Life Tables

How a year's death rates by age become probabilities of dying, survivors, person-years and a life expectancy, with every step derived. The probability of dying derived from the death rate in general and under uniform deaths and a constant hazard, with the gap between the two bounded; the columns of the table and life expectancy proved to be the mean remaining lifetime; the open last interval and why one rate for it misleads; period and cohort tables; and the standard error of a life expectancy by the delta method, Chiang's formula with the last interval added. The 2019 U.S. table is rebuilt from CDC WONDER and compared with the published NCHS table, and the difference of half a year in life expectancy at birth is traced to its sources. Worked problems on a ten-year interval, a three-row table by hand and the precision of life expectancy at 85.

In 2019, 2,854,838 residents of the United States died. The National Center for Health Statistics (NCHS) turned that year's deaths and population counts into a single number: a life expectancy at birth of 78.8 years (Arias and Xu, 2022). No one followed a group of babies for 78.8 years to find it, and no baby born in 2019 is promised it. How does one year's set of death rates become a number of years of life, what exactly does that number mean, and how far can it be trusted?

A life table answers the first question. It takes a death rate for each age, converts each rate into a probability of dying within the age interval, and follows an imaginary group of 100,000 newborns through those probabilities, counting how many survive to each age and how many years they live in total. Life expectancy is the total years lived divided by the number of people. Every step needs an assumption about how deaths fall within an interval, and the published table adds several more about the oldest ages. This article derives the steps, rebuilds the 2019 U.S. table from public data, and traces where it differs from the published one. It is the unit's home for the conversion of rates to probabilities and for the life table functions; Censoring, Survival Functions and Hazards generalizes them to the survival function and the hazard of any lifetime.

Here is the route. First the notation, and the difference between a period table and a cohort table. Then the conversion from a death rate to a probability of dying, in general and under the two standard assumptions, with a bound on how far apart they can be. Then the columns of the table, with life expectancy proved to be a mean remaining lifetime, and the open last interval. Then the 2019 table rebuilt from CDC WONDER, its comparison with the NCHS table and a decomposition of the difference. Then abridged tables with wide intervals. Finally the precision of a life expectancy, derived by the delta method and checked by simulation. The article assumes Rates in Public Health (person-time, and the risk a constant rate implies) and Crude Rates and Their Precision (the Poisson model for a count of deaths); the precision section uses The Delta Method.

Notation and the scope of the proofs

Ages are in years. An age interval $[x, x+n)$ starts at exact age $x$ and has length $n$; the left subscript $n$ in ${}_nq_x$ names the length, and is dropped when $n = 1$. The columns of a life table are:

  • ${}_nm_x$, the death rate of the interval: deaths per person-year lived in it.
  • ${}_nq_x$, the probability of dying in the interval for a person alive at exact age $x$, and ${}_np_x = 1 - {}_nq_x$.
  • $l_x$, the number alive at exact age $x$ out of a radix of $l_0 = 100{,}000$ births.
  • ${}_nd_x = l_x - l_{x+n}$, the number dying in the interval.
  • ${}_nL_x$, the person-years lived in the interval by the table's population.
  • ${}_na_x$, the average number of years lived in the interval by those who die in it.
  • $T_x = \sum_{y \ge x} {}_nL_y$, the person-years lived beyond age $x$, and $e_x = T_x/l_x$, the life expectancy at age $x$.

The letter $T$ has two uses in this unit. In this article $T_x$ is the demographers' column of person-years; in the rest of the unit $T$ is a survival time. Where both appear, the survival time is called the age at death.

Read as a function of exact age, $l(x)/l_0$ is the survival function of the age at death in the table's population: the probability that a newborn is alive at age $x$. Its hazard, $\mu(x) = -l'(x)/l(x)$, is called the force of mortality by demographers; Censoring, Survival Functions and Hazards defines the hazard for any lifetime and proves the relation $l(x)/l_0 = \exp(-\int_0^x \mu(u)\,du)$. The proofs here use only calculus on $l$, assumed continuous and differentiable inside each interval where a derivative appears.

A period table and a cohort table

A cohort life table follows people born in one year through their actual lives: the rate at age 50 is the rate this cohort experienced fifty years later. A period life table applies the age-specific rates of one calendar year to a hypothetical cohort, as if a newborn would meet at age 50 the death rate that 50-year-olds had in that year. NCHS's report describes the period table as a "snapshot" of current mortality, and its annual U.S. tables are period tables (Arias and Xu, 2022, Introduction). A complete cohort table needs a century of data, and for any cohort with members still alive it needs rates that have not happened yet.

Example (what 78.8 years does not mean). A common reading is that babies born in 2019 will live 78.8 years on average. That holds only if every age-specific rate stays at its 2019 level for the next century. Suppose instead that every rate falls by 1 percent a year, so that the cohort born in 2019 meets at age $x$ the 2019 rate at $x$ times $0.99^x$. The period life expectancy computed below from the 2019 rates is 79.3 years; the same calculation along the cohort's diagonal gives 91.1 years (both from the table rebuilt later in this article). In the cohort calculation the open group's rate, like every other, is the 2019 rate times $0.99^{85}$, so a cohort member reaching 85 expects 17.8 more years; that one assumption carries much of the difference, and with the open group's 2019 rate kept the cohort figure would be 84.3. Falling rates make the period figure an underestimate of a real cohort's lifetime, and rising rates make it an overestimate. A period life expectancy is a summary of one year's rates on a scale of years, which is how it is used: to compare years and populations without the distortions of age structure that Direct Standardization removes by another route.

From a death rate to a probability of dying

Vital statistics supply deaths $D$ and a population $P$ for each age interval, and their ratio is the age-specific death rate $M = D/P$. Rates in Public Health shows that the midyear population is an estimate of the person-years lived in the year, so $M$ is a rate per person-year. The life table needs a probability: of those alive at exact age $x$, what share die before $x + n$? The two differ because the people who die in the interval are not exposed for the whole interval.

The table's own rate is defined in the same way as the observed one, deaths over person-years:

$${}_nm_x = \frac{{}_nd_x}{{}_nL_x}.$$

A life table built from observed rates sets ${}_nm_x = M$ for each interval. The question is how to get ${}_nq_x$ from ${}_nm_x$.

Here is the idea. The person-years in the interval come from two groups: the survivors, who live all $n$ years of it, and those who die, who live ${}_na_x$ years on average. Writing that down links the rate to the probability.

Theorem 1 (the general conversion). For an interval of length $n$,

$${}_nL_x = n\,l_{x+n} + {}_na_x\,{}_nd_x, \qquad\text{and}\qquad {}_nq_x = \frac{n\,{}_nm_x}{1 + (n - {}_na_x)\,{}_nm_x}. \tag{1}$$

Proof. Of the $l_x$ people alive at $x$, $l_{x+n}$ survive the interval and each lives $n$ years in it; the ${}_nd_x$ who die live ${}_na_x$ years on average, by the definition of ${}_na_x$. Adding gives the first identity. For the second, write $d = {}_nd_x$, $q = d/l_x$ and $m = {}_nm_x$, and use $l_{x+n} = l_x - d$:

$$m = \frac{d}{n(l_x - d) + a\,d} = \frac{q}{n(1 - q) + a\,q}.$$

Solving for $q$: $m\,n - m\,n\,q + m\,a\,q = q$, so $q\,(1 + n m - a m) = n m$, which is (1). $\square$

Formula (1) is exact. All the assumptions enter through ${}_na_x$, which is not observed in ordinary vital statistics. Chiang (1967, equation 7) writes the same relation with ${}_na_x$ expressed as a fraction of the interval, and records that it had been used to construct life tables; the NCHS writes the same conversion for single years (Arias and Xu, 2022, Technical Notes, equation 3). Two assumptions about the shape of $l$ inside the interval fix ${}_na_x$.

Theorem 2 (uniform deaths). If $l$ falls linearly on $[x, x+n]$, that is, if deaths are spread uniformly over the interval, then ${}_na_x = n/2$ and

$${}_nq_x = \frac{2n\,{}_nm_x}{2 + n\,{}_nm_x} = \frac{n\,{}_nm_x}{1 + \tfrac12 n\,{}_nm_x}. \tag{2}$$

The formula gives a probability below 1 exactly when $n\,{}_nm_x < 2$.

Proof. With $l(x + t) = l_x - t\,d/n$ for $0 \le t \le n$, each death happens at a time uniformly distributed over the interval, whose mean is $n/2$. Substituting ${}_na_x = n/2$ in (1) gives (2). The value $2x'/(2 + x')$ with $x' = n\,{}_nm_x > 0$ is below 1 exactly when $2x' < 2 + x'$, that is, $x' < 2$. $\square$

Theorem 3 (constant hazard). If the hazard $\mu(t) = -l'(t)/l(t)$ is constant, equal to $\mu$, on $[x, x+n)$, then ${}_nm_x = \mu$,

$${}_nq_x = 1 - e^{-n\,{}_nm_x}, \tag{3}$$

and ${}_na_x = 1/\mu - n e^{-n\mu}/(1 - e^{-n\mu})$, which is less than $n/2$.

Proof. A constant hazard gives $l(x + t) = l_x e^{-\mu t}$ for $0 \le t \le n$, by the argument of Rates in Public Health (the product $l(x+t)e^{\mu t}$ has derivative zero). Then ${}_nd_x = l_x(1 - e^{-n\mu})$ and

$${}_nL_x = \int_0^n l_x e^{-\mu t}\,dt = l_x\,\frac{1 - e^{-n\mu}}{\mu},$$

so ${}_nm_x = {}_nd_x/{}_nL_x = \mu$, and ${}_nq_x = 1 - e^{-n\mu}$ is (3). By the first identity of (1), ${}_na_x = ({}_nL_x - n\,l_{x+n})/{}_nd_x = 1/\mu - n e^{-n\mu}/(1 - e^{-n\mu})$. To compare it with $n/2$, put $z = n\mu > 0$: then ${}_na_x/n = 1/z - 1/(e^z - 1)$, and ${}_na_x < n/2$ is the inequality $1/(e^z - 1) > (2 - z)/(2z)$. For $z \ge 2$ the right side is at most 0 and the left side is positive. For $0 < z < 2$ it is equivalent to $2z > (2 - z)(e^z - 1)$, that is, to $e^z < (2 + z)/(2 - z)$, which Theorem 4 proves. $\square$

The two theorems answer one question with two assumptions. How far apart are the answers?

Theorem 4 (the two conversions compared). For $0 < z < 2$,

$$1 - e^{-z} < \frac{2z}{2 + z}, \qquad\text{and}\qquad \frac{2z}{2 + z} - (1 - e^{-z}) = \frac{z^3}{12} + O(z^4).$$

So for the same rate, uniform deaths always give the larger probability of dying, and the gap grows with the cube of $n\,{}_nm_x$.

Proof. The first inequality says $e^{-z} > (2 - z)/(2 + z)$, that is, $g(z) = \ln(2 + z) - \ln(2 - z) - z > 0$. Now $g(0) = 0$ and $g'(z) = \frac{1}{2+z} + \frac{1}{2-z} - 1 = \frac{4}{4 - z^2} - 1 = \frac{z^2}{4 - z^2} > 0$ on $(0, 2)$, so $g(z) > 0$ there. For the expansion, $\frac{2z}{2+z} = z\,(1 + z/2)^{-1} = z - \frac{z^2}{2} + \frac{z^3}{4} - \cdots$ and $1 - e^{-z} = z - \frac{z^2}{2} + \frac{z^3}{6} - \cdots$; the difference of the cubic terms is $\frac{z^3}{4} - \frac{z^3}{6} = \frac{z^3}{12}$. $\square$

For a single year of age the gap does not matter. At $z = 0.1$, a death rate of 100 per 1,000, the two probabilities differ by $7.6 \times 10^{-5}$. The rate at age 84 in 2019 was below 0.08 per person-year. For wide intervals the gap grows: at $z = 1$ the probabilities are 0.6667 and 0.6321. The NCHS uses uniform deaths for single years from age 1 to 99, so that $q_x = M_x/(1 + \tfrac12 M_x)$ and $L_x = \tfrac12(l_x + l_{x+1})$ (Arias and Xu, 2022, Technical Notes, equations 4, 5 and 17).

Which value of ${}_na_x$ is right? Neither assumption holds exactly, and the NCHS complete table shows by how much. Its single-year columns give, for any wider interval, ${}_na_x = ({}_nL_x - n\,l_{x+n})/{}_nd_x$ by (1). For the interval from 75 to 85 that is $a_{75} = 5.48$ years, above the uniform value of 5. A constant hazard at the observed 2019 rate of 0.04309 per person-year would give $a_{75} = 4.64$, below 5. Within those ten years the hazard more than doubles (the NCHS death rate at 84 is 2.59 times that at 75), so the deaths of the interval lean toward its end, while a constant hazard makes them lean toward its start, because it only shrinks the number of survivors. For the first year of life the NCHS uses a separation factor $f$, the share of a year's infant deaths that occur to infants born in the previous year, and sets $L_0 = f l_0 + (1 - f) l_1$ (Technical Notes, equation 18), which is (1) with $a_0 = f$; in 2019, $f = 0.127$ (Table III). The report's reason is that infant deaths are concentrated at the beginning of the first year of life rather than in its middle, so $a_0$ is far below $\tfrac12$. For other ten-year intervals the complete table gives $a_{15} = 5.96$ and $a_{45} = 5.64$.

Example (a misconception: the constant hazard is the realistic choice). It is tempting to prefer (3) because a constant hazard is a model of how deaths occur, while uniform deaths sound like a bookkeeping convention. Over a ten-year interval at older ages the evidence above points the other way: the true ${}_na_x$ is above $n/2$, uniform deaths put it at $n/2$, and the constant hazard puts it further below. The abridged tables later in this article measure the effect on life expectancy.

The columns of the table

With ${}_nq_x$ in hand, the table follows by arithmetic. Start from $l_0 = 100{,}000$ and set

$$l_{x+n} = l_x(1 - {}_nq_x), \qquad {}_nd_x = l_x - l_{x+n}, \qquad {}_nL_x = n\,l_{x+n} + {}_na_x\,{}_nd_x,$$

then sum the person-years from the top down, $T_x = {}_nL_x + T_{x+n}$, and divide, $e_x = T_x/l_x$. The NCHS report describes the same columns (Arias and Xu, 2022, Data and Methods): $T_x$ is "the total number of person-years that would be lived after the beginning of the age interval" by the table's cohort, and $e_x$ is $T_x/l_x$.

Why is $T_x/l_x$ an expectation? Here is the idea: $T_x$ adds up the time lived beyond $x$ by everyone in the table, and $l_x$ counts the people who share it, so the ratio is an average remaining lifetime. The tail formula for the mean makes this exact.

Theorem 5 (life expectancy is the mean remaining lifetime). Let the age at death $A$ of a member of the table's population have survival function $P(A > u) = l(u)/l_0$, and let $L$ and $T$ be computed from $l$ exactly, ${}_nL_x = \int_x^{x+n} l(u)\,du$ and $T_x = \int_x^\infty l(u)\,du$. Then, for $l(x) > 0$,

$$e_x = \frac{T_x}{l_x} = E[A - x \mid A > x].$$

Proof. Given $A > x$, the remaining lifetime $R = A - x$ is never negative and has $P(R > r \mid A > x) = P(A > x + r)/P(A > x) = l(x + r)/l(x)$. By the tail formula, which Expectation, Variance and Covariance takes as the definition of the mean of a variable that is never negative (and proves equal to the integral of $r$ times the density when there is one), $E[R \mid A > x] = \int_0^\infty l(x + r)/l(x)\,dr = T_x/l_x$. That ${}_nL_x$ is the integral of $l$ over the interval is the definition of person-years: at each instant $u$ the table has $l(u)$ people alive, each living $du$ years. $\square$

The function $x \mapsto E[A - x \mid A > x]$ is the mean residual life; Censoring, Survival Functions and Hazards proves that it determines the whole distribution. The table computes the integrals with the interval rule (1), which is exact under its assumption about ${}_na_x$.

The precision section needs the table written as a recursion. Using $T_x = {}_nL_x + T_{x+n}$, ${}_nL_x = n\,l_{x+n} + a\,d$ and $T_{x+n} = l_{x+n}e_{x+n}$, and dividing by $l_x$:

$$e_x = {}_np_x\,(n + e_{x+n}) + {}_nq_x\,{}_na_x. \tag{4}$$

A person alive at $x$ survives the interval with probability ${}_np_x$, and then lives its $n$ years and $e_{x+n}$ more; or dies in it and lives ${}_na_x$ years of it.

The open last interval

Data end in an open group, such as 85 and over in CDC WONDER's population counts. The last row of the table has $l_w$ people at its start, all of whom die in it, so ${}_\infty q_w = 1$, and it needs ${}_\infty L_w$ without an end to integrate to. The usual closing is ${}_\infty L_w = l_w/M_w$, so that $e_w = 1/M_w$. When is that right?

Theorem 6 (closing the table). (a) If the hazard is constant, $\mu$, at all ages from $w$ on, then $e_w = 1/\mu$, and the observed rate of the open group has expected value $\mu$. (b) In a stationary population, one whose number of people at each age $u \ge w$ is proportional to $l(u)$, the death rate of the open group is exactly $l_w/T_w$, so $1/M_w = e_w$ whatever the hazard. (c) If the hazard increases with age and the population of the open group is younger than the stationary one, in the sense that for every age $v \ge w$ its share aged $v$ and over is at most the stationary share, then $M_w \le l_w/T_w$, so $1/M_w$ is at least $e_w$; it is strictly larger when, at some age $v > w$ at which the hazard rises ($\mu_v > \mu_{v-1}$), the open group's share aged $v$ and over is below the stationary share.

Proof. (a) With $l(u) = l_w e^{-\mu(u - w)}$, $T_w = l_w/\mu$ and $e_w = 1/\mu$; every person-year in the group carries the same hazard, so expected deaths per person-year equal $\mu$. (b) In a stationary population with $c\,l(u)$ people at age $u$, the person-years per year at ages $u \ge w$ are $c\int_w^\infty l(u)\,du = c\,T_w$, and the deaths per year are $c\int_w^\infty l(u)\mu(u)\,du = c\int_w^\infty -l'(u)\,du = c\,l_w$. The rate is $l_w/T_w = 1/e_w$. (c) Work with single years $v = w, w+1, \ldots$ (the last one open). Let $\mu_v$ be the death rate at age $v$, increasing in $v$; let $\pi_v$ be the share of the open group's actual population at age $v$, and $s_v$ the share in the stationary population. Then $M_w = \sum_v \mu_v \pi_v$ and $l_w/T_w = \sum_v \mu_v s_v$. Write $\Pi_v = \sum_{u \ge v}\pi_u$ and $S_v = \sum_{u \ge v} s_u$, so $\Pi_w = S_w = 1$ and, by assumption, $\Pi_v \le S_v$. Summing by parts, with $\mu_{w-1} = 0$,

$$\sum_v \mu_v(\pi_v - s_v) = \sum_v (\mu_v - \mu_{v-1})(\Pi_v - S_v) \le 0,$$

because each $\mu_v - \mu_{v-1} \ge 0$ (the first is $\mu_w > 0$, multiplied by $\Pi_w - S_w = 0$) and each $\Pi_v - S_v \le 0$. So $M_w \le l_w/T_w = 1/e_w$. Every term of the sum is at most 0, so the sum is 0 exactly when each term is, that is, when at every $v > w$ either $\mu_v = \mu_{v-1}$ or $\Pi_v = S_v$; otherwise $M_w < l_w/T_w$. $\square$

A real population is stationary only by accident, since births and migration differ from year to year; part (c) says which way a younger open group pushes $1/M_w$. In (c) the actual population is assumed to have the table's death rate at each single age, which is how a period table is built. The 2019 table below measures how far $1/M_w$ can be off.

The 2019 table rebuilt from public data

CDC WONDER's Underlying Cause of Death database gives the deaths of U.S. residents in 2019 by single year of age, and the population by single year up to 84, with one figure, 6,604,958, for ages 85 and over. Of the 2,854,838 deaths, 147 had age not stated. The NCHS distributes such deaths over the ages in proportion, multiplying each age's count by $F = D/D_a$, the ratio of all deaths to those with age stated (Arias and Xu, 2022, Technical Notes, equation 1), and so do we: $F = 1.00005$. The table is then built as follows.

  • Ages 1 to 84: $q_x = M_x/(1 + \tfrac12 M_x)$, the NCHS formula, with $M_x$ the WONDER deaths (times $F$) over the WONDER population.
  • Age 0: the rate gives $m_0 = 0.005530$ and, with $a_0 = 1/2$, $q_0 = 0.005515$. Infants are better counted against births than against a midyear population, and the NCHS uses $q_0 = D_0(1 - f)/B_t + D_0 f/B_{t-1}$, with $B_t$ and $B_{t-1}$ the births of 2019 and 2018 (Technical Notes, equation 6). With the WONDER infant deaths, $f = 0.127$ and the 3,747,540 and 3,791,712 births of Table III of the report, this gives $q_0 = 0.005575$, the published value. We use it, with $a_0 = f$.
  • Ages 85 and over: one open interval with $M_{85} = 132.3$ deaths per 1,000 person-years (873,746 deaths, times $F$, over 6,604,958 people), closed by Theorem 6 with $e_{85} = 1/M_{85} = 7.56$ years.

The result is a life expectancy at birth of 79.32 years, against the 78.85 years of the NCHS table (78.848 at full precision, published as 78.8).

The 2019 table against the published one

Age $l_x$, rebuilt $l_x$, NCHS $e_x$, rebuilt $e_x$, NCHS
1 99,443 99,443 78.76 78.29
25 98,531 98,531 55.34 54.86
45 95,338 95,336 36.81 36.32
65 83,878 83,868 20.15 19.59
75 69,892 69,854 13.08 12.42
85 44,014 43,295 7.56 6.67

The survivors agree to within a few people until age 65 and then drift apart: by 85 the rebuilt table has 1.7 percent more survivors. The life expectancies differ by 0.47 to 0.51 years at every age up to 55, and the gap grows to 0.56 at 65 and 0.89 at 85. Because $e_x$ at every age includes the years lived after 85, a difference in the last row reaches every row above it.

Left: death rates by single year of age in 2019 on a log scale, from CDC WONDER (dashed) and from the NCHS table (solid, its deaths over its person-years). Notice that the two curves coincide until the 80s, and that WONDER's single rate for everyone aged 85 and over (the flat dashed segment) lies below the NCHS rates from age 89 on. Right: survivors in the rebuilt table divided by survivors in the NCHS table. The ratio stays at 1.000 until the mid-60s and climbs to 1.017 by age 85, inside the age range where the NCHS blends in Medicare data.
Left: death rates by single year of age in 2019 on a log scale, from CDC WONDER (dashed) and from the NCHS table (solid, its deaths over its person-years). Notice that the two curves coincide until the 80s, and that WONDER's single rate for everyone aged 85 and over (the flat dashed segment) lies below the NCHS rates from age 89 on. Right: survivors in the rebuilt table divided by survivors in the NCHS table. The ratio stays at 1.000 until the mid-60s and climbs to 1.017 by age 85, inside the age range where the NCHS blends in Medicare data.

Where the difference comes from

The NCHS method differs from ours in four places, which its Technical Notes describe (Arias and Xu, 2022):

  • It smooths the population and death counts over age with Beers' ordinary minimized fifth difference formula before dividing.
  • For ages 66 to 99 it blends the vital statistics death rates with death rates of Medicare enrollees, with weight $(95 - x)/30$ on vital statistics and $(x - 65)/30$ on Medicare at ages 66 to 94, and Medicare alone at 95 to 99 (equation 7). The report's reason: "Medicare data are considered more accurate than vital statistics and census data at the oldest ages because Medicare enrollees must have proof of age in order to enroll."
  • It smooths the rates from about age 85 (86 for the total population in 2019, Table IV) with a logistic model of the hazard proposed by Kannisto, fitted by maximum likelihood to ages 85 to 99, and extrapolates it to age 120, where the table runs out of survivors. The last published row is 100 and over.
  • It computes $q_0$ from births, as above.

Each of these can be switched off or on in the rebuilt table, one at a time, to see how much of the half year it explains.

Infancy. Replacing the births-based $q_0$ by the rate-based one with $a_0 = 1/2$ moves $e_0$ from 79.317 to 79.324: under a hundredth of a year. The two probabilities differ by about 1 percent, and a 1 percent change in a probability of 0.0056 moves few people.

The open interval. Keeping our rows to 84 but closing the table with the NCHS value $e_{85} = 6.67$ instead of $1/M_{85} = 7.56$ lowers $e_0$ to 78.93. The closing contributes 0.39 of the 0.47 years. The size follows from the recursion (4): changing $e_{85}$ alone changes $e_0$ by $(l_{85}/l_0)$ times the change, and $l_{85}/l_0 = 0.4401$ here, so $0.4401 \times 0.89 \approx 0.39$.

Ages below 85. Keeping our closing but using the NCHS probabilities $q_x$ for every age below 85 lowers $e_0$ to 79.23. The rows below 85 contribute 0.08 years. Averaged over single ages, our $q_x$ is 1.007 times the NCHS value at ages 1 to 29 (where the Beers smoothing and the small counts of single ages make the ratio noisy in both directions), 1.000 at ages 30 to 64, and 0.981 at ages 70 to 84, inside the Medicare blend. The figure shows the result: our survivors pull ahead only after 66.

So most of the half year is the last row. Why is one rate for 85 and over so far from the NCHS figure? There are three candidates: the population of 85 and over may be younger than the stationary one (part (c) of Theorem 6); its count may be off; or the NCHS rates at these ages, which rest on Medicare deaths and enrollees and on the smoothing model, may differ from the rates of the population that WONDER counts for other reasons (a different enrolled population, misstated ages on death certificates, the model itself). If the NCHS rates are taken as the true rates, the first two can be separated with the NCHS table. Its rates for single years from 85 to 99, and for 100 and over, are $m_x = d_x/L_x$. Dividing the WONDER deaths at each of those ages by the NCHS rate gives the population that those deaths and rates imply: 5,957,429 people aged 85 and over, against the census-based 6,604,958, which is 1.109 times as many. The implied population's mean age, 89.7 years, is almost that of the table's stationary population above 85, 89.8 years. With the implied population as denominator the open group's rate would be 146.7 per 1,000, close to the stationary rate $l_{85}/T_{85}$ of the NCHS table, 149.9 per 1,000, and far from the 132.3 per 1,000 that WONDER's denominator gives. If the NCHS rates are taken as the rates of this population, the larger part of the gap comes from the size of the denominator, the count of people aged 85 and over, not from its age structure. Without that condition the computation shows only that the WONDER deaths and population and the NCHS rates are inconsistent at these ages. The NCHS resolves the question in practice by preferring Medicare data, for the reason it states; the computation here does not decide which is right.

Abridged tables and the choice of ${}_na_x$

Older practice used wide intervals: an abridged table. The NCHS annual tables were abridged and "closed at age 85 and over" from 1945 to 1996 (Arias and Xu, 2022, Technical Notes). The rates unit's data come in the 11 NCHS age groups (under 1, 1 to 4, 5 to 14, then ten-year groups to 75 to 84, and 85 and over), so an abridged table can be built from them and set beside the complete one.

With the same infant row and open interval as before, the three conversions give:

  • uniform deaths, (2) with ${}_na_x = n/2$: $e_0 = 79.27$, with 45,388 survivors at 85;
  • a constant hazard in each interval, (3): $e_0 = 79.20$, with 45,725 survivors at 85;
  • the ${}_na_x$ of the NCHS complete table in (1): $e_0 = 79.45$.

The complete table, from the same deaths and populations by single year, gave 79.32. Two things are visible. The constant hazard is the furthest from the complete table among the two standard assumptions, as the discussion of ${}_na_x$ predicted: it underestimates the probability of dying in every ten-year interval of later life. And the "correct" ${}_na_x$, taken from a complete table, does not by itself reproduce the complete table: the ${}_na_x$ of our own complete table, $a_{75} = 5.45$, also gives 79.45.

The reason is the other half of an abridged table's approximation. The table's own death rate for an interval weights the single-year rates by the table's stationary person-years ${}_nL_x$, while the observed rate weights them by the actual population. For the interval from 75 to 84 our complete table has ${}_{10}m_{75} = 0.04452$ per person-year, and the observed rate is 0.968 times that; for 65 to 74 the ratio is 0.979. The observed rates are lower than the table's, which is what the argument of part (c) of Theorem 6 predicts when the population within an interval is younger than the stationary one. With the table's own rates and ${}_na_x$ the abridged calculation reproduces the complete $e_0$ exactly (the script checks it), as it must, since (1) is an identity. For an abridged table built from observed rates, uniform deaths happen to offset part of the error of the rates. That is a property of these data, not a rule.

Example (2019 against 2020). The period table compares years. The abridged table with uniform deaths, built in the same way from the WONDER deaths and populations of 2020 (except the infant row: without the 2020 births, $q_0$ is the 2020 infant rate times the 2019 ratio of the births-based $q_0$ to the infant rate), gives $e_0 = 77.15$, a fall of 2.11 years from the 79.27 of 2019 by the same method. In 2020 the death rate rose in every age group from 5 on, by between 2 and 25 percent, and fell under age 5. The fall is a statement about the rates of 2020, not a forecast that babies born in 2020 will live two years less.

How precise is a life expectancy?

A life expectancy computed from all the deaths of a country has no sampling error in the survey sense, but its deaths are still the outcome of chance: the same population, with the same underlying risks, could have had a few more or a few fewer deaths at each age. The question is how much $e_0$ would move.

Chiang modeled the deaths $D_i$ in interval $i$ as binomial among the $N_i$ people alive at its start, with probability $q_i$, and derived from that the variance of the estimated probability, $S^2_{\hat q_i} = \hat q_i^2(1 - \hat q_i)/D_i$ (Chiang, 1967, equation 10). A Poisson model for $D_i$ with mean $m_i P_i$, as in Crude Rates and Their Precision, leads to nearly the same thing. For a one-year interval, (1) is $q = m/(1 + (1 - a)m)$, whose derivative is $1/(1 + (1 - a)m)^2 = (1 - (1 - a)q)^2$, and $\operatorname{Var}(\hat m) = m/P \approx m^2/D$. The delta method gives $\operatorname{Var}(\hat q) \approx (1 - (1 - a)q)^4\,m^2/D$, and since $m = q/(1 - (1 - a)q)$ this is $q^2(1 - (1 - a)q)^2/D$. With $a = \tfrac12$ it is $q^2(1 - q + q^2/4)/D$, which differs from Chiang's $q^2(1 - q)/D$ only by the term $q^4/(4D)$. Either way, the probabilities of different intervals are treated as independent, since they rest on different people's deaths.

Here is the idea of the derivation. Life expectancy is a smooth function of the probabilities $q_i$ and of the last row's rate. Its sensitivity to $q_i$ is the number of person-years that a death at age $i$ removes, weighted by the share of the table alive at $i$. The delta method turns sensitivities and variances into a variance.

Theorem 7 (the variance of life expectancy). Let the table have intervals $i = \alpha, \ldots, w - 1$ of lengths $n_i$ with fixed ${}_na_i$ (written $a_i$), and an open last interval closed by $e_w = 1/M_w$. If the estimates $\hat q_\alpha, \ldots, \hat q_{w-1}, \hat M_w$ are independent, each approximately normal around its true value with small variance, then $\hat e_\alpha$ is approximately normal around $e_\alpha$ with

$$\operatorname{Var}(\hat e_\alpha) \approx \sum_{i=\alpha}^{w-1}\Big(\frac{l_i}{l_\alpha}\Big)^2\big[e_{i+1} + (n_i - a_i)\big]^2\operatorname{Var}(\hat q_i) + \Big(\frac{l_w}{l_\alpha}\Big)^2\frac{1}{M_w^2 D_w}, \tag{5}$$

where $e_{i+1}$ is the life expectancy at the start of the next interval and $D_w$ the number of deaths in the open interval.

Proof. By (4), $e_i = (1 - q_i)(n_i + e_{i+1}) + q_i a_i$, and $e_{i+1}$ does not involve $q_i$. So $\partial e_i/\partial q_i = a_i - n_i - e_{i+1} = -[e_{i+1} + (n_i - a_i)]$. For $\alpha < i$, $e_\alpha$ depends on $q_i$ only through $e_{\alpha+1}$, with $\partial e_\alpha/\partial e_{\alpha+1} = 1 - q_\alpha = p_\alpha$; by the chain rule, $\partial e_\alpha/\partial q_i = p_\alpha p_{\alpha+1}\cdots p_{i-1}\,\partial e_i/\partial q_i = (l_i/l_\alpha)\,\partial e_i/\partial q_i$. In the same way $\partial e_\alpha/\partial e_w = l_w/l_\alpha$, and $e_w = 1/M_w$ has $de_w/dM_w = -1/M_w^2$. Under the Poisson model $\operatorname{Var}(\hat M_w) = M_w/P_w = M_w^2/D_w$, using $D_w \approx M_w P_w$, so the last term's variance is $M_w^{-4}\cdot M_w^2/D_w = 1/(M_w^2 D_w)$. The delta method for several estimates (The Delta Method, Theorem 3) gives approximate normality with variance $\nabla^{\mathsf T}\Sigma\nabla$, where $\nabla$ is the vector of these partial derivatives and $\Sigma$ the covariance matrix of the estimates; independence makes $\Sigma$ diagonal, and the quadratic form is the sum (5). $\square$

Chiang's formula (Chiang, 1967, equation 17) is the sum in (5) without its last term, with $a_i$ written as a fraction of the interval, so that his $(1 - a_i)n_i$ is our $n_i - a_i$. Chiang (1967) states that the derivation is given in his two papers of 1960 on the stochastic study of the life table (Chiang, 1960a, 1960b). The omitted term matters when the open interval is wide and its deaths are few.

For the United States in 2019 the precision is not a concern. With $D_i$ the deaths at each single age and Chiang's binomial variance $\hat q_i^2(1 - \hat q_i)/D_i$ in (5) (here and in the simulation below, an age with no deaths contributes 0), the sum gives a standard error of $e_0$ of 0.0083 years by Chiang's formula and 0.0090 years with the last interval. Chance moves the national figure by days. The half year between the rebuilt and the published tables is a matter of method, and no confidence interval reflects it.

For a small population the picture changes. Take a county of 49,997 people with the U.S. age structure of 2019 (the national population scaled to 50,000, rounded at each age) and the U.S. death rates: about 435 deaths a year. We simulated 30 seeds (20261119 to 20261148) of 2,000 such years each, with the deaths at each age drawn from a Poisson distribution, built the table for each simulated year with uniform deaths at every age (including age 0, for simplicity), and computed (5). The true value of $e_0$ for this simplified table is 79.32 years. Over the 30 seeds (median, with the interquartile range in brackets):

  • the standard deviation of $\hat e_0$ across simulated years was 0.733 (0.724 to 0.741) years;
  • the median standard error from (5) was 0.731 (0.729 to 0.731), and its ratio to the standard deviation 0.996 (0.987 to 1.009);
  • without the last term the ratio was 0.913 (0.904 to 0.924);
  • the interval $\hat e_0 \pm 1.96$ standard errors covered the true value in 0.951 (0.947 to 0.955) of the years with the last term and 0.925 (0.924 to 0.930) without it.

In this simulation the delta-method standard error with the open interval matched the spread of $\hat e_0$, and dropping the open interval cost about 8 percent of the standard error and 2.5 points of coverage: 0.44 of the table's newborns reach 85, and the open group's deaths are few.

What is proved here, what is cited, what the experiment shows

Proved here: the general conversion from a death rate to a probability of dying (Theorem 1); its forms under uniform deaths (Theorem 2) and a constant hazard (Theorem 3), with ${}_na_x < n/2$ under a constant hazard; the comparison of the two, with the cubic gap (Theorem 4); life expectancy as the mean remaining lifetime (Theorem 5), given the tail formula of Expectation, Variance and Covariance, and the recursion (4); the closing of the open interval, exact in a stationary population and never too short for a younger population under an increasing hazard (Theorem 6); the delta-method variance of life expectancy with the last interval (Theorem 7), given Theorem 3 of The Delta Method.

Cited. Chiang's binomial variance of $\hat q_i$ and his formula for the variance of $\hat e_\alpha$ (Chiang, 1967, equations 10 and 17), which Theorem 7 rederives. The NCHS methods and its table: the not-stated factor, Beers' graduation, the births-based $q_0$ and the separation factor, the Medicare blend, the Kannisto model and the closing at 100 and over, and every published value (Arias and Xu, 2022, Table 1, Table III, Table IV and Technical Notes).

What the computations show. The 2019 U.S. table rebuilt from CDC WONDER, its agreement with the NCHS table to age 65 and the half-year gap in $e_0$, of which 0.39 years come from closing the table at 85 and 0.08 from the Medicare years below 85; that the census-based population of 85 and over is 1.109 times the population the NCHS rates imply; the abridged tables under three conversions; and, in the simulation of a small population, that the standard error (5) matched the spread of $\hat e_0$ while Chiang's formula without the last interval fell short. These are observations of these data and this simulation.

Worked problems

These problems were written for this article, on the 2019 data used above and on small exact cases; the article's script computes every answer.

Problem 1 (a ten-year interval). In 2019 the death rate of U.S. residents aged 75 to 84 was ${}_{10}M_{75} = 0.04309$ per person-year. Convert it to a probability of dying within the ten years under uniform deaths, under a constant hazard, and with the NCHS value $a_{75} = 5.48$. Compare with the NCHS table's ten-year probability.

Here $z = 10 \times 0.04309 = 0.4309$. Uniform deaths, (2): ${}_{10}q_{75} = 0.4309/(1 + 0.2154) = 0.3545$. A constant hazard, (3): $1 - e^{-0.4309} = 0.3500$. With $a_{75} = 5.48$ in (1): $0.4309/(1 + 4.52 \times 0.04309) = 0.3606$. Theorem 4 orders the first two, and the third is larger still because a larger ${}_na_x$ means each death removes fewer person-years, so the same rate needs more deaths. The NCHS table has $1 - l_{85}/l_{75} = 0.3802$. The remaining difference is not the conversion: it is the death rate itself, which in the NCHS table rests on the Medicare blend and weights the single ages by the table's stationary person-years rather than by the 2019 population.

Problem 2 (a three-row table by hand). A population has death rates $m_0 = 1/100$ in the first year of life (with $a_0 = 1/10$), ${}_4m_1 = 1/400$ from 1 to 5 (with ${}_4a_1 = 2$), and $M_5 = 1/20$ at all ages from 5 on. Build the table and find $e_0$.

By (1), $q_0 = \frac{1/100}{1 + 0.9/100} = \frac{10}{1009} = 0.009911$ and ${}_4q_1 = \frac{4/400}{1 + 2/400} = \frac{2}{201} = 0.009950$. Then $l_1 = 100{,}000(1 - 10/1009) = 99{,}008.9$ and $l_5 = l_1(1 - 2/201) = 98{,}023.8$. Person-years: $L_0 = l_1 + 0.1(l_0 - l_1) = 99{,}108.0$; ${}_4L_1 = 4l_5 + 2(l_1 - l_5) = 394{,}065.4$; the open row by Theorem 6, $L_5 = l_5/M_5 = 20\,l_5 = 1{,}960{,}475.1$. So $T_0 = 2{,}453{,}648.5$ and $e_0 = T_0/l_0 = 24.536$ years. As a check, the table's own rates are the given ones: $(l_0 - l_1)/L_0 = 1/100$ and $(l_1 - l_5)/{}_4L_1 = 1/400$ exactly, as Theorem 1 guarantees. The answer is dominated by the open row: everyone who reaches 5 lives 20 more years on average.

Problem 3 (the precision of life expectancy at 85). Find the standard error of $e_{85} = 1/M_{85}$ for the United States in 2019, and for a population with the same rate but 500 deaths a year at 85 and over.

From Theorem 7 with $\alpha = w$, the standard error is $1/(M_{85}\sqrt{D_{85}})$, that is, $e_{85}/\sqrt{D_{85}}$. With $e_{85} = 7.56$ years: for the United States, with 873,746 deaths, $7.56/\sqrt{873{,}746} = 0.0081$ years. For 500 deaths, $7.56/\sqrt{500} = 0.34$ years. The relative standard error of the rate, $1/\sqrt{D}$, passes unchanged to its reciprocal. Neither number reflects the difference of 0.89 years between the rebuilt and the NCHS values of $e_{85}$, which is a matter of method and data at the oldest ages (see "Where the difference comes from"), not of chance.

Summary

  • A death rate becomes a probability of dying through the years lived by those who die: ${}_nq_x = n\,{}_nm_x/(1 + (n - {}_na_x){}_nm_x)$ exactly (Theorem 1).
  • Uniform deaths give ${}_na_x = n/2$ (Theorem 2), a constant hazard gives $1 - e^{-n m}$ with ${}_na_x < n/2$ (Theorem 3), and the uniform answer is always larger, by about $(nm)^3/12$ (Theorem 4). At older ages in ten-year intervals the true ${}_na_x$ exceeds $n/2$.
  • $e_x = T_x/l_x$ is the mean remaining lifetime of the table's population (Theorem 5). Closing the table with $1/M_w$ is exact for a stationary population and is at least $e_w$ for a younger one when the hazard rises with age (Theorem 6).
  • A period table is a summary of one year's rates, not a forecast; under rates falling 1 percent a year a 2019 cohort would live 91.1 years against the period 79.3.
  • Rebuilt from CDC WONDER, the 2019 table gives $e_0 = 79.32$ against the NCHS 78.85. Closing at 85 with WONDER's population explains 0.39 years, the Medicare years below 85 another 0.08, and infancy almost nothing.
  • The delta-method variance of $e_\alpha$ is Chiang's sum plus a term for the open interval (Theorem 7). For the U.S. it is negligible (0.0090 years); for a population of 50,000 it is about 0.73 years, and omitting the open interval understates it by about 8 percent.

What we have learned

A life table is a chain of conditional probabilities, one per age interval, joined by a rule for the years lived inside each interval. The rule hardly matters for single years and matters a little for decades, while the last, open interval can move life expectancy at birth by almost half a year: the oldest ages are where the data are least reliable and where every life expectancy in the table collects its last years. The precision of a national life expectancy is days; its accuracy is set by method.

The table here was built from rates, with everyone's death or survival in a year counted. A study that follows people, as the NCHS follow-up of NHANES participants does, sees some deaths and loses sight of everyone else at the end of follow-up. Censoring, Survival Functions and Hazards defines the survival function and the hazard for any lifetime, shows what such incomplete observation can and cannot reveal, and writes the likelihood that the rest of the unit uses; The Kaplan–Meier Estimator is the life table's product of conditional probabilities rebuilt for individual follow-up times.

References

  • Arias, E., and Xu, J. (2022). United States life tables, 2019. National Vital Statistics Reports, 70(19). National Center for Health Statistics. https://doi.org/10.15620/cdc:113096. Introduction, Data and Methods, Table 1 (from NCHS's spreadsheet of the table), Tables III and IV, Technical Notes (equations 1, 3 to 7, 17 and 18). record · 3 pages
  • Centers for Disease Control and Prevention, National Center for Health Statistics. Underlying Cause of Death 1999–2020, CDC WONDER online database, released 2021, from the Multiple Cause of Death Files compiled through the Vital Statistics Cooperative Program. https://wonder.cdc.gov/ucd-icd10.html (queried November 2026: deaths of 2019 by single year of age, and deaths and populations of 2019 and 2020 by the 11 NCHS age groups). record · 6 pages
  • Chiang, C. L. (1960a). A stochastic study of the life table and its applications: I. Probability distributions of the biometric functions. Biometrics, 16(4), 618–635. (Cited through Chiang, 1967, references 3 and 4, as the source of the derivations; not read.) record
  • Chiang, C. L. (1960b). A stochastic study of the life table and its applications: II. Sample variance of the observed expectation of life and other biometric functions. Human Biology, 32(3), 221–238. (Cited through Chiang, 1967; not read.) record
  • Chiang, C. L. (1967). Variance and Covariance of Life Table Functions Estimated From a Sample of Deaths. Vital and Health Statistics, Series 2, No. 20. Public Health Service Publication No. 1000-Series 2-No. 20. National Center for Health Statistics. https://www.cdc.gov/nchs/data/series/sr_02/sr02_020acc.pdf. Section II, equations 3 to 17. record