Intergenerational Mobility in the Fourth Industrial Revolution

Thor Berger Research Institute of Industrial Economics, Stockholm Oxford Martin School, University of Oxford Department of Economic History & Centre for Economic Demography, Lund University

Per Engzell Nuﬃeld College, University of Oxford Leverhulme Centre for Demographic Science, University of Oxford Swedish Institute for Social Research (SOFI), Stockholm University

January 2020

Abstract

The maturation of industrial society has long been seen as an engine of occupational upgrading and opportunity. Following the rise of the factory, the assembly line, and the oﬃce computer, some claim that we are now entering a fourth industrial revolution where autonomous systems are transforming the nature of work. What are the consequences of this transformation for intergenerational income mobility?

Examining variation across 722 U.S. labor markets, we find that intergenerational persistence is higher in areas heavily exposed to industrial automation. These effects are rooted in childhood experiences and concentrated among men from disadvantaged homes. Unequal labor relations appear to exacerbate the association, while affordable access to college ameliorates it. The received view of industrial change as an engine of mobility should be revised to consider the institutional context of automation.

1 Introduction

Scholars since at least Parsons (1951) have emphasized the maturation of industrial society as an engine of occupational upgrading and intergenerational mobility (Blau and Duncan, 1967; Lipset and Bendix, 1959; Treiman, 1970). As structural transformation kept children from following in their parents’ footsteps, increasingly universal modes of organization would ensure the dispersal of new opportunities in the population. While this so-called “liberal theory” of industrialization never enjoyed undivided support (Goldthorpe, 1960; Hout, 1989; Jackson and Grusky, 2018), now more than ever it is timely to revisit some of its tenets. In the past few decades, increased automation of jobs has displaced workers and exacerbated wage inequality as less-educated men in particular saw their advantages in the labor market deteriorate. Many observers believe that we are living through a “fourth industrial revolution” in which a wide range of automation technologies is transforming the nature of work to the same extent as the rise of the factory, the assembly line, and computer technologies did in the past (Frey, 2019).

Arguably, the readiness of societies to endure such transformations depends on whether the next generation can expect to be better off than the last. If structural upheaval harms those currently in the labor force but ultimately leaves their children with better prospects—as the liberal theory would have it—then concerns may be overblown. While we lack systematic evidence on this issue, work across a range of fields suggests that we have reason to be less than cheerful. Economists have documented the pervasive impact of automation technologies on employment and earnings (Acemoglu and Restrepo, 2019). Important qualitative work testifies to the wide-reaching consequences that technological job loss has for affected workers, families, and communities (e.g., Goldstein, 2017). There is also a rigorous impact evaluation literature studying the consequences of creative destruction for the social-psychological development and life-course attainment of children to displaced workers (Brand, 2015). Yet, we lack a comprehensive account of what these developments mean for intergenerational mobility in the first decades of the twenty-first century.

Given this gap in the literature, one might expect that social stratiﬁcation scholars would

2 be scrambling to inform the debate. Instead, recent work on intergenerational mobility has tended to revisit age-old questions such as the role of past educational expansion in explaining historical mobility trends (Bloome et al., 2018; Breen and Müller, 2020; Pfeﬀer and Hertel, 2015). Important as such questions are, this means that we have failed to contribute many of the basic empirical facts that could help move the debate forward. This is not without reason, as addressing the consequences of ongoing structural transformation remains a methodological challenge. With few exceptions, research into the nexus between industrial change and mobility has been inductive, making inferences based on national-level patterns and trends (Breen and Jonsson, 2005; Ganzeboom et al., 1991; Grusky and DiPrete, 1990; Treiman and Ganzeboom, 2000). In contrast, nascent technological shifts may be too recent to have left an imprint at these aggregate levels. In this study, we take on this challenge by adopting a meso-comparative approach (Grusky, 1983; VanHeuvelen, 2018b), studying cross-sectional variation across local labor markets that diﬀer in industrial composition.

Our analytical strategy links local mobility outcomes to differences in the exposure to automation across 722 U.S. commuting zones: ecologically meaningful units that span the mainland (Tolbert and Sizer, 1996). For each commuting zone, we measure the exposure to automation by leveraging historical differences in industrial specialization combined with data on the adoption of a key automation technology—industrial robots—across U.S. industries. With the exception of historical work (Knigge, Maas, and van Leeuwen, 2014), intergenerational stratification research has rarely observed industrial change at the local community level as we do. We pair these data with data on income attainment for children born in the early 1980s focusing both on relative mobility as well as upward mobility out of the bottom of the distribution (Chetty, Hendren, Kline, et al.,

2014). This local-level analysis supplies us with considerable variation and allows for a rich set of covariates. That we construct our automation measure based on initial industrial composition also allows us to compare commuting zones with the same initial potential for automation, but where local institutional conditions diﬀer. The principal ﬁnding is that a higher exposure to automation erodes chances for upward mobility,

3 thereby perpetuating the transmission of economic status across generations. To understand what explains these results, we distinguish two mechanisms: diminished labor market prospects for entering cohorts and developmental consequences of community job loss that manifest themselves earlier in life. In fact, we show that mobility deficits associated with automation appear already in children’s educational attainment and increase with the proportion of childhood spent in an area more exposed to automation. This allows us to rule out that labor market prospects alone drive the relationship and suggests that developmental factors are at play. We also show that these effects are largely concentrated among sons rather than daughters, while patterns by race are more complex: Blacks appear less disadvantaged by automation but part of this is explained by their lower mobility chances to begin with. Taken together, these results support a long-held view of local job loss as a trauma with community-wide ramifications (Conger and Elder Jr, 1994; Jahoda et al., [1933]1971; Wilson, 1996). Our results paint a stark contrast to the liberal theory of industrialization and its sanguine view of the consequences of structural transformation. How come, then, that the liberal view appears less relevant for understanding our time than industrial revolutions of the past? Clearly, earlier structural shifts occurred against a backdrop of educational expansion and rapid economic growth, conditions largely absent today. But this observation merely amounts to restating the question. Some might argue that the answer lies in the nature of technology itself: for example, in that current technologies are mainly labor displacing, not labor enhancing as was the case for much of the twentieth century

(Frey, 2019). We emphasize instead that the distributional and intergenerational consequences of technology adoption are a political choice and depend on institutional arrangements. In this sense, the liberal theory of industrialization was always incomplete (Hout and DiPrete, 2006). To substantiate our interpretation, we show that the negative consequences of automation are largely concentrated in areas with high top income shares, declining labor unions, and lack of access to aﬀordable college education.

This article advances the literature on several fronts. Social mobility research has long been uneasy with the industrialization thesis but failed to oﬀer a compelling alternative. Instead, recent

4 mobility research seems to have abandoned its former interest in industrial change in favor of other processes such as educational expansion (Bloome et al., 2018; Breen and Müller, 2020; Pfeﬀer and Hertel, 2015). Regrettably, this means that current labor market transformations have become the nearly exclusive domain of economists (Acemoglu and Restrepo, 2019; Autor, 2015). This work establishes several useful facts, but has failed to take sociological considerations seriously. Family and community consequences of job loss are the subject of a separate literature, but often with a focus on case studies or natural experiments (Brand, 2015), meaning that there is a dearth of representative evidence. We unite these various strands of literature to provide national-level data on this crucial issue. Ultimately, our work recommends a conditional view where the mobility consequences of automation are contingent on political context. Future work should therefore extend this research to other institutional settings than the U.S., and to industrial revolutions of the past.

Background

The coming of the fourth industrial revolution

Over the past decades, labor markets in the U.S. and other wealthy countries have seen a “hollowing-out” with employment growth concentrated at both ends of the income or education spectrum, while formerly well-paid jobs for the middle class have disappeared (Autor et al., 2003;

DiPrete, 1993; Morris and Western, 1999; Neckerman and Torche, 2007). Although there is scholarly disagreement about the relative contribution of factors such as globalization, industry deregulation, or union decline, there is ample evidence that one important driver is the increased substitution for human labor of digital technology in repetitive and ruled-based routine work (Jacobs and Dirlam, 2016; Kristal, 2013; Powell and Snellman, 2004). In particular, the spread of industrial robots has led to increased unemployment and reductions in wages and labor force participation

(Acemoglu and Restrepo, 2019; Autor, 2015). These changes have been heralded as a “fourth industrial revolution” following the adoption of steam power in the nineteenth century, the rise of the assembly line in the twentieth century,

5 and recent decades’ rapid computerization—with the present innovation consisting in the use of robots and other self-governing systems that communicate in fully automated production chains (Frey, 2019). The manufacturing sector in particular has undergone pervasive automation due to the spread of robots performing a range of tasks such as assembly, material handling, and welding. While the first industrial robot was introduced by General Motors already in the 1960s, the robotization of manufacturing mainly took place in the last decades of the twentieth century, and few robots were in use as late as 1980 (Office of Technology Assessment, 1984). Understanding the role of automation in shaping intergenerational mobility remains a critical challenge in light of recent estimates that a substantial share are likely automatable in the next decades. Frey and Osborne (2017), for example, estimate that it is technologically feasible to automate 47% of U.S. aggregate employment. Far from being of mere economic significance, these transformations have cultural meaning. Automation has eroded the economic prospects of groups who used to enjoy great stability just decades ago. Working-class men in particular were once the backbone of a bustling economy but are increasingly seen, by themselves and others, as economic losers (Gest, 2016; Hochschild, 2018). As observers from Willis (1977) onwards have emphasized, perhaps no other group cultivated an identity so intimately bound up with their employment as white, working-class men (Lamont, 2009). As an object of sociological study, the disappearance of work is a shock that extends beyond the individual or even family to send ripples throughout entire communities (Conger and Elder Jr, 1994;

Jahoda et al., [1933]1971; Wilson, 1996). The workplace is an arena for communal engagement, the destruction of which can have negative consequences for social capital and collective eﬃcacy (Brand, 2015). Funding available to schools and other commons may diminish with the erosion of the local tax base, or as funds need to be diverted for remediation programs (Gassman-Pines et al., 2015). Above all, the loss of industries can be a collective trauma that leaves local communities struggling to regain their sense of collective identity and purpose (Goldstein, 2017).

The accelerated pace of robot adoption in the 1990s and 2000s and its attendant consequences in the labor market have been extensively studied by economists (Acemoglu and Restrepo, 2019). This

6 work establishes beyond doubt the persistent eﬀects of automation for earnings and employment

among low-educated workers. At the same time, it remains incomplete in at least two respects. First, an intergenerational perspective has been conspicuously absent. We therefore do not know whether the detrimental eﬀects of automation reﬂect a temporary loss, or a more enduring blow

to aﬀected families. Secondly, economists have tended to treat technology as an exogenous force, independent of political dynamics. But ultimately, the distributional eﬀects of automation depend on the institutional landscape surrounding its implementation (DiPrete, 2002; DiPrete and McManus,

1996). In the following sections, we remedy each of these shortcomings before putting our account to empirical test. Doing so, we build on and extend the conventional narrative in social stratiﬁcation research—the liberal theory of industrialization—to understand how automation may

shape mobility chances in the twenty-ﬁrst century.

Intergenerational mobility and industrial change

Historically, industrial change was seen as a driver of occupational upgrading and growing

opportunity (Lipset and Bendix, 1959; Treiman, 1970). Scholars of social stratiﬁcation have long been uneasy with the simplistic character of this predominant view (Goldthorpe, 1960; Hout, 1989; Treiman, 1970), and its failure to account for actually observed patterns (Grusky and Hauser,

1984; Hazelrigg and Garnier, 1976; Knigge, Maas, and van Leeuwen, 2014). Yet, it remains the primary foil for social mobility research and so we outline it here before showing how an institutional view lends itself to a more plausible account. The so-called OED triangle of origins, education, and destinations has guided much of this discussion in the past (Blau and Duncan, 1967; Grusky, 1983). This heuristic, depicted in Figure 1, separates the association between social origins O and destinations D into a direct eﬀect, and an indirect one mediated by education, E.

Social transformation can modify the strength of transmission via any of the paths OE, ED, or OD. While past studies mainly conceptualize social status, whether at origin or destination, in terms of occupational status (Blau and Duncan, 1967) or social class (Erikson and Goldthorpe, 1992), the

OED model is equally relevant for the case of income transmission that we study here (Bloome

7 et al., 2018; Harding and Munk, 2019).

[Figure 1 about here.]

The industrialization thesis holds that with technological change, allocation to social positions will follow increasingly meritocratic criteria (Treiman, 1970). As children are pushed away from

their parents’ footsteps, labor market and state institutions are to ensure access to new opportunities. This prediction comes about through several steps. First, the functional needs of complex societies will make recruitment depend more and more on formal qualiﬁcations, strengthening the association

ED. As schooling expands to accommodate these demands, access is equalized and the association OE weakens. At the same time, there is a compositional effect whereby the direct association OD tends to be less pronounced at higher levels of qualification (Hout, 1988; Torche, 2011). While there is by now little question that mobility increased during the last century (Breen, 2004), the difficulty is to pin this change to technological change as such. Improvements in social class mobility have been far from uniform, and by most accounts have taken their fullest form in societies that offset disequalizing effects of the market (Breen, 2004; Erikson and Goldthorpe, 1992). Research on income mobility finds similar results, across and within countries (Blanden, 2013; Chetty, Hendren, Kline, et al., 2014).

Thus, while historical trends seem to favor the liberal narrative, geographic differences lend greater weight to other institutional factors that may have overlapped with industrialization in time: educational expansion, the rise of welfare state institutions, and widely shared redistribution of economic growth (Hout and DiPrete, 2006). This conclusion is echoed by research studying the nineteenth century’s industrial revolution. In the Netherlands, Knigge et al. (2014a,b) find that while it coincided temporally with increased mobility, industrialization at the local level actually brought higher intergenerational persistence, contrary to the liberal view. Instead, what eventually appears to have improved mobility is the growth of modern infrastructure and public goods—railways, postal system, and so on (cf. Pérez, 2017). In sum, while there is agreement that mobility increased over the twentieth century it remains debatable whether this was driven by industrial change as such. The current industrial revolution offers a compelling test case for the

8 industrialization thesis, as it encompasses many elements of previous transformations but without the background of expanding state institutions and shared economic growth.

Toward an institutionally informed account

As the language of a “fourth industrial revolution” suggests, current structural transformations bear many of the hallmarks of similar changes in the past. Despite misgivings, previous industrial revolutions eventually brought with them increased prosperity, urbanization, and—while the precise causal pathways remain disputed—a decreasing persistence of status (Breen, 2004). Why would the liberal theory be any less relevant in understanding our present age? The problem with the industrialization thesis is perhaps not so much that it is wrong as that it is incomplete—by treating technology as an immutable force it ignores the human arrangements and relations that govern its implementation. The liberal theory shares this shortcoming with functionalist theories of stratiﬁcation more broadly (Davis and Moore, 1945; Parsons, 1951), as well as with many economic accounts. The challenge is to understand when and under what condition technological development hinders or helps mobility, and here an analysis that fails to account for institutional factors is bound to be misleading. Among economists, skill-biased technological change has emerged as the favored theory to account for the surge in inequality from the 1970s (Autor, 2014). Meanwhile, sociological work stresses that these changes do not take place in a political vacuum (DiPrete and McManus, 1996; Horowitz, 2018; Kristal, 2013), and studies of the interplay between automation and organizational inequality highlight that there is no deterministic relationship between the two (Jürgens et al., 1993; Krzywdzinski, 2017). Instead, industry is a contested arena where politics shape not only how new production technologies are deployed, but to whose beneﬁt (Busemeyer and Iversen, 2012;

Gallie, 2017). In the U.S., worker representation and collective bargaining have been in steady decline since the 1970s, but with variation at the local and industry level (Brady, Baker, et al., 2013; Wallace et al., 2009; Western and Rosenfeld, 2011). Accompanying this shift is a broader pattern of norm change (Kim et al., 2015; Mizruchi, 2013; VanHeuvelen, 2018a) and a plethora

9 of measures favoring the accumulation of corporate proﬁt and marginalization of workers (Jacobs and Myers, 2014; Tope and Jacobs, 2009; Volscho and Kelly, 2012). Another key factor to consider is the supply of education to keep pace with the changing nature of skill demand (Goldin and Katz, 2009). Since the early 1980s, post-secondary education in the

U.S. has followed a worrying trend where men’s attainment remained stagnant while that of women increased (Mitnik, Cumberworth, et al., 2016; Roksa et al., 2007). This is not due to a decrease in the demand for education—that is, the proportion aspiring for a college education—which has steadily risen (Reynolds and Johnson, 2011). Instead, research points to diﬀerences in academic preparedness and family vulnerability which systematically work to the disadvantage of boys (Buchmann and DiPrete, 2006; Entwisle et al., 2007; Reardon, 2011). Meanwhile, soaring education costs and income and wealth concentration entail that low-income families are increasingly reliant on governmental and other support schemes to be able to aﬀord college (Jerrim et al., 2015; Mayer, 2001, 2010).

This overview suggests why, this time around, technological change may have the effect of undermining rather than empowering workers and their children. In our analysis, we attend to these concerns by studying the interplay between automation on the one hand and the distribution of bargaining power and access to educational opportunity on the other. We focus on local labor markets within the U.S., which exhibit significant variation in economic and social makeup. To the extent that we find meaningful institutional variation here, it would underscore the role of broader national economic and social institutions in ameliorating the harm of structural transformation. For example, European countries have seen rates of robot adoption surpassing that of the U.S., but without the same detrimental effects on employment, wages, and quality of work (Dauth et al.,

2018; Graetz and Michaels, 2018). Our analysis can therefore be seen as a ﬁrst step in advancing a comparative research agenda.

10 Hypotheses

Drawing on the model sketched above and previous work, we derive a set of hypotheses. The prevailing discourse among economists and policy-makers construes technological change foremost as an increase in the returns to human capital, that is, a strengthening of the ED link in the OED triangle (Figure 1). Empirical work lends support to this view: starting in the 1980s, there has been a steep growth in the earnings premium associated with post-secondary education (Autor, 2014). Absent educational expansion that drove mobility in the past, however, it is hard to envision a corresponding weakening of either the OE or OD link. Holding everything constant, and simply based on the increasing returns to education, therefore, we would expect the persistence of income from one generation to the next to rise.

Hypothesis 1.—In areas more exposed to automation, income levels persist more strongly from parent to child.

In principle, it is possible to expect higher persistence to be driven both by the top and bottom of the parental income distribution. As for the top, decompositions of the rising skill premium have found it to be concentrated among occupations such as engineers, engineering managers, and computer and systems analysts (Y. Liu and Grusky, 2013)—prime examples of “microclasses” that are also highly inheritable (Jonsson et al., 2009). As for the bottom, it is likely that automation will push children of manual workers into more precarious employment, such as that found in low-skilled service jobs. In practice, previous work on automation ﬁnds detrimental eﬀects on wages throughout the earnings distribution (Acemoglu and Restrepo, 2019) which makes the rich-getting-richer scenario seem less likely. This leads us to think that, if anything, stickiness at the bottom will be the dominant driver.

Hypothesis 2.—Higher intergenerational persistence in areas exposed to automation is driven by lower upward mobility.

The popular perception of automation is that it struck hardest among demographics whose employment would have oﬀered them life-long security and growing earnings prospects a generation ago—foremost, non-Hispanic white men. Without doubt, the casualties of automation are

11 male-dominated industries, while the rise of women in education leads us to expect that they

would be more shielded from structural change at any rate (Buchmann and DiPrete, 2006; DiPrete and Nonnemaker, 1997). The expectation with regard to race is more complicated. The loss of industrial jobs is a prominent theme in the work of Wilson (1987, 1996) on urban Black poverty and

the “spatial mismatch” hypothesis became a popular explanation for racial disparities in the 1990s (Mouw, 2000). As Cherlin (2014, p. 7-9) also points out, a larger proportion of Black than white men were working in manufacturing at the peak of industrial employment. However, our earliest

income measurements are from the late 1990s, after the suburban exodus of industrial employers. Therefore, and in keeping with recent qualitative work (Edin et al., 2019; Gest, 2016; Hochschild, 2018), we predict the following.

Hypothesis 3.—The link between automation and intergenerational persistence is most pronounced among non-Hispanic white men.

So far, the mechanisms we have discussed refer to shifting fortunes in the labor market.

However, when technological disruption strikes a community it may affect children’s life course already from an earlier age. We test this in two ways. First, by comparing children who through family moves spent a different amount of their childhood in a given area. Second, by looking at educational outcomes directly. If detrimental effects of automation on mobility work solely through opportunities in the labor market, we would expect similar effects for those who live in a given area when entering adulthood, regardless of where they grew up. On the other hand, if adverse effects of automation are rooted in childhood experiences, it should be stronger the larger the proportion of childhood spent in a given area, and manifest itself before entry into the labor market. Hypothesis 4.—The link between automation and intergenerational persistence is stronger the earlier the onset of exposure, and visible for both education and income.

In principle, workers can exert considerable power over how technology is wielded in the labor market but this is contingent on local politics and norms (Fernandez, 2001; Krzywdzinski, 2017).

While structural transformation may create opportunities for mobility, whether that mobility is upward or downward depends on the alternatives that become available instead. Studies of European

12 countries ﬁnd that industrial robots do not necessarily displace workers, who often remain with

the same employer but shift occupational roles (Dauth et al., 2018). In our analysis, we measure commuting-zone variation in the top 1% income to capture norms and political measures that favor corporate enrichment at the expense of communities and workers (Kim et al., 2015; Volscho

and Kelly, 2012). In addition, we also inspect variation across states who experienced diﬀerential decline the share of workers covered by unions prior to the takeoﬀ of industrial automation. A robust body of work documents that unionization is associated with outcomes such as lower poverty rates,

higher average worker earnings, greater job security, and lower elite compensation (Brady, Blome, et al., 2016), and it is reasonable to think that unions may inﬂuence to what extent technology crowds out labor as well.

Hypothesis 5.—The link between automation and intergenerational persistence is stronger in areas with higher top income shares, and in states that experienced larger union decline.

Another important aspect to consider is the availability of institutional remediation, foremost, educational opportunities. The weaker eﬀect of class origins on status attainment at higher levels of education is well known (Hout, 1988; Torche, 2011), although recent research suggests that a college degree may act as less of a leveler for income than for occupational access (Karlson, 2019;

Zhou, 2019). Nevertheless, when former industrial jobs disappear, getting a foot in the door of more skilled sectors is a vast improvement over the alternative: to join the ranks of a growing “new service proletariat” (Esping-Andersen, 1993) or lumpenproletariat (Jackson and Grusky, 2018). Beyond serving as an escalator for individual mobility, colleges may also spur local transformation by attracting public revenue, private investments, and stemming population decline (Cermeño, 2018; S. Liu, 2015). We capture these separate mechanisms by examining both the number of colleges per capita, which may capture agglomeration effects for the local economy, and the average college tuition, which may capture the accessibility of higher education for students from disadvantaged homes. We expect that both the supply of college education and its affordability may moderate the effects of automation on mobility. Hypothesis 6.—The link between automation and intergenerational persistence is weaker the

13 greater the number of colleges in an area, but stronger the more costly their tuition rates.

Data and measures

Figure 2 displays the rise of industrial robots from 1982 to 2011, measured as the total number of units in operation U.S.-wide. The U.S. remains a relative laggard in the adoption of industrial robots relative to its Asian and European counterparts—for example, while the U.S. stock of robots per thousand workers hovered between 1 and 2 in the first decade of the new millennium, it rose from 3 to 5 in Germany during the same period (Acemoglu and Restrepo, 2019). Whereas the adoption of robots in the U.S. is still relatively limited, it has been heavily concentrated to certain sectors. We use this sectoral variation together with local industrial composition to study variation at the level of U.S. commuting zones—rural and urban labor markets delineated based on commuting patterns (Tolbert and Sizer, 1996). Combining differences in the adoption of robots across industries with initial differences in industrial specialization uncovers significant local variation in the susceptibility to automation that allows us to study how automation affected intergenerational mobility for children born in the early 1980s.

[Figure 2 about here.]

Several studies have noted the strengths that accrue to this type of community-level analysis (Knigge, Maas, and van Leeuwen, 2014; Moller et al., 2009). Counties or local labor markets are “social microcosms” (Moller et al., 2009) with vast differences in internal social and economic makeup. Not only do they offer rich variation in exposures and outcomes; they also provide a window on changes at the level where many of the theoretical mechanisms are expected to occur (Grusky, 1983). By drawing on administrative data collected with uniform procedures and aims, our analysis is less susceptible to the sampling, definition, or measurement artefacts that inevitably plague cross-country research. We now go on to describe the measurement of intergenerational income mobility in our data, and thereafter the exposure to automation.

14 Intergenerational mobility

Intergenerational mobility data come from the Equality of Opportunity Project (Chetty, Hendren, Kline, et al., 2014), which has estimated a range of mobility metrics using individual federal tax records from the Internal Revenue Service (IRS). Most mobility metrics pertain to cohorts born

in 1980–1982 and their parents, with children assigned to the commuting zone where they resided at age 16. Child income is measured as mean family income in 2011–2012 when children are approximately 30 years old, while parent income is measured by mean family income between

1996–2000 (Figure 2). In some analyses, we also inspect children’s individual income separately, measured in the same years. While family income is the more encompassing measure of living standards, individual incomes are useful to disentangle the separate implications for men and

women’s labor market prospects. The literature on economic mobility has generated a variety of measures, of which the most common is the intergenerational elasticity of incomes. This statistic, which simply reﬂects the

derivative of expected log child income with respect to log parent income is usually estimated using ordinary least squares, where the elasticity becomes the regression coeﬃcient (Mitnik, Bryant, et al., 2019). Sensitivity to marginal distributions and the ages at which income is measured has

led recent research to prefer the rank-order correlation (Bloome et al., 2018; Chetty, Hendren, Kline, et al., 2014), which represents a similar derivative where each variable has instead been transformed to percentile ranks. Letting intercept and slope vary by commuting zone j, we have:

C P Yij = αj + βj Yi j + εij,

C P where Yi j and Yij represent child and parent income ranks in the national income distribution. The main parameter of interest here is βj which represents how strongly income rank is transmitted from parent to child. We use the term “rank correlation” to refer to this parameter, although national ranks may deviate from a uniform distribution at the commuting-zone level.

Rank correlations are conceptually close to the path models that have informed much previous research on the industrialization thesis (Blau and Duncan, 1967; Grusky, 1983; Knigge, Maas,

15 Van Leeuwen, et al., 2014). However, they are uninformative about whether mobility is driven by

persistence at the bottom, top, or somewhere inbetween. As the impact of automation has mainly been felt in middle-income jobs (Acemoglu and Restrepo, 2019; Autor et al., 2003; Graetz and Michaels, 2017), a single parameter may not be enough to capture the complexity of mobility

patterns. In a next step we therefore inspect transition probabilities for the full 5 × 5 mobility table between quintiles of parent and child income. That is, we deﬁne a set of measures:

C P Ppq,j = Pr(Q = q | Q = p), p, q = 1,..., 5, i j i j ∀ where QC and QP represent child and parent income quintiles, and j indexes the commuting zone as before. This, in turn, allows us to define more specific mobility patterns such as “rags-to-riches” mobility [Pr(QC = 5 | QP = 1)], poverty persistence [Pr(QC = QP | QP = 1)], elite persistence [Pr(QC = QP | QP = 5)], or downward mobility from the middle class [Pr(QC < QP | QP ∈ {2, 3, 4})]. In practice, we find that much of the consequences of automation are located in the bottom half of the distribution. A useful summary measure is therefore what Chetty, Hendren, Kline, et al. (2014) term “absolute upward mobility,” defined as the income rank expectation for a child born into the bottom half of the distribution—which, given the approximate linearity of the rank-rank relationship, is equivalent to the predicted rank of children born to parents at the 25th percentile:

C P Aj = E(Yij | Yij < 50) = αj + 25βj .

This measure also offers a convenient way to look at racial differences which are not well captured by the rank correlation (Chetty, Hendren, Jones, et al., 2018). We use upward mobility for the 1980–1982 birth cohorts when studying the whole population, but expand the window to 1978–1983 birth cohorts to overcome small cell sizes when studying mobility separately by race, and 1980–1986 birth cohorts for the age-specific effects that we describe next. To test our hypothesis that the link between automation and income attainment is rooted in childhood events rather than mere labor market prospects, we use estimates from a specification comparing children who move at different ages (Chetty and Hendren, 2018). These estimates

16 net out time-constant variation across commuting zones and use only variation in the length of

childhood spent in a given area. The intuition is as follows. Consider all children of a given cohort who reside in commuting zone j at the time they turn 16. Some of those children will have lived their whole life there, others have moved there with their families at any time between age 0 and 16.

The fixed-effects specification discards with the population of permanent residents and compares only the outcomes of children who have moved to the area, depending on when they arrived. To the extent that earlier arrivals achieve a lower level of income we attribute this to experiences before the age of 16, since local labor market prospects are the same for all children regardless of when they arrived. These estimates are scaled to reflect the expected percentage decrease (or increase) in adult income from spending one additional year of childhood in a given commuting zone. Because of the later birth of these cohorts (1980–1986), income is here measured at age 26. Finally, we look at the probability of completing various educational transitions conditional on childhood income: high school, at least some college, and a four-year college degree. Information on educational attainment is from decennial Censuses or the 2005–2015 American Community Survey (ACS) that have been linked to the IRS data underlying the income mobility estimates (Chetty, Friedman, et al., 2018). Educational attainment is as reported by the child, with priority given to more recent ACS data if available, and excluding all respondents younger than age 24 at the time of questionnaire completion. High school degree holders include those with a General Educational Development (GED) certificate, the next level includes those who report “at least some college credit” or higher, while four-year college completion is defined as having at least a Bachelor’s degree.

Exposure to automation

A central challenge in identifying the impacts of technological change is the a lack of data on the spread of new technologies, which has led most studies to adopt an indirect task-approach (Autor et al., 2003; Denice and Rosenfeld, 2018). In contrast to prior literature, we instead obtain data from the International Federation of Robotics (IFR) that provide industry-level data on the use of

17 industrial robots in a number of countries including the United States.1 Speciﬁcally, we use U.S. data

from the IFR on the robot stock in 13 manufacturing industries and six broad non-manufacturing sectors.2 Our interest is in the local exposure to automation, however, which requires mapping the aggregate industry-level data from the IFR to the level of commuting zones through the distribution of local employment across industries. To measure local industrial composition, we rely on individual-level data from the 1980 Census that includes information about individuals’ place of residence and employment by industry.3 By collapsing the individual-level Census data, we can then simply calculate employment shares by industry for each individual commuting zone, which in turn can be linked to the IFR data on robot use in each industry listed above.

Formally, after pairing the industry-level data from the IFR and the commuting-zone level industrial shares we deﬁne the exposure to automation between 1980 and 2011 for each commuting zone j as follows:

k ! Õ Robots E xposure = Industryk × US,2011 , j j,1980 k k∈K WorkersUS,2011

k where Industryj,1980 corresponds to the share of a commuting zone’s workers employed in industry 1The IFR follows the International Organization for Standardization in defining an industrial robot as an “automatically controlled, reprogrammable multipurpose manipulator programmable in three or more axes” (see: https://ifr.org/industrial-robots). 2In manufacturing, there are consistent data on the use of robots for 13 industries: food and beverages; textiles; wood and furniture; paper; plastic and chemicals; glass and ceramics; basic metals; metal products; metal machinery; electronics; automotive; other vehicles; and other manufacturing industries. Outside of manufacturing, we construct the data for the use of robots in six broad industries: agriculture, forestry, and fishing; mining; utilities; construction; education, research, and development; and other non-manufacturing industries (e.g. services and entertainment). 3To preserve confidentiality, the Census reports individuals’ place of residence for “county groups” that in some cases span multiple commuting zones. In cases where individuals are enumerated in such county groups, we use a probabilistic-weighting approach to assign them to commuting zones (Dorn, 2009).

18 Robotsk k k in 1980 computed from the 1980 Census, and US,2011/WorkersUS,2011 denotes the national level of robot usage per thousand workers in that industry in 2011 based on data from the IFR and the

2011 ACS. Intuitively, this measure reflects differences in exposure to robots across commuting zones driven by variation in automation across U.S. industries in 2011 and initial differences in industry specialization across commuting zones in 1980, which predates the measurement of both child and parent income in all our main mobility measures (Figure 2). In other words, a higher level of exposure to automation is thus driven by local specialization in industries that experienced a subsequent greater penetration of robots. To reduce the skewed distribution of robot exposure across commuting zones, we enter this variable in logged form and standardize it to have a mean of zero and standard deviation of one throughout the empirical analysis. Figure 3 displays the geographical distribution of our baseline exposure measure, documenting the significant spatial variation in exposure to automation across U.S. commuting zones and that the highest levels of exposure are heavily concentrated to the Rust Belt, as we would expect.

[Figure 3 about here.]

To be clear, the exposure measure reflects the potential for automation given the industrial composition of a commuting zone in 1980, not the actual number of robots installed locally. This is integral to our design, as ultimately what we are interested in is how technological frontiers and institutions interact. We are thereby able to compare commuting zones with the same automation potential in 1980 but where other local conditions differ. Moreover, the fact that we use this relatively indirect proxy means that it will almost certainly be less confounded by unobserved factors that may cause both automation and intergenerational mobility. Another important caveat is that detailed data on the use of industrial robots is limited throughout the 1980s (see Figure 2). However, we know that the number of robots by 1982 was as low as 6,300 nationwide (Office of Technology Assessment, 1984), which allows us to establish that robot use in the early 1980s was negligible. This means that our measure of robot exposure can be interpreted as capturing over-time change in exposure. To corroborate this interpretation, we also calculate the actual change in exposure until 2011 using the starting points of 1982 (Office of Technology

19 Assessment, 1984) and 1993, respectively—the latter being the ﬁrst year in which data are available

from the IFR. Because robot use is not disaggregated by industry for these years, we allocate robots to industries based on the observed industry shares in the early 2000s. Reassuringly, using these alternative data sources to estimate changes in robot exposure between 1982–2011 and 1993–2011

yields nearly identical results and they are both highly (r > 0.99) correlated with our baseline measure.

Control variables and moderators

Automation is, however, only one of the major shocks that have hit local labor markets and an important question is to what extent it is conflated with other variables. We address this by controlling for a range of other demographic and structural characteristics. As basic commuting zone controls, we include the log of population size, the log of average household income, and whether the commuting zone intersects a metropolitan statistical area. To adjust for local demographic composition we further control for the share of Black residents, the share of population in four different age bins (below 15, age 15–24, 25–64, and 65 or above), and the share female. In a further step we add a control for the share of college educated. Thereafter we control for initial differences in employment composition, measured as the share of workers employed in manufacturing in 1980. Finally, in a last step we include controls for Census region and division as a set of fixed effects (U.S. Census Bureau, n.d.). To avoid conditioning on posttreatment variables, all these measures are observed at baseline in the 1980 Census.

The expected relationship between industrial change and mobility is contingent on institutional arrangements as we have argued above. We use two measures of local political arrangements to test this: the share of income accruing to the top 1% of earners in a commuting zone, and the statewide trend in union decline leading up to the takeoﬀ of industrial automation. The top income share is calculated on the distribution of family incomes in the parent generation as measured in 1996–2000 Chetty, Hendren, Kline, et al. (2014). Ideally, we would have earlier measures which are not available in our data. Note, however, that as late as in 1996 the trend in robot adoption

20 had yet to take oﬀ (see Figure 2), and top income shares in our data correlate weakly with both

automation and upward mobility (r = −0.05 and r = 0.07 respectively in our weighted regressions; cf. Appendix Table 7). As a measure of power resources on the shop floor, we use statewide declines in unionization for the years leading up to the takeoff of industrial automation (Brady, Baker, et al., 2013; Western and Rosenfeld, 2011). We gather data on union decline from the Union Membership and Coverage Database (UMCD), a resource first compiled by Hirsch et al. (2001) and maintained annually since.

The UMCD reports union coverage density at the state level annually since 1977 when Current Population Survey (CPS) coverage begins. Although the UMCD contains data on detailed industries and occupations from 1983 onwards, we focus on the earlier period to capture longer-term political trends that predate the rise of industrial automation. We calculate a measure of annual statewide decline in coverage throughout the first ten years for which data are available (1977–1986). To gauge the supply and accessibility of higher education, we measure the number of colleges per capita and the average college tuition at the commuting-zone level. These measures are constructed using data from the Integrated Postsecondary Education Data System (IPEDS) by Chetty, Hendren, Kline, et al. (2014). Colleges per capita is defined as the number of Title IV, degree-granting colleges in the year 2000, divided by total commuting zone population. Average tuition is a measure of the “sticker price for in-state, full-time undergraduates” for all colleges included in the first instance, weighted by enrollment (Chetty, Hendren, Kline, et al., 2014, Appendix H). As data on colleges and tuition only exist for a subset of commuting zones (574 and 570, respectively), the number in these regressions is smaller than the total universe of commuting zones.

Results

To assess Hypothesis 1 of an inverse association between automation and intergenerational mobility, we plot the bivariate association between commuting-zone level exposure to automation and the intergenerational rank correlation βj in Figure 4. Recall that the rank correlation is a measure of persistence, so higher values indicate that mobility is lower. The correlation between automation

21 and persistence in income rank is sizeable and of the expected sign: 0.41, or 0.35 when weighted by population size. This association is comparable to that of several key mobility correlates (Chetty, Hendren, Kline, et al., 2014). While Figure 4 conﬁrms that income levels persist more strongly in areas heavily exposed to automation, it does not tell us where in the distribution that persistence is most pronounced. To answer that question, we turn to inspecting conditional transition probabilities throughout the full mobility table among parent and child income quintiles.

[Figure 4 about here.]

Our Hypothesis 2 states that automation of jobs translates to a more marked persistence in the bottom of the distribution. To test this, we fit a linear model predicting each conditional probability in the 5 × 5 mobility matrix between quintiles of parent and child income in Figure 5. The linear probability model is tractable and has the added benefit that associations can be interpreted as average marginal effects (Mood, 2010). Here we standardize automation so that coefficients contrast the difference between areas one standard deviation below and above the mean of exposure. There is clearly higher persistence in the bottom two income quintiles in areas more exposed to automation, as well as lower rates of entry into the top from all quintiles Q1–Q4. Interestingly, there are also signs of disproportional downward mobility from the three middle quintiles Q2–Q4 in areas with higher automation exposure, indicating a “downslide” from the middle class. In contrast, there is no association between automation and mobility for those who grow up in the top income quintile. Thus Hypothesis 2 is supported: the consequences of automation for mobility are most marked in the bottom of the distribution.

[Table 1 about here.]

As a further test of Hypothesis 2, in Table 1 we estimate commuting-zone level regressions using diﬀerent mobility outcomes—the intergenerational rank correlation, absolute upward mobility, and rags-to-riches mobility—with stepwise addition of the control variables described earlier. All regressions are weighted by population size with standard errors clustered at the Census division

22 level. In Panel A, a one standard deviation diﬀerence in exposure to automation is associated with a 0.02 point higher rank correlation. This represents a 6% diﬀerence relative to the mean, or a correlation of 0.35 as shown in Figure 4. The association remains robust to controls for demographics and other commuting zone attributes throughout columns 2–4, but is reduced by control for the initial share of manufacturing employment and Census division. In contrast, results for absolute upward mobility—the expected percentile rank for children born in the bottom half of the distribution—is not diminished by these controls (Panel B). Neither is rags-to-riches mobility, the probability of moving to the top quintile conditional on starting in the bottom (Panel C). In the rest of our analysis, we focus particularly on upward mobility and on models estimated under the full set of controls.

[Figure 5 about here.]

Our next question concerns demographic differences, where Hypothesis 3 states that the consequences of automation for mobility are most pronounced among non-Hispanic white men. Table 2 shows regressions of absolute upward mobility separately by gender and race and this time we measure individual as opposed to household income in the child generation, as described above. As minorities are not represented in sufficient numbers in all commuting zones, the number of observations vary from 550 (Black males) to 722 (whites). The results reveal that effects of automation are more pronounced among men, while the story for race differences is more complex. On the one hand, associations are nominally larger for non-Hispanic and Hispanic white men compared to their Black counterparts. Inspection of the correlation matrix in Appendix Table 8 also reveals a markedly stronger correlation for non-Hispanic white men (−0.51) than Hispanics (−0.30) or Blacks (−0.30). However, this difference should be interpreted in light of the fact that

Blacks as a group experience less upward mobility regardless of circumstances. When viewed in relation to their average mobility levels reported in Table 2, percentage diﬀerences for the three groups are more similar and only slightly higher among non-Hispanic (1.53/49.27 = 3.1%) or

Hispanic white men (1.58/47.54 = 3.3%) as compared to Blacks (1.09/38.89 = 2.8%).

23 [Table 2 about here.]

Automation could hamper income attainment via several mechanisms: through the opportunity structure children face when entering the labor market, or exposure to poor environments earlier in life. In Table 3 we analyze results from the fixed-effects specification described above, where outcomes are only compared among children who moved to an area at different ages. We distinguish between family and individual income for both men and women and by parental income, with Panel A providing estimates for children from households below the median and Panel B for those above the median. Associations reflect the per-year percentage difference in adult income that results from spending a longer part of one’s childhood in a given commuting zone. Table 3 reveals a differential impact of automation by length of childhood exposure for men but not women. For men in Panel A, moving to a commuting zone with a one standard deviation higher automation exposure a year earlier in life results in a 0.12% penalty in adult family income. These effects are smaller among children from better-off homes (Panel B). We should also recall that these estimates adjust for a range of overlapping disadvantages that disproportionately affect lower-income homes. These results support Hypothesis 4: the association between automation and mobility is stronger the earlier the onset of exposure.

[Table 3 about here.]

If disadvantages among exposed children start before labor market entry, this points toward explanations rooted in life-course development, especially perhaps educational attainment. To test this mechanism, in Table 4 we show regressions of the fraction of children born in each commuting zone that attain a high school degree (or GED), some college, and a four-year college degree.

Exposure to automation has detrimental eﬀects on each transition among white men and women, and on college graduation for Black women. Some of the most pronounced eﬀects, however, are found on the college entry of white men. A one standard deviation higher exposure to automation is associated with 1.8 percentage points lower probability of enrolling in college among non-Hispanic white men (Panel B, column 7). We further distinguish this group by childhood income in Table

24 5, inspecting educational attainment at five points in the parental income distribution: the bottom and top 1%, and percentiles 25, 50, and 75. This shows that effects are concentrated to the bottom: the size of the association decreases linearly with each step up the parental income ladder and is no longer significant at the top of the distribution.

[Table 4 about here.]

[Table 5 about here.]

In our analyses so far, we have allowed the relationship between automation and mobility to vary by demographic but not by local context. Our theoretical expectations are that adverse consequences of automation will be more pronounced in areas with an unbalanced distribution of power and limited access to educational opportunities (Hypotheses 5 and 6). We now turn to examining these hypotheses. To proxy for power resources and political norms, we measure the commuting-zone level top 1% income share and statewide declines in unionized labor during the decade leading up to industrial automation. As measures of college access, we use the number of colleges per capita and the (enrollment-weighted) average cost of college tuition. We standardize each of these moderators to have a mean of zero and standard deviation of one, similarly to the measure of automation exposure.

[Table 6 about here.]

In Table 6, we interact automation exposure with each of these institutional variables in isolation as well as in a joint model, with and without commuting-zone level controls. We ﬁnd interactions of the expected sign for all measures except one (colleges per capita), but not all of them equally robust across speciﬁcation. A negative interaction with the top 1% income share is evident only when controlling for other commuting-zone level factors and not when all interactions are included jointly. A negative interaction with union decline is evident in models without controls, with or without other interaction included. The number of colleges per capita shows no interaction with automation exposure but college tuition shows an interaction of the expected sign: the negative

25 mobility consequences of automation are worse where local college tuition is high. This association is only apparent when adjusting for other commuting-zone level covariates, but remains signiﬁcant in a model with all interaction terms entered jointly. In other words, mobility deﬁcits associated with automation are stronger in areas with high top income shares, greater union decline, and more expensive college tuition. Overall, these results are suggestive of a moderation by local political context consistent with Hypotheses 5 and 6.

Discussion

Since the 1980s, increased automation of jobs has displaced workers, reduced average workers’ wages, and exacerbated wage inequality. Industrial upheavals witnessed today have created new winners and losers and rapidly eroded the employment and earnings prospects of groups who used to enjoy great stability. Public debate on these issues often revolves around the question of whether people are left better or worse off than their family a generation ago. Yet, scholarly work on the consequences of automation has largely tended to focus on inequality in the cross-section, and social scientists have failed to contribute many of the empirical facts that could help move the debate forward. Here we have sought to remedy this situation and bring social stratification research into the current era. Our study is the first to link U.S. intergenerational mobility to automation in the twenty-first century. The results are sobering. Examining variation in the exposure to automation across local labor markets, we find a persistent link between automation in the workplace and the chances of low-income children to attain higher incomes in adulthood: a standard deviation higher automation exposure is associated with a 0.9–1.2-point reduction in upward mobility, the percentile rank that a child from the bottom half of the income distribution can expect to attain in adulthood. This corresponds to one tenth of the distance that separates a place like Charlotte from one like San Jose, representing the very bottom and top of the urban mobility hierarchy (Chetty, Hendren,

Kline, et al., 2014). These results are robust to a wide range of statistical confounders, and using sibling comparisons of mobility that address the potential sorting of families across local labor

26 markets. Our results thus confirm that job destruction brought about by technological change can have significant intergenerational repercussions for families’ economic well-being. To explain our results, we distinguished between two mechanisms. One concerns the prospects that children face in the labor market. The other is that automation impacts on children’s life course earlier on by eroding the ability of families to invest in their economic attainment. Two findings point toward the latter explanation. First, mobility deficits associated with automation increase with the proportion of childhood spent in an affected area. To be clear, our findings are rooted in childhood environments: each year of childhood spent in an area more exposed to automation accumulates to lower income in adulthood. Second, mobility deficits manifest themselves already in children’s educational attainment. Thus, in circumstances where high education is increasingly a condition for economic well-being, low-income children are becoming less likely to attain it. Taken together, these results suggest that labor market prospects alone are insufficient to explain the relationship between automation and mobility and point to developmental factors earlier in the life course. These results run counter to the liberal thesis of industrialization that has long remained the primary foil for social mobility research (Jackson and Grusky, 2018; Pfeffer and Hertel, 2015)—that is, the conjecture that technological progress will reduce the transmission of status. Already three decades ago, Hout (1989, p. 7) called this “the most widely tested and least widely supported among the theories in this literature,” and discontent goes back at least another three decades (Goldthorpe,

1960). Yet research has been inconclusive as industrial change historically was confounded by the background conditions of shared economic growth and expanding state institutions. The present-day U.S. provides a strategic test for the industrialization thesis, as a context that lacks these background conditions. Here, the industrialism thesis is squarely rejected: automation in the workplace today instead erodes life chances and perpetuates the transmission of economic status. While our findings fit poorly with the industrialism thesis, they confirm a large body of quantitative and qualitative work outside the area of social stratification. Economists have documented far-reaching consequences of ongoing structural transformations for cross-sectional

27 inequality. But this work construes inequality mainly as complementarities between human and other capital, sidestepping political considerations. In contrast we have argued that technological change must be studied through the lenses of institutions and power, using tools familiar to sociologists. An abundant literature has also documented the pervasive negative effects of the disappearance of work on local communities (Brand, 2015). With the erosion of living standards that job loss brings, parents will be worse placed to provide a nurturing environment, access to good neighbourhood or schools, or pay for their children’s way through college. This naturally raises question of whether the increasing displacement of human workers has intergenerational repercussions. But until now, there has been little representative national-level data on this issue. Our findings also confirm some of the public perceptions of how technology disrupts communities once built on the foundation of well-paid and stable manufacturing jobs. In line with arguments that men are the “economic losers” of ongoing structural change, mobility deficits associated with automation is largely a male phenomenon. The spatial diffusion of automation also mirrors the recent precipitous rise in alcohol- and drug-related deaths and suicides (Dwyer-Lindgren et al., 2018). One proposed explanation for this troubling phenomenon focuses on intergenerational comparisons. As Case and Deaton (2015, p. 15081) note: “many of the baby-boom generation are the first to find, in midlife, that they will not be better off than were their parents.” However, while these “deaths of despair” are predominantly a non-Hispanic white phenomenon, our race differences show a more complex pattern. The raw mobility deficits are larger among non-Hispanic and Hispanic white men, but these demographic differences shrink once we factor in Black men’s overall lower prospects for upward mobility. To return to the liberal thesis of industrialization: what should take its place? The failure to provide a persuasive answer to this question is perhaps the main reason for its lingering in the sociological imagination. Instead of rejecting the thesis wholesale, we would recommend a conditional view where the mobility consequences of technological change depend on political context. When jobs in outmoded sectors make their demise there is a mechanical sense in which mobility increases: by pushing children away from following in their parents’ footsteps. But the

28 distributional consequences depend on how prepared society is to provide remediation through

education or other means. Societies have a choice to adapt and invest in their members. Ultimately, the challenge is to describe the institutional conditions under which industrial transformation promotes mobility, and those under which it hampers it. While ongoing technological changes

appear to displace rather than enhance labor, the question remains why. We have pointed to two broad sets of institutional determinants. The first is the balance of market power: laissez-faire policies may incentivize companies to invest in machinery but disincentivize them from investing in workers. The second concerns the availability of institutional remediation through education. We have provided some tentative evidence that these conditions moderate the effect of automation across U.S. labor markets, but it remains to study how automation has shaped mobility in more encompassing welfare states such as Germany or Scandinavia (DiPrete, 2002). This is not least motivated by findings that the labor-market impact of new technology seems to differ significantly across countries (Graetz and Michaels, 2018). Our work is also relevant for the question of how modernization shaped status attainment in times past, which is gaining new relevance with the increased availability of historical microdata (Song and Campbell, 2017). Our analysis has several limitations that offers opportunities for further research. We rely on administrative data that overcome many difficulties of sampling, definition, or measurement common to survey-based research. Yet, these data are not without constraints. Crucially, our variation is cross-sectional in nature and the observation window for parent and child incomes is limited. This is perhaps more likely to bias the level of intergenerational correlations than their geographic distribution on which we base our analysis (Mazumder, 2016). Yet, the fact that we document similar gradients in education suggests that, if anything, income differences might be larger with opportunity to follow exposed cohorts until old age. Another limitation of our study is that it takes an ecological view. To some extent, this is the appropriate level of analysis as mobility and structural transformation are by nature ecological phenomena. At the same time, further work is needed in connecting the meso- and micro-level in this regard. These limitations notwithstanding, our work opens up important new avenues for future research.

29 While mobility scholars have debated the implications of social changes in the nineteenth and

twentieth century, they have yet to consider imminent transformations of labor with the same seriousness. It has been claimed that we are at the brink of a “fourth industrial revolution” where human labor risks becoming redundant in many sectors. The public response to these

transformations will be shaped by whether they make the next generation better or worse oﬀ in the long run. An important implication of our ﬁndings is that automation technologies interact in complex ways with existing economic and social institutions in shaping intergenerational

transmission. While we focus on local labor markets within the U.S. which exhibit signiﬁcant variation in economic and social makeup, this highlights the potential role of broader national policies. More work on comparative country studies will be needed to understand the conditions

that determine how automation shapes the prospects of intergenerational mobility.

[Table 7 about here.]

[Table 8 about here.]

References

Acemoglu, Daron and Pascual Restrepo (2019), “Robots and jobs: Evidence from US labor

markets,” Journal of Political Economy, forthcom. Autor, David (2014), “Skills, education, and the rise of earnings inequality among the “other 99 percent”,” Science, 344, 6186, pp. 843-851.

— (2015), “Why are there still so many jobs?: The history and future of workplace automation,” Journal of Economic Perspectives, 29, 3, pp. 3-30. Autor, David, Frank Levy, and Richard J Murnane (2003), “The skill content of recent technological

change: An empirical exploration,” Quarterly Journal of Economics, 118, 4, pp. 1279-1333. Blanden, Jo (2013), “Cross-country rankings in intergenerational mobility: a comparison of approaches from economics and sociology,” Journal of Economic Surveys, 27, 1, pp. 38-73.

30 Blau, Peter M and Otis Dudley Duncan (1967), The American Occupational Structure, John Wiley

& Sons. Bloome, Deirdre, Shauna Dyer, and Xiang Zhou (2018), “Educational Inequality, Educational Expansion, and Intergenerational Income Persistence in the United States,” American

Sociological Review, 83, 6, pp. 1215-1253. Brady, David, Regina S Baker, and Ryan Finnigan (2013), “When unionization disappears: State-level unionization and working poverty in the United States,” American Sociological

Review, 78, 5, pp. 872-896. Brady, David, Agnes Blome, and Hanna Kleider (2016), “How politics and institutions shape poverty and inequality,” The Oxford handbook of the social science of poverty, p. 117.

Brand, Jennie E (2015), “The far-reaching impact of job loss and unemployment,” Annual Review of Sociology, 41, pp. 359-375. Breen, Richard (ed.) (2004), Social mobility in Europe, Oxford University Press.

Breen, Richard and Jan O Jonsson (2005), “Inequality of opportunity in comparative perspective: Recent research on educational attainment and social mobility,” Annual Review of sociology, 31, pp. 223-243.

Breen, Richard and Walter Müller (2020), Education and Intergenerational Social Mobility in Europe and the United States, Stanford University Press. Buchmann, Claudia and Thomas A DiPrete (2006), “The growing female advantage in college

completion: The role of family background and academic achievement,” American Sociological Review, 71, 4, pp. 515-541. Busemeyer, Marius R and Torben Iversen (2012), “Collective skill systems, wage bargaining, and

labor market stratiﬁcation,” in The political economy of collective skill formation, ed. by Marius R Busemeyer and Christine Trampusch, Oxford University Press, pp. 205-33. Case, Anne and Angus Deaton (2015), “Rising morbidity and mortality in midlife among white

non-Hispanic Americans in the 21st century,” Proceedings of the National Academy of Sciences, 112, 49, pp. 15078-15083.

31 Cermeño, Alexandra L (2018), “Do universities generate spatial spillovers?: Evidence from US

counties between 1930 and 2010,” Journal of Economic Geography. Cherlin, Andrew J (2014), Labor’s love lost: the rise and fall of the working-class family in America, Russell Sage Foundation.

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, tech. rep., National Bureau of Economic Research.

Chetty, Raj and Nathaniel Hendren (2018), “The impacts of neighborhoods on intergenerational mobility II: County-level estimates,” Quarterly Journal of Economics, forthcom. Chetty, Raj, Nathaniel Hendren, Maggie R Jones, and Sonya R Porter (2018), Race and Economic

Opportunity in the United States: An Intergenerational Perspective, National Bureau of Economic Research. 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, 4, pp. 1553-1623. Conger, Rand D and Glen H Elder Jr (1994), Families in Troubled Times, New York: Aldine de

Gruyter. Dauth, Wolfgang, Sebastian Findeisen, Jens Suedekum, and Nicole Woessner (2018), Adjusting to robots: Worker-level evidence, tech. rep., Federal Reserve Bank of Minneapolis, Opportunity

and Inclusive Growth Institute. Davis, Kingsley and Wilbert E Moore (1945), “Some principles of stratiﬁcation,” American Sociological Review, 10, 2, pp. 242-249.

Denice, Patrick and Jake Rosenfeld (2018), “Unions and Nonunion Pay in the United States, 1977-2015,” Sociological Science, 5, 23, pp. 541-561, issn: 2330-6696, doi: 10.15195/v5. a23, http://dx.doi.org/10.15195/v5.a23.

DiPrete, Thomas A (1993), “Industrial restructuring and the mobility response of American workers in the 1980s,” American Sociological Review, pp. 74-96.

32 DiPrete, Thomas A (2002), “Life course risks, mobility regimes, and mobility consequences: A

comparison of Sweden, Germany, and the United States,” American Journal of Sociology, 108, 2, pp. 267-309. DiPrete, Thomas A and Patricia A McManus (1996), “Institutions, technical change, and diverging

life chances: Earnings mobility in the United States and Germany,” American Journal of Sociology, 102, 1, pp. 34-79. DiPrete, Thomas A and K Lynn Nonnemaker (1997), “Structural change, labor market turbulence,

and labor market outcomes,” American Sociological Review, pp. 386-404. Dorn, David (2009), Essays on inequality, spatial interaction, and the demand for skills, PhD dissertation no. 3613, University of St. Gallen.

Dwyer-Lindgren, Laura, Amelia Bertozzi-Villa, Rebecca W Stubbs, Chloe Morozoﬀ, Shreya Shirude, Jürgen Unützer, Mohsen Naghavi, Ali H Mokdad, and Christopher JL Murray (2018), “Trends and patterns of geographic variation in mortality from substance use disorders and

intentional injuries among US counties, 1980-2014,” Jama, 319, 10, pp. 1013-1023. Edin, Kathryn, Timothy Nelson, Andrew Cherlin, and Robert Francis (2019), “The Tenuous Attachments of Working-Class Men,” Journal of Economic Perspectives, 33, 2, pp. 211-28.

Entwisle, Doris R, Karl L Alexander, and Linda S Olson (2007), “Early schooling: The handicap of being poor and male,” Sociology of Education, 80, 2, pp. 114-138. Erikson, Robert and John H Goldthorpe (1992), The constant ﬂux: A study of class mobility in

industrial societies, Oxford University Press. Esping-Andersen, Gøsta (1993), Changing classes: Stratiﬁcation and mobility in post-industrial societies, SAGE Publications.

Fernandez, Roberto M (2001), “Skill-biased technological change and wage inequality: Evidence from a plant retooling,” American Journal of Sociology, 107, 2, pp. 273-320. Frey, Carl B (2019), The Technology Trap: Capital, Labor, and Power in the Age of Automation,

Princeton University Press.

33 Frey, Carl B and Michael A Osborne (2017), “The future of employment: how susceptible are jobs

to computerisation?” Technological Forecasting and Social Change, 114, pp. 254-280. Gallie, Duncan (2017), “The quality of work in a changing labour market,” Social Policy & Administration, 51, 2, pp. 226-243.

Ganzeboom, Harry BG, Donald J Treiman, and Wout C Ultee (1991), “Comparative intergenerational stratiﬁcation research: Three generations and beyond,” Annual Review of sociology, 17, 1, pp. 277-302.

Gassman-Pines, Anna, Christina M Gibson-Davis, and Elizabeth O Ananat (2015), “How economic downturns aﬀect children’s development: an interdisciplinary perspective on pathways of inﬂuence,” Child Development Perspectives, 9, 4, pp. 233-238.

Gest, Justin (2016), The new minority: White working class politics in an age of immigration and inequality, Oxford University Press. Goldin, Claudia D and Lawrence F Katz (2009), The race between education and technology,

Harvard University Press. Goldstein, Amy (2017), Janesville: An American Story, Simon and Schuster. Goldthorpe, John H (1960), “Social stratiﬁcation in industrial society,” The Sociological Review,

8, S1, pp. 97-122. Graetz, Georg and Guy Michaels (2017), “Is modern technology responsible for jobless recoveries?” American Economic Review, 107, 5, pp. 168-73.

— (2018), “Robots at work,” Review of Economics and Statistics, 100, 5, pp. 753-768. Grusky, David B (1983), “Industrialization and the status attainment process: The thesis of industrialism reconsidered,” American Sociological Review, 48, 4, pp. 494-506.

Grusky, David B and Thomas A DiPrete (1990), “Recent trends in the process of stratiﬁcation,” Demography, 27, 4, pp. 617-637. Grusky, David B and Robert M Hauser (1984), “Comparative social mobility revisited: models of

convergence and divergence in 16 countries,” American Sociological Review, 49, 1, pp. 19-38.

34 Harding, David J and Martin D Munk (2019), “The Decline of Intergenerational Income Mobility in

Denmark: Returns to Education, Demographic Change, and Labor Market Experience,” Social Forces, forthcom. Hazelrigg, Lawrence E and Maurice A Garnier (1976), “Occupational mobility in industrial

societies: a comparative analysis of diﬀerential access to occupational ranks in seventeen countries,” American Sociological Review, 41, 3, pp. 498-511. Hirsch, Barry T, David A Macpherson, and Wayne G Vroman (2001), “Estimates of union density

by state,” Monthly Labor Review, 124, 7, pp. 51-55. Hochschild, Arlie Russell (2018), Strangers in their own land: Anger and mourning on the American right, The New Press.

Horowitz, Jonathan (2018), “Relative Education and the Advantage of a College Degree,” American Sociological Review, 83, 4, pp. 771-801. Hout, Michael (1988), “More universalism, less structural mobility: The American occupational

structure in the 1980s,” American Journal of Sociology, 93, 6, pp. 1358-1400. — (1989), Following in father’s footsteps: Social mobility in Ireland, Harvard University Press. Hout, Michael and Thomas A DiPrete (2006), “What we have learned: RC28’s contributions to

knowledge about social stratification,” Research in Social Stratification and Mobility, 24, 1, pp. 1-20. Jackson, Michelle and David B Grusky (2018), “A post-liberal theory of stratification,” British

Journal of Sociology, 69, 4, pp. 1096-1133. Jacobs, David and Jonathan C Dirlam (2016), “Politics and economic stratiﬁcation: Power resources and income inequality in the United States,” American Journal of Sociology, 122, 2, pp. 469-500.

Jacobs, David and Lindsey Myers (2014), “Union strength, neoliberalism, and inequality: Contingent political analyses of US income diﬀerences since 1950,” American Sociological Review, 79, 4, pp. 752-774.

Jahoda, Marie, Paul F Lazarsfeld, and Hans Zeisel ([1933]1971), Marienthal: The sociography of an unemployed community, Chicago: Aldine Atherton.

35 Jerrim, John, Anna K Chmielewski, and Phil Parker (2015), “Socioeconomic inequality in access

to high-status colleges: A cross-country comparison,” Research in Social Stratiﬁcation and Mobility, 42, pp. 20-32. Jonsson, Jan O, David B Grusky, Matthew Di Carlo, Reinhard Pollak, and Mary C Brinton (2009),

“Microclass mobility: Social reproduction in four countries,” American Journal of Sociology, 114, 4, pp. 977-1036. Jürgens, Ulrich, Thomas Malsch, and Knuth Dohse (1993), Breaking from Taylorism: Changing

forms of work in the automobile industry, Cambridge University Press. Karlson, Kristian Bernt (2019), “College as equalizer?: Testing the selectivity hypothesis,” Social Science Research, 80, pp. 216-229.

Kim, Jerry W, Bruce Kogut, and Jae-Suk Yang (2015), “Executive compensation, fat cats, and best athletes,” American Sociological Review, 80, 2, pp. 299-328. Knigge, Antonie, Ineke Maas, Marco HD Van Leeuwen, and Kees Mandemakers (2014),

“Status attainment of siblings during modernization,” American Sociological Review, 79, 3, pp. 549-574. Knigge, Antonie, Ineke Maas, and Marco HD van Leeuwen (2014), “Sources of sibling (dis)

similarity: Total family impact on status variation in the Netherlands in the nineteenth century,” American Journal of Sociology, 120, 3, pp. 908-948. Kristal, Tali (2013), “The capitalist machine: Computerization, workers’ power, and the decline in

labor’s share within US industries,” American Sociological Review, 78, 3, pp. 361-389. Krzywdzinski, Martin (2017), “Automation, skill requirements and labour-use strategies: high-wage and low-wage approaches to high-tech manufacturing in the automotive industry,” New

Technology, Work and Employment, 32, 3, pp. 247-267. Lamont, Michèle (2009), The dignity of working men: Morality and the boundaries of race, class, and immigration, Harvard University Press.

Lipset, Seymour Martin and Reinhard Bendix (1959), Social mobility in industrial society, University of California Press Berkeley.

36 Liu, Shimeng (2015), “Spillovers from universities: Evidence from the land-grant program,”

Journal of Urban Economics, 87, pp. 25-41. Liu, Yujia and David B Grusky (2013), “The payoﬀ to skill in the third industrial revolution,” American Journal of Sociology, 118, 5, pp. 1330-1374.

Mayer, Susan E (2001), “How did the increase in economic inequality between 1970 and 1990 aﬀect children’s educational attainment?” American Journal of Sociology, 107, 1, pp. 1-32. — (2010), “The relationship between income inequality and inequality in schooling,” Theory and

Research in Education, 8, 1, pp. 5-20. Mazumder, Bhashkar (2016), “Estimating the Intergenerational Elasticity and Rank Association in the United States: Overcoming the Current Limitations of Tax Data,” in Inequality: Causes and

Consequences, Emerald Group Publishing Limited, pp. 83-129. Mitnik, Pablo A, Victoria Bryant, and Michael Weber (2019), “The Intergenerational Transmission of Family-Income Advantages in the United States,” Sociological Science, 6, pp. 380-415.

Mitnik, Pablo A, Erin Cumberworth, and David B Grusky (2016), “Social mobility in a high-inequality regime,” The ANNALS of the American Academy of Political and Social Science, 663, 1, pp. 140-184.

Mizruchi, Mark S (2013), The fracturing of the American corporate elite, Harvard University Press. Moller, Stephanie, Arthur S Alderson, and Francois Nielsen (2009), “Changing patterns of income inequality in US counties, 1970–2000,” American Journal of Sociology, 114, 4, pp. 1037-1101.

Mood, Carina (2010), “Logistic regression: Why we cannot do what we think we can do, and what we can do about it,” European Sociological Review, 26, 1, pp. 67-82. Morris, Martina and Bruce Western (1999), “Inequality in Earnings at the Close of the Twentieth

Century,” Annual Review of Sociology, 25, 1, pp. 623-657. Mouw, Ted (2000), “Job relocation and the racial gap in unemployment in Detroit and Chicago, 1980 to 1990,” American Sociological Review, pp. 730-753.

Neckerman, Kathryn M and Florencia Torche (2007), “Inequality: Causes and consequences,” Annual Review of Sociology, 33, pp. 335-357.

37 Oﬃce of Technology Assessment (1984), Computerized Manufacturing Automation: Employment,

Education, and the Workplace, OTA-CIT-235. Washington, DC: US Government Publishing Oﬃce, https://digital.library.unt.edu/ark:/67531/metadc39537/m2/1/high_ res_d/8408.pdf.

Parsons, Talcott (1951), The Social System, New York: Free Press. Pérez, Santiago (2017), “Railroads and the Rural to Urban Transition: Evidence from 19th-Century Argentina,” Job Market Paper.

Pfeﬀer, Fabian T and Florian R Hertel (2015), “How has educational expansion shaped social mobility trends in the United States?” Social Forces, 94, 1, pp. 143-180. Powell, Walter W and Kaisa Snellman (2004), “The knowledge economy,” Annual Review of

Sociology, 30, pp. 199-220. Reardon, Sean F (2011), “The widening academic achievement gap between the rich and the poor: New evidence and possible explanations,” in Whither opportunity?: Rising inequality,

schools, and children’s life chances, ed. by Greg J Duncan and Richard J Murnane, Russell Sage Foundation, pp. 91-116. Reynolds, John R and Monica Kirkpatrick Johnson (2011), “Change in the stratiﬁcation of

educational expectations and their realization,” Social Forces, 90, 1, pp. 85-109. Roksa, Josipa, Eric Grodsky, Richard Arum, and Adam Gamoran (2007), “United States: Changes in Higher Education and Social Stratiﬁcation,” in Stratiﬁcation in higher education: A comparative

study, ed. by Yossi Shavit, Richard Arum, and Adam Gamoran, Stanford University Press Stanford, CA, pp. 165-194. Song, Xi and Cameron D Campbell (2017), “Genealogical microdata and their signiﬁcance for

social science,” Annual Review of Sociology, 43, pp. 75-99. Tolbert, Charles M and Molly Sizer (1996), US commuting zones and labor market areas: A 1990 update, tech. rep.

Tope, Daniel and David Jacobs (2009), “The politics of union decline: The contingent determinants of union recognition elections and victories,” American Sociological Review, 74, 5, pp. 842-864.

38 Torche, Florencia (2011), “Is a college degree still the great equalizer?: Intergenerational mobility

across levels of schooling in the United States,” American Journal of Sociology, 117, 3, pp. 763-807. Treiman, Donald J (1970), “Industrialization and social stratiﬁcation,” Sociological Inquiry, 40, 2,

pp. 207-234. Treiman, Donald J and Harry BG Ganzeboom (2000), “The fourth generation of comparative stratiﬁcation research,” The international handbook of sociology, pp. 122-50.

U.S. Census Bureau (n.d.), Census Bureau Regions and Divisions with State FIPS Codes, https: //www2.census.gov/geo/pdfs/maps-data/maps/reference/us_regdiv.pdf. VanHeuvelen, Tom (2018a), “Moral Economies or Hidden Talents?: A Longitudinal Analysis of

Union Decline and Wage Inequality, 1973–2015,” Social Forces, 97, 2, pp. 495-530. — (2018b), “Recovering the Missing Middle: A Mesocomparative Analysis of Within-Group Inequality, 1970–2011,” American Journal of Sociology, 123, 4, pp. 1064-1116.

Volscho, Thomas W and Nathan J Kelly (2012), “The rise of the super-rich: Power resources, taxes, ﬁnancial markets, and the dynamics of the top 1 percent, 1949 to 2008,” American Sociological Review, 77, 5, pp. 679-699.

Wallace, Michael, Andrew S Fullerton, and Mustafa E Gurbuz (2009), “Union organizing eﬀort and success in the US, 1948–2004,” Research in Social Stratiﬁcation and Mobility, 27, 1, pp. 13-34. Western, Bruce and Jake Rosenfeld (2011), “Unions, norms, and the rise in US wage inequality,”

American Sociological Review, 76, 4, pp. 513-537. Willis, Paul E. (1977), Learning to labour: How working class kids get working class jobs, Saxon House.

Wilson, William Julius (1987), The truly disadvantaged: The inner city, the underclass, and public policy, University of Chicago Press. — (1996), When work disappears: The world of the new urban poor, New York: Alfred A. Knopf.

Zhou, Xiang (2019), “Equalization or Selection?: Reassessing the ‘Meritocratic Power’ of a College Degree in Intergenerational Income Mobility,” American Sociological Review, forthcom.

39 E

O D

Figure 1: The OED triangle: origin, education, destination.

40 250000

200000

150000

100000

50000

1980 1990 2000 2010 2020

Children born Parent income measured Child income measured Number of robots

Figure 2: Automation and Measurement of Parent and Child Income in the Data

41 Figure 3: Exposure to Automation across Commuting Zones

42 .5

Cleveland .4 New Orleans Chicago JacksonvilleAtlanta Detroit Washington DC Fort Worth .3 Denver Las VegasSalt Lake City San Jose Rank correlation Rank .2

Linear fit .1 Unweighted (r =.41) Weighted (r =.35)

-4 -2 0 2 4 Exposure to automation

Figure 4: Automation and Intergenerational Persistence in Income Rank across Commuting Zones.

43 Origin income quintile Bottom 2 3 4 Top

.05

0 Difference in

transition probability transition -.05

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Destination income quintile

Figure 5: Automation and Income Quintile Transition Probabilities

44 Table 1: Intergenerational Mobility and Exposure to Automation (OLS).

Panel A. Outcome: Relative mobility (rank-rank slope)

(1) (2) (3) (4) (5) (6)

Exposure to automation 0.021∗∗∗ 0.026∗∗∗ 0.024∗∗∗ 0.018∗∗∗ 0.011∗∗ 0.000 (0.004) (0.006) (0.002) (0.002) (0.004) (0.006) Mean of outcome 0.333 0.333 0.333 0.333 0.333 0.333 Observations (CZs) 693 693 693 693 693 693

Panel B. Outcome: Absolute upward mobility (p25)

(1) (2) (3) (4) (5) (6)

Exposure to automation -0.883∗ -0.987∗ -1.272∗∗∗ -1.165∗∗ -1.502∗∗ -1.153∗∗ (0.440) (0.467) (0.364) (0.397) (0.486) (0.441) Mean of outcome 41.624 41.624 41.624 41.624 41.624 41.624 Observations (CZs) 693 693 693 693 693 693

Panel C. Outcome: Rags-to-riches mobility

(1) (2) (3) (4) (5) (6)

Exposure to automation -0.009∗∗ -0.011∗∗ -0.012∗∗∗ -0.010∗∗∗ -0.012∗∗∗ -0.008∗ (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) CZ controls No Yes Yes Yes Yes Yes CZ demographics No No Yes Yes Yes Yes CZ education No No No Yes Yes Yes CZ manufacturing No No No No Yes Yes Census region FE No No No No No Yes Mean of outcome 0.081 0.081 0.081 0.081 0.081 0.081 Observations (CZs) 712 712 712 712 712 712

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.

45 Table 2: Absolute Upward Mobility and Exposure to Automation (OLS): Gender and Race.

Outcome: Absolute upward mobility (p25) Panel A. Male Panel B. Female

Black Hispanic White Black Hispanic White

(1) (2) (3) (4) (5) (6)

Exposure to automation -1.086∗∗ -1.578∗∗ -1.529∗∗ -0.347 -1.060∗∗ -0.927∗ (0.460) (0.539) (0.560) (0.316) (0.446) (0.484) CZ controls Yes Yes Yes Yes Yes Yes CZ demographics Yes Yes Yes Yes Yes Yes CZ education Yes Yes Yes Yes Yes Yes CZ manufacturing Yes Yes Yes Yes Yes Yes Census region FE Yes Yes Yes Yes Yes Yes Mean of outcome 38.894 47.540 49.265 41.382 40.850 40.431 Observations (CZs) 550 670 722 555 676 722

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.

46 Table 3: Childhood Eﬀects and Exposure to Automation (OLS)

Panel A. Outcome: Childhood exposure eﬀect on upward mobility (p25) Family income Individual income

(1) (2) (3) (4) (5) (6) All Male Female All Male Female

Exposure to automation -0.075∗∗ -0.118∗∗∗ -0.018 -0.077∗∗ -0.106∗∗∗ -0.023 (0.025) (0.030) (0.031) (0.030) (0.030) (0.030) Mean of outcome -0.072 -0.109 -0.056 -0.080 -0.107 -0.047 Observations (CZs) 702 676 673 702 676 673

Panel B. Outcome: Childhood exposure eﬀect on upward mobility (p75) Family income Individual income

(1) (2) (3) (4) (5) (6) All Male Female All Male Female

Exposure to automation -0.071∗ -0.078∗∗ -0.079 -0.108 -0.094∗∗ -0.048 (0.032) (0.030) (0.047) (0.058) (0.037) (0.110) Mean of outcome -0.025 -0.049 -0.033 -0.050 -0.075 -0.066 Observations (CZs) 702 676 673 702 676 673

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.

47 Table 4: Educational Transitions and Exposure to Automation (OLS)

Outcome: Fraction of children attaining HS or college degree All Black Hispanic non-Hisp. White

Male Female Male Female Male Female Male Female

Panel A. High school or GED (1) (2) (3) (4) (5) (6) (7) (8)

Exposure to automation -0.007∗ -0.008∗∗ 0.002 -0.005 -0.015∗∗ -0.012∗∗ -0.012∗ -0.011∗ (0.003) (0.003) (0.006) (0.003) (0.005) (0.005) (0.007) (0.005) Mean of outcome 0.751 0.823 0.705 0.812 0.684 0.768 0.781 0.842 Observations 712 711 367 365 431 443 705 703

Panel B. At least some college (1) (2) (3) (4) (5) (6) (7) (8)

Exposure to automation -0.013∗∗ -0.007 -0.008 -0.010 -0.016∗∗ -0.012∗∗ -0.018∗∗∗ -0.009 (0.004) (0.005) (0.006) (0.005) (0.006) (0.004) (0.003) (0.006) Mean of outcome 0.469 0.621 0.415 0.626 0.422 0.550 0.486 0.633 Observations 706 705 313 316 347 370 696 694

Panel C. Four-year college degree (1) (2) (3) (4) (5) (6) (7) (8)

Exposure to automation -0.008∗∗ -0.010∗ -0.006 -0.010∗∗ -0.009∗∗ -0.016∗∗∗ -0.013∗∗ -0.016∗ (0.002) (0.005) (0.003) (0.003) (0.003) (0.003) (0.004) (0.008) CZ controls Yes Yes Yes Yes Yes Yes Yes Yes CZ demographics Yes Yes Yes Yes Yes Yes Yes Yes CZ education Yes Yes Yes Yes Yes Yes Yes Yes CZ manufacturing Yes Yes Yes Yes Yes Yes Yes Yes Census region FE Yes Yes Yes Yes Yes Yes Yes Yes Mean of outcome 0.155 0.234 0.111 0.204 0.110 0.176 0.166 0.247 Observations 706 705 313 316 347 370 696 694

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.

48 Table 5: Educational Transitions and Exposure to Automation: White Males by Childhood Income

Outcome: Fraction of children attaining HS or college degree

Parental income percentile: p1 p25 p50 p75 p100

Panel A. High school or GED (1) (2) (3) (4) (5)

Exposure to automation -0.019∗ -0.012∗ -0.008 -0.005 -0.003 (0.010) (0.007) (0.005) (0.003) (0.003) Mean of outcome 0.671 0.781 0.855 0.908 0.942 Observations 705 705 705 705 705

Panel B. At least some college (1) (2) (3) (4) (5)

Exposure to automation -0.021∗∗∗ -0.018∗∗∗ -0.015∗∗∗ -0.010∗∗∗ -0.005 (0.004) (0.003) (0.003) (0.003) (0.004) Mean of outcome 0.370 0.486 0.604 0.760 0.956 Observations 696 696 696 696 696

Panel C. Four-year college degree (1) (2) (3) (4) (5)

Exposure to automation -0.013∗∗ -0.013∗∗ -0.013∗∗ -0.013∗ -0.014 (0.004) (0.004) (0.004) (0.006) (0.013) CZ controls Yes Yes Yes Yes Yes CZ demographics Yes Yes Yes Yes Yes CZ education Yes Yes Yes Yes Yes CZ manufacturing Yes Yes Yes Yes Yes Census region FE Yes Yes Yes Yes Yes Mean of outcome 0.111 0.166 0.247 0.399 0.806 Observations 696 696 696 696 696

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.

49 Table 6: Interactions with institutional covariates (OLS).

Outcome: Absolute upward mobility (p25)

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Exposure to automation -0.878∗ -1.000∗∗ -0.852∗ -1.193∗∗ -0.906∗∗∗ -1.011∗ -0.991∗∗ -1.001∗∗ -0.965∗ -0.720∗ (0.392) (0.419) (0.416) (0.451) (0.243) (0.470) (0.346) (0.421) (0.431) (0.379) Top 1% share 0.198 0.399 0.451 0.441∗ (0.303) (0.264) (0.298) (0.212) Top 1% share * automation 0.136 -0.257∗∗∗ -0.216 -0.144 (0.342) (0.065) (0.439) (0.101) Union decline 1.308 0.085 1.361 0.223 (0.714) (0.278) (0.754) (0.187) Union decline * automation -0.886∗∗∗ 0.011 -0.914∗∗ -0.063 (0.217) (0.222) (0.273) (0.205)

50 Colleges per capita 0.872 0.702 1.003 0.670∗ (1.141) (0.396) (1.161) (0.342) Colleges per capita * automation -0.216 0.288 -0.418 0.250 (0.509) (0.212) (0.473) (0.245) College tuition 0.391 0.434 0.167 0.298 (0.404) (0.255) (0.415) (0.241) College tuition * automation 0.139 -0.629∗∗ 0.068 -0.592∗∗ (0.281) (0.190) (0.312) (0.183) CZ controls No Yes No Yes No Yes No Yes No Yes CZ demographics No Yes No Yes No Yes No Yes No Yes CZ education No Yes No Yes No Yes No Yes No Yes CZ manufacturing No Yes No Yes No Yes No Yes No Yes Census region FE No Yes No Yes No Yes No Yes No Yes Observations (CZs) 570 570 570 570 570 570 570 570 570 570

Note: Standard errors are given in parentheses and are clustered at the Census division level. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1. Table 7: Automation Exposure, Mobility Outcomes, and Institutional Covariates

Quintiles of exposure to automation

All Q1 Q2 Q3 Q4 Q5

Exposure to automation -0.00 -1.47 -0.48 0.12 0.56 1.29 Income mobility

Absolute mobility (p25) 44.02 46.79 46.31 42.49 42.92 42.05 Relative mobility (rank-rank correlation) 0.33 0.28 0.30 0.34 0.34 0.36 Rags-to-riches mobility (Q1-to-Q5) 0.10 0.14 0.12 0.08 0.09 0.08

Absolute mobility, Black males (p25) 39.78 42.18 41.44 39.55 39.51 38.04 Absolute mobility, Hisp. males (p25) 47.64 50.39 49.79 46.89 46.77 45.09 Absolute mobility, non-Hisp. white males (p25) 50.32 54.60 52.08 49.26 48.43 47.18

Absolute mobility, Black females (p25) 39.61 39.82 40.00 39.78 39.60 39.10 Absolute mobility, Hisp. females (p25) 38.32 38.79 38.92 37.93 38.19 37.93 Absolute mobility, non-Hisp. white females (p25) 39.51 41.63 40.57 38.56 38.94 37.84

Childhood eﬀects on income (p25) 0.29 0.57 0.54 0.10 0.16 0.11 Childhood eﬀects on income (p75) 0.21 0.23 0.28 0.16 0.20 0.18 Educational attainment

At least some college (males) 0.47 0.53 0.50 0.46 0.45 0.42 At least some college (females) 0.63 0.67 0.64 0.62 0.62 0.58 Institutional covariates

Top 1% share 10.93 9.91 10.63 11.58 11.57 10.84 Union decline 0.71 0.71 0.65 0.60 0.65 0.91 Colleges per capita 0.02 0.04 0.02 0.02 0.02 0.02

College tuition (/1000) 4.38 3.09 3.44 3.67 4.84 6.20

51 Table 8: Pearson’s Correlation Matrix of Automation Exposure, Mobility Outcomes, and Institutional Covariates

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18)

(1) Exposure to automation 1.00 (2) Absolute mobility (p25) -0.31 1.00 (3) Relative mobility (rank-rank correlation) 0.41 -0.70 1.00 (4) Rags-to-riches mobility (Q1-to-Q5) -0.40 0.92 -0.69 1.00 (5) Absolute mobility, Black males (p25) -0.33 0.35 -0.28 0.37 1.00 (6) Absolute mobility, Hisp. males (p25) -0.30 0.47 -0.40 0.48 0.32 1.00 (7) Absolute mobility, non-Hisp. white males (p25) -0.51 0.78 -0.50 0.76 0.43 0.54 1.00 52 (8) Absolute mobility, Black females (p25) -0.06 0.15 -0.16 0.14 0.14 0.10 0.19 1.00 (9) Absolute mobility, Hisp. females (p25) -0.08 0.18 -0.24 0.13 0.08 0.25 0.18 0.20 1.00 (10) Absolute mobility, non-Hisp. white females (p25) -0.33 0.65 -0.43 0.58 0.19 0.26 0.68 0.32 0.35 1.00 (11) Childhood eﬀects on income (p25) -0.29 0.96 -0.67 0.88 0.30 0.44 0.71 0.11 0.17 0.57 1.00 (12) Childhood eﬀects on income (p75) -0.09 0.74 -0.15 0.65 0.18 0.28 0.58 -0.02 -0.03 0.38 0.76 1.00 (13) At least some college (males) -0.39 0.64 -0.54 0.61 0.27 0.29 0.59 0.14 0.21 0.67 0.59 0.32 1.00 (14) At least some college (females) -0.34 0.52 -0.43 0.48 0.16 0.20 0.50 0.20 0.19 0.62 0.48 0.25 0.72 1.00 (15) Top 1% share 0.10 -0.21 0.01 -0.18 -0.08 -0.01 -0.14 0.07 0.17 -0.05 -0.22 -0.34 -0.03 -0.01 1.00 (16) Union decline 0.20 0.08 -0.09 0.03 -0.06 -0.06 -0.04 -0.03 -0.06 -0.09 0.09 0.01 -0.02 -0.06 -0.05 1.00 (17) Colleges per capita -0.31 0.24 -0.13 0.26 0.15 0.04 0.37 0.01 -0.19 0.34 0.18 0.22 0.33 0.36 -0.19 -0.06 1.00 (18) College tuition (/1000) 0.31 -0.02 0.10 -0.06 -0.17 -0.18 -0.11 0.11 0.08 0.12 -0.02 -0.01 -0.10 -0.08 0.10 0.23 0.01 1.00