Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

A Different Land of Opportunity:

The Geography of Intergenerational


Mobility in the Early Twentieth-
Century United States
Hui Ren Tan, National University of Singapore

Has the geography of intergenerational mobility in the United States


changed over time? Constructing a large historical linked sample, I
show that upward mobility in the early twentieth century was greater
for those who grew up in the coastal and industrial regions, in contrast
to more recent times, where mobility is higher among persons who
were raised in the middle of the country. The historical patterns are
not driven by imperfections in record linkage or measurement error
in economic status.

I. Introduction
Does a child’s chances of success later in life depend on where he or she
was raised? Focusing on the United States today, Chetty et al. (2014, 2020)
and Chetty and Hendren (2018) document substantial variation in inter-
generational income mobility across different parts of the country. Those

I am indebted to Robert Margo, Kevin Lang, Daniele Paserman, James Feigen-


baum, and Pascual Retrespo for their patience and guidance. For detailed comments
and helpful discussions, I thank Leah Boustan, Ellora Derenoncourt, Claudia
Goldin, Yi Jie Gwee, Nathaniel Hendren, Jessica Pan, Jan Stuhler, Lin Tian, and
Marianne Wanamaker. I am also grateful to Ran Abramitzky, Philipp Ager, Jeremy
Atack, David Autor, Samuel Bazzi, Brian Beach, Hoyt Bleakley, William Collins,
Björn Eriksson, Daniel Fetter, Martin Fiszbein, Paola Giuliano, Adam Guren,
Trevon Logan, Nathan Nunn, Claudia Olivetti, John Parman, Martin Saavedra,

Submitted September 30, 2020; Accepted December 16, 2021; Electronically published November 3, 2022.
Journal of Labor Economics, volume 41, number 1, January 2023.
© 2022 The University of Chicago. All rights reserved. Published by The University of Chicago Press in
association with The Society of Labor Economists and The National Opinion Research Center. https://
doi.org/10.1086/718595

77
78 Tan

who grew up in the nonindustrial Midwest, in particular, tend to experience


higher rates of upward mobility. Has this always been the case?
To shed light on the geography of income mobility in the past, Connor
and Storper (2020) construct a longitudinal data set from the historical cen-
suses, from which the extent of upward mobility in each state economic area
(SEA) can then be determined.1 They observe significant changes to the spa-
tial patterns of mobility in some regions and stability in others.2 However,
their analysis does not consider several issues that are key to establishing the
robustness of these findings, including measurement error in the economic
status of parents and the uncertainty associated with the mobility estimates.
My paper attempts to make progress on these fronts.
I begin by building a large linked sample of White sons from the 1910 and
1940 censuses, and I then use it to estimate the degree of upward mobility in
each commuting zone (CZ).3 CZs are chosen over SEAs as the unit of anal-
ysis to allow for a direct comparison with contemporary studies on inter-
generational mobility (Chetty et al. 2014, 2018, 2020; Chetty and Hendren
2018). The data reveal a markedly different land of opportunity in the early
twentieth century. While the South has persistently lagged behind other ar-
eas, upward mobility was initially higher for those who were raised along
the coastal regions and in the industrial Midwest, in contrast to more recent
times. Such patterns are broadly consistent with Connor and Storper (2020)
and are not driven by imperfections in record linkage or measurement error
in the economic status of individuals.
Because CZs are compared on the basis of estimates rather than the true
values of upward mobility, there is some uncertainty over their relative posi-
tions in terms of mobility. How informative might the geographic variation
in intergenerational mobility then be? Applying the inference procedure de-
veloped by Mogstad et al. (2020), I illustrate that the key regions of high and
low mobility can still be identified even after taking uncertainty into account.

Na’ama Shenhav, Mark Shepard, Allison Shertzer, and Nico Voigtläender as well as
participants at the Economic History Association Annual Meeting, the World Eco-
nomic History Congress, the Harvard Economic History Workshop, the Stanford
Economic History Seminar, the Northwestern Seminar in Economic History, and
the Boston University Empirical Microeconomics Lunch Seminar for useful feed-
back. This work was supported by the 2018 Sokoloff Dissertation Fellowship. Con-
tact the author at huiren@nus.edu.sg. Information concerning access to the data
used in this paper is available as supplemental material online.
1
SEAs are groups of contiguous counties with similar economic characteristics
in 1950 (Bogue 1951).
2
In addition to documenting the historical geography of income mobility, Con-
nor and Storper (2020) also attempt to explain why it changed over time. Given the
focus of my paper, I will reference only the first part of their piece.
3
CZs are clusters of counties characterized by strong commuting ties within a
CZ and weak commuting ties between CZs (Autor and Dorn 2013; Tolbert and
Sizer 1996).
A Different Land of Opportunity 79

The final part of this paper then shifts attention from income to education
mobility, the latter of which is not discussed in Connor and Storper (2020).
While income is positively related to education, it also depends on other fac-
tors; the two metrics of mobility could thus exhibit different spatial patterns.
Indeed, I find that the geographic correspondence between income and
education mobility within a given period is not perfect. During the early twen-
tieth century, for example, the nonindustrial Midwest had low rates of in-
come mobility but high rates of education mobility. When focusing solely
on education mobility, I again observe substantial changes in geographic pat-
terns over time, though the changes are not always in sync with the spatial
evolution in income mobility.
This paper contributes to a large and growing literature on intergenera-
tional mobility. Traditionally, much of the work in this field has centered
on addressing the issues of life cycle and attenuation bias (Solon 1992;
Mazumder 2005; Haider and Solon 2006; Nybom and Stuhler 2017). Recent
developments, however, have opened new avenues for research. First, the
availability of large-scale administrative data has enabled researchers to
study intergenerational mobility not just nationally but at finer levels of ge-
ography as well (Chetty et al. 2014, 2018, 2020; Chetty and Hendren 2018).
Second, the release of the complete census records from 1850 to 1940 has led
to an explosion in the use of historical linked samples, some of which have
shed light on how intergenerational mobility has changed over time (Long
and Ferrie 2013; Olivetti and Paserman 2015; Feigenbaum 2018; Ward
2020). My paper synthesizes these developments by constructing estimates
of upward mobility at the CZ level from historical census data.

II. Data
To characterize the land of opportunity in the past, I require information
on the economic status of children and their parents, from which the degree
of intergenerational mobility in each locality can then be estimated.
I start by constructing a large linked sample of individuals, matching native-
born White boys aged 0–18 from the 1910 census to their older selves in the
1940 enumerations.4 The earlier census records the economic standing of

4
I make five points on the population of interest. First, the sample is restricted to
native-born White individuals in order to obtain the widest geographic coverage
while simplifying the analysis by excluding cross-area variation in intergenerational
mobility that is driven by differences in race-nativity composition. Second, I focus
on males, as females tend to change their last names upon marriage, making it dif-
ficult to track them over time by name. Third, an age ceiling of 18 years is imposed
because sons need to be residing with their fathers in 1910 for the latter’s outcomes
to be observed. The older a child, the more likely he is to have left home. Fourth,
while it is possible to construct linked samples from older censuses, a 1910–40 data
set is preferred, as wages were first recorded in 1940. These will be used to impute
the average earnings in each occupation category when measuring the economic
80 Tan

fathers, while the later census contains the outcomes of sons when they are
adults. Both years are available from Integrated Public Use Microdata Series
(IPUMS; Ruggles et al. 2019). To link each person, I use the iterative pro-
cedure popularized by Abramitzky et al. (2012, 2014, 2019), which searches
for exact matches by name, birthplace, and age, failing which an age differ-
ence of 1 and then 2 years is allowed.5 This yields a preliminary panel of
4,235,230 observations, with a match rate of 27.3%.6 In the robustness sec-
tion below, I show that the baseline findings are not sensitive to alternative
methods of linking.
With the linked sample, one can then estimate the extent of absolute up-
ward mobility in each CZ. I focus primarily on absolute rather than relative
mobility, as the latter may be driven by worse outcomes for the rich (Chetty
et al. 2014).7 CZs are chosen as the unit of analysis to facilitate a direct com-
parison with contemporary work on intergenerational mobility (Chetty et al.
2014, 2018, 2020; Chetty and Hendren 2018), as alluded to earlier.8 I define
upward mobility in a given CZ as the expected adult rank of sons who
grew up there and whose fathers are at the 25th percentile.9 Sons are ranked
relative to other sons from the same birth cohort, while fathers are ranked
against other fathers with sons of the same age, both at the national level.10

status of individuals. Fifth, my sample deviates from Connor and Storper (2020),
who match 5–18-year-olds from 1920 to 1940 instead.
5
There are different versions of the iterative procedure (Abramitzky et al. 2012,
2014, 2019). The basic approach uses the New York State Identification and Intel-
ligence System (NYSIIS) algorithm to standardize names on the basis of pronunci-
ation before searching for matches. I adopt the variant that matches individuals on
the basis of actual names instead, as NYSIIS-standardized names tend to increase
the rate of false positives (Bailey et al. 2020).
6
This refers to the sample size before further restrictions are imposed, such as
dropping those with missing information. My match rate is comparable to other
work using similar iterative procedures. For example, Abramitzky et al. (2012) link
29% of Norwegian-born men from the 1865 Norwegian census to the 1900 Nor-
wegian and US censuses. Hoehn-Velasco (2021) matches 29% of males from rural
parts of the United States in 1920 to the 1940 census.
7
Nonetheless, for completeness, the subsequent analysis will also briefly consider
how the geographic patterns of absolute and relative mobility are related.
8
CZs are delineated on the basis of commuting patterns in 1990. The use of
1990-based CZs here does not imply that these were the areas within which people
commuted to work in the early twentieth century. They simply allow one to divide
the country into subunits that can be directly compared with contemporary studies.
The main results in this paper can also be replicated at the county level.
9
I make two points here. First, the CZ of residence in 1910 is taken to be where a
child was raised. Second, much of this paper will focus on sons with fathers at the
25th percentile for consistency with other work on the geography of mobility
(Chetty et al. 2014, 2018, 2020; Chetty and Hendren 2018). To complement this,
the later analysis will also briefly study sons with fathers at the 75th percentile.
10
Not all individuals in the current release of the 1910 census have had their occu-
pation strings translated into standardized occupation codes. For simplicity, fathers
A Different Land of Opportunity 81

Because the US Census Bureau did not ask for total income prior to 1950,
the rankings will be based on imputed occupation earnings.11 Specifically,
I adopt the approach in Collins and Wanamaker (2020) by computing the
average income of White men in each occupation category using wage data
from the 1940 census, stratified by region of residence.12 These imputations,
termed occupation income scores, also adjust for earnings differences be-
tween self-employed and wage workers as well as for perquisites accruing
to agriculture workers.13
In practice, I estimate the rate of upward mobility in each CZ by regress-
ing the ranks of sons on the ranks of their fathers and then using the regres-
sion constant and slope to predict the expected rank of sons whose fathers
are at the 25th percentile, following Chetty et al. (2014, 2020) and Chetty
and Hendren (2018).14 This implicitly assumes that the rank-rank relation
between sons and fathers is approximately linear even at the CZ level. While
true in more recent settings (Chetty et al. 2014, 2020), how plausible might
this be historically?15 Existing work has yet to explore this. To make prog-
ress, figure 1 takes the 1910–40 linked data and displays binned scatterplots

whose occupations have not yet been classified are dropped from the sample, and
those who remain are then ranked among each other.
11
The 1940 census records wages but not income from businesses or other
sources, which are particularly important for self-employed persons and farmers.
12
While IPUMS provides its own version of occupation-based earnings, these
are calculated from the 1950 census and do not vary within occupation categories.
The imputations proposed by Collins and Wanamaker (2020) are thus more rele-
vant (closer in time) and refined (stratified by gender-race-region within occupa-
tions), the latter of which improves the reliability of cross-group comparisons
(Inwood et al. 2019; Saavedra and Twinam 2020). For simplicity, I also code all per-
sons with nonoccupational responses (e.g., retiree, student) as having no earnings.
13
To account for earnings differences between self-employed and wage workers,
Collins and Wanamaker (2020) use the earnings ratio of the two groups in the 1960
census. To adjust for perquisites, they rely on information from the US Depart-
ment of Agriculture.
14
Two points are worth noting here. First, to minimize life cycle bias, I restrict
the linked sample to sons whose fathers were aged 30–50 in 1910, around the mid-
dle of their life cycles. While rank-rank relations can also be biased by when the
outcomes of sons are measured (Chetty et al. 2014; Nybom and Stuhler 2017), this
is unlikely to pose a serious problem, as the sons in my data are around 30–48 years
of age in 1940, falling within the age range where rank-rank associations are most
stable (Chetty et al. 2014; Nybom and Stuhler 2017). Second, Chetty et al. (2018)—
which will later be used to characterize the contemporary geography of mobility—
adopt an alternative estimation approach that accommodates possible nonlinearities
in rank-rank relations. I opt for the simpler linear regression method here since the
rank-rank relations are likely to be reasonably linear, as discussed below.
15
In more recent settings, Chetty et al. (2014) find a linear relation between the
mean income ranks of all children and their parents at the CZ level, while Chetty
et al. (2020) further verify this by race. However, neither paper checks for linearity
at the CZ-race-gender level, the last dimension of which is relevant here, given the
focus on sons.
82 Tan

FIG. 1.—Rank-rank relation between sons and fathers is approximately linear.


Data are from the 1910–40 linked sample. The figure displays binned scatterplots
of the percentile ranks of sons and fathers in the full sample (A) as well as for three
CZs drawn from the upper end (B), middle (C), and lower end (D) of the popula-
tion size distribution. The bins are obtained by dividing the population into 50
comparably sized groups where possible, based on the percentile rank of fathers.
All rankings are measured at the national level and are based on imputed earnings.
Best-fit linear lines are shown, weighted by the size of each bin. A reasonably linear
trend can be observed in each case, although the associations are noisier than those
in contemporary studies. A color version of this figure is available online.

of the ranks of sons and fathers in the full sample (fig. 1A) as well as for three
CZs drawn from the upper end (fig. 1B), middle (fig. 1C), and lower end
(fig. 1D) of the population size distribution. Reasonably linear trends can be
observed across the different cases, although they are much noisier than those
in contemporary studies (Chetty et al. 2014, 2020).16 More generally, to as-
sess linearity for all CZs, I will also consider nonlinear rank-rank specifica-
tions in the subsequent analysis. These produce results that are broadly sim-
ilar to the baseline, suggesting that the linear assumption may be reasonable.
16
Collins and Wanamaker (2020) also observe a fair amount of noise when plot-
ting the ranks of sons and fathers against each other, using samples linked across
1880–1900 and 1910–30. Their plots are for the United States as a whole rather than
local areas. That the historical associations are noisier than those in more modern
studies may not be surprising, as the former are based on imputed rather than actual
earnings.
A Different Land of Opportunity 83

III. The Historical Land of Opportunity


This section presents the geography of intergenerational mobility in the
early twentieth century, assesses the robustness of these patterns, and then
accounts for the uncertainty associated with the mobility-based rankings
of CZs.

A. Baseline Results
The land of opportunity in the past was markedly different from the pre-
sent. Figure 2A maps the historical estimates of upward mobility across the
country, while figure 2B shows the contemporary rates constructed by
Chetty et al. (2018), the latter of which are based on administrative tax re-
cords for the 1978–83 birth cohorts.17 For accuracy and comparability,
CZs with fewer than 500 observations in the historical data are dropped
from both the historical and the contemporary samples, leaving a common
set of 587 CZs.18 These CZs are then ranked by upward mobility and divided
into sextiles, with darker shades indicating higher levels of mobility. I find
that children who grew up in the coastal regions and industrial Midwest
had more opportunities for upward mobility during the early twentieth
century (fig. 2A). In contrast, mobility in more recent times is greater

17
Chetty et al. (2018) provide several versions of upward mobility; I use the rates
based on individual income and non-Hispanic White sons for comparability with
the historical data. Besides Chetty et al. (2018), contemporary estimates of upward
mobility for each CZ are also available from Chetty et al. (2014, 2020) and Chetty and
Hendren (2018). Several factors, however, favor the use of Chetty et al.’s (2018) data.
First, of the aforementioned studies, only Chetty et al. (2018, 2020) provide race-
specific statistics, an important consideration, given the focus on White individuals
in this paper. Second, every CZ has a unique value of upward mobility in Chetty et al.
(2018) but not in Chetty et al. (2020), reflecting the use of different methods to ensure
confidentiality. Third, standard errors are available only for Chetty et al.’s (2018) es-
timates; these will be needed to account for the uncertainty in mobility-based rankings
below. For the aforementioned reasons, the data in Chetty et al. (2018) are preferred.
18
To be precise, the threshold of 500 observations refers to the number of effective
observations. Because the 1910 census contains information on the county but not
the CZ of residence, the former needs to be mapped over to the latter. However, a
1910 county is not always fully contained within the boundaries of a single 1990-
based CZ. To address this, I assign each son one or more weights, calculated as
the share of his county’s area falling within a given CZ for all intersecting CZs.
The sum of these weighted individuals gives the effective number of observations.
These weights are applied when estimating the rank-rank regressions for each CZ.
Had a similar exercise been performed for Black individuals, the sample would com-
prise just 101 CZs, over 77% of which are in the Deep South (Alabama, Arkansas,
Florida, Georgia, Louisiana, Mississippi, North Carolina, South Carolina, and Texas).
This narrow geographic scope partly reflects the clustering of Black individuals in
the South, but it is also a function of the smaller population size and poorer match
quality for Black relative to White individuals. The restricted range of areas is partic-
ularly limiting when attempting to study the geography of opportunity across the
country, thus reinforcing the choice to focus on White individuals here.
84 Tan

FIG. 2.—Geography of opportunity: sons from poorer households. Data are


from the 1910–40 linked sample (A) and Chetty et al. (2018; B). The figure shows
the extent of absolute upward mobility in a given CZ, defined as the predicted adult
rank of sons who grew up there and whose fathers (A) or parents (B) are at the 25th per-
centile. For accuracy and comparability, CZs with fewer than 500 observations in the
historical data are dropped from both the historical and the contemporary samples.
Sample size is 587 CZs. A color version of this figure is available online.

among those who were raised in the nonindustrial Midwest (fig. 2B). While
the South has persistently lagged behind other areas, the spatial correlation
between the past and present patterns on the whole is still low at 0.119, sug-
gesting that much change has occurred over time. The findings here can also
A Different Land of Opportunity 85

be observed at the county level (fig. A1; figs. A1–A13 are available online)
and are broadly consistent with Connor and Storper (2020).
While not the primary focus of this paper, similar spatial patterns are evi-
dent among richer families as well.19 For each CZ, figure 3 shows the pre-
dicted rank of sons who grew up there and whose fathers or parents are at the
75th percentile.20 Within a given period (historical or contemporary), richer
children tend to do better in places where their poorer counterparts experience
more upward mobility (figs. 2, 3).21 Across periods, one again observes a shift
in where opportunities are more abundant, from the outer parts of the coun-
try in the past to the central regions more recently (fig. 3). As before, much of
the South persistently lags behind other areas, but the overall correlation be-
tween the past and present geographies remains low at 0.103. That the land of
opportunity has evolved in a similar way for the rich and poor suggests that
the underlying drivers of change are likely to transcend income groups.
Would the same geographic patterns be observed if relative rather than ab-
solute mobility had been used instead? I explore this in three steps.22 First, I
use the 1910–40 linked sample to predict the ranks of sons at 50 different per-
centiles across the entire distribution of fathers. These capture the degree of
absolute mobility at various percentiles (Chetty et al. 2014). Second, for each
of the 50 percentiles, I calculate the CZ-level correlation between the extent
of absolute and relative mobility.23 The latter is given by the slope of the son-
father rank-rank regression used previously and is common across percen-
tiles, as in Chetty et al. (2014). A negative association between the two met-
rics would imply that CZs with greater relative mobility also generate better
outcomes for those with fathers at a given percentile. Third, I conduct a sim-
ilar exercise for more recent cohorts, based on the mobility estimates from
Chetty et al. (2018).24 Figure 4 then plots the respective correlations for the
historical and contemporary periods.
19
Connor and Storper (2020) do not consider this subset of the population.
20
For consistency, fig. 3 uses the same set of CZs as fig. 2.
21
The correlation between figs. 2A and 3A is 0.807, while the correlation be-
tween figs. 2B and 3B is 0.891.
22
Chetty et al. (2014) provide a nonparametric version of the exercise that follows.
23
As a robustness check, Connor and Storper (2020) compare absolute mobility
at the 25th percentile and relative mobility across SEAs. However, their use of ab-
solute mobility at just 1 percentile masks important heterogeneity in this relation, as
documented below.
24
For each CZ, Chetty et al. (2018) provide the predicted ranks of sons whose par-
ents are at the 1st, 25th, 50th, 75th, and 100th percentiles. Assuming a linear relation
between the ranks of sons and fathers in a given CZ, I back out a constant and slope
for every CZ from the predictions at the 25th and 75th percentiles. These are then
used to obtain the predictions for all 50 percentiles. To check whether my results
are sensitive to the linearity assumption, I redo the predictions on the basis of the
1st and 25th percentiles as well as the 75th and 100th percentiles separately. The re-
sulting estimates of absolute and relative mobility are highly correlated (well over
0.9) with the original measures.
FIG. 3.—Geography of opportunity: sons from richer households. Data are
from the 1910–40 linked sample (A) and Chetty et al. (2018; B). The figure shows
the predicted adult rank of sons who grew up in a given CZ and whose fathers
(A) or parents (B) are at the 75th percentile. For accuracy and comparability, CZs
with fewer than 500 observations in the historical data are dropped from both the
historical and the contemporary samples. Sample size is 587 CZs. A color version
of this figure is available online.

86
FIG. 4.—Spatial relation between absolute and relative mobility. Data are from the
1910–40 linked sample (A) and Chetty et al. (2018; B). Each marker denotes the CZ-
level correlation between the degree of absolute mobility at a specific percentile of
fathers (A) or parents (B) and the extent of relative mobility (common across percen-
tiles). The former is measured as the predicted adult rank of sons whose fathers (A) or
parents (B) are at a given percentile, based on son-father rank-rank regressions. The
latter is given by the slopes of the same regressions. Dashed lines indicate zero cor-
relations. For accuracy and comparability, CZs with fewer than 500 observations in
the historical data are dropped from both the historical and the contemporary sam-
ples. Sample size is 587 CZs. A color version of this figure is available online.

87
88 Tan

The spatial correlation between absolute and relative mobility varies mono-
tonically with the economic standing of fathers. Figure 4 shows that this as-
sociation starts off strongly negative at low percentiles and becomes less neg-
ative as the ranks of fathers increase, eventually crossing the zero line in the
upper part of the distribution.25 The negative correlations across most per-
centiles indicate that places with higher relative mobility produce better out-
comes for the majority of the population. This is particularly evident among
sons from poorer households. In some instances, greater relative mobility
may also be related to worse outcomes, as indicated by the positive correla-
tions at higher percentiles. Nonetheless, the results as a whole point to a fair
amount of overlap in the geographic patterns of absolute and relative mobility.

B. Robustness Checks
Having characterized the geography of upward mobility in the early
twentieth century, I now assess the robustness of these patterns. This is done
by implementing a suite of checks, each of which makes specific adjustments
to the data or estimates before reconstructing the map of mobility; figure A2
displays these maps.26 The alternative maps are then compared with the base-
line in figure 2A, and the resulting correlations are presented in figure 5.27 I
find reasonably consistent patterns across the different exercises, suggesting
that the geographic divide documented earlier likely reflects the true land of
opportunity in the past.

1. Linked Sample Quality


I start by investigating whether the quality of the 1910–40 linked data af-
fects the reliability of the geographic patterns of upward mobility. Bailey
et al. (2020) show that automated linking methods are susceptible to false pos-
itives and do not generate representative samples, both of which could bias
inferences on intergenerational mobility.28 I consider the importance of false

25
For the contemporary period (fig. 4B), the zero line crossing is somewhat sen-
sitive to how absolute and relative mobility are constructed. Footnote 24 described
three ways of obtaining these values; they yield crossing points that range from the
80th to 99th percentile. Despite the uncertainty here, the change in sign always oc-
curs at the upper end of the distribution.
26
Figure A2 focuses on sons from poorer households (primarily those with fa-
thers at the 25th percentile). Figure A3 gives the corresponding maps for sons from
richer families (mainly persons with fathers at the 75th percentile).
27
Figure 5 considers sons with fathers at the 25th percentile or from the bottom
half of the occupation income distribution. Figure A4 shows the results for sons
with fathers at the 75th percentile or from the top half of the distribution.
28
Two points are worth noting here. First, automated linking methods do not pro-
duce perfectly accurate or representative samples, as there are only a few individual-
level characteristics that can be used to match each person, the combination of which
may not even be unique, unlike contemporary administrative records containing
A Different Land of Opportunity 89

FIG. 5.—Robustness checks for historical geography of opportunity. Data are


from the 1910–40 linked sample. The figure is based on a set of robustness checks
(y-axis). Each exercise makes specific adjustments to the data or estimates before
reconstructing the map of opportunity for sons whose fathers are at the 25th per-
centile or from the bottom half of the occupation income distribution. The alterna-
tive maps are then compared with the original patterns in figure 2A, and the result-
ing correlations are presented in the figure. Dashed line indicates a perfect positive
correlation. Sample size (in brackets) ranges from 361 to 587 CZs. A color version
of this figure is available online.

positives and unrepresentativeness in turn; neither appears to have distorted


the original findings substantially.
To understand the role of false positives, I use different linking techniques
to reconstruct the 1910–40 sample. First, I adopt a conservative version of
the baseline iterative procedure that requires names to be unique within
5-year age bands in both the 1910 and the 1940 censuses as well as an exact
match on age (Abramitzky et al. 2012, 2014, 2019). This lowers the incidence
of false positives substantially but at the cost of fewer matches (Bailey et al.
2020).29 A similar check is performed in Connor and Storper (2020). Second,
I go further by implementing the machine learning approach pioneered by

unique individual identifiers. These characteristics may also be measured with error.
Second, Bailey et al. (2020) show that imperfections in record linkage can bias esti-
mates of the intergenerational income elasticity obtained from log-log income regres-
sions, but they do not consider the rank-rank relations used in this paper.
29
With fewer matches, the set of CZs for which upward mobility can be accu-
rately estimated is also smaller. Figure 5 indicates (in brackets) the number of
CZs with at least 500 observations in both the baseline and the alternative samples.
90 Tan

Feigenbaum (2016), which generates relatively more accurate links (Bailey


et al. 2020). This is carried out by manually matching a random sample of
individuals, using it to train a machine algorithm, and then allowing the al-
gorithm to select matches in the full space of potential links.30 I calibrate both
a less and a more conservative machine algorithm, where the former places
equal emphasis on accuracy and efficiency while the latter gives thrice the
weight to accuracy. Third, I take the intersection of links that are identified
by different methods. This helps avoid idiosyncratic errors but reduces the
match rate dramatically (Bailey et al. 2020). To ensure that each CZ has a suf-
ficient number of observations for reliable estimation and that the set of
these CZs remains reasonably large, I focus on links that are common to just
the basic iterative procedure and the less conservative machine learning algo-
rithm. Reassuringly, the alternative linked samples consistently yield spatial
patterns of mobility that are highly correlated with the baseline, as indicated
in the top panel of figure 5.
To assess the influence of unrepresentativeness, I follow Bailey et al.’s
(2020) recommendation by using inverse propensity weights to adjust for
observable differences between matched and unmatched persons, after
which the geography of upward mobility is reconstructed.31 Figure 5 reveals
a sizable overlap between this map and the baseline, reinforcing the original
findings.32

30
More precisely, the algorithm assigns a probabilistic score to each potential
match. A potential match is then declared to be true if its score is (1) the highest
for a given individual in 1910, (2) sufficiently high in absolute terms, and (3) suffi-
ciently high relative to the second-highest alternative, if any. The thresholds used in
points 2 and 3 are calibrated to balance the trade-off between accuracy and efficiency
in the matching process.
31
Adapting the approach in Bailey et al. (2020), I compute the inverse propensity
weights in two steps. First, I run a probit model of match status on a vector of char-
acteristics: the length of first and last names separately, the share of persons with the
same first name, the proportion of individuals having the same last name, a quadratic
in age, an indicator for those who have siblings in their 1910 households, the number
of siblings, a quadratic in the age of one’s father, the father’s occupation category
(white-collar, farmer, skilled/ semiskilled, and unskilled), an indicator for house-
holds residing in urban areas, and a dummy for farm households. Second, the probit
coefficients are used to estimate each individual’s probability of being matched, and
inverse propensity weights are computed as ½ð1 2 pÞ=p  ½m=ð1 2 mÞ, where p is
the propensity score and m is the match rate.
32
The observed robustness is not surprising; table A1 (tables A1, A2 are available
online) shows that the linked data are not particularly unrepresentative to begin
with, at least along the dimensions considered. This could partly reflect the focus
on native-born White individuals in this paper, which eliminates potentially large
differences in representation along race and nativity. For example, Ward (2020)
links both Black and White individuals across historical census records and finds
that Black individuals have to be given several times the weight of White individuals
in order to make the sample representative of the underlying population.
A Different Land of Opportunity 91

In summary, while the 1910–40 linked sample is not perfect, its limita-
tions are unlikely to have distorted the spatial patterns of upward mobility
substantially.

2. Measurement Error
Next, I explore whether errors in the measurement of individual economic
status could have biased the local estimates of upward mobility.
One limitation of the analysis thus far is the use of occupation income
scores to measure socioeconomic standing. By construction, these scores
are imperfect proxies for actual earnings—if the degree of imperfection var-
ies systematically across the country, it could distort the true geographic var-
iation in upward mobility. To shed light on this, I restrict the 1910–40 linked
sample to sons who were wage and salary workers in 1940. Such persons can
be ranked by their annual wages instead of just occupation scores.33 Using
this subset of the population and their wage-based rankings generates spatial
patterns of mobility that are comparable to the baseline, as reflected by the
high correlation in the middle panel of figure 5.34
A more traditional concern in the intergenerational mobility literature is
attenuation bias due to the mismeasurement of parents’ permanent economic
status (Solon 1992), an issue that is not addressed in Connor and Storper
(2020). Ward (2020), in particular, constructs his own historical linked data
and shows that mobility estimates can be severely biased when the occupa-
tion scores of fathers are observed just once. This is because a single obser-
vation of economic status is subject to much measurement error arising from
the misreporting of occupations (Ward 2020). Again, if the resulting bias
varies systematically across space, it could distort the true patterns of up-
ward mobility. To make progress, I average the economic status of fathers
across two observations, similar in spirit to Ward (2020). This requires some
additional information. I thus link fathers in 1910 forward to 1920 and back-
ward to 1900 separately and combine them with the original 1910–40 links
for sons to generate 1910–1920–1940 and 1910–1900–1940 linkages.35 The
33
Even with individual wage income, the geography of upward mobility could
still be distorted for two reasons. First, one can adjust the measure of economic sta-
tus for sons in 1940 but not for fathers in 1910, when wage data are unavailable.
Second, wage income might also be measured with error.
34
This exercise shrinks the underlying sample by about 30%; such a large and
nonrandom selection naturally imposes an upper bound on the spatial correlation
with the baseline. To illustrate this, I rank the same wage workers by occupation
scores and find a correlation of 0.916 between the resulting map of mobility and
the original geography. The reduced sample also restricts the set of CZs for which
upward mobility can be reliably estimated.
35
Each additional linkage reduces the sample size substantially. This matters be-
cause unlike much of the historical mobility literature where the analysis is con-
ducted at the national level (Long and Ferrie 2013; Olivetti and Paserman 2015;
Feigenbaum 2018; Ward 2020), my paper focuses on local statistics. Constructing
92 Tan

extended data allow one to calculate the mean occupation scores of fathers
over 1910–20 or 1910–1900, rerank them, and then reconstruct the local
rates of upward mobility. Reassuringly, this yields spatial patterns that are
similar to the baseline, as indicated in figure 5.

3. Alternative Estimates of Upward Mobility


I now consider whether the geographic patterns are sensitive to how up-
ward mobility is measured. As a start, I redefine the level of upward mobility
in a given CZ as the mean adult rank of sons who grew up there and whose
fathers are from the bottom half of the national occupation income distribu-
tion.36 This does not impose any parametric assumptions on the rank-rank
relation between sons and fathers. However, such a metric could be distorted
by differences in income distribution across CZs: the spatial variation in
mobility might simply reflect areas where fathers tend to be richer within the
bottom half of households and whose sons did better accordingly, and vice
versa. Nonetheless, the high correlation in figure 5 indicates that the alterna-
tive definition yields spatial patterns that mirror the baseline. As a separate
exercise, I also use quadratic and cubic models rather than a linear specifica-
tion to predict the original measure of upward mobility, thus allowing for
potential nonlinearities in the rank-rank associations. Both leave the geogra-
phy of opportunity largely unchanged, as seen from the high correlations in
figure 5.

4. Excluding Sons of Farmers


Finally, I check whether the spatial variation in upward mobility is pri-
marily driven by local differences in farming intensity. To do this, I exclude
sons of farmers from the 1910–40 sample, rerank the remaining individuals,
and then reconstruct the map of opportunity. Figure 5 reveals a strong cor-
relation between this alternative map and the baseline, suggesting that the geo-
graphic divide in mobility documented above captures features of the general
population and not just the presence of farmers. By extension, the evolution
in spatial patterns over time is also unlikely to reflect the shift away from ag-
riculture alone.

reliable estimates of upward mobility for each CZ requires a sufficiently large num-
ber of observations in a given locality. I thus avoid using a 1910–1900–1920–1940
sample, where fathers are linked to both the 1900 and the 1920 censuses simulta-
neously. Not doing so could leave some bias unaccounted for. However, it is worth
noting that the reduction in bias when going from one to two father observations is
larger than the change going from two to three observations (Ward 2020). If the
geographic patterns remain stable when switching from one to two observations,
this might indicate that the remaining bias has limited influence as well.
36
When studying sons from richer households (figs. A3, A4), I compute the mean
rank of sons with fathers from the top half of the occupation income distribution
instead.
A Different Land of Opportunity 93

C. Uncertainty in Mobility-Based Rankings


How informative are the geographic patterns of upward mobility in the
early twentieth century? The preceding sections used heat maps to illustrate
this variation, where CZs were ranked by mobility rates and then divided
into sextiles. Because the rankings are based on estimates rather than the true
values of upward mobility, there is some uncertainty over whether a CZ has
more or less mobility than other CZs. A high-mobility area might simply
reflect statistical noise instead of a genuine advantage and similarly for a
low-mobility locality. Mogstad et al. (2020) raise this concern with Chetty
et al.’s (2018) contemporary estimates of mobility: they find that it is often
not possible to determine whether a given place has more or less mobility
than other places after accounting for the uncertainty in mobility-based
rankings, except for the Great Plains and Southeast.37 Connor and Storper
(2020) do not consider whether this might be an issue in the past. In this sec-
tion, I show that the historical results are subject to a similar limitation but
that one can still identify the key regions of high and low mobility even after
uncertainty has been factored in.
To shed light on the value of mobility-based comparisons, I proceed in
three steps. First, I generate bootstrapped standard errors for the historical
rates of upward mobility; these capture the uncertainty around the local
measures of mobility.38 Second, I apply Mogstad et al.’s (2020) inference pro-
cedure to the point estimates of upward mobility and the corresponding
standard errors, constructing random sets that contain the mobility-based
ranks of all CZs with at least 95% probability, termed simultaneous confi-
dence sets. Third, I divide CZs into several groups: a high-mobility group
comprising CZs with simultaneous confidence sets that lie entirely in the
top half of the mobility-based rankings (likely top half), a low-mobility
group covering CZs with confidence sets that are completely within the bot-
tom half (likely bottom half), and an undetermined group encompassing
CZs with confidence sets that do not allow them to be assigned to either
group. Figure 6A then maps the different groups of CZs. As a point of com-
parison, figure 6B presents the analogous map for the more recent period,
using the point estimates and standard errors from Chetty et al. (2018).39

37
Mogstad et al. (2020) find that mobility-based rankings are even less informa-
tive at the county level as the degree of uncertainty tends to be greater. They also
show that the causal effects of local areas on adult outcomes estimated in Chetty
and Hendren (2018) do not provide informative rankings even at the CZ level.
38
The bootstrapped standard errors are based on 500 replications separately for
each CZ. For this section, the point estimates of upward mobility and standard er-
rors are both obtained from unweighted rank-rank regressions.
39
Mogstad et al. (2020) provide a similar map with Chetty et al.’s (2018) data.
However, they use the estimates derived from the household income of all races
and both genders, whereas I draw on the rates constructed from the individual in-
come of non-Hispanic White males for comparability with the historical data.
FIG. 6.—Likely top and bottom halves of CZs based on mobility. Data are from
the 1910–40 linked sample (A) and Chetty et al. (2018; B). Each figure identifies
three groups of CZs: a high-mobility group comprising CZs with simultaneous
confidence sets that lie entirely in the top half of the mobility-based rankings (likely
top half), a low-mobility group covering CZs with confidence sets that are com-
pletely within the bottom half (likely bottom half), and an undetermined group en-
compassing CZs with confidence sets that do not allow them to be assigned to ei-
ther group. Simultaneous confidence sets are constructed for each CZ using the
predicted adult rank of sons who have fathers (A) or parents (B) at the 25th percen-
tile and the corresponding standard errors. Sample size is 587 CZs.
A Different Land of Opportunity 95

There are two takeaways from figure 6.40 First, one often cannot tell whether
a CZ has more or less upward mobility than other CZs in both the past and the
present. This can be seen from the large set of CZs falling into the undeter-
mined category.41 Second, in spite of this uncertainty, the main findings from
the baseline analysis can still be observed here. With 95% confidence, upward
mobility during the early twentieth century was higher along the coasts and
lower in some sections of the nonindustrial Midwest and South (fig. 6A). In
more recent times, parts of the nonindustrial Midwest and South can be con-
fidently identified as regions of high and low mobility, respectively (fig. 6B).
Thus, one can still conclude that the land of opportunity in the past was mark-
edly different from the present even after accounting for uncertainty.42
The above results also reveal that mobility-based rankings tend to be more
informative at the upper and lower ends of the distribution. Why might this
be? Following Mogstad et al. (2020), I consider two possible explanations.
First, upward mobility may be more precisely estimated for CZs at the top
and bottom. Figure A6, however, shows that the standard errors of the mo-
bility estimates are not particularly small for CZs at either end. Second, the
point estimates of upward mobility could be more dispersed in the upper
and lower parts of the ranking compared with the middle. Figure A7 plots
the differences in mobility between CZs that are ranked adjacent to each other
and finds a larger spread at the tail ends, consistent with this hypothesis.

IV. Income versus Education Mobility


Thus far, this paper has focused on the historical geography of income
mobility. How does it compare with the spatial variation in education mo-
bility, broadly defined as a child’s education attainment by adulthood con-
ditional on the economic status or human capital of one’s parents? Connor
and Storper (2020) do not explore this aspect of mobility. While income is
positively related to education (Angrist and Krueger 1991; Ashenfelter and
Krueger 1994; Feigenbaum and Tan 2020), it is also shaped by other factors.43
The two dimensions of mobility could thus exhibit different geographic pat-
terns. This matters because historical data on individual income are hard to
come by, in which case one might turn to education as a measure of socio-
economic standing in the past. Card et al. (2018), for example, compare

40
Figure 6 is based on the predicted rank of sons with fathers or parents at the
25th percentile. Figure A5 provides the corresponding maps for sons whose fathers
or parents are at the 75th percentile.
41
Specifically, 49.2% and 70.0% of the 587 CZs in fig. 6A and 6B, respectively,
fall into the undetermined category. The corresponding proportions in fig. A5 are
even higher at 76.8% and 82.5%.
42
Reasonably similar conclusions can be reached with sons from richer house-
holds, although the results are relatively less robust (see fig. A5).
43
Heckman (1995) observes that measured cognitive ability explains only a small
part of the variation in earnings.
96 Tan

the geography of education mobility in 1940 with the spatial variation in in-
come mobility more recently, the latter of which is based on Chetty et al.
(2014).44 They observe some overlap in geographic patterns and conclude
that the local drivers of mobility are highly persistent, seemingly at odds with
the findings in my paper.45 In this section, I show that the spatial correspon-
dence between income and education mobility in the same period is not per-
fect. Caution is thus necessary when using different metrics to study how the
land of opportunity has evolved over time.
To study the spatial relation between income and education mobility, I
proceed as follows. For the historical period, I take the 1910–40 linked sam-
ple, restrict it to sons whose fathers are from the bottom half of the occupa-
tion income distribution, and then calculate the share of sons from each CZ
who finished at least grade j by 1940, where j ∈ f8, 9, 10, 11, 12g, as well as
their average years of schooling.46 These measures of education mobility are
compared against the geography of income mobility in figure 2A, and the
resulting correlations are presented in figure 7A.47 For the contemporary pe-
riod, I use the predicted share of sons from each CZ who completed high
school, acquired some college experience, had at least a community college
degree, obtained a 4-year college degree, or finished a graduate degree, con-
ditional on their parents being at the 25th percentile of the income distribu-
tion.48 These are juxtaposed against the patterns of income mobility in fig-
ure 2B, and the correlations are shown in figure 7B.49 For completeness,
figures A10 and A11 provide the maps of education mobility by period.50

44
For each locality, Card et al. (2018) define upward mobility in education as the
fraction of 16–18-year-olds who completed at least 9 years of schooling among
those whose parents have 5–8 years of schooling (roughly the middle of the parent
education distribution).
45
Derenoncourt (2021) presents a similar comparison for Black individuals. She
finds a shift in geographic patterns over time, in contrast to Card et al. (2018). This
difference could partly stem from the more restricted geographic coverage in
Derenoncourt (2021), where much of the Midwest and West are excluded because
of the small number of Black individuals.
46
Two points are worth noting here. First, the conditioning is based on the eco-
nomic status of fathers instead of their education levels, as the latter are not recorded
in 1910. Second, my education metrics are constructed as simple averages rather than
regression-based predictions because the education-rank relation between sons and
fathers varies with the specific measure of education, as illustrated by fig. A8. While
this differs from how the local rates of income mobility are estimated, it is unlikely to
pose a serious problem since the spatial patterns of income mobility are similar under
the regression-based and simple average measures (see fig. 5).
47
Figure A9A shows the corresponding correlations when the historical sample
is restricted to sons with fathers from the top half of the national occupation income
distribution and when the income mobility patterns are those in fig. 3A.
48
The five education-related variables are from Chetty et al. (2018).
49
Figure A9B shows the results for sons with parents at the 75th percentile.
50
Figures A12 and A13 display the maps for sons from richer households.
FIG. 7.—Spatial relation between income and education opportunity. Data are
from the 1910–40 linked sample (A) and Chetty et al. (2018; B). Each marker denotes
the correlation between the geographic patterns of income and education opportunity.
The former is given by figure 2. The latter is measured as the share of sons from a given
CZ that meet the respective education thresholds along the y-axis, among those whose
fathers are from the bottom half of the occupation income distribution (A) or whose
parents are at the 25th percentile of the income distribution (B). Sample sizes are
484 CZs (A) and 587 CZs (B), except in the case of graduate degree completion, where
there are 586 CZs. A color version of this figure is available online.
98 Tan

Table 1
Intertemporal Correlation in Geography of Education Opportunity
Contemporary Measures
Historical High Some Community College 4-Year College Graduate
Measures School College Degree Degree Degree
At least grade 8 .430 .453 .450 .345 .035
At least grade 9 .069 .299 .199 .286 .164
At least grade 10 .135 .342 .252 .315 .156
At least grade 11 .188 .384 .298 .339 .138
At least grade 12 .245 .414 .341 .358 .116
Years of schooling .355 .464 .425 .380 .093
NOTE.—Data are from the 1910–40 linked sample (historical measures) and Chetty et al. (2018; contem-
porary measures). Each cell shows the CZ-level correlation in education opportunity between the past and
present, based on the thresholds for the historical period (historical measures) and the levels for the con-
temporary period (contemporary measures). The focus is on sons with fathers from the bottom half of
the occupation income distribution (historical measures) or with parents at the 25th percentile of the in-
come distribution (contemporary measures). Sample size is 484 CZs.

Figure 7 reveals two things. First, there is a positive association between the
spatial variation in income and education mobility in both the past and the
present.51 This may not be surprising, given the positive relation between in-
come and education. Second, however, the correlations often fall below 0.6.52
This might be consistent with Rothstein (2019), who shows that geographic
differences in the child-parent education-income gradient explain only part
of the variation in income mobility. While the less than perfect correlations
could partly be due to statistical noise, they also reflect more fundamental re-
gional differences. For example, during the early twentieth century, the non-
industrial Midwest lagged behind other areas in terms of income mobility
(fig. 2A) but was one of the strongest performers on the basis of education mo-
bility (fig. A10). The results here highlight the need for caution when using dif-
ferent measures to compare the patterns of mobility across space and over time.
When focusing on just education mobility, one again observes significant
changes in geographic patterns from the past to present, though not always
in the same direction as the evolution in income mobility. Table 1 presents a
matrix of correlations between the historical and contemporary rates of ed-
ucation mobility at the CZ level, using the metrics in figure 7.53 While

51
There is only one instance of a negative correlation: when the comparison is be-
tween income and education mobility in the present, where the latter is based on grad-
uate degree completion, among sons with parents at the 75th percentile (see fig. A9B).
52
The correlations are even weaker among richer households (see fig. A9).
53
Table 1 focuses on sons with fathers from the bottom half of the occupation
income distribution (for the historical period) or with parents at the 25th percentile
of the income distribution (for the contemporary period). Table A2 gives the cor-
responding results for sons with fathers from the top half or with parents at the
75th percentile.
A Different Land of Opportunity 99

positive, the associations usually fall below 0.4, pointing to substantial changes.54
Again, the nonindustrial Midwest provides an example of how the geogra-
phies of income and education mobility differ over time. Section III found
that this region moved from a low to a high position in terms of income mo-
bility, whereas figures A10 and A11 show that large parts of the nonindus-
trial Midwest retained their leading positions in education mobility. These
observations reinforce the need for caution when comparing the spatial var-
iation in mobility along different dimensions.

V. Conclusion
What was the geography of intergenerational mobility like in the early
twentieth-century United States? Constructing a large linked sample of in-
dividuals from the 1910 and 1940 censuses, I show that the spatial patterns of
mobility in the past were markedly different from the present. While chil-
dren growing up in the middle of the country experience some of the highest
rates of upward mobility today, there were relatively more opportunities in
the coastal regions and industrial areas historically. These findings are
broadly consistent with Connor and Storper (2020). To take the analysis
further, I also examine the robustness of the observed patterns in light of
a range of concerns that are not addressed in Connor and Storper (2020), in-
cluding measurement error in the economic standing of parents and the un-
certainty associated with the mobility estimates.
I conclude by proposing two directions for further work. First, this paper
studies White sons. Has the geography of opportunity for Black individuals
and daughters evolved in a similar way? A limiting factor for both groups is
how well one can match them across historical data sets, which in turn af-
fects the size, geographic coverage, reliability, and representativeness of
the linked sample. This constraint may be gradually relaxed as newer link-
ing methods come to the fore. Feigenbaum and Gross (2021), for instance,
develop a genealogy-based approach that is able to match a large number of
women in the early twentieth century.
Second, why did the land of opportunity change? Recent work has begun
to explore this question, but much remains to be done. Derenoncourt (2021),
for example, finds that the Great Migration of African Americans increased
crime and policing in northern destinations, which lowered the gains from
growing up in those places. Focusing on the Great Migration, however, nec-
essarily excludes much of the nonindustrial Midwest, where few Black indi-
viduals lived but where opportunities for upward mobility have improved
substantially compared with the rest of the country (fig. 2). Handy and Shester
(2020), on the other hand, argue that changes in the geographic patterns of

54
The correlations are lower and can even be negative when looking at sons from
richer households (see table A2).
100 Tan

mobility can largely be explained by changing labor market conditions and


housing prices. However, their analysis is primarily restricted to cohorts
from the 1980s, missing out on developments in the earlier decades. Under-
standing the factors behind the spatial evolution in mobility across a much
wider area and over a longer time horizon is likely to be a fruitful avenue
for future research.

References
Abramitzky, Ran, Leah Platt Boustan, and Katherine Eriksson. 2012.
Europe’s tired, poor, huddled masses: Self-selection and economic out-
comes in the Age of Mass Migration. American Economic Review 102,
no. 5:1832–56.
———. 2014. A nation of immigrants: Assimilation and economic out-
comes in the Age of Mass Migration. Journal of Political Economy 122,
no. 3:467–506.
———. 2019. To the New World and back again: Return migrants in the
Age of Mass Migration. Industrial and Labor Relations Review 72, no. 2:
300–322.
Angrist, Joshua D., and Alan B. Krueger. 1991. Does compulsory school
attendance affect schooling and earnings? Quarterly Journal of Econom-
ics 106, no. 4:979–1014.
Ashenfelter, Orley, and Alan Krueger. 1994. Estimates of the economic re-
turn to schooling from a new sample of twins. American Economic Re-
view 84, no. 5:1157–73.
Autor, David H., and David Dorn. 2013. The growth of low-skill service
jobs and the polarization of the US labor market. American Economic
Review 103, no. 5:1553–97.
Bailey, Martha J., Connor Cole, Morgan Henderson, and Catherine Mas-
sey. 2020. How well do automated linking methods perform? Lessons
from US historical data. Journal of Economic Literature 58, no. 4:997–
1044.
Bogue, Donald J. 1951. State economic areas: A description of the procedure
used in making a functional grouping of the counties of the United States.
Washington, DC: US Government Printing Office.
Card, David, Ciprian Domnisoru, and Lowell Taylor. 2018. The intergen-
erational transmission of human capital: Evidence from the Golden Age
of upward mobility. NBER Working Paper no. 25000, National Bureau
of Economic Research, Cambridge, MA.
Chetty, Raj, John N. Friedman, Nathaniel Hendren, Maggie R. Jones, and
Sonya R. Porter. 2018. The opportunity atlas: Mapping the childhood
roots of social mobility. NBER Working Paper no. 25147, National Bu-
reau of Economic Research, Cambridge, MA.
A Different Land of Opportunity 101

Chetty, Raj, and Nathaniel Hendren. 2018. The impacts of neighborhoods


on intergenerational mobility I: Childhood exposure effects. Quarterly
Journal of Economics 133, no. 3:1107–62.
Chetty, Raj, Nathaniel Hendren, Maggie R. Jones, and Sonya R. Porter.
2020. Race and economic opportunity in the United States: An intergen-
erational perspective. Quarterly Journal of Economics 135, no. 2:711–83.
Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez. 2014.
Where is the land of opportunity? The geography of intergenerational
mobility in the United States. Quarterly Journal of Economics 129,
no. 4:1553–623.
Collins, William J., and Marianne H. Wanamaker. 2020. African American
intergenerational economic mobility since 1880. NBER Working Paper
no. 23395, National Bureau of Economic Research, Cambridge, MA.
Connor, Dylan Shane, and Michael Storper. 2020. The changing geography
of social mobility in the United States. Proceedings of the National Acad-
emy of Sciences of the USA 117, no. 48:30309–17.
Derenoncourt, Ellora. 2021. Can you move to opportunity? Evidence from
the Great Migration. Working paper.
Feigenbaum, James J. 2016. A machine learning approach to census record
linking. Working paper.
———. 2018. Multiple measures of historical intergenerational mobility:
Iowa 1914 to 1940. Economic Journal 128, no. 612:F446–F481.
Feigenbaum, James J., and Daniel P. Gross. 2021. Automation and the fu-
ture of young workers: Evidence from telephone operation in the early
20th century. NBER Working Paper no. 28061, National Bureau of Eco-
nomic Research, Cambridge, MA.
Feigenbaum, James J., and Hui Ren Tan. 2020. The return to education in
the mid-twentieth century: Evidence from twins. Journal of Economic
History 80, no. 4:1101–42.
Haider, Steven, and Gary Solon. 2006. Life-cycle variation in the associa-
tion between current and lifetime earnings. American Economic Review
96, no. 4:1308–20.
Handy, Christopher, and Katherine L. Shester. 2020. Accounting for
changes in intergenerational mobility. Working paper.
Heckman, James J. 1995. Lessons from the bell curve. Journal of Political
Economy 103, no. 5:1091–120.
Hoehn-Velasco, Lauren. 2021. The long-term impact of preventative public
health programs. Economic Journal 131, no. 634:797–826.
Inwood, Kris, Chris Minns, and Fraser Summerfield. 2019. Occupational
income scores and immigrant assimilation. Evidence from the Canadian
census. Explorations in Economic History 72:114–22.
Long, Jason, and Joseph Ferrie. 2013. Intergenerational occupational mo-
bility in Great Britain and the United States since 1850. American Eco-
nomic Review 103, no. 4:1109–37.
102 Tan

Mazumder, Bhashkar. 2005. Fortunate sons: New estimates of intergener-


ational mobility in the United States using Social Security earnings data.
Review of Economics and Statistics 87, no. 2:235–55.
Mogstad, Magne, Joseph P. Romano, Azeem Shaikh, and Daniel Wilhelm.
2020. Inference for ranks with applications to mobility across neighbor-
hoods and academic achievement across countries. NBER Working Pa-
per no. 26883, National Bureau of Economic Research, Cambridge, MA.
Nybom, Martin, and Jan Stuhler. 2017. Biases in standard measures of in-
tergenerational income dependence. Journal of Human Resources 52,
no. 3:800–825.
Olivetti, Claudia, and M. Daniele Paserman. 2015. In the name of the son
(and the daughter): Intergenerational mobility in the United States,
1850–1940. American Economic Review 105, no. 8:1–31.
Rothstein, Jesse. 2019. Inequality of educational opportunity? Schools as
mediators of the intergenerational transmission of income. Journal of La-
bor Economics 37, no. S1:S85–S123.
Ruggles, Steven, Sarah Flood, Ronald Goeken, Josiah Grover, Erin Meyer,
Jose Pacas, and Matthew Sobek. 2019. Integrated Public Use Microdata
Series. Version 9.0. Minneapolis: Minnesota Population Center.
Saavedra, Martin, and Tate Twinam. 2020. A machine learning approach to
improving occupational income scores. Explorations in Economic History
75:101304.
Solon, Gary. 1992. Intergenerational income mobility in the United States.
American Economic Review 82, no. 3:393–408.
Tolbert, Charles M., and Molly Sizer. 1996. US commuting zones and labor
market areas: A 1990 update. Economic Research Service Staff Paper
no. 11614.
Ward, Zachary. 2020. Intergenerational mobility in American history: Ac-
counting for race and measurement error. Working paper.

You might also like