The income-mortality gradient is the empirical relationship between household income (or socioeconomic status more broadly) and life expectancy (LE): higher income is associated with longer life throughout the income distribution. This within-country association should be distinguished from the cross-country Preston Curve (income vs. LE across nations), which has shifted upward over time primarily due to technology diffusion rather than income growth — indicating that income is an intermediate, not ultimate, determinant of mortality at the global level. In the United States, the gradient is steep, continuous, and widening over time. Waldron (2007) established the widening trend across successive birth cohorts using Social Security Administration (SSA) administrative cohort data; Chetty et al. (2016) quantified it at national scale using period methods and a near-universal income coverage.
| Metric | Men | Women |
|---|---|---|
| Expected age at death, bottom 1% | 72.7 | 78.8 |
| Expected age at death, top 1% | 87.3 | 88.9 |
| LE gap, top vs. bottom 1% | 14.6 yrs | 10.1 yrs |
| LE gain 2001–2014, top 5% | yrs | yrs |
| LE gain 2001–2014, bottom 5% | yrs | yrs |
The -year gap for women between top and bottom is equivalent to the mortality decrement from lifetime smoking.
Waldron (2001) is the immediate precursor to the 2004 study, using the same 1973 Exact Match dataset but focusing on education as the socioeconomic status (SES) proxy (the earnings variable was not yet available). Key gradient finding: men with less than years of education face greater odds of dying than college-educated men — the same magnitude as retiring at exactly age 62 vs. age 65+. Education differentials widen across birth cohorts 1906–1932, foreshadowing the widening gradient that Waldron (2007) and Chetty et al. (2016) document at larger scale. The compounding of education and early retirement risk is a direct implication: low-educated early retirees face both penalties simultaneously. See Select and Ultimate Mortality Tables for the age-62 spike mechanism.
Waldron (2004) provides an earlier look at the earnings gradient within a specific Social Security subpopulation: men who claimed retired-worker benefits before age 65. The paper finds that the small group whose health equaled that of age-65 retirees (Group C in the trichotomy model) was concentrated in the top earnings quartile — men retiring at exactly age 64 with high lifetime earnings. Below-median earners who claimed early were categorically worse off in health and mortality. This is an early direct demonstration of a steep SES-health gradient within a Social Security subpopulation, predating and foreshadowing the broader earnings-rank analyses.
Waldron (2007) is the empirical precursor to Chetty et al. Using SSA's Continuous Work History Sample (CWHS) matched to Numident death records — male Social Security–covered workers, birth cohorts 1912–1941, deaths observed at ages – from 1972–2001 — she measures mortality by whether peak-age earnings (ages 45–55) fell in the top or bottom half of the covered-worker distribution.
Key findings:
Waldron uses cohort LE (following birth cohorts over time); Chetty uses period LE (cross-sectional snapshot). They measure different things but are complementary: Waldron establishes the generational dimension of the widening (1972–2001 data); Chetty confirms it continues through the most recent period (2001–2014). Together they form a continuous evidentiary record from the early 1970s through 2014.
An important limitation: Waldron's sample excludes zero earners at ages 45–55 ( of men, the most at-risk group). Men not in the labor force at age 45 have fewer years of LE than participants (Rogot et al. 1992). The Waldron estimates therefore understate the true population LE gap, particularly at the bottom. Within birth cohorts, the earnings-mortality gap also narrows at older ages — not because of health convergence but because frailer low-earners die first, leaving a selectively robust survivor group. See Period Mortality for a fuller treatment of this frailty problem.
The choice of SES proxy matters enormously for interpreting mortality trends, not just for cross-sectional estimates but especially for time trends.
Earnings rank (Waldron 2007; Chetty et al. 2016) classifies individuals by their position in the earnings distribution — a measure that is inherently cohort-relative and stable as a population concept across time. Rising real wages or changing income distributions do not distort comparisons across years because the researcher always compares the same relative position.
Education credential (Olshansky et al. 2012 and much of the education-mortality literature) classifies individuals by whether they completed a given credential — typically a high school diploma. This appears straightforward but creates a serious compositional problem over time: as educational attainment rises, the non-diploma group becomes increasingly adversely selected.
In the US, high school graduation rates rose from for cohorts born 1903–1921 to for cohorts born 1919–1937. Concretely:
Comparing the "less than high school" group in 1990 to the same group in 2008 means comparing very different segments of the population. The 2008 non-diploma group is a smaller, more adversely selected residual — people who failed to complete an increasingly universal credential — whose mortality would be elevated by selection alone, regardless of any causal effect of the credential.
Bound, Geronimus, Rodriguez, and Waidmann (2015) formalize this critique using the same Multiple Cause of Death data as Olshansky et al. (2012), which had reported dramatic LE declines of – years for low-education whites. Reclassifying education by bottom-quartile rank within each birth cohort, they find:
| Group | Olshansky (absolute credential) | Bound et al. (relative rank) |
|---|---|---|
| White women, 1990–2010 | years LE | years LE |
| White men, 1990–2010 | years LE | years LE |
| Black men and women | No decline | No decline |
The dramatic decline largely collapses. A modest, real stagnation for white women of low SES survives — genuine, but not the crisis the credential-based measure implied. Dowd and Hamoudi (2014) proved the artifact mechanism formally via simulation: rising educational attainment alone can reproduce the Olshansky pattern even in data where every group's LE is improving.
The 2008 cross-sectional LE estimates from Olshansky et al. (2012) — which are not subject to the compositional bias critique, since they are within-year snapshots — provide the most detailed race × sex × education breakdown available for this period:
| Group | Within-group LE gap (highest vs. lowest education) |
|---|---|
| White men | 12.9 years |
| White women | 10.4 years |
| Black men | 9.7 years |
| Black women | 6.5 years |
| Hispanic men | 5.5 years |
| Hispanic women | 2.9 years |
The widest single disparity: years between white men with years of education and black men with years. For women: years between the same credential groups.
Cross-race education reversal: Highly educated blacks and Hispanics ( years) outlive low-education whites ( years) by – years — education dominates race at the credential extremes. But a residual racial gap persists even after conditioning on education: highly educated black men live approximately fewer years than equally educated white men, indicating that race operates through channels — weathering, chronic discrimination stress, residential segregation — not fully captured by educational attainment.
What this means for the gradient: The earnings-rank approach of Waldron and Chetty is the more robust methodology for tracking mortality trends over time. Studies using absolute education credentials to track trends (as opposed to cross-sectional levels) are vulnerable to this compositional bias whenever attainment is rising. The underlying inequality story — widening LE gaps between high and low SES, stagnation at the bottom — is confirmed by both approaches; the magnitude and the narrative (decline vs. stagnation) differ.
Corroboration via ordered probit (Sanzenbacher et al. 2017): Using National Longitudinal Mortality Study (NLMS) data (1.5M individuals, 1979–2011), Sanzenbacher et al. develop a more principled relative-rank approach: an ordered probit assigns each individual's probability of completing each education level, then reassigns individuals until each quartile represents exactly 25% of their birth cohort. Under this procedure, all four quartiles gained LE (men Q1 +4.1 yrs, Q4 +5.9 yrs at age 65; women Q1 +1.3 yrs, Q4 +3.1 yrs). Annual mortality improvement ranges from 0.5% (women Q1) to 2.4% (men Q4). The Olshansky apparent declines are entirely explained by the two simultaneous forces the ordered probit separates: rising SES-mortality inequality plus rising compositional disadvantage in the absolute low-education group. The corroboration of Bound et al. on an independent dataset using a more rigorous method effectively closes this methodological debate. See Sanzenbacher et al 2017 — Rising Inequality in Life Expectancy by Socioeconomic Status and Education-Mortality Gradient.
Chetty et al. estimate period life expectancy at age (see Period Mortality) using Gompertz extrapolation above age . Income is measured as household earnings percentile rank within sex and age-year, lagged years to mitigate reverse causality. Estimates are race- and ethnicity-adjusted using NLMS data. Data source: billion person-year observations from Internal Revenue Service (IRS) tax and SSA death records, 1999–2014.
For individuals in the bottom income quartile, life expectancy varies years across U.S. commuting zones — a larger range than many national health interventions achieve. The strongest predictors of high LE for low-income individuals are:
| Correlate | r | Direction |
|---|---|---|
| % immigrants | + | |
| Smoking rate | − | |
| Median home values | + | |
| Government expenditures per capita | + | |
| Obesity rate | − | |
| % college graduates | + |
Not significantly correlated: healthcare access (insurance rates, Medicare spending, preventive care quality), income inequality (Gini index), residential segregation, local labor market conditions (unemployment, population change).
The implication: the proximate drivers of geographic variation in low-income mortality are health behaviors, not healthcare access or economic structure.
The lowest-LE commuting zones for the bottom income quartile form a geographic belt: Indiana, Ohio, Michigan (industrial Midwest), Kentucky, Tennessee, Arkansas, Oklahoma, Kansas. The highest-LE zones are in California, New York, and Vermont. This pattern overlaps directly with Disability Insurance (DI) incidence geography — see below.
The same commuting zone (CZ) geography appears in the Intergenerational Income Mobility literature (Chetty et al. 2014c): absolute upward mobility is lowest in the Southeast and Appalachian belt and highest in the Mountain West and rural upper Midwest. The geographic overlap is near-total. The shared low-mobility, low-LE corridor reflects common upstream drivers — manufacturing decline, low educational attainment, concentrated poverty — rather than two separate phenomena.
Pampel, Krueger, and Denney (2010) synthesize the behavioral science literature into a nine-mechanism taxonomy organized by a motives vs. means framework. Health behaviors — primarily smoking, physical inactivity, diet, and heavy alcohol use — account for of the income-mortality gradient, making them an important but not dominant intermediate pathway.
The strongest empirical decomposition (Cutler and Lleras-Muney 2010) partitions the education-behavior gap: income, insurance, and family structure account for ; knowledge and cognition ; social networks ; time discounting and personality . The near-zero contribution of time discounting directly undermines rational addiction theory as the primary explanation for SES disparities in health behaviors.
Cutler, Lange, Meara, Richards-Shubik, and Ruhm (2011) make a methodologically distinct contribution: they ask not whether behaviors explain the cross-sectional education-mortality gap (they partly do) but whether changes in behaviors explain why the gap widened between the early 1970s and late 1990s. Using National Health and Nutrition Examination Survey (NHANES) I (1971–75) and National Health Interview Survey (NHIS) (1987–2000) with Oaxaca-style hazard decompositions, they find that behavioral risk factor changes explain of the widening for men and at most – for women — and the latter is not statistically significant. The paradox: smoking declined more for the college-educated, but the differential is concentrated among adults under 60; among adults 60+, who account for most deaths, the smoking differential is absent or reversed. Obesity grew in parallel across education groups; hypertension and cholesterol improved similarly. More than of the gradient widening is explained by growing coefficient effects — rising returns to education conditional on behaviors, and rising mortality hazards from the same risk factors. The policy implication is direct: even complete equalization of behavioral risk factor distributions across education groups would not substantially close the gap. See Education-Mortality Gradient for full detail.
For the income-mortality gradient specifically, three mechanisms are most relevant to the geographic and structural patterns documented above:
The geographic correlates Chetty et al. (2016) identify as the strongest predictors of low-income mortality (smoking rate , obesity rate ) are precisely mechanisms 1 and 8 expressed at the commuting-zone level. Health behaviors are not independent causes of the gradient but are themselves downstream of structural conditions — manufactured job loss, residential segregation, disinvestment — that create both low SES and unhealthy built environments simultaneously. See SES Health Behavior Gradient for the full taxonomy.
The SES stratification of obesity is visible in Olshansky et al.'s (2005) race-sex-specific counterfactual: the obesity LE penalty is largest for black males (– years vs. – for white males), and the projected – year future burden falls heaviest on minority and youth cohorts already carrying high obesity prevalence — the same population clusters concentrated at the bottom of the income-mortality gradient.
DI disabled-worker beneficiaries are overwhelmingly low-income — by definition, they cannot engage in Substantial Gainful Activity. They cluster at or below the bottom income quartile. The worst-LE tier documented by Chetty — expected age at death – for bottom-quartile men — overlaps substantially with the mortality experience of DI beneficiaries documented by Meseguer 2021 — DI Beneficiary Mortality vs. General Population.
The commuting zones with the worst life expectancy for low-income individuals — Gary IN, Detroit MI, Toledo OH, Oklahoma — are the same regions with the highest DI application and award rates. This is not coincidental: concentrated low-income populations with high rates of smoking, obesity, and musculoskeletal disease drive both DI incidence and premature mortality through the same underlying conditions.
Woolf and Schoomaker (2019) provide quantitative confirmation: Ohio, Pennsylvania, Indiana, and Kentucky — only of the U.S. population — accounted for of all estimated excess midlife deaths 2010–2017. This four-state Ohio Valley cluster is the geographic epicenter of both the DI high-incidence belt and the deaths-of-despair crisis, driven by the same industrial decline, low educational attainment, and opioid exposure.
Top men claim Social Security (SS) and Medicare for more years than bottom men ( more years for women). The progressive benefit formula substantially understates redistribution toward the rich, because high earners collect far longer. Bosworth and Burke (2014) estimate that differential mortality offsets approximately half of overall SS system progressivity; for men, the offset is even larger because the LE gap by SES is wider at older ages. The National Academy of Sciences (NAS)/Auerbach et al. (2017) cohort comparison quantifies the trajectory: between the 1930 and 1960 birth cohorts, top-quintile men's lifetime SS benefits rose from $229k to $295k (2009 dollars) while bottom-quintile men's benefits fell in absolute real terms. The growing mortality gap has therefore produced a secular erosion of Social Security's redistributive effectiveness — even with an unchanged benefit formula. See Social Security Progressivity.
However, DI is different: DI benefits reach individuals before retirement, in the worst-LE tier, before they have accumulated the survival advantage that makes retirement benefits flow disproportionately to high earners. Cash DI benefits also directly reduce mortality — Gelber, Moore, and Strand (2018) estimated that a $1,000/year increase in DI benefits reduces annual mortality by percentage points (cited in DI Beneficiary Mortality). The income-LE gradient makes this finding mechanically coherent: shifting low-income individuals up the income distribution even modestly reduces mortality risk. Including DI in progressivity analyses restores combined Old-Age and Survivors Insurance (OASI)+DI system progressivity even at the household level (Steuerle et al. 2004a; Gustman and Steinmeier 2001).
Woolf and Schoomaker document that state LE gaps widened to years by 2017 (Hawaii highest, Mississippi lowest), with divergence beginning in the 1980s and accelerating in the 1990s. Adjacent state pairs diverged sharply: Colorado/Kansas gap grew from to years; Alabama/Georgia from to years. The authors attribute this to neoliberal devolution of federal social policy authority to states in the early 1990s, creating systematic differences in minimum wage, Earned Income Tax Credit (EITC) generosity, Medicaid expansion, and gun laws between otherwise comparable neighboring states.
For DI mortality research, this state divergence matters on two levels: (1) the U.S. general population mortality — the denominator in DI/general-population ratio studies — is increasingly heterogeneous and deteriorating in exactly the regions with the highest DI incidence, making the ratio harder to interpret without regional stratification; (2) state-level social policy variation affects mortality through the same social determinants (economic stress, health insurance access, violence) that also drive DI application rates, suggesting common confounders for both outcomes.
Proposals to index Social Security and Medicare eligibility ages to LE gains are explicitly flagged by Chetty et al. as regressive: LE gains have been concentrated in the top of the income distribution. For populations near DI eligibility age thresholds (50, 55 under the Vocational Grid), raising those thresholds would deny benefits to individuals whose LE is already far below the national average — compounding an existing inequality rather than correcting it.
Chetty et al. explicitly address the tension with Case and Deaton's Deaths of Despair findings. Chetty finds increasing LE across most income groups during 2001–2014, which appears to contradict Case and Deaton's finding of rising midlife mortality for white men. The difference is methodological: Chetty pools all races, excludes $0-income individuals (who have mismeasured mortality in SSA data), and focuses on LE at 40 rather than age-specific mortality rates at 45–54. Within their framework, Chetty confirms that LE gains have been far smaller for low-income individuals — consistent with widening inequality rather than aggregate decline.
Case and Deaton (2017) extend this by testing the income hypothesis head-on. The aggregate co-movement of white non-Hispanic (WNH) median household income and all-cause mortality (both appearing to turn down around 1999) initially looks like evidence for income causality. But three tests falsify it:
The racial control: Black non-Hispanic (BNH) and Hispanic incomes tracked WNH incomes through the full period, and college-educated BNHs experienced steeper post-1999 income declines than comparable WNHs. Yet BNH and Hispanic mortality fell at to /year. No contemporaneous income story can explain this without invoking a race-specific mechanism.
The within-WNH education split: Median household incomes of high school (HS)-or-less and bachelor's degree (BA)+ WNHs do not diverge in the way mortality does. By education group, income trends do not produce a "scissors" pattern; mortality does. Income stagnation alone cannot explain why HS-or-less WNH mortality rose while BA+ WNH mortality fell.
The European evidence: Countries hit hardest by the Great Recession (Ireland, Spain, UK, Netherlands) experienced income declines as large as the U.S. turnaround, but showed no slowdown in their mortality decline. Countries unaffected (France, Germany, Sweden) had stable incomes but the same mortality trajectory — steady improvement.
The aggregate U-shape in income co-moving with the aggregate U-shape in mortality is a statistical coincidence: deaths of despair rise smoothly since 1990, while heart disease decline decelerates smoothly — together generating an inflection around 1999 that happens to coincide with the median income peak. Neither component has a structural link to income timing.
Implication for the wiki: The income-mortality gradient documented by Chetty et al. (2016) — a cross-sectional, level relationship — is real and large. But Case and Deaton (2017) establish that changes in contemporaneous income explain very little of the temporal mortality reversal. The gradient operates through long-run structural channels (education, health behaviors, health capital accumulation over a lifetime) rather than through the kind of year-to-year income sensitivity that a purely contemporaneous story would require.
Bor, Cohen, and Galea (2017) — a systematic review in The Lancet covering US evidence 1980–2015 — make an original empirical contribution by reanalyzing the Chetty et al. (2016) supplementary data to decompose why the income–LE gap widened between 2001 and 2014. Their two-channel framework is the key methodological contribution:
Channel 1 — Income distribution shift: Falling real incomes at the bottom (25th percentile household incomes fell for both men and women 2001–14) would widen the top–bottom LE gap mechanically, even if the gradient slope were unchanged.
Channel 2 — Gradient steepening: The same income level became a more powerful predictor of mortality — i.e., the LE penalty per unit of income became steeper — independently of any income distribution change.
Their decomposition shows that gradient steepening is the dominant factor. Specifically, the income–survival gradient (measured as life expectancy per log-point of household income) shifted:
| Sex | 2001 gradient | 2014 gradient |
|---|---|---|
| Women | yrs/log-point | yrs/log-point |
| Men | yrs/log-point | yrs/log-point |
Income distribution changes explain approximately one-third (men) to one-sixth (women) of the divergence; the steepening gradient explains the remainder.
The $60,000 threshold: The gradient steepened only for households earning below approximately $60,000/year. Above that level — rich vs. upper-middle-class — the income–LE relationship was unchanged. This threshold finding is consistent with a tipping dynamic in which the feedback loops of the Health-Poverty Trap operate most intensely below a specific income floor.
Policy implication of the decomposition: If income distribution changes were the dominant channel, redistribution alone might close the gap. If gradient steepening dominates, the mechanism is more complex — the same income level is generating worse health outcomes over time, suggesting that the environment in which low-income individuals live (neighborhood conditions, healthcare access, behavioral norms, substance availability) deteriorated independently of income levels. Breaking the gap therefore requires targeting the slope of the gradient, not just the income distribution.
Global Burden of Disease (GBD) 2021 (US Burden of Disease Collaborators 2024) provides the most granular documentation of geographic inequality in US health outcomes, extending through 2021 and using a -country comparison set.
| Metric | Best (Hawaii) | Worst |
|---|---|---|
| Life expectancy | 81.2 years | Mississippi: 71.9 years |
| HALE | 68.3 years | West Virginia: 59.5 years |
| Age-std. mortality rate | 433.2/100,000 | Mississippi: 867.5/100,000 |
| YLD rate | 12,085.3/100,000 | West Virginia: 14,832.9/100,000 |
The Hawaii–Mississippi mortality rate ratio of approximately captures the full within-country range. West Virginia's healthy life expectancy (HALE) of 59.5 years ranked st globally (out of other countries) — a WV male can expect fewer healthy life-years than the average male in other countries.
states lost life expectancy. The six worst: WV ( y), Oklahoma ( y), Mississippi ( y), Kentucky ( y), New Mexico ( y), Arkansas ( y). These are the same states with the highest DI incidence, highest drug use disorder burdens, and lowest educational attainment — confirming the DI geography-income gradient overlap documented in Chetty et al. (2016).
GBD 2021 observes that "improvements in health outcomes as a result of better access to medical care appear to have stalled during the period of expanded access compared with the preceding two decades (1990–2010)." The preceding 20 years (1990–2010) showed larger per-decade gains in LE and HALE rank than the post-2010 decade despite lower insurance coverage. The paper explicitly concludes: "health policies that only increase the proportion of the US population with insurance are only one part of the comprehensive strategy that is needed."
This finding is methodologically important for the income-mortality gradient literature: if the gradient operates through health behaviors, built environment, and cumulative disadvantage rather than through healthcare access per se (as Chetty et al. 2016 found for geographic variation), then expanded insurance coverage would be expected to have limited impact on gradient-driven mortality — consistent with the GBD 2021 stagnation finding.
West Virginia, Kentucky, Oklahoma, Pennsylvania, New Mexico, Ohio, Tennessee, and Arizona had higher age-standardised years lived with disability (YLD) rates than all 204 GBD countries in 2021. This is the YLD analog of the LE/HALE rank decline — not just mortality is worse, but disability burden is also more severe than in the rest of the world. The geographic overlap with the DI high-incidence belt is near-total.
Poterba, Venti, and Wise (2013a) extend the education-health-mortality gradient into the post-retirement wealth domain using Health and Retirement Study (HRS) panel data (1992–2008, ). The paper decomposes the education-wealth gap into five pathways: health levels at retirement, portfolio returns, Social Security income, defined benefit (DB) pension income, and a residual.
Health as pathway: Education produces a percentage point gap in health at retirement between college graduates and high school dropouts, but post-retirement health trajectories are education-independent. Education's health effect is a pre-retirement stock, not an ongoing retirement flow. This implies that the SES-health gradient (at least as it affects wealth) operates through health selection into retirement, not through differential aging.
Portfolio return gradient: College-educated households earn an average two-year return of vs. for those with less than high school — a differential. This is the single most powerful identified pathway to the education-wealth gap. The mechanism is portfolio composition: higher-education households hold more equities; lower-education households hold more low-yield assets (savings accounts, certificates of deposit [CDs]).
Social Security income as asset protector: A $10,000/year increase in Social Security income is associated with $7,000–$27,000 more non-annuity wealth (depending on quintile). SS income reduces the need for precautionary drawdown, allowing non-annuity assets to compound longer. DB pension income shows a much smaller and less consistent protective effect.
Gap magnitudes: The total education wealth gap ranges from (Q1) to (Q5). Identified pathways explain roughly of the gap in upper quintiles; remains unexplained in the lower quintiles, suggesting uncaptured channels (consumption behavior, inheritance, marriage) matter more at lower wealth levels.
Lifecycle model puzzle: Subjective survival probability has no statistically significant effect on asset drawdown — a direct rejection of a core lifecycle model prediction that individuals expecting shorter lives draw down faster.
These findings complement the earnings-rank gradient documented by Waldron (2007) and Chetty et al. (2016): the SES wealth gap documented through life at mortality is the endpoint of a process that begins with health at retirement entry, amplifies through portfolio returns, and is cushioned (unequally) by Social Security income.
Pijoan-Mas and Ríos-Rull (2014) demonstrate that standard period life expectancy (LE) grossly overstates the true within-cohort SES gradient by holding SES characteristics fixed at a given level for the entire remaining life — ignoring that income, wealth, health, and marital status all evolve stochastically over time. Using a structural hazard model with time-varying endogenous covariates (health, wealth, income, labor market status, marital status, smoking) estimated on HRS white Americans (1992–2008), they compute expected longevity — the true within-cohort gap in remaining life-years.
Overstatement ratios:
Three-channel decomposition (education, white men, Table 4):
| Channel | Years | Share |
|---|---|---|
| Initial health at 50 | ||
| Health trajectory after 50 | — dominant | |
| Conditional mortality (given health) | null |
For wealth: (a) yrs (), (b) yrs (), (c) yrs (). For labor market: initial health dominates ( of yrs total).
Marital status and smoking differ: conditional mortality accounts for years of the marital gap (between one-third and one-half of the total) and years of the smoking gap — indicating these characteristics have biological/social survival effects beyond their health impacts.
Education: no direct mortality effect conditional on health. The entire education-mortality gradient is mediated through initial health stock and subsequent health deterioration; once health is controlled, college and no-HS white men die at the same rate. This is the sharpest identification of the causal structure: education operates through health investment pathways, not through a direct biological survival advantage.
Widening gradients 1992–2008 (change in expected longevity gap, Table 3):
Methodological implication: Many descriptive studies reporting large SES-mortality gradients (including early Chetty et al. framing based on period LE) substantially overstate the within-individual gap. A inflation of the income gradient means that income-mortality correlation studies conflate the within-person effect with the cross-sectional compositional pattern. See Pijoan-Mas and Ríos-Rull 2014 — Heterogeneity in Expected Longevities.