Educational Disparities in Adult Mortality Across U.S. States: How Do They Differ, and Have They Changed Since the Mid-1980s?

Abstract

Adult mortality varies greatly by educational attainment. Explanations have focused on actions and choices made by individuals, neglecting contextual factors such as economic and policy environments. This study takes an important step toward explaining educational disparities in U.S. adult mortality and their growth since the mid-1980s by examining them across U.S. states. We analyzed data on adults aged 45–89 in the 1985–2011 National Health Interview Survey Linked Mortality File (721,448 adults; 225,592 deaths). We compared educational disparities in mortality in the early twenty-first century (1999–2011) with those of the late twentieth century (1985–1998) for 36 large-sample states, accounting for demographic covariates and birth state. We found that disparities vary considerably by state: in the early twenty-first century, the greater risk of death associated with lacking a high school credential, compared with having completed at least one year of college, ranged from 40 % in Arizona to 104 % in Maryland. The size of the disparities varies across states primarily because mortality associated with low education varies. Between the two periods, higher-educated adult mortality declined to similar levels across most states, but lower-educated adult mortality decreased, increased, or changed little, depending on the state. Consequently, educational disparities in mortality grew over time in many, but not all, states, with growth most common in the South and Midwest. The findings provide new insights into the troubling trends and disparities in U.S. adult mortality.

Introduction

Educational attainment is one of the strongest social determinants of health and mortality among U.S. adults (Galea et al. 2011; Woolf et al. 2007). Higher-educated adults report better self-rated health and have lower age-specific rates of chronic conditions, such as diabetes and disability, than those with less education (see review in Zajacova and Lawrence 2018). Higher-educated adults tend to live longer and spend a greater portion of their years in good health (Montez and Hayward 2014). For instance, in 2010, non-Hispanic white women aged 25 could expect to live another 51 years if they did not have a high school credential compared with 60 years if they had a bachelor’s degree or higher (Sasson 2016).

Educational disparities in mortality have been extensively documented at the national level. The disparities have grown considerably since the mid-1980s (Bound et al. 2015; Case and Deaton 2015; Cutler et al. 2011; Dowd and Hamoudi 2014; Hendi 2015; Jemal et al. 2008; Lauderdale 2001; Masters et al. 2012; Meara et al. 2008; Miech et al. 2011; Montez and Zajacova 2013; Montez et al. 2011; Olshansky et al. 2012; Preston and Elo 1995; Sasson 2016). For instance, Sasson (2016) reported that the gap in life expectancy among 25-year-old nonHispanic white women who had not completed high school and those with a college education increased from 2.5 years in 1990 to 9.3 years in 2010. Most studies have found that the growing disparities reflect the combined effect of declining mortality among higher-educated adults alongside rising mortality among lower-educated adults, particularly among lower-educated white women (e.g., Meara et al. 2008; Montez et al. 2011; Sasson 2016).

Explanations for the disparities and their growth over time remain contested. Most commonly, researchers have focused on individuals’ lifestyle behaviors, such as smoking (e.g., Cutler et al. 2011). Others have focused on compositional or selection effects, where growing disparities may simply reflect an increasingly small and inherently disadvantaged group of low-educated adults (e.g., Dowd and Hamoudi 2014). Still others have emphasized structural factors, such as changes in the social and economic landscape over the last half-century (e.g., Hayward et al. 2015). All these explanations are likely at play (Montez and Zajacova 2014). That is, structural changes, such as deindustrialization, a shrinking safety net, and greater educational homogamy in marriage have made education more critical for garnering beneficial resources, regardless of compositional changes among lower-educated adults.

This study expands knowledge about educational disparities and their growth by examining them at the U.S. state level. As the first study with this focus, it provides new information about how the disparities vary across the country. Evidence from states that have experienced growing, shrinking, or stable disparities can shed new clues about how certain state environments shaped the disparities, providing more granular context for understanding national trends in the relationship between education and adult mortality.

Contextualizing Educational Disparities in U.S. Adult Mortality

From Theorizing to Describing Contextual Variation

We posit that educational disparities in U.S. adult mortality, and their growth since the mid1980s, remain poorly understood for two primary reasons. The first reason is the individualist paradigm that characterizes most U.S. studies on this topic. Under this paradigm, studies have almost exclusively examined individual-level explanations (e.g., smoking behavior), even when they mention contextual-level explanations (e.g., deindustrialization). The focus on individual-level explanations may reflect the dominant view in U.S. studies—whether implicit or explicit— that education is a personal resource. Explanations for educational disparities in health and mortality typically point to the ways in which schooling provides individuals the agency and human capital to acquire resources and avoid risks for mortality (Cutler and Lleras-Muney 2010; Mirowsky and Ross 2003). That is, schooling enhances individuals’ capacity to coalesce healthy lifestyles, obtain good jobs, seek out medical knowledge, avoid financial hardship, develop social ties, and so on. Although these individualist explanations are important, they overlook the fact that context also matters. Individuals are embedded in epidemiological, socioeconomic, and policy contexts that may condition the effects of education on mortality (Baker et al. 2017; Clouston et al. 2016; Hayward et al. 2015).

The U.S. focus on individual-level explanations for growing educational disparities in mortality is also inconsistent with dominant conceptual frameworks for elucidating health disparities in general—for example, the theory of fundamental social causes (Link and Phelan 1995), ecosocial theory (Krieger 2001), institutional theories of health inequality (Beckfield and Krieger 2009; Beckfield et al. 2015), and the theory of constrained choices (Bird and Rieker 2008)—which emphasize that context also matters. These perspectives assert that contexts shape not only how resources, such as education, are distributed in the population but also the importance of those resources for health via policies (e.g., policies on taxation, labor unions, and unemployment) and economic opportunities. Both context and agency play a role and share space on the causal pathway. For instance, contexts (e.g., labor markets) can weaken or strengthen the individual pathways (e.g., employment) between educational attainment and health.

Drawing on these contextual perspectives, research outside the United States has often compared educational disparities in health and mortality across countries, finding that the magnitude of the disparities differs markedly across them (Bambra et al. 2010; Brennenstuhl et al. 2012; Cambois et al. 2016; Frohlich et al. 2006; Mackenbach et al. 2008; Navarro and Shi 2001). Evidence for certain welfare regimes being associated with particularly small or large disparities in health is mixed. However, a systematic review by Brennenstuhl and colleagues (2012) found that about one-half of studies reported that social democratic countries have smaller health inequalities and better population health than other regimes, suggesting that economic and policy contexts shape the importance of individuals’ socioeconomic resources for health and longevity.

From National to Subnational Analyses

The second reason for inadequate understanding of educational disparities in mortality and their growth since the mid-1980s is that studies have generally examined the United States as one homogenous unit. This assumption is increasingly untenable. A small but growing number of studies have found that socioeconomic disparities in health and longevity differ greatly across the country at several geographic levels. A study of the four U.S. census regions found that educational disparities in mortality are larger in the South than the Northeast (Montez and Berkman 2014). Another study concluded that educational disparities in disability varied considerably across states, even after the researchers adjusted for the race and socioeconomic characteristics of individuals and their local area of residence (Montez et al. 2017b). A study of men’s mortality found that urban areas had larger educational disparities in mortality than did rural areas (Hayward et al. 1997). Examining an even more granular geographic level, another study reported that mortality disparities by income differed markedly across commuting zones (Chetty et al. 2016). In sum, having low socioeconomic status is more problematic in some parts of the country than others, and understanding why carries important research and policy implications.

Several contextual factors can shape educational disparities in mortality. For one, overall levels of mortality differ by geography. Differences in mortality have widened since around 1980 across U.S. regions (Fenelon 2013), states (Wilmoth et al. 2011), and counties (Ezzati et al. 2008). Also important is the geographic distribution of education. Consequently, national-level analyses may confound the mortality risk of having low education with the mortality risk of residing in certain parts of the country. Additionally, the widening cleavages in the economic and policy contexts of U.S. regions (Newman and O’Brien 2011) and states (Montez 2017) are critical, as we discuss later. This study takes an important step forward in disentangling these complex processes by examining educational disparities in mortality and their growth across U.S. states.

U.S. States as Context

U.S. states vary in policies, resources, and opportunity structures in ways that affect population health (Grusky et al. 2015). For instance, decisions made by governors and state legislatures affect employment, housing, transportation, social integration, lifestyles, and numerous other social determinants of health. States may have become increasingly important in shaping population health since the 1980s, when structural changes such as deregulation, devolution, and state preemption movements gained momentum (Montez 2017). Deregulation has affected the industrial and economic base in many parts of the country. The devolution of federal programs to the states has elevated states’ discretion over the prioritization and funding of policies and programs. State preemption laws forbid or limit local governments from legislating on certain issues, many of which benefit health, such as indoor smoking bans (Pomeranz and Pertschuk 2017).

The combination of transferring federal authority to states (via devolution) and reducing heterogeneity in local policies (via preemption) has given states an increasingly important role in population health. In fact, states’ economic and policy environments have diverged quite dramatically since the 1980s. Consider the divergence between 1980 and 2016 in just three state policies that shape health: minimum wage (Van Dykea et al. 2018), tobacco taxes (Tauras 2004), and the Earned Income Tax Credit (EITC) (Muennig et al. 2016). In 1980, just two states had a minimum wage above the federal minimum; by 2016, the minimum wage in 29 states was above the federal level ($7.25), ranging from $7.50 to $10.00 per hour. In 1980, state excise taxes on tobacco ranged from $0.02 to $0.21 per pack of cigarettes; by 2016, state excise taxes ranged from $0.17 to $4.35 per pack. In 1980, no state offered its own EITC; by 2016, one-half of states offered one and had done so for 2 to 31 years.

The tendency for state policies to “cluster” (e.g., for states with low tobacco taxes also to have a lower minimum wage and to offer no EITC¹) may have magnified the impact on population health. Take the example of New York and Mississippi. Since the early 1980s, New York has imposed a substantial excise tax on cigarettes ($4.35 per pack in 2016), implemented its own EITC, participated in Medicaid expansion, and set a minimum wage above the federal level ($9.00 per hour in 2016). In contrast, Mississippi has a negligible cigarette tax ($0.68 per pack in 2016); does not offer its own EITC; opted out of Medicaid expansion, defaulted to the federal minimum wage; and has preempted local governments from implementing health-promoting legislation, such as paid sick days, a higher minimum wage, stricter firearm regulations, and nutrition labeling in restaurants (Montez 2017; Pomeranz and Pertschuk 2017).

These tectonic policy shifts might help explain why disparities in adult mortality across states widened after 1980 (Wilmoth et al. 2011). By the end of the twentieth century, roughly one-half of the variation in life expectancy across counties was attributable to the state within which they were located (Arcaya et al. 2012). In 2000, the range of life expectancy at age 50 across states exceeded the range across comparable high-income countries (Wilmoth et al. 2011).

State contexts appear to have an independent effect on population heath, even after accounting for local area of residence (Montez et al. 2017a) and individuals’ characteristics (e.g., Kawachi et al. 1997; Montez et al. 2016; Subramanian et al. 2001). For instance, a study examining state variation in women’s mortality during the 1990s found that 53 % of the variation was related to states’ characteristics (e.g., income inequality), whereas 30 % was related to women’s characteristics, such as race (Montez et al. 2016).

Heterogeneity in the Importance of State Contexts for Mortality

States may not affect the mortality of all residents equally, however. In terms of educational attainment, states may have their strongest effect on low-educated adults. These individuals possess fewer resources of their own and face greater risks in their daily lives, all of which make the resources, opportunities, and protections provided by their state particularly salient. In fact, and often by design (as with means-tested programs), many state policies disproportionately impact adults of lower socioeconomic status, including those with low education. For example, state EITC, minimum wage, Medicaid, welfare, tobacco taxes, unemployment, and incarceration policies are all more salient for low-educated adults. In contrast, higher-educated adults may be able to marshal their own resources to maximize health and longevity across state contexts. Higher education may act as a “personal firewall” (Montez et al. 2017b). Consistent with this notion, studies of the United States and European countries have found that the variation in age at death is smallest for higher-educated adults (Brown et al. 2012; Sasson 2016; van Raalte et al. 2011).

If higher-educated adults are better equipped to maximize their longevity regardless of state contexts, and if lower-educated adults are disproportionately affected by state contexts, then state-level variation in the educational gradient in mortality should mainly reflect variation in the mortality of lower-educated adults. Findings of two recent studies support this assertion. Montez and colleagues (2017b) found that educational disparities in disability varied across U.S. states primarily because disability levels among low-educated adults varied. Similarly, Chetty and colleagues (2016) found that the size of disparities in mortality by income varied across commuting zones mainly because the mortality of low-income adults varied.

These patterns represent an additional dimension of inequality from both a population and an individual perspective (Sasson 2016). Across states, lower-educated individuals may experience varying degrees of uncertainty about their longevity. Such doubts can exert independent influences on mortality by, for example, altering health behaviors (Scott-Sheldon et al. 2010) and economic decisions (Hurd et al. 2004). For these reasons, we examine educational disparities in mortality across states and how they relate to the mortality of low- and higheducated adults. This study is a necessary first step toward integrating individualist and contextual explanations for the disparities.

Aims

To our knowledge, this is the first study to document educational disparities in adult mortality by U.S. state and to examine within-state changes since the mid-1980s. We address three questions:

1. How do educational disparities in mortality vary across U.S. states in the twenty-first century?

2. Do the disparities vary mainly because the mortality of low-educated adults varies?

3. How have the disparities changed since the mid-1980s?

Answering these questions will identify states where growth in disparities has been sizable, minimal, or absent. This descriptive study carves out the contours of the problem so that subsequent studies can investigate which state-based policy levers can reduce the disparities.

Data and Methods

Data

We analyzed data from the 1985–2011 restricted-use National Health Interview Survey Linked Mortality File (NHIS-LMF). The NHIS is an annual, cross-sectional, nationally representative survey of the noninstitutionalized U.S. population. The NHIS-LMF links adults aged 18 and older in the 1985–2009 NHIS with death records in the National Death Index through December 31, 2011. The linkage is based primarily on a probabilistic matching algorithm, which correctly classifies the vital status of 98.5 % of survey records (National Center for Health Statistics 2009). We analyzed the restricted-use version because it identifies state of residence and, among U.S.-born respondents, state of birth.

The NHIS-LMF offers several advantages over vital statistics in addressing our aims. Educational attainment is provided by survey respondents, which is considered more reliable than education reported on death certificates (Rostron et al. 2010). It provides the numerator (deaths) and denominator (population) for mortality rates; death certificate data provide only the numerator and must be combined with another source, such as the U.S. Census, to obtain the denominator. Another benefit is that the NHIS-LMF provides education-specific mortality data starting in 1985, whereas death certificates did not do so until 1989 or later (depending on the U.S. state). The NHIS-LMF has some limitations, however. Its size is smaller than vital statistics, the sampling frame excludes institutionalized adults, and mortality linkages are less reliable for some demographic groups (Ingram et al. 2008). Nevertheless, its strengths outweigh its weaknesses to address our aims (for more detailed discussions about the strengths and weaknesses of both data sources, see Hendi 2017; Sasson 2016, 2017).

We created a person-year data structure in which respondents contribute one observation for each year, beginning with their NHIS interview year, until their year of death (or 2011 for survivors). Decedents were censored after death. A death indicator for each person-year observation was set to 0 for adults who survived the calendar year and to 1 for adults who died.

Sample

We retained person-year records for years when respondents were between ages 45 and 89. Respondents could “age in” and “age out” of the sample. For instance, a respondent who was age 35 years in 1985 would “age in” to the sample in 1995 when they turned age 45. The lower limit was set at age 45 years because of the sparse number of deaths in some states among young adults within certain education levels. The upper limit was set at age 89 years because the NHIS-LMF mortality matches are not as reliable at older ages (Ingram et al. 2008). An important additional benefit of our age range is that it is similar to the range examined in previous national-level studies on educational disparities in mortality. We excluded respondents who were ineligible for mortality follow-up, lived in the District of Columbia, or were missing information on education (fewer than 2 % of respondents). Our final sample contained 721,448 persons; 13,225,421 person-years of exposure; and 225,592 deaths during the follow-up period.

Educational Attainment and Period

We categorized educational attainment at three levels: less than high school, high school credential (diploma or GED), and more than high school. This categorization is often used in mortality analyses of older birth cohorts (e.g., Brown et al. 2012). In supplementary analyses, we disaggregated the more than high school category into some college and bachelor’s or higher, given the rising importance of a college degree in accessing resources such as income and marriage (DiPrete and Buchmann 2006) and protecting against mortality (Hayward et al. 2015). We used this four-category measure to examine a subset of 24 states with sufficient deaths for all four education levels in each period.

We divided the 27 years of mortality follow-up into two periods: 1985–1998 and 1999–2011. These periods are roughly evenly split and represent the end of the twentieth century and the beginning of the twenty-first century, a division that has been used previously to describe changes in disparities at the national level (Hayward et al. 2015).

Covariates

All analyses adjusted for age, race, and nativity. Age was measured in years from 45 to 89. Race was dichotomized into white/nonwhite for consistency across the 27 survey years. The 0.1 % of adults missing race information were combined with the nonwhite group. Nativity = 1 if the respondent was born in the 50 U.S. states or 0 otherwise. Sex was either included in the model, with male as the reference, or used to stratify models.

Methods

We estimated Poisson regression models from the person-year data, stratified by U.S. state. Equation (1) provides the basic form of the model. It estimates the natural logarithm of the mortality rate π in period p for state s as a function of two education dummy variables (LTHS for less than high school; HS for high school; with MTHS, more than high school, as the reference group) and the covariates contained in the b₃ vector. Equation (2) was used to examine changes in educational disparities in mortality between the two periods. The time variable T was coded 0 for 1985–1998 and 1 for 1999–2011. All models were estimated with Stata 14.1.

$$\text{ln}\left({\text{π}}_{p,s}\right)=b_0+b_{1}LTHS+b_{2}HS+{\mathit{b}}_{3}covariates$$ (1)

$$\text{ln}\left({\text{π}}_s\right)=b_0+b_{1}LTHS+b_{2}HS+{\mathit{b}}_{3}covariates+b_{4}T+b_{5}LTHS\times T+b_{6}HS\times T$$ (2)

Preliminary Analyses

Before the main analysis, we conducted extensive preliminary work to determine whether and how to account for the length of time a respondent had lived in, or had been exposed to, their state of residence. Given its importance, we describe this work here in some detail.

We used two key pieces of geographic information available in the restricted-use NHISLMF: state of residence and state (or country) of birth. First, we assessed whether including state of birth was more predictive of adult mortality than state of residence alone. We estimated three models, restricting the sample to U.S.-born respondents. We estimated mortality from state of birth in Model 1, from state of residence in Model 2, and from both states in Model 3. All three models adjusted for age, sex, and race. We considered the model with the smallest Bayesian information criterion (BIC) (Raftery 1995) to provide the best fit. We used BIC instead of the Akaike information criterion (AIC) because of the large sample and because we consider type I and II errors to be similarly undesirable (Dziak et al. 2012). The BIC values are shown in Table A1 in the online appendix.

We found that for both women and men and during both periods, state of residence was a better predictor of adult mortality than state of birth. Further, including both state of residence and birth did not provide a better fit than state of residence alone, given the BIC’s penalty for overfitting the model. Consequently, we included only state of residence in our main analyses.

In the second phase of the preliminary analyses, we examined the sensitivity of the education coefficients—b₁ and b₂ in Eq. (1)—to five schemes for defining the sample based on respondents’ potential interstate migration histories; we refer to these as potential because we do not have complete migration histories. Schemes 1–3 gradually restricted the sample: Scheme 1 included all respondents, Scheme 2 included only U.S.-born respondents, and Scheme 3 included U.S.-born respondents who resided in their state of birth. These three schemes did not adjust for the NHIS-LMF sample weights. Schemes 4 and 5 included all respondents but, unlike Scheme 1, included sample weights. Scheme 4 adjusted for NHIS-provided sample weights, and Scheme 5 used modified weights that we created. The modified weights allowed respondents with greater exposure to a state to have a larger contribution to the mortality estimates. Respondents born in their state of residence retained their NHIS weights (effectively multiplying them by 1.0), whereas those born in a U.S. state different than the one they resided in or outside the United States received weights of 0.75 and 0.50, respectively. The 1.0/0.75/0.50 weighing strategy reflects our assumptions about ages of (im)migration into the state of residence among these cohorts. We explored different weights and found the results to be robust to different multipliers.

We then estimated Eq. (1) for each of the five schemes. The results (shown in the online appendix, Fig. A1) revealed little difference in the b₁ or b₂ coefficients across the five schemes. That Scheme 1 (no weights) and Scheme 4 (original sample weights) produced similar coefficients reflects the fact that the weights are largely a function of model covariates. That Scheme 5 (modified weights) also produced similar coefficients reaffirms our finding that state of residence is a better predictor of adult mortality than state of birth. Accounting for exposure to a state among in-migrants had little effect. Given the robustness of these results, and our desire to maximize the sample size, we retained all respondents in our analyses and used the modified sample weights.

Results

Summary characteristics of the sample across all 50 states are provided in Table 1. The average age was 61 years in both periods. In Period 1, 28 % of respondents lacked a high school credential, 38 % had a high school credential, and 35 % had attained at least one year of college. In Period 2, these percentages were 19 %, 36 %, and 45 %, respectively.

Table 1.

Characteristics of analytic sample of U.S. adults aged 45–89 years by period

	Total Period (1985–2011)	Period 1 (1985–1998)	Period 2 (1999–2011)
Average Age (years)	61.1	61.5	61.0
Male (%)	45.5	45.1	45.7
Education Level (%)
Less than high school	21.3	27.7	19.0
High school	36.7	37.6	36.0
More than high school	42.3	34.7	45.0
Race (%)
White	83.2	84.2	82.9
Nonwhite	16.8	15.8	17.1
U.S.-born (%)	88.0	90.6	87.0
Number of Respondents	721,448	427,956	293,492
Number of Person-Years	13,225,421	3,361,420	9,864,001
Number of Deaths	225,592	63,431	162,161

Source: 1985–2011 restricted National Health Interview Survey Linked Mortality File.

For most analyses, we focus on the 36 states with the largest numbers of deaths in the NHIS-LMF. We defined large as at least 100 deaths at every education level within each period. Most states far exceeded this threshold. The online appendix provides model results for all states except North Dakota because of its small sample.

How Do Educational Disparities in Mortality Vary Across U.S. States in the Twenty-First Century?

Using Eq. (1) for Period 2 (1999–2011), we found that in all 36 large-sample states, the risk of death for adults without a high school credential (LTHS) was significantly higher than for adults with more than high school (MTHS), adjusting for age, sex, race, and nativity. For example, in New York, the risk of death was 70 % (relative risk (RR) = 1.70) greater for LTHS adults than for MTHS adults. The magnitude of the disparity varied considerably across states, as shown in Fig. 1, where darker shades of red indicate wider disparities. They were smallest in Arizona (RR = 1.40), Utah (RR = 1.46), and Colorado (RR = 1.50), and largest in Maine (RR = 1.97), South Carolina (RR = 1.97), and Maryland (RR = 2.04). Although the geographic pattern is not particularly strong, states in Appalachia had comparatively large disparities. Of the 12 states with an RR of 1.80 or greater, six were in Appalachia (Virginia, Kentucky, Georgia, Ohio, South Carolina, and Maryland); the other six were distributed across the Northeast (Maine and Massachusetts), Midwest (Michigan and Missouri), South (Arkansas), and West (Oregon).

Fig. 1.

Educational disparities in adult mortality during the twenty-first century (1999–2011) and how the disparities have changed since the late twentieth century (1985–1998). The educational disparity in mortality is defined here as the relative risk (RR) of death for adults without a high school credential compared with adults who have more than a high school education. States shaded gray were excluded because of sample size concerns. Asterisks indicate that the disparity in Period 2 (1999–2011) is significantly greater than it was in Period 1 (1985–1998):*p < .05; **p < .01; ***p < .001. The A, B, C, and D grades indicate how absolute mortality risk changed between the periods. States with an A had declining mortality for all education levels; B states had declining mortality for adults with at least a high school credential and little to no change for adults without a high school credential; C states had declining mortality for adults with at least a high school credential but increasing mortality among adults without a high school credential; D states had increasing mortality for most, if not all, education levels.

Do the Disparities Vary Mainly Because the Mortality of Low-Educated Adults Varies?

We estimated Eq. (2) and then set the values of the demographic covariates to U.S.-born = 1, white = 1, sex = 0.45, and age = 65 to obtain the annual mortality risks displayed in Fig. 2. This figure plots the risks by education level within each of the 36 states. States are sorted by mortality of the low-educated group. For instance, in South Carolina (left), the annual risk of death was 0.011 for U.S.-born white adults aged 65 with more than a high school credential but was 55 % higher (0.017) for their peers with a high school credential and 100 % higher (0.022) for those without a credential.

Fig. 2.

Annual mortality risk by education level and U.S. state of residence in the twenty-first century. Estimated rates shown are for U.S.-born white adults 65 years of age. The figure includes 36 large-sample states (see the Methods section).

Figure 2 supports our expectation that the size of the gradient is largely determined by the mortality of low-educated adults. Their risk of death ranges from 0.014 in Minnesota to 0.022 in South Carolina, a difference of 0.008. Among high-educated adults, it ranges from 0.009 in Minnesota and Montana to 0.013 in Alabama, a difference of 0.004. The difference is twice as large for low-educated adults as for high-educated adults. Importantly, the greater dispersion in low-educated adult mortality across states is not simply a function of their higher absolute mortality risk. Taking their higher mortality into account, the coefficient of variation (CV = 100[s / x], where s is the standard deviation of the 36 mortality rates, and x is their average), is also greater for lower-educated (12.3 %) than higher-educated (10.5 %) adults.

Another way to quantify this pattern is to estimate the correlation between the size of the gradient and the absolute mortality risk of lower-educated adults across the 36 states. Table 2 includes these estimates. The correlation between the gradient and the mortality of lower-educated adults was .78 in Period 1 and.89 in Period 2. In contrast, the correlation between the gradient and the mortality risk of higher-educated adults was weaker in both periods (–.06 in Period 1 and .40 in Period 2). In sum, the gradient is more closely associated with the mortality experience of lower-educated adults than with their higher-educated peers.

Table 2.

Correlation between the absolute difference in mortality across education levels and the absolute mortality risk for each education level (N = 36 U.S. states)^a

	Period 1 (1985–1998)			Period 2 (1999–2011)
	LTHS	HS	MTHS	LTHS	HS	MTHS
Women and Men	.78	.40	−.06	.89b	.69	.40
Women	.71	.12	−.30	.91	.46	.11
Men	.75	.40	−.14	.87	.70	.31

^aThe absolute difference is defined as the mortality of adults without a high school credential minus the mortality of adults with more than a high school credential.

^bInterpretation: the correlation between the absolute difference in mortality in each state, and the mortality of adults without a high school credential in each state, was .89 during Period 2.

It is also interesting that the correlation between the gradient and the mortality of lower-educated adults increased over time for both sexes. The correlation increased from .71 to .91 among women (p = .002, using a Z test) and from .75 to .87 among men (p = .081). In other words, the variation in educational disparities in mortality across states is increasingly determined by the mortality risk of lower-educated adults.

Another intriguing pattern in Table 2 is that in Period 1, states could have a large disparity because the mortality of higher-educated adults was relatively low compared with other states and/or because mortality of lower-educated adults was relatively high compared with other states. This pattern is evident from (1) the negative correlation between high-educated adult mortality and the disparity, and (2) the positive correlation between low-educated adult mortality and the disparity. However, by Period 2, both correlations were positive and were especially large for low-educated adults. In other words, by Period 2, states with large disparities tended to have relatively high mortality across all education levels, especially for low-educated adults.

How Have the Disparities Changed Within Each State Since the Mid-1980s?

Table 3 presents model coefficients (RRs from Eq. (2)) for education and education-by-period interactions. (Covariate coefficients are available on request; results for all states are shown online in Table A2.) States were separated into two groups. Group 1 contains states where a mortality disadvantage existed in Period 1 only for adults without a high school credential; Group 2 contains states where the disadvantage existed for adults with or without a high school credential. States were then sorted by the growth in the disparity over time, defined as the RR of the interaction term between less than high school and period (i.e., column Period 2 × LTHS).

Table 3.

Relative risk of death associated with less than high school (LTHS) and high school (HS), compared with more than high school (MTHS), across time, within U.S. states

State	LTHS	HS	Period 2	Period 2 × LTHS	Period 2 × HS
Group 1
Maine	1.29	1.19	0.80	1.53*	1.14
Montana	1.22	1.22	0.81	1.42*	1.09
Missouri	1.34**	1.15	0.85	1.40**	1.06
Oklahoma	1.31**	1.09	0.91	1.29**	1.08
Arkansas	1.42**	1.13	0.73**	1.29*	1.16
Virginia	1.49***	1.16	0.78***	1.25**	1.28*
Oregon	1.44**	1.20	0.78*	1.24	1.14
Georgia	1.22***	1.16	0.82*	1.22*	1.17
Louisiana	1.42***	1.10	0.86	1.20	1.16
Michigan	1.61***	1.10	0.89	1.11	1.08
Florida	1.41***	1.11	0.89*	1.10	1.08
Alabama	1.49***	1.12	0.88	1.09	1.14
Iowa	1.47**	1.17	0.93	1.09	1.12
Connecticut	1.45***	0.97	0.84	1.08	1.29*
Arizona	1.46***	1.15	0.90	0.97	0.97
Utah	1.55**	1.03	0.89	0.94	1.07
Group 2
Minnesota	1.28**	1.21*	0.80*	1.25*	1.13
Kansas	1.39**	1.27*	1.02	1.25	0.97
Texas	1.5***	1.16**	0.86***	1.23***	1.13*
California	1.37***	1.22***	0.87***	1.21***	1.07
Illinois	1.41***	1.21**	0.80***	120**	1.10
Indiana	1.41***	1.23*	0.83*	1.18	0.98
Wisconsin	1.38***	1.18*	0.88	1.14	1.05
Maryland	1.82***	1.31*	0.87	1.12	1.13
Massachusetts	1.65***	1.33***	0.85*	1.12	1.03
New Jersey	1.59***	1.31***	0.86*	1.11	0.96
Pennsylvania	1.58***	1.17**	0.87*	1.08	1.07
South Carolina	1.83***	1.91***	1.05	1.07	0.77
New York	1.61***	1.25***	0.85***	1.05	1.02
Ohio	1.73***	1.30***	0.97	1.04	1.00
Washington	1.63***	1.38**	0.98	1.03	0.97
Kentucky	1.86***	1.37*	1.05	1.01	1.00
West Virginia	1.64*	1.55*	1.04	1.00	0.79
North Carolina	1.62***	1.19*	0.92	1.00	1.07
Colorado	1.50**	1.36*	0.92	0.99	0.96
Tennessee	1.83***	1.36***	1.09	0.96	0.95
United States	1.52***	1.21***	0.88***	1.15***	1.07***

Notes: Data are from adults aged 45–89 in the 1985–2011 restricted NHIS-LMF. The table includes the 36 large-sample states. Period 1 = 1985–1998; Period 2 = 1999–2011. Relative risks are from Poisson regression models, stratified by state, including education, period, education-by-period interactions, age, sex, race, and nativity, adjusted by the modified sample weights. Group 1 includes states where the mortality disadvantage (vs. MTHS) in Period 1 was significant for LTHS; Group 2 includes states where the disadvantage was significant for LTHS and HS.

^*p < .05;

^**p < .01;

^***p < .001

Table 3 shows that educational disparities in mortality grew significantly in 11 of the 36 states between Periods 1 and 2, and the increase was often large. For example, the Period 2 × LTHS column shows that the disparity grew by 53 % in Maine, 42 % in Montana, and 40 % in Missouri. Group 1 states generally exhibited more growth than did Group 2: 44 % of states in Group 1 experienced a significant increase in the disparity, compared with 20 % of states in Group 2. That is, growth occurred especially in states where previously little, if any, mortality disadvantage had been associated with having a high school credential. In several other states, the education gradient showed little to no growth. In fact, in seven states (Arizona, Utah, Kentucky, West Virginia, North Carolina, Colorado, and Tennessee) the RR was 1.01 or lower for the Period 2 × LTHS interaction.

Figure 3 displays the change in the gradient between periods by state. The top panel, which plots RRs for each state by period, clearly shows the marked increase in the disparities in the 11 states (concentrated on the left side of the figure) contrasted against the relatively stable disparities on the right. For instance, the disparity in Maine was relatively small in Period 1 but became one of the largest by Period 2.² To better understand these changes in the relative risks, we also plotted the absolute mortality risks (middle panel). This panel shows, for example, that relative disparities grew in Maine because the mortality of higher-educated adults declined while the mortality of lower-educated adults rose: that is, mortality in both tails of the education distribution was pulled outward. To make this point even clearer, the bottom panel plots absolute differences in mortality rates by education level between the two periods. Of the 11 states where the relative disparities grew significantly, the growth in 6 states (Maine, Montana, Missouri, Oklahoma, Texas, and California) reflected declining mortality of higher-educated adults but rising mortality of lower-educated adults. The growth in the other five states (Arkansas, Virginia, Georgia, Minnesota, and Illinois) reflected declining mortality of higher-educated adults alongside smaller or negligible declines among lower-educated adults.

Fig. 3.

Relative and absolute mortality risk by education level across the two periods. LTHS = less than high school, HS = high school, MTHS = more than high school. Mortality estimates shown are for U.S.-born white adults 65 years of age.

We overlay all this information onto Fig. 1 for a complete picture. Asterisks identify the 11 states where the disparity significantly grew between the periods. Letters A, B, C, and D summarize the main ways that mortality changed across education groups. For instance, C states had increasing mortality among lower-educated adults but declining mortality for higher-educated adults, and D states had increasing mortality for most, if not all, education levels. Interestingly, the rising mortality of low-educated adults (the C and D states), which has received much attention lately, occurred in just 13 of the 36 states we examined, which were concentrated in Appalachian (Kentucky, Tennessee, West Virginia, and South Carolina) and central (Kansas, Oklahoma, Texas, Louisiana, Missouri, and Iowa) states.

Supplementary Analyses

Robustness Check on Interstate Migration

To ensure that our findings were not simply a function of (im)migration patterns, we separately estimated Eq. (2) for (1) U.S.-born adults and (2) adults born in their state of residence. We focus here on two high-immigration states as examples. (All states are available in Table A4 and Fig. A2 in the online appendix.) In Texas, educational disparities in mortality grew by 23 % (p < .001) in the preceding main analyses, by 23 % (p < .001) among U.S.-born adults, and by 22 % (p < .01) among those born in Texas. The corresponding rates in California were 21 % (p < .001), 21 % (p < .001), and 29 % (p < .01). Results were also robust in states that did not experience growth in disparities. For example, in West Virginia, the RR for Period 2 × LTHS was 1.00 for all adults, 1.00 for U.S.-born adults, and 0.97 for those born in the state. That our findings are robust across these subsamples concurs with our preliminary analyses finding that incorporating state of birth does not materially improve our ability to predict mortality.

Distinguishing College Graduates

The top section of Table 4 identify states where in Period 1, only adults without a high school credential had a significant mortality disadvantage compared with adults with a bachelor’s degree or higher. That is, in those (mostly southern) states—including Missouri, Oklahoma, Georgia, Louisiana, Alabama, Florida, and Virginia—having a bachelor’s degree did not confer a significant mortality benefit compared with a high school credential or some college. In the bottom group of states in the table, all education levels were disadvantaged compared with having a bachelor’s degree in Period 1. Between the two periods, disparities grew significantly in 12 of the 24 states. For instance, they grew by 50 % in Missouri, 47 % in Virginia, and 39 % in Oklahoma (column Period 2 × LTHS). Again, the growth mainly occurred in states where previously little mortality advantage had been associated with having a college degree.

Table 4.

Relative risks of death associated with less than high school (LTHS), high school (HS), or some college (SC), compared with a bachelor’s degree or higher (CO), across time, within U.S. states

State	LTHS	HS	SC	Period 2	Period 2 × LTHS	Period 2 × HS	Period 2 × SC
Group 1
Missouri	1.48**	1.27	1.21	0.79	1.50**	1.14	1.12
Virginia	1.53***	1.20	1.05	0.67***	1.47***	1.50***	1.39*
Oklahoma	1.41**	1.18	1.14	0.84	1.39*	1.16	1.13
Georgia	1.60***	1.22	1.10	0.73**	1.37*	1.31*	1.24
Louisiana	1.53***	1.18	1.14	0.78*	1.32*	1.28	1.18
Alabama	1.52***	1.14	1.03	0.74*	1.29	1.35*	1.36
Florida	1.47***	1.16	1.08	0.82**	1.19*	1.17	1.15
Group 2
Minnesota	1.47**	1.39*	1.28	0.73*	1.38*	1.24	1.16
Wisconsin	1.49***	1.28*	1.15	0.77*	1.30*	1.19	1.24
Michigan	1.79***	1.23*	1.21	0.77**	1.29*	1.26*	1.28
Massachusetts	1.73***	1.39***	1.09	0.75**	1.27*	1.17	1.30*
Indiana	1.54***	1.34*	1.18	0.80	1.22	1.02	1.05
Maryland	1.90***	1.36*	1.10	0.82	1.19	1.20	1.13
Washington	1.88***	1.59***	1.29	0.85	1.18	1.11	1.24
North Carolina	1.85***	1.36**	1.25	0.86	1.07	1.14	1.10
Group 3
California	1.54***	1.37***	1.24***	0.81***	1.30***	1.15**	1.13*
Texas	1.55***	1.33***	1.30***	0.86*	1.24**	1.14	1.00
Illinois	1.68***	1.43***	1.35**	0.82*	1.17	1.07	0.97
Pennsylvania	1.75***	1.30***	1.23*	0.82*	1.14	1.14	1.13
Ohio	1.95***	1.46***	1.27*	0.90	1.11	1.08	1.11
New Jersey	1.96***	1.61***	1.53***	0.88	1.09	0.95	0.98
New York	1.77***	1.38***	1.23**	0.84**	1.07	1.04	1.03
Tennessee	2.13***	1.58***	1.34*	1.03	1.02	1.01	1.10
Arizona	1.82***	1.44**	1.50**	1.00	0.87	0.88	0.82
United States	1.71***	1.36***	1.25***	0.83***	1.22***	1.14***	1.11***

Notes: Data are froms adults aged 45–89 in the 1985–2011 restricted NHIS-LMF. The table includes the 24 states with samples large enough to estimate mortality for four education levels. Period 1 = 1985–1998; Period 2 = 1999–2011. Relative risks are estimated from Poisson regression models, stratified by state, including education, period, education-by-period interactions, age, gender, race, and nativity, adjusted by the modified sample weights. Group 1 includes states where the mortality disadvantage (vs. CO) in Period 1 was significant only for LTHS; Group 2 includes states where the disadvantage was significant for LTHS and HS; and Group 3 includes states where the disadvantage was significant for LTHS, HS, and SC.

^*p < .05;

^**p < .01;

^***p < .001

Patterns for Women and Men

The online appendix shows results stratified by sex (Tables A5 and A6; Figs. A3–A5). Although our main analyses found significant growth in the disparities for 11 of the 36 states, the sex-stratified analyses revealed that in some states (Montana, Illinois, Florida, and Ohio), the growth was significant for women only. We also found that when the disparities widened for men, they typically did so because higher-educated men experienced a disproportionate decline in mortality compared with other men; but when the disparities grew for women, they typically did so because the mortality of higher-educated women declined while the mortality of women without a high school credential rose.

In summary, in states where the disparities grew, the growth reflects one of two patterns. Taking the 12 states where disparities grew when using a four-category measure of education, absolute mortality declined for adults with a bachelor’s degree or higher in all states; however, seven of those states had little to no decline in mortality among adults without a high school credential (Virginia, Georgia, Florida, Minnesota, Wisconsin, Michigan, and Massachusetts), and five states had rising mortality among those adults (Missouri, Oklahoma, Louisiana, California, and Texas). Finally, our finding that the disparities did not change in some states does not mean that absolute mortality was stable. For instance, although the disparities were stable in Tennessee, Pennsylvania, New Jersey, and New York, absolute mortality increased for all education groups in Tennessee but declined for all education groups in the other three states.

Discussion

This study takes an important and innovative step toward explaining educational disparities in U.S. adult mortality and their growth since the mid-1980s by examining them at the state level. It provides new evidence and suggests new hypotheses about why disparities exist, why they have grown, and how they might be reduced. Because states appear to be an increasingly consequential context, disparities research should integrate contextual explanations and more conventional individualist approaches.

Our findings support five major conclusions. First, educational disparities in mortality in the twenty-first century are large and statistically significant in all 36 states we examined, but their size varies considerably. For instance, compared with adults with more than a high school credential, adults without a high school credential had a 40 % higher risk of death in Arizona but a 104 % higher risk in Maryland. National estimates that ignore such variation obscure considerable heterogeneity across states and insights that can be discovered in state-level analyses.

Second, the size of the education gradient varies across states mainly because the mortality of low-educated adults varies. This is consistent with a recent study of disparities in disability where state-level variation in the disparities was mainly due to the disability among low-educated adults (Montez et al. 2017b). Having a low level of education appears to be a particularly large mortality threat in certain states, where attaining higher levels of education is especially important. Higher education seems to provide a personal firewall against states’ contexts (Montez et al. 2017b). Higher-educated adults may be better able to marshal their own social and economic resources to garner health advantages across these vastly different socioeconomic and policy contexts, while resources available to lower-educated adults may be more closely tethered to their states’ contexts.

The strong correlation across states between the size of the education gradient and the mortality rates of low-educated adults has strengthened since the 1980s because the mortality of higher-educated adults—particularly those with at least at a bachelor’s degree—has declined in nearly all states to fairly similar levels, while mortality of lower-educated adults has varied more widely by decreasing, increasing, or remaining stable, depending on the state. Consequently, by the twenty-first century, states with the largest disparities tended to have relatively high mortality across all education levels, especially for low-educated adults. This scenario, where greater inequality in state-level mortality is associated with higher overall mortality, comports with Wilkinson’s (1996) thesis linking inequality to overall level of health and longevity.

Third, educational disparities in mortality in the twenty-first century have grown significantly beyond those in the late twentieth century in many, but not all, states. When we used a three-category measure of education (less than high school, high school, or more than high school), 11 of the 36 states we examined experienced a significant growth in the disparities. The growth was significant in six additional states when we used a four-category measure of education (including having a bachelor’s degree). These 17 states were concentrated in the South and Midwest; educational disparities widened less in the Northeast and West. A study that examined U.S. regions reached similar conclusions (Montez and Berkman 2014). Of the states we examined, 57 % of Southern states and 56 % of Midwestern states had significant growth in the disparities, compared with 33 % in the Northeast and 29 % in the West.

Fourth, states exhibited two patterns of growing disparities. In 7 of the 17 states with significant growth, the changes resulted from the combination of declining mortality among high-educated adults alongside increasing mortality among low-educated adults (Maine, Montana, Missouri, Oklahoma, Texas, California, and Louisiana). In the other 10 states (and in states where the growth was not statistically significant), mortality declined mostly for higher-educated adults; it declined less or remained relatively stable for lower-educated adults (Arkansas, Virginia, Georgia, Minnesota, Illinois, Alabama, Florida, Wisconsin, Michigan, and Massachusetts).

Fifth, our findings contribute new information to the debate about the increasing mortality of lower-educated adults. This increase has received considerable attention since it was first reported a decade ago (Meara et al. 2008). Some have argued that the increase is an artifact of negative selection in the low-educated group: with rising educational attainment in the U.S. population, the proportion of adults with less than a high school has diminished, and those left behind may have higher risks for poor health and premature death. However, we found that the mortality rates in this group did not increase in the majority of states we examined. Thus, the worsening mortality may not be an inevitable manifestation of selection.³ Accordingly, it is reasonable to conjecture that the increasing mortality reflects the confluence of having a low education level and residing in an unfavorable state context, regardless of compositional changes.

Our finding raises other questions: What has happened in states like Massachusetts, New York, and Oregon, where mortality of low-educated adults has decreased? And how does this differ from states like Missouri, Oklahoma, and Texas, where their mortality has increased? Although answering these questions is well beyond the scope of our study, we offer some possibilities for future research. For one, systematic differences in policies exist between these two groups of states. Consider the three example policies discussed in the Introduction. As of 2016, Massachusetts, New York, and Oregon had, on average, a minimum wage of $9.42 per hour, a tobacco tax of $3.06 per pack of cigarettes, and a 21-year-old EITC program. In contrast, Missouri, Oklahoma, and Texas averaged a lower minimum wage ($7.38 per hour) and tobacco tax ($0.87 per pack), and had offered an EITC for only five years. These policies are salient mostly for low-educated adults and are associated with outcomes such as birth weight, infant mortality, teenage pregnancy, employment, smoking, and adult disability and mortality (Bullinger 2017; Komro et al. 2016; Montez et al. 2017a; Muennig et al. 2016; Strully et al. 2010; Van Dykea et al. 2018). A growing divergence in these policies (and others) since the 1980s may help explain why lower-educated adults have fared better in some states than others. These cases illustrate new research questions opened by our study.

In sum, the reasons for the growth in the education gradient are likely to be complex and evolve over time. Nevertheless, national analyses may be inadequate for elucidating the reasons because those analyses may obscure countervailing forces and trends that exist within states. This study lays the necessary groundwork for future analyses to systematically test explanations for the patterns we report here. Explanations will need to jointly consider contextual-level and individual-level factors. As stated previously, our aim here was to carve out the contours of this troubling problem and provide insights that will allow for more refined hypothesis testing.

Limitations and Future Directions

Our study has some shortcomings. The survey data are observational and preclude formal testing of causal associations between education and mortality; we rely on recent reviews of the evidence that support a causal interpretation (e.g., Hummer and Lariscy 2011; Zajacova and Lawrence 2018). We also do not have complete migration histories on the respondents. Although no U.S. mortality data set provides birth-to-grave migration histories, our data source provides proxies, including state of residence and state of birth. Our preliminary analyses suggest, however, that having incomplete migration histories is unlikely to change our conclusions. Results were robust across different weighting schemes and subgroups such as U.S.-born respondents and nonmigrants. Other studies have similarly reported that interstate migration does not explain geographic patterns or trends in mortality at the county and state levels (Ezzati et al. 2008; Fenelon 2013).

Although we focused on states for reasons listed in the Introduction, local areas (e.g., counties, neighborhoods) may additionally shape mortality and the importance of education for mortality. Nonetheless, prior studies have found that even after local variation was accounted for, disability and mortality differ strongly by state (Arcaya et al. 2012; Montez et al. 2017b).

We did not formally test how much of the cross-state pattern in the disparities was due to specific characteristics of states versus specific characteristics of individuals. Such a test is well outside the aims of our study, and the necessary data are not included in our restricted-use NHISLMF data agreement. Nevertheless, our findings suggest that states matter, net of key individual-level determinants of mortality. That is, even after accounting for age, sex, race, nativity, and education (which are tightly correlated with other determinants, such as income and lifestyles), we found that the importance of states for the disparities grew over time in many parts of the country.

Our study is a starting point; many more questions remain. For instance, how do the disparities vary for demographic subgroups? Is it possible to pinpoint the calendar years when the disparities began to widen in each state? How do the disparities vary by cause of death? Which state-level characteristics explain the patterns and trends? On this last question, we suggest several directions for such research. As mentioned earlier, these include decades of structural changes, such as deregulation, decentralization of political authority, and state preemption laws, in addition to the formation of organizations such as the American Legislative Exchange Council (MacLean 2017; Montez 2017). Within states, factors such as the attitudes of constituents, political leanings of legislators, and industry pressures might also have played a role (for an example on tobacco control, see Golden et al. 2013). States also have deeply embedded historical, cultural, and social characteristics, many of which reflect a common ethos; these factors can shape the diffusion of information, technologies, and policies within states (Skinner and Staiger 2007).

Conclusion

Educational disparities in adult mortality at the national level obscure large and important differences in the disparities across U.S. states. States differ substantially in the magnitude of the disparities as well as in the degree of changes in the gradient since the mid-1980s. Efforts to elucidate the disparities and their growth over time should leverage this heterogeneity and identify key state-level contextual characteristics—including social and economic policies—that shape these trends. The complexity of these questions requires a major refocusing of research efforts to understand the context in which educational disparities in mortality emerge.