Childhood Growth and Socioeconomic Outcomes in Early Adulthood Evidence from the Inter-War United States

Abstract

Childhood malnutrition and its later life effects were important concerns in European and North American social policy in the early twentieth century. However, there have been few studies of the long-term socioeconomic consequences of malnutrition in childhood. We use a unique longitudinal dataset to provide credible causal estimates of the effects of childhood nutrition on early-adult educational and employment outcomes. Our dataset includes 2,499 children in Saint Paul, Minnesota who were weighed and measured in a national children’s health survey in 1918/1919 at 0–6 years of age. We observe those same people in the 1920, 1930 and 1940 U.S. censuses allowing us to measure childhood socioeconomic status (1920), adolescent school attendance (1930) and early-adult wages, and employment and educational attainment (1940). Examining variation between biological siblings, we are able to obtain credibly causal estimates of the relationship between childhood stature and weight and later life outcomes, largely canceling out the bias otherwise resulting from their joint correlation with genes and socioeconomic background. Because the initial survey located children within households, we identify the effect of differences in early childhood nutrition from differences between male siblings. Consistent with contemporary evidence from developing countries we find that being taller and heavier in early childhood is associated with better educational and labor market outcomes. Identifying all effects within families to control for socioeconomic background and family structure we find a standard deviation increase in BMI in early childhood was associated with a 3% increase in weekly earnings and that boys who were heavier for their age at the initial survey were 10% less likely to be unemployed in 1940. Taken together, these results confirm the importance of investments in early life health for later-life outcomes.

1. Introduction

The correlated development of body and mind has intrigued social scientists and educators since the emergence of mass education and modern research methods in the mid-nineteenth century. For example, John Dewey, the famous philosopher of education, noted that “… physical growth is not identical with mental growth but the two coincide in time, and normally the latter is impossible without the former.” In 1913 the United States Commissioner of Education wrote that there was “a growing conviction on the part of teachers and students of education that there must be a close relation between the physical and mental development of children and that this relation must be known and respected. As yet, however, we have little definite knowledge on the subject” (Baldwin, 1914). Acquiring more definite knowledge on the relationship between different facets of growth and the consequences for children’s later lives motivated a significant amount of pioneering longitudinal research in the United States (Sontag, 1971) and Europe (Martin, Gunnell, Pemberton, Frankel, & Smith, 2005) before World War II. Birth and early-childhood origin longitudinal studies have themselves grown in intellectual stature and become an important part of modern social science (Lawlor, Andersen, & Batty, 2009) since the 1940s.

Yet because the findings of birth cohort studies are specific to their cohort an important puzzle remains: what were the consequences of physical growth for children’s later lives in the early twentieth century? Significant changes in social policy, nutrition, and physical activity suggest that growth stagnation may have had more serious consequences in the past as compared to modern high-income countries with more generous social support policies. Thus, findings from post-World War II cohorts in high-income countries provide a valuable starting point for comparison. Beyond significant changes in income and welfare, nutrition and physical activity has changed significantly for both children and adults. Post-World War II cohorts have had access to increasingly calorie-dense and processed foods, and physical activity for transportation has declined significantly. Moreover, adolescents have reduced participation in the labour market since the early twentieth century. With reduced adolescent labour in physically intensive jobs children whose growth is slower in early childhood have more chances of catching up in their teens, which is likely to dampen the effects of early-life growth challenges in modern societies.

Modern evidence from low- and middle- income countries suggests that the socioeconomic consequences of compromised physical growth are significant (McGovern, Krishna, Aguayo, & Subramanian, 2017). Shorter stature and being underweight are both associated with a range of socioeconomic outcomes in adolescence and early adulthood including reduced educational attainment and earnings, and poorer marriage outcomes. The association with educational attainment is particularly important because it suggests a plausible causal pathway through the life course. Children who are shorter or lighter do worse in school, with subsequent impact on adult outcomes related to education such as marriage, job opportunities, and earnings.

Much of the robust modern evidence on the importance of stature and weight for adult outcomes comes from countries with a different childhood growth pattern than seen in early twentieth century North America or Europe. In many lower and middle income countries where this research has been conducted the average child is 1–2 standard deviations below contemporary healthy height-for-age norms between the ages of 2 and 6 (Victora, de Onis, Hallal, Blössner, & Shrimpton, 2010). By contrast, children in the early twentieth century United States were often 0.5 – 1 standard deviations below modern age-specific norms. Conversely, in modern high income countries where children’s height is closer to its genetic potential the link between stature differences and social outcomes runs through different channels, including higher ascribed social status for taller people (Lindqvist, 2012) rather than reflecting growth failure in shorter people. Moreover, the distribution and direction of childhood growth concerns in high income countries today is significantly different than it was a century ago. Compromised growth in high income countries is very low, with stunting rates below 2% in the United States. There are now significantly greater risks for childhood overweight and obesity than growth failure.

We build on the prior literature in two ways. First, we measure the consequences of reduced stature in a context of moderate nutritional shortfalls where the average child was 0.5 – 1 standard deviations below modern healthy norms. Second, we extend the temporal understanding of how nutrition affects adult outcomes by examining childhood growth in a significantly earlier birth cohort than previously studied. Following recent recommendations in the literature about causal inference in longitudinal studies we take an “outcome wide” approach, and discuss how the combination of results can be interpreted (VanderWeele, Mathur, & Chen, 2020). We find the direction and magnitude of modern estimates from contemporary low-income countries is broadly matched in the United States in the early twentieth century, suggesting these findings are relatively robust across quite different social contexts. Children who were significantly lighter for their age fared worse on important indicators of early-adult social status. Early twentieth century social policy makers were thus broadly correct in paying significant attention to improving childhood nutrition.

The paper proceeds as follows: First, we first discuss how childhood growth can be used as a measure of wellbeing and discuss its relationship to other measures of nutrition and health. We then give an overview of the longitudinal microdata developed for our analyses. The dataset we develop for the paper is both broadly representative of white children in the United States around 1920 and allows us to credibly identify causal effects of stature. Drawing from a city-wide household health survey our data allows us to enumerate siblings and identify the causal effect of childhood growth from the differences between siblings. We examine the consequences of reduced childhood stature for four adolescent and adult outcomes: 1) remaining in school past the required leaving age, 2) educational attainment, 3) adult earnings, and 4) adult unemployment duration towards the end of the Great Depression. Consistent with modern evidence, we find that being shorter or lighter for age in childhood is associated with reduced educational attainment, lower earnings, and longer spells of unemployment. Taken together our results confirm in a new context that compromised growth has important social and economic effects on individuals that persist over several decades.

2. Childhood growth as a measure of wellbeing

2.1. Growth and well-being

Stature is a useful population-wide indicator of nutrition, environmental conditions, disease, and physical workload (Bogin, 1999; Eveleth & Tanner, 1990). A significant literature has examined adult stature in the past as an indicator of well-being. Children’s stature can also be informative, but because children are growing a slightly different analytical approach is needed. Although genetics largely determine maximum potential stature, ongoing nutritional deficits during the growth years result in individual stature falling short of the maximum height that could be achieved at any given age. Growing children first use the calories they consume to replenish energy and fight disease, before being able to grow taller or wider. Children who are persistently sick or expending more calories than they take in will grow more slowly. If these deficits are widespread through the population, average completed stature in adults, or average height-for-age scores in children, will fall (Jelenkovic et al., 2011; Silventoinen, 2003).

There are three main causes of nutritional deficits. While not mutually exclusive because they can occur together for both individuals and populations, they are analytically separate: (1) limited available calories, (2) persistent infectious disease while energy intake is constant, and (3) elevated physical activity while food intake and illness are constant. The incorporation of energy expenditure and loss, alongside food intake, leads to the idea of “net nutrition” (Steckel, 1986). Because changes in children’s growth are subject to several influences, as described above, the existence of faster or slower growth does not in itself identify causes of change. Instead, changing childhood growth patterns show when different cohorts or sub-populations—presumed to have the same average genetic potential for growth—have grown at varying speeds. Once periods of changing growth are identified, the potential influences on changing net nutrition can be investigated to identify what caused growth to change. Our findings in this paper suggest the early twentieth century was an important period for American children’s growth: Children from better-off and smaller families were well-nourished and close to modern growth standards. However, a significant minority of American children experienced periods of food insecurity and poor nutrition and suffered growth shortfalls as a result.

Much of the literature on stature as a measure of population wellbeing in the past has focused on adults, and thus inferred growth patterns from terminal height. Samples drawn from military and prison sources have been important sources. In these settings institutional requirements to classify and identify individuals led to the collection of heights—and after the 1870s weights—from large numbers of people, mostly men. In the United States data on women are somewhat more abundant owing to the records of slave sales (Richard H. Steckel, 2008, 2009). Notably, the slave manifests include children as well as adult men and women. Steckel and colleagues, and Schneider, have used the slave manifest to show that slave children had a distinctive growth pattern. Slave children were deprived of nutrients at a very young age when they could not make a contribution as workers but caught up somewhat later in childhood when fed more because they were working (Schneider, 2017, 2020; Richard H. Steckel, 1986; Richard H Steckel & Ziebarth, 2016).

2.2. Children’s growth as a measure of well being

Even in circumstances less dire than slavery, children’s stature can reveal important trends in differential treatment. Differences in the growth pattern between boys and girls can be used to infer potential gender discrimination. In modern well-nourished societies with low fertility, and at least a nominal commitment to equal treatment, boys’ and girls’ growth patterns are very similar through age 10. Deviations from a pattern of equality can identify situations in which parents favour some of their children. Discrimination against girls suggests a reasonable starting hypothesis of relatively shorter girls than boys.

Empirical findings complicate this easy assumption in both historical and modern populations. Girls growth may indeed be more robust to nutritional deficits than boys (Bogin, 1999). Moreover European and North American research on children’s growth since the mid-nineteenth century has detected only slight differences between boys’ and girls’ stature (Schneider, 2016). Before the decline in child labour in Europe and North America in the early twentieth century, children’s earnings were an important share of family income. If boys were fed more to reflect their earning potential, one would see boys do better within the family and be taller for their age than girls in the same family. At a population level this would be revealed in the standardized height-for-age of boys being greater than girls across a large number of families. Conversely, child labour may retard growth potential, so children in the labour market may end up shorter. Children from families of different sizes may also have a differing growth pattern. Lower-order children may be taller for their age than younger siblings because more of the critical growth period for earlier children occurs with less competition within the household for resources. An offsetting effect is that younger children may benefit from the earnings of adolescent siblings in the labour market. Thus, multiple influences including gender, family and household composition, and labour market opportunities and needs impact children’s ability to grow at their genetic potential. The net effect of various influences across different families is reflected in average height-for-age, conditional on family circumstances and gender. As in the anthropometric literature more broadly, average height-for-age in children identifies not an individual, but a societal effect.

Because children are growing a different approach to measurement and comparison is required than used for adults. Stature by itself is not an appropriate measure of how children are faring nutritionally. The shortest six-year-old will almost certainly be taller than the tallest two-year-old. Girls and boys also have slightly different growth norms, though male and female infants begin life with very similar distributions of height and weight, and absolute differences are small through age ten. Even this summary varies over time: with better nutrition the adolescent growth spurt in both sexes starts at an earlier average age. Yet because girls start their adolescent growth spurt earlier, the difference in stature at a given age between boys and girls has changed over time.

The main complication is age. To compare children’s stature across age and sex we normalize stature, and measure deviation from contemporary norms. The standard approach in recent anthropometric research is to compare the historical stature of children to modern growth norms (Schneider, 2016, 2017) from the World Health Organisation, an approach followed in this paper to facilitate comparisons to other scholars studying long-term change. Comparing children from previous cohorts to modern norms does not impose modern expectations on the past. Rather, this approach allows consistent comparison across a range of different contexts. As mentioned above, the most serious challenge to using modern norms to study historic growth is that the onset of the adolescent growth spurt has shifted. In this paper we focus on a population of children measured through age six to avoid the need to confront this issue.

Research on children’s stature in the late nineteenth and early twentieth century is most abundant for Britain. Harris found early twentieth century British girls were slightly taller for their age than boys, but after puberty boys overtook girls in comparisons to modern growth standards (Harris, 1995). In an institutional setting Schneider found that girls who entered orphanages in late nineteenth century London were, on average, at lower percentiles of modern height standards than boys, suggesting girls were worse off (Schneider, 2016). However, by the inter-war era boys and girls in the Boyd-Orr survey of working class families in Britain reached the same height-for-age compared to modern standards (Hatton & Martin, 2010). These findings suggest that gender differences in the treatment of children declined in the early twentieth century. The effects of family size and structure on childhood stature were more significant during an era of fertility decline that varied by social class and parental education. Both birth order and family size affected children’s growth: children born later in larger families were shorter than their peers born earlier to the same parents (Hatton & Martin, 2010). Thus, despite rapid growth in the use of stature to study wellbeing in the past there is only a relatively small literature on children’s stature before World War II. Among this literature on children’s growth in Europe and North America in the past, the study of growth and stature in preschool children has been particularly small.

The small size of the literature on preschool children’s growth from the mid-nineteenth century through World War II does not reflect a lack of data. Data on stature for millions of children was collected before World War I (Baldwin, 1921), reflecting significant contemporary interest in growth from physiologists, psychologists and the broader social policy and public health community (Tanner, 1981). But relatively little individual level data has survived with measurements of children’s stature, particularly for preschool children. Indeed, the uniqueness of the data we use in this paper highlights the extensive destruction of original records. In 1918 the United States government set itself the ambitious target of measuring stature and weight of five million children for Children’s Year (Peixotto, 1918). That only a mere two million cards were received by the Children’s Bureau in Washington shows that even a failure to reach ambitious goals may still deliver substantial accomplishments. Data quality was variable, and national growth norms published based on the selection of the best 172,000 measurements (Woodbury, 1921). The data that we use from Saint Paul, Minnesota, appears to be the only surviving microdata from the nationwide data collection.¹

2.3. Childhood growth and adult outcomes in modern societies

By contrast there is extensive research in low- and middle- income countries today about the correlates of nutritional status and physical growth, and their effects on later-life outcomes. This line of research has persisted because growth faltering—where growth slows after infancy— and stunting—where children are significantly below WHO age-specific standards for optimal stature—remains high even though it has declined over the past 20 years (Vaivada et al., 2020; Victora et al., 2010). At both an individual and population level poorer nutrition in early childhood leading to slower growth is associated with worse socioeconomic outcomes in adolescence and adulthood. Summarizing research across Brazil, Guatemala, India, and the Philippines Victora et al. (2008) estimate that children who were one standard deviation taller for their age at 2 years completed an additional half-year of schooling by their early 20s. Similar results are evident across a wider range of countries in a more recent study, which estimates that globally children in developing countries lose an average of six months of schooling due to growth faltering (Fink et al., 2016).

With schooling being an important determinant of incomes in adulthood, several studies have continued this line of research by estimating the causal association of height with earnings. Because of the strong genetic component to height, the most reliable evidence comes from twin or sibling studies that can difference out genetic influences, or randomized controlled trials assessing nutritional interventions. Consistent with these findings, McGovern and colleagues estimate across a range of contexts that an increase in adult stature of 1cm—around 1/6 of a standard deviation—is causally associated with 4% higher incomes (McGovern et al., 2017). The effects of stunting and growth faltering compound across generations. Children born to stunted parents are themselves likely to be shorter (Walker, Chang, Wright, Osmond, & Grantham-McGregor, 2015) and score lower on developmental measures at an early age.

The consistent finding in modern studies in low-income countries that nutrition and growth have significant consequences for early-adult social and economic outcomes suggests we should see a link between compromised growth in early childhood and adult outcomes in nineteenth and early twentieth century Europe and North America. Stunting and growth faltering were common in the United States before World War II (Roberts & Warren, 2017), but at lower rates than observed in developing countries today (Victora et al., 2010). The national growth norms for pre-school children published in 1921 show a pattern of declining height-for-age relative to modern standards (Woodbury, 1921) beginning in infancy and continuing through 36 months. Both boys and girls averaging nearly one standard deviation below modern standards between the ages of 3 and 6. Similar patterns of growth faltering are evident in data collected by Baldwin and colleagues in Iowa for children born in the 1920s (Iowa Child Welfare Research Station, 1929) among a middle-class population. The first American cohorts in which growth faltering itself faltered were born in the 1930s (Simmons & Todd, 1938; Vickers & Stuart, 1943).

Until at least the 1930s growth faltering was common in the United States. With the average child nearly one standard deviation below modern standards, between 10 and 30% of American children born before 1920 were likely to have been stunted. The scale of growth faltering and stunting in the early twentieth century United States was closer to levels observed in low- and middle- income countries in the past 30 years than to the modern United States. Thus, the consequences are likely to have been in the same direction as in low- and middle-income countries today. Using data on children who were similar in height and weight to national norms at the time, we use modern methods of causal inference through sibling comparisons to investigate the consequences of growth failure in the United States in the early twentieth century.

3. Data and methods

Our data come from a unique set of surviving records from a children’s health survey from 1918/19, previously linked to the 1920 census (Roberts & Warren, 2017). The Saint Paul study was part of a larger collection of national height-weight data (Woodbury, 1921). The height- and weight- for age profile of Saint Paul children was nearly identical to national norms published from data collected around the country. In this paper we extend the record linkage by using the IPUMS Multigenerational Longitudinal Data (MLP) which contain identifiers linking records of men and women in the full-count IPUMS datasets of the 1900–1940 censuses. With links to the 1930 census we are able to observe our sample between the ages of 12 and 18, and at a point in their lives when children in Minnesota could obtain a work permit (at age 14) or choose to leave school (at age 16) (Lleras-Muney, 2005). Links to the 1940 census find our sample in early adulthood, and beyond the median age at marriage, prompting a focus on men only. The near-universal practice of women changing their surname at marriage means we cannot link women to census records after they have married. From the 1940 census we obtain information on adult wage and salary earnings, educational attainment, and unemployment. We outline the origins and construction of the data, before proceeding with the analysis.

3.1. Saint Paul children’s health survey

The health survey conducted in Saint Paul in 1918/19 was part of a much larger national effort to examine children’s health during World War I (Ruis, 2013). Shortly after the United States entered the war the Women’s Committee of the National Council of Defense and the Federal Children’s Bureau examined how foreign countries had monitored children’s nutrition. Seemingly successful examples of children’s welfare work in wartime Britain and France prompted President Wilson to allot $150,000 for measuring children and promoting their health, and aiming to save the lives of 100,000 children a year (Rude, 1919) (Titzel, 1919). The “Weighing and Measuring Test” was the first phase of the campaign, beginning in 1918. Seven million cards were distributed around the country for children under six years of age to be weighed and measured by parents, (Lathrop, 1921). Researchers at the Children’s Bureau knew this example of “citizen science” would not be entirely accurate. Thus, they worked to ensure doctors and nurses would make at least 100,000 physical examinations to estimate the reliability of data from parents. Contemporary analyses of the data examined variation across different regions of the country, urban and rural settings, and the national origins of parents or grandparents (Woodbury, 1921).

Data collection in Saint Paul (Minnesota) was part of the broader Children’s Year activities. As in other parts of the country, data collection was a partnership between public and private agencies. The Women’s Branch of the Ramsey County division of the Minnesota Commission of Public Safety organized data collection in Saint Paul. Little documentation can be found about administration of the survey in Ramsey County. The effort was substantial, with more than 14,000 children weighed and measured in the city. Survey cards describing children show where they were measured and who measured them, providing insight into the data collection process. Nearly all children in Saint Paul were measured by registered nurses or doctors. One in six were measured at offices of the Child Welfare Association or at a child welfare “station” at one of the city’s big department stores. This evidence of substantial professional involvement by public health workers, doctors, and nurses, provides greater confidence in measures of stature and weight. Data collection was completed before the influenza pandemic beginning in autumn 1918. It is unlikely that health conditions during the summer of 1918 affected who was weighed in the initial survey, but post-survey influenza mortality may affect who was linked to the 1920 census. We discuss these issues in conjunction with other potential biases in the Discussion.

Saint Paul was a representative urban environment for American children at the time. In 1920—the closest U.S. decennial census to the health survey—Saint Paul had a population of 235,000. With its neighbour, Minneapolis, the Twin Cities urban population was more than 600,000. As in other Midwestern and East Coast cities the population was largely white, and either immigrants themselves or the descendants of immigrants. In Saint Paul 77% of the population was native-born white, with 57% of those having at least one foreign-born parent (Ruggles, Genadek, Goeken, Grover, & Sobek, 2015). The African American population was small: at the time of the 1918 survey few African Americans had settled in Saint Paul. Only 61 African American children were measured in the survey; too few to derive conclusions about racial differences.

Although it was the state capital, Saint Paul was better characterized as industrial, with a quarter of working men in manufacturing. Railroads were the biggest employers in Saint Paul: Nearly 20 per cent of working men had a job with a railroad company. Clerical work for the state government and distribution and transportation associated with railroads were the final important aspects of Saint Paul’s economic structure (Ruggles et al., 2015). The health environment of the city was slightly better than for other American cities. Infant mortality rates from 1915–18 averaged 78/1000 live births, compared to 95 for white children in other cities (Bureau of the Census, 1916–1920). Retail food prices increased in the Twin Cities during World War I, as they did around the United States. Prices for protein intensive foods important for growth, such as meat and dairy, were 15% lower than national urban averages (Bureau of Labor Statistics, 1922). Across a range of factors, Saint Paul children who were weighed and measured in 1918 were demographically similar to children in other large cities. But Saint Paul children experienced somewhat better health and nutritional conditions than children in larger and denser cities such as New York, Boston, or Chicago.

3.2. Creation of baseline health data

Original survey cards for 14,252 Saint Paul children in the Weighing and Measuring Test were preserved at the Minnesota Historical Society. The cards record the name, address, age or date of birth, height and weight, and serious illness of children. Additional demographic information about the family was written on some cards by social workers and nurses who took the survey, but these details were not collected systematically. The archiving of the cards facilitated straightforward identification of families and siblings. Cards were sorted by city ward, and then ordered alphabetically within each ward, so siblings were clearly identifiable: living at the same address and sharing a surname. In nearly every case that we found a sibling their cards were in sequential order, and siblings separated by an unrelated child’s card were rare. Thus, the sort order of data entry was useful in making initial determination of family relationships. Family relationships were directly enumerated in the census, and were confirmed during record linkage to the 1920 census (Roberts & Warren, 2017).

We linked 70% of children’s records from the 1918 survey to the 1920 census. Links were made in two phases: A set of initial links were made by a cohort of trained undergraduate interns searching on Ancestry.com who then entered data on the children and their household co-residents. The initial hand-linked data were supplemented by machine-linking to an electronic dataset of the complete 1920 census. For children with full anthropometric data, and thus usable in the analyses, the match rate was 72%. Accounting for under-enumeration (Hacker, 2013) and likely mortality over 18 months we estimate linking 78% of surviving and enumerated children to a census record.

Matched subjects did not differ substantively from their peers who we could not match to the 1920 census. Mean height-for-age Z scores for unmatched children were slightly lower (−0.77) than for matched children (−0.72). Weight-for-age Z scores differed by the same amount. While both differences are statistically significant at standard levels, the magnitude is small. Thus, we are confident that our inferences are not substantially biased by differences in the initial linking phase to the 1920 census. Characteristics of the households in which children were found are presented in Table 1.

Table 1.

Characteristics of the matched sample and Saint Paul children in 1920

Characteristic	Proportion of sample	Saint Paul children < 7
Resident in MN in 1920	0.97	N/A
Head of household’s nativity
Born in MN	0.39	0.37
Born in USA	0.21	0.18
Born abroad	0.40	0.45
Head of household’s occupation
Professional, manager, owner	0.21	0.18
Clerical, sales, service	0.17	0.20
Trades, craftsmen, operatives	0.45	0.48
Laborers	0.11	0.11
No occupation / out of labor force	0.06	0.04
Owned house	0.49	0.49
Ward of city
1	0.13	0.12
2	0.08	0.11
3	0.01	0.01
4	0.01	0.01
5	0.11	0.11
6	0.11	0.11
7	0.08	0.08
8	0.13	0.13
9	0.08	0.07
10	0.10	0.08
11	0.05	0.11
12	0.12	0.07

Children in the survey and linked to the census were similar to the broader population of children under 7 in Saint Paul. We make a comparison to Saint Paul only, as most of the linked sample had not moved far. The low migration rate reflects both that non-migrants are easier to link than migrants, and a short period at risk in which to move. Some measurements were carried out as late as 1919, and the 1920 census was conducted around a 1 January enumeration date.

Occupational and housing characteristics of the linked sample’s parents were very similar to the broader population of parents of 0–6-year-olds in 1920 in Saint Paul. Household heads were largely employed in industrial and service occupations, with one in five being professionals or managers. Labour markets were still tight in 1920 following World War I, with only 6% of household heads not recording an occupation. Across the entire city there was a near-equal split between homeowners and renters. The greatest threat to the broader generalizability of our results is likely to be the slightly better urban health and nutritional environment in Saint Paul compared to other American cities at the same time.

3.3. Record linkage to the 1930 and 1940 census

Our analysis of later-life outcomes is based on newly released IPUMS Multigenerational Longitudinal Data (MLP) (Helgertz, Ruggles, et al., 2020), which contain indentifiers linking records of men and women in the full-count IPUMS datasets of the 1900–1940 censuses (Ruggles et al., 2019). The design of the MLP facilitates straightforward integration of non-census demographic data sources with longitudinal data from the census, after links have been made to one census. In our case, initial links were made to the 1920 census. From this starting point we can obtain MLP links of our sample to the 1930, and subsequently 1940 census. The restricted use versions of the IPUMS complete-count censuses contain census-specific individual identifiers. Records from additional censuses can be attached (e.g. 1920 to 1930) by using the crosswalks constructed by MLP, consisting of pairs of unique individual identifiers. Record linkage in the MLP was carried out forwards between pairs of adjacent censuses. Thus, the 1940 census is linked to the 1930 census, and an independent set of links is made between the 1930 and 1920 censuses. There is no direct link made over the 20 year span between 1920 and 1940, so that subjects who were alive in both years, but not enumerated in 1930 will not be linked over the 20-year span.

Links in the MLP were generated through a two-step probabilistic linking algorithm, first identifying high-confidence links in the data through the use of an elaborate set of linking variables. The subsequent step exploits already declared confident matches to identify less certain links among other household members. The linking algorithm not only links a higher share of males compared to other methods used within the social sciences, but is also characterized by a higher precision (Helgertz, Price, et al., 2020). For our population of individuals who are children in 1920, often found with siblings, and adolescents in 1930, the MLP process is likely to make a high number of matches between the two censuses because of the confirmation provided by finding people in similar households ten years apart. Between 1930 and 1940 our sample is likely to experience more attrition, as adolescents moved out of the household. While we identify our sample through to 1940 by chaining links made between 1920 and 1930, and then between 1930 and 1940, we would face similar challenges if trying to make direct links between 1920 and 1940. Children observed at ages 1–7 in 1920 are likely to appear in a different houshold context in 1940 because of the change in age, and typical ages for leaving home in the United States in this era. In 1940 the median age for men leaving home was 23, and 20 for women. With women changing their name at marriage, we are unlikely to find a representative sample of women in 1940, because surnames are used to complete links between American census records. Given the age of our sample in 1940 (22–28), the women that we can link would be slightly unusual as they married later than average.

3.4. Adolescent and early adult outcomes

We analyse a range of relevant adolescent and early-adult outcomes taken from the 1930 and 1940 censuses, relating outcomes to childhood height-for-age measured in 1918/19. In 1930 we assess whether boys and girls were in school. In 1940, focusing on young men only, we examine 1) educational attainment, 2) weekly earnings for men in waged or salaried positions, and 3) the duration of any unemployment spell, and 4) whether young men were married. Because of the strong genetic component in stature, our preferred specification for all outcomes applies sibling fixed effects. However, a slight majority (57%) of our sample in 1918 that we matched to 1920 census had no other co-resident children in the household. As we link sequentially to the 1930 and 1940 census the linked sample also contains singletons. In particular, siblings who were measured together in 1918 become singletons if one of the children were girls because we do not consider girls in our links to 1940. To increase the efficiency of our estimates we apply a new fixed-effects estimator (Bruno, Magazzini, & Stampini, 2020; Magazzini, Bruno, & Stampini, 2020) that includes information from singletons, whether these children were found alone in 1918–20, or become singletons because only one member of the sibling group is found in 1930 or 1940. The assumption behind this estimator is that the OLS bias is the same for singletons as for pairs or triples of siblings. The intuition of the approach is that information from singletons cannot be used for point estimates because there are no other subjects within the group (family) on which characteristics differ. But there is information about the distribution of singleton characteristics for relevant variables, and thus information on singletons can reduce standard errors.

4. Results

4.1. Record linkage rates and lack of apparent bias

We were successful in linking a high share of children to the 1930 and 1940 census using the MLP links, obtaining sufficient cases for analysis. Having previously linked 72% of the children from the 1918/19 survey to the 1920 census, we begin with a sample of 9,963 children who could potentially be matched to the 1930 census (Table 2). Conditional on linking to 1920, we linked 70.5% of boys forward to the 1930 census, and 49.6% (2,499) to the 1940 census. We linked a smaller share of girls to 1930 (62%), likely reflecting the broader social pattern that young women left home at earlier ages, and just 22.7% (1,118) to 1940. Because we focus on 1940 outcomes for men, the sample size in this year is most relevant for comparison to other studies. Our sample size sits in the middle of samples used for similar studies: significantly larger than intensive studies of nutritional interventions in low- and middle- income countries, yet smaller than panel and cohort studies in high-income countries (McGovern et al., 2017). Baseline characteristics of groups by their record-linkage are shown in Table 3. There were no significant biases in the physical status of children who could be linked onward from the 1920 census. On both height- (Figure 1) and weight- (Figure 2) for age standards, children who were not linked were slightly smaller. This small bias was similar for boys and girls, and we show combined results. Unfortunately, we cannot link to mortality registers to explore whether smaller children died earlier, as these have not been digitized in a format suitable for record linkage.

Table 2.

Number and share of subjects linked to censuses

	Male	Female	Total
	N %	N %	N %
Appear in health survey only	2,193	2,041	4,234
	30.34	29.30	29.83
Linked to 1920	1,482	1,873	3,355
	20.51	26.89	23.64
Linked through 1930	1,053	1,934	2,987
	14.57	27.76	21.05
Linked through 1940	2,499	1,118	3,617
(includes links to 1940)	34.58	16.05	25.48
Total	7,227	6,966	14,193
	100.00	100.00	100.00

Table 3.

Baseline characteristics of match groups

	Last year subject linked to
	1918	1920	1930	1940	Total
	N=4289	N=3355	N=2987	N=3619	N=14250
Male	2193 (51.8%)	1482 (44.2%)	1053 (35.3%)	2499 (69.1%)	7227 (50.9%)
Female	2041 (48.2%)	1873 (55.8%)	1934 (64.7%)	1118 (30.9%)	6966 (49.1%)
Age in years	2.6 (1.8)	2.8 (1.7)	3.0 (1.7)	2.6 (1.7)	2.7 (1.7)
1920 residence
Not resident in MN, 1920		184 (5.5%)	88 (2.9%)	68 (1.9%)	340 (3.4%)
Resident in MN, 1920		3171 (94.5%)	2899 (97.1%)	3551 (98.1%)	9621 (96.6%)
Head’s nativity
Born in MN		1190 (35.5%)	1183 (39.6%)	1515 (41.9%)	3888 (39.0%)
Born in US, not MN		756 (22.5%)	651 (21.8%)	723 (20.0%)	2130 (21.4%)
Foreign born		1409 (42.0%)	1153 (38.6%)	1381 (38.2%)	3943 (39.6%)
Head’s occupation
Professional		120 (3.6%)	132 (4.4%)	185 (5.1%)	437 (4.4%)
Farmers & farm managers		103 (3.1%)	133 (4.5%)	142 (3.9%)	378 (3.8%)
Managers and proprietors		409 (12.2%)	404 (13.5%)	449 (12.4%)	1262 (12.7%)
Clerical		262 (7.8%)	235 (7.9%)	298 (8.2%)	795 (8.0%)
Sales		156 (4.6%)	162 (5.4%)	174 (4.8%)	492 (4.9%)
Skilled trades		971 (28.9%)	899 (30.1%)	1090 (30.1%)	2960 (29.7%)
Operatives and related workers		508 (15.1%)	461 (15.4%)	582 (16.1%)	1551 (15.6%)
Service workers		150 (4.5%)	109 (3.6%)	142 (3.9%)	401 (4.0%)
Farm laborers		35 (1.0%)	37 (1.2%)	24 (0.7%)	96 (1.0%)
Laborers and no occupation		372 (11.1%)	265 (8.9%)	325 (9.0%)	962 (9.7%)
Unclassifiable occupation		9 (0.3%)	3 (0.1%)	4 (0.1%)	16 (0.2%)
No occupation/out of labor force		260 (7.7%)	147 (4.9%)	204 (5.6%)	611 (6.1%)
Head’s industry
Agriculture		106 (3.2%)	142 (4.8%)	166 (4.6%)	414 (4.2%)
Construction or extractive		179 (5.3%)	178 (6.0%)	180 (5.0%)	537 (5.4%)
Manufacturing, durable		378 (11.3%)	301 (10.1%)	414 (11.4%)	1093 (11.0%)
Manufacturing, non-durable		518 (15.4%)	491 (16.4%)	551 (15.2%)	1560 (15.7%)
Transportation, construction, utilities		707 (21.1%)	679 (22.7%)	828 (22.9%)	2214 (22.2%)
Wholesale and retail trade		464 (13.8%)	419 (14.0%)	518 (14.3%)	1401 (14.1%)
Finance, insurance, real estate		80 (2.4%)	59 (2.0%)	73 (2.0%)	212 (2.1%)
Services		220 (6.6%)	162 (5.4%)	242 (6.7%)	624 (6.3%)
Government		172 (5.1%)	168 (5.6%)	178 (4.9%)	518 (5.2%)
Unclassifiable industry		214 (6.4%)	185 (6.2%)	207 (5.7%)	606 (6.1%)
No industry		317 (9.4%)	203 (6.8%)	262 (7.2%)	782 (7.9%)
Order of child in family group (~ birth order)
One		1369 (40.8%)	1006 (33.7%)	1314 (36.3%)	3689 (37.0%)
Two		819 (24.4%)	757 (25.3%)	913 (25.2%)	2489 (25.0%)
Three		476 (14.2%)	480 (16.1%)	545 (15.1%)	1501 (15.1%)
Four		294 (8.8%)	284 (9.5%)	343 (9.5%)	921 (9.2%)
Five		165 (4.9%)	174 (5.8%)	186 (5.1%)	525 (5.3%)
Six		232 (6.9%)	286 (9.6%)	318 (8.8%)	836 (8.4%)

Figure 1.

Height-for-age scores were similar across linking status

Figure 2.

Weight-for-age scores were similar across linking status

Another key factor in assessing whether the group linked to later life records can be used for broader inference is whether their families were, at baseline, similar on socioeconomic dimensions. The best measure of these characteristics is the occupation of the relative (typically the father) heading the child’s household in 1920 (Table 3). As with physical characteristics, we see no large differences between children linked only to 1920, or those who we could link further to 1930 and 1940. The distribution of head’s occupations is remarkably similar. The most significant difference is that we were able to link a large share of children resident (in 1920) with heads born in Minnesota all the way to 1940. However, children living with foreign-born heads in 1920 were linked to 1940 at a higher rate than children of American, but not Minnesota-born, heads. Thus, comparing the children of native and foreign-born households, we see only a small and unimportant difference in record linkage rates to 1930 and 1940. With only small differences in physical status for our links through 1920, and small differences between groups linked or not to subsequent censuses, we can assess the trajectories of the linked sample more confidently as being representative of the underlying population surviving across the interval.

4.2. Adolescent outcomes: school attendance in 1930

Lying squarely on the pathway between childhood development and adult socioeconomic outcomes, school attendance has been found to be related to physical growth and nutrition in low- and middle- income countries in recent decades. Despite the theoretical and empirical importance of schooling, we find no evidence schooling was related to within-family differences in height- or weight- for age. In the first instance, we see only minor differences in school attendance rates by sex (Table 4). Through age 15 which covers much of our sample, attendance was nearly universal, and only drops to below 80% for 16-year-olds. In a pattern reflective of broader cohort-specific trends in the region (92% of our linked sample was still living in Minnesota in 1930, and more than 95% in the Midwest), school attendance diminished significantly at age 17 (Goldin, 1998). School attendance rates in our sample were relatively high through age 15, but these rates are consistent with broader population patterns. As we find a significant number of our sample in the Midwest and California, we are picking up the “high school movement” in the United States described by Goldin where high school attendance rose significantly in the 1920s and 1930s in these regions, compared to the South and Northeast.

Table 4.

School attendance rates by age and sex in 1930

Age	Male	Female
12	.97	.94
13	.97	.95
14	.95	.93
15	.90	.89
16	.79	.76
17	.56	.60
18	.45	.50

Using a sibling fixed effects approach, we found no impact of physical growth on adolescent school attendance (Table 5). The only difference between siblings that we were able to detect was between siblings of different “sibling order:” Being the second child in a family compared to the first increased the chance of school attendance by 11% in our linear probability estimates. At first glance this result appears to contradict findings that parents may favour the educational attainment of older children. We note from Table 4 that the differences between sequential ages in school attendance rates differ by at least 11% after age 15. Thus, the estimate of the effect of sibling order is largely picking up an artefact of our linked sample being measured in the same year (1930), and school attendance declining from age 15. To properly identify the effect of physical growth on school attendance we would ideally observe children from the same family at the same age, or multiple ages. By observing children from the same family at different adolescent ages, our regression is merely identifying something close to the observed population difference in school attendance between siblings 2–3 years apart.

Table 5.

Family fixed effects models of school attendance in 1930

Logit estimates: School attendance in 1930
	Coef.	Std. Err.	z	P>|z|	[95% Conf. Interval]
Z, height for age	−0.025	0.117	−0.210	0.834	−0.254	0.205
Sibling order	1.621	0.164	9.887	0.000	1.300	1.942
Female	−0.106	0.263	−0.404	0.686	−0.623	0.410
Linear probability estimates: School attendance in 1930
		Robust
	Coef.	Std. Err.	z	P>|z|	[95% Conf. Interval]
Beta
Z, height for age	−0.004	0.006	−0.689	0.491	−0.017	0.008
Sibling order	0.113	0.011	9.887	0.000	0.090	0.135
Female	0.007	0.015	0.492	0.622	−0.022	0.037
Constant	0.556	0.039	14.132	0.000	0.479	0.633
Bias
Z, height for age	0.019	0.006	3.012	0.003	0.007	0.032
Sibling order	−0.124	0.012	−10.767	0.000	−0.146	−0.101
Female	−0.017	0.014	−1.213	0.225	−0.045	0.011
Constant	0.367	0.038	9.662	0.000	0.293	0.442

4.3. Adult outcomes

Educational attainment

We examine adult outcomes for boys in our initial sample who were between the ages of 22 and 28 when we observe them in 1940. At this age, young men of all ages in our sample have had the opportunity to graduate from college. The modal boy in this birth cohort from Minnesota attended high school but did not graduate; the median grade completed was the 10^th grade (Table 6). In our longitudinally linked sample, the modal boy did not graduate. However, the median grade completed was 12^th grade, equivalent to graduation from high school. A similar proportion of boys in our linked sample and those born in Minnesota in the same era attended college or graduated from it.

Table 6.

Educational attainment measured in 1940 (boys only)

	Freq.	Percent	Boys born in MN, aged 22–28
< HS	1032	41.82	55.58
HS	904	36.63	28.30
Some coll.	320	12.97	9.13
College grad	212	8.59	6.99
Total	2468	100.00

Compared to state averages, our linked sample appears to have higher educational attainment. However, this reflects that our panel is comprised of boys who were already living in an urban area, where high school attendance rates were higher. Moreover, 75% of our panel were still living in Saint Paul in 1940. The Twin Cities metropolitan area was distinguished by a high number of colleges, including the flagship campus of the public university. The college attendance (9.3%) and completion (8.3%) rates in our linked sample are very similar to the rates for men of the same age living in urban areas outside the South. Thus, initial apparent differences between our longitudinal panel and its own birth cohort in Minnesota reflect the urban composition of the sample and to some extent the availability of college in the metropolitan area. High school graduation rates are slightly higher than in the same age cohort more broadly, but similar to boys with similar urban origins to our group.

We find no effects of differential height-for-age on the years of schooling once accounting for shared family characteristics (Table 7). Our naïve estimates under OLS without accounting for family relationships shows that boys who were one standard deviation taller completed an additional quarter year of school, and this estimate is statistically significant. However, the size of this estimate halves when we account for shared family characteristics using our preferred approach of incorporating the additional information from singletons. We suspect the fairly tight clustering of educational attainment contributes to this null finding: more than 50% of the linked sample attended high school and thus the variation in completed years of schooling is relatively low.

Table 7.

Effect of physical characteristics on years of schooling

OLS	Coef.	Std. Err.	t	P>|t|	[95% Conf. Interval]
Z, height for age	0.272	0.038	7.235	0.000	0.198	0.346
Constant	11.468	0.062	185.728	0.000	11.347	11.590
Family fixed effects, height only
beta
Z, height for age	0.135	0.112	1.207	0.227	−0.084	0.354
Constant	11.365	0.099	115.306	0.000	11.172	11.558
bias
Z, height for age	0.137	0.113	1.208	0.227	−0.085	0.358
Constant	0.102	0.084	1.206	0.228	−0.064	0.267
Family fixed effects, height and family characteristics
beta
Z, height for age	0.086	0.105	0.815	0.415	−0.120	0.292
Birth order index	1.188	0.896	1.326	0.185	−0.568	2.945
First born	0.376	0.383	0.981	0.327	−0.375	1.126
Age (integer)	0.077	0.126	0.611	0.541	−0.170	0.324
Constant	9.590	1.449	6.620	0.000	6.750	12.429
bias
Z, height for age	0.167	0.106	1.583	0.113	−0.040	0.375
Birth order index	−1.082	0.894	−1.211	0.226	−2.835	0.670
First born	0.093	0.372	0.250	0.803	−0.637	0.822
Age (integer)	−0.152	0.127	−1.194	0.232	−0.401	0.097
Constant	1.778	1.438	1.236	0.216	−1.040	4.597

We turn finally to examining whether differences in physical growth in childhood affects work and wages in early adulthood. Here the timing of the 1940 census is fortuitous for our analysis, coming at the end of the Great Depression. Thus, we see significant levels of unemployment with the potential to identify effects at a moment when unemployment was highly variable. By contrast, if the economy had been closer to full employment, we would be harder pressed to identify within families what distinguished the employed from the unemployed. But with unemployment rates in our sample at 13% we have greater analytical opportunities (Table 8). Earnings are available in the 1940 census for people working for wages or salary, comprising just over 2,000 of our sample. We convert the annual earnings provided in the census to a measure of logged weekly wages for comparability with other work using these sources (Feigenbaum & Tan, 2020).

Table 8.

Employment characteristics of sample in 1940

Class of worker	Freq.	Percent	Cum.
Not available	303	12.13	12.13
Self-employed	134	5.36	17.49
Works for wages	2061	82.51	100.00
Total	2498	100.00
Unemployed at 1940 census
Not unemployed	2161	86.47	86.47
Unemployed at 1940 census	338	13.53	100.00
Total	2499	100.00

We find stronger evidence that physical growth measured in 1918 is associated with adult outcomes in relation to unemployment and weekly earnings. Following the literature from low- and middle- income countries we assess the role of both height and BMI as measures of physical growth. Children who were lighter for their age, or lighter for their build may have had compromised nutrition even if their height was not falling far below modern norms. In our preferred specifications that account for shared family characteristics we find that body mass index affects both unemployment and weekly earnings. The direction of effects is consistent: being heavier in childhood was associated with better outcomes. For unemployment we find, using a linear probability model, that an increase in BMI for age of one standard deviation reduces the chance of being unemployed by 2.6 percentage points (Table 9). Relative to the mean unemployment rate of 13%, this is a significant effect. For men who were employed it was also advantageous to have been heavier in childhood relative to one’s siblings. Men who were one standard deviation heavier in childhood earned 5.8% more per week (Table 10). In supplementary analyses that controlled for the broad occupational sector in which men were employed these effects were stronger with an 8% increase in weekly earnings for each standard deviation increase in BMI for age relative to siblings. This suggests that early-life nutrition also influenced occupational choices. At face value these effects are sizeable and greater than the returns to an additional year of education for men of a similar birth cohort (Feigenbaum & Tan, 2020).

Table 9.

Unemployment at 1940 census, Family FE, BMI

		Robust
	Coef.	Std. Err.	z	P>|z|	95% Conf. Interval
beta
Z, BMI for age	−0.026	0.013	−2.031	0.042	−0.051	−0.001
Z, BMI for age square	0.001	0.007	0.095	0.924	−0.013	0.014
Birth order index	0.057	0.109	0.523	0.601	−0.156	0.270
First born	0.035	0.046	0.761	0.447	−0.055	0.126
Age	0.000	0.018	0.015	0.988	−0.036	0.036
Constant	0.054	0.183	0.296	0.767	−0.304	0.412
bias
Z, BMI for age	0.015	0.013	1.216	0.224	−0.009	0.040
Z, BMI for age square	0.004	0.006	0.551	0.582	−0.009	0.016
Birth order index	−0.022	0.111	−0.201	0.841	−0.240	0.196
First born	−0.008	0.044	−0.190	0.849	−0.095	0.079
Age	−0.004	0.018	−0.220	0.826	−0.039	0.032
Constant	0.034	0.186	0.180	0.857	−0.331	0.398

Table 10.

Weekly earnings and physical growth

Weekly earnings at 1940 census, OLS
	Coef.	Std.	Err.	t	P>|t|	[95% Conf. Interval]
Z, height for age	0.013	0.009	1.496	0.135	−0.004	0.030
Constant	3.018	0.014	214.775	0.000	2.990	3.046
Family fixed effects, height only
beta
Z, height for age	−0.066	0.031	−2.147	0.032	−0.127	−0.006
Constant	2.962	0.026	113.531	0.000	2.911	3.013
bias
Z, height for age	0.075	0.030	2.468	0.014	0.015	0.134
Constant	0.055	0.022	2.446	0.014	0.011	0.098
Family fixed effects: BMI, family structure, and education
beta
Z, BMI for age	0.058	0.029	1.978	0.048	0.001	0.115
Z, BMI for age square	−0.002	0.009	−0.174	0.862	−0.019	0.016
Birth order index	−0.048	0.158	−0.301	0.763	−0.357	0.262
First born	−0.087	0.085	−1.014	0.311	−0.254	0.081
Age	0.070	0.067	1.036	0.300	−0.062	0.202
Age squared	0.002	0.010	0.184	0.854	−0.018	0.022
Years of schooling	0.028	0.023	1.223	0.221	−0.017	0.073
Constant	2.563	0.320	8.011	0.000	1.936	3.190
bias
Z, BMI for age	−0.035	0.028	−1.223	0.221	−0.091	0.021
Z, BMI for age square	−0.002	0.008	−0.196	0.845	−0.018	0.015
Birth order index	0.020	0.157	0.125	0.901	−0.289	0.328
First born	0.084	0.082	1.030	0.303	−0.076	0.244
Age	0.003	0.066	0.038	0.970	−0.126	0.131
Age squared	−0.004	0.010	−0.454	0.650	−0.023	0.015
Years of schooling	0.000	0.022	0.018	0.986	−0.043	0.044
Constant	−0.013	0.312	−0.040	0.968	−0.624	0.599

5. Discussion

In this paper we bring together two strands of the literature—the historical analyses of changing growth patterns and modern analyses of the long-term consequences of compromised childhood growth—to ask whether children in the early twentieth century United States suffered long-term consequences from being smaller or lighter in childhood. We take advantage of a large and broadly representative population survey of children’s health, linked to later life records. With our sample including a large number of siblings we use sibling-comparison methods to identify plausibly causal effects of differences in early childhood stature and BMI. Because sibling comparison methods require a significant amount of complete longitudinal data on both siblings, we apply new techniques for combining singletons and complete sibling sets in estimation (Magazzini et al., 2020).

While sibling comparison methods offer important advantages for causal inference by allowing researchers to difference out shared, measured family characteristics they are also subject to important limitations. Researchers should not assume that the fundamental problem of causal inference has found its solution in within-family comparisons. Sibling comparison methods control for shared experiences but assume homogenous experiences within the family that may not occur (East & Jacobson, 2000). Although two of the authors (who are only children) cannot personally attest to it, it is well established in historical (Parman, 2015a) and contemporary (Conley & Lareau, 2008; Hsin, 2012) studies that parents treat siblings differently. Parents’ differential treatment may be responsive to birth order or indicators of children’s health, and either reinforce or compensate for initial endowments or health shocks. Parman’s study (2015a, 2015b) of a similarly-aged cohort to ours suggests two ways of interpreting our findings in light of these concerns. First, children like those in this cohort, born before the influenza pandemic, who had siblings in utero during the pandemic tended to benefit from parental re-allocation of resources towards pre-pandemic children. Our cohort is composed entirely of pre-pandemic children, so our within-family comparisons are not affected by a comparison between pre-pandemic children and those whose health was affected by being in-utero during the influenza pandemic. Secondly, Parman’s work shows that because family fixed effects models are affected by attenuation bias from measurement error (Griliches, 1979) our estimated coefficients on height and body mass are likely to be under-estimates of the true effects.

While sibling comparison methods are likely to be robust for our sample, the particular birth years of the cohort place them in historical time in a way that may affect our results. Specifically, our results could be influenced by

1. Childhood or parental mortality during the 1918/19 influenza pandemic

2. Father’s mortality during World War I

3. The cohort being of high school age during the beginning of the Great Depression

We discuss these in turn to assess the extent to which they are likely to affect our results. All three issues may affect the extent to which the cohort can be linked first to the 1920 health survey, and then to subsequent censuses. Parental mortality and the onset of the Great Depression will also affect families’ choices about education and labour force entry. Education is one of our outcome variables, and the timing of labour force entry and workplace experience affects earnings.

Mortality has both direct and indirect impacts on record linkage for this cohort. The obvious direct effect of childhood mortality in this sample—weighed and measured in 1918—is to make it impossible to obtain socioeconomic information from the 1920 census because the child is dead and not enumerated. We noted in our previous work that there were only small anthropometric differences between children matched to the census and links that failed (Roberts & Warren, 2017). Moreover, childhood mortality can account for only a small fraction of failed links. We matched 9,963 children to the 1920 census from of an initial sample of 14,252. Influenza or other forms of childhood mortality can account for, at most, 23 per cent of record linkage failure. There were 990 deaths of children in Saint Paul in 1918 and 1919 at ages consistent with being in our sample (Table 11). However, as the Saint Paul data only weighed and measured about half the city’s eligible children it is unlikely all the deaths came from children in the sample. If the share of study children among deaths were proportionate to their share in the city population, then mortality would account for approximately 11% of record linkage failures. In short, mortality among the sample is likely to be an important contributor to record linkage failure, but other issues including under-enumeration in the census, transcription challenges, and common names account for much more.

Table 11.

Number of deaths of Saint Paul children and adults of child-bearing age in 1918–19 relative to population

	1918a		1919		Population b			Mortality rate, 1918c
Age	Flu	Total	Flu	Total	1910	1920	1918
0	25	449	11	331	3,831	4,345	4,242	106
1	23	94	5	56	3,369	4,493	4,268	22
2	15	55	2	35	3,876	4,429	4,318	13
3	8	39	6	33	3,774	4,469	4,330	9
4	6	31	1	21	3,596	4,366	4,212	7
5–9	18	92	6	85	17,535	20,441	19,860	5
20–29	219	506	37	254	53,192	49,047	49,876	10
30–39	190	536	37	304	35,257	41,214	40,023	13
40–49	50	329	11	277	25,213	27,627	27,144	12

Notes:

^a)Mortality figures from Mortality Statistics 1918 (p.454) and Mortality Statistics 1919 (p.471) 19^th and 20^th Annual Reports, Bureau of the Census, Washington. Available at https://www.cdc.gov/nchs/products/vsus/vsus_1890_1938.htm.

^b)1910 and 1920 Saint Paul population figures from IPUMS complete-count census data. Available at https://usa.ipums.org/usa/complete_count.shtml. 1918 population estimated by linear interpolation.

^c)Mortality rate per 1,000 calculated relative to 1918 estimated population.

For similar reasons parental mortality from World War I or the influenza epidemic is unlikely to be a major source of bias relative to other causes. The mortality rate of soldiers from Ramsey County—of which Saint Paul was 96% of the 1920 population—in World War I was low. Just 148 men from Ramsey County died abroad during the war, with casualty rates slightly below the American average. Moreover, even in the unlikely event that all of these men were fathers of children in the Saint Paul health survey, they would account for only a small fraction of the more than 4,000 children who were not linked to the 1920 census. While World War I impacted American children, its impact was much smaller than for children in other combatant nations. The death toll for American soldiers was low, and men who served in the American Expeditionary Force were younger and less likely to be married than in other armies. A question in the 1930 census on veteran status allows us to assess this indirectly, by examining how many veterans residing in Minnesota had children born in the 1912–18 period. Clearly this is not the true population of interest, since we would prefer to examine veterans in 1920 and be able to identify men resident in Minnesota in 1918. However, the census did not ask about World War I military service until 1930. Few men born before 1880 entered military service in the American forces other than as officers, so the tables focus on men born between 1880 and 1900. Men who had served in World War I were significantly less likely to have children born in the 1912–18 period living with them in 1930. For both veterans and non-veterans, the number of children observed with them in 1930 will be an under-estimate of fertility since children born 1912–18 could have left home by 1930. However, it is unlikely the propensity of children to leave home differs significantly between the children of veterans and non-veterans. Veterans of the Great War were much less likely to have children born 1912–18 living with them in 1930 (Table 12), suggesting lower fertility among men who served. Just 5 per cent of ever-married veterans had a child born 1912–18 living with them, compared to 45 per cent of non-veterans. This suggests that relatively few children in the study group would have had a father who served.

Table 12.

Children born to veterans residing in Minnesota in 1930

	All men		Ever-married men
Children born 1912–18	Did not serve	WWI Veteran	Did not serve	WWI Veteran
0	0.64	0.96	0.55	0.95
1	0.18	0.03	0.22	0.04
2	0.11	0.01	0.15	0.01
3 or more	0.07	0.00	0.08	0.00
N	297,870	83,001	235,231	63,407

Note: Veterans selected from men born 1880–1900

Finally, we can examine changing family structure among Saint Paul children of the age examined in the health survey and in our sample to assess whether parental mortality may be a major source of bias. For children aged 0–6 family structure did not change significantly in Saint Paul from 1910 to 1930 (Table 13). At all three census enumerations 93 per cent of 0–6-year-old children were living with both parents. Among all children of eligible birth years (1912–18) for the health survey the proportions were similar. At the point study children were initially linked to the 1920 census, we found 96 per cent of our sample were living with both parents, slightly higher than among the general population. The differences were magnified when we compared our linked sample’s family structure to their birth cohort in 1930: 90 per cent of the sample linked to 1930 were living with both parents reflecting that record linkage between censuses is more likely when people remain with their families.

Table 13.

Family structure of Saint Paul children, 1910–1930

	Born 1912–18		Sample linked to 1930		0–6 year old children
	1920	1930	Boys	Girls	1910	1920	1930
Both parents	0.91	0.78	0.88	0.91	0.93	0.93	0.93
Father only	0.02	0.03	0.02	0.02	0.01	0.01	0.01
Mother only	0.05	0.11	0.07	0.05	0.03	0.04	0.04
Neither parent	0.02	0.08	0.03	0.02	0.03	0.02	0.02

Thus, while parental or child mortality is likely to be a modest share of unlinked cases our sample has a notably higher rate of remaining at home in 1930. The context of this census enumeration occurring early in the Great Depression is critical. Our linked sample over-represents children who remained at home, whether to stay in school or to help their parents by earning an income.

6. Conclusion

In prior work we showed that children’s growth in this era was characterized by a pattern of growth faltering from modern norms in the toddler and pre-school years (Roberts & Warren, 2017). This pattern was observed for both height and weight. Children from larger families were smaller, and later born children in large families even more disadvantaged. Because of socioeconomic variation in fertility—men with professional and white-collar jobs tended to have smaller families—the effects of growth faltering largely fell on children of poorer backgrounds. In the conclusion to our previous paper we speculated that modern evidence from low- and middle- income countries suggested that the impact of growth faltering was likely to have been noticeable (McGovern et al., 2017). In this paper we confirm that speculation: Childhood physical status has significant effects on early-adult earnings and unemployment.

The estimated effects we observe are broadly consistent with the findings from contemporary lower income countries. A standard deviation change in height- or BMI- for age is associated with differences in adult earnings or the likelihood of unemployment of 2–6%. These effect sizes are in the range of those found by McGovern and colleagues’ review of similar causal estimates from modern developing countries (McGovern et al., 2017). As with all social science findings, an important question is “how large is large?” Early-childhood growth does not entirely explain early-adult educational attainment or employment outcomes entirely, but it has an identifiable impact. These estimates imply a difference in earnings of more than 10% between men one standard deviation below and one standard deviation above the norm. Our sample is observed in early adulthood, near the start of men’s working and earning careers. A common finding in the literature on wages and earnings is that wages early in adulthood are predictive, though do not determine, later earnings (Brunner & Kuhn, 2014; Cheng, 2014; Fuller, 2008): Men who earn more in their first job in their early 20s will tend to earn more throughout their career. Finding that early-childhood stature and BMI had important effects on wages in early adulthood makes it likely that the cumulative effect was even larger.

Identifying the causal effects of childhood stature and BMI from differences within families (between siblings) has important inferential advantages by holding all shared aspects of family background constant. However, one limitation of inferential approach in that we lose the ability to identify effects between families—comparing a shorter child from a poorer family with a taller child from a richer family. Our previous paper showed that the differences in stature between children of different social backgrounds were significant: children of professionals and white collar workers were half a standard deviation taller for their age than children of laborers (Roberts & Warren, 2017). Taking the results from our within-family estimates implies that the compromised physical growth of working-class children contributed to differences in adult earnings 20 years later. Thus, nutrition was an identifiable channel for the reproduction of income and employment disadvantage between generations. It is useful to contrast our findings with the concerns of contemporaries who saw childhood nutrition as an urgent issue and were concerned hunger affected children’s ability to do well in school, and the chances of children from working-class backgrounds to improve their standard of living. We do not find large differences in educational attainment between children from the same family of different body mass for age, but do find differences in earnings and the chances of being unemployed. Years of schooling is a coarse measure of educational achievement, and in both historical and contemporary settings children who attained different grade levels had varying levels of understanding and achievement. Wages and job performance that contribute to employability are more plausibly associated with these finer distinctions in achievement conditional on reaching the same grade level in school.

The context in which we study these issues was one in which childhood malnutrition was significant, but declining. Children born a decade later and reaching their early adult years after World War II were significantly better nourished, and evidence of growth faltering is more limited. In part these improvements in nutrition and childhood growth resulted from programmes such as school lunches to improve children’s diets (Levine, 2010; Ruis, 2013). The cohort we study was not particularly affected by wartime conditions despite the motivations for the health survey. The stature and weight of our cohort measured in 1918 was, in fact, slightly better than cohorts born 10 or 20 years earlier. Our findings confirm using robust causal methods that widespread growth faltering and poor nutrition in urban areas in the early twentieth century United States had significant long-lasting consequences for children, and that the magnitudes of these effects were similar to those seen in low and middle income countries in the last two decades.