Can the names parents gave their children give us insights into how parents in historical times planned their families? In this study, we explore whether the names given to the firstborn child can be used as indicators of family-size preferences and, if so, what this reveals about the emergence of intentional family planning over the course of the demographic transition. We analyze historical populations from 1850 to 1940 in the United States, where early fertility control and large sample sizes allow separate analyses of the White and Black populations. We also analyze Norway from 1800 to 1910, where there was a much later fertility transition. A split-sample method allows automated scoring of each name in terms of predicted family size. We find a strong relationship between naming and family size in the U.S. White population as early as 1850, for the Black population beginning in 1940, and for the Norwegian population in 1910. These results provide new evidence of the emergence of “conscious calculation” during the fertility transition. Our methods may also be applicable to modern high-fertility populations in the midst of fertility decline.
Demographers and historians have long tried to measure the extent to which fertility levels are the result of intentional behavior. Louis Henry introduced the fundamental distinction between fertility that is “natural” and that in which childbearing is limited after a target family size is reached (Henry 1961). Levels of “natural” fertility can vary between populations, as a result of coital frequency, breastfeeding practices, and other proximate determinants (Bongaarts 1978). A major tenet of demographic transition theory is the irreversible transition from “natural” fertility to parity-specific control (Coale 1973; Coale and Watkins 2017). The onset of thinking of childbearing in terms of family-size targets has been said to represent a fundamental change for parents, the moment when fertility enters the realm of “conscious calculation” (Coale 1973).
The traditional way to measure departures from “natural” fertility is with the measurement of parity progression ratios (the conditional probability of having a child of order i, given a birth of order i − 1) for women who have completed their childbearing (David et al. 1988; Henry 1961). Sudden drop-offs in these ratios at higher parities beyond what can be expected from the effects of reproductive aging indicate the practice of parity-specific control. Coale and Trussell (1974) developed methods for inferring parity-specific control from the age pattern of marital fertility, with declines at older ages proxying for parity-specific control. Both of these approaches are indirect in the sense that they infer parental intentionality from the age and parity pattern of fertility behavior, not from statements by parents about the children they would like to have.
Direct reports of intentions are available in modern surveys such as the World Fertility Surveys and the more recent Demographic and Health Surveys. van de Walle (1992) reported that many women in pretransitional fertility populations did not think about childbearing in terms of numeric targets, instead replying to interviewers with answers such as “as many as god will grant.” In historical populations, statements of intended family size are extremely rare. van de Walle wrote:
It would be useful to have a survey taken in 1800 in England or in 1750 in France, in which women were asked: “If you could start your life all over again, how many children would you like to have?” Of course we have no such survey, but it is not entirely fanciful to consider whether the question itself would have made sense to our ancestors. (van de Walle 1992:497)
In this article, we demonstrate a new, indirect method that tries to reconstruct the answers to such a “survey” of family-size preferences. Our idea is to use the first names that parents chose for their children as informative of how many children a family intends, with certain names associated with large families and others predicting small families. Naming patterns offer a potential window into parental aspirations and ambitions for their children at the time of birth of their first child. With endless choices possible and no direct financial constraints limiting parents, the selection of a name “could” present an opportunity to express parental thinking at the key point in time of the starting of the family. Our approach takes advantage of the recent progress in transcribing historical censuses, which makes available the names of children and the number of children in the household (Ruggles et al. 2021).
Our approach of using population censuses offers a critical advantage in allowing us to go beyond the expressions of norms and desires of selective elites, such as those found in writings and art (van de Walle 1992), to consider the expression of child preferences by the entire population. We focus on historical census data for White and Black populations in the United States, where the fertility transition began relatively early. For comparison, we look at the Norwegian population, which experienced a much later fertility transition, to see if there was a pretransitional time when names and family size were unrelated. Our approach could be expanded to other countries, both for past and perhaps even to present transitions for those with access to modern censuses or administrative data systems providing first name transcriptions.
In the United States, we find that family sizes varied systematically with the name chosen for the oldest child, with parents of small families naming their oldest child names like “Kenneth” or “Donald,” while parents of large families were more likely to choose names like “Thomas” or “William.” Names, of course, follow the contours of social class (Lieberson 2010). But we find that even within economic and cultural groups, the choice of name is able to predict family size independently. Our interpretation is that parents in similar circumstances varied in the relative importance they placed on child quantity and quality (Becker and Lewis 1973)—where “quantity” refers to the number of children and “quality” refers to the amount of parental investment per child. These preferences are reflected in the names they gave their children and in their decisions about how many children to have and how much to invest in each child.
The study of children's names and fertility was pioneered by Hacker (1999), who manually classified names by their biblical origin and argued that naming reflected religiosity. Our study is different in that it builds on recent innovations in the automated classification and scoring of names based on their statistical distributions (Abramitzky, Boustan, and Eriksson 2020; Connor 2021; Fryer and Levitt 2004; Goldstein and Stecklov 2016).
Figure 1 provides a first overview of the relationship between names of the firstborn child and observed family size from 1900 for Whites in the United States. The panels show the average number of children in the household for each name, with gray bars indicating the relatively narrow 95% confidence interval for each mean.1 We call the average family size associated with each name the “F-score,” for “fertility score.”
Before using the quantitative scores, it is interesting to consider the family-size pattern of some specific names. In the girls panel, we see that “Helen,” “Ruth,” “Hazel,” and “Mildred” were all low-fertility names in 1900, whereas “Nancy,” “Cora,” and “Martha” were high-fertility names. For boys, we see that “Harold,” “Ralph,” and “Donald” were low-fertility names, whereas “Peter,” “Oscar,” and “James” were high-fertility names. There are hints of many patterns, such as more biblical (perhaps Catholic?) names among high-fertility names, but this pattern, like many others, has exceptions—in this case, the saint's name “Paul” is a low-fertility name. Over time, we find that the F-scores of particular names often vary. For example, by 1940 (not shown), “Mildred” had shifted from being the low-fertility name it is in 1900 to a high-fertility name.2
These exploratory plots reveal clearly structured variation in family size according to the name given to the first child, with differences averaging as large as two children per family between high- and low-fertility names. Why this might be, how it can be analyzed, and what it reveals about the history of fertility decline is the subject of the rest of this article.
Theory: Fertility Decline and Child Naming
The analysis of child name choice and family size offers a new approach for reconsidering long-standing theoretical debates in the fertility transition literature. As we explain in the following, it is not the names per se that are the focus of our interest but rather how the choice of child names sheds light on parental preferences for child quantity and quality over the course of the fertility transition. We argue that the link between naming and family size is evidence of “conscious calculation” of family size by parents.
Adaptationist and Ideational Theories
Theories of the fertility transition offer contrasting views of whether changes in attitudes and preferences were an essential part of the historical fertility transition (Goldstein and Klüsener 2014). According to the “adaptationist” perspective (Carlsson 1966), the primary change that led to fertility decline was one of new circumstances that inspired parents—without any necessary ideational change—to have smaller families. On the other hand, the ideational perspective, championed most prominently by researchers associated with Princeton's European Fertility Project (Coale and Watkins 2017), argued that a change in the consciousness of parents—the emergence and diffusion of the idea of fertility control—is essential for understanding the variation in the geography and timing of fertility control. These rather extreme perspectives can be, and are often, combined. Coale's emphasis on “conscious calculation” was part of a broader framework of “ready, willing, and able,” with importance given to circumstances, ideas, and contraceptive technology. Even Becker and Lewis's (1973) quantity–quality trade-off can be seen as a combination of changes in circumstances that alter the relative prices of child quantity and quality and social and ideational mechanisms that, for example, make it hard to reverse course once there has been a shift toward increased investments in children.
Under the purely adaptationist perspective, any link between child naming and family size would be fully accounted for by accurate measurement of socioeconomic status, and there would be no need for additional ideational change to explain changes in the costs and benefits of children. The ideational perspective, on the other hand, would predict a persistent unexplainable association of naming and family size. The choice of first name may then reflect the variability in preferences among parents who are otherwise similar, revealing their conscious decisionmaking that goes beyond simply adaptation to circumstances and revealing different ways of thinking about their children.
More recently, theorists (Bachrach and Morgan 2013) have argued that family size is just one part of a bundle of choices that parents make, guided by overarching “schemata” that structure their thinking and life choices. This perspective is useful for understanding how names and fertility might be linked, because in this framework the signification of a child's name is not limited to ideas about family size, but also extends to a larger set of ideas about parents' goals in life and what kind of people they want to represent themselves as being.
Putting these theoretical perspectives together suggests several ways to think about what first names can reveal about fertility decline. In its simplest form, an absence of association between names and family size would suggest pretransitional fertility in which there is little or no “conscious calculation” about family size, whereas the presence of an association would suggest active choices by parents in controlling their family size. A more complicated approach is to take into account the persistence of such effects net of socioeconomic circumstances. In this case, the theories allow us to distinguish between fertility control that simply reflects adaptation to this new reality and behavior that is also influenced by new ways of thinking about fertility, family, and the larger ambitions parents may have for their own and their children's lives. From this perspective, naming can reveal the orientation (or “schema”) of parents. Finally, our combination of theoretical frameworks also allows for fertility change without any change in naming: this might be the outcome if social mobility and ambitions were strongly constrained and fertility was primarily driven by a response to circumstances and conditions.
As we argued earlier, name choice offers a potential indicator of underlying family-size preferences, which are otherwise unobservable. A challenge in measuring preferences in this way is the confounding effect of social class and cultural background, which influence both naming and fertility. Figure 2 shows our conception of how preferences and status influence naming and fertility. Here, the term “preferences” is shorthand for the ambitions, desires, and intentions that may consciously influence fertility.3 We use “status” as shorthand for socioeconomic as well as cultural variables that might influence both naming and fertility.
In practice, we observe status, naming, and fertility, but we do not observe preferences. Without controls for status, the predictive power of names on fertility may reflect the confounding influence of status on naming and fertility, which is potentially supportive of the “adaptational” perspective. However, with effective controls for status, the effect of names, we would argue, reveals the effect of preferences on fertility outcomes, supporting the connection between ideas, preferences, and fertility decline advocated by the proponents of the ideational perspective.
We use two separate data sets for our three analyses. All analyses build directly on individual-level 100% full-count census files from the IPUMS project (Ruggles 2014; Ruggles et al. 2021). The first two analyses are based on the historical U.S. Census files, while the third analysis builds on historical census data from Norway from the IPUMS International project. Together, these files allow us to consider the relationship between name choices for oldest children and the number of children born in households in both pretransition fertility settings and in contexts where fertility levels have substantially declined.
Our main analysis is based on individual-level records from the U.S. decennial censuses of 1850 through 1940. We employ restricted-use 100% data of these transcriptions, made available by the IPUMS project (Ruggles 2014; Ruggles et al. 2021). Because these census records are beyond the 72-year confidentiality threshold, they include the full transcriptions of first and last names as written down by census takers and transcribed by the IPUMS project. Additional variables include urban/rural residence, father's occupation, ages of all household members, sex, state and county of residence, and last names.
In our initial analysis based on U.S. Census data, we focus on native-born, White respondents (based on the father's characteristics), with the goal of reducing confounding between group-specific naming practices and the differential fertility among immigrants and natives as well as those between mostly White, majority groups and other racial/ethnic groups. The 100% data also allow us to do a separate analysis of African Americans, who have a different fertility trajectory (Tolnay 1981) and who experienced both slavery and intense discrimination during the early parts of the fertility transition in the United States. We consider the relationship between naming and fertility separately for Black individuals.4
In addition to the United States, we extend our investigation to Norway, which has transcribed censuses beginning in the early nineteenth century. We make use of the publicly available data from IPUMS International (https://international.ipums.org/). This analysis allows us to explore the relationship between naming and individual-level fertility in a pre-fertility-transition context. The Norwegian data set is smaller and does not have harmonized variables for occupational income. We use province-level fixed effects for Norway, in contrast to county-level effects for the United States. We do not use last names for the Norwegian analysis because of patronymic naming practices (Abramitzky, Mill, and Pérez 2020).
Measurement of fertility from historical census data requires a careful balance of assumptions because census files rarely collect the sort of information allowing straightforward measurement of childbearing. Our measure of fertility is the count of children enumerated in the household. To make this measure more meaningful, we limit our sample to households in which the mother is aged 35–44 and in which the oldest child is 21 or younger. This restricts us to women old enough to have begun childbearing but also limits the number of children who may have already left the household.5
The use of “children in the household” as a measure of lifetime fertility has several potential issues. Children may have died before enumeration, children may have left the household before enumeration, and more children could be born after enumeration. To check our fertility measure, we repeated our analysis for 1900 and 1910, when census questions were asked on children ever born and children surviving. By restricting the reanalysis to households in which all of the children ever born are enumerated, we are more certain that the eldest child is indeed the firstborn and that the number of children in the household is the number of children ever born. We find that the results of this analysis are essentially identical to those using the unrestricted samples (see Table A2 in the online appendix), supporting the use of the number of children in the household as a useful, even if imperfect, indicator of the number of children ever born.
Our analysis uses the name of the oldest child as a proxy measure of the parents' lifetime fertility intentions. The names of later-born children may reflect the effect of birth order on names, with higher birth order children receiving systematically different names. By using only the name of the first child, we assure the temporal ordering of cause and effect, with the name of the first child preceding the arrival of subsequent births.
We use the first names transcribed in the IPUMS data. The names were first reported by census respondents to census takers, who wrote them in this era of data collection in long-hand on census manuscript forms. These forms were then transcribed (Ruggles 2014). Our data-cleaning approach involves dropping all individuals whose given names begin with a single initial. We recode abbreviations and nicknames back to standardized names to best approximate the names given at birth.6 Thus, for example, “Charlie,” “Charley,” “Chuck,” and “Chas.” are all coded as “Charles.” We also remove first names that are relatively rare, which removes some misspellings or possible transcription errors. We limit our analysis to first names that appear 30 times in the 50% training file.
Fertility Scores for Names
A challenge of working with name data is the large number and heterogeneity of first names and the many dimensions and characteristics of names. Determining which features are important for parents in terms of thinking about the quality and quantity of children they would like is not obvious. One approach would be to use human coders to classify key features of names (Gerhards and Hans 2009; Sue and Telles 2007), but it is not clear what these features would be, particularly in historical settings in which the valence of names may not be familiar to contemporary coders. In contrast, our approach uses information built into the data to score names according to one or more characteristics. This approach was employed in early work in economics to study discrimination (Fryer and Levitt 2004) and has been further developed in demographic studies in recent years (Abramitzky, Boustan, and Eriksson 2020; Connor 2021; Goldstein and Stecklov 2016).
Our current approach aims to address both limitations noted above that are inherent in using coders. We use an automated approach, using split-sample methods to assign and test the fertility scoring of names. We create a training data set by randomly selecting 50% of the households in each census. We estimate fertility scores calculating the average family size of households by eldest child's name in the training data. For example, households with an oldest child named “John” may have an average size of 3.2, whereas households in which the oldest child is named “Kenneth” may have an average size of 2.4. The fertility scores from the training data are then assigned as independent variables in the test data.
The training–test split-sample approach allows a principled approach to separate systematic from random variation and is increasingly common with supervised machine learning approaches. This method is reapplied to each round of census as well as to subpopulations in our analysis (Black and White estimates are derived separately). For relatively common names, the variation in F-scores we find in our training sample is likely to reflect true differences in mean fertility at the population level. But for many low-frequency names, the F-scores estimated in the training data will include a substantial amount of chance variation. Splitting the sample and reporting only the results of the regressions on the test sample (which was not used to estimate the F-scores) allows us to measure the predictiveness of names on family size without risk of overfitting.7
We use ordinary least-squares regression to predict the number of children in the household (“fertility”) and its relationship to the F-score (“FScore” in the equations below) value of names. For individual i, whose oldest child has a name with FScorei and a set of characteristics Xk,i, we estimate the set of coefficients β0, β1, . . . using the estimation equation
In certain cases, we also use a second model to control for within-family effects. This model uses sibling fixed effects for brothers from the same origin household j. The ith brother from origin household j is modeled as
Our statistical analysis proceeds in three parts. First, we estimate the F-score values of family size by name of the oldest child in the training sample and create an additional variable in the test data using these scores. For example, if the average size of families in the training data with an oldest child named “Kenneth” was 2.4, then this same score would be applied to families in the test data with oldest children named “Kenneth.”
Second, we examine whether the estimated name scores are predictive of family size in the test sample. We do this by regressing F-score on the family sizes of the set-aside test data. The criterion for prediction is whether the coefficient on F-score is statistically distinguishable from zero.
Finally, we introduce a series of control variables to explore the extent to which the effect of names on family size is due to the tendency of names to be associated with other predictors of family size and the extent to which names appear to have an independent effect. The goal of this third analysis is to see if the choice of name for the eldest child is predictive of ultimate family size when comparing families that are otherwise similar in every measurable way but that might have different preferences for child quantity relative to quality. We are using the terms quantity and quality in the sense used by Becker and Lewis (1973) to differentiate investments in additional children versus higher per-child investments. If the effect of names persists after introducing controls, then we argue this is evidence that the choice of first name by parents is revealing the role of parental preferences (for quality versus quantity) in influencing family size, providing the measure of intentionality that we seek.
The transcribed censuses allow us to control for a wide set of important influences, including small geographic units, father's occupational status, and last name fixed effects that proxy for ethnic, cultural, and religious origins.
In all of our models that include control variables, we include measures for the age of the eldest child. This is important to do, since family size will tend to be mechanically correlated with age of the eldest child (if the oldest child is age 2, then he or she cannot have as many siblings as if he or she were 12 years old). Because name frequencies change over time, leaving out controls for age will make names that are rising in popularity—and thus are more common among young children—appear to be linked to small family size. This is a potential confounder in earlier research on biblical names, although Hacker (1999) did limit his analysis to eldest children aged 5–9.
The most stringent test we have is to compare the family sizes of adult siblings. Because siblings share a large number of unobservable background influences, evaluating the effect of names on fertility outcomes controls for many effects that otherwise might confound our estimates. If the first names given to the eldest child of each of two brothers are predictive of the difference in family sizes of those brothers, then we consider this very strong evidence that first names are providing a choice-based signal that can be used to confirm intentionality of family size. Of course, since the desire to invest in more educated but smaller families may be transmitted to both siblings, the sibling comparison may in fact be an overcorrection in that it may remove much of the variation that we intend to capture in the explanatory side of the model. In this context, the sibling fixed-effects model may be an overly conservative approach, but one that helps us to consider the robustness of our findings.
The sibling fixed-effects analyses are based on substantially smaller samples, which is a limitation owing to the inherent challenge in matching sibships from earlier census rounds. We identified male siblings in 1870 aged 5–15 using the 1900 census to see how many children each brother had had. The 1870 and 1900 censuses gave us a sample of more than 53,000 fathers in 1900 who were the offspring of families in 1870. We then analyzed whether first names of oldest children, based on the F-scores calculated in our earlier 1900 analysis, were predictive of the number of children in the household in 1900. The children of the brothers are cousins, but our comparison is of the family sizes of brothers. We repeated this procedure for two additional rounds as well. We estimated fertility in 1930 by matching siblings in 1900 and fertility in 1940 by matching siblings in 1920.
We begin by presenting the results for native-born U.S. White individuals, showing a variety of models for 1900 (Table 1) and then applying the most detailed model to each of the separate censuses available from 1850 to 1940 (Table 2). These first analyses show that naming patterns have consistently predicted fertility, even after controlling for geographic and socioeconomic covariates. Next, we present sibling models to show that the predictive power of names on number of children is present even within families (Table 3).
To look at other examples of the connection between naming and fertility—which we interpret as evidence of intentional childbearing behavior—we extend our analysis first to the post–Civil War African American population in the United States from 1870 to 1940 (Table 4) and then to Norway (Table 5), for which censuses are available from 1801 to 1910.
Overall, our findings suggest that in all three of these groups, names eventually became predictive of family size. The U.S. native-born White example suggests that intentionality has been a feature of fertility since at least 1850. The U.S. African American example suggests that names were not correlated strongly with fertility until very late in the fertility transition, perhaps reflecting the limited opportunities of the children of African American parents. The Norwegian data provide an example of a society in which the link between naming and fertility was entirely absent before the demographic transition but emerged with the onset of fertility decline.
Native-Born U.S. White Population
Detailed Analysis of 1900
The basic results of our analysis based on the full-count IPUMS census file for 1900 are shown in Table 1. The first column, Model 0, verifies the correct estimation of F-score for the training data, showing the expected coefficient of 1.000. The R2 value indicates that first names predict about 4% of the variation in family size in the training data, comparing favorably with the explanatory power of other important predictors of fertility that are introduced in subsequent models. For example, occupational income and last name each additionally explain about 1% of the variation in family size (see adjusted R2 values in Models 4 and 5).
The second (Model 1) and later columns (Models 2–4) regress the F-scores estimated in the training data on observed family sizes in the set-aside, test data set. The bivariate regression coefficient of 0.962 tells us that a one-child change in F-score—that is, in having an eldest child with a name that has an average completed family size that is higher by one child—is associated with a change of nearly one child in average family size in the out-of-sample test data. The important result here is not the value of this coefficient—which is largely determined by sample sizes—but rather its statistical significance. The coefficient is many standard errors from zero, demonstrating clearly that the predictive power of first names in the set-aside data is not due simply to chance. The squared correlation coefficient in this model is now reduced to 3.7%, slightly less than the original, showing that a small portion of the predictive power of names was indeed overfitting and does not carry over from training data to the test data.
The next step in our analysis is to assess whether the association between names and fertility is due to confounding variables such as time trends in naming, geography, and socioeconomic status. Model 2 repeats the regression on the test data with the addition of a control variable for age of oldest child. We find that including this variable greatly increases the predictive power of the regression—owing to the mechanical effect of, for example, 13-year-olds having more younger siblings than 3-year-olds.
The apparent signal of family size from names could result from a correlation between the pattern of name-giving and known predictors of fertility, such as urban/rural residence and other measures of socioeconomic status. On the other hand, if first names retain an independent influence on fertility beyond these covariates, then names may be informing us about the otherwise unobserved preferences of parents for family size and child investments. The attribution of the F-score effect to unobserved preferences depends on the degree to which one believes one has completely controlled for other covariates. In our initial models (3 and 4), we use basic demographic controls, yet in subsequent models, we add a number of fixed effects including geographic area (states and also counties), last name, and even estimation of sibling fixed-effects models, comparing the family size of brothers by name of their firstborn children.
Models 3 and 4 add more substantive, potentially confounding, variables. “Geography” includes rural/urban status, state of residence, and county fixed effects; “occupation” is a measure used to capture the father's occupational income score (Ruggles et al. 2021). These variables are predictive of family size in the expected manner, with urban dwellers on average having about 0.5 fewer children in the household, and an additional $1,000 of father's occupational income—a useful metric given that it is slightly below a standard deviation increase in the father's income score variable—predicting 0.2 fewer children. The geographic and occupational predictors appear to be responsible for part of the observed effect of choice of first name. The coefficient on F-score declines to about 0.32 but remains highly significant, clearly distinguishable from zero.
We interpreted the correlation between naming of the firstborn and subsequent family size in Table 1 as evidence of intentional family planning behavior. An alternative explanation is that imperfect controls for background characteristics related to both naming and fertility may be responsible for the effect we are finding. To address this, we estimated several additional models. First, we estimated a model including fixed effects for last names, allowing us to control for last name–related cultural variables such as ethnicity, religion, and country of origin that are known to influence name choice and fertility (Model 5). Our large sample sizes allow us to estimate effects of first names among individuals with the same last names (e.g., O'Donnell, Lombardi, Smith, or Cohen). The results from Table 1 show evidence of stability of our F-score coefficient, with only modest decline in its magnitude and continued high significance, and signal the robustness of the main effect we are considering despite aggressive controls.
Change Over Time
Table 2 shows the results of estimating the last name fixed-effects model (similar to Table 1, Model 5, for 1900 though with a smaller sample size) for all available full-count census data from 1850 through 1940. We equalized sample sizes using the smallest available sample across years, based on 1850, to make the coefficients on F-score more comparable. These models also include dummy variable coding for the age of the oldest child as well as for each state.
Two key points emerge in Table 2. First, the evidence supports the effect of F-score on fertility throughout the entire period. For Whites in the United States since at least 1850—the earliest transcribed census with names—name choice of the eldest child has predicted family size, even after controlling for urban residence, occupational income of the father, state of residence, and all of the cultural, religious, and socioeconomic information contained in last names. The second key point is that the coefficients in Table 2 offer evidence on temporal trends. The size of the F-score coefficient rises from 1850 to 1900 and then falls slightly in subsequent census rounds between 1900 and 1930 before rising again in 1940.
The rise and subsequent decline of the importance of fertility intentionality as inferred from child naming is broadly consistent with increased differentiation during the fastest part of the fertility transition (Coale and Zelnik 1963; Hacker 2003). However, the magnitude of the coefficient also reflects the precision of the F-score estimates, itself a function of the frequency distribution of names in the training sample. We are thus cautious about any strong interpretation of coefficient values across populations or time periods. Interestingly, the effect of father's occupation, as measured by occupational income score, is remarkably constant from 1850 to 1940, with each additional $1,000 in occupational income (in 1950 dollars) being associated with 0.1–0.2 fewer children. Occupation thus seems to have had a consistent structuring effect on fertility throughout the course of the demographic transition (Jones and Tertilt 2008). Urban/rural status has a large effect throughout all periods, although reaching its highest value in 1930. Overall, our analysis from 1850 to 1940 tells a story of stability in the role of these various factors influencing fertility. Of course, important changes within the American population—in terms of urbanism, occupation, and the ambitions that parents had for their children—did occur over time.
Sibling Fixed Effects
We adopt a more conservative approach to further ensure identification of the effect of names in light of unobservable and potentially confounding factors. Our approach takes advantage of the longitudinal nature of the repeated censuses using sibling links between pairs of men in the 1900 census, based on being identified as brothers as children in the earlier census of 1870 (Abramitzky, Boustan, and Rashid 2020).8 Estimation of a within-sibship model allows us to effectively control for omitted socioeconomic status and cultural characteristics that distinguish families by using the extent these characteristics are shared by siblings. Our estimate of the effect of first names is then based on the variation in both naming of firstborn children and fertility outcomes within adult sibling groups.
Table 3 shows our estimates of the impact of F-score on the number of children that brothers have for each of the three census rounds. In each case, we present two models: the first shows the sibling fixed-effects model and the second includes sibling occupational income and rural/urban residence. Results are similar in all three time periods, with the name of the first child (F-score) being predictive of family size in both controlled and uncontrolled models. The model for 1900 shows the effect of F-score using a sibling fixed-effects model with only age of oldest child as a control. The effect of F-score (about 0.14) is about half the magnitude of the surname fixed effect estimated earlier (Table 1, Model 5). Adding controls for each sibling's occupational income and urban/rural residence changes the coefficient only slightly (to about 0.13). Subsequent models for 1930 and 1940 show similar results, with larger sample sizes reducing standard errors and increasing statistical significance. Overall, this sibling analysis provides additional evidence that the predictive power of the name of the first child on family size is due to differences in intentions among otherwise similar individuals.
U.S. Black Population
The strong and robust findings for White individuals raise the issue of whether names are always a reliable indicator of fertility decline. The African American population began seeing declines in total fertility in the late nineteenth century (Tolnay 1981). The large sample sizes of the U.S. censuses allow us to carry out a separate analysis for African Americans, using F-scores specific to the Black population to account for distinctive naming practices (Cook et al. 2014; Lieberson and Mikelson 1995).
In Table 4, we see that the covariates of urban/rural residence and occupation consistently predict fertility starting in 1870. However, names are at best very weak predictors of fertility. From 1870 to 1930, the effects are all positive but small and statistically (when each is considered individually) indistinguishable from zero. Only in 1940 does the coefficient reach statistical significance.
One could interpret the small effect of names, despite the decline in fertility, as a refutation of our general claim, that names are evidence of conscious calculation necessary for fertility decline. However, one could also interpret the small effects within the framework of parental ambition. The theory we put forward suggests that when parents see the potential for investing more in their children, then they may choose to make a quantity–quality trade-off. However, for African Americans, whose access to education and more generally upward mobility was extremely limited, the quantity–quality trade-off was less salient (Card and Krueger 1992; Tolnay 1981). While fertility was limited as African Americans became an increasingly urban, nonfarming population, as evidenced by the coefficients on occupation and urban/rural residence, it appears that there was not the same kind of differentiation within these groups by name (and inferred ambition) that we saw among White individuals. Without opportunity, it may be the case that there was less of an intentional trade-off between quantity and quality, and less of a connection with names.
We now turn to an analysis of a population for which we have a longer history of transcribed censuses. This is more challenging than it might seem, given that once a state has a full-functioning and complete census apparatus it is likely already undergoing some key dimensions of modernization. Our goal is to see if there was ever a “pretransitional” society in which names were not predictive of fertility at an individual level. We chose to work with Norway because (1) it had some of the earliest transcribed censuses, dating back to 1801 (see the North Atlantic Population Project (NAPP) page: https://www.ipums.org/projects/ipums-napp), and (2) the fertility transition in Norway happened relatively late within Europe, with most observers dating it to the end of the nineteenth century (Shorter et al. 1971) or even the early twentieth Century (Sogner 2003).
Our analysis of Norway has some limitations. First, the number of individuals is much smaller than in the United States, even when using the 100% census. Second, there is no standardized occupational income variable.9 Third, Norwegian naming had its own idiosyncratic traditions of naming, which tended to be conservative. In particular, it was common to name children after grandparents (Abramitzky, Boustan, and Rashid 2020).
Table 5 shows the results of regressions for Norwegians, and they are similar to those shown earlier for White and Black Americans. The controls are limited to geography (at the province level beginning in 1865), age of eldest child, and urban/rural residence. The general finding is that there is barely any evidence of a link between child naming and fertility before 1910, and the first emergence of a significant link appears in 1910. None of the coefficients on F-score before 1910 are significant. Even urban/rural fertility is not significant in 1865 and 1875. Then, in 1910, names start having predictive power, as does urban/rural residence.10
The general pattern suggests very little structuring of individual-level fertility through 1900 and the emergence of intentional fertility control only around 1910. It is true that name choice may have been limited in the Norwegian context, but the simultaneous emergence of fertility predictability from names along with fertility decline would suggest that the changes in fertility coincided with a new pattern of differentiation.
The main finding of this article is that eventually the names given to firstborn children are predictors of the family size in all of the populations we studied. The predictive power of names persisted after introducing controls for socioeconomic and cultural confounding, and even when comparing the number of children that brothers have, which we were able to do for the U.S. White population. Our findings, we argue, constitute evidence of emergence of intentionality in reproductive decisions, with parents' choices at the time of birth of their first child being connected to their eventual family size. Parents are not answering van de Walle's hypothetical historical survey question about how many children they would like in a direct manner, but they are giving us enough information to measure their intentions.
The approach we introduce here shows how thousands of distinct names can be scored and used as a predictor of family size. Our methods, which we applied to populations from the nineteenth and first part of the twentieth centuries, are also potentially applicable to modern populations, for example, those in sub-Saharan Africa that are still going through their first fertility transition. One contribution of our study is to show how qualitative textual data like names can be used in quantitative demographic research.
Each of the three cases we studied revealed a different aspect of the relationship between parental ambitions, intentions, and fertility outcomes. In Norway, there was clearly a pretransitional era, before fertility began to decline, when there was no relationship between naming and fertility. Only at the end of the nineteenth century did differentiation among large and small families become associated with name choices. The Norwegian case illustrates the emergence of intentionality as fertility begins to fall.
In the United States, the native-born White population shows a clear association between names and fertility since at least 1850, the date of the earliest nominal census. The United States is known as an example of early onset of fertility decline (Hacker et al. 2021),11 and the continuous predictive power of names from 1850 through 1940 is consistent with this. The richness of the U.S. census data also allows us to include a fuller set of covariates as well as to compare the child name choices and family sizes of brothers, demonstrating that the different preferences of people who are otherwise quite similar are also predictive of the number of children they have.
Finally, in the African American population in the United States, despite having a long history of fertility decline (Haines 2008; Tolnay 1981), there is little or no association of child name choice with family size until 1940. This is not a statistical issue arising from small sample sizes, since the sample sizes are larger than in Norway. Rather, we suspect that the independence of names and family size is due to the constraints faced by African American parents. White parents had the choice to invest more in a fewer number of children, and those who chose smaller families were likely to see a payoff in terms of their children's occupational and life success; but Black parents, living in a highly segregated, racist environment, had fewer opportunities to invest in their children via schooling, and any investments they could make were less likely to provide upward mobility for their children. Thus, African American parents reduced fertility, perhaps as the costs of raising children rose, but they did not do so in a manner that signaled the kind of “lifestyle choice” that White parents appeared to be making.
Regarding the classic debate between ideational and adaptationist explanations of historical fertility decline, our findings are generally supportive of the ideational perspective, although they do not exclude that adaptation also played an important role. In the analyses for Norway and for the U.S. White population, it seems clear that even for those in similar circumstances, differentiated preferences for family size emerged as part of the demographic transition, with an increasing number of parents choosing to have smaller families and to invest more in each child.
Although our analysis, like any observational study, cannot compare individuals who are perfectly alike in every respect, our goal in controlling for economic and cultural variables and in comparing sibling outcomes was to show that it was not just the differences in circumstances that parents faced that led them to having larger or smaller families, but it was also differences in preferences themselves. In his quantity–quality model, Becker emphasized the role of price differences of child quantity and quality in explaining fertility change over time (Becker and Lewis 1973). Here, we try to control for such price differences, explaining individual variation in terms of differences in preferences.12
Finally, the connection between child names and family sizes is consistent with the reasons for fertility decline put forward by early observers such as Arsène Dumont (Dumont 1890) in France and Warren S. Thompson in the United States (Thompson 1942). What we find is that the orientation of parents, as captured by the names they gave their children, was an important factor in fertility decline. As Warren Thompson put it:
If we use the term “ambition” to describe the state of mind aroused in people by the desire to bridge the gap between their actual status and that which they would like and by the will to maintain and transmit to their children a more or less enviable position already achieved, then, in the author's opinion, personal ambition is a basic cause of the decline in the birth rate [emphasis added]. (Thompson 1942:206)
More modern, less traditional naming was an expression of this ambition. In the context of Ireland, recent work has shown that the factors driving the slow pace of decline of fertility were tied to traditional norms, evident also in names, but that they were largely rooted in lagging levels of urbanization and the relationship to “new information and aspiration” (Connor 2021).
The large-scale use of text strings in census data provides a new opportunity for investigating the causes of demographic behavior. This study has shown how child names can be used, when we are able to harness sufficiently large data sets, to gain insight into what was in the minds of parents as they were starting a family. We used an automated procedure for scoring names and did not delve further into the many interesting features and characteristics of names. Alternative approaches are given in the example by Connor (2021) in Ireland, which compares name frequencies across generations, and in the early paper by Hacker (1999) using biblical naming.
The methods used in this study could be improved. Further studies that included complete lists of children ever born for women who had completed childbearing could improve the measurement of fertility. A methodological weakness of our current statistical method is that while our automated split-sample scoring methods are useful in testing whether names have predictive power, they are less useful in comparing the strength of prediction across populations. This is because the variance explained by names is a function of both the between-name variation in family size and the precision of our F-score estimates. The latter depends on sample sizes and name frequency distribution of the training data. We have standardized comparisons by using the same sample sizes across time periods, but our methods still do not take into account the changing distribution of name frequencies.
In terms of future research, we believe that the analysis of naming patterns at a population level offers promise as an early indicator of social change that leads to fertility decline. For example, a number of studies have found that the content of soap operas seems to influence fertility decline in India (Jensen and Oster 2009). The analysis of names may provide more evidence of this connection. In Africa, the study of naming patterns in the midst of what appears to be a stalled fertility transition (Bongaarts and Casterline 2013; Shapiro and Gebreselassie 2013) may provide additional insights into the possible resumptions of fertility decline.
We benefited from outstanding research assistance from Maria Osborne.
Here, as in the analysis, we limit households to those in which mothers are aged 35 to 44. These figures use the 100% data. In our train–test framework used later in the paper, we estimate scores based on a subsample.
This pattern of nontraditional names being associated with smaller families was explored by Connor (2021), who used relative frequency in the previous generation rather than some fixed feature (i.e., whether the name appears in a list or fixed reference) to define which names are “traditional.”
In general, preferences can be influenced by status, but here we are narrowing the focus to differences in preferences among people of equal status. In economics, such preferences would be the values of parameters of a utility function that would predict different choices by people facing the same prices.
We do not separately study immigrants, leaving the potentially interesting analysis of the relationship of child naming and fertility among immigrants for future research.
Following a reviewer’s suggestion, we checked to see if our results would be affected if we restricted our analysis to women aged 40 and older, who would have been nearly finished with their childbearing. The results in Table A1 (online appendix) show that our findings are not sensitive to the exclusion of younger women ages 35–39.
Some small portion of nicknames will actually be given at birth. So our approach is conservative, omitting some potential differentiation in names in order to preserve temporal ordering and be consistent with the logic of predicting later outcomes (family size) with earlier predictors (name given to first child).
The problem is much like the estimation of the predictive power of single-nucleotide polymorphisms (SNPS) in Genome-Wide Analysis (GWAS), in which some of the SNPS that appear to be predictive are the artefact of chance occurrences. The solution in the GWAS case is to test whether the polygenic score estimated from all loci in one data set is still predictive when applied to another population. Statistical significance in this later case is interpreted as validation of the predictive effect of the polygenic score. Similarly, the predictiveness of the name of the first child on family size can be measured by whether regressions of F-score on family size are statistically significant in the set-aside test sample.
The link option we chose was ”ABE-NYSIIS conservative.”
In the U.S. sample, this is a constructed IPUMS variable.
The small sample sizes mean that the absence of statistically significant effects is not definitive evidence that there is no effect, but we note that when we artificially limit the U.S. analysis to sample sizes similar in size to that of Norway, we do find statistically significant associations between F-score and family size in the United States.
Indeed, Hacker et al. (2021) found evidence of even earlier fertility decline, beginning in 1830.
In Becker’s language, this would be differences in the elasticity of substitution between quantity and quality, which would be parameters in individual utility functions.