Paper
Cutler, David, and Grant Miller (2005). Demography, 42(1): 1–22
Read the original paper → Opens at the publisher; use the DU library / JSTOR login if it asks for access.
A difference-in-differences study of the staggered introduction of water filtration and chlorination across major US cities in the early twentieth century, and what it did to total, infant and child mortality. The explainer below reconstructs the design and the headline decomposition.
An interactive, section-by-section guide to The Role of Public Health Improvements in Health Advances: The Twentieth-Century United States by David Cutler and Grant Miller, Demography (2005)
How much did clean water matter?
Between 1900 and 1936, big American cities began filtering and chlorinating their drinking water. Pick a city and watch what happened to typhoid.
1936
1.2typhoid deaths per 100,000
Values traced from the paper’s Figure 2, so they are approximate. Before-and-after averages are raw comparisons, not the paper’s estimates.
Typhoid was only one disease. Cutler and Miller use the uneven timing of filtration and chlorination across 13 cities to ask a bigger question: how much of the era’s fall in total, infant and child mortality came from clean water?
The great mortality decline
Covers the paper’s introduction
It merely removes the bulk of our excreta from our houses to choke our rivers with foul deposits and rot at our neighbors’ door. It introduces into our houses a most deadly enemy.A chemist on the water closet, quoted in Scientific American, July 24, 1869, and in the paper’s epigraph
No documented period in American history saw mortality fall as fast as the late nineteenth and early twentieth centuries. From 1900 to 1940, death rates fell by about 40%, roughly 1% a year, and life expectancy at birth rose from 47 to 63. Nearly all of the decline came from infectious disease, and the “urban penalty”, the extra mortality of living in a city, disappeared. The authors argue that forces outside medical care must have been at work.
The paper lists three candidate explanations, which are not mutually exclusive:
- Nutrition and living standards
- Robert Fogel and Thomas McKeown argued that better diets and rising incomes made people more resistant to disease.
- Private hygiene
- Campaigns urged households to wash hands and food, boil milk and breast-feed; infant and child health varies strongly with mothers’ education.
- Public health works
- Cities built water purification, sewers, refuse collection, milk pasteurization and meat inspection.
The third explanation had the least evidence behind it. Earlier studies linked sanitation spending or infrastructure to mortality, but big projects were often bundled with other reforms, launched in times of crisis, and built over decades, so their effects were hard to isolate. Cutler and Miller focus on two discrete technologies, water filtration and chlorination, whose start dates can be pinned down city by city.
Total mortality in the 13 cities fell by about 30% between 1900 and 1936. What share of that fall do you think clean water explains?
The headline results are large: clean water explains nearly half of the fall in total mortality in these cities, three-quarters of the fall in infant mortality and nearly two-thirds of the fall in child mortality, and it nearly wiped out typhoid fever. The authors put the return on the investment at about 23 to 1. Later work has questioned how large these effects are; Section 11 covers that debate.
Bottom line. If cheap public works, rather than income growth or private behavior, drove much of the historic mortality decline, that matters for how poor countries today spend their health budgets.
Dirty water in American cities
Covers “Background on Public Health Advances”: the disease environment and early water systems
In 1900, infectious diseases caused 44% of deaths in the study cities; by 1936, about 18%. The fall had started earlier: the crude death rate was about 22 per 1,000 in the 1850s, about 18 by 1900 and about 11 by 1940.
Share of all deaths, by cause, 1900 and 1936
Table 1, the paper’s major cities. The two group rows (“Major infectious diseases” and “Childhood infectious diseases”) are the paper’s totals. Diarrhea and enteritis were not reported consistently, so there is no 1936 figure.
Until the bacteriology revolution of the 1870s, most people held the miasma theory: disease came from foul-smelling vapors. People noticed that bad smells went with sickness, so concern about sewage came before anyone understood germs, but it did not produce clean water. Philadelphia built the first large municipal water system at the start of the nineteenth century, and other cities followed, often after years of squabbling over sources. The water still carried disease.
Sewers made things worse. Built to carry storm water, they clogged once water closets spread in the 1870s, and critics called them “elongated cesspools”. Many cities emptied their sewers upstream of their own water intakes, or into lakes near them, so the cities with the most extensive sewers had the greatest potential to pollute their drinking water. Even careful cities drank the untreated sewage of towns upstream.
The urban penalty: how much higher city mortality was than rural mortality
From seven states with good data before 1900. In 1900, life expectancy at birth for white males was also 10 years higher in rural areas than in cities.
Bottom line. Growing cities had built water and sewer systems that spread waterborne disease, and the damage fell hardest on infants and young children.
Filtration and chlorination arrive
Covers “Clean Water Technologies” and “Adoption of Clean Water Technologies”
Filtration was first designed to clear up cloudy, discolored and bad-tasting water, not to fight disease. It was used in the United States as early as 1872, in Poughkeepsie, New York, but was not common before the 1890s. Filtration removed many bacteria but not all, and of the disinfection methods tried, chlorination was the cheapest. The first large-scale chlorination came in 1908 at the Boonton Reservoir of Jersey City’s waterworks.
How cities got cleaner water
Adoption dates varied widely. Cincinnati, Philadelphia and Pittsburgh began building large filtration plants around 1900, while Chicago and Milwaukee waited decades. Chlorination spread faster: most big cities started within ten years of 1908.
Which of the 13 cities had adopted each technology, year by year?
Drawn from Table 3, the same dates as the paper’s Figure 1. Sewage chlorination reached only one city in this period (Cleveland, 1922). Jersey City’s filtration date is printed as 1978; Chicago filtered after 1940 and Memphis adopted nothing before 1937.
The authors argue that the exact year each city adopted was close to arbitrary. Sanitarians fought for years or decades to persuade city councils, and even after a consensus formed, cities quarreled over whether the city or a private contractor should do the work and which filtration method to use. Philadelphia is typical: more than 20 years passed between its first water-quality studies in the 1880s and filtered water. Cheaper fixes, such as moving sewer outfalls below the water intakes, failed; corruption and bribery around private water contracts stalled the project through the 1890s; and after the money was approved, building the plants took nearly a decade.
Bottom line. Cities adopted filtration and chlorination at different times for reasons largely unrelated to their mortality that year, which is what the research design needs.
Thirteen cities, 37 years
Covers “Data and Sample Selection”
The United States had no national system of death records before 1933. From 1900, the Census Bureau published annual deaths by city, cause and age for a “death registration area” of 10 states plus registration cities elsewhere. The series runs to 1936; a new one began in 1937, and 1936 also comes just before modern antibiotics. Deaths from diarrheal disease were reported inconsistently, but typhoid fever, a marker of waterborne disease, was reported throughout. In 1900, diarrheal deaths outnumbered typhoid deaths by about three to one.
Intervention dates come from water-system censuses in engineering journals (1924, 1932 and 1943), other periodicals and histories. The authors started from every city with at least 100,000 people in 1900, plus a few with unusually good records, and kept only those with published dates for all four interventions, because phone calls to water departments produced conflicting dates and answers like “circa 1910”. That left 13 cities: Baltimore, Chicago, Cincinnati, Cleveland, Detroit, Jersey City, Louisville, Memphis, Milwaukee, New Orleans, Philadelphia, Pittsburgh and St. Louis. Each date is the year in which most of the city’s population first received the treated water.
Average death rates in the 13 cities
Table 2: mean across cities, with the whiskers showing one standard deviation either side. The table’s heading says “deaths per 1,000”, but the text makes clear the rates are per 100,000.
The cities also changed in other ways, which is why the regressions control for population structure. On average they nearly doubled in size between 1900 and 1940, and the foreign-born share halved:
Selected rows of Table 4, from the decennial censuses. Values between censuses are interpolated linearly.
Bottom line. A small panel, but with precise, published adoption dates and consistent annual mortality data by cause and age.
The research design
Covers “Empirical Strategy”
The design is difference-in-differences: compare how mortality changed in a city when it adopted clean water with how it changed, over the same years, in cities that had not. Anything that moved all cities at once, such as the spread of hygiene knowledge, drops out; so does anything fixed within a city, such as a warm climate. A confounder would have to change in the same cities in the same years as filtration and chlorination.
Line the cities up by when they filtered
Typhoid deaths per 100,000 for the nine cities that filtered between 1900 and 1936, traced from Figure 2, on a logarithmic scale, like the regressions’ logged mortality. Year 0 is the filtration year in Table 3.
Formally, for city c in year t, the paper estimates equation (1):
Equation (1), term by term
Tap any part of the equation.
Pick a term to see what it does.
Which diseases should respond? Typhoid and diarrhea spread through water directly. But around 1900 Hiram Mills in Massachusetts and J. J. Reincke in Hamburg independently noticed that overall mortality fell by more than waterborne deaths alone when filtration began. This “Mills–Reincke phenomenon” is usually explained by dirty water weakening the immune system, which would make other infections deadlier. Chronic diseases such as cancer should respond little, at least at first.
What could go wrong, and how the paper responds
Bottom line. The effect is identified by abrupt changes in each city’s mortality in the year it adopted, compared with cities that had not yet adopted.
How much did clean water do?
Covers “Results”: Figure 2 and Table 5
Figure 2, the typhoid series in the hero above, shows no general rise in typhoid before cities adopted, which argues against cities acting only after a bad outbreak, though some cities show a dip in the years just before. Table 5 turns this into estimates. Because mortality is logged, the paper reads each coefficient as a percentage change.
The effect of clean water on mortality
Table 5, with 95% confidence intervals from Huber–White standard errors; stars in the paper are noted in each row. Values are in log points, with the exact percentage change, eβ − 1, in brackets. The paper reports the joint effect with an F statistic; its interval here uses the standard error implied by that statistic, |β|/√F. The two lead rows are indicators for the five years before adoption.
Filtration alone reduced typhoid by about 46%, total mortality by 16%, infant mortality by 43% and child mortality by 46%. Chlorination alone shows no detectable effect. The positive interaction for typhoid and total mortality suggests the two were substitutes: once water was filtered, chlorinating it added less. The authors caution that this partly reflects timing. Most cities filtered first, and cheap chlorination spread so quickly that there is little variation left to measure it.
Combining both technologies, clean water cut total mortality by about 13%, while total mortality in these cities fell by 30% between 1900 and 1936. That is the source of the headline share: 13 ÷ 30 ≈ 43%. The same arithmetic gives 74% for infants and 62% for children.
Bottom line. Filtration produced large, sharp drops in typhoid, infant and child mortality. The “43%” headline uses a simple approximation; measured consistently, the share is somewhat smaller, especially for infants and children.
Beyond typhoid
Covers “Results”: Table 6 and the decomposition of the total effect
Typhoid alone cannot explain the drop in total mortality. The fall in typhoid deaths due to clean water accounts for only about 2% of the total decline. If deaths from all waterborne diseases were about three times typhoid deaths, as in 1900, they account for about 8%. So where did the rest come from? Table 6 runs the same regression for other causes of death.
Effects on other causes of death
Table 6, in log points, with 95% confidence intervals; stars in the paper are noted in each row. Axis clipped at ±2; smallpox, with only 95 observations, has intervals far wider than the axis. Malaria, smallpox and measles drop years with no deaths.
Among diseases reported consistently, pneumonia, tuberculosis, meningitis and diphtheria respond to clean water, consistent with the Mills–Reincke phenomenon. Cancer and diabetes, the only chronic diseases reported throughout, do not, which is the placebo test one would want.
Where does the 43% come from?
Percentage points of the total mortality decline, from the last column of Table 6 and the text. The paper’s text gives meningitis 5 and tuberculosis 6, the reverse of its Table 6; this chart follows the table. Either way, specific causes account for 32 of the 43 points.
Bottom line. Most of clean water’s effect on total mortality came through infectious diseases that are not waterborne, while chronic diseases were unaffected.
Did the drop come first?
Covers “Specification Test”: the lead dummies in Table 5 and Table 7
If adoption timing really is as good as random, mortality should not behave unusually just before a city adopts. Rising mortality beforehand would suggest cities acted after bad years, and the estimates might pick up regression to the mean. Falling mortality beforehand would suggest the estimates confuse ongoing declines with the effect of clean water.
None of the five-year lead indicators in Table 5 is significantly positive, but for filtration they are significantly negative for total, infant and child mortality. Mortality was already falling in the years before filtration. The authors’ response is Table 7, which estimates a separate effect for each year around adoption.
Effects by year relative to adoption
Table 7, in log points, with 95% confidence intervals from Huber–White standard errors. Year 0 is the adoption year. The specification also includes five-year leads and lags, sewage dummies, fixed effects, city trends and demographic controls.
For total mortality, filtration’s coefficient is between −0.05 and −0.07 in each of the four years before adoption and falls to −0.17 in the adoption year itself. The authors read this as an abrupt drop at the true date, and note two reasons for earlier declines: some cities phased filtration in over several years, and the coded date is when most of the population was served.
Bottom line. There is a sharp break in the adoption year, but also a smaller downward drift beforehand, which a careful reader should keep in mind.
Over time and across cities
Covers “Long-Run Effects and Behavioral Responses” and “Distributional Effects of Clean Water”: Tables 8 and 9
Would the effect grow or fade? It could grow if people learned to use clean water well, or if public investment made private precautions such as hand washing, boiling milk and careful food storage worth more. It could fade if clean water crowded out costly private precautions, or if the people it saved were frail and died soon after anyway.
The cumulative effect of both technologies, by years since adoption
Table 8: joint effects of filtration and chlorination in the adoption year, and the extra effects from five and nine years on, added up and converted to a percentage change. Several coefficients in the second typhoid column are printed without significance stars; the effects shown here use the joint effects, which are starred.
The typhoid effect grows strongly: a joint effect of −0.26 log points at first and a further −0.65 from five years on, which the paper reads as 26% and 65%, bringing typhoid close to eradication by 1936. The authors read growing effects as evidence against crowd-out or frail survivors, and as a hint that public health works and private health practices are complements. The pattern is much weaker for total mortality, where the later increments are small.
Who benefited most? Good measures of poverty are scarce in historical data, so the authors use the share of each city’s residents who were illiterate, interpolated between censuses, and interact it with the clean water indicators.
The joint effect of clean water, by the city’s illiteracy rate
Table 9: joint effect plus joint effect × illiteracy share, converted from log points to a percentage change. No confidence intervals are shown because the paper does not report the covariance needed. The paper does not report the cities’ illiteracy rates, so the slider covers a wide range, and values far from the cities’ actual rates are extrapolations. The values under the joint effects are labeled F statistics but include negative numbers, so they are probably t statistics.
Cities with more illiterate residents gained much more. This fits clean water helping the poor most, but the authors note they cannot tell whether poor people benefited more or whether everyone in poorer cities did. Reverse causation would work the other way, since falling mortality should raise literacy, not lower it.
Bottom line. The typhoid effect deepened over time, and clean water appears to have mattered most in poorer, less literate cities.
Was it worth it?
Covers “Social Rate of Return to Clean Water” and Table 10
To be conservative, the authors charge clean water with the value of each city’s entire water system, just under $300 million in 2003 dollars for the average big-city system in 1915, and assume it had to be rebuilt every 10 years: about $29 million a year. On the benefit side, they apply the 13% mortality reduction to the average city in 1915, assume each life saved would otherwise have ended at age 27, about half of life expectancy then, and count the 39 further years a 27-year-old could expect to live.
They value a year of life at about $100,000 today, then adjust it to 1930 incomes: real GDP per person grew about fivefold from $7,496 in 1930 to $37,600 in 2003, and the value of life rises faster than income. That gives $11,723 to $13,280 per person-year in 1930, and they use the lower figure.
Rebuild the rate of return
deaths averted a year, average city
person-years saved a year
annual benefits
annual cost
benefits per dollar of cost
cost per person-year saved
Table 10’s inputs, with the average city’s 1915 population and death rate implied by its figures (1,484 deaths averted ÷ 13.26% ≈ 11,190 deaths a year). The system’s value is taken as $290 million in 2003 dollars and spread evenly over its life, with no interest, as in the paper. Results match Table 10 up to rounding.
At the point estimate, clean water saved about 1,484 lives and 57,922 person-years a year in the average city, worth about $679 million against a cost of $29 million: a return of 23 to 1, with a 95% confidence interval of 7 to 40, or about $500 per person-year saved. These figures leave out lower illness and higher productivity.
Bottom line. Even with deliberately pessimistic assumptions about costs, benefits exceed costs many times over across the whole confidence interval.
Clean water today, and a correction
Covers the conclusion, plus research published since
The paper’s main results, for its 13 cities from 1900 to 1936:
- Typhoid fever
- Nearly eradicated; filtration alone cut it by about 46%, and the effect grew over time.
- Total mortality
- About 13% lower with both technologies; 43% of the decline, by the paper’s arithmetic.
- Infant mortality
- About 46 log points lower; 74% of the decline, by the paper’s arithmetic.
- Child mortality
- About 50 log points lower; 62% of the decline, by the paper’s arithmetic.
- Other causes
- Pneumonia, meningitis, tuberculosis and diphtheria fell; cancer and diabetes did not.
- Return
- About 23 to 1, or about $500 per person-year saved, in 2003 dollars.
The authors turn to developing countries. When they wrote, about 1.1 billion people lacked clean water and 2.4 billion lacked adequate sanitation; one of the Millennium Development Goals was to halve the share without both by 2015, and the UN declared 2005–2015 the “Water for Life” decade. They note that much aid has financed new water supply while neglecting sanitation, and that in 2000 more than a fifth of drinking water samples from existing systems failed quality standards, even as developing countries reported that over 90% of their supplies were adequately disinfected. Their message is that cheap disinfection can pay off enormously even without full sanitation.
The paper’s back-of-the-envelope for today
value of one year’s deaths averted
Footnote 24: 1,767,326 deaths from diarrheal disease worldwide in 2002 (WHO), 30 person-years lost per death and $100,000 per person-year. The paper rounds its result to about $160 billion.
Bottom line. Clean water clearly ended urban typhoid. How much of the wider mortality decline it explains is still debated, and the answer depends on data choices this paper made.
Test yourself
Eight questions on the paper and the debate around it
Glossary
Terms used in the paper and this guide
- Chlorination
- Disinfecting water by adding chlorine, which kills bacteria. Cheap; first used at scale in Jersey City in 1908.
- Death registration area
- The states and cities whose death records the Census Bureau compiled before national registration was complete in 1933.
- Difference-in-differences
- Comparing the change in an outcome for units that receive a treatment with the change, over the same period, for units that do not.
- Epidemiological transition
- The shift from infectious diseases to chronic diseases as the main causes of death.
- Filtration
- Passing water through sand and gravel to strain out particles and bacteria. Slow sand filters rely on gravity; rapid (mechanical) filters use pressure and chemical coagulants.
- Fixed effects
- Indicators for each city or year that remove everything constant within that city or common to that year.
- Huber–White standard errors
- Standard errors that remain valid when the variance of the error differs across observations.
- Lead (placebo) test
- Checking whether the outcome changes before a treatment starts, which a causal effect should not produce.
- Log points
- Differences in natural logarithms. For small values they approximate percentage changes; a change of b log points equals a (eb − 1) × 100% change.
- Miasma theory
- The pre-germ belief that disease came from foul-smelling “bad air”.
- Mills–Reincke phenomenon
- The observation that purifying water reduces deaths from diseases other than waterborne ones, by more than the waterborne decline alone.
- Social rate of return
- Total benefits to society divided by total costs.
- Typhoid fever
- A bacterial infection spread by food or water contaminated with human waste; a marker of waterborne disease in the paper.
- Urban penalty
- The higher mortality of city dwellers compared with rural residents in the nineteenth century.
- Value of a person-year
- The money value placed on one year of life, used to express mortality reductions in dollars.
Cutler, D., & Miller, G. (2005). The Role of Public Health Improvements in Health Advances: The Twentieth-Century United States. Demography, 42(1), 1–22. https://doi.org/10.1353/dem.2005.0002
This guide paraphrases the published paper and redraws its results from its tables. The typhoid series for 12 cities are traced from the paper’s Figure 2, so they are approximate; averaged across cities they match Table 2’s means for 1900, 1920 and 1936. Figure 2 marks Baltimore’s and Detroit’s filtration a year later than Table 3; this guide uses Table 3’s dates, as the regressions do. The later research cited is Anderson, Charles and Rees (2022), “Reexamining the Contribution of Public Health Efforts to the Decline in Urban Mortality”, American Economic Journal: Applied Economics 14(2): 126–57; Cutler and Miller’s comment (pp. 158–65); and the authors’ reply (pp. 166–69).