Household Spending during the 1918 Influenza Pandemic Free

Pandemics affect consumer preferences. Sickness, death, and restrictions on commerce can reduce household income and opportunities for spending. Just how much spending is affected, overall and for particular categories, has been an area of key concern during the covid-19 pandemic. To mitigate against the worst effects, governments in the United States and elsewhere provided relief to households likely to be most affected.¹

But what happens to consumer spending in the absence of such government largesse? To answer this question, we turn to the 1918 influenza, the largest twentieth-century pandemic. Influenza initially reached New Orleans by September 4, carried by three seaman who had arrived from Boston. In the next few days, it arrived at ports along the Atlantic and Gulf coasts then rapidly spread inland, following rivers and railways. By October it had spanned the entire country with devastating consequences. From March 1918 through March 1920, the pandemic killed between 500,000 and 675,000 people in the United States and between 50 and 100 million people worldwide.²

Uncertainty surrounds not only the death count but also the economic effects. Barro, Ursúa, and Weng find that the pandemic was associated with an average decline in gross domestic product and consumption of 6–8 percent across a sample of forty-eight countries. But death rates were higher in many of these countries, and in the United States the declines were smaller, at about 2 percent. Burns and Mitchell find a recession in the United States of “exceptional brevity and moderate amplitude.” Velde finds a sharp but short-lived decline in industrial output, little effect on retail, and no evidence of increased business failures or a stressed financial system. Baker and coauthors find no evidence that the pandemic can explain any of the twenty-three daily stock market jumps between March 1918 and June 2020. Arthi and Parman, in a review of lessons from the pandemic, refer to the “relatively mild and short-lived effects on the economy.” Brainerd and Siegler, as well as Correa and coauthors, find no lasting negative effects into the 1920s.³

It is in the context of this mixed macroeconomic evidence that we investigate household-level spending behavior during the pandemic in the United States. To do so, we use a cost-of-living survey administered in late 1918 and early 1919 by the Bureau of Labor Statistics (hereinafter bls), a surprisingly neglected source. One possible reason for neglect is that cities were only surveyed once; as a result, it is impossible to construct a city-level panel to measure changes over time in the same city. We instead compare households located in cities that were surveyed before, during, and near the end of the largest wave of deaths.

We first consider several possible mechanisms that could drive spending. The money that households use to purchase goods and services can come from income or other sources, such as borrowing and accumulated savings. We find that households surveyed before or in the first few weeks of the main wave of the pandemic had incomes of about $1,720 on average. Among households surveyed as the pandemic waned in late 1918 and early 1919, average income was about $100 lower. Average spending by these households exhibited the same difference: $1,640 on average among households surveyed before or at the start of the pandemic, $1,540 on average among households surveyed later. No matter when they were surveyed, households withdrew at most $20 from savings on average and borrowed even less. Furthermore, across the income distribution, the relationship between income and spending was similar no matter when households were surveyed. Differences in average income explain differences in average spending.

We next consider whether differences in average income were associated with variation in experience of the pandemic. The survey sample consisted of households with a husband, wife, and at least one child, meaning that these were households in which the primary wage earner did not die from influenza. However, significant medical expenses may signal households in which one or more members became sick with influenza. Among households with large medical expenses, the average income was lower for households surveyed later, just as it was for households with few medical expenses. Likewise, in cities with shorter public interventions and higher pandemic mortality, average income was lower for households surveyed later, just as it was in cities with less robust public interventions and more deaths. To the extent that the pandemic explains differences in average income (and thus average spending), it was substantially a nationwide effect and was not just concentrated among particular households or in particular cities.

We next consider whether the changing composition of cities during the survey’s rollout could explain differences in average spending between households surveyed at different times. The differences varied by region of the country, but in every region, average spending was lower among households surveyed as the pandemic started to wane in November and December 1918 compared to households surveyed as deaths surged in September and October 1918. Using a standard regression analysis that controls for geographical region as well as the race of each household head, we confirm that average spending was about $100 lower on average among households surveyed later.

We next track spending on various categories of goods and services. There was little difference in spending on durable goods, but spending on nondurable goods and services was up to 12 percent lower among households surveyed later. The quantity of life insurance sold door to door was lower and the price paid was similar for households surveyed later, suggesting that both the supply of and demand for this type of insurance were lower. By contrast, for a type of life insurance that required less personal contact, the quantity purchased was higher and the price lower among households surveyed later, suggesting greater supply. Finally, cluster analysis allows us to identify older, wealthier households and those located in the Midwest and West as jointly associated with large spending differences.

Much of what we know about the economic consequences of the pandemic is reconstructed from industry-level information. In perhaps the most comprehensive survey of American economic performance during the pandemic, Velde uses information on industrial output, retail sales, business failures, industry reports, and stock market prices. The same is true elsewhere. From Brazil to Spain, Indonesia to Sweden, Italy to South Africa, influenza had adverse but mild and generally short-lived economic effects. But those studies tend to focus on changes to production or wages. They rely predominantly on population or manufacturing census data, often aggregated at the regional level. The long-term effects of the pandemic are better understood, owing to more comprehensive, individual-level census data in later years.⁴

Using a household survey conducted during the 1918 influenza pandemic, we can offer for the first time a detailed account of spending differences between households observed at different times during the pandemic. From July 1918 through February 1919, the bls administered a cost-of-living survey to 12,817 households in 99 cities across 42 states.  Appendix I lists the cities. The Inter-university Consortium for Political and Social Research (icpsr) maintains two records of the survey: icpsr 8299, which provides household characteristics, income, and spending, and icpsr 6276, compiled by Olney, which provides additional variables about spending and saving.⁵

Each household was observed once, and all households in a city were typically observed in the same month. Figure 1 shows the rollout of the survey. It was first administered in Baltimore in July, then elsewhere in the Northeast in August, after which it expanded to the West Coast, and then broadly across the Midwest and South by early 1919. The month of observation is not recorded for fifteen households, and fifty households were observed in a month different from most of the other households in their city. Our final sample of 12,731 excludes these households, as well as twenty-one others that do not have race recorded.

The survey recorded income earned by each household member over the previous twelve months. The survey also collected information about amounts spent on more than 400 categories of goods and services over the previous twelve months. (The survey is identified as covering 1917–1919 because the retrospective income and spending covered parts of 1917 for households observed in 1918.) Households were selected to represent “typical” working-class families in industrial locales; every family had at least one wage earner or salaried worker, and every family consisted of a husband, a wife, and at least one child. The survey excluded households that had recently moved or that had recently immigrated and did not speak English. Of the household heads surveyed, 93 percent were white, and the survey included few Black households outside the Midwest and South. The average household had 4.5 members. Beyond the husband, wife, and child present in every household, 1.2 of the additional 1.5 members on average were children.

The survey offers one clear advantage for studying the pandemic: It was administered throughout much of the largest wave of deaths, which lasted from September 1918 into early 1919. The survey, therefore, observed households with a range of exposure to the pandemic, from those surveyed before the pandemic began to those for whom half of their year’s income and spending—September 1918 to February 1919—occurred during the pandemic.

The survey also has several limitations. First, the survey’s exclusively urban, working-class sample means that it does not represent people living in rural areas, or poorer or richer people living in cities. Second, the survey’s sample of households with a husband, wife, and child present excludes families in which one of the principal members died during the pandemic; we use spending on medicine and medical care to identify households that may have experienced illness.

Third, all households in a city were surveyed in the same month, so we cannot measure changes within a city as the pandemic progressed. (The survey also does not record when in a month a household was observed, so we cannot compare households observed in the days before or after the flu spread to their city.) Instead, we use the six-month rollout of the survey to compare households observed at different times during the pandemic. Because the survey was administered in only a few cities in some months, we group the months of the survey as July–August 1918, September–October 1918, November–December 1918, and January–February 1919.

Fourth, spending values (as well as income, withdrawals from saving, and borrowing) are totals over the previous twelve months. A shorter recall period would be susceptible to seasonal differences in spending. But this longer recall period obscures when within the previous year the spending occurred. We cannot compare average spending in each pair of months and instead can only compare average annual spending among households surveyed in each pair of months. However, the overlap of these recall periods is suggestive of the timing of any spending differences. For example, among households surveyed in July and August 1918 and households surveyed in September and October 1918, the twelve-month recall period has ten months in common (November 1917 through August 1918). Any difference in spending between these two groups of households is therefore suggestive of a difference in spending by the first group in September and October 1917 to spending by the second group in September and October 1918.

Table 1 describes the final sample grouped by month of survey. Among the 1,790 households in the 7 cities surveyed in July and August 1918, average spending over the previous year was $1,626 per household. Among households surveyed as the main wave of the pandemic began in September and October, average spending was slightly higher at $1,641 per household. Among households surveyed in November and December, average spending was $103 lower at $1,538 per household. Among households surveyed toward the end of the main wave of the pandemic in January and February 1919, average spending was similarly low, at $1,549 per household. Panel (a) of Figure 2 plots these average spending values. (These dollar values and, unless otherwise noted, all other dollar values in this article are converted to real January 1919 dollars using the monthly consumer price index, assuming that spending was spread evenly over the previous twelve months. The price level grew by more than 15 percent per year in the late 1910s, a sustained high rate of inflation that has not since been matched.)⁶

The money that households spend can come from a variety of sources, three of which were recorded by the survey: income, withdrawals from accumulated savings, and borrowing. Panel (b) of Table 1 reports the average values of these sources, and panel (a) of Figure 2 depicts these averages. Among households surveyed in July and August, average income was $1,686 over the previous twelve months. This average was $1,732 for households surveyed in September and October and was then about $100 lower for households surveyed at the end of 1918 or in early 1919. Throughout the period, average income remained between $60 and $95 higher than average spending.

Households withdrew little from savings and borrowed even less. Average withdrawals from savings over the previous twelve months were $12 for households surveyed in July and August, and just $8 higher for households surveyed in early 1919. Average household borrowing over the previous twelve months remained less than $8 regardless of when households were surveyed. Differences in income, and not these other sources of money, explain differences in spending for households surveyed at different times.  Appendix II shows that there was little variation in several forms of saving, such as cash in the bank. Rather, differences in average spending reflected differences in average income.

A concern with these comparisons is that they might reflect differences in ability to recall the true amounts of spending (and income, withdrawals from savings, and borrowing) among households surveyed as the pandemic progressed. A signal of such inaccuracy is the share of reported values that are round numbers ending in zero or five. For example, ages in a population tend to be evenly distributed across numbers ending in 0 through 9, so when more than one-fifth of people report ages ending in 0 or 5, it indicates that some are reporting approximate rather than exact ages. Of course, many prices, quantities, and spending amounts may truly be round numbers, so some heaping is expected. But differences in heaping among households surveyed at different times could signal measurement error. Across all the individual categories of goods for which spending was recorded, panel (c) of Table 1 reports the share of the nominal amounts that are positive and end in zero or five. Among households surveyed in July and August 1918, this share was 30.4 percent. The share then declined slightly for households surveyed later, reaching a value of 28.5 percent among households surveyed in early 1919. Conversely, 36.8 percent of income values reported by households surveyed in July and August 1918 were round numbers, and this share rose to 39.3 percent for households surveyed in early 1919. These estimates suggest that measurement error in spending declined in households surveyed later while measurement error in income rose. But both differences were small and are unlikely to drive the patterns in panel (b). Heaping is more common for reported withdrawals from savings and borrowing, but again few households relied on these sources of money.⁷

Yet, it is possible that the consistent relationship between average income and average spending, demonstrated in panel (b) of Table 1, masks variation across the income distribution. Panel (d) reports average spending for households grouped by income, from $800 through $2,600, roughly the fifth and ninety-fifth percentiles of household income. Panel (b) of Figure 2 depicts these consumption functions for each month of survey. In each case, the average slope of the consumption function is between 0.83 and 0.86, meaning that households spent, on average, about $6 of every additional $7 of income. At each level of income, average spending did not vary by survey timing.

Households surveyed before the main wave of the pandemic or just as deaths surged spent about $1,630 on average. Households surveyed later spent about $100 less on average, a difference of 6 percent. Differences in average income explain these differences in average spending. There were no concurrent differences in households’ withdrawals from savings or borrowing. There were also no concurrent differences in spending as a share of income for households across the income distribution.

Can variation in experience of the pandemic account for differences in average income? Most consequentially, the pandemic led to the death of wage earners, but because the survey sample consists exclusively of households in which the husband and wife were alive, we cannot observe households that lost a primary wage earner. We are also unable to identify directly whether any household member had been sick with influenza. Instead, we use medical expenses to infer experience of illness.

Average household medical spending was $55 among households surveyed in July and August 1918 and $71 among households surveyed in early 1919. This difference might reflect increased pandemic-related medical needs and also highlights that households surveyed early on, before the main wave of the pandemic, still had substantial medical expenses. More than 97 percent of households had some medical expenses. Medical spending is, at best, a coarse signal of experience of influenza.

We group households by above- and below-median medical spending—$45—and measure average household income in each group in each month of the survey. Panel (a) of Figure 3 reports these average values. Among households with above-median medical spending, the average income was $1,808 for households surveyed in July and August, nearly the same for households surveyed in September and October, and then about $100 lower for households surveyed in late 1918 and early 1919. Among households with below-median medical spending, average income was consistently lower, but differences in average income followed a similar pattern: Compared to households surveyed in July and August, average income was about the same (or slightly higher) for households surveyed in September and October, then was about $100 lower for households surveyed later. Households with more income spent more on average, including medical spending, but differences in average income across the survey period did not depend on medical spending. This comparison suggests that experience of illness (as coarsely measured by medical spending) alone does not explain differences in income and, by extension, differences in spending.

We next use two aggregate measures of the pandemic at the city level—public interventions and mortality rates. Markel and coauthors collected information about the public policy response to the 1918 pandemic in forty-three cities, thirty-three of which were included in the bls Cost of Living Survey. They focus on three categories of non-pharmaceutical interventions—school closures, bans on public gatherings, and isolation and quarantine measures—and tally the total number of days that interventions were imposed. In nearly every city, interventions were in effect by early October. In some cities, interventions lasted just a few weeks; in others, several months.⁸

Because the h1n1 virus was not directly detectable at the time, deaths due to the pandemic are commonly measured as deaths exceeding the baseline number of deaths in the preceding and following years. Using annual mortality statistics reports, Collins and coauthors tabulate excess due to influenza and pneumonia in excess of the average in the same city and month between 1910 and 1916. They do so for fifty cities, thirty-eight of which were included in the bls Cost of Living Survey. We sum these monthly values for September 1918 through February 1919, the period of elevated mortality during the pandemic. On average, cities with longer interventions experienced fewer deaths.⁹

Because the duration of interventions and mortality were recorded in fewer than half of the cities, we group the months of the survey as July through September, October and November, and December through February. The included households were surveyed before or just as the pandemic began, at the height of the pandemic, and as pandemic deaths waned. Panel (b) of Figure 3 reports the average income for households in cities with below- and above-median duration of non-pharmaceutical interventions, ninety-nine days. Panel (c) reports the average income for households in cities with above- and below-median pandemic mortality, 500 deaths per 100,000 people.

Again, cities with shorter interventions tended to have higher mortality. As shown in both panels (b) and (c), in these cities there was little difference in average income between households surveyed before and at the height of the pandemic. As we have shown using contemporary census records, labor market changes during the pandemic, such as the widespread death of working-age adults, drew surviving women into the labor force, and their additional earnings may have blunted any overall decrease in income during the pandemic. But there was then a larger difference for households surveyed as the pandemic waned, particularly among cities with many deaths. In cities with longer interventions and lower mortality, there was a larger difference in income between households surveyed before and at the height of the pandemic. Average income was again lower among households surveyed near the end of the pandemic.¹⁰

Together, this evidence indicates that longer interventions (and lower mortality) were associated with larger immediate differences in average household income. Yet, the patterns across all the panels in Figure 3 are clear. Like individual households’ experience of illness, city-level pandemic characteristics do not completely explain differences in income. In every case, households surveyed later spent less.

The evidence presented thus far shows that households surveyed later spent less on average, that these differences closely match differences in average income, and that households’ and cities’ varied experiences of the pandemic do not fully explain these differences. It is possible that the cumulative experience of the pandemic in general explains lower spending by households surveyed later. It is also possible that these differences in spending are unrelated to the pandemic and are instead due to other differences between households surveyed at different times.

To examine whether differences in sample composition can explain differences in spending, we consider two concerns about sample composition. First, the survey’s rollout, from the Northeast to the West and then Midwest and South, offers a non-random cross-section of the country at any particular time. If cities surveyed later were in regions with generally lower spending, then average spending would appear to fall over time even without the pandemic. Second, Black households spent less than white households on average. If cities surveyed later had a higher proportion of Black residents, this changing racial composition of the survey sample could explain lower average spending among households surveyed later.

To address these concerns, we first split the sample by census region. Panel (a) of Table 2 provides overall sample sizes and average spending, repeated from Table 1. Panel (a) of Figure 4 plots average spending. Panel (b) of Table 2 provides equivalent sample sizes for households living in the Northeast. Most Northeastern households were surveyed by October 1918, and none in 1919. As depicted in panel (b) of Figure 4, average spending was $1,684 among households in the Northeast surveyed in July and August 1918. This average was $45 lower among households surveyed as the pandemic began in September and October and was a further $119 lower among households surveyed in November and December. Within the Northeast, spending was lower among households surveyed as the pandemic progressed.

Panels (c) through (e) of Table 2 and Figure 4 present equivalent sample sizes and spending averages for the other census regions. In the Midwest and South, average spending was higher among households surveyed in September and October compared to households surveyed in July and August. However, in each region, only one city was surveyed in July and August. In all regions, average spending was at least $94 lower among households surveyed in November and December compared to households surveyed in September and October.

We next split the sample by race. Nearly 94 percent of household heads were white. As shown in panel (f) of Table 1 and Figure 2, spending averages for white households grouped by month of the survey were nearly the same as for the whole sample. As shown in panel (g), among Black households surveyed in July and August, average spending was $1,389. The average was nearly $200 lower for Black households surveyed in September and October. There was little difference in average spending for Black households surveyed in late 1918 or early 1919. This comparison suggests that, unlike for white households, there was little difference in average spending for Black households surveyed in September and October compared to Black households surveyed in November and December. Unfortunately, though, these averages for Black households are calculated using small samples, concentrated in just a few cities. For example, just sixty-eight Black households were surveyed in September or October, and fifty-five of them were in Cincinnati. ( Appendix I provides sample sizes by race for all ninety-nine survey cities.)

There is one observation per household. Spending records the total amount spent by the household over the previous twelve months. The main estimated coefficients of interest, β₁ through β₃, estimate the difference in average spending between households observed in each of the three final month groups and households observed in July and August. Midwest, South, and West are dummy variables equal to one if the household lives in the Midwest, South, or West (Northeast is the omitted region). Black equals one if the household head is Black (white is the omitted race). We cluster standard errors at the city level. Clustering does not change the point estimates but widens the confidence intervals, generally by a factor of two or three.

Column 1 of Table 3 presents the regression results, first estimated without the region- and race-fixed effects. The estimated coefficients on each month group calculate the average spending among households surveyed in those months minus the average among households surveyed in July and August. The differences between the average spending values are presented in panel (b) of Table 1. Column 2 of Table 3 presents the full regression estimates, including region and race fixed effects. Comparing within regions and within racial groups, average spending was $26 lower among households surveyed in September and October compared to households surveyed in July and August. This difference grew to −$107 among households surveyed in November and December and was −$78 in January and February 1919. Average household spending was lower among households surveyed later as the pandemic progressed, especially among households surveyed in November and December. These differences are large. For example, after considering differences in average spending by census region and race, spending by households surveyed in in November and December was, on average, 6.6 percent (107 ÷ 1,626) lower compared to households surveyed in July and August, a difference that is statistically distinguishable from zero at the 5-percent level of significance.

Again, though, these regression estimates do not prove that the pandemic caused changes in spending. Other differences in sample composition, beyond geography and race, could explain the differences in spending. In  Appendix III, we show that differences in average spending by month of the survey are generally consistent when the sample is split by family size and exposure to World War I. In  Appendix IV, we show that, for a subset of cities in which a previous cost of living survey was conducted thirty years earlier, the findings are robust to controlling additionally for city-level average spending in 1888. Yet, there may be unobservable features of cities and households surveyed at different times that explain the differences in average spending. We therefore continue to interpret estimated coefficients β₁ through β₃ as measuring differences in average spending between households grouped by the month of the survey.

Total household spending was lower among households surveyed later in the pandemic. Across categories of spending, there was variation in the size and even direction of these differences. Newspapers at the time reported a change in spending patterns. We document one illustrative example from Philadelphia, which had suffered severely from the pandemic because of the slow public health response. In January 1919, the Woman’s Exchange in the Evening Public Ledger made evident the urgent need to adjust spending to the changing circumstances:

The bls Cost of Living Survey recorded spending on hundreds of different items, which we categorize in several ways. For each category of spending, we estimate equation 1 to measure differences in average spending between households surveyed at different times and report the regression estimates in Figure 5. First, we separately measure spending on durable goods, nondurable goods, and services, as shown in the first three panels of Figure 5. Durable goods, including jewelry and furniture, accounted for less than 5 percent of spending, and there was little difference in average spending between households surveyed at different times. Nondurable goods, including food and clothing, accounted for more than 60 percent of spending. Compared to average spending on nondurable goods among households surveyed in July and August, average spending was $11 lower among households surveyed in September and October, and another $27 lower among households surveyed in November and December. Nondurable goods and services drive the overall differences in total spending.

We next measure spending on various other categories of goods. Food accounted for more than 40 percent of household spending in July and August, and average spending on food was $53 lower for households surveyed in November and December. (Eating out does not substantially affect spending on food over the course of the pandemic; more than 97 percent of spending on food was for food to be prepared at home. Conversely, by the time of the 2020 covid-19 pandemic, eating out accounted for a third of spending on food.) Spending on clothing, housing, rent, life insurance, and liquor were all also lower among households surveyed later. There was little difference in spending on home ownership, furniture and furnishings, amusement and vacations, and (as expected given that the sample consisted of intact families) funerals. Households surveyed later spent more on fuel and lighting, tobacco, and medical expenses.¹²

It is important to consider how factors other than the pandemic could explain the differences between households surveyed at different times. One concern is the seasonality of expenses. For example, if spending on fuel and lighting is higher in the winter months, then the survey’s conclusion in wintertime means that households surveyed later would be expected to spend more, regardless of the pandemic. However, because the survey measured spending over the previous twelve months, the differences between households observed at different times are not an artefact of the seasonality of consumption.

A second concern is the diversion of food to the war effort. The resulting food shortages could explain the lower observed spending on food among households surveyed later. However, such shortages also contributed to higher food prices, which, as Hall documents, grew by up to 50 percent per year for some items. If households surveyed later bought less food but at higher prices, the overall effect on spending is ambiguous. Spending would have been lower if food prices rose at a slower rate than the quantity of food purchased declined, and spending would have been higher if prices rose faster than quantity purchased declined.¹³

Although we cannot conclusively determine whether food shortages contributed to lower food spending among households surveyed later, we perform the following comparison to provide a suggestive answer. For eighty-seven individual food items, ranging from butter to wheat bread, the survey recorded each household’s quantity purchased and the amount spent on the item, allowing us to infer the average nominal price paid as spending divided by quantity. For each item, we calculate the percentage change in average quantity purchased between households surveyed in July and August 1918 and households surveyed in January and February 1919. We do the same for percentage change in average price paid. Figure 6 provides a scatterplot of these percentage changes along with a linear best-fit line. This line has a slope of −1.6, meaning that, on average, across the eighty-seven food items, every 1 percentage point increase in price over the survey period was associated with a 1.6 percentage point decrease in quantity purchased. Given that the percentage change in quantity purchased exceeded the percentage change in price, wartime food shortages and resulting price increases could explain reduced spending on food as the pandemic progressed.

A third concern is variation in prohibition across cities. By the start of the pandemic, the sale of alcoholic beverages was prohibited statewide in twenty-seven states and in parts of sixteen more. (The Wartime Prohibition Act further restricted the production and sale of alcohol starting in May 1919, and the Volstead Act fully prohibited the production and sale of alcohol starting in January 2020. Both nationwide changes took effect after the bls Cost of Living Survey.) We divide the sample into the fifty-seven cities where alcohol sales were prohibited and the forty-one cities where alcohol sales were allowed in 1918 (information about prohibition was compiled by Sechrist and is not available for Norfolk, Virginia). We estimate equation 1 for each sample. None of the cities where alcohol sales were prohibited were surveyed in July or August 1918, so we instead use the three-month groupings from panels (b) and (c) of Figure 3: July–September 1918, October–November 1918, and December 1918–February 1919.¹⁴

Column 1 of Table 4 presents the regression results using the ninety-nine-city sample. As is shown in panel (n) of Figure 5, spending on liquor was lower for households surveyed later, particularly those surveyed at the end of 1918 and early 1919. Column 2 presents the regression results for cities with prohibition in place. Average spending on alcohol varied by survey timing, but never by more than about $2 per household. Column 3 presents the regression results for cities where alcohol sales were allowed. Average spending was again lower as the pandemic progressed, but by at most $3, not the $6 in the full sample. This comparison suggests that at least part of the apparent difference in spending on liquor was, in fact, due to the sequence of cities surveyed—those surveyed earlier tended to allow alcohol sales, and those surveyed later tended to prohibit them.

We showed above that total spending on life insurance was lower among households surveyed later. Here we decompose this total spending on life insurance by type. Old-line life insurance consisted of whole-life policies, whose value the insured person could borrow against. For industrial life insurance policies, fees were collected weekly door to door. Fraternal life insurance was offered through religious, occupational, or other affiliation groups. We focus on these three most common types of life insurance; establishment and other types accounted for less than 6 percent of spending on life insurance.¹⁵

Panel (a) of Figure 7 shows differences in spending on each type, again estimated using equation 1. Households surveyed in July and August spent an average of $14.30 on old-line policies. Average spending was nearly the same for households surveyed late in 1919 and in early 1919. Spending on industrial policies was $28.20 among households surveyed in July and August and was $7 lower among households surveyed in January and February. Spending on fraternal policies was $5 on average for households surveyed in July and August and was $2.60 higher for households surveyed in early 1919. Because of their weekly door-to-door sales, industrial policies required the kind of close contact that may have contributed to the spread of influenza. Lower spending on industrial but not on other types of policies may have been caused by people avoiding close contact with others.

Were the differences between households surveyed at different times due to differences in supply, demand, or both? To answer this question, we separately measure changes in the quantity of insurance purchased and its average price. Panel (b) reports the share of households holding each type of insurance. Compared to households surveyed in July and August, the share with old-line life insurance was slightly lower for households surveyed later. The share with industrial life was lower among households surveyed later, by about 18 percentage points. The share with fraternal life insurance was higher among households surveyed later, by up to 4 percentage points. Panel (c) shows differences in the average amount of coverage held, by households with any insurance. Average old-line insurance coverage was lower in September and October compared to July and August, then rebounded among households surveyed later. The average industrial insurance coverage was $190 lower among households surveyed in January and February than in July and August. For fraternal insurance, average coverage was $310 higher among households surveyed in in January and February compared to July and August. Panel (d) shows little difference in the average price paid for old-line and industrial insurance. Households surveyed later paid a lower price for fraternal insurance.

Together, the findings in panels (b) through (d) show that there was little variation by survey timing in the amount of old-line insurance purchased and the price paid. For industrial insurance, the amount of insurance purchased was lower in cities surveyed later, but there was little difference in price. For fraternal insurance, the amount of insurance purchased was higher and the price paid was lower in cities surveyed later. These differences suggest similarity between cities in the market for old-line insurance, lower supply of and demand for industrial insurance among cities surveyed later, and greater supply of fraternal insurance. Again, industrial life insurance was sold door to door and fraternal life insurance was sold through affiliation groups. These findings are consistent with an explanation in which salespeople were less willing to go door to door (or were prohibited from doing so), households were less willing to receive such salespeople, and social organizations compensated by increasing supply.

An important additional change in the market for life insurance was the introduction of War Risk policies. These were life insurance policies offered to service members during World War I. It is possible that these policies increasingly crowded out supply of and reduced demand for other forms of life insurance. For example, greater availability of War Risk policies over the survey period could explain the lower quantity of industrial insurance purchased in cities surveyed later. Again, though, the bls Cost of Living Survey sample consisted of families in which the husband was present and employed as a civilian. These were not families for which War Risk policies would have been available. Unfortunately, we are not able to measure directly whether War Risk policies crowded out private life insurance for families with service members.¹⁶

To discover which factors best explain differences in spending during the pandemic, we use a two-step machine-learning process to identify the household characteristics that are jointly associated with spending differences. Households that spend similarly to one another are divided into two clusters across all individual spending categories. We use k-means clustering to group households into two clusters based on spending across all individual categories of spending. This clustering begins with two means, then adds cities to the cluster it is most like (as measured using Euclidian distance), recalculating the mean of each cluster after every city is added. This first step yields two clusters of 3,252 households and 9,479 households.

We then use a classification tree to identify the combinations of household characteristics that explain membership in the two clusters. Because the clusters are unbalanced, we triple each observation in the smaller cluster so that the clusters are roughly the same size. We then consider the following characteristics observed for each household: census region, race, person-years spent in the household by all members, age of the husband, age of the wife, and income. We use ten-fold cross-validation to prune the variables and identify the tree of optimal complexity. Households in the first cluster were generally located in the Northeast or South census regions, with a wife under age 35 or household income below $1,591. Those in the second cluster were generally located in the Northeast or South with a wife aged 35 or older and income above $1,591, or were located in the Midwest or West.

Figure 8 presents estimates of equation 1 for each cluster (because census region is used to define the clusters, we do not include it as a control variable in the regressions). Because the survey moved inland from the coasts, some regions have only a few cities observed in a given pair of months, so we group months into three larger samples rather than four smaller ones. We find little difference in average total household spending between households observed in the Midwest and West in July through September and those observed in the Northeast and South. As shown in panel (a), the difference in spending for households surveyed in October and November, and then again in December through February, was similar for both groups. Variation by region alone does not explain lower average spending among households surveyed later. As shown in panels (b) and (c), variation by wife’s age alone and income alone also do not explain lower average spending among households surveyed later.

The difference between groups emerges when we consider these three characteristics (region, wife’s age, and household income) jointly. For households in the Northeast or South with a younger wife or with lower income, spending was about $50 lower for households surveyed after mortality surged in October compared with those surveyed before the surge. For all other households (with an older wife, or higher income, or located in the Midwest or West), spending was more than $150 lower for households surveyed after mortality surged compared to those surveyed earlier. This comparison highlights the usefulness of cluster analysis for studying differences in spending between households surveyed at different times. By using a range of household characteristics to identify households that spend similarly to one another, we can identify a set of characteristics that jointly explain differences in average spending between households surveyed at different times, even though none individually explains the differences in spending.

This article provides microeconomic evidence from the bls Cost of Living Survey of spending differences during the 1918 pandemic at the household level. Compared to households surveyed before the pandemic, households surveyed later spent about 6 percent less on average. However, the sample that we analyze is not a stratified random sample of all Americans during the 1918 influenza and instead consists exclusively of households with an employed prime-aged male head with wife and children. As spending differences were largest among older, wealthier households in the Midwest and West—groups more likely to be represented in our sample—we expect that our estimates of spending differences might be attenuated by the inclusion of a more stratified sample. Again, that is what existing macroeconomic evidence shows: a mild and short-lived recession with a decline in consumption of around 2 percent.¹⁷

The advantage of our microeconomic evidence is therefore not that it captures the full experience of all Americans, but rather that it offers a rich set of information about income and various categories of spending at the household level. We show that differences in income explain differences in total spending during the 1918 influenza pandemic. Income, and therefore spending, was lowest in cities that imposed longer school closures, bans on public gathering, and other nonpharmaceutical interventions. Although there was little difference over time in spending on durable goods, spending on nondurable goods and services especially was lower among households surveyed later. Reluctance to interact closely with other people may explain this reallocation in spending. For example, spending was lower for life insurance sold door to door on a weekly basis but higher for other forms of life insurance sold more impersonally. Variation in both the price and quantity for different types of life insurance suggests differences between cities in both supply and demand.

Many of these findings resemble early evidence from the covid-19 pandemic. Given the origins of covid-19, Chinese consumers were the first to be studied. Using daily transaction data across 214 cities, Chen and coauthors found that offline consumption fell by 42 percent during the eight-week period between January and March 2020. Dining and entertainment (72 percent) and travel (64 percent) saw the largest declines. For the United States, Baker and coauthors use transaction-level household financial data to show that spending declined overall but increased for groceries, changes that were largest in states that issued shelter-in-place orders. Tauber and Van Zandweghe show that spending on services in the United States declined in the first year of the covid-19 pandemic, whereas spending on durable goods remained constant or even increased. Together, the evidence from 1918 and 2020 shows declines in spending associated with pandemic recessions. But, beyond these aggregate declines, pandemics shift consumer priorities. Unlike business cycle recessions—during which spending on durable goods typically declines—pandemics appear to be associated with a particular reallocation of spending away from services and nondurable goods.¹⁸

By focusing on household-level data, this study provides new insight into how income shocks and behavioral adjustments interacted during the 1918 pandemic. Spending declined primarily due to lost earnings rather than a fear of contagion, with the sharpest reductions in in-person services and goods requiring direct interaction. Pandemic-related disruptions—such as the decline in door-to-door insurance sales—coincided with shifts toward lower-contact transactions, illustrating how supply-side frictions reinforced changes in demand. Local public health interventions shaped short-term economic conditions: Cities with longer restrictions saw greater immediate income losses but lower mortality.

These results extend the broader view of a short-lived, moderate recession by detailing how households adjusted consumption in response to both economic and epidemiological shocks. The findings highlight the role of public health interventions in shaping household consumption and underscore how local variation in policy responses translated into differences in economic outcomes.

Future research could expand this approach by incorporating more diverse samples, including rural households, single-earner families that lost a wage earner, or households from different racial and immigrant backgrounds. The extent to which declines varied across these groups remains an open question. Another avenue is linking household expenditure data to local labor market conditions to assess how earnings shocks transmitted across sectors and regions. Such extensions would improve our understanding of the economic mechanisms underlying pandemic responses and inform policies aimed at mitigating household-level income shocks in future crises.

Table A1 lists the ninety-nine bls Cost of Living Survey cities, grouped by month of survey. The table also tallies the number of white and Black households in each city included in the final sample.

This article shows that, for households surveyed at different times, differences in average spending closely followed differences in average income. Throughout the entire survey period, average spending was between $60 and $95 lower than average spending, and households withdrew little from savings and borrowed even less. This appendix reports average values of several forms of saving. The first is two forms of precautionary saving—cash holdings on hand and in a bank. Panels (a) and (b) of Figure A1 report average values of these savings for households grouped by month of survey. For both forms of saving, households surveyed in November and December reported the largest average amounts. This comparison suggests that these forms of liquid saving were highest for families surveyed just after the pandemic peaked. This comparison is consistent with an explanation in which families, worried about pandemic-related economic uncertainty, engaged in precautionary saving. However, these average values are small, less than $20 combined, suggesting that precautionary saving was not a major component of household budgeting on average.

The second form of saving is two war-related savings options—liberty bonds and war savings stamps. The United States government began issuing liberty bonds in 1917 to finance its involvement in World War I, and two-thirds of households would come to own the bonds. The government sold the bonds in four campaigns, the last of which lasted into October 1918. The government also issued war savings stamps in small denominations, which could be accumulated and then exchanged for a bond. Households surveyed in July and August spent $53 on average on liberty bonds and $6 on war savings stamps. Among households surveyed in September and October, purchases of bonds were $20 higher, and purchases of stamps were $2 higher. Among households surveyed in November and December, purchases of bonds and stamps were $3 lower and $5 higher. Among households surveyed in January and February 1919, purchases of both were lower. As with cash on hand, these comparisons suggest highest average war savings for households surveyed in late 1918.¹⁹

Average spending was lower for households surveyed later, particularly for households surveyed in November and December 1918 compared to households surveyed in September and October 1918. A concern with this comparison is that, because the bls Cost of Living survey reached different census regions at different times, the difference in average spending between households surveyed at different times could simply reflect differences in spending between regions. This article addresses this concern by showing that, within each region, households surveyed later spent less on average.

However, any additional differences in sample composition over time could raise similar concerns. This appendix deals with two possibilities—family size and exposure to World War I. First, smaller households spent less on average. If households surveyed later were smaller, this characteristic alone could explain lower spending among households surveyed later. Second, although all households in the sample had a husband present and employed as a civilian, the ninety-nine survey cities had different exposure to World War I draft rates, casualty rates, and industrial production. For example, the war effort could have led to increased industrial output in some cities, driving up income and spending. If those cities were surveyed earlier, this difference could explain lower average spending among households surveyed later.

Panel (a) of Table A2 provides information by month of survey for the full sample, repeated from panel (a) of Table 2. Average spending values are shown in panel (a) of Figure A2. Panels (b) and (c) of Table A2 report average spending by month of survey for households above and below the median family size of four family members. (Family size is measured as the number of person-years spent in the household by all household members, allowing for fractional values if a family member was present for less than the whole year.) Although larger families spend more than smaller ones on average, the differences by month of survey are consistent: Regardless of family size, households surveyed in November and December spent about $100 less on average than did households surveyed in September and October.

The remaining panels of Table A2 and Figure A2 perform equivalent comparisons based on city-level exposure to World War I. We divide the sample of ninety-nine cities into above- and below-median county-level draft rate and above- and below-median casualty rate. We also divide the sample into cities located in states heavily involved in wartime industrial production (Michigan, New Jersey, New York, and Ohio) and all other states. (Unfortunately, we do not know of a county or city-level measure of wartime industrial output.) Again, in every case, average spending was lower for households surveyed in November and December compared to households surveyed in September and October. However, in cities with high draft rates, average spending was then about $100 higher for households surveyed in early 1919. In states heavily involved in wartime industrial production, no households were surveyed in 1919. As with comparisons by census region, these comparisons again show that the level of spending varies by household size and exposure to World War I, but, with the possible exception of cities with high World War I casualty rates and wartime production, differences by survey timing are roughly consistent across groups.²⁰

Column 1 of Table A3 presents the main findings estimated using equation 1, repeated from column 2 of Table 3. A concern with the estimates above is that they reflect pre-existing differences in spending levels between cities surveyed at different times. The closest comparable measurement of spending of which we are aware was another cost-of-living survey administered in 1888. This survey was administered only in the Northeast, Midwest, and South, and only records the respondent’s state, not city. Figure A3 shows the fifty-seven cities included in the 1917–1919 bls Cost of Living Survey that are located in states in which the 1888 survey was conducted.²¹

Column 2 of Table A3 reports estimates of equation 1 using only households in these fifty-seven cities. As with the full sample, average spending was lower for households in these cities surveyed in November and December compared to households surveyed in July and August. Column 3 reports estimates of equation 1, additionally including state-level average household spending in 1888 as a control variable to account for pre-existing differences in spending levels. Although the regression point estimates change, this regression again indicates that households surveyed in November and December spent less on average than households surveyed in July and August. Pre-existing differences in spending levels between cities surveyed at different times, at least as measured by statewide spending averages in 1888, do not explain the main findings.