Period Mortality

Definition

Period mortality (also "period life expectancy") measures mortality using the age-specific death rates observed in a single calendar period (e.g., a 5-year window), applied as if a hypothetical person would experience those rates for the rest of their life. It is a cross-sectional snapshot, not a longitudinal cohort estimate.

Key Ideas

Period vs. cohort: Period life expectancy averages across different birth cohorts alive in a given year. Cohort life expectancy follows one birth cohort over time. Period estimates lag behind improving cohort experience and can mask widening generational gaps.
Gompertz's Law: Human mortality increases exponentially with age — linear in the logarithmic scale. This is why mortality is typically plotted on a log scale for wide age ranges.
Select vs. ultimate: For populations with a qualifying event (like disability insurance (DI) entitlement), mortality depends on both attained age and time since the event. The "select period" (10 years for recent DI studies) captures duration-specific mortality; after that, only attained age matters ("ultimate" mortality).
Empirical regularities of the US all-cause age profile (Meseguer 2006): Three features are consistent across decades in Office of the Chief Actuary (OACT) period life table data. (1) The "elongated V" shape: log mortality declines smoothly from birth to ages 10–14, then increases near-linearly for all remaining ages — a consequence of Gompertz's law plus the childhood mortality decline. (2) An approximately linear downward trend in the Lee-Carter $k_t$ index across the $1928$ – $2001$ sample, reflecting uniform secular improvement across age groups. (3) Cross-age correlation in $\Delta$ log(mortality) that declines with age distance: adjacent age groups (e.g., $30$ – $34$ and $35$ – $39$ ) have correlations near $0.88$ , while distant groups (e.g., age 0 and ages 95+) have correlations near $0.19$ . The "accident hump" — a mortality spike at ages $15$ – $19$ from motor-vehicle fatalities — is visible in the age profile from the second half of the 20th century onward. These regularities have direct implications for forecasting model design: models that exploit cross-age correlation structure (such as Bayesian vector autoregression (BVAR) with correlation-calibrated cross-equation weights) outperform univariate models. See SSA Mortality Forecasting.

Cohort Life Expectancy: Construction and Limits

Cohort life expectancy (LE) is often described as if it were purely observed, but in practice it is a hybrid of observation and projection. Waldron (2007) provides a concrete illustration. Using Social Security Administration (SSA)'s Continuous Work History Sample (CWHS) matched to the Numident death file, she observes deaths at ages 60–89 from 1972–2001. For the 1912 birth cohort, deaths are fully observed across the entire age range; for the 1941 birth cohort, deaths are observed only at age 60 — everything beyond that must be projected from a regression model extrapolating 30 years of observed trend into the next 30 years. This means that as birth cohort increases, cohort LE estimates become increasingly reliant on projection assumptions rather than data.

The frailty problem: Within any given birth cohort, earnings-mortality differentials narrow markedly at older ages. Men in the bottom half of the earnings distribution at ages 60–64 face $\approx 50$ – $80\%$ greater odds of dying than men in the top half; by ages 80–84, the gap may be statistically indistinguishable. This does not mean the two groups converge in health — it reflects survivor selection: frailer low-earners die before reaching older ages, leaving a selectively robust survivor group. Because the frailer members are no longer observed at older ages, cohort LE estimates for low socioeconomic status (SES) groups at older ages overstate their true health status relative to younger cohorts entering the observation window. Waldron explicitly declines to control for this and designates her widening-differential findings as preliminary on that account.

The practical implication: the widening cohort LE gap documented by Waldron ( $1.2 \to 5.8$ years at age 60 across birth years 1912–1941) is likely an underestimate of the true gap, because (a) zero earners — the most at-risk group — are excluded from the sample, and (b) sample frailty pushes the apparent LE of low-earner survivors upward at older ages. See Income-Mortality Gradient.

Compositional Selection Bias in Subgroup Mortality Trends

A closely related artifact — distinct from frailty/survivor selection but sharing the same logical structure — arises when a subgroup's definition drifts in composition over time. The measured trend in subgroup mortality then conflates genuine change in mortality with change in who belongs to the subgroup.

The clearest example is education-credential mortality studies. Bound, Geronimus, Rodriguez, and Waidmann (2015) showed that the widely reported finding of dramatic LE declines among US whites without a high school diploma (Olshansky et al. 2012: $-3.4$ years for men, $-5.3$ years for women, $1990$ – $2008$ ) is largely such an artifact. Because US high school graduation rates rose sharply over the 20th century, "less than high school" captured the bottom $55\%$ of the population in 1990 but only the bottom $37\%$ by 2008. The credential group shrank and grew increasingly adversely selected — not because underlying mortality worsened, but because the definition captured progressively more disadvantaged individuals each year. Reclassifying by bottom educational quartile within birth cohorts, the apparent $3.9$ -year LE decline for white women shrinks to $1.2$ years; the apparent $2.2$ -year decline for white men becomes a $0.4$ -year gain.

The general principle: any subgroup defined by an absolute threshold (credential, income cutoff, diagnosis code) will experience compositional drift if the underlying distribution shifts over time. Trend analysis within such groups conflates true mortality change with selection change. The remedy is to classify by relative rank within cohort rather than by absolute threshold — which is why earnings-rank approaches (Waldron; Chetty et al.) are more robust for tracking mortality trends than credential-based approaches. See Income-Mortality Gradient.

Third Form: Within-Band Age Aggregation

A subtler variant arises when a broad age band (e.g., $45$ – $54$ ) is treated as a homogeneous unit for mortality rate calculation. Gelman and Auerbach (2016a) identified this artifact in Case and Deaton's deaths of despair paper. Case and Deaton computed group mortality as total deaths / total population for ages 45–54, without adjusting for the internal age distribution. As the baby boom generation passed through, the average age inside the band shifted upward over $1999$ – $2013$ — mechanically raising the group mortality rate even if no individual's age-specific rate changed.

The correction: compute age-specific rates for each single year of age within the band and take a weighted average under a fixed reference age distribution (the choice of reference distribution barely matters — uniform, 1999, or 2013 weights all yield the same conclusion). After adjustment:

The post-1998 increase splits into a real rise ( $1999$ – $2005$ ) and an artifactual rise ( $2005$ – $2013$ — explained entirely by age composition)
The trend is sex-asymmetric: genuine increase for women through 2013; near-reversal for men after 2005

The bias magnitude is modest ( $\approx 5\%$ in the $1999$ – $2003$ trend), so cross-country comparisons (which show $\approx 20\%$ differences) survive. But the absolute within-U.S. trend is substantially overstated.

This mechanism shares the logical structure of the credential drift problem: the denominator group's composition changes over time, and the measured trend conflates genuine mortality change with compositional change. The remedy is age-standardization to a fixed reference — conceptually identical to cohort-relative ranking in the credential context. See Deaths of Despair.

Woolf and Schoomaker (2019): 60-Year LE Joinpoints and State Divergence

Woolf and Schoomaker (2019) provide the most comprehensive picture of U.S. period life expectancy over the modern era, using Joinpoint Regression Program version 4.7.0.0 on the U.S. Mortality Database (LE 1959–2016) and Centers for Disease Control and Prevention (CDC) WONDER cause-specific data (1999–2017).

LE joinpoints: the long-run arc

U.S. LE at birth (both sexes combined): $69.9$ (1959) $\to$ $78.9$ peak (2014) $\to$ $78.6$ (2017). The overall increase of 9 years was not uniform:

Segment	Annual percentage change (APC)	Interpretation
$1959$ – $1969$	$+0.06$	Near-stagnation
$1969$ – $1979$	$+0.48$	Fastest growth — cardiovascular revolution
$1979$ – $2003$	$+0.19$	Slowdown — early human immunodeficiency virus (HIV)/acquired immunodeficiency syndrome (AIDS), then stabilization
$2003$ – $2011$	$+0.27$	Renewed improvement
$2011$ – $2016$	$\approx 0.00$	Plateau
2014–	negative	First sustained 3-year decline since 1918 flu

The 1969–1979 decade — when statins, hypertension treatment, and emergency cardiac care were introduced — contributed the largest per-year LE gains in the modern record. By 2011, gains had stalled. By 2017, the United States was on a trajectory to take more than a century to reach the average LE that peer high-income countries had achieved by 2016.

Midlife mortality acceleration

All-cause mortality among adults 25–64 rose from $328.5$ to $348.2$ /100,000 ( $+6.0\%$ ) $2010$ – $2017$ . The $25$ – $34$ age group saw the worst increase: $102.9 \to 132.8$ /100,000 ( $+29.0\%$ ). Cause-specific increases had begun as early as the 1990s in some states but were masked until $\approx 2010$ by concurrent declines in ischemic heart disease, cancer, and HIV.

State LE divergence

State LE gaps widened to $7.0$ years by 2017 (Hawaii highest, Mississippi lowest). Divergence began in the 1980s, with substantial jumps in the 1990s coinciding with neoliberal devolution of federal social policy authority to states. Adjacent state pairs diverged sharply: Colorado/Kansas gap grew from $0.3$ to $1.5$ years; Alabama/Georgia from $0.1$ to $2.3$ years. State policy choices (minimum wage, Earned Income Tax Credit (EITC), Medicaid expansion, gun laws) are candidate explanations, though the paper does not establish causation. This divergence is methodologically important for DI mortality research: the "U.S. general population" is not a homogeneous mortality benchmark — it is increasingly a composite of rapidly diverging state populations.

Olshansky et al. (2005): Obesity and the Stalling of Life Expectancy

Olshansky et al. (2005) provide two empirical contributions directly relevant to period mortality methodology: a life-table counterfactual quantifying obesity's current burden on US life expectancy, and an audit of SSA's historical LE forecasting record.

Life-table counterfactual

Linking National Health and Nutrition Examination Survey (NHANES) III obesity prevalence data to body mass index (BMI)-specific mortality tables, the authors estimate that if all obese Americans (BMI $\geq 30$ ) acquired the mortality risk of individuals with BMI $= 24$ , US period life expectancy at birth would increase by:

$0.33$ – $0.93$ years (white males)
$0.30$ – $0.81$ years (white females)
$0.30$ – $1.08$ years (black males)
$0.21$ – $0.73$ years (black females)

At current prevalence, the obesity LE penalty exceeds the combined burden of all accidental deaths (accidents, homicide, suicide). Looking forward, if obese youth cohorts carry elevated risk into middle and old age without medical offset, the population-level LE reduction could rise to $2$ – $5$ years — shifting obesity from a headwind on LE growth to an active reversal.

SSA historical forecasting record

Olshansky et al. document SSA's age-65 LE forecasts from published actuarial studies spanning $1952$ – $2003$ . The pattern: SSA consistently underestimated pre-1980 LE gains (assuming rapid improvement could not be sustained), then switched to optimistic extrapolation after 1980 — precisely when LE at age 65 for US women began to stall. This is one of three documented sources of SSA's systematic post-1980 optimistic bias; see Lee (2003) and Soneji-King (2012) for the other two. See SSA Mortality Forecasting.

Prediction and partial confirmation

The paper warned that US LE would "soon come to an end" absent effective obesity interventions — a prediction realized when US LE peaked in 2014 and fell in 2015, 2016, and 2017. The mechanism differed: the actual decline was driven by deaths of despair (drug overdose, suicide, alcohol) rather than obesity. Medical management of obesity-related conditions (statins, antihypertensives, improved glycemic control) partially offset the obesity-mortality pathway in the intervening decade. The structural fragility Olshansky identified was correct; the proximate cause was not. See Deaths of Despair.

GBD 2021: US Rank Decline Through 2021

GBD 2021 (US Burden of Disease Collaborators 2024) extends the cross-national LE/healthy life expectancy (HALE) comparison through 2021, incorporating COVID-19 and using 204 countries as the comparison set.

National LE and HALE levels

US period LE in 2021: $77.1$ years ( $74.3$ male, $80.0$ female), down from a pre-COVID peak of $79.1$ years in 2019. US HALE in 2021: $64.4$ years ( $63.2$ male, $65.7$ female), down from a peak of $\approx 66.7$ years in 2010 — meaning the healthy years component of LE stagnated earlier than total LE and declined more sharply.

Rank decline: 1990–2021

Metric	1990 rank	2021 rank	Direction
LE, males	$35$ th of $204$	$46$ th	↓ $11$
LE, females	$19$ th of $204$	$47$ th	↓ $28$
HALE, males	$42$ nd of $204$	$69$ th	↓ $27$
HALE, females	$32$ nd of $204$	$76$ th	↓ $44$

HALE rank declined faster than LE rank — confirming that morbidity worsened at a rate exceeding mortality change. The female HALE decline ( $32$ nd $\to$ $76$ th) is the starkest single number in the paper.

State-level rank comparisons

Best-performing state (Hawaii): LE rank fell from $6$ th/ $4$ th globally (1990) to $28$ th/ $22$ nd (2021); HALE rank from $14$ th/ $5$ th to $39$ th/ $34$ th. Even the nation's healthiest state deteriorated relative to peers.

Worst-performing state (West Virginia for HALE): ranked $141$ st for males, $137$ th for females in HALE among all 203 other countries and territories — worse than the vast majority of the world. WV LE was $71.9$ years (not disaggregated by sex), down $2.7$ years from 1990.

14 US states lost life expectancy 1990–2021. States with the largest LE losses:

West Virginia: $-2.7$ years
Oklahoma: $-1.8$ years
Mississippi: $-1.5$ years
Kentucky: $-1.4$ years
New Mexico: $-1.2$ years
Arkansas: $-1.0$ years

Washington, DC gained the most LE ( $+9.7$ years), driven by ischemic heart disease (IHD) decline and neonatal improvements.

COVID-19 contribution

COVID-19 deaths were the largest single cause of LE decline across states in 2021, with Arizona losing $2.8$ LE years to COVID and Hawaii only $0.8$ years. This geographic heterogeneity reflects pre-existing COVID-19 vaccination rates, underlying comorbidities, and health system capacity differences. The COVID shock sits on top of a pre-existing LE deterioration: the US was already on a declining trajectory vs. peers before 2020.

Historical Age Distribution of US Mortality Decline (Cutler and Meara 2001)

Cutler and Meara (2001) decompose the $\approx 67\%$ fall in age-adjusted US mortality from 1900 to 1998 (life expectancy rising from $\approx 48$ to $\approx 80$ years; crude death rate falling from $1$ -in- $42$ to $1$ -in- $125$ Americans/year) into three mechanistically distinct eras using contributions to life expectancy at birth by age group (Table 1):

Era	Total LE gain	Dominant age group	Mechanism
$1900$ – $40$	$+15.9$ yrs	Infants ( $+4.7$ ), children $1$ – $14$ ( $+4.4$ ), young adults $15$ – $34$ ( $+3.6$ )	Public health infrastructure, improved nutrition
$1940$ – $60$	$+6.4$ yrs	Young adults ( $+1.6$ ), infants ( $+1.4$ ), elderly begin ( $+1.2$ )	Sulfa drugs (late 1930s) and penicillin (late 1940s)
$1960$ – $90$	$+5.1$ yrs	Elderly ( $+1.8$ ), older adults $35$ – $64$ ( $+1.4$ )	Cardiovascular treatment: statins, antihypertensives, coronary care units (CCUs), coronary artery bypass grafting (CABG)

Annual rates of decline for elderly mortality: $0.3\%$ /year (Era 1) $\to$ $1.1\%$ /year (Era 2) $\to$ $1.1\%$ /year (Era 3). The antibiotic era inaugurated fast elderly improvement; the cardiovascular disease (CVD) era sustained it.

Antibiotic identification. Cutler and Meara isolate the antibiotic contribution via difference-in-differences (DiD): mortality from pneumonia and influenza (treatable by sulfa drugs and penicillin) vs. dysentery and diarrheal diseases (not treatable) before and after drug introduction. Net DiD effect: approximately $-3.8\%$ /year attributable to antibiotics (Table 4).

CVD dominance. CVD accounted for $\approx 98\%$ of the 1960–90 mortality reduction. CVD death rates fell $\approx 2\%$ /year from their mid-1960s peak, with roughly two-thirds of the cumulative decline by 1995. Behavioral factors (especially smoking cessation) contributed but the authors argue medical care is the primary driver based on timing: CVD decline began before widespread exercise and dietary change.

The "medicalization of death." Post-1960, medical care displaced public health and nutrition as the primary mechanism of mortality decline — a qualitatively new regime requiring ongoing healthcare delivery and expenditure rather than one-time infrastructure. This shift helps explain why US healthcare spending on the elderly grew so dramatically in the second half of the 20th century. See Medicalization of Death.

Connection to the $b_x$ tilt. The three-era shift — from infant/child dominance to elderly dominance — is the historical realization of the $b_x$ tilt described by Lee (2003). As elderly mortality became the locus of improvement, the age-specific sensitivity weights $b_x$ in the Lee-Carter model tilted toward older ages, generating more life-years per unit of percentage improvement and sustaining linear $e_0$ growth. See below and Lee-Carter Model.

Linear Life Expectancy Trend and the $b_x$ Tilt (Lee 2003)

Period life expectancy in high-income countries has grown at a remarkably stable linear rate for most of the 20th century — roughly $+0.2$ years/year in leading nations and $+0.144$ years/year in the United States under Lee-Carter extrapolation. This poses an analytical puzzle: if age-specific death rates decline at a constant percentage rate (as the Lee-Carter $k_t$ structure implies), and if mortality is already very low at young ages, diminishing returns should eventually cause $e_0$ gains to decelerate. Yet the linear trend has persisted.

Ronald Lee (2003) resolves this paradox through the $b_x$ tilt mechanism. In the Lee-Carter (LC) model, each age group's sensitivity to the overall mortality trend is captured by $b_x$ — the age-specific loading on the latent factor $k_t$ . Over the 20th century, the pattern of mortality improvement shifted:

Early 20th century: Improvement concentrated at young ages (infectious disease control, infant and child mortality declines). Young-age $b_x$ weights were large.
Mid–late 20th century: Improvement concentrated at older ages (cardiovascular treatment, cancer survival). Old-age $b_x$ weights grew.

Vaupel (1986) demonstrated that the life-years contribution of a given percentage decline in mortality at age $x$ is proportional to $l_x \times e_x$ (survival to that age times remaining life expectancy) — a quantity that is substantially larger for older adults than the corresponding calculation at young ages, once baseline infant and child mortality is already low. As $b_x$ tilted toward older ages, each percentage point of mortality improvement at the new dominant ages generated more $e_0$ gain per unit improvement — offsetting the diminishing-returns effect from young-age improvement approaching zero. The net result: near-linear $e_0$ growth despite sub-linear implied gains from any single age group's rate trajectory.

Cross-National $b_x$ Universality (Tuljapurkar, Li and Boe 2000)

Tuljapurkar, Li, and Boe (2000) established that the $b_x$ profile and the single-factor LC structure are not US-specific but cross-national empirical regularities. Applying the LC SVD decomposition to all G7 countries (1950–1994), they found that in every country the first singular value explains $>94\%$ of temporal variance in log death rates, and $k_t$ declines linearly. The $b_x$ profile is broadly similar across all seven — higher at younger ages, lower at old ages — confirming the $b_x$ tilt as a universal feature of 20th-century mortality change, not a statistical accident of US data. Japan stands out for having the most accurate old-age data (no lumping at ages 85+) and the fastest $k_t$ drift ( $z = 0.79$ /yr vs. $0.26$ /yr for the US). See Lee-Carter Model.

The Oeppen-Vaupel Best-Practice Record

Oeppen and Vaupel (2002) compiled the best-practice female life expectancy (the highest nationally observed in any given year) from 1840 to 2000. The result: a near-linear trend at $+0.24$ years/year with no sign of deceleration, directly contradicting repeated claims of approaching biological limits to longevity. The record has been held by different nations in different periods (Sweden in the 19th century, New Zealand and Norway in the mid-20th century, then Japan from the 1960s onward), suggesting that each nation's trajectory represents a local realization of a common underlying trend, not a global ceiling.

The practical implication: the best-practice trend is a legitimate upper-bound scenario for national forecasting. Nations below the frontier converge toward it at a rate that, when estimated, implies US life expectancy in 2030 approximately $\approx 3.8$ years above SSA's 2002 intermediate projection. See SSA Mortality Forecasting.

The Broken Limits Saga

The same paper that documented the best-practice trend documented its complement: a catalogue of expert predictions that the trend was approaching a ceiling. Oeppen and Vaupel (2002) compiled asserted upper bounds on life expectancy from Dublin (1928) through Olshansky et al. (2001) and plotted them against the best-practice line:

Author(s)	Year	Asserted ceiling
Dublin	1928	$64.75$ yr (for both sexes)
Dublin & Lotka	1936	revised slightly upward
Bourgeois-Pichat	1952	$\approx 77$ yr
Siegel	1980	$\approx 79$ yr
Coale	1996	$\approx 85$ yr
Coale & Guo	1989	$\approx 85$ yr
Olshansky, Carnes, Cassel	1990	$35$ yr remaining at age $50$ ( $\approx 85$ yr)
World Bank, UN	multiple	various, all exceeded
Olshansky, Carnes, Désesquelles	2001	$\approx 85$ yr

Every ceiling from Dublin (1928) through Olshansky et al. (1990) was broken, on average 5 years after publication. Dublin's 1928 ceiling was already exceeded at the time of publication: New Zealand non-Maori women had $e_0 = 65.93$ in 1921. Olshansky et al.'s 1990 assertion that life expectancy "should not exceed … 35 years at age 50 unless major breakthroughs occur in controlling the fundamental rate of aging" was surpassed by Japanese women in 1996.

The Oeppen-Vaupel analysis adds three further findings beyond documenting the trend: (1) expert ceiling predictions have been repeatedly proven wrong; (2) apparent leveling off of life expectancy in specific countries is an artifact of laggards catching up and leaders falling behind — no individual country holding the record shows plateauing at the time it held the record; (3) if life expectancy were approaching a maximum, the record growth rate should be slowing, but it has been steady at $1/4$ year/year for 160 years.

International Convergence Toward Best Practice

Lee (2003) estimates convergence models for a panel of high-income countries. National $e_0$ converges toward the best-practice frontier at rate $\alpha \approx 0.06$ – $0.08$ per year (implying a half-life of convergence of approximately 10 years). A slightly positive gap $^2$ coefficient suggests the convergence is slightly accelerating — nations far from the frontier close the gap faster than those already near it. The convergence model provides an external anchor grounded in the cross-national distribution of mortality outcomes rather than extrapolation of a single country's time series.

Stochastic Model Bias at Old Ages (Meseguer 2008)

Meseguer (2008) provides direct empirical evidence that standard stochastic mortality models fail precisely where the $b_x$ tilt places the highest weights: ages $65$ – $95+$ . Using Human Mortality Database (HMD) data for 16 countries with a 1980 jump-off year, both the bias-corrected Lee-Carter model and the $\text{AR}(1)$ model of Denton, Feaver, and Spencer (2005) systematically overestimate mortality at old ages — meaning they underestimate the pace of improvement at the ages that now dominate $e_0$ growth. Retirement ages ( $65$ – $95+$ ) account for $97$ – $99\%$ of total age-profile mean squared error across all countries and forecast horizons.

The practical consequence is that $e_0$ forecasts are systematically biased downward not because of the trend rate assumption (the gap Lee (2003) documents) but because of the age-specific covariance structure. This is a distinct and complementary failure mode. Moreover, the LC model's interval forecasts collapse to near-zero empirical coverage for dependency ratios at long horizons, while $e_0$ coverage remains $\approx 100\%$ — demonstrating that $e_0$ aggregation masks age-specific interval failure through cancellation of errors across ages. For Old-Age, Survivors, and Disability Insurance (OASDI) solvency purposes, dependency ratios (not $e_0$ ) are the correct validation metric. See SSA Mortality Forecasting.

Rau et al. (2006): 10 Years After Kannisto — Continued and Accelerating Old-Age Improvement

Rau, Jasilionis, Jdanov, and Vaupel (2006) updated Kannisto et al.'s (1994) 27-country study of advanced-age mortality using the Kannisto-Thatcher Database (KTDB) through 2000 — a decade extension that produced three major findings: improvement rates accelerated monotonically, no biological minimum was detectable, and the US emerged as a prominent stagnation outlier.

Data and Method

The KTDB collects population and death counts for ages 80+ by sex, birth year, age, and calendar year via Lexis Triangles, maintained at the Max Planck Institute for Demographic Research. The primary metric is the annual improvement rate $\rho = -\left((\bar{m}_2/\bar{m}_1)^{1/\delta} - 1\right)$ , where $\bar{m}_1$ and $\bar{m}_2$ are age-standardized average death rates in successive decades, $\delta$ is the interval between midpoints (10 years), and the standard population is Sweden $1950$ – $1990$ (to ensure stable weights across countries with different age compositions within the 80–89 group).

Population Counts: Every Country, Every Age Group (Table 1)

Across 27 countries, the 80+ population grew from $8.9$ million to $29.7$ million ( $+3.02\%$ /year on average, $1960$ – $2000$ ). The centenarian (100+) population grew from $8{,}181$ to $93{,}307$ — an $11$ -fold increase at $+6.09\%$ /year. Japan's centenarians grew at $+10.76\%$ /year, the fastest of any country-group combination. Critically, every single country in the 27-country sample showed an increase in both its 80+ and its 100+ population — not a single exception. This directly and comprehensively refutes Fries's (1980) prediction that "the number of very old persons will not increase" and that rectangularization of the survival curve was approaching completion. See Compression of Morbidity.

Accelerating Improvement: The Table 2 Results (19 Countries)

For the 19 countries with sufficient data across all four decade pairs (1950s–1960s, 1960s–1970s, 1970s–1980s, 1980s–1990s), annual improvement rates $\rho$ for ages 80–89 were monotonically accelerating in both sexes:

Decade pair	Males 80–89	Females 80–89
1950s $\to$ 1960s	$0.81\%$ /yr	$0.91\%$ /yr
1960s $\to$ 1970s	$0.89\%$ /yr	$1.35\%$ /yr
1970s $\to$ 1980s	$1.20\%$ /yr	$2.00\%$ /yr
1980s $\to$ 1990s	$1.88\%$ /yr	$2.45\%$ /yr

Female improvement roughly tripled over the full period; male improvement more than doubled. No country in the 19-country sample showed deceleration at ages 80–89 across the full period. This directly extends Kannisto et al.'s (1994) finding — which covered through approximately 1990 — by documenting that the acceleration not only continued but intensified.

No Biological Minimum: The Japan Paradox

The clearest test of a biological-floor hypothesis is the cross-country relationship between current death rates and improvement rates. If countries were approaching a minimum, those with the lowest current rates would show the slowest improvement (least room left). The data show the opposite: Japan, with the lowest old-age death rates of any country, simultaneously shows among the fastest improvement rates. A scatter of 1980s standardized death rates (x-axis) against 1990s improvement rates (y-axis) has no positive slope — it is flat or slightly negative. The biological-limit argument fails its most direct empirical test.

This result also implies that the best-practice nation (Japan) is not converging toward a plateau; it is pulling away from the rest. The Oeppen-Vaupel (2002) best-practice trend applies not only to period $e_0$ at all ages but specifically to the 80+ age group that drives the $b_x$ tilt mechanism.

The US as Outlier

The US held the world's lowest old-age death rates through approximately the mid-1980s — it was the global mortality leader. Since then, improvement has stagnated while virtually all other high-income countries continued to improve. Of the 27 countries in the study, $22$ showed acceleration in the 1990s relative to the 1980s; the US was among the $5$ that did not. Rau et al. explicitly note that the US stagnation "is not yet understood." This finding independently corroborates Wilmoth's (2005) GENUS paper, which used HMD data for $15$ high-income countries and found that $11$ of $15$ showed accelerating old-age improvement while the US was one of only $4$ deviants. See SSA Mortality Forecasting.

The SSA forecasting implication is direct: OACT's practice of anchoring long-run mortality improvement projections to the recent US-specific trend (rather than the historical or cross-national trend) codifies an anomalous and unexplained national experience as the baseline. The two studies together make the same methodological argument with independent data.

East Germany: Largest Country-Level Catch-Up

East Germany showed the largest single-country improvement jump in the dataset, concentrated in the 1980s→1990s decade pair following reunification. Women's death rates fell at $>3\%$ /year; men's rates improved at $\approx 2.5\%$ /year. This is consistent with rapid convergence in health system access, pharmaceutical availability, and living standards after 1990. East Germany's trajectory illustrates that stagnating improvement (as seen in the pre-1990 Eastern Bloc) can be reversed rapidly once systemic barriers are removed.

Age Gradient and $b_x$ Connection

Improvement rates at ages $80$ – $84$ consistently exceed those at $95$ – $99$ , which exceed those at 100+. This declining improvement gradient with age does not indicate a biological floor — it mirrors the $b_x$ loading profile in the Lee-Carter model, where older ages have lower loadings (i.e., historically lower proportional improvement). Crucially, improvement rates at every age group rose across successive decades; there is no crossover where very old ages cease improving. For the four best-performing countries, women show acceleration across the entire 80–100+ range; men show no regression at the oldest ages. The gradient is a structural feature of 20th-century mortality decline, not evidence of an approaching ceiling. See Lee-Carter Model.

Age-Adjusted Death Rates and the 1940/2000 Standard Break (Anderson and Rosenberg 1998)

The age-adjusted death rate (AADR) — the summary statistic used for nearly all US mortality surveillance — is not a uniquely defined number. It depends on which standard population is used as the weighting scheme. The US switched from the 1940 standard population to the 2000 standard population for all national vital statistics, effective with data year 1999.

Magnitude of the level change: Under the 1940 standard, the US all-cause AADR in 1995 was $503.9$ /100,000. Under the 2000 standard, the same 1995 data yield $918.5$ /100,000 — an $82\%$ increase. The crude rate was $880.0$ . The doubling reflects the aging of the standard: the 1940 standard weights ages 65+ at $6.8\%$ of the population; the 2000 standard weights ages 65+ at $12.4\%$ . Because mortality rises steeply with age, using an older standard produces a higher aggregate rate.

Trend attenuation: Both standards show falling mortality 1979–1995, but the implied improvement is smaller under the 2000 standard ( $-9.2\%$ ) than the 1940 standard ( $-12.7\%$ ). The 2000 standard places greater weight on older ages where percentage improvements were smaller in this period, attenuating the apparent improvement rate.

Cause-specific amplification: Causes of death concentrated at older ages are most affected. Stroke AADR in 1995: $26.7$ (1940 standard) $\to$ $63.9$ (2000 standard)/100,000 — a $2.4\times$ increase. Causes primarily affecting younger people change proportionally less.

Life expectancy immunity: Period life expectancy at birth ( $e_0$ ) uses its own internal life-table stationary population as weights, not an external standard. It is unaffected by the 1940/2000 switch and remains consistent across the series break.

Series incomparability: Pre-1999 CDC-published AADR series used the 1940 standard; post-1999 series use the 2000 standard. The two series are not directly comparable. Any trend analysis spanning the 1999 break must apply one standard consistently across all years or flag the break explicitly. See Age Standardization for methodological detail on direct vs. indirect standardization and gamma-distribution confidence intervals.

Period Mortality in Social Security (SS) Solvency Forecasting (Soneji and King 2012)

Period mortality projections are also the central input to OASDI trust fund solvency calculations. The SSA Office of the Chief Actuary produces 75-year age-and-sex-specific mortality forecasts annually, using a process based on linear extrapolation of historical cause-specific death rates plus subjective selection of 70 "ultimate rates of decline" approved by OASDI trustees (political appointees). Soneji and King (2012) showed this process is undocumented, unreplicable, and systematically overoptimistic about solvency relative to a formal Bayesian alternative.

The key finding connects period mortality methodology to SS finance: SSA's intermediate-cost scenario produces expected ages at death that are lower than Soneji/King's model — SSA assumes more people die earlier than the Bayesian model predicts. Because people living longer means more years of benefit payments, SSA's intermediate solvency projection is too optimistic by $\$730$ billion in trust fund balance by 2031. The mechanism runs through declining smoking: as smoking rates fall, the cohort of longer-lived non-smokers expands, raising future benefit costs by an amount SSA's qualitative forecast fails to capture. See SSA Mortality Forecasting and Office of the Chief Actuary.

How It Works

For DI beneficiaries, SSA constructs mortality tables using a multiple-decrement framework: beneficiaries can exit via death, recovery, or conversion to retirement benefits. The probability of death, $q_{[x]+t}$ , represents the conditional probability of dying in year $t+1$ given entitlement at age $x$ and survival to year $t$ .

Important limitation: Claimants who die during the 5-month DI waiting period are never observed. This means year-0 mortality estimates are slightly understated.

Why It Matters

Period estimates are used for program cost projections and benefit design.
Because period life expectancy discounts cohort improvements, it understates the true generational gaps in longevity — particularly relevant given widening socioeconomic mortality disparities.
Comparing DI beneficiary period mortality to historical general-population mortality provides intuitive benchmarks (e.g., "equivalent to the general population in 1900").

Period vs. Cohort: The Wilmoth (2005) Taxonomy

Wilmoth (2005) provides the most systematic formal treatment of the relationship between period and cohort mortality, including a taxonomy of six mean-lifespan measures that resolves a prominent methodological controversy.

The period/cohort distinction. Period life expectancy $e_0$ uses age-specific death rates observed at a single calendar time $t$ , applied to a hypothetical synthetic cohort. Cohort life expectancy follows an actual birth cohort through its entire life. The two measures answer different questions — neither is "correct" in the abstract — but they diverge systematically under improving mortality because period survival conditions on current (favorable) rates while cohort survival conditions on past (higher) rates.

Period-to-cohort speed conversion. If period $e_0$ improves at rate $r$ per calendar year, cohort $e_0$ improves at $s = r/(1-r)$ per birth year. Under the linear shift model — in which each percentile of the age-at-death distribution shifts linearly in time — this conversion is exact. For $r = 0.2$ years/year (a plausible estimate for leading countries), $s = 0.25$ years/birth-year — $25\%$ faster improvement from the cohort perspective. See Linear Shift Model and Period vs. Cohort Life Expectancy.

CAL and the tempo-adjustment controversy. Bongaarts and Feeney (2002, 2003) proposed that period $e_0$ is "tempo-distorted" upward when mortality is declining, because deaths are shifted to later periods. Wilmoth (2005) rejects this: $e_0$ is a timing measure (expected mean age at death), not a count measure, so no analogous distortion to the total fertility rate (TFR) tempo effect applies. Their "generalized tempo-adjusted $e_0$ " is mathematically identical to CAL (cross-sectional average length of life, independently derived by Brouard 1986 and Guillot 2003) — a valid population-dynamics measure, not a bias correction. CAL is necessarily lower than period $e_0$ under improving mortality (reflecting the cohort survival shortfall), but this gap is interpretable, not artifactual. See Cross-Sectional Average Length of Life (CAL) and Tempo Effects in Mortality.

Swedish validation. Wilmoth fits the linear shift model to Sweden 1751–2003 (HMD data): period $e_0$ grows approximately linearly at $\bar{r} \approx 0.17$ years/year over the long run. CAL tracks period $e_0$ with a 2–4 year lag — empirically confirming that the gap is sustained and stable under continued improvement.

Sex Mortality Differences and the Cohort Smoking Etiology

Preston and Wang (2006) establish that U.S. sex differences in all-cause mortality from 1948 to 2003 follow a cohort diagonal rather than a period pattern: the male/female death rate ratio is best explained by birth cohort, not by calendar year. The peak occurs for cohorts born approximately 1903–1908, tracking diagonally through the Lexis diagram as those cohorts age into the mortality-relevant years.

Three Lines of Evidence for Smoking Attribution

Timing: The cohort pattern in sex mortality differences precedes the rise of lung cancer mortality and mirrors the period when male smoking prevalence surged far above female prevalence.
Cause specificity: The male excess is concentrated in causes associated with smoking (lung cancer, cardiovascular disease, chronic obstructive pulmonary disease (COPD)) and absent in non-smoking-related causes.
Within-sex convergence: Birth cohorts where women approached men's smoking rates show narrowing sex differentials on the expected schedule.

Age-Period-Cohort (APC) Regression Results

Using negative binomial regression with age, period, and cohort fixed effects and mean years smoked before age 40 as the cohort covariate:

Parameter	Estimate	SE	Interpretation
Smoking (male)	$0.0230$	$0.0022$	$+2.33\%$ /year of pre-40 smoking raises male/female death ratio
Sex $\times$ Smoking interaction	$-0.0100$	—	Women face $\approx 57\%$ of men's per-year smoking mortality risk
Sex ratio amplification, $1900$ – $1904$ cohort	$+41\%$	—	Peak cohort smoking differential
Sex ratio amplification, $1945$ – $1949$ cohort	$+18\%$	—	Later cohort, smaller differential

The $23$ -percentage-point narrowing across cohorts is attributed to declining male-to-female smoking ratios among later birth cohorts — a behavioral convergence that mechanically reduces the sex mortality gap as those cohorts age.

Period Distortion from Smoking

Without controlling for smoking, the estimated all-cause mortality decline over 1948–2003 is $48\%$ . Controlling for cohort smoking histories, the estimate rises to $56\%$ — a difference of $8$ percentage points representing genuine underlying improvement that was masked by men's heavier smoking burden. Any period-based mortality analysis omitting a smoking covariate will systematically underestimate the pace of baseline improvement. See SSA Mortality Forecasting.

Forecasting Implications

As younger cohorts with lower male-to-female smoking ratios enter the high-mortality age range (65+), men aged 50–85 are projected to see survival probabilities rise $\approx 23\%$ from cohort smoking decline alone by 2030; women only $\approx 2\%$ . This asymmetric gain drives further sex gap convergence independently of any additional behavioral change. Period-based mortality projections that ignore cohort smoking histories will overstate future sex mortality differences and understate the pace of male improvement.