Is the 1918 pandemic over? Almond (2006), Long-Term Effects of In Utero Influenza Exposure

The fetal origins hypothesis, associated with the epidemiologist David Barker, holds that some chronic health conditions trace back to how a fetus developed. Animal experiments support it, but testing it in people is hard for two reasons. Families whose babies have better conditions in the womb also differ in genes and resources, and the damage may take decades to appear.

The 1918 influenza pandemic offers an unusually clean test. About 25 million Americans caught the flu and survived, including roughly one in three women of childbearing age. The deadly fall wave arrived without warning in October 1918 and had largely passed by January 1919, so babies born just months apart had very different experiences in the womb. Its severity also varied erratically across states. And because the 1960, 1970 and 1980 Censuses record each person’s quarter and state of birth, those babies can be followed into middle age.

The damage runs through almost every outcome the census records. The 1919 birth cohort, in the womb at the pandemic’s peak, got less schooling, earned less, was more often poor and disabled, held lower-status jobs, and received more in welfare payments. The effects appear for men and women, whites and nonwhites. Men born between January and September 1919, for example, are estimated to have had about 20% higher rates of work-preventing disability at age 61 because of their prenatal exposure.

Two features limit the worry that only the hardiest babies survived. The pandemic killed about 0.5% of the U.S. population, far fewer than the famines studied before. And if it did weed out the weakest, the survivors would look healthier, so the estimates would understate the damage rather than overstate it.

Bottom line. If prenatal health shapes adult earnings and disability, then improving the health of pregnant women pays off decades later, a return that conventional cost-benefit calculations leave out.

Economists knew that adult health and earnings move together, and studies had linked maternal smoking and low birth weight to worse schooling and health. The worry is that such measures of fetal health stand in for unobserved family traits, such as genes. Twin studies in California, Norway and Canada hold the family fixed and still find that the heavier twin does better. But parents may treat twins with different endowments differently, which can bias those estimates too.

The strongest evidence came from famines. No lasting mortality effects were found for people in the womb during the Finnish famine of 1866–68 or the siege of Leningrad. Both were long and very deadly, which leaves room for the weakest to die and for fewer, healthier couples to conceive. The Dutch Hunger Winter of 1944–45 was shorter and less deadly, and people in the womb during it later showed more heart disease, poorer self-reported health and more antisocial personality disorder.

Almond’s design keeps the Dutch famine’s best feature, a short and sharp shock, and adds two more: very low mortality, and outcomes measured in schooling, income and work rather than only in health.

Bottom line. Earlier evidence was either tangled up with family background or clouded by high mortality and long exposure. The 1918 pandemic avoids both problems.

Doctors have long suspected that influenza can leave lasting damage. For infection during pregnancy, the most studied link is with schizophrenia in the child, supported by blood samples preserved from pregnant women in Oakland in 1959–66 and by experiments in mice. Earlier work by Almond and Mazumder also linked the 1919 birth cohort to more diabetes and stroke. The medical pathway remains debated.

The fall wave crossed the United States in about a month and, by its end a few months later, had killed more Americans than all U.S. combat deaths of the twentieth century. Where it hit hardest looked arbitrary: St. Paul’s death rate was 70% higher than neighboring Minneapolis’s, and Dayton’s 80% higher than Columbus’s. Investigators at the time could not tie the variation to population density, location or economic development; only weather seemed to matter.

Unlike ordinary flu, the 1918 virus struck adults in their prime: deaths by age formed a W, with peaks among infants, the elderly and people around 30. About 28% of the population reported being infected, and about a third of women of childbearing age. Pregnant women were especially vulnerable, and the damage reached the babies they carried.

Bottom line. An unforeseen, brief and geographically erratic shock that hit pregnant women hard is close to what an experiment on prenatal health would look like.

A pandemic can raise infant deaths in two ways, with opposite implications for the survivors. If it damages everyone’s health (scarring), more babies die and more of the survivors end up disabled as adults. If it mainly kills babies who were already the weakest (culling), the survivors are healthier than usual.

Almond captures this with an unobserved health index, h*, fixed at birth. A person dies before adulthood if h* falls below a survival threshold δ₀, and is disabled as an adult if h* lies between δ₀ and a higher threshold δ₁. With F the distribution of health:

This matters for reading the results. Studies that used mortality rates as a proxy for fetal health implicitly assumed scarring. To the extent that the pandemic instead culled weak infants, the survivors would look better, and estimates of the damage would be biased toward zero. Neither contemporary sources nor the paper’s results suggest culling was the main effect.

Bottom line. More infant deaths together with more adult disability point to scarring. Culling would push adult outcomes the other way.

Only three censuses after the pandemic, those of 1960, 1970 and 1980, record adults’ quarter of birth, which is what lets the analysis separate babies born months apart. Almond uses the public-use samples: 1% of the population for 1960, 3% for 1970 and 5% for 1980. He keeps people born in the United States, which ties them to the timing of the U.S. pandemic, and drops records whose age (or, for the state analysis, birthplace) was imputed by the Census Bureau.

Birth year has to be reconstructed, because age is recorded as of 31 March: it is census year − age − 1 for people born April to December, and census year − age for those born January to March. For 1980 disability questions, Almond uses a Berkeley extract of the census, because the standard IPUMS coding merged two questions that were no longer mutually exclusive in 1980.

The two groups have the same average age in each census, and they shared the same history, including World War II, a year apart. Yet the 1919 cohort comes out worse in 27 of the 28 comparisons. The one exception is Social Security income in 1970, slightly lower for the 1919 cohort.

For the geographic analysis, the best available measure of the pandemic’s toll on pregnancy is the maternal mortality rate: deaths from childbirth per birth. In 1917, 12,528 mothers died in childbirth, 0.66% of live births. Influenza only became a reportable disease after the pandemic began, so there are no reliable infection counts by state, and births were recorded in just 19 states.

Bottom line. Large census samples with quarter and state of birth make it possible to compare adults born months apart, and born in places the pandemic hit differently.

No two birth cohorts can be compared with everything else equal: a cohort born a year later is a year younger at each census, and cohort differences tend to change smoothly. The pandemic makes a sharp prediction instead. Only the cohort in the womb at the peak should break from the trend; those born just before, exposed as infants, and those conceived after it ended should not.

The dip appears for men and for women, which rules out military service as the explanation, and in all three censuses. The 1919 cohort was 4–5% less likely to finish high school than its trend predicts. Since only about a third of pregnant women caught the flu, the effect on children of infected mothers is roughly three times larger: 13–15% less likely to graduate and about five months less schooling. Relative to their lower average schooling, effects for nonwhites are about twice as large.

The larger 1980 sample allows a look by quarter of birth. Graduation rates fall below trend for babies born in the first two quarters of 1919, significantly so for each, and recover afterward.

Could the 1919 babies simply have had different parents? Those born in the first half of 1919 were conceived before the pandemic arrived, so any change in who had children would have had to start months before the virus appeared and stop right after it left. The share of people with a foreign-born parent, which strongly predicts schooling and income, shows no break for 1919 (t-statistic 0.20).

The dip does shrink by about a quarter between 1960 and 1980 for men and women, and more for nonwhites, which suggests that the worst affected tended to die younger. And because schooling is decided in the teens, these effects fit “life course” models of health better than a narrow reading of the fetal origins hypothesis, which expects damage to surface after the reproductive years.

Bottom line. The 1919 cohort breaks sharply from an otherwise smooth rise in schooling, and nothing the census records about its parents explains the break.

Disability questions first appear in the 1970 Census. That year, men born in 1919 were 6% more likely to report a disability that limited work, and 8% more likely to report one that prevented it: 17% and 25% more likely for sons of infected mothers.

With a milder, work-limiting measure, the 1980 data suggest that babies born in the last quarter of 1918, exposed late in pregnancy, may have been affected too.

Men born in 1919 earned $700–900 a year less in wages (2005 dollars), about $2,500 or 5–9% for sons of infected mothers. Part of that follows from their shorter schooling. How much?

Transfers rose too. Disabled workers could receive Social Security disability payments before retirement, and most of these were recorded as welfare income. In 1980, men born in 1919 received about 8% more in welfare. For women and nonwhites in 1970, average welfare payments were 12% higher, roughly a third higher for children of infected mothers. Babies born April–June 1919, in the first trimester at the pandemic’s peak, received the highest average welfare payments of any birth quarter from April 1911 to April 1925.

Less schooling and lower wages may also have made crime relatively more attractive. In the 1940 Census, which catches these cohorts near the peak ages for crime, people born in 1919 were significantly more likely to be in jail than the surrounding cohorts.

Bottom line. Beyond schooling, prenatal exposure raised disability and welfare receipt and cut earnings, and perhaps half of the wage loss has nothing to do with schooling.

The second approach compares people born in the same year but in places the pandemic hit differently. Timing alone is little help: the census does not record month of birth, and the virus crossed the country in about a month. But contemporaries reported that it grew milder as it spread, which allows two indirect tests.

According to a U.S. Public Health Service map, the fall wave reached seven western states last:

If the virus had weakened by then, people born there in 1919 should show milder long-term effects. Going by the signs of the estimates, they generally do, and the difference is significant for four outcomes: poverty in 1960, work-preventing disability in 1970, and wage income and Social Security income in 1980. With only 3% of U.S. births in these states, statistical power is limited, and the state-level data behind the map have been lost.

A 1918 survey by Sydenstricker also recorded when each of the nine census divisions reached the epidemic stage, and, for Baltimore, how often infections turned into pneumonia each week, a measure of virulence. Combining the two gives an average virulence score by division: New England and the East South Central states met the most virulent phase, the Mountain and Pacific states the mildest.

In the small 1960 sample, β₂ mostly has the expected sign but is imprecise. The 1970 results are more consistent: people born in 1919 in more virulent divisions were more often poor, lived in lower-income neighborhoods, and were more often and longer disabled. In 1980 the estimate is imprecise except for wage income, which is lower as predicted.

Bottom line. The timing and virulence data are thin, but they point in the same direction as the cohort comparison.

The main geographic test uses how much maternal mortality rose in each state. Almond gives four reasons to prefer it to overall death rates: those depend on each state’s age mix, given the W-shaped age profile of deaths; men died more, partly because of tuberculosis; deaths in childbirth have few close substitutes as a recorded cause; and maternal mortality has been used before as a marker of conditions in the womb. The cost is coverage: only the 19 birth-registration states, mostly northeastern, wealthier and whiter, have the data, covering about half of the U.S.-born population.

Across those states, maternal mortality rose from 0.66% of births in 1917 to 0.92% in 1918, with the rise concentrated among births late in the year, roughly a doubling that fall. The increase varied erratically: Kansas mothers saw more than ten times the rise that Wisconsin mothers did.

Most maternal deaths, even in 1918, were not caused by the flu, so Almond converts the rise into an infection rate. If a share w of pregnant women was infected, and infected mothers died at rate k, the 1918 rate is a weighted average of k and the normal rate. Rearranging gives equation (7):

Each adult’s census outcome is then regressed on the infection rate in their state of birth in the year before they were born, to capture the time in the womb. Birth-year fixed effects remove differences between cohorts, and state fixed effects remove fixed differences between states, so the effect is identified by how sharply infection rose and fell in each state around the pandemic.

Every outcome moves in the expected direction. At the average one-third infection rate, schooling falls by about a quarter of a year, income by about 6%, and poverty rises by about 1.5 points. These effects are about 50% larger than the cohort-comparison estimates for 1960, though the difference is significant only for income. Infant mortality enters with the same sign as infection, again pointing to scarring rather than culling. The 1980 estimates are about half as large as those for 1960. In 1970 they are not significant, partly because many states are missing from some of the 1970 samples.

Bottom line. A different source of variation, differences between states within the same birth cohort, gives effects of similar or larger size, at least in 1960 and 1980.

To explain the cohort results, a rival factor would have to jump for the 1919 birth cohort and nowhere else. To explain the state results, it would have to follow the pandemic’s erratic geography as well. Pick a suspect to see how the paper tests it.

Bottom line. Within the data available in 2006, adding controls moves the estimates only modestly, and the checks that should strengthen the effect do so.

The paper’s main estimates for men born in 1919, relative to trend, and what they imply for sons of infected mothers:

High school graduation

1.4–2.1 points lower; 13–15% less likely to graduate for children of infected mothers.

Years of schooling

0.12–0.18 years fewer; about five months for children of infected mothers.

Wage income

$700–900 a year lower; about $2,500, or 5–9%, for sons of infected mothers.

Poverty

0.6–1.0 points more likely to be below 150% of the poverty line.

Socioeconomic status

0.6–0.8 index points lower on Duncan’s occupational scale.

Disability, 1970

8% more likely to have a work-preventing disability; 25% for sons of infected mothers.

Welfare, 1970

12% higher payments for women and nonwhites born in 1919.

Almond draws three lessons. Congenital traits, like inherited ones, can have hidden effects that surface only decades later. Unlike genes, they are shaped by the prenatal environment, so economic outcomes are more open to improvement than usually assumed. And differences in fetal health may explain part of the strong link between adult health and income.

He singles out racial disparities. At the time of writing, a black infant in the United States was more than twice as likely as a white infant to die before age one, and he argues that improving the early-life health of black infants could narrow later gaps in health and income.

Bottom line. Health before birth can shape human capital for a lifetime, which makes it an economic investment and not only a medical one.

Attrition

People missing from a later data source, here the share of a birth cohort that does not appear in the census, mostly because they had died.

Birth cohort

Everyone born in the same period, such as the same year or quarter.

Birth-registration states

The 19 states that recorded births in 1917–19, for which maternal mortality rates are available.

Culling

When a shock kills the weakest, leaving survivors who are healthier on average; also called selective mortality.

Duncan socioeconomic index

A score of occupational status based on the education and income typical of each occupation.

Fetal origins hypothesis

The idea, associated with David Barker, that conditions in the womb shape chronic health in adulthood.

Fixed effects

Indicators for each group, such as each state or birth year, that remove everything constant within that group.

In utero

In the womb, before birth.

IPUMS

The Integrated Public Use Microdata Series, harmonized samples of census records used by researchers.

Latent variable model

A model in which an unobserved variable, here health at birth, determines observed outcomes such as death or disability by crossing thresholds.

Maternal mortality rate

Deaths of mothers related to childbirth, divided by the number of births.

Natural experiment

An event that assigns exposure as if at random, so that exposed and unexposed groups can be compared.

Sequelae

Lasting aftereffects of a disease.

Treatment effect on the treated

The effect on those actually exposed. Because only about a third of pregnant women were infected, the paper scales cohort-wide effects by three.

This guide paraphrases the published paper and redraws its results from its tables. The monthly death wave, the stillbirth series and the two cohort charts are traced from the paper’s Figures 1b, A1, 3 and 2, so their values are approximate; the culling-versus-scarring diagram uses illustrative numbers. The note on later research cites Beach, Brown, Ferrie, Saavedra and Thomas (2022), Journal of Political Economy 130(7): 1963–1990.