 2006). Among older age groups that entered adolescence many decades earlier during periods of growing cigarette use, high SES groups should show relatively high rates of smoking, education disparities in smoking should be modest, and smoking should have a relatively small role in accounting for mortality disparities. In contrast, among younger age groups that entered adolescence during periods of declining cigarette use, lower SES groups should show substantially higher rates of smoking, education disparities in smoking should be greater, and smoking should have a large role in accounting for educational mortality disparities. These age group differences thus reflect the diffusion of smoking down the status hierarchy over the past century. Again, however, because the diffusion process began later and has proceeded less far for women, the importance of smoking for education disparities in mortality should prove weaker for women than men.

Hypotheses

These general arguments predict variations in the mediating impact of smoking on the education-mortality relationship. To summarize,

• H1: Smoking should have less impact on mortality disparities among older than younger age groups because education disparities in smoking have increased in more recent decades.

• H2: Smoking should have less impact for explaining women’s educational differences in mortality than men’s because, consistent with the timing of diffusion, education disparities in smoking have widened more among men than women.

The contingent hypotheses imply that smoking will show the weakest mediating impact on the education-mortality gap among older women and the strongest mediating impact among younger men. The average across all age and gender groups will fall between the extremes.

Although other components of SES may have mediating influences as well, education is best suited for testing the hypotheses. Educational attainment has stronger influences on both smoking status (Barbeau, Krieger and Soobader 2004) and overall mortality (Muller 2002) than other components of SES, such as occupation and income. Education is associated with permanent income (Cowell 2006) and is established early enough in life to represent a cause rather than a result of smoking and ill health (Hummer and Lariscy 2010). Link and Phelan (1995 1996) establish a framework for understanding educational attainment as a social process through which persons form new relationships and support systems, create resources that enable them to acquire good health, and gain knowledge surrounding the benefits and detriments of a host of healthy or unhealthy behaviors, thereby reducing risks and lengthening life (see also Mirowsky and Ross 2003). This process begins early in life, cumulates through adulthood, and reflects the lifelong influence of education.

METHODS

Data

Data come from the public-use National Health Interview Survey Linked Mortality File (NHIS-LMF) for the year 1995, and years 1997 through 2000, merged with prospective mortality follow-up through December 2002 (NCHS 2007a; NCHS various years). NHIS-LMF is well-suited for our analyses because it includes detailed information on educational attainment and cigarette smoking and allows examination of mortality risk among a large non-institutionalized adult population. To conduct the analyses, individuals from the NHIS sample years are pooled and their records are linked to prospective mortality status. NCHS uses a probabilistic mortality matching scheme that assigns weights to multiple factors, including social security number, first and last name, and date of birth (NCHS 2007a).

The 1995 and 1997–2000 years of the NHIS contain a standard set of sociodemographic and health variables that vary little over time (NCHS various years). Information on cigarette smoking is available in the 1995 NHIS Year 2000 Objective supplement and in the 1997 to 2000 NHIS Sample Adult Files (SAF). NHIS annual household response rates are consistently 90% or better (NCHS various years) and are typically a little lower in topical and sample adult supplements. The response rate for the 1995 Year 2000 supplement was 81% and rates for the 1997 to 2000 SAF ranged from 70% in 1999 to 82% in 1997 (NCHS various years). The 1996 NHIS is excluded from the analyses because it does not include information on cigarette smoking.

The sample includes 123,817 adults age 25 and older, ages when most individuals have completed their education. Linking the 1995 supplemental data and the 1997 to 2000 SAF to mortality results in 7,030 deaths from all causes. Females contribute 53% of all deaths and the average age at death is just over 69 years. A small number of cases, 2.4%, are dropped because they are missing data on the key variables or because NCHS designates them ineligible to be linked to prospective mortality. NCHS (2007a) provides weights that adjust for the exclusion of ineligible records.

Variables

A consistent measure of educational attainment is constructed over the sample years that reflects whether each adult has more than 4 years of college (the referent), completed 4 years of college, some college but not 4 years completed, a high school education or GED, or less than a high school education. As expected, death is not evenly distributed among the educational categories. Just over 41% of all deaths occurred among individuals with a high school education or less.

Initial models control for age, sex, race/ethnicity, marital status, employment status, and income. Age is a continuous 5-year interval measure. Sex is coded with male as the referent. Sex-specific models are also included and adults are separated into three age groups: 25–49, 50–69, and 70 or older. Whereas the oldest age group includes 60.6% of all deaths, there remain a sufficient number of deaths for analyses in each of the age groups.

Race/ethnic dummy variables reflect whether a person is non-Hispanic white (the referent), non-Hispanic black, Hispanic, or some other race/ethnicity. Because previous work has established important mortality differentials by marital status (Rogers, Hummer and Nam 2000), the models include whether the person is married (the referent), divorced or separated, never married, or widowed.

Although education has some of the strongest SES effects on health and mortality (Lleras-Muney 2005), it is important to simultaneously consider the effects of employment and income. The independent effects of each of the socioeconomic indicators—education, employment, and income—are no doubt affected by each other. Because education affects mortality directly but also indirectly by increasing employment opportunities and income, its direct effect (with other SES controls) and total effects (without other SES controls) are examined.

Employment status represents whether the person is working (the referent) or not working or in the labor force. For income, the reference person for the household reports the total family income in categories defined by NCHS and then each member of the household is assigned that value. Income data (detailed below) in the NHIS person files is constructed and then merged to the individuals included in the sample adult and supplemental data files used here.

To approximate a continuous measure of income, a number of sequential steps are taken. First, the midpoint of each category is taken and a median value for the open-ended income category is estimated (see Parker and Fenwick 1983). Next, to account for the purchasing power of different sized families, the value is adjusted (see Van der Gaag and Smolensky 1982) and the consumer price index is used to adjust for changes in purchasing power over time. About 18% of the family income data were missing. Consequently, missing values are estimated using a number of covariates in the data, including a less detailed income measure that asked whether family income was above or at or below $20,000. To better represent the variability in the actual income data, stochastic variation is incorporated into the predicted values (Gelman and Hill 2007).

Terciles of income are generated and dummy variables are included to compare the mortality prospects of those in the lowest and middle income categories to those in the highest in the multivariate analyses. The robustness of the measure was investigated by using finer categories of income, categories identical to those used in Lantz et al.’s analysis (1998), and a logged transformation of the continuous measure to account for its skewed distribution. All these measures produce similar results to those using the generated terciles and, ultimately, the tercile variables are included to assist in interpretation of effects. Finally, estimating models with and without the missing income data check the consistency of the income measure. No discernible differences were found in these results.

To investigate the mediating effects of smoking status on the education-mortality relationship, smokers are delineated into never (the referent), former, and current statuses. Second, among current smokers and consistent with the literature (Rogers et al. 2005), smokers are coded as light (those who smoke 19 or fewer cigarettes a day), moderate (20 to 39 cigarettes a day), and heavy (40 or more cigarettes a day). A pack is typically 20 cigarettes.

Statistical Analyses

Cox proportional hazard models are used to examine the risk of overall mortality (Allison 1984). Duration in quarter-year intervals indicate the time to death in the models, and records are censored if respondents survived the follow-up period. Cox models are particularly useful because they do not impose a distribution of death across age nor do they require the analyst to choose a particular distributional form for the times of survival specified. The model declares that the hazard rate for the jth respondent is

$$\text{h}(\text{t}\mid {\text{x}}_{\text{j}})=\text{h}0(\text{t})exp({\text{x}}_{\text{j}}{\beta }_{\text{x}})$$ (1)

where the coefficients β_x are estimated from the data using a partial likelihood approach (Hoffman 2004).

Though Cox models make no assumption about the shape of the hazard of experiencing the event over time, estimates are generated assuming that, whatever the shape of the hazard, it is the same for all respondents, recognized as the proportional hazards assumption. This proportionality assumption was tested by generating Schoenfeld and scaled Schoenfeld residuals in Stata (Cleves, Gould, and Gutierrez 2004; StataCorp 2007) and calculating variable-by-variable and global chi-square statistics and associated p-values for the full model as well as the models stratified by age and sex (available upon request). Global tests indicate the full model and the full models by gender violate the proportionality assumption. The models stratified by age groups generally hold to the global proportionality assumption, and the variable-specific tests reveal that the effects of our main measures of interest, education and smoking, do not vary over the study period. These results and evidence that Cox estimates are sturdy despite proportionality violations when based on large nationally representative samples (Therneau and Grambsch 2000) provides some confidence in the analytic approach used here.

Model fit is evaluated by calculating G statistics as

$$\text{G}=-2^{\ast }(log\text{likelihood}\text{of}\text{model}1-log\text{likelihood}\text{of}\text{model}2)$$ (2)

with a χ² distribution and degrees of freedom equal to the number of new variables included between models (Hosmer and Lemeshow 2000). All results are reported as hazard ratios (HRs). Stata 9.0 (StataCorp 2007) incorporates sample weights and estimates robust standard errors that account for the NHIS stratified and clustered sampling design (NCHS various years).

RESULTS

Table 1 highlights important differences in smoking status by education for all adults, across age groups, and by sex. Panel A provides a description of the entire sample and shows that current smoking status decreases with increasing educational attainment. For example, 29.2% of those with less than a high school education are current smokers, whereas only 9.9% of those with post-baccalaureate education are current smokers. Panel A reveals important age differences in the proportions of smokers by educational attainment. For 25 to 49 year olds with less than a high school education, 43.9% are never smokers, 13.2% former smokers, and 42.9% current smokers. Similarly educated 50 to 69 year olds have over twice the percentage of former smokers but a substantially lower percentage of current smokers. Compared to the younger age groups, fewer individuals aged 70 and older are current smokers, regardless of educational attainment. Similar to the younger age groups, the proportion of current smokers declines with increasing educational attainment among those 70 and older, but only 10.0% of those with the lowest educational attainment at these ages currently smoke.

Table 1.

Smoking Status by Education, Age Group, and Gender, U.S. Adults, 1995–2000.^a, ^b

		Age Groups
	All Ages	25 to 49 years	50 to 69 years	70 years and older
Panel A. Full Sample (N = 123,817)
Never Smoker
Less than High School	45.2 %	43.9 %	37.4 %	54.4 %
High School, GED	45.8	45.2	42.5	53.6
Some college	49.0	52.3	39.9	49.5
4 years of college	61.6	68.2	44.7	49.6
More than 4 years of college	64.8	72.5	53.7	51.1
Former Smoker
Less than High School	25.6	13.2	30.3	35.6
High School, GED	24.7	16.5	32.8	36.5
Some college	25.7	18.9	36.7	41.3
4 years of college	24.1	17.0	39.7	43.9
More than 4 years of college	25.3	17.3	36.0	42.7
Current Smoker
Less than High School	29.2	42.9	32.3	10.0
High School, GED	29.5	38.3	24.7	9.9
Some college	25.3	28.8	23.4	9.2
4 years of college	14.3	14.8	15.6	6.5
More than 4 years of college	9.9	10.2	10.3	6.2
Panel B. Males (N = 53,393)
Never Smoker
Less than High School	30.1 %	36.1 %	23.5 %	28.7 %
High School, GED	36.2	41.1	27.5	31.3
Some college	42.7	49.4	29.1	30.1
4 years of college	57.1	65.9	38.1	38.3
More than 4 years of college	61.8	71.3	50.1	44.7
Former Smoker
Less than High School	35.0	16.1	39.6	59.0
High School, GED	30.1	18.1	44.0	58.6
Some college	29.8	19.7	45.9	60.8
4 years of college	27.2	17.6	44.7	57.2
More than 4 years of college	28.3	17.9	39.9	50.1
Current Smoker
Less than High School	34.9	47.8	36.9	12.3
High School, GED	33.7	40.8	28.5	10.1
Some college	27.5	30.9	25.0	9.1
4 years of college	15.7	16.5	17.2	4.5
More than 4 years of college	9.9	10.8	10.0	5.2
Panel C. Females (N = 70,424)
Never Smoker
Less than High School	56.4 %	50.9 %	48.4 %	68.5 %
High School, GED	52.9	48.9	52.2	63.4
Some college	53.7	54.6	48.1	60.5
4 years of college	65.8	70.2	52.0	60.5
More than 4 years of college	68.1	73.5	58.1	60.1
Former Smoker
Less than High School	18.6	10.6	22.9	22.7
High School, GED	20.7	15.2	25.5	26.7
Some college	22.5	18.3	29.7	30.2
4 years of college	21.2	16.5	34.2	31.2
More than 4 years of college	22.1	16.8	31.2	32.1
Current Smoker
Less than High School	25.0	38.5	28.7	8.8
High School, GED	26.4	35.9	22.3	9.9
Some college	23.8	27.1	22.2	9.3
4 years of college	13.0	13.3	13.8	8.3
More than 4 years of college	9.8	9.7	10.7	7.8

Source: derived from NCHS: 1995, 1997–2000.

^aSmoking statuses for each education group add to 100 percent.

^bAll results represent weighted proportions.

Panels B and C of Table 1 describe distributional differences of smoking status by educational attainment stratified by sex and age. A greater proportion of males with less than a high school education are current smokers compared to the full sample. This is true for all males and in all age groups with comparable levels of educational attainment. Although descriptive, the smoking by education gradient is most visible and consistent with diffusion arguments for the youngest group of males, those aged 25 to 49. Among this age group there are large proportions of current smokers in the lowest educational categories—47.8% of less than high school educated, 40.8% of high school educated, and 30.9% of some college educated males in this age group are current smokers. Although the proportions of current smokers by educational attainment for all females are smaller than for males, patterns similar to males emerge in the two youngest age groups, 25 to 49 and 50 to 69 (compare Panels B and C). For example, 38.5% of females aged 25–49 with less than a high school education and 35.9% with a high school education or GED currently smoke. Further, the proportion of current smokers among females in this age group steadily declines to 9.7% for those with greater than 4 years of college.

Table 2 presents hazard ratios for the risk of death for adults of all ages and for three broad age groups. There is an inverse association between educational attainment and mortality risk among all adults, net of other important confounding factors (Panel A, Model 1). Compared to those with more than 4 years of college education, those with 4 years of college experienced a 14% higher risk of death, and those with less than high school education experienced a 56% higher risk of death over the follow-up period. Panel A, Model 2 shows the impact that smoking has on mortality and on the education gap in mortality. Compared to never smokers, former smokers exhibit 38% higher mortality, and current heavy smokers show a 2.6-fold increased risk of death over the follow-up period. Including smoking status in Model 2 decreases the relative risk of death for the least compared to the most educated from a 1.56 to 1.44 HR; thus, controlling for smoking reduces the education gap in mortality by over 20% ([1.56−1.44]/[1.56−1]*100). Similar but slightly larger percentage reductions in the relative risk of death by education occur for high school educated individuals, with a decline from a 1.40 to a 1.29 HR, and some college educated individuals, from a 1.39 to a 1.30 HR.

Table 2.

Hazard Ratios for Overall Mortality among All Adults and Select Age Groups, U.S., 1995–2002 (N=123,817).

	A. All Ages (N = 123,817)		B. Age 25–49 (N = 70,449)		C. Age 50–69 (N = 33,961)		D. Age 70+ (N=19,407)
	Model 1	Model 2a	Model 1	Model 2b	Model 1	Model 2c	Model 1	Model 2d
Education (more than 4 years college)
Less than High School	1.56 **	1.44 **	2.36 **	1.96 **	1.49 **	1.31 **	1.46 **	1.42 **
High School, GED	1.40 **	1.29 **	1.96 **	1.69 **	1.31 *	1.21	1.30 **	1.25 **
Some college	1.39 **	1.30 **	1.90 **	1.70 **	1.38	1.26	1.23 **	1.19 **
4 years of college	1.14 **	1.11 **	1.34 **	1.31 **	1.06	1.02	1.14 *	1.12
Age (continuous 5 yr intervals)	1.39 **	1.43 **	1.47 **	1.46 **	1.30 **	1.34 **	1.51 **	1.58 **
Sex (female)
Male	1.78 **	1.59 **	1.82 **	1.71 **	1.78 **	1.58 **	1.76 **	1.55 **
Race/Ethnicity (non-Hispanic white)
non-Hispanic black	1.09 **	1.11 **	1.09	1.18 **	1.07 **	1.11 *	1.06 *	1.07 **
Hispanic	0.82 **	0.89 **	0.96	1.11	0.76 **	0.85 **	0.76 **	0.79 **
Other	0.94	1.01	1.30 **	1.39 **	0.91	0.99	0.81 **	0.86 **
Marital status (married)
Divorced or separated	1.40 **	1.28 **	1.62 **	1.52 **	1.36 **	1.26 **	1.25 **	1.16 **
Never married	1.41 **	1.42 **	1.54 **	1.48 **	1.45 **	1.47 **	1.22 **	1.23 **
Widowed	1.21 **	1.20 **	1.09	1.04	1.30 **	1.26 **	1.16 **	1.13 **
Employment (working)
Not working	1.98 **	1.90 **	2.54 **	2.48 **	1.91 **	1.87 **	1.59 **	1.55 **
Income (highest third)
middle third	1.20 **	1.18 **	1.37 **	1.32 **	1.24 **	1.20 **	1.13 **	1.12 **
lowest third	1.40 **	1.37 **	1.73 **	1.61 **	1.68 **	1.61 **	1.21 **	1.20 **
Smoking status (Never smoker)
Former smoker		1.38 **		1.18 **		1.45 **		1.42 **
Current light smoker		1.96 **		1.60 **		1.99 **		2.10 **
Current moderate smoker		2.06 **		1.86 **		1.96 **		2.25 **
Current heavy smoker		2.57 **		1.92 **		2.42 **		3.08 **
Log Likelihood	−31352.0	−31135.7	−3399.6	−3371.8	−7395.2	−7327.3	−14294.9	−14175.3

Source: derived from NCHS: 1995, 1997–2000.

Note: referent in parentheses.

^aModel 2 significantly better than Model 1, X² = 432.6, p ≤ . 01.

^bModel 2 significantly better than Model 1, X² = 55.6, p ≤ . 01.

^cModel 2 significantly better than Model 1, X² = 135.8, p ≤ . 01.

^dModel 2 significantly better than Model 1, X² = 239.2, p ≤ . 01.

^*p ≤ .05;

^**p ≤ . 01

The mediating effect of smoking on the relationship between education and mortality varies by age group. Educational attainment is most strongly linked to prospective mortality for the youngest age group (compare Model 1 in Panels B, C, and D). In fact, those with some college or less education experience nearly double the risk of death compared with those in the highest educational category. Smoking exhibits some of its strongest mediating effects on the education-mortality relationship among young adults—the group predicted by diffusion arguments to show the clearest class distinctions. Introducing smoking in Model 2 reduces the HR for individuals aged 25–49 by 22% for those with some college, by 28% for high school graduates, and by 29% for those with less than a high school degree, compared to those with more than 4 years of college.

Smoking also mediates the effect of educational attainment on mortality for those aged 50–69. Although the risk of death for the less than high school educated in Panel C, Model 1 for 50–69 year olds is not as high as in the youngest cohort, smoking reduces the educational gap in mortality by over 30% or more for this age group (see Panel C, Model 2). For persons aged 70 and older, educational attainment is related to mortality risk, but exerts more modest effects. The educational gap in mortality is reduced by 10 to 17% with the introduction of smoking (compare Panel D, Models 1 and 2) among those in the oldest age group. Using less than high school educated persons as an example, Figure 1 shows the differences in these reductions for all age groups and for the young and old.

Figure 1.

Percent Reduction to Overall Mortality for Less Than High School Educated Persons after Accounting for Cigarette Smoking Status, U.S. Adults, 1995–2002.

Finally, Table 2 shows that the mediating effects of smoking on income disparities in mortality vary by age and are important but not as important as the mediating effects on the education gap in mortality. For adults of all ages, smoking mediates the income disparity in mortality for the lowest income group by about 7%, from a 1.40 to a 1.37 HR. That reduction is 10% for 50 to 69 year olds and reaches 16% for the youngest age group, from a 1.73 to a 1.61 HR. Similarly small evidence of mediation by smoking appears for employment status.

Table 3 disaggregates the age-specific relationships between smoking and education from Table 2 by gender. Consistent with the lag in smoking adoption among females posited by diffusion arguments, the mediating effect of smoking on the relationship between education and mortality among adults of all ages is stronger for men than for women (compare all ages, Models 1 and 2, Panels A and B). For example, smoking reduces the relative risk of death for less than high school educated males compared to males with more than 4 years of college by 31%, from a 1.52 to a 1.36 HR. The same comparison for females yields a mere 6% reduction, from a 1.68 to a 1.64 HR.

Table 3.

Hazard Ratios for Overall Mortality by Age and Gender, U.S. Adults, 1995–2002.^a

	All Ages		Age 25–49		Age 50–69		Age 70 and older
	Model 1	Model 2b	Model 1	Model 2c	Model 1	Model 2d	Model 1	Model 2e
Panel A. Males (N = 53,393)
Education (more than 4 years college)
Less than High School	1.52 **	1.36 **	2.29 **	1.84 **	1.43 **	1.24 **	1.50 **	1.39 **
High School, GED	1.35 **	1.22 **	1.86 **	1.57 *	1.35 **	1.21	1.26	1.17
Some college	1.39 **	1.29 **	1.80 **	1.59 *	1.51 **	1.38 **	1.19	1.12
4 years of college	0.99	0.97	1.08	1.05	0.99	0.96	1.00	0.98
Smoking status (Never smoker)
Former smoker		1.31 **		1.25 **		1.34 **		1.34 **
Current light smoker		2.05 **		1.60 **		2.13 **		2.20 **
Current moderate smoker		2.16 **		2.00 **		1.97 **		2.46 **
Current heavy smoker		2.53 **		1.78 **		2.45 **		3.25 **
Log Likelihood	−11807.8	−11702.2	−1482.6	−1470.1	−3186.1	−3150.0	−4335.6	−4286.7

Panel B. Females (N = 70,424)
	Model 1	Model 2f	Model 1	Model 2g	Model 1	Model 2h	Model 1	Model 2i
Education (more than 4 years college)
Less than High School	1.68 **	1.64 **	2.65	2.26	1.62 *	1.45 **	1.49 **	1.57 **
High School, GED	1.52 **	1.47 **	2.20	1.91	1.37	1.29	1.38 **	1.42 **
Some college	1.45 *	1.41	2.15 *	1.93	1.38	1.27	1.26 **	1.29 **
4 years of college	1.40	1.39	1.84	1.80	1.34	1.27	1.32 **	1.34 **
Smoking status (Never smoker)
Former smoker		1.47 **		1.11		1.68 **		1.52 **
Current light smoker		1.94 **		1.65 **		1.91 **		2.16 **
Current moderate smoker		1.90 **		1.71 *		2.00 **		1.87 **
Current heavy smoker		2.62 **		2.61 **		2.09 **		2.96 **
Log Likelihood	−14824.3	−14710.9	−1384.0	−1373.2	−2923.1	−2887.9	−7212.5	−7143.4

Source: derived from NCHS: 1995, 1997–2000.

Note: referent in parentheses.

^aAll models control for age, race/ethnicity, marital status, employment, and income.

^bModel 2 significantly better than Model 1, X² = 211.2, p ≤ . 01.

^cModel 2 significantly better than Model 1, X² = 25.0, p ≤ . 01.

^dModel 2 significantly better than Model 1, X² = 72.2, p ≤ . 01.

^eModel 2 significantly better than Model 1, X² = 97.8, p ≤ . 01.

^fModel 2 significantly better than Model 1, X² = 226.8, p ≤ . 01.

^gModel 2 significantly better than Model 1, X² = 21.6, p ≤ . 01.

^hModel 2 significantly better than Model 1, X² = 70.4, p ≤ . 01.

ⁱModel 2 significantly better than Model 1, X² = 138.2, p ≤ . 01.

^*p ≤ .05;

^**p ≤ .01.

Males and females are further differentiated by age. Smoking has a considerable mediating effect among the youngest and least educated males. Indeed, compared to males aged 25–49 with more than four years of college, smoking reduces the mortality gap for comparably aged males by 26% for those with some college, 34% for those with a high school degree, and 35% for those with less than a high school degree. Smoking also reduces the male mortality gap between the education extremes (those with more than four years of college compared to those with less than a high school degree) by 44% among those aged 50–69. Although the relative risks of death by education are high for females aged 25–49, only females with some college education are at a significantly higher risk of death compared to females with more than 4 years of college. Notably, controlling for smoking reduces that risk by 26%, from a 2.15 to a 1.93 HR.

Finally, the results for the oldest cohorts of men and women reveal very different findings. For ages 70 and older, smoking moderately reduces the risk of mortality by education for men but has no effect for women. For example, for the least educated oldest men compared to the most educated oldest men, smoking reduces the mortality risk by 22%, from a 1.50 to a 1.39 HR. But for women in this same age range, the relative risk of death actually increases when smoking status enters the equation. Using less than high school educated men and women as examples, Figure 2 shows the magnitude of the gender differences for all ages and for the selected age groups.

Figure 2.

Percent Reduction to Overall Mortality by Age and Gender to All-Cause Mortality for Less Than High School Educated Persons after Accounting for Cigarette Smoking Status, U.S. Adults, 1995–2002.

It is possible that the mediating effects of cigarette smoking on the education mortality gap are heightened when considering education as the only indicator of SES. Therefore, analyses were conducted excluding employment status and income to see if the total effects were larger than the direct effects controlling for other SES indicators. The results from these analyses proved similar to the results reported in Tables 2 and 3 (available upon request). For example, in models excluding employment status and income, cigarette smoking reduces the education mortality gap for less than high school educated individuals by 20%, a reduction nearly identical to that in Model 2 for all ages in Table 2. For young males with less than a high school degree, smoking reduces the education mortality gap by 36% in the education only models, compared to a 35% reduction in the models presented in Table 3. As expected, HRs are larger when excluding the employment and income controls, but the mediating effect of smoking is virtually identical.

To extend the results for smoking, other health behaviors and detrimental outcomes were analyzed to determine their effects on the education mortality gap. Items are available in different supplement years of the NHIS and include obtaining flu shots, getting routine physician checkups, wearing seatbelts in automobiles, alcohol consumption, body mass index, and weekly vigorous physical activity. Unfortunately, no NHIS supplements contain all of these items. With the exception of weekly vigorous physical activity, these behaviors did little to mediate the education gap in mortality. The mediating effects of weekly vigorous physical activity are similar in size to the effects of smoking for the entire sample but differed less across age and gender groups (available upon request). For example, controlling for vigorous physical exercise mediates the education mortality gap for less than high school educated individuals by 20%, nearly identical to the mediating effect of smoking in Model 2 for all ages in Table 2. For young and middle aged males with less than a high school education, vigorous physical activity mediates the education mortality gap by 14 and 27%, respectively, whereas smoking in Table 3 for those same individuals shrinks the gap by 35 and 44%, respectively.

DISCUSSION

The education gap in overall mortality is multifaceted, complex, and due in part to cigarette smoking. Controlling for cigarette smoking, one of the most pernicious individual behaviors, explains part of the education gap in mortality, especially among younger adults. Among all adults, smoking accounts for about 20% of the educational gap in U.S. adult mortality. This overall result proves similar to findings reported by Lantz et al. (1998) and Marmot (2006), a result that might be expected given use of the same general methodological strategy. However, our results show also considerable variation by gender and age. At the extreme, smoking explains up to 44% of the education-mortality gap for middle-aged men, results that come closer to those of Jha et al. (2006). In contrast, controlling for smoking actually widens rather than explains education mortality disparities among older women—the opposite of what occurs for younger women and males. Thus, stratifying by age and gender initiates a detailed look into the potential consequences of variations in health behaviors and their impacts on mortality inequality by education.

The results provide indirect support for diffusion perspectives (Glied and Lleras-Muney 2008; Rogers 2003) that have been applied to smoking adoption (Pampel 2001; Pampel 2005; Pampel 2007; Preston and Wang 2006). These studies show that smoking adoption follows a process of change by SES. Consistent with these perspectives and in support of our hypotheses, the mediating effects of smoking on educational differences in mortality are greater among males than females, and greater among younger than older age groups.

The stronger mediating effect of smoking on the education gap in mortality at younger ages may be a consequence of stronger disparities in smoking status by educational attainment among younger populations. The robust disparities in smoking appear to result from a shift in class-based rejection of smoking and preferences for healthier lifestyles. Our results suggest that those with higher levels of education are increasingly living healthier lifestyles, whereas individuals with lower levels of education may rely more heavily on smoking as a relatively cheap and convenient way to escape, obtain immediate gratification and pleasure, cope, relax, and relieve stress (Waldron 1991). These changes greatly affect educational inequalities in mortality.

Our results show that smoking mediates the educational effects of mortality for younger but not older women (ages 70 and above). This finding is consistent with a lag in female adoption of smoking and with current smoking trends, whereby smoking is becoming more heavily concentrated among younger and less-educated women. Though both males and females of all ages generally experience a strong, graded effect of smoking on mortality, older females lagged behind older males in the early stages of cigarette diffusion in the United States and consequently show little variation in smoking adoption by educational attainment and less gradation of the effects of smoking on mortality.

Based on these findings, social policies could increase life expectancy by reducing the prevalence of smoking among all adults, and further close the education-mortality gap by reducing the disproportionately high smoking prevalence rates among the least educated. Although the age-adjusted cigarette smoking prevalence rate of adults aged 25 and older in the United States has experienced tremendous declines—from 36.9% in 1974 to 20.3% in 2005—it did not change between 2005 and 2006 (NCHS 2007b). Meanwhile, government laws, regulations, restrictions, warnings, taxes, and bans have all succeeded in reducing smoking prevalence (Shafey, Dolwick and Guindon 2003), but have not reduced smoking disparities by education. Further concentration of smoking among persons of lower socioeconomic statuses may result in increasingly difficult future reductions in the overall smoking prevalence rate. Individuals with lower levels of education may have fewer reasons to quit, may adopt ineffective therapies, and may be less responsive to antismoking campaigns (Cummings and Hyland 2005; Honjo et al. 2006; Waldron 1991). Aggressive efforts to curb and ultimately eliminate smoking among the less educated could help close socioeconomic differences in U.S. adult mortality.

Because education is a fundamental cause of health and longevity, individuals with lower levels of education may quit smoking but substitute another risky behavior for cigarettes, which may dampen the overall impact of smoking cessation (see Link and Phelan 1996). Nevertheless, it is especially important to reduce smoking prevalence rates because tobacco consumption is addictive, often clusters with other risky behaviors (such as binge drinking), and places individuals “at risk of risks” (Link and Phelan 1995). Furthermore, the risks of tobacco consumption accumulate over time, affect multiple organs, and result in multiple causes of death, including ischemic heart disease, chronic obstructive pulmonary disease, and lung cancer (Himes 2010). For these reasons, most risky behaviors that individuals may adopt to compensate for quitting smoking would be far less life-threatening.

To our knowledge, no previous work has investigated the mediating effect of smoking on overall mortality by detailed age and gender groups. The NHIS-LMF provides a unique resource through its prospective mortality follow-up. On the negative side, the NHIS years used here are pooled repeated cross-sectional data so our results do not assess changes, for example, in smoking status between the time of interview and time of censoring or death. Our relatively short follow-up period suggests that any time-varying smoking effects would be modest and would most likely result in a slight underestimate of the effects of smoking on mortality, because most transitions would be from current to former smokers or from reductions in consumption among current smokers (see Jousilahti et al. 1999).

Our results underscore the importance of better understanding the interrelationships between SES, risky health behavioral profiles, and survival disadvantage. Future work should explore the mediating effects of other health behaviors, social support networks, and social resources that may also help to differentiate the mortality patterns of the more and less highly educated segments of the population and provide insights on how such a gap could be closed. Following Link and Phelan (1996) and Phelan and colleagues (2004), if we wish to reduce health and mortality inequalities, we must address the social inequalities that drive the differences. Evidence suggests that educational inequalities in mortality are increasing (Meara, Richards, and Cutler 2008) and patterns of detrimental health behaviors like cigarette smoking are partially to blame.

We add to the debate within mortality literature centering on the effects of health behaviors on inequality in two important ways. First, we show that the impact of health behaviors on mortality inequality is an age- and gender-specific narrative. Second, with regard to health damaging behaviors and in this case cigarette smoking, social patterns of behavior carry on to patterns of mortality. Our results highlight the value of social perspectives on adoption of harmful behaviors. We extend the diffusion of innovation perspective by showing that the patterns of change operating to influence smoking adoption are reflected in mortality prospects, further contributing to inequalities in length of life.