Disease and Development: Evidence from Hookworm Eradication in the American South^*

2.1 Hookworm Disease

Hookworm is an intestinal parasite that lodges itself in the human intestine and absorbs nutrients from the victim’s bloodstream. The symptoms of hookworm infection (or uncinaria) are lethargy and anemia. In rare cases, the anemia can become so severe as to cause death. The life cycle of the hookworm is dependent on unsanitary conditions. The nematodes lay their eggs in the intestine, but the larvae are passed out of the digestive system in feces. Hookworm is therefore transmitted through skin contact with infected fecal matter. The larvae then burrow their way in through the skin. The lifespan of a hookworm is much shorter than that of a human, and so continuous reinfection is required to generate any sustained worm load.

There are two angles for managing hookworm: treatment and prevention. The treatment consists of simply taking a deworming medicine. Preventative measures include limiting skin contact with polluted soil (through the use of shoes, for example) and dealing with excrement in ways that minimize soil pollution in the first place (e.g., the use of sanitary latrines).

2.2 The Eradication Campaign

The Rockefeller Sanitary Commission for the Eradication of Hookworm Disease was formed in 1910 with the donation of one million dollars by John D. Rockefeller. Some years before, an American doctor (Charles W. Stiles) had recognized hookworm symptoms in Southerners. Through intermediaries, Dr. Stiles had convinced Rockefeller that taking on hookworm was a good foray into large-scale charity. The Commission began by conducting surveys of hookworm-infection rates among children across the region. The RSC surveyed over six hundred counties in the South and found hookworm infection to be over forty percent among children.

Soon after, the treatment campaign began. First, the RSC sent teams of health-care workers to counties to administer and dispense deworming treatments free of charge. RSC dispensaries visited a large and mostly contiguous fraction of the South and the campaign treated over 400,000 individuals with deworming medication.¹ Second, the RSC sought to educate doctors, teachers, and the general public on how to recognize the symptoms of hookworm disease so that fewer cases would go untreated. Another part of this publicity campaign included education about the importance of hygiene, especially with regard to the use of sanitary privies. In this period, oftentimes even public buildings such as schools and churches did not have such hygienic facilities. Follow-up surveys conducted afterward showed a substantial decline in hookworm infection [RSC, 1915]. Although the stated goal of eradication was not achieved, the hookworm-infection rate of the region did drop by more than half, and fewer extreme cases of the disease went unnoticed and untreated.

Because the deworming treatments are short-term solutions, eradication requires (a) sustained monitoring (and treatment as needed) and (b) a reduction in the probability of reinfection. Follow-up efforts by private and governmental actors likely played a key role in consolidating the gains from the RSC and continuing the progress toward complete eradication.² Even after the RSC formally disbanded, the effort to eradicate hookworm infection continued. State governments ramped up their funding of anti-hookworm campaigns as the RSC was winding down. Local and state governments eventually took over some of its activities. The successor to the RSC, the Rockefeller Foundation’s International Health Board (IHB), continued to be involved at a lower level of funding. The IHB sponsored a handful of demonstration projects of the “intensive method,” which combined the deworming treatments and publicity campaigns of the RSC with technical assistance in building latrines at homes and public buildings. The state boards of health largely adopted this method and applied it to a degree throughout their jurisdictions. Harder to measure, but of considerable importance, the hookworm problem had entered into the public consciousness.

2.3 Testimonials Following the Campaign

Anecdotal evidence suggests that the RSC had an impact on human capital. Periodically educators would write the Commission thanking it for its efforts and describing the improvements following hookworm treatment. The following letter is from the school board of Varnado, La. [RSC, 1912].

As a result of your treatment for hookworm in our school, we find that children who were ranking fifth and sixth in their classes now rank second and third. Their lessons are not so hard for them; they pay better attention in class and they have more energy. […] In short, we have here in our school-rooms today about 120 bright, rosy-faced children, whereas had you not been sent here to treat them we would have had that many pale-faced, stupid children.

Farmer [1970] relates the following testimonials from the same period:

Teachers, school officials, and editors continued to be amazed at the difference in children after treatment for hookworm disease. A. J. Caldwell, Principal of Hammond High School in Louisiana, wrote that there was a decided improvement in the students in his school. One girl, who was in the fifth grade and did not attend school regularly because she was so pale and weak, started regaining her color and strength after treatment and finished the school term at the top of her class. C. C. Wright, Superintendent of Schools in Wilkes County, North Carolina, was an ardent supporter of the eradication program after examination of the pupils in his district revealed over 50 percent infection. Treatment cured the majority of these cases and the quality of performance in the county schools was raised considerably.

Typical of school officials’ attitude was that of W. H. Smith, State Supervisor of Rural Schools in Mississippi, who was thoroughly convinced that the economic prosperity of the people and the progress of educational development of the state depended largely on the successful eradication of the hookworm. The mental and physical growth of hundreds of children was evident. Smith asked for expansion of the program so that the thousands of children who were still suffering from mental and physical retardation might be saved. An editorial in a county newspaper of Hardin County, Texas, congratulated the County Commissioners for appropriating $300 toward the expense of operating five dispensaries throughout the county. According to the editorial, “300 dollars was never spent for a better purpose.”

And a report [RSC, 1915] describes the experience of one community in Virginia (in the sandy-soil Tidewater area of the state):

Henry Thrift, of Village, Va., [the school’s headmaster] told me in simple words an appealing story of how the treatment of these children had transformed the school. Children who were listless and dull are now active and alert; children who could not study a year ago are not only studying now, but are finding joy in learning. These children were born of anemic parents; were themselves infected in infancy; for the first time in their lives their cheeks show the glow of health. With this has come a new light to the eye, a new spring to the step, a new outlook on life. All this shows itself in a new spirit in the school. […] Some of the 45 children who had never attended school, having been treated, have come in during the year. Others have declared their intention to enter in the fall.

2.4 Identification Strategy

The first factor for identifying the effect of the hookworm-eradication campaign is that different areas of the South had distinct incidences of the disease. Hookworm larvae were better equipped to survive in areas with sandy soil and a warm climate. Broadly, this meant that the residents of the coastal plain of the South were much more vulnerable to infection than were those from the piedmont or mountain regions. Populations in areas with high (pre-existing) infection rates were in a position to benefit from the newly available treatments, whereas areas with low prevalence were not. This heterogeneity allows for a treatment-control strategy.

Second, the initiation of the campaign by the RSC was largely a function of factors external to the Southern states;³ The eradication campaign was made possible by critical innovations to knowledge: understanding how the disease worked and more importantly recognizing its presence. This contrasts with explanations that might have troublesome endogeneity problems, such as changes in government spending or positive income shocks in the infected areas. But even with the knowledge of the hookworm problem, there would have been formidable obstacles to taking action. The public-health infrastructure of this period was extremely limited. Rockefeller’s donation was an important precondition for attacking the problem.

Thirdly, the anti-hookworm campaign achieved considerable progress against the disease in less than a decade. This is a sudden change on historical time scales. Moreover, I examine outcomes over a fifty-year time span, which is unquestionably long relative to the five-year RSC intervention.

These factors combine to form the central variable in the present study:

More compactly, call this variable ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$), where j indexes the geographic area and t indicates j the year. The variable $H_j^{\mathit{pre}}$ denotes the level of hookworm infection among school-aged children j in area j at the time of the RSC’s initial survey, and Post_t is a dummy variable indicating whether year t is later than the active years of the RSC campaign (1910–1915).

I compare the evolution of outcomes (e.g., investment in human capital) across counties with distinct hookworm-infection rates, in order to assess the contribution of the eradication campaign to the observed changes. Estimating⁴ equation (1) measures the reduced-form differences by pre-eradication hookworm for some outcome Y_ijt for person i in area j at time t.

in which Y_ijt is the outcome of interest, the δ_t are time dummies, the δ_j are geographic fixed effects, and X_ijt is some vector of individual-level controls.⁵

How realistic is the assumption that areas with high infection rates benefited more from the eradication campaign? Resurveys found a decrease in hookworm infection of thirty percentage points across the infected areas of the South. Such a dramatic drop in the region’s average infection rate, barring a drastic reversal in the pattern of hookworm incidence across the region, would have had the supposed effect of reducing infection rates more in highly infected areas than in areas with moderate infection rates. Figure I presents data on this issue.⁶ The basic assumption of this section — that areas where hookworm was highly endemic saw a greater drop in infection than areas with low infection rates — is born out across states and across counties.

2.5 Related Studies

Several pieces of contemporaneous evidence also complement the results from the present study. Summarizing evidence from randomized trials in developing countries, Dickson et al. [2000] find mixed evidence of the effect of hookworm infection on schooling, whereas Miguel and Kremer [2004] estimate the impact to be strong and positive using an experiment in Kenya. Miguel and Kremer argue that infection spillovers contaminated the earlier mixed results. Specifically, previous studies often randomized within schools, but fail to deal with the reinfection problem. As a result, they argue, follow-up surveys often found limited effects; no increase in school attendance is observed because there is little persistent difference in infection rates between control and treatment groups. (Philipson [2000] also discusses this evaluation issue in a general context.) Small-scale interventions that do not manage reinfection are therefore less likely to succeed. The RSC intervention, on the other hand, was of such a scale that it brought about large reductions in hookworm disease in entire areas, and these gains were further consolidated through improvements in sanitation. In the context of economic development, it is precisely such a large and persistent reduction in disease burden that we would wish to consider. This is a further advantage of examining the RSC campaign: enough time has passed since its inception that we can assess its long-term consequences.

There are several other recent studies that consider the early-twentieth-century reduction in tropical diseases in the American South. While childhood effects are the focus of this study, Brinkley [1994] examines the role hookworm played in agricultural productivity. He finds a negative conditional correlation between hookworm infection and agricultural income per capita, although he does not specifically use the RSC intervention to identify this relationship. Bleakley [2002a] examines the interaction between malaria and hookworm. Bleakley and Lange [2004] consider the hookworm-related increase in returns to schooling in a quantity-quality model, and examine the fertility behavior of households in response to hookworm eradication. Additionally, an earlier version of the present study was found in the first chapter of Bleakley [2002b], and those results were partially summarized by Bleakley [2003]. The latter, summary piece discussed results for schooling and income, but did not treat literacy or regular school attendance nor did it consider whether hookworm-related changes in adult income were an artifact of some alternative time-series process, nor did it report results considering possible omitted variables.

4.1 Main Results

In this subsection, I estimate the effect of the hookworm-eradication campaign using equation (1) above. This involves comparing the census microdata on human capital before and after the RSC. As described above, the variable of interest ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$) is the interaction of pre-period hookworm infection, $H_j^{\mathit{pre}}$, with a dummy, Post_t, indicating whether the year comes after the RSC.

Using the two-period comparison, I find a substantial increase in school enrollment among children living in areas that had high levels of hookworm infection in 1910. This is true in absolute terms and also relative to areas with lower levels of infection. Specifically, the coefficient on ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$) implies that a county with a 1910 infection rate of 50% would experience an increase in schooling enrollment of five percentage points, relative to a county with no infection problem. In 1910, the mean of school enrollment in the sample was 0.78 and the standard deviation across SEAs was 0.11. Moreover, the standard deviation of hookworm infection rates across SEAs in 1910 was 0.23; so a one-standard-deviation increase in lagged hookworm infection is associated with a post-RSC increase in schooling enrollment of one quarter of a standard deviation.

These empirical results are presented in Table III. Estimates of the variable of interest, pre-period hookworm × post, are displayed for various outcomes and specifications. Panel A presents the estimates using the 1910 and 1920 censuses, which bracket the RSC intervention. In addition to the results on school enrollment mentioned in the previous paragraph, I estimate positive effects of hookworm eradication on full-time school attendance and literacy as well. Panel B contains similar estimates using the Census microdata from 1900–1950. (The literacy variable is not available in the later Censuses, so Column 3 is blank in Panels B-D; literacy results in Panels E-G use the 1910–20 Censuses.)

The surge in schooling attendance in high-hookworm counties coincided with the campaign for hookworm eradication. This is reassuring because it indicates that the differences estimated above were not due to pre-existing differential trends across the region. Such trends — coming from a number of possible factors — might have caused differential movements in school attendance even in the absence of the anti-hookworm campaign. However, my estimates of the effect of the reduced hookworm infection are robust to allowing for these trends.

The surge in school attendance in high-hookworm counties coincided with the campaign for hookworm eradication. This can be seen in Figure III.¹⁰ As shown in the graph, areas with more hookworm infection had lower levels of school attendance prior to the RSC, but these groups converge markedly thereafter. This lends credence to the view that this separation was caused by the anti-hookworm intervention.

The “timing” hypotheses can be further investigated by augmenting the regression specification in equation 1. Specifically, we can add an addition term to capture these possible pre-existing trends. I construct this term by interacting $H_j^{\mathit{pre}}$ with t, the time variable. The resulting equation is as follows:

in which the δ_t are time dummies, the δ_j are geographic fixed effects, and ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$) is the interaction of $H_j^{\mathit{pre}}$ with Post_t, as above. Differences between the areas that arise due to pre-existing trends will load onto γ, whereas differences that coincide with the anti-hookworm campaign will load onto β.

This comparison of “sudden shift” versus “trend” continues to attribute the increase in schooling among high-hookworm counties to the period following the RSC intervention. These results are seen in Panel C of Table III. Allowing for a differential trend by SEAs with different preexisting infection rates does not alter the main conclusion: the higher-hookworm counties saw relative increases in school attendance immediately following the RSC. This result is also robust to allowing the trends to be SEA-specific (Panel D).

The specification in Panel E contains controls for state-level shocks and policy changes, most notably the compulsory-schooling and child-labor laws that were imposed in the first half of the twentieth century. Since these shifts were at the level of state × year, this specification implements a simple fix to purge the estimates of this effect: including (state × year) fixed effects. Throwing out all of the cross-state variation yields estimated effects that are essentially unchanged.

Another concern is mean reversion across areas: if some counties had high hookworm infection and low schooling because of some (temporary) income shock, we might expect rises in school attendance in the following period even if hookworm had not affected the schooling decision. In Panel F of Table III, I incorporate the interaction of Post_t with 1910 average school attendance by SEA into the specification. Differential incidence of state policies (by average school-attendance rates) are also absorbed by the interaction of state × year dummies with lagged school attendance. There is evidence of mean reversion in schooling, but this mechanism does not account for the rise in schooling in higher-hookworm areas.

In Panel G, I re-estimate equation 2 using only state-level variation in the anti-hookworm campaign. Because the RSC did not attempt a systematic survey of hookworm across the whole country, I employ a measure of hookworm infection by state from Kofoid and Tucker (1921), who surveyed infection rates among army recruits. I match individuals to this variable using their state of birth. Because the full set of states is a much more heterogeneous sample, I also control for mean reversion as above.

Restricting the analysis to the state level excludes a substantial fraction of the useful variation: in all cases, the standard error on the estimate of ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$) is approximately twice those found above in the SEA-level analysis. There are two likely reasons for this: (i) I am reducing the number of geographic units and adjusting the standard errors to account for intraclass correlation; (ii) the standard deviation of infection rates across states is substantially smaller that the standard deviation across county groups.

On the other hand, point estimates of the effect of hookworm eradication are approximately the same magnitude as those in the county-level results. The result for enrollment is smaller than the estimates above, and we can reject neither zero nor the estimates from the county-group-level variation. On the other hand, attendance and literacy do show statistically significant responses to hookworm, with magnitudes that are larger than previous estimates.

The results indicate that, at the state level, the effect of hookworm eradication worked principally through the intensive margin of human-capital formation (literacy and full-time school attendance). This suggests that, if we examine these cohorts as adults, we will see increases in human capital, but the estimates (of years of schooling especially) may well be statistically insignificant. These results provide a natural benchmark for the cohort-based analysis in Section 6 below, since the retrospective-cohort analysis employs precisely the state-of-birth variation in hookworm.

4.2 Sensitivity Analysis

The finding that highly infected counties experienced surges in school attendance is not sensitive to controlling for a variety of alternative hypotheses. These include allowing for changes in health and health policy, educational resources, race and race relations, urbanization and land use, and parental background. I contrast these hypotheses with the effect of hookworm and the RSC by starting with equations 1 and 2 and adding plausible proxies for the supposed confounds. These results are found in Table IV. The baseline results from above are shown in Panel A, and the subsequent panels present specifications with additional controls. In every case, the added control variables are jointly significant at conventional confidence levels. (The aggregate control variables enter into the specification interacted with Post_t. All aggregate main effects are absorbed by the area fixed effects.)

These results are not sensitive to controlling for various health-related measures, as seen in Panel B of Table IV. First, I include the number of individuals examined by the RSC, per capita, to consider direct effects of the RSC campaign of dispensaries, which, in addition to hookworm treatments, provided hookworm exams. The stated purpose of the examinations was to determine if children had hookworm disease, but it seems unlikely that trained doctors and nurses would do absolutely nothing if they recognized some other disease. I also use data on sanitation collected by the RSC concurrently with its hookworm survey of Southern counties. Part of the Rockefeller campaign was a campaign of sanitation education for the whole region. It is therefore a reasonable alternative hypothesis that improvements in sanitation had benefits above and beyond simply reducing hookworm. Additionally, I include three variables measuring increased governmental involvement in public health: (i) the number of full-time health officers per capita; (ii) changes in health and sanitation spending by county governments over various intervals¹¹; and, because the federal government involved itself in sanitation efforts in the areas surrounding troop cantonments during World War I, (iii) WWI camp size per capita. Next, I account for the presence of malaria, the other major tropical disease in the South at the turn of the century. Finally, I incorporate the 1900–10 changes in total fertility rates, which might reflect a latent shift in parental attitudes towards the quantity-quality tradeoff.

In Panel C, I control for a number variables related to education and race, and obtain essentially similar estimates on the hookworm variable. If the results that I ascribe to hookworm eradication were instead due to some shift in local educational policy, the school-resources variables should correct for that. Consequently, I include logarithmic changes in the school term, teachers per school, schools per square mile, pupil/teacher ratio, value of school property, and educational spending. I use average literacy rates among children and adults to proxy for mean reversion. Additionally, I control for the estimated 1910–20 change in returns to literacy among adults, by SEA, as a measure of demand-side factors that influence the decision about human-capital investment. Another particular policy shift was the (partial) correction of the resource imbalance between white and black schools, as documented by Margo (1990) and Donohue et al. (2002). While much of this increase took place well after the decade of the 1910s, it is nonetheless useful to consider whether the earlier phases of this transition affect the estimates of the impact of hookworm eradication. The first thing to note is that the above estimates already incorporate (black × year) effects. However, since these extra funds were presumably directed towards with areas with larger black populations, I also allow for effects at the aggregate level. In particular, I interact the percentage black of the SEAs 1910 population with Post_t. I also control directly for the schools built through the philanthropy of Julius Rosenwald (see Donohue et al. (2002) for more on this program), which appear to have affected enrollment and attendance. Finally, I proxy for poor race relations using a measure of lynchings per capita. Adding these controls makes little difference for the estimated effect of hookworm eradication.

Hookworm was a relatively rural disease, but changes in school attendance for urban versus rural children are not responsible for the results presented above. In Panel C of Table IV, I allow the human-capital outcomes to change across counties differentially by the fraction of that area’s population living in an urban area, by changes in urbanization in the decade prior to the intervention, and by farm acreage and value per capita. Additionally, the high-hookworm counties tended to be in the coastal plain where crop mixes were different, so I control for the acreages per capita sharecropped and planted with cotton or tobacco.

In Panel D, I include controls for parental background, but these do not materially affect the estimated hookworm coefficient. I proxy for each parent’s income with the occupational income score, and include a binary indicator for whether that parent is present (missing parents getting an imputed income of zero). While these parental SES variables are generally highly statistically significant, their inclusion results in changes of the hookworm coefficient that are less than a standard error. Similar results are obtained when using literacy or Duncan’s socioeconomic indicator as the measures of parental background, as well as if allowing these parental variables to vary across year.

Finally, Panel E of Table IV presents the estimated coefficient on hookworm × Post_t when all of the above controls are included in the specification. The estimates are similar to the baseline.

4.3 Demographic Decompositions

The estimated relationship between hookworm and human capital was not simply concentrated on one particular demographic group, although there are noteworthy differences as well. These results are seen in Table V. For preteens and adolescents, the estimates for enrollment and attendance are close in magnitude, which suggests a balancing of two offsetting effects: younger children were more likely to be infected, but adolescents were closer to the margin of enrolling and/or attending regularly. The literacy results are different by age group but, because literacy is a stock variable, the larger coefficient for preteens means that children, following the eradication campaign, became literate earlier in life. Males and females were similarly affected by the intervention (results not shown).

There was also an important difference between how blacks and whites responded to the anti-hookworm campaign. Whites appeared to have positive responses to hookworm eradication by all three measures of human capital but, for blacks, the estimated effects of hookworm were uniformly larger. There are several candidate explanations for this result. One is that the general health of blacks was more sensitive to a given level of (own) hookworm infection. However, this explanation is inconsistent with existing medical evidence (Vance, 1932; Brinkley, 1994). The other possibility is that whites, because of higher average incomes and therefore better sanitary conditions, had lower rates of infection. Unfortunately, there is no direct published evidence on this hypothesis. Racial decompositions of infection data were apparently intentionally concealed by the RSC in the hopes of not generating controversy in what was already a racially charged climate (Ettling, 1981). A third explanation is that whites, who were more likely than blacks to go to school and be literate, simply had less scope to increase much along these measures of human-capital investment.¹²

The long-term consequence of these racial differences is less clear because, as noted by Margo (1990), the return to schooling was lower for blacks than whites during this period. Therefore, I revisit this issue in Section 6 when measuring the effect of childhood exposure to hookworm on adult income.

4.4 Interpretation

The estimates presented above imply plausible numbers for the effect of hookworm infection on school attendance. We can compare the reduced-form effect of ( $H_j^{\mathit{pre}}\times {\text{Post}}_t$) (about 0.09) to the estimated decline in infection as a function of the same variable (0.44).¹³ Dividing the first number by the second gives us the Indirect Least Squares (ILS) estimate of infection on schooling: 0.20. This indicates that a child infected with hookworm is 20% (i.e., percentage points) less likely to be attending school. Similarly, ILS estimates imply a 0.13 lower probability of being literate and 0.33 reduction in the probability of attending school full time.¹⁴

These estimates indicate that hookworm played a major role in the South’s lagging behind the rest of the country, as shown in Table VI. In computing the depressing effect of hookworm on the region’s human-capital accumulation, I multiply the ILS estimates from above with an estimate of the area’s hookworm burden. I assume a 40% regional hookworm-infection rate, as reported by the RSC. The resulting numbers account for around half of the human-capital gap.

Finally, a simple calculation shows that being hookworm free in childhood would have translated into substantial gains in income as an adult. I perform this calculation by accumulating the flow of school attendance over the sample ages. This is done simply by multiplying the ILS estimate times the age range of the sample (9 years). The estimates from above imply an increase of 2.1 years of school attended and 3.2 years of school attended full time. For the 10% return to schooling estimated in 1940, the 2.1 extra years of schooling would translate into a 21% wage gain. This number would presumably be a lower bound because the more intensive school attendance should increase the return to the inframarginal schooling as well. These extrapolations are confirmed with direct evidence in Section 6 below.

6.1 Results for Earnings, Schooling and Literacy

I start with graphical comparisons of the outcomes of different cohorts. First, I compare the 1939 wages among cohorts with different degrees of exposure to the hookworm-eradication campaign. Cohorts older than age twenty in 1910 were too old to have benefited directly from the RSC’s activities, and therefore I define them as unexposed. In contrast, I define fully exposed cohorts to be those born after 1910. Figure IV displays the difference in the log earnings between these two cohorts for each state of birth. I graph this wage difference against the state’s hookworm infection rate. As is evident from the graph, states with higher hookworm burdens saw larger increases in wages saw on average larger increases in wages.

Next, I focus on a simple parameterization of the cross-cohort comparison: the number of childhood years potentially exposed to the anti-hookworm campaign, times the hookworm intensity in the state of birth. Exposure to the RSC, Exp_ik, will be zero for older cohorts, rise linearly for those born in the nineteen years prior to 1910, and stop at 19 for younger cohorts.¹⁶ Nineteen is chosen because most individuals in this period would have completed their schooling investments by that age, and Smilie and Augustine (1926) found that hookworm infection was negligible at older ages. This regression equation is therefore

for which it should be noted that the main effects of H_j (hookworm in state of birth j) and Exp_ik (childhood temporal exposure of cohort k to the RSC) are absorbed by the cohort and state fixed effects. The additional demographic controls consist of indicator variables for age × black × female cell, as well as interactions of state-of-birth dummies with age, black, and female.

Children with more exposure to the campaign, by being born either later and in a state with greater treatment intensity, are more likely to be literate and earn higher incomes as adults. Results are mixed for years of schooling, but this is well within the range of normal statistical variation. Table IX contains these results. Panel A presents the estimates from the basic specification of equation 3.

The estimates do not appear to be an an artifact of mean reversion. The Southern states had markedly lower income before the Rockefeller campaign, and the hookworm problem was undoubtedly made worse by the low productivity of the region. If the oldest cohorts had high hookworm infection and low productivity because of some mean-reverting shock, we might expect income gains for the subsequent cohorts even in the absence of a direct effect of hookworm on productivity. I use data on labor earnings by state in 1899 from Lebergott (1964). I interact the natural logarithm of this wage measure with age and include the interaction in the even-numbered columns of Table IX. This analysis yields mixed evidence of mean reversion in the data, but the inclusion of these controls does not affect the main result: labor income and literacy is still estimated to be substantially higher as a result of the anti-hookworm campaign.¹⁷ In the analysis that follows, I estimate the hookworm coefficient both with and without this correction for mean reversion.

I estimate effects of hookworm eradication that are similar to the baseline when I incorporating a series of health, demographic, and educational variables into the analysis. These results are seen in Panel B of Table IX. Health controls (all measured prior to 1910 except as noted) include doctors per capita, the natural logarithm of spending by state boards of health, the fraction of WWI recruits rejected for health reasons (c. 1918), the child mortality rate, and the 1890–1932 change in the child mortality rate. Demographic controls include 1910 state-level rates for fertility, literacy, racial composition, urbanization, and unemployment (from 1930). I also control for changes in educational policy with the c. 1905–1925 changes in school term, (log) expenditures per pupil, pupil/teacher ratio, and (log) teacher salaries.

I argue that the earnings results are not contaminated by hookworm-induced changes in the probability of self-employment. The major difficulty in using the earnings data from the 1940 Census is that it was incomplete: labor income from self-employment was excluded. This would complicate the interpretation of the coefficient on earnings if the hookworm also induced changes in choice to be self-employed. I show in Panel C that the hookworm × exposure variable does not predict missing data on earnings. In results not reported here, I also estimate regression equations identical to those used for Panel B of Table IX, but with two new dependent variables: binary indicators for (i) self employment and (ii) non-wage/salary income being greater that fifty dollars. In doing so, I find no statistically significant relationship between the hookworm measure and self employment.¹⁸ Finally, I find evidence of hookworm-related increase in the total time worked (either for wage/salary or not), although, once mean reversion controls are included, the increased labor supply does not account for a large fraction of the earnings effect.

I also consider the role played by the quantity of and returns to of schooling in the wage results. Recall that the results for years of schooling were insignificantly different from zero. Consistent with this, controlling directly for education does not significantly change the estimated effect of hookworm treatment. We can nevertheless easily reject, for conventional returns to schooling, the hypothesis that the wage effect is due entirely due to a rise in education.¹⁹ However, the fact that I estimate increases in literacy without concomitant rises in the quantity of schooling suggests an alternative hypothesis: changes in quality. In particular, it may be that students spend the same number of years in school, but the time is better spent. For example, there might be less absenteeism, or students might be better equipped to absorb the material while in school. As was shown above, students were less likely to work while in school and were more more likely to be literate, as a consequence of hookworm eradication. This suggests that the return to schooling might have been raised by the hookworm intervention.

There is indeed evidence of an increase in the return to schooling as a result of the intervention. This can be seen in Panel D of Table IX. The crucial interaction is the triple: between years of schooling and the treatment-intensity variable (H_j × Exp_ik). (The additional second-order interactions are absorbed with a series of dummies for birth state × education and birth year × education. The first-order effects of education, state of birth, and year of birth are also absorbed with indicator variables.) The triple interaction is positive and statistically significant for labor earnings.

This hookworm-related change in the return to schooling can potentially explain a large fraction of the increase in earnings described above. This regression, by comparing individuals with different terminal levels of attainment, estimates the average marginal effect of schooling in the sample (and how it changes following hookworm eradication). If the intervention had similar effects on the return to inframarginal schooling, we can compute the overall contribution through this channel. Multiplying the triple interaction (0.0022) by the average years of schooling in the South (7.72) yields 0.0170, almost eighty percent of the coefficient on (H_j × Exp_ik) in the first row of Panel D. (Because the education variable was de-meaned before interaction, the second-order term is evaluated at the mean of education.) Moreover, I cannot reject the hypothesis that all of the earnings effects worked through the rising return to schooling.

These disparate results for human capital are consistent with the evidence from the sequential cross sections above. In Table III, I show that the effect of hookworm eradication in the cross-state evidence played out primarily on the intensive rather than extensive margin of education. Consequently, we would expect relatively little change in years of schooling reported. While years of schooling is an important input measure, the intervention had the effect that children attended school more intensively. It not surprising, then, to find an effect on an output measure like literacy.

Several differences emerge among demographic groups. Results estimated from subsamples of males, females, whites, and blacks are contained in Table IX, Panels E–H, respectively. For no subgroup is there a robustly significant relationship between years of schooling and (H_j × Exp_ik). Estimates for literacy, on the other hand, are positive and significantly different from zero for all demographic groups. Literacy responses are larger for females than for males, as well as for blacks than for whites, possibly because females and blacks had lower pre-existing literacy rates. The estimate of (H_j × Exp_ik) in the earnings equation yields a positive and significant number for males, while the result for females is not sensitive to the inclusion of a mean-reversion control. Whites and blacks show similar earnings responses, on the other hand. This may be because blacks, while gaining more on measured human capital, faced lower returns to skill in the labor market.

As a further check, I assemble information on states’ laws covering compulsory schooling and child labor for various years. These data are based on the work of Acemoglu and Angrist (2000), and I copy their variables verbatim for the years 1914 and forward. Then, using similar methodology, I extend the coverage back to 1894, where possible. (Details on the extension to earlier periods are found in the Data Appendix.) These data consist of two sets of variables: CA and CL. The CA variables refer to laws on compulsory school attendance and the CL variables to laws on child labor restrictions. The base indices are constructed to be in units of implicitly required years of school attended. The dummies used in the regressions are coded as follows:

Compulsory-schooling data are then assigned according to the laws in force in the state of birth of the individual when he/she was 14 years old.²⁰

Controlling for measures of compulsory-schooling laws does not significantly affect the estimates of the effect of hookworm eradication. These results are seen in Table X. In Panel A, there are controls for compulsory-attendance and child-labor laws. The CA dummies show generally positive effects of compulsory-attendance laws on the income and human-capital variables. The CL, however, are mixed in sign. The CL11 variable is excluded from the literacy regressions for lack of within-state variance in that sample. Because I judged the early CA data to be of higher quality, in Panel B I include only the CA dummies as controls for compulsory-schooling laws. In neither Panel do the estimates of the effect of hookworm change substantially from those found above in Table IX.

Similar results are obtained after recoding the early child-labor laws to include only those laws with broad sectoral coverage. These new estimates are shown in Panel C of Table X. Before 1914, many of the Southern states had in force laws restricting youth employment in mines and manufacturing. But these industries represented, at the time, small fractions of state employment. There is therefore the nontrivial concern that such laws provided little to no constraint on the behavior of children desiring to work in the agriculture. Consequently, I also created an alternate coding of the CL variable that indicated the presence of child-labor restrictions across all sectors of the economy.

6.2 Cohort-Specific Relationship Between Income and Pre-Eradication Hookworm

In this subsection, I show that the shift in the hookworm-income relationship coincides with childhood exposure to the eradication efforts. This can be seen graphically, and I also provide statistical tests that indicate that the shift is indeed coincident with exposure to eradication rather than with some pre-existing trend or autoregressive process.

I use two proxies for labor productivity that are available for a large number of Censuses. The occupational income score and Duncan socioeconomic index are both average indicators by disaggregated occupational categories that were calibrated using data from the 1950 Census. The former variable is the average by occupation of all reported labor earnings. The measure due to Duncan (1961) is instead a weighted average of earnings and education among males within each occupation. Both variables can therefore measure shifts in income that take place between occupations. The Duncan measure has the added benefit of picking up between-occupation shifts in skill requirements for jobs. Occupation has been measured by the Census for more than a century, and so these income proxies are available for a substantial stretch of cohorts.

I combine microdata from ten Censuses to construct a panel of average income by cohort. The base sample consists of native-born males in the IPUMS and North Atlantic Population Project (NAPP, 2004) datasets between the ages of 25 and 55, inclusive, for the census years 1880–1990, which results in year-of-birth cohorts from 1825 to 1965. The individual income proxies are projected on to dummies for year-of-birth × Census year (cohorts can appear in multiple Censuses in this design). I then take the average residual from this procedure for each cell defined by year of birth and state of birth.

For each year of birth, OLS regression coefficients are estimated on the resulting cross section of states of birth. Consider a simple regression model of an average outcome, Y_jk, for a cohort with state of birth j and year of birth k:

in which β_k is year-of-birth-specific coefficient on hookworm, X_j is a vector of other state-of-birth controls, and δ_k and Γ_k are cohort-specific intercept and slope coefficients. I estimate this equation using OLS for each year of birth k. This specification allows us to examine how the relationship between income and pre-eradication hookworm (β̂ _k) differs across cohorts.

I start with a simple graphical analysis using this flexible specification for cross-cohort comparison. Figure V displays a plot of the estimated β_k. The x axis is cohort’s year of birth. The y axis for each graphic plots the estimated cohort-specific coefficients on the state-level hookworm measure. Each cohort’s point estimate is marked with a dot. The top row of graphs contain estimates from the basic specification, in which this state-of-birth average residual is regressed on to hookworm infection, Lebergott’s measure of 1899 wage levels, and a dummy for the Southern region. The bottom row displays estimates from the “full controls” specification, which, in addition to the basic variables, contains the various control variables from Table IX, Panel B.²¹

Because hookworm was principally a childhood disease, cohorts that were already adults in 1910 were too old to have benefited from the reduction in hookworm. On the other hand, later cohorts experienced reduced hookworm infection during their childhood. This benefit increased with younger cohorts who were exposed to the RSC’s efforts for a greater fraction of their childhood. The dashed lines therefore measure the number of years of potential childhood exposure (defined above) to the Rockefeller Sanitary Commission’s activities. (The line is rescaled such that pre-1880 and post-1930 levels match those of the β̂k. The exposure line is not rescaled in the x dimension.) Cohorts born late enough to have been exposed to the campaign during childhood generally have higher income than earlier cohorts, and this shift correlates with higher potential exposure to the RSC.

Formal statistical tests confirm that the shift in the income/pre-period hookworm relationship coincided with exposure to hookworm eradication, rather than with some trend or autoregressive process. This can be seen by treating the estimated β_k as a time series and estimating the following regression equation:

in which Exp_k is exposure to hookworm eradication (defined above), the kⁿ terms are nth-order trends, and Φ (L) is a distributed lag operator. To account for the changing precision with which the generated observations are estimated, observations are weighted by the inverse of the standard error for β̂k. Table XI reports estimates of equation 5 under a variety of order assumptions about trends and autoregression. The dependent variables are the cohort-specific regression estimates of income proxies on hookworm that are shown in Figure V for the specifications with broad sets of controls. Panel A contains estimates using the occupational income score as a proxy for income. The estimates on the exposure term are broadly similar across specifications, and there is no statistically significant evidence of trends or autoregression in these β_k. When the Duncan SEI is used instead (Panel B), there is evidence of a downward trend, but estimates of the exposure coefficient are stable once this is accounted for.

Table XII reports estimates of equation 5 for alternative specifications and constructions of the underlying sample. Panel A repeats the result from above. For Panel B, I restrict the base sample of Census data to a narrower age range, when individuals are more likely to be working and in a job that reflects their permanent income. Coefficients rise for both income proxies. Major revisions to the coding of occupations occurred with the 1980 and 2000 Censuses, but results are not sensitive to including data from after these breakpoints, as seen in Panels C and D. Panels E and F zoom in to narrower ranges of year of birth. Restricting the analysis to one hundred and then forty years of birth cohorts yields similar estimates on exposure, albeit with larger standard errors. Casual inspection of Figure V suggests why: it becomes more difficult to distinguish between exposure and trend when viewing a narrower set of birth years. Indeed, estimates using the occupational income score are not significantly different from zero for the narrowest range.²²

6.3 Interpretation

In this section, I characterize the magnitude of the effect of the hookworm reduction in more easily interpretable units, and contrast the estimates with cross-areas differences in income per capita. For ease of the analysis that follows, I focus on the contrast between the cohort with zero childhood exposure to the RSC and the cohort with full exposure. For example, comparing these cohorts in two areas that were one standard deviation (within the RSC-targeted area) apart in hookworm-infection rates, we would expect wages to have increased 11% more in the area with the higher pre-period infection rate.

Using Indirect Least Squares (ILS), I estimate the approximate effect of childhood hookworm infection on adult wages to be around forty-three percent. Again, I compare the fully exposed and non-exposed cohorts to construct the estimate. The increase in wages as a function of the Kofoid measure of $H_j^{\mathit{pre}}$ is 0.32, when comparing zero to full RSC exposure. Since hookworm was largely eradicated in the time span considered, I regress the pre-RSC hookworm (reported by Jacocks at the state level for certain states) on the Kofoid measure and estimate that the decrease in infection rates as a function of $H_j^{\mathit{pre}}$ to be 0.748.²³ This yields an ILS estimate of −0.43 in natural-log terms. That is, I estimate that hookworm infection during childhood caused a drop of approximately 0.43 in log adult wages.

The ILS estimates using the occupational proxies for income bracket the wage result. The estimated shift in income related to RSC exposure was estimated in Table XI. I compare these estimated changes with their respective averages for men born in the South between 1875 and 1895, and then construct the ILS coefficient as above. For the occupational income score, I estimate that the proportional change in income related to childhood hookworm infection is −0.32. The same estimate for the Duncan SEI is −0.5.

These results point to changes in the returns to schooling as well. I compute the drop in returns to years of schooling due to hookworm to be approximately 0.047 = 0.0022 × 16/0.748, the ILS estimate for the changing returns to a year of education. This represents a substantial drop due to hookworm—around fifty percent of the estimated return to schooling in this period.

The estimated impact is large enough that it bears consideration in a macroeconomic context, although is clearly not so large that it unreasonably explains “everything.” The log-income gap between the North and the South at 1900 was approximately 0.75. For a 40% infection rate in the South and an effect of hookworm on wages of 0.43, we would expect a reduction in Southern incomes of approximately 17 log percentage points. In other words, some 22% of this income gap could be attributed to hookworm infection in the South. On the other hand, if we turn to contemporaneous evidence from developing countries, Miguel and Kremer estimate that the prevalence of intestinal-parasitic infection in the Busia region of Kenya is around 90% among school-aged children. Applying our estimates to this area of Kenya, we would expect a log-income gain of approximately 0.38 from a complete eradication of intestinal worms from the country. This would be enough to raise income per capita to match the level of Zimbabwe, but obviously well short of the 200–300 log percentage points needed to reach OECD levels.²⁴