Internal migration is fundamental to the American narrative. It has been seen for centuries as a tool for individuals to improve their economic situation (de Tocqueville 1835 [Reference de Tocqueville, Harvey and Winthrop2000]; Turner Reference Turner1921; Ward Reference Ward2022), with “great opportunities [lying] just over the horizon” (Brooks Reference Brooks2003, p. WK15). It was also a fundamental driver of the United States’s transformation into the world’s industrial and economic powerhouse, providing an avenue for labor to reallocate to the most productive sectors (Caselli and Coleman 2001; Kuznets Reference Kuznets1966).Footnote 1 Because of its importance on both the micro and macro scales, internal migration has been the subject of large literatures in economics, economic history, and history, which have provided insights into the nature, causes, and effects of internal migration in specific contexts in U.S. history.Footnote 2 But constraints on the data previously available to study internal migration have severely limited scholars’ ability to study its long-run patterns, leaving significant and fundamental blind spots in economic historians’ understanding of this formative phenomenon in U.S. economic and social history.
This paper documents and describes, for the first time, the rates of, selection into, and destination choice patterns of native-born white men’s inter-county migration in the United States over the period 1850–1940. My analysis is enabled by recent advances in the availability of complete-count data from the U.S. censuses of this period (Ruggles et al. Reference Ruggles, Sarah Flood, Ronald Goeken, Schouweiler and Sobek2021) and in the technology by which to make links between them (Abramitzky et al. Reference Abramitzky, Leah Platt Boustan, Feigenbaum and Pérez2021a; Bailey et al. Reference Bailey, Cole, Henderson and Massey2020). Building on these advances, I construct 13 datasets linking native-born white men, aged 18–40 when first observed, over all possible 10- and 20-year spans in the period 1850–1940.Footnote 3 These datasets enable me to overcome limitations faced in prior studies of the long-run trends in U.S. internal migration: only with linked data is it possible to observe inter-county migration, to separate the flow of migration from its stock, and to measure the selectivity of migration.
I find that the rates of inter-county migration of the native-born adult white male population were remarkably stable between the 1850s and the 1920s, at about 33 percent for 10-year spans and about 40 percent for 20-year spans. Selection into migration on the basis of occupational rank was also largely constant over time, with migrants either neutrally or slightly negatively selected.
This constancy contrasts with substantial changes over time in the orientation of internal migration, coming from changes in internal migrants’ origins and destination choice patterns. Both the deterrent effect of distance in destination choices and the relative attractiveness of the west increased over the study period. At the same time, the average distance of a move declined, and intra-state moves grew to comprise a greater share of inter-county moves, implying that a focus on inter-state moves alone would miss an increasingly large share of migration.Footnote 4 Most strikingly, the relationship between internal migration and urbanization changed over my study period. Urbanites were initially more likely than observationally similar ruralists to migrate, but by the twentieth century were either less likely or approximately as likely, depending on the definition of an urban place. The attractiveness of urban areas as destinations for internal migrants also increased over time. The combination of these patterns resulted in a steady increase from the beginning of my study period through the 1920s in the degree to which internal migrants’ increase in urbanization over a linkage span exceeded that of stayers. That is, internal migration increasingly became a force driving the urbanization of the economy.
The 1930s marked a change in these patterns in all respects. Selection on the basis of both occupational rank and initial urban status was moderated relative to the earlier twentieth century. More dramatically, a substantial decline in the rates of inter-county migration occurred for the first time, with 10-year migration rates declining by nearly 8 percentage points, or about 25 percent. This decline was coupled with a substantial retrenchment in the degree to which the urbanization growth or labor demand growth experienced by internal migrants exceeded that of stayers as urban areas became less attractive as destinations. That is, whereas internal migration was a force driving urbanization in the earlier parts of the twentieth century, this was not true in the 1930s.
The contribution of this paper is predicated on the advantages arising from the newfound ability to make links between all complete-count U.S. censuses from 1850 to 1940. But recent scholarship (e.g., Abramitzky et al. Reference Abramitzky, Leah Platt Boustan, Feigenbaum and Pérez2021a; Bailey et al. Reference Bailey, Cole, Henderson and Massey2020) has brought attention to potential bias due to false links arising from automated linking methods. This challenge is particularly apposite in studying internal migration because any incorrect match will, in all likelihood, appear as an observation of inter-county migration. This challenge will spuriously increase observed migration rates, conflate selection into migration with selection into false matching, and confound true destination choice patterns with spurious ones generated by false matches. To address this concern, I repeat the main results with alternative matching methods of various strictness and draw only conclusions that are robust to the choice of method. For my estimates of the rates of inter-county migration, which are most sensitive to this danger, I also propose a method to estimate the rate of false matches directly and to correct my estimates for them. The principle of this method is that the difference between two estimates of the same quantity, one of which depends on linkage and the other of which does not, is informative of the rate of false linkage. Specifically, I use information on the ages and birthplaces of children in the household to generate an alternate measure of inter-state migration that does not require linkage (Collins and Zimran Reference Collins and Zimran2019; Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004), though it can be applied only to a select sample of individuals.Footnote 5 Comparing this measure to that arising from linkage enables me to estimate the rate of false matches for each linkage method and to correct my estimated inter-county migration rates. The resulting estimates are largely invariant to the strictness of the linkage method.
The main contribution of this paper is to update, deepen, and expand existing descriptions of U.S. internal migration in the period 1850–1940 (Ferrie Reference Ferrie1997a, 2006a, 2006b; Hall and Ruggles Reference Hall and Ruggles2004; Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004).Footnote 6 Prior work in this vein has relied on unlinked census data and on information on individuals’ state of birth. It has therefore only been able to quantify the stock of inter-state migrants or the rate of inter-state migration of individuals with young children over spans of 10 years or shorter, and has been extremely limited in its ability to describe the selection and destination choice patterns of migrants. By using linked census data, I am able to provide the first comprehensive description of the rates, selection, and sorting of U.S. inter-county migration. Indeed, given the constraints on prior studies, this is the first description of the rates, selection, and sorting of internal migration flows at any geographic level that is not limited to families with children. On the whole, my findings provide an entirely new view of U.S. internal migration, documenting, for the first time, facts that are interesting and important for their own sake and for the better understanding of U.S. history that they provide. Importantly, the story of internal migration arising from my analysis differs from the one arising from earlier studies of U.S. internal migration over this period—constant rather than declining in frequency before the 1930s and neutrally or slightly negatively rather than positively selected—providing a different interpretation of U.S. internal migration and its interaction with the development of the U.S. economy.
The main limitation of this contribution is that I constrain attention to native-born white men. This is in part necessitated by the shortcomings of the data. Name changes at marriage make linkage impractical for women,Footnote 7 and the vast majority of the black population was enslaved in 1850 and 1860 and therefore was not included in the census and cannot be linked. Beyond data constraints, I exclude immigrants from the analysis because Zimran (Reference Zimran2022c) has already studied immigrants’ internal migration in the United States in general and in comparison to that of natives.Footnote 8 I exclude black men from the analysis, although in principle they could be included from 1870 onward, for three main reasons. First, doing so enables me to restrict attention to the same population throughout the study period. Second, internal migration patterns among blacks have already received substantial scholarly attention, including with linked census data.Footnote 9 Finally, as Zimran (Reference Zimran2022c) shows for immigrants, the unique questions surrounding the internal migration of blacks and the comparison of these patterns to those of whites merit attention beyond the scope of the present paper.
In establishing fundamental facts regarding long-run patterns in the internal migration of native-born white males, this paper also adds to several literatures beyond that seeking to describe long-run trends in internal migration. First, it contributes to the large literature studying specific instances of internal migration in the United States, such as frontier migration (Ferrie Reference Ferrie1997b; Stewart Reference Stewart2006), the Great Migration of African Americans (Collins and Wanamaker (Reference Collins and Wanamaker2014, Reference Collins and Wanamaker2015), and Dust Bowl and Great Depression migration of the 1930s (Boone and Wilse-Samson 2023; Fishback, Horrace, and Kantor Reference Fishback, Horrace and Kantor2006; Gutmann et al. Reference Gutmann, Daniel Brown, Cunningham, Susan Hautaniemi Leonard, Jeremy Mikecz, Rhode and Sylvester2016; Hornbeck Reference Hornbeck2023; Long and Siu Reference Long and Siu2018; Sichko Reference Sichko2024). Although it has been possible to study these specific instances of migration in detail, limited understanding of the broad patterns of U.S. internal migration implies that the context into which the findings of this literature fit has not been clear. This paper brings this backdrop into sharper focus. This paper also dovetails with papers describing the rates of, selection into, and sorting patterns of modern internal migration (e.g., Greenwood Reference Greenwood1975; Jia et al. 2023; Molloy, Smith, and Wozniak Reference Molloy, Smith and Wozniak2011; Sprung-Keyser, Hendren, and Porter Reference Sprung-Keyser, Hendren and Porter2022). In combination with them, it enables a description of internal migration patterns in the United States over nearly 175 years. Moreover, given the weight assigned to internal migration in the literatures on U.S. economic growth (Kuznets Reference Kuznets1966) and on intergenerational mobility (e.g., Ward Reference Ward2022), the clearer understanding that this paper provides of internal migration adds to these literatures.
Finally, in exploiting linked data, seriously considering the biases that might characterize them, and proposing a method to estimate and adjust for bias due to linking error, this paper also relates to recent work taking advantage of the increased ease in making links across complete-count censuses to better describe patterns in mobility in U.S. history, broadly defined to include intergenerational mobility (e.g., Pérez Reference Pérez2019; Ward Reference Ward2023) and immigrant assimilation (e.g., Abramitzky et al. Reference Abramitzky, Platt Boustan, Jácome and Pérez2021b; Collins and Zimran Reference Collins and Zimran2019, Reference Collins and Zimran2023; Zimran Reference Zimran2022c) in addition to internal migration.
Ultimately, this paper addresses basic questions with relatively simple answers. But these questions and the new answers that I provide are fundamental and essential to a complete economic history of internal migration in the United States specifically and to a complete economic history of the United States more generally.
BACKGROUND
Most of what is known about U.S. internal migration before 1935 is based on the census’s question on individuals’ state of birth, which, through the census of 1930, was the only systematically available information on internal migration. The simplest application of these data uses information in census publications to determine the share of the population in each census year that lived outside of the state of birth—that is, the stock of inter-state migrants. Ferrie (Reference Ferrie1997a, 2006a, 2006b) and Hall and Ruggles (Reference Hall and Ruggles2004) report the results of such an analysis.Footnote 10 They find that the stock of white male inter-state migrants was effectively constant (as a share of population) from 1850 to 1940. Some improvement over these tabulations is possible with census microdata. These enable a focus on particular age cohorts, reducing the impact of concerns such as changing age composition.Footnote 11 Hall and Ruggles (Reference Hall and Ruggles2004) and Rosenbloom and Sundstrom (Reference Rosenbloom and Sundstrom2004) do this, finding a strong decline in the inter-state migration rates of native-born white men throughout the nineteenth century, with a slight increase from 1900 to 1920 for the young and a continuing decline in this period for the old.
The foremost shortcoming of analyses of this type is that they describe only the stock of inter-state migrants, not the flow.Footnote 12 Relatedly, if an individual moved several times between birth and observation, only one move would be observed. As a result, conclusions drawn from these data regarding migration are largely incomparable to those in, for instance, the literature on the Age of Mass Migration (Abramitzky and Boustan 2017; Hatton and Ward Reference Hatton, Ward, Diebolt and Haupert2019; Hatton and Williamson Reference Hatton and Williamson1998) or the literature on modern internal migration (e.g., Jia et al. 2023; Molloy, Smith, and Wozniak Reference Molloy, Smith and Wozniak2011; Sprung-Keyser, Hendren, and Porter Reference Sprung-Keyser, Hendren and Porter2022). Rosenbloom and Sundstrom (Reference Rosenbloom and Sundstrom2004) attempt to overcome this constraint. They use information on the birthplaces and ages of children in order to determine whether their parents had moved over the prior decade, enabling them to observe migration rates rather than stocks. The main limitation of this approach is that it can be applied only to families with young children and can be used only for relatively short spans due to the tendency of children to leave their parents’ household around age 18.Footnote 13 This analysis, like those focusing on individuals’ birth state-residence state comparisons, yields evidence of a sharp decline in the migration rates of native-born white men through the nineteenth century, followed by a slight increase through the twentieth. Estimates of internal migration based on birth state-residence state comparisons also do not permit the observation of intra-state moves. Such moves are likely to be particularly important in studying rural-to-urban migration (e.g., Department of Commerce 1933, p. 135; Ferrie Reference Ferrie2005) and thus to shedding light on the role of such flows in U.S. development.
The available census data also limit what can be learned even about the moves that can be observed because individuals are observed only after any migration has taken place. It is, therefore, not possible to study selection into migration on any but the most basic characteristics, or without assuming (implicitly or explicitly) that post-migration education and occupational information reflect an individual’s pre-migration characteristics. Hall and Ruggles (Reference Hall and Ruggles2004) and Rosenbloom and Sundstrom (Reference Rosenbloom and Sundstrom2004) interpret the available information to indicate positive selection into inter-state migration throughout U.S. history. The ability to study migrants’ destination choices is also limited. While individuals’ destinations can, of course, be observed, the absence of detailed data on the prior place of residence limits the extent to which the distance of a move can be determined, meaning that it is difficult to determine the drivers of destination choice while accounting for the cost of migration. Moreover, the inability to determine the timing of migration means that conditions in the destination and potential alternative destinations at the time that the move occurred are not known.
All of these limitations can be overcome by using linked census data, which provide an alternative way to measure individuals’ internal migration simply by comparing their places of residence in the initial and final census. Such data make it possible to bound the timing of a move, meaning that the flow rather than the stock can be observed, and the finer residence data enable the observation of intra-state inter-county moves. The pre-migration information available in the initial census enables the direct measurement of migrant selection and the determination of the distance of the move.
Several studies have exploited such data to study internal migration.Footnote 14 Steckel (Reference Steckel1988, Reference Steckel1989), for instance, uses data on census records linked between 1850 and 1860 to analyze westward migration. A number of other scholars (e.g., Collins and Wanamaker Reference Collins and Wanamaker2014, Reference Collins and Wanamaker2015; Ferrie Reference Ferrie1999; Long and Siu Reference Long and Siu2018; Stewart Reference Stewart2006, Reference Stewart2009) also study the rates, selection, and sorting of specific instances of internal migration with linked data. But until recently, it was not possible to construct linked datasets with sufficient coverage to reveal the broad, long-term patterns of U.S. internal migration. As a result, there is no systematic study of inter-county migration in the United States for the period before 1935, and even what is known about internal migration is extremely limited as a result of the constraints described earlier.
Fully digitized data on the characteristics of every individual in every census have only recently become available, enabling for the first time linkage of the white male population as a whole from 1850 to 1940 and equipping me to improve on the limitations of existing research.
Beyond Native-Born White Men
To maintain comparability to my empirical analysis, the preceding discussion has focused on the internal migration of native-born white men. The existing literature shows, however, that internal migration patterns for other groups may have been different in important ways.
The group most overlooked by internal migration studies based on birth state-residence state comparisons is immigrants, for whom birthplace information is not informative of prior place of residence in the United States. Zimran (Reference Zimran2022c) uses linked census data covering 1850–1930 to study immigrant men’s internal migration over the Age of Mass Migration and to compare it to that of natives. He finds that immigrant men were at least as likely as non-southern native-born white men to make an inter-county move, were more likely than natives to remain in and move to urban areas, and, like natives, became increasingly attached to urban areas over time.
Differences have also been documented between the internal migration of black and white men through the course of U.S. history, and in particular during the Great Migration (1910–1965). Unlinked census data reveal a substantially stronger pattern of rising migration rates for blacks after 1910 than for whites (Hall and Ruggles Reference Hall and Ruggles2004), and it has been argued that the end of mass European immigration was responsible for this surge (Collins Reference Collins1997). This movement was, as I will find for whites, slowed in the 1930s due to the Great Depression (e.g., Derenoncourt Reference Derenoncourt2022). Collins and Wanamaker (Reference Collins and Wanamaker2014, Reference Collins and Wanamaker2015) have studied the peak of the Great Migration in the pre-1940 period using data linked between the censuses of 1910 and 1930. They find that selection into migration for both black and southern white men was neutral, and that, in their destination choices, blacks were “more deterred by distance, attracted to manufacturing, and responsive to labor demand” (Collins and Wanamaker Reference Collins and Wanamaker2015, p. 947) than were whites.
Perhaps the main advantage of unlinked census data in the study of internal migration is the ability to include women in the analysis, though their experience is generally not emphasized. Hall and Ruggles (Reference Hall and Ruggles2004, figure 3, p. 836) show that inter-state migration stocks of white and black women evolved similarly to those of men of the same race, though they were always slightly lower, indicating somewhat less frequent inter-state migration for women.Footnote 15
Figure 1 INTER-COUNTY MIGRATION RATES BY LINKAGE METHOD AND SPAN
Note: Each figure shows the probability that an individual in the linked sample beginning in the year on the x-axis was observed living in a different county 10 or 20 years later, according to each linkage method. All observations are weighted by inverse linkage probability. No correction is made for false matches.
Source: See the text.
On the whole, the limited application of linked census data to the study of U.S. internal migration prior to 1940 leaves a number of fundamental questions unanswered. This paper takes a first step toward answering them.
Modern Internal Migration
In contrast to the limited picture that exists of internal migration before 1935, internal migration since then is extremely well documented due both to the prior-place-of-residence question that has been included in the census since 1940 and to the availability of more detailed data, such as the CPS, the ACS, and tax records, some of which directly link individuals’ residences over time. As in the historical literature, there is a substantial body of work studying specific instances of internal migration in the United States or using internal migration as a setting in which to study other questions.Footnote 16
There is also a literature describing the basic characteristics of U.S. internal migration, though the more comprehensive data in this regard imply that measurement is more straightforward than in historical settings. The broad findings in this regard in terms of rates, selection, and sorting are summarized by Jia et al. (2023) and Molloy, Smith, and Wozniak (Reference Molloy, Smith and Wozniak2011). The modern literature largely begins around 1980 and shows evidence of a decline in internal migration rates since then. Combined with results from Ferrie (Reference Ferrie and Susan2006a, 2006b), Hall and Ruggles (Reference Hall and Ruggles2004), and Rosenbloom and Sundstrom (Reference Rosenbloom and Sundstrom2004), which describe the intervening decades, the post-WWII picture of internal migration is of an increase until about 1980, followed by a decrease since then. There is also evidence of positive selection into internal migration (Wozniak Reference Wozniak2010). By creating the first comprehensive series of migration rates (rather than stocks) with known timing for the period before 1940, this paper enables, for the first time, the dovetailing of modern and historical internal migration rate series and, therefore, the construction, for the first time, of a series of internal migration rates spanning nearly 175 years of U.S. history.
DATA
My analysis is based on 13 datasets making all possible 10- and 20-year links between the U.S. censuses of 1850–1940.Footnote 17 I begin the analysis in 1850 because that year’s census was the first to enumerate the entire free population. The analysis ends in 1940 because, at the time of writing, this was the most recent complete-count census that had been fully digitized. The 1870–1890, 1880–1890, 1890–1900, and 1890–1910 spans are omitted because the vast majority of the 1890 census records were destroyed by fire and are unavailable for linkage or analysis.
I created these datasets by merging complete-count census records provided by Ruggles et al. (Reference Ruggles, Sarah Flood, Ronald Goeken, Schouweiler and Sobek2021) with the “basic” linkage crosswalks provided by Zimran (Reference Zimran2022a).Footnote 18 Zimran’s (2022a) method, which implements suggestions made in the literature (Abramitzky et al. Reference Abramitzky, Leah Platt Boustan, Feigenbaum and Pérez2021a; Bailey et al. Reference Bailey, Cole, Henderson and Massey2020) to increase the quality of matches by requiring the uniqueness of individuals in a particular age band and using an orthographic distance measure to compare the names of potential matches, provides my preferred links, referred to as the Main links in the analysis later. The datasets created by merging the census records with the linkage crosswalks provide information on an individual’s county of residence in each of the two censuses, which enables me to determine whether an individual made an inter-county move between them.Footnote 19
Throughout my analysis, I restrict attention to native-born white men aged 18–40 in the initial census. Beyond the restriction to native-born white men discussed previously, the restriction to those aged 18–40 in the initial census is intended to ensure that men are observed in the labor force in the initial census of the linkage span (to enable construction of the occupational rank measure) while ensuring that they are also not so old that mortality is an important concern. It also limits attention to ages in which there are likely to be children in the household at the time of the final census, enabling the alternate household structure-based method for estimating migration and thus for estimating false match rates.Footnote 20
There are two main concerns that arise in the use of linked census data to study internal migration. The first is the danger of false matches—that is, the concern that the linked datasets may not actually describe the same person in the two census years. This concern has been highlighted recently by Abramitzky et al. (Reference Abramitzky, Leah Platt Boustan, Feigenbaum and Pérez2021a) and Bailey et al. (Reference Bailey, Cole, Henderson and Massey2020) and touches on all aspects of the analysis of internal migration.Footnote 21 Since nearly all false matches are to individuals living in a different county, the observed rate of migration conflates true migration with false matches. Selection into migration is also, therefore, conflated with selection into false matching. Finally, under the assumption that a false match links an individual in one census to a random individual in a subsequent one, true destination choice patterns are conflated with a tendency for false matches to show spurious migration toward more populous areas.
I address the danger posed by false matches in two ways. First, I repeat all of my analysis using data sets constructed by four alternate linkage methods, drawing only conclusions that are robust to the choice of linkage method. Two of the alternative linkage methods (referred to as ABEE and ABEN) are simply different from the main method in their linkage parameters, requiring uniqueness of links in a somewhat smaller age band than my preferred linkage method and requiring either an exact match of names or a match of the NYSIIS standardization of names rather than making an orthographic distance comparison. The other two linkage methods (referred to as Int and Int+) are stricter, reducing the danger of false matches at the potential risk of a less representative sample by limiting matches to the intersection of the set of matches by the main method and the ABEE and ABEN methods, and additionally by requiring agreement across censuses of information not used in linkage. Online Appendix B.1 presents further details of the samples constructed by these alternative linkage methods. In studying the rates of internal migration, I also propose a method, described in detail in the next section, to estimate the rate of false matches for each linkage method and to correct my estimated internal migration rates.
The second main concern in the use of linked data is that they may not be representative of all individuals at risk for linkage. Indeed, this is one reason why, even though it is tempting to simply use the strictest linkage method to minimize the danger of false matches, I do not do so—this would increase the danger of constructing an unrepresentative sample. To address this concern, I reweight each linked dataset so that its observable characteristics match (as closely as possible) the distribution of observable characteristics of those at risk for linkage in the initial census. Online Appendix B.1 provides details.
The initial census of each span also provides data on a variety of individuals’ pre-migration characteristics, including occupation, literacy, and initial urban status, which I use to study migrant selection. Literacy, though flawed and potentially changing in definition over time, is the only consistently available measure of human capital. Occupation is the only measure of economic status that is available in a consistent way over the complete 1850–1940 period. As described in Online Appendix B.2, I use Ruggles et al.’s (Reference Ruggles, Sarah Flood, Ronald Goeken, Schouweiler and Sobek2021) occupational codes to construct a measure of an individual’s occupational rank relative to the white male population aged 18–64 in each census. To determine whether an individual resided in an urban area in the initial and final census of each linkage span, I use the official Census Bureau definition of an urban place as one with at least 2,500 inhabitants, as well as an alternative definition using a population cutoff of 25,000.
Finally, I construct a number of measures of the characteristics of an individual’s move and of his counties of initial and final residence. These include the distance of the move and whether it crossed state or regional (i.e., census divisions) boundaries; the share of each county’s population residing in urban areas under various definitions; and a Bartik (Reference Bartik1991)-type measure of labor demand growth. I address changing boundaries using Hornbeck’s (2010) method.
INTER-COUNTY MIGRATION RATES, 1850–1930
I begin by answering the most basic, but perhaps the most fundamental, question about U.S. internal migration—what was the rate of inter-county migration? That is, how likely was it that an individual observed in a given census year would move to a different county over the next 10 or 20 years?
Uncorrected Estimates
Figure 1 presents the estimated uncorrected rates of inter-county migration over 10- and 20-year spans. The results vary depending on the linkage method—the stricter methods show uniformly lower rates and the patterns over time are also different, with the more permissive methods finding a slight decline over time before a sharper decline in the 1930s, and the stricter methods showing general stability before a final decline. However, the danger of bias from false matches challenges the validity of these results.Footnote 22
Figure 2 ESTIMATED RATES OF FALSE LINKAGE
Note: This figure shows the estimated rate of false matching for each linkage method and 10-year span according to the comparison of inter-state migration estimates by the linkage method and by the method based on the birth places and ages of children. The rate is computed according to the law of total probability, as explained in the text.
Source: See the text.
Quantifying and Correcting for False Matches
To address concerns about spurious migration due to false matches, I propose a method to estimate the rate of false matches and to correct for the bias that they induce.Footnote 23 The principle of this method is that, where a quantity can be estimated for the same sample by two methods—one based on linkage and the other not—the difference between the two estimates is informative of the rate of false matches.
In this case, for individuals in the linked sample whose household structure enables it, I use the birthplace and age composition of children in an individual’s household to create a non-linkage-based measure of inter-state migration over the previous 10 years.Footnote 24 Movers are defined as those with a child aged less than 10 years old born in a different state than the current state of residence and no older child born in the state of residence. Stayers are defined as those with a child born in the state of residence at least 10 years old and no children younger than 10 born in a different state. This categorization is performed for the latter census of each 10-year span (i.e., in 1860 for the 1850–1860 span) for the linked sample; it is not applied to 20-year spans because children are unlikely to remain in their parents’ household for over 20 years. This procedure results in a subsample of the linked dataset—which I refer to as the corroboration sample—composed of individuals whose inter-state migration status over a 10-year span can be measured in two ways—by comparing their state of residence in the initial and final censuses and according to their household composition.Footnote 25
According to the law of total probability, the probability that an individual in the corroboration sample is observed to have made an inter-state move (whether he truly moved or not) according to his residence state in the initial and final census (i.e., according to linkage), which I denote as P(moved state), can be written as
$$\matrix{ {P({\rm{moved}}\,{\rm{state}}) = P({\rm{moved}}\,{\rm{state|true}}\,{\rm{match}})[1 - P({\rm{false}}\,{\rm{match}})]} \cr {\quad \quad \quad \quad \quad \quad + P({\rm{moved}}\,{\rm{state}}|{\rm{false}}\,{\rm{match}})P({\rm{false}}\,{\rm{match}}).} \cr } $$
Rearranging Equation (1), I can express the probability of a false match as
$$P({\rm{false}}{\mkern 1mu} \, {\rm{match}}) = {{P({\rm{moved}}{\mkern 1mu} \, {\rm{state}}){\rm{ - }}P({\rm{moved}}{\mkern 1mu} \, {\rm{state|true}}{\mkern 1mu} \, {\rm{match}})} \over {P({\rm{moved}}{\mkern 1mu} \, {\rm{state|false}}{\mkern 1mu} \, {\rm{match}}){\rm{ - }}P({\rm{moved}}{\mkern 1mu} \, {\rm{state|true}}{\mkern 1mu} \, {\rm{match}})}}.$$
In calculating Equation (2), I use the following quantities from the corroboration sample. As P(moved state), I use the linkage-based estimated rate of inter-state migration according to the comparison of the initial- and final-year residence state. As P(moved state|true match)—the true probability of an inter-state move—I use the estimated rate of inter-state migration according to the household composition method. Finally, as P(moved state|false match)—the probability of observing inter-state migration in the case of a false match—I use an individual’s birthplace and age to determine the average probability that a person to whom he could be linked would live in a different state, which is straightforward to estimate.Footnote 26
Figure 2 presents the estimated rates of false linkage for each method and initial census year. The estimates presented in Figure 2 fit expectations. The estimated false match rates are higher for the more permissive linkage methods and decline over time from about 15 percent to under 10 percent by the end of the study period.Footnote 27 For the more restrictive methods, the estimated false match rates are initially about 5 percent and fall to approximately zero in the twentieth century.Footnote 28
Figure 3 INTER-COUNTY MIGRATION RATES BY LINKAGE METHOD AND SPAN, CORRECTED FOR FALSE MATCHES
Note: Each figure shows the probability that an individual in the linked sample beginning in the year on the x-axis was observed living in a different non-overlapping county 10 or 20 years later, according to each linkage method. All observations are weighted by inverse linkage probability. The estimates are then adjusted for false matches according to the method presented in the text.
Source: See the text.
Having estimated P(false match) based on inter-state migration in the corroboration sample, I use this estimate to correct my estimates of inter-county migration for the full linked sample. I rearrange Equation (1) and replace inter-state moves with inter-county moves to yield
$$\matrix{ {P({\rm{moved}}\,{\rm{country|true}}\,{\rm{match}})} \cr { = {{P({\rm{moved}}\,{\rm{country}}) - P({\rm{moved}}\,{\rm{country}}|{\rm{false}}\,{\rm{match}})P({\rm{false}}\,{\rm{match}})} \over {1 - P({\rm{false}}\,{\rm{match}})}}.} \cr } $$
To compute this value, I use the following quantities. As P(false match), I use the estimate computed from Equation (2); for 10-year spans, I use the estimate from the analogous span; for 20-year spans, I use the estimate for the 10-year span beginning in the same year; there is no 1880–1890 span, and so I must omit the 1880–1900 span. As P(moved county), I use the probability of observing an inter-county move in the full linked dataset. For inter-county migration, P(moved county|false match) is sufficiently close to one that, with minimal loss, I can write Equation (3) as
$$P({\rm{moved}}\,{\rm{country}}|{\rm{true}}\,{\rm{match}}) = {{P({\rm{moved}}\,{\rm{country}}) - P({\rm{false}}\,{\rm{match}})} \over {1 - P({\rm{false}}\,{\rm{match}})}}.$$
The estimates coming from computing Equation (4) are my benchmark estimates of inter-county migration rates that are corrected for false matches. Despite my efforts to correct these estimated migration rates, it is inevitable that there will remain some error. There are many possible causes. One in particular is that the corroboration sample is not representative of the broader linked sample. This is a somewhat less severe issue than when using the household composition method to directly measure migration (Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004). Although the internal migration of men with children is likely different from that of all men, an issue arises for this method only insofar as the probability of a false match differs between these groups. I investigate the degree to which selection into the corroboration sample is potentially problematic in Online Appendix B.3, where I show that such selection is unlikely to have led to biased estimates of false match rates.
This method also relies on the assumption that the true migration rates of the correctly linked and incorrectly linked are the same. Since it is well established that selection into linkage is non-random, it is not likely that this assumption holds in reality—characteristics determining success in linkage likely also determine the propensity to move. The method also assumes that there is no error (on average) in determining individuals’ migration status using the household composition method. This method is also not informative as to which observations are incorrectly linked since it is used simply to deflate an aggregate quantity.Footnote 29 This implies that it cannot be used in more detailed analyses of selection and destination choice.
Corrected Estimates
Figure 3 presents the estimated corrected inter-county migration rates for native-born white men for each linkage method and span. For both 10- and 20-year spans, the resulting predictions of inter-county migration rates and their changes over time are similar across the linkage methods.Footnote 30 For 10-year spans, the estimated migration rates by my preferred linkage method are consistently between 31.5 and 35.7 percent from the 1850–1860 span to the 1920–1930 span, with no clear trend. I discuss the 1930s in more detail later. The estimated 20-year migration rates are higher, at 39.6 to 44.1 percent, and evolve largely without trend, save for a slight decline between the two spans of the nineteenth century. The main takeaway is that the frequency of inter-county migration over the period 1850–1930 was largely constant.Footnote 31
MIGRANT SELECTION, 1850–1930
Characterizing migrant selection is crucial to a basic description of any flow of migration. It provides insight into the potential effects of internal migration on individuals, on the economies of sending and receiving areas, and on the broader economy. By providing information on the pre-migration characteristics of prospective migrants, linked data enable me to delve further into this question than has previously been possible. However, it remains possible that errors in linkage may influence conclusions if the probability of a false matche varies along a dimension on which I study selection. I present here the results using only my preferred linkage method, limiting my conclusions to those that are robust to the use of alternate and stricter linkage methods (results in Online Appendix C).
I focus first on two measures of migrant selection—occupational rank and literacy in the initial year of the span—that speak to whether internal migrants were positively or negatively selected. These correspond roughly to the typical focus in the economics of migration on earnings (e.g., Borjas Reference Borjas1987; Chiquiar and Hanson Reference Chiquiar and Hanson2005) and education (e.g., Card Reference Card2005). To measure selection, I estimate, separately for each linked dataset, the equation
$${y_{it}} = {\alpha _t} + {\beta _t}{r_{it}} + {\gamma _t}{\ell _{it}} + {\delta _t}{u_{it}} + {{\bf{x'}}_{it}}\varphi + {\varepsilon _{it}},$$
where y it is an indicator equal to one if individual i migrated across county lines in span t, r it is individual i’s initial-year occupational rank, ℓ it is individual i’s initial-year literacy, u it is individual i’s initial-year urban residence, and x it is a vector of initial-year controls. The β t , γ t , and δ t coefficients provide measures of migrant selection in span t, indicating whether greater occupational rank, literacy, or urban status were associated with a greater probability of migration.Footnote 32 To ensure that changes in the β t , γ t , and δ t coefficients across censuses are the product of actual changes in selection patterns rather than of changing availability of controls in the census, and to avoid confounding the interpretation of the coefficients, I limit the vector x it to the variables listed in Online Appendix Table A.1 that are available in all initial censuses.Footnote 33
Figure 4 presents results on migrant selection on the basis of occupational rank and literacy. Each panel contains two sets of estimates—one with no controls and one with controls, including initial-county fixed effects and the other dimensions of selection.Footnote 34 Panels (a) and (b) focus on selection on the basis of occupational rank. Unconditional selection into inter-county migration was consistently negative, moving from very close to zero in the beginning of the study period to more strongly negative in the late nineteenth century and in the twentieth.Footnote 35 Such unconditional comparisons, however, ignore differences in age, sector, distance to potential destinations, and other factors likely to influence migration. When conditioning on all observables, including the initial county of residence, selection patterns were more stable over time. For spans beginning in 1850, conditional selection on the basis of occupational rank was zero or positive. A notable decline in spans beginning in 1860 then occurred, indicating more negative selection into migration in this span than earlier. From this point through the 1920s, the coefficients hover around –0.05. With the occupational rank measure ranging from zero to one with a standard deviation of roughly 0.25, these coefficients are small but not negligible: a one-standard deviation increase in occupational rank was associated with about a 1.25-percentage point decline in migration probability on a base ranging from 38 to 52 percent.Footnote 36 The variation in these coefficients over time is even smaller.Footnote 37 In sum, beginning in the 1860s, individuals of greater occupational rank were somewhat less likely or as likely to make an inter-county move than otherwise similar individuals from the same initial county but of lower rank,Footnote 38 and this difference remained largely constant over time.
Figure 4 MIGRANT SELECTION BY MEASURE AND SPAN, INDIVIDUAL CHARACTERISTICS
Note: Each figure shows the coefficient on the variable in question in a selection regression for migration over the stated span. All observations are weighted by inverse linkage probability. Vertical lines indicate 95-percent confidence intervals. Controls are all controls available in all census years, as well as county fixed effects.
Source: See the text.
Panels (c) and (d) of Figure 4 focus on selection into internal migration on the basis of literacy. All estimates indicate that literate individuals were less or as likely as otherwise similar individuals to make an inter-county move (including those for alternate linkage methods). The patterns are similar across 10- and 20-year spans: a decline in the magnitude of selection through the nineteenth century and stabilization in the twentieth. The coefficients are again small but non-negligible, with literate individuals about 2 to 4 percentage points less likely to migrate than illiterates.
On the whole, these results paint a picture of migration that was neutrally or slightly negatively selected, with this selection, by most indications—including on the basis of the best available measure of socioeconomic status—largely constant from the 1860s through the 1920s.Footnote 39
CHANGING ORIGINS AND ORIENTATION, 1850–1930
In this section, I study selection into internal migration on the basis of initial urban residence as well as the destination choices of internal migrants. I find that the results of general constancy in the rates and selection of internal migration conceal important changes in the nature of U.S. internal migration. Again, I present results here for my main linkage method, with results from alternate methods presented in Online Appendix C.
Figure 5 focuses on selection into inter-county migration on the basis of urban residence in the initial year of each span using two different measures of urbanization—the 2,500-inhabitant cutoff (Panels (a) and (b)) and an indicator for being in a city of at least 25,000 (Panels (c) and (d)).Footnote 40 For each measure and span, I include two sets of estimates—an unconditional estimate and one that conditions on all controls, but I use state fixed effects instead of county fixed effects.Footnote 41 In spans beginning in 1850, urbanites were 5 to 10 percentage points more likely to migrate than otherwise similar ruralists, depending on the linkage method. This conditional urban migration premium declined over the nineteenth century. By spans beginning in 1880 for 20-year spans or in 1900 for 10-year spans, this pattern reversed, with urbanites approximately as likely (for the 2,500-person definition) or up to 5 percentage points less likely (for the 25,000-person definition) than otherwise similar ruralists from the same state to migrate. Ruralists’ conditional migration premium then declined over the twentieth century. Thus, in contrast to selection on the basis of occupational rank and literacy, selection on the basis of urban residence was relatively larger in magnitude, changed throughout the 1850–1930 period, and for larger cities changed in sign such that urbanites were initially more and then less likely than ruralists to move.Footnote 42
Figure 5 MIGRANT SELECTION BY MEASURE AND SPAN, URBAN RESIDENCE
Note: Each figure shows the coefficient on the variable in question in a selection regression for migration over the stated span. All observations are weighted by inverse linkage probability. Vertical lines indicate 95-percent confidence intervals. Controls are all controls available in all census years, as well as state fixed effects.
Source: See the text.
Changes over time are also evident in internal migrants’ destination choice patterns. Figure 6 focuses on the distance distribution of moves. Panels (a) and (b) divide moves into inter-county but intra-state, inter-state but intra-region, and inter-region. This division is particularly important as the fraction of moves that are intra-state measures the extent to which internal mobility is not observed when using state of birth to determine migration. Inter-county intra-state moves were an important component of U.S. internal migration, accounting for about 40 percent of moves at the beginning of the study period and rising over time to about 60 percent by the end of the study period. This increase came at the expense of inter-region moves, with the share of inter-state but intra-region moves remaining largely constant in frequency over time. Panels (c) and (d) present violin plots for distance of moves. The distribution of move distance has a consistent peak below 50 miles. In the earlier periods, a second peak around 500 miles is also evident and fades over time.
Figure 6 DISTANCE OF MOVES
Note: Panels (a) and (b) show the cumulative distribution of intra-state, inter-state but intra-region, and inter-region moves. Panels (c) and (d) present the distribution of move distance for each linkage span with box plots presenting the mean and quartiles.
Source: See the text.
To what extent were these unconditional changes the product of changing individual characteristics over time? Using a pooled dataset of all migrants from all linkage spans of a particular length, I estimate a regression of the form
$${y_{it}} = {\beta _t} + {{\bf{x'}}_{it}}\delta + {\varepsilon _{it}},$$
where y it is some measure of individual i’s move distance in span t, β t is a series of span fixed effects, and x it is a set of initial-year controls, including all observables available in all census years and either initial-year state or initial-year county fixed effects. Figure 7 plots the β t resulting from estimating Equation (6) with no controls, with controls and initial-state fixed effects, and with controls and initial-county fixed effects. Omitting the indicator for spans beginning in 1850, these estimates show how migration distance changed over time. Paralleling Figure 6, all measures show a large, unconditional decline in the distance of the move or in the probability of moving across states or regions over time. The decline, however, was moderated by changes in the demographics of migrants, meaning, for instance, that the move made by an individual in 1850–1860 was about 100 miles further than that made by an observationally similar person in 1920–1930.
Figure 7 CHANGES IN DISTANCE OF MOVES
Note: These figures show the coefficients on initial-year fixed effects in regressions of destination characteristics with and without controls, with 1850 as the excluded year. “Controls” includes all controls available in all censuses, including state fixed effects. “Controls and FE” also includes county fixed effects. Vertical lines are 95-percent confidence intervals. Observations are weighted to correct for selection into linkage.
Source: See the text.
To determine which characteristics of destinations attracted migrants, I restrict attention to movers and estimate a conditional logit model of the form
$${P_{iotj}} = {{\exp ({\beta _t}{d_{oj}} + {{\bf{z'}}_{jt}}{\delta _t})} \over {\sum\nolimits_k {\exp ({\beta _t}{d_{ok}} + {{\bf{z'}}_{kt}}{\delta _t})} }},$$
where P iotj is the probability that migrant i initially from county o in linkage span t chose destination j, z jt is a vector of characteristics of county j in the initial year of span t, d oj is the distance between counties o and j, and k indexes all potential destinations. The vector z includes a county’s initial-year urbanization (i.e., the fraction of the population living in urban areas) and census-division fixed effects (with New England as the excluded category).Footnote 43 The coefficients β t quantify the deterrent effect of distance in span t, and the coefficients δ t quantify the attractiveness of the various county characteristics in that period. The coefficients have the usual interpretation of coefficients in a logit model—the marginal effect of the variable in question on the log odds of selecting a particular destination, conditional on its distance; where P iotj is near zero (as is the case for most county pairs), the coefficient can also be interpreted as the percent change in the probability of choosing a particular destination for a unit increase in the regressor. Online Appendix B.4 presents additional details on the estimation of this model.
Figure 8 presents the results of this estimation. Panels (a)–(d) focus on the deterrent effect of distance and on the attractiveness of urban areas.Footnote 44 An increase over time in the deterrent effect of distance is evident from the increasing magnitude of its negative coefficient over time: in the nineteenth century, the elasticity of migration probability with respect to distance was about –1.5; by the twentieth century, it was about –1.7. This result is consistent with the declining distance of moves described earlier. A large increase over time in the attractiveness of urban destinations is also clear through the 1920s. In the nineteenth century, a completely urban county had a 1.5 greater log odds of being selected than an entirely rural destination.Footnote 45 By the 1920s, this figure had risen to about 3 or 3.5, depending on the definition of an urban place.Footnote 46
Figure 8 CONDITIONAL LOGIT RESULTS
Note: Panels (a)–(d) present coefficients on distance and the measure of urban residence from the conditional logit estimation. Panels (e) and (f) present the coefficients from the regressions of Panels (a) and (b), respectively, on the census division indicators, with New England excluded. Note that there is no data for spans that would include 1890; the lines span this period simply to enable a clearer reading of the trend over time.
Source: See the text.
Panels (e) and (f) of Figure 8 compare the attractiveness of each census division as a destination relative to New England. For the most part, there is little change over time. The main exceptions are the Mountain and Pacific divisions, however, which increased substantially in their relative attractiveness, indicating that there was some characteristic of these areas other than their urbanization that increasingly attracted internal migrants.Footnote 47
Finally, I compare the growth over the linkage span of the urbanization or labor demand growth of movers’ and stayers’ residence counties by estimating an equation of the form
$${y_{i{t_2}}} - {y_{i{t_1}}} = {\alpha _t} + {\beta _t}{m_{it}} + {\bf{x}}_{it}{\gamma _t} + {\varepsilon _{it}}$$
where
$${y_{i{t_2}}}$$
is the value for individual i’s residence county in the final year of span t,
$${y_{i{t_1}}}$$
is the residence county in the initial year, m
it
is an indicator equal to one if individual i moved during span t, and x
it
is a vector of individual i’s observables in the initial year of span t, including indicators for initial county of residence. For stayers, the y
it
values describe the same county in both periods, meaning that stayers capture the change in the urbanization of their home county. A positive β
t
implies that movers experienced a greater gain in urbanization than stayers from the same county—that is, that they moved to areas growing more quickly.
Figure 9 presents the results.Footnote 48 In essence, these results combine the selection and sorting patterns described previously, in which urbanites became relatively less likely to migrate as compared to ruralists and urban areas increased in attractiveness over time. In the nineteenth century, I find that movers and stayers experienced approximately the same increase in urban status and labor demand growth. By the twentieth century, the growth experienced by movers was greater than that experienced by stayers. Thus, whereas nineteenth-century migration was relatively neutral with respect to urbanization, twentieth-century migration was a force driving the contemporaneous urbanization of the economy by shifting population into more urban (or at least more rapidly urbanizing) areas.
Figure 9 MOVERS’ CHANGE IN URBANIZATION AND LABOR DEMAND GROWTH
Note: This figure compares the growth in urbanization or labor demand growth experienced by movers to that experienced by stayers, with either no controls, controlling for all observables, or controlling for all observables and county fixed effects.
Source: See the text.
THE 1930S
All of these patterns changed in the 1930s. For the first time, this decade was marked by a substantial decline in the rate of internal migration (Figure 3 Panel (a)), which fell from 34.2 percent for 1920–1930 to 27.6 percent for 1930–1940 according to my preferred linkage method. The 1930s also marked a moderation in selection into migration on the basis of occupational rank (Figure 4 Panel (a)). Relative to the 1920s, there was a considerable increase in the coefficient that was larger than its change in any other period in the twentieth century. This change marked either a moderation of negative selection or a transition from somewhat negative to somewhat positive selection, depending on the linkage method; that is, internal migrants were less negatively selected in the 1930s than in the 1920s. A similar, though somewhat weaker, moderation is present in terms of selection into internal migration on the basis of urban residence in that urbanites and ruralists were more similar in terms of migration propensity than ever before in the twentieth century, though this was the product of a trend that evolved throughout the twentieth century (Figure 5 Panels (a) and (c)). Similarly, the distance of moves was lowest in the 1930s, again continuing a declining trend (Figure 7 Panels (a), (c), and (e)).
A particularly striking change concerns the attractiveness of urban areas as destinations for internal migrants. From the 1850s to the 1920s, the attractiveness of urban areas as destinations for internal migrants increased. But the 1930s marked a reversal of this trend, with a decline in this measure to levels approximately equal to those of the 1910s (Figure 8 Panels (a) and (c)). Even more dramatically, the 1930s marked a sharp reversal in the degree to which the urbanization growth or labor demand growth of movers exceeded that of stayers (Figure 9 Panels (a), (c), and (e)). This figure had climbed from the 1850s to the 1920s, but reversed sharply in the 1930s. For instance, in the 1920s, movers’ growth in the likelihood of living in a city of 25,000 or more inhabitants was about 8 percentage points greater than that of stayers. By the 1930s, this figure had fallen to zero—a level not seen since the late nineteenth century—implying that the urbanization growth of movers and stayers was nearly identical and that internal migration was no longer a force driving urbanization.
SUMMARY OF ROBUSTNESS CHECKS
In addition to verifying, in Online Appendix C, that the conclusions that I draw are robust to the choice of linkage method, I also verify the robustness of my results to a number of other permutations of the sample or definitions of variables. In Online Appendix D, I redefine migration such that an individual must both cross county lines and move at least 150 miles to be considered an internal migrant. Naturally, the estimated migration rates are lower, in the vicinity of 15 percent over 10-year spans and 20 percent over 20-year spans. There is also a slight decline (about 2 to 3 percentage points) in these rates in the nineteenth century, which on the whole results in a slight downward trend, even when omitting the 1930s. Selection into migration was also different, with urbanites more likely than ruralists to make longer-distance moves. Other results are largely unaffected.
In Online Appendix E, I omit individuals with foreign-born fathers from the sample. The intention of this exercise is to ensure that the continued assimilation of second-generation immigrants does not affect the results. In Online Appendix F, I repeat the main results using imputed occupational codes from Zimran (Reference Zimran2022b) in cases where the occupations given by Ruggles et al. (Reference Ruggles, Sarah Flood, Ronald Goeken, Schouweiler and Sobek2021) are listed as “Not Yet Classified.”
DISCUSSION AND CONCLUSION
Internal migration is one of the fundamental forces that contributed to the development of the American economy and identity. But a lack of suitable data has made it difficult to establish even the most basic facts of internal migration in U.S. history. In this paper, I exploit recent advances in the availability of complete-count census data and in the technology to make links between censuses in order to describe the trends in the rates, selection, and sorting of the inter-county migration of native-born white men over the period 1850–1940. I find that the rates of and selection into inter-county migration were generally constant over time, with largely neutral or slightly negative selection into migration. But the distance of moves declined over time, migrants became increasingly attracted to the west, and the origins and orientation of migration shifted considerably over this period to become increasingly oriented toward driving a flow of population toward urban areas. The 1930s then marked a change in all regards, with migration declining in frequency, becoming somewhat more neutrally selected, and sharply declining in the degree to which it drove increases in urbanization.
These findings deepen our understanding of U.S. internal migration. Regarding the rates of internal migration, existing research (Ferrie Reference Ferrie1997a; Hall and Ruggles Reference Hall and Ruggles2004; Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004) has identified a decline in inter-state migration over the period 1850–1940, though the studies differ on the precise timing of this decline. My findings for inter-county migration paint a different picture—one of stable migration rates until the 1930s. In part, this difference in results can be attributed to the greater ability of linked data to bound the timing of migration. But it can also be attributed to my finding that inter-county but intra-state migration became increasingly important as average migration distances declined, meaning that a focus solely on inter-state migration would overlook an increasingly large share of internal population movements over time.
My results also give an understanding of the selection and sorting of internal migration that goes into far greater depth than has previously been possible. There are few existing estimates of migrant selection over a broad span to which my results can be compared. But the estimates that do exist (Hall and Ruggles Reference Hall and Ruggles2004; Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004) point to positive selection—something for which I find no evidence in my analysis. The existing descriptions of migrants’ destination choices in the long run (Hall and Ruggles Reference Hall and Ruggles2004; Rosenbloom and Sundstrom Reference Rosenbloom and Sundstrom2004) are even more limited, generally focusing on the region of destination or the urban or rural status of the destination. My analysis provides a richer description than has previously been possible.
The results of this paper also help to better understand the development of the U.S. economy over the period that I study and, in particular, shed new light on Turner’s (1921) influential interpretation of internal migration in U.S. history and later critiques of this interpretation. Turner (Reference Turner1921) famously argued that the United States’s high internal migration rates in global perspective in the nineteenth century were the product of the availability of land on the frontier, which provided opportunities for surplus labor from eastern cities.Footnote 49 Turner (Reference Turner1921) also predicted that these high migration rates would be temporary, declining when the frontier ceased to exist. Others (e.g., Shannon Reference Shannon1945; Weber Reference Weber1899) challenged claims of uniquely high migration rates in the nineteenth century, arguing that the growth of urban areas would draw migration, counteracting this decline.
Consistent with Turner (Reference Turner1921), I find evidence that urbanites were more likely to migrate in the nineteenth century. But I find no evidence of strongly negative selection into internal migration,Footnote 50 nor do my conditional logit results indicate that rural areas were particularly attractive in the nineteenth century (i.e., the coefficient on urban is always positive). I also find no evidence of a decline in migration rates after the closing of the frontier, contrary to findings for inter-state migration. Instead, I find that the closing of the frontier was followed by a twentieth century marked by a shift of internal migration into a force driving the urbanization of the economy from the perspective of both selection and destination choice. Beyond validating critiques of Turner’s (1921) claims that the nineteenth century was exceptional in its high rates of internal migration, this finding speaks to the structural transformation of the economy as it shifted from agricultural to industrial (Caselli and Coleman 2001; Kuznets Reference Kuznets1966), with internal migration helping to allocate labor to the nation’s growing industrial sector in urban areas, as evidenced by my results regarding the labor demand growth experienced by movers.
My results also shed light on the unique nature of internal migration in the 1930s. The unique economic and climatic circumstances of the 1930s, combined with the U.S. census’s first direct query on migration in the 1940 census, have attracted substantial attention to internal migration in this period.Footnote 51 Despite the common view of prevalent migration in this period in response to the unique shocks of the 1930s, my results show that migration rates were, in fact, lower during this period than in preceding decades (see also Long and Siu Reference Long and Siu2018). This result is consistent with Saks and Wozniak’s (2011) finding that internal migration is generally pro-cyclical. My findings regarding the relationship between urbanization and migration also complement existing results. Fishback, Horrace, and Kantor (Reference Fishback, Horrace and Kantor2006), for instance, use data on counties’ aggregate population growth to show that more urban areas experienced relatively greater outmigration in the 1930s—a phenomenon due in part to “moves of despair” from depressed urban areas to farms (Boone and Wilse-Samson 2023; Boyd Reference Boyd2002).Footnote 52 My analysis complements this literature by putting these patterns into the broader context of U.S. history, comparing them to prior decades characterized by prior economic shocks.
Finally, the results of this paper also introduce or deepen a number of puzzles. The first concerns the relatively constant frequency of internal migration over the first 80 of the 90 years that I study. The structural transformation of the U.S. economy, combined with changes in transportation technology, land availability, international immigration, and labor market integration from 1850 to 1930, would be expected to lead to some change in the frequency or selectivity of internal migration. Indeed, there is evidence that these forces were associated with substantial changes in immigrant assimilation (Collins and Zimran Reference Collins and Zimran2023) and intergenerational mobility (Long and Ferrie Reference Long and Ferrie2013; c.f., Ward Reference Ward2023). Yet such a change in the frequency of internal migration did not occur until the basket of shocks of the 1930s. A potential explanation is that this expected change came in the form of the increase in black migration after the onset of World War I, combined with changes in the frequency of international immigration (Collins Reference Collins1997). But, in general, it is possible that linked data-based studies of the internal migration of groups other than native-born white men can illuminate this issue. Another puzzle concerns the increase over time in the deterrent effect of distance in the destination choices of internal migrants. This increase came despite improvements over the study period in transportation technology. This specific result can be rationalized by observing that the share of migrants who moved shorter distances increased,Footnote 53 but a larger question concerns why the increase in short-distance migration occurred. Explaining either of these puzzles is beyond the scope of this paper, but my documentation of them lays them out as the targets of future research.
