This article explores how patterns of health, morbidity, and disability have changed across successive generations of older adults in the United States. Using a novel method for comparing state-specific partial life expectancies—that is, total life expectancy (LE), and health expectancies (HEs) in different health states, bounded between two ages—I explore changes in healthy life expectancy across successive birth cohorts of the U.S. population. Results show that little compression of disability is occurring across cohorts, LE with chronic morbidities has expanded considerably, and self-rated health is improving across cohorts, but only at ages 70+. These findings suggest that successive cohorts in the U.S. population may be on divergent paths in terms of late-life health and well-being. Exploring heterogeneity in these patterns, I find that less educated individuals have substantially lower partial LE and disability-free LE than those with more schooling, and that disability-free life is declining among those with less than a high school diploma. Differences in HEs are pervasive across racial and ethnic groups, and both disabled LE and unhealthy LE are expanding in some disadvantaged subgroups. The continued increases in partial LE with morbidities across successive cohorts, and the broad stagnation of disability-free and healthy LE, present a broad view of a U.S. population in which successive generations are not living healthier lives.
Nearly a half-century of inquiry from researchers in demography, epidemiology, and population health has focused on the consequences of increases in lifespan for population-level disability and morbidity. This area of research has been spurred on by two competing frameworks for understanding the prospects of continued progress in population health. In the compression of morbidity framework, expansion in life expectancy (LE) was thought to result in a compression of disability to a smaller portion of late life, with medical advances and improved health behaviors increasing the age of onset of morbidity or disability more quickly than rises in LE (Fries 1980, 2005). A different theoretical viewpoint, predicting an expansion of disability or morbidity, has also been put forward (Gruenberg 1977). Succinctly put, this theory posits that reductions in mortality (and thus increases in LE) may result in a shift in frailty over successive cohorts, as individuals with worse health are now likely to survive longer than they would have in the past.
It is worth noting that these two viewpoints are not necessarily opposing—within a population, there is likely to be substantial heterogeneity, with different subpopulations experiencing expansions or compressions of disability and/or morbidity at the same time (Crimmins 1996; Robine and Michel 2004; Zeng et al. 2017). In addition, different facets of health (i.e., disability, self-rated health, chronic morbidities) may be changing in different directions across cohorts (Crimmins 2004). The dynamic equilibrium model, developed by Manton (1982), recognizes this heterogeneity, hypothesizing that improvements in medical technology and early diagnoses may lead to diseases being discovered, and controlled, at early stages, resulting in an increase in the proportion of the population with chronic conditions but steady or declining rates of disability and mortality.
For the United States, the considerable body of evidence on the compression of morbidity and disability is mixed (Beltrán-Sánchez et al. 2016; Crimmins and Beltrán-Sánchez 2011; Cutler et al. 2013; Martin et al. 2010). Broadly, studies suggest that older individuals are living longer, but find mixed evidence on whether or not this additional lifespan is subject to morbidities and disabling conditions. In part, these findings are likely due to different definitions of morbidity and disability. Studies investigating morbidity, as defined by the presence of a set of biomarker or physician-assessed health conditions, have generally found an increasing trend across time in the United States (Beltrán-Sánchez et al. 2016; Crimmins 2004, 2015; Crimmins and Beltrán-Sánchez 2011; Crimmins et al. 2019), though some research finds the opposite (Cutler et al. 2013). Studies focused on disability, generally defined by limitations on activities of daily living (ADLs), have more often found that time spent with these sorts of limitations is stable, or declining, as a share of total life expectancy (Bardenheier et al. 2016; Crimmins 2015; Freedman et al. 2016).
Although examining population-level changes in LE and health expectancies (HEs) is quite important for monitoring health at the national level, health and well-being in later life are known to be substantially patterned by sociodemographic characteristics. Inequalities in mortality by educational attainment are substantial in the United States (Hendi 2015; Sasson 2016), and extensive prior research has found substantial gaps in longevity and healthy life expectancy between individuals with high and low levels of education (Chiu et al. 2016; Chiu et al. 2019; Crimmins and Saito 2001; Molla et al. 2004; Solé-Auró et al. 2015). These gradients are hypothesized to emerge from a variety of sources, including differentials in income and wealth, differential access to (and quality of) preventive health care over the life course, and educational gradients in health behaviors such as smoking, diet, and physical activity. The existing literature has also found a consistent Black–White gap in disability and healthy life expectancy, with White Americans living longer, healthier lives than Black Americans (Berkman et al. 1989; Crimmins and Saito 2001; Crimmins et al. 1989; Freedman and Spillman 2016; Hayward and Heron 1999; Soneji 2006). Researchers have attributed these differences to a combination of gradients in access to health care, income, and health behaviors between White and minority populations, although it is likely that structural factors, including entrenched racism, differential patterns of treatment by the healthcare system, and cumulative stress from discrimination at both the individual and institutional levels, are just as culpable (Boen 2019; Williams et al. 2019). A number of studies have found that gradients in HEs have widened over time by educational group and have found mixed evidence of trends by racial and ethnic groups (Cantu et al. 2019; Chiu et al. 2019).
A potential limitation of the existing research base on trends in HEs in the United States is that much of this work relies on period comparisons—that is, measuring mortality and morbidity/disability conditions in a population at different points in time, and observing changes in the trends over time (Cantu et al. 2019; Chiu et al. 2019; Crimmins and Saito 2001; Freedman and Spillman 2016; Freedman et al. 2016). Although this synthetic-cohort approach may be useful for monitoring aggregate trends in population-level disability, these results do not easily translate to the experience of any given birth cohort of individuals. Tracking disability using only period estimates may obscure substantial heterogeneity between birth cohorts and provides limited information on how morbidity, disability, and life expectancy have shifted over time and across cohorts. Measuring cohort changes provides results that match more closely with the lived experience of individuals in a population (Manton et al. 2008). Indeed, previous work has found that period estimates of disability-free life expectancy (DFLE) have relatively poor correspondence with cohort DFLE (Manton and Land 2000). A small but important body of research investigated trends in health expectancies across cohorts over the 1980s, 1990s, and early 2000s (Manton et al. 2008; Manton and Land 2000; Manton et al. 1997; Soneji 2006). Briefly summarized, these studies generally found that successive cohorts were surviving longer and remaining healthier. However, with the exception of a small number of studies focused on specific diseases (Bardenheier et al. 2016; Payne and Kobayashi 2022) or on specific birth cohorts (Beltrán-Sánchez et al. 2016; Crimmins et al. 2019), little recent research has investigated cohort patterns in the compression of morbidity or disability in the United States.
In part, this relative lack of cohort-focused research on morbidity, disability, and healthy life expectancy is a result of data limitations. Estimating full-cohort HEs—that is, the years an individual from a given birth cohort can expect to live in specific states such as disability-free, morbidity-free, or with good self-reported health—is only possible for extinct cohorts. Few countries have the sort of long-term, intensive health surveillance systems that allow for the estimation of full-cohort–based evidence on the morbidity compression. However, though data requirements for estimating full-cohort HEs are quite high, data requirements for monitoring cohort HEs during discrete portions of the life course are much less stringent. This article demonstrates a method for estimating state-specific partial-cohort life expectancies—that is, total LE, and HEs in discrete states of health, morbidity, or disability bounded between two ages. This method allows the exploration of LE and HEs within a given time period across successive birth cohorts, and thus can ascertain whether a contraction or an expansion of HEs is occurring in different portions of the life course (Liu et al. 2019; Payne and Wong 2019).
Indeed, there are reasons why these partial LE and HE measures may be more useful than full-cohort HEs. For one, completed cohort HEs are predominantly a historical measure—that is, by the time a full-cohort LE and HEs can be estimated, the cohort is necessarily extinct. Hence full-cohort HEs are less useful for monitoring recent trends in a population. Focusing on cohort estimates also makes for cleaner contrasts when comparing groups that may differ in composition over time. The potential bias introduced by secular changes occurring in the population is minimized by a partial-cohort approach. For example, comparing HEs in a cohort framework limits the potential bias introduced from changes in the prevalence of health conditions and health behaviors such as smoking, obesity, and diabetes (Fishman et al. 2014; Holford et al. 2014; Lee et al. 2010; Masters et al. 2013; Preston and Wang 2006) and cohort differences in environmental exposures in early life (Reynolds 2019). Cohort-based HE estimates also have the advantage of applying to an actual group of individuals in the population, rather than to a purely synthetic population as in period HE estimates. In addition, partial-cohort HE measures allow for the combined monitoring of how mortality and health conditions are changing across cohorts during key portions of the life course—such as late middle ages, where recent scholarship in the United States has shown rising incidence of physical limitation, poor health, and mortality (Case and Deaton 2015; Freedman et al. 2013).
This study takes a comprehensive view to exploring population health by monitoring changes in three distinct facets of well-being across birth cohorts in the U.S. population: disability-free life expectancy, morbidity-free life expectancy (MFLE), and healthy life expectancy (HLE). These measures distinguish between years that are free of disabilities, morbidities, or poor health and years lived with these conditions, providing metrics that combine mortality and health into a single measure. I compare cohort changes in morbidity, disability, health, and longevity to understand whether, in cohort terms, the U.S. population is experiencing a compression or an expansion of years lived with these conditions. In addition, I explore whether these trends are patterned by social inequalities in educational attainment and race or ethnicity.
Data and Methods
These analyses use data from 1998–2016 biannual waves of the U.S. Health and Retirement Survey (HRS), a long-running longitudinal sample survey of older adults (Sonnega et al. 2014), to examine whether successive birth cohorts in the U.S. population are living more of their life with physical limitations or disability. The HRS is a cohort study of adults aged 51+ in the United States. The cohort is refreshed with a new six-year birth cohort every three waves. Analyses used the RAND HRS Longitudinal File 2016 (V2) (Health and Retirement Study 2020).
Three different outcomes are used to monitor cohort changes in HEs over the 1998–2016 period: disability-free life expectancy, morbidity-free life expectancy, and healthy life expectancy. The online Supplemental Figure S1 illustrates the three outcomes. For each outcome, I explore whether there has been an expansion or compression of time spent with disability/morbidities/poor health, both in proportionate terms (i.e., whether DFLE, MFLE, or HLE have grown or shrunk as a proportion of total partial LE within an age-group across cohorts) and in absolute terms (i.e., whether the total years of LE spent with disability, morbidities, or poor health have increased or decreased in an age-group across cohorts).
Self-reported disability is based on questions from the activities of daily living scale (Katz et al. 1963). Disability is conceptualized within the framework of the disablement process (Verbrugge and Jette 1994), in which disability is defined as a gap between an individual's capacities (physical, sensory, or cognitive) and the demands of a given activity in that individual's particular environment. Thus, it is influenced both by physical declines due to age and health conditions and by an individuals' environment and available accommodations. Individuals are classified as ADL-disabled if they report difficulty or inability in doing any of the following five activities: bathing, eating, getting in/out of bed, dressing, and walking across a room. Where necessary, proxy responses on ADL disability are used (Sonnega et al. 2014). Individuals who reported no physical limitations or ADL disabilities are classified as disability-free. Information on mortality and date of death came from the HRS tracker file, which draws on both the National Death Index as well as exit interviews with a spouse or knowledgeable family member (Sonnega et al. 2014).
Morbidity is conceptualized as the presence of diagnosed diseases or conditions that may directly contribute to impaired functioning and premature mortality (World Health Organization 2014). Self-reported morbidity is defined as receiving a physician diagnosis of any of the following five chronic morbidities: cancer, diabetes mellitus, chronic heart disease, chronic lung disease, or prior stroke. In the United States, these conditions represent the top five causes of death due to chronic noncommunicable disease (Kochanek et al. 2020) and were the underlying cause of approximately 60% of all deaths and 85% of chronic disease deaths in 2019. Individuals are classified as morbidity-free if they report never being diagnosed with these conditions, and individuals are classified as having chronic morbidities if they have been diagnosed with one or more conditions.
Self-rated health is defined using a question asking the respondent to report on their current general health status, with responses of “excellent,” “very good,” “good,” “fair,” and “poor.” Individuals are classified as healthy if they select one of the first three options and as unhealthy if they select “fair” or “poor.” Self-reported health is widely used as a measure of general health and well-being and is known to be highly predictive of survival (DeSalvo et al. 2006).
Covariates include age (continuous), sex (men vs. women), and birth cohort (six-year birth cohorts spanning those born in 1912–1917 to 1954–1959). Level of attained education encompasses four categories: less than a high school diploma, high school graduate or GED holder, some college, or completed bachelor's degree or more. Self-reported racial and ethnic identity was grouped into three categories: non-Hispanic Black, Hispanic, and non-Hispanic White. Individuals reporting other identifications were not included in the analysis of racial/ethnic subgroups owing to small sample sizes.
This study compares partial-cohort life expectancy (PC-LE) and partial-cohort health expectancies (PC-HEs) in six separate six-year age ranges (54–59, 60–65, 66–71, 72–77, 78–83, and 84–89) across three six-year time periods (1998–2004, 2004–2010, and 2010–2016). Eight six-year birth cohorts from the HRS are included in these analyses:
1954–1959, the “middle baby boomers”;
1948–1953, the “early baby boomers”;
1942–1947, the “war babies”;
1936–1941, the late HRS original cohort;
1930–1935, the early HRS original cohort;
1924–1929, the “children of the Depression”;
1918–1923, the late “Asset and Health Dynamics Among the Oldest Old” cohort; and
1912–1917, the early “Asset and Health Dynamics Among the Oldest Old” cohort.
Details on the structure of the 18 period–cohort groupings compared in these analyses are provided in Table 1. The primary focus is to compare partial LE and partial HEs within a given age range across three successive cohorts in the U.S. population; for example, the first row of Table 1 shows that the analyses in ages 54–59 compare partial LE and partial HEs across cohorts born in 1942–1947 (the “early” cohort, observed in years 1998–2004), 1948–1953 (the “middle” cohort, observed in years 2004–2010), and 1954–1959 (the “later” cohort, observed in years 2010–2016).
As described in the following analyses, I focus on comparing PC-LE and PC-DFLE across three cohorts in each of the six age-groups under study. To simplify the presentation of results, these cohorts are referred to as the early, middle, and later cohorts for comparisons made across a given age range—that is, the early, middle, and later cohorts compared in ages 54–59 were born in 1942–1947, 1948–1953, and 1954–1959, respectively. For the comparison in ages 84–89, the cohorts compared are the 1912–1917 cohort (early), the 1918–1923 cohort (middle), and the 1924–1929 cohort (later).
The structure of these comparisons is perhaps best understood graphically. Figure 1 translates the first row of Table 1 into a Lexis diagram, visualizing the cohort comparison in ages 54–59. For each cohort comparison, I use all observed data for members of each cohort during each observation period—so the “early” cohort in Figure 1 would comprise all HRS sample members born between 1942 and 1947, and observed between the years 1998 and 2004. Note that each six-year period–cohort group summarizes transitions spanning 12 years of age; in the example in Figure 1, the youngest member of each cohort is exactly age 51 at the start of the period, and the oldest member of each cohort is just under 63 by the end of the observation period. To increase interpretability, and to account for potential variation in the age composition within the six-year cohorts, the analysis model predicts transition probabilities and estimates PC-LE and PC-HEs for the central age-trajectory in each cohort. In the analyses described in Figure 1, this means that PC-LE and PC-HEs are estimated for each cohort starting at age 54 and ending at age 60, as shown by the diagonal dashed lines. In all analyses, the cohorts are distinct—that is, there are no individuals contributing person-years of observation to more than one of the early, middle, or later cohorts.
The analyses focus on comparing period–cohort groupings rather than age–cohort groupings, a choice that is largely driven by the somewhat limited period of HRS follow-up. For a six-year birth cohort, the period–cohort approach requires data spanning only six years of calendar time, but members of the cohort are observed spanning 12 years of age. In contrast, an age–cohort approach would require observing a cohort over 12 years of calendar time, but only six years of age. Given that only 18 years of longitudinal HRS data (1998–2016) are available, this period–cohort approach was chosen to maximize the number of cohort comparisons that could be made. This use of six-year birth cohorts also aligns with the HRS sample design—since 1998 the HRS has added in a new six-year birth cohort every six years—the war babies (born 1942–1947, added in 1998), the early baby boomers (born 1948–1953, added in 2004), and the middle baby boomers (born 1954–1959, added in 2010).
To estimate PC-LE and PC-HEs for each cohort–period combination, I initially convert the HRS data to a person-year time scale, assuming that transitions between states occur at a random time between observations. Annual transition probabilities are estimated separately for each of the six age-groups under study (see Table 1); for example, the analyses of PC-LE and PC-HEs in ages 54–59 are based on data from the 1942–1947 (early), 1948–1953 (middle), and 1954–1959 (later) birth cohorts. Transition probabilities are modeled using a multinomial logistic regression model that is stratified by initial state. In the population-level analyses (results in Figures 2–4), the model includes age, age squared, sex, dummy variables for birth cohort, and interactions between birth cohort, age, and sex. Analyses exploring disparities between educational groups (results in Figures 5–7) include all covariates from the population-level analysis and add covariates for each educational group analyzed excepting a reference group (those with a high school education) and two-way interactions between education group and sex and between education group and birth cohort. Analyses exploring disparities between racial/ethnic groups (results in Figures 8–10) include all covariates from the population-level analysis and add covariates for each racial/ethnic group analyzed excepting a reference group (non-Hispanic Whites) and two-way interactions between racial/ethnic groups and sex and between racial/ethnic groups and birth cohort.
From these models, I generate matrices of age-specific transition probabilities for each combination of cohort and sex. Note that, because the models are run separately for each age-group, transition probabilities are not forced to align where ages overlap. Tables of these estimated transition probabilities are available in the online Supplemental Materials. Transition probability estimates were obtained using PROC SURVEYLOGISTIC in SAS version 9.4.1 The underlying modeling strategy, and HE estimation methods, are based on an adapted version of the Stochastic Population Analysis for Complex Events (SPACE) suite of SAS programs (Cai et al. 2010).
I use these observed transition probabilities as the input for the microsimulation-based multistate life table model (Cai et al. 2010; Kohler et al. 2017; Liu et al. 2019; Payne 2018; Payne and Wong 2019). I first generate synthetic cohorts of 100,000 individuals for each of the early, middle, and later cohorts, who have the same sex and initial state distribution as in the observed data (which are presented in online Supplemental Table S1). These individuals are then “aged” forward year by year using age-, gender-, and birth-cohort–specific mortality rates and probabilities of transitioning in and out of health/disability/morbidity estimated from the data. This process is repeated until the end of the age range under study—that is, when investigating PC-LE and PC-HEs for a given cohort in ages 54–59, the model will microsimulate the life courses of 100,000 individuals starting at age 54 and ending at exact age 60, applying the transition probabilities estimated from the data. The resulting synthetic cohort is analyzed to estimate health expectancies and total LE bounded between ages 54 and 60. Point estimates shown are from transition probabilities estimated from the full sample. Conﬁdence intervals, which reﬂect both the uncertainty of the estimated transition probabilities and the uncertainty from the microsimulation, were created by reestimating the aforementioned analysis sequence on 199 bootstrap resamples from each of the age range×birth cohort groups in the study. The central 95% of the distribution of these 200 parameters (the point estimate from the full data set, and the 199 estimates from the bootstrap resampling) is taken as the 95% confidence interval.
Inverse-probability (IP) weights are included to correct for potential bias introduced from differential loss to follow-up. This method weights complete cases (those who do not attrit) by the inverse of their probability of being a complete case, and include age and all sociodemographic and health variables included in online Supplemental Table S1. IP weights were generated separately for each time period and birth cohort included in these analyses. The final weight used in the analyses is the resulting IP weight multiplied by the HRS combined respondent weight and nursing home resident weight (averaged over each period–cohort combination) (DuGoff et al. 2014; Payne and Wong 2019). This final weight was used in the WEIGHT statement of the PROC SURVEYLOGISTIC model described earlier (Groves et al. 2009).
The online Supplemental Table S1 presents the baseline characteristics of each six-year cohort at the start of each six-year observation period. Briefly, there is a general trend of increasing educational attainment over successive cohorts within each age and declines in the proportion of respondents requiring the assistance of a proxy to complete the interview. No substantial patterns are apparent in the proportion of each cohort reporting individual ADL limitations within age-groups. Within each age range, more recently born cohorts are generally more likely to report a diagnosis of each chronic morbidity, though these gradients are quite substantial for diabetes and cancer and somewhat smaller for heart disease, stroke, and lung disease. Overall, within each age range, the most recently born cohort has the highest percentage of individuals reporting one or more chronic morbidity.
Disability-Free Life Expectancy
Figure 2 provides estimates of partial-cohort LE (PC-LE) and partial-cohort disability-free life expectancy (PC-DFLE) across birth cohorts in the U.S. population. PC-LE increases across cohorts—that is, within each age-group, more recently born cohorts appear to be living slightly longer than cohorts born further in the past. These differences are not universally significant (particularly in earlier age-groups; see online Supplemental Table S2), but are quite consistent across cohorts. In older age ranges, these increases in PC-LE are more substantial, with rises of 0.15 years across cohorts in ages 72–77, 0.25 years in ages 78–83, and 0.12 years in ages 84–89. Note that, because of the inherent random nature of the microsimulation process used for estimation, PC-LE estimates may not precisely match between models investigating cohort differences in self-rated health, disability, and morbidity—though differences will almost always be quite small.
PC-DFLE generally rose across cohorts along with PC-LE, though at a slightly slower rate (Figure 2 and online Supplemental Table S3). Outside of the oldest age range under study, findings suggest that although PC-DFLE slightly increases in total years, it is relatively unchanged as a proportion of total life in each age range across successive cohorts. The proportion of partial LE spent disability-free does appear to rise over cohorts in ages 84–89, largely owing to gains in disability-free life among women.
Morbidity-Free Life Expectancy
Life spent morbidity-free declines across successive cohorts (Figure 3 and online Supplemental Table S4). The proportion of total partial LE spent without chronic morbidities declines substantially across cohorts, dropping from 72% of life in ages 54–59 in the cohort born in 1942–1947 to 65% in the cohort born in 1954–1959 in the combined sample. The gradient from the early to late cohort is even steeper in later age ranges, where in both proportionate and absolute terms, more recently born cohorts are living significantly fewer years free of chronic morbidities. Men lived substantially fewer years of morbidity-free life compared with women. Above age 60, men could expect to spend fewer than half of their years free of chronic morbidities.
Healthy Life Expectancy
Healthy life expectancy provides an estimate of the years of life an individual can expect to live with good self-rated health—in this case, reporting “good,” “very good,” or “excellent” self-rated health. Patterns of PC-HLE across successive birth cohorts in the U.S. population are presented in Figure 4 and online Supplemental Table S4. At younger old ages (ages 54–59 and 60–65), little change is evident across cohorts, with individuals expected to spend about three quarters of these years with good self-rated health. In older ages, a gradient begins to emerge. In ages 66–71, 72–77, 78–83, and 84–89, successive cohorts report living significantly more years in good self-rated health between the early and later cohorts. Although the total proportion of healthy life declines with increasing age, more recently born cohorts report spending proportionately more years in good health than cohorts born further in the past. These gradients are more pronounced among men at older ages, with a comparatively smaller gradient seen at all ages for women.
Educational Differences in PC-LE and PC-HEs
I find substantial heterogeneity in different birth cohorts' experiences with disability depending on educational attainment (Figure 5). A steep gradient exists in the proportion of partial life expectancy spent free of ADL disability between those without a high school diploma and those with a bachelor's degree or more. For example, a person with less than a high school diploma born in 1954–1959 could expect to live only 74% of life in ages 54–59 disability-free, while a member of this cohort with a bachelor's degree could expect to live 95% of these ages free of ADL disabilities. Total partial LE increases across successive cohorts, though magnitudes are small at younger ages compared with older ages.
Importantly, the pattern of DFLE across successive cohorts differs by level of schooling attainment. Individuals with less than a high school diploma are broadly experiencing an expansion of life spent with disabilities across successive cohorts, with significant declines in the proportion of life spent disability-free between the early and later cohorts in ages 60–65, 66–71, 72–77, and 78–83. In all other education groups, successive cohorts have comparable proportions of life spent disability-free across successive cohorts, suggesting that no substantial compression of disability is occurring in these groups. Tables of DFLE, MFLE, and HLE estimates by educational attainment and sex are available in the online Supplemental Materials.
Educational gradients in MFLE (Figure 6) are also substantial. Individuals with lower schooling attainment expected to live between 0.5 and 0.8 fewer years free of major morbidities than did individuals with the highest schooling attainment in ages 54–59, 60–65, 66–71, 72–77, and 78–83. As a proportion of partial LE, these gradients are largest at younger ages and diminish quite substantially after age 72—although individuals with a bachelor's degree or more retain some advantage at older ages. Focusing on changes across cohorts, patterns are comparable at each level of schooling attainment—with a few exceptions, successive birth cohorts appear to experience proportionately more time spent with chronic morbidities, regardless of their level of schooling.
Patterns of healthy life expectancy by educational attainment display considerable variation, with successive cohorts showing markedly different patterns depending on their level of completed schooling (Figure 7). At younger adult ages (ages 54–59 and 60–65), individuals with a high school diploma or less are experiencing a significant decline in the proportion of life spent healthy across successive cohorts. These declines are quite substantial for those with less than a high school diploma: PC-HLE dropped from .55 to .49 as a proportion of total LE in ages 54–59 between the early and later cohorts, and from .52 to .40 in ages 60–65. In other educational groups, PC-HLE is similar across cohorts in ages 54–77. At older ages, most education groups show improvements in the proportion of life spent healthy across cohorts, although this pattern is not universal. However, there are differences in the proportion of these years spent healthy across different levels of completed schooling—those without a high school diploma could expect to spend just over half of these years with good self-rated health, compared with 70–80% of these years for those with some college or a college degree.
Racial/Ethnic Differences in PC-LE and PC-HEs
Differences in both level and trend emerge when exploring PC-DFLE by race/ethnicity (Figure 8). Overall, non-Hispanic White individuals expect to spend proportionately more life disability-free than either non-Hispanic Black or Hispanic individuals. In ages 54–59, non-Hispanic Whites live about 90% of total life years disability-free, compared to about 80% for non-Hispanic Blacks and around 83% for Hispanics. The magnitude of these differences is similar in other age ranges, although smaller in the 84–89 age-group. Patterns across cohorts show that increases in disability-free life are tracking with overall increases in PC-LE among non-Hispanic Whites, leading to a consistent proportion of partial LE spent disability-free across cohorts. These patterns are more varied among the non-Hispanic Black and Hispanic populations. Successive cohorts of non-Hispanic Black individuals are experiencing proportionate declines in PC-DFLE in ages 54–59 and 84–89, while the proportion of partial LE spent disability-free is relatively stable across cohorts at other ages. Successive cohorts of Hispanic individuals experience similar levels of PC-DFLE below age 72, but appear to see proportionate decreases in disability-free life after this age (though these differences are not universally significant, owing in part to small sample sizes of older Hispanics in the HRS). Tables of DFLE, MFLE, and HLE estimates by racial and ethnic groups and sex are available in the online Supplemental Materials.
Figure 9 displays patterns of PC-LE and PC-MFLE across racial and ethnic groups. The proportion of life in each age spent with chronic morbidities is similar across groups, although at older ages Hispanics can expect to live slightly fewer years with diagnosed chronic morbidities than non-Hispanic Blacks or non-Hispanic Whites. For all three groups, the proportion of partial LE spent with chronic morbidities and the absolute life-years spent with morbidities increase across successive cohorts in almost every age range.
The patterns of PC-HLE across successive cohorts by race and ethnicity are provided in Figure 10. Substantial differences exist in the proportion of partial LEs spent healthy between racial and ethnic groups, with non-Hispanic Whites reporting spending more years with good health than non-Hispanic Blacks and Hispanics. In ages 54–59, non-Hispanic Whites spend around 80% of LE with good self-rated health, compared with 65% for non-Hispanic Blacks and 60% for Hispanics. These disparities are similar across age-groups—that is, even at older ages where the overall proportion of life spent in good self-rated health declines for all groups, non-Hispanic Whites still expect to spend 10–20% more partial LE with good self-rated health than do non-Hispanic Blacks and Hispanics. At older adult ages (70+), more recently born cohorts of non-Hispanic Blacks and Whites appear to be living proportionately more years with good health, although the gradient across cohorts is not universally significant. Patterns are more mixed for Hispanics, with little change seen over successive cohorts—and confidence intervals are quite wide at older ages.
This article evaluates cohort patterns in population-level healthy, disability-free, and morbidity-free life in the U.S. population and finds that six key trends are occurring across successive birth cohorts. First, partial life expectancies are increasing across cohorts, particularly at older ages. Second, though partial life expectancies are increasing, I find little evidence for a substantial compression of time spent disabled as a proportion of total partial LE. Third, across all age-groups investigated, life expectancy without chronic morbidities declined substantially over successive cohorts. Fourth, patterns of PC-LE and PC-HEs across cohorts are generally more favorable at older ages than at younger ones—that is, PC-LE is increasing more rapidly at older ages, and these ages are also seeing proportionate increases in PC-HLE and PC-DFLE. In contrast, gains in PC-LE are smaller at younger ages, and I find little evidence of proportionate improvements in PC-DFLE or PC-HLE in these ages. Fifth, substantial heterogeneity in PC-LE and PC-HEs exists by both level of schooling and race/ethnicity. Finally, patterns of later-life health differ substantially by gender in the U.S. population, with men on average living proportionally more years disability-free than women, but also living with a greater burden of chronic morbidities.
These analyses find substantial differences in PC-HEs by race/ethnicity and level of schooling, with less advantaged groups generally having shorter partial LE and living proportionately more years with disabilities, morbidities, and poor self-rated health. In addition to these overall lower levels of PC-LE and PC-HEs, more disadvantaged groups experienced worsening trends in HEs across cohorts. Successive cohorts of lower educated individuals and non-Hispanic Black individuals experienced an expansion of life spent in poor health and with disabilities among younger age-groups. These differential health and disability patterns may result from more educated or more advantaged groups having better access to infrastructure, accommodations, and services that may act to support their ability to conduct basic activities of daily living. Alternatively, they could represent the deterioration of living conditions among lower socioeconomic status groups in the United States both during and after the Great Recession (Bauer and Shambaugh 2018; Kochhar and Fry 2014; Muennig et al. 2018).
A number of trends in population health were occurring over the observation period. Rates of obesity among older adults, though already climbing through much of the late twentieth century, rose quickly over the course of the study period (Hales et al. 2020; Parikh et al. 2007). Although this increase in obesity was primarily driven by period effects (i.e., secular increases in obesity rates across all ages in the population), BMI increases were also patterned by birth cohorts, with age-specific probabilities of obesity increasing across successive cohorts (Reither et al. 2009). In conjunction with rises in obesity, the proportion of the U.S. population diagnosed with diabetes also rose steadily between 1998 and 2016, a rise also seen across the cohorts used in this analysis (see online Supplemental Table S1). Though smaller in magnitude to the rise in diabetes, the shares of the sample reporting cancer, lung disease, and heart disease within each age-group all increased over successive cohorts in the HRS sample (Supplemental Table S1). Treatment availability and control of chronic conditions rose in the U.S. population in tandem with these rises in chronic conditions (Crimmins 2015). I also find that patterns over cohorts are more favorable in older age-groups than in younger ones, a finding that is consistent with recent evidence of troubling trends in health, disability, and mortality identified in late middle ages in the United States (Case and Deaton 2015; Freedman et al. 2013; King et al. 2013).
By any of the metrics used in this study, successive cohorts in the U.S. population are not experiencing the compression of morbidity predicted by Fries (1980), nor are they experiencing a universal expansion of time spent in poor health or with disabilities, as predicted by Gruenberg (1977). Broadly construed, this study's findings suggest that successive cohorts are experiencing patterns that most closely align with Manton's dynamic equilibrium framework (Manton 1982). That is, I find substantial increases in partial LE spent with chronic morbidities across successive cohorts, but much smaller variations in time spent with ADL limitation or with self-reported poor health. This aligns with Manton's suggestion that life expectancy improvements have resulted from reductions in the severity and rate of progression of chronic diseases (Manton 1982:227). The analyses of morbidity in this study rely on a somewhat blunt measure of being ever diagnosed with a serious chronic morbidity, and thus do not account for whether the condition is currently under control via medication or behavior modification. At the population level, a key open question remains: Is the later-life health of individuals with controlled conditions the same as that of individuals with the absence of these conditions? Future, more focused work is needed to investigate period and cohort patterns in morbidity control and multimorbidity in the U.S. population, as well as lower severity physical functioning impairments.
There is also the potential that changes in mortality selection in the U.S. population could play a role in shaping HEs across cohorts. If mortality selection is related to individual frailty, cohorts with higher mortality selection into older ages may be less frail and at lower risk of poor health and disabling conditions (Vaupel et al. 1979). Considerable debate exists as to whether these selection processes are deterministic of later-life health, or whether frailty in later life is malleable (Kannisto 1991). Although the present study is not focused on directly exploring the impact of mortality selection, the cohorts under study do vary considerably in their mortality histories, especially with regard to survivorship to older ages. The online Supplemental Figure S2 shows the proportion of each birth cohort that survived to ages 54, 60, 66, and 72 by gender, based on data from the Human Mortality Database (Human Mortality Database 2020). From the 1933 birth cohort to the 1948 birth cohort, the proportion of male newborns surviving to age 66 rose nearly 10% for men (from 65% to 74%) and over 5% for women. Similar rapid increases in survivorship to age 72 are seen across successive birth cohorts, as well as somewhat more gradual increases in survivorship to ages 54 and 60. These reductions in mortality selection could lead to successive cohorts at higher risk of disability and chronic disease, and may also play a role in shaping the growing disparity between educational groups (Zheng 2020).
Limitations and Strengths
Several limitations need to be considered when evaluating the results of this study. The current analyses follow a ﬁrst-order Markov chain and are thus not state-duration dependent—that is, transition probabilities are not adjusted by duration of stay in each state. Individuals who experience a health transition between waves of data collection are assumed to experience only a single transition during the period between surveys, which likely misses shorter term transitions between health statuses. There is also the potential that these findings, which compare cohort LE and HEs within bounded age ranges, may not fully align with patterns in completed HEs across cohorts. The most obvious situation where this could occur is the introduction of a substantial period shock that affects disablement or chronic morbidities, such as the introduction of a highly effective or curative treatment for diabetes. In such a situation, inferences based on these partial-cohort LE and HE measures could be poor representations of completed cohort life and health expectancies. Hence, these findings should not be understood as a precise substitute for measuring full-cohort life expectancies, but rather a way of monitoring how disability, morbidity, health, and mortality in key portions of the life course are changing across cohorts.
One of the main strengths of these analyses is the focus on understanding disability change across successive birth cohorts, rather than simply change over time. With a few notable exceptions (Bardenheier et al. 2016; Beltrán-Sánchez et al. 2016; Liu et al. 2019; Manton et al. 2008; Manton and Land 2000; Payne and Wong 2019), most work investigating the compression has relied on period comparisons—that is, measuring mortality and disability conditions in a population at different points in time, and observing changes using data aggregated across many cohorts in a population. Though this approach may be useful for monitoring aggregate population-level disability, these results do not easily translate to the experience of any given cohort of individuals. Disability and mortality are strongly linked to cohort-specific life course exposures, but period-based models provide estimates aggregated across cohorts, obscuring the substantial heterogeneity between birth cohorts (Manton et al. 2008; Palloni and Beltrán-Sánchez 2017). This study's estimates center on measuring how LE and HEs are changing within age-groups over successive birth cohorts, a metric more directly applicable to understanding trends in population health. Where similar data exist, the methods developed in these analyses could be used to elucidate patterns in health over time and across birth cohorts.
These results suggest that successive cohorts in the U.S. population may be on divergent paths in terms of late-life health and well-being. Although my findings are agnostic as to whether these changes are primarily driven by cohort or period effects, the continued increases in partial life expectancy spent with morbidities across successive cohorts, the stagnation of DFLE, and modest increases in HLE present a broad view of a twenty-first century in which successive cohorts in the United States are not living healthier lives. These findings stand in marked contrast to research from previous decades, which broadly found increases in LE and a compression of morbidity over successive cohorts through the 1980s and 1990s (Manton et al. 2008; Manton and Land 2000; Manton et al. 1997; Soneji 2006). Although partial life expectancies have risen for most groups in the population, these increases in lifespan are being matched by increases in disabled life expectancy and exceeded by increases in life expectancy with chronic diseases. If these patterns continue, the U.S. population faces the very real prospect of seeing the end of continued improvements in healthy longevity in coming decades.
C.F.P. is supported by an Australian Research Council Discovery Early Career Researcher Award (DE210100087) funded by the Australian Government, and by an ANU Futures Scheme Award funded by the Australian National University. The Health and Retirement Study is sponsored by the National Institute on Aging (grant NIA U01AG009740) and is conducted by the University of Michigan. The author wishes to thank the editors and four anonymous reviewers for their helpful comments on this research.
Note that these same transition probability point estimates could be estimated using the simpler PROC LOGISTIC model with a WEIGHT statement. The SURVEYLOGISTIC command was used to draw valid inferences for variable selection when testing interaction terms in the analysis model, but taking account of sample design in both the regression model and bootstrapping raises potential issues of overaccounting for the HRS sample design in variance estimation. However, only the point estimates from the PROC SURVEYLOGISTIC output are used in the HE calculations, as the confidence intervals are estimated via bootstrapping. Any potential implications of overaccounting for sample design would only impact variance estimates around the point estimates, which play no role in the HE estimation process. This issue is discussed in more depth in Cai et al. (2010).