Task Specialization in U.S. Cities from 1880 to 2000

Task Specialization in U.S. Cities from 1880 to 2000 Abstract We develop a new methodology for quantifying the tasks undertaken within occupations using over 3,000 verbs from more than 12,000 occupational descriptions in the Dictionary of Occupational Titles (DOTs). Using micro data from the United States from 1880 to 2000, we find an increase in the employment share of interactive occupations within sectors over time that is larger in metro areas than nonmetro areas. We interpret these findings using a model in which reductions in transport and communication costs induce urban areas to specialize according to their comparative advantage in interactive tasks. We present suggestive evidence relating increases in employment in interactive occupations to improvements in transport and communication technologies. Our findings highlight a change in the nature of agglomeration over time toward an increased emphasis on human interaction. 1. Introduction Agglomeration forces are widely understood to play a central role in sustaining the dense concentrations of population observed in urban areas. Much less is known about the detailed tasks undertaken in urban areas and how these have changed over time. Yet understanding the task content of employment in urban and rural areas is central to evaluating alternative theories of agglomeration and assessing the likely impact of improvements in transport and communication technologies on spatial concentrations of economic activity. In this paper, we provide new evidence on the detailed tasks undertaken by workers in urban and rural areas over a long historical time period in the United States. We develop a new methodology for measuring the individual tasks undertaken within occupations using the verbs from occupational descriptions in the Dictionary of Occupational Titles (DOTs). We implement this methodology using micro data on employment in disaggregated occupations and sectors in metro and nonmetro areas from 1880 to 2000. To measure the individual tasks undertaken within each occupation, we use over 3,000 verbs from more than 12,000 occupational descriptions in both historical and contemporary editions of the DOTs. Using these verbs, we find a systematic change in the composition of employment across tasks in urban versus rural areas over time. We quantify this change in task content using the meaning of verbs from Roget’s Thesaurus, as the standard reference for English usage. In both metro and nonmetro areas, we find a systematic reallocation of employment over time toward interactive occupations, which involve tasks described by verbs that appear in thesaurus categories concerned with thought, communication and intersocial activity. At the beginning of our sample period, metro areas actually have lower shares of employment in interactive occupations than nonmetro areas. Over time, employment growth in interactive occupations is much higher in metro areas, so that by the end of our sample period the initial pattern of specialization is reversed, and metro areas are more interactive than nonmetro areas. This increasing interactiveness of employment at higher population densities is observed not only between metro and nonmetro areas but also across metro areas of different population densities. Although in 1880 there is little relationship between specialization in interactive occupations and population density, by 2000 this relationship is positive, strong, and statistically significant. Taken together, these results suggest that human interaction has become increasingly important in agglomerations of economic activity over time. To interpret these empirical results, we develop a model of the spatial distribution of employment across occupations, sectors, and locations. The model features trade in the final good produced by each sector and in the tasks undertaken by the occupations within each sector. When the costs of trading tasks between locations are prohibitively high, all tasks are performed where the final good is produced. As the costs of trading tasks fall, it becomes feasible to unbundle production across locations and trade tasks between them. If agglomeration forces are stronger for interactive tasks, densely populated urban locations have a comparative advantage in interactive tasks, which implies that reductions in task trade costs induce them to specialize in more-interactive occupations, whereas more sparsely populated rural locations specialize in less-interactive occupations. We provide empirical evidence on the relationship between changes in the interactiveness of employment and the major innovations in transport and communication technology that occurred during our sample period. Following the award of Alexander Graham Bell’s patent for the telephone in 1886, the U.S. telephone network grew rapidly in the opening decades of the 20th century.1 After the award of Karl Benz’s patent for the internal combustion engine in 1879 and after the passage of the Federal Aid Road Act of 1916 and the Federal Highway Act of 1921, the U.S. road network and automobile use expanded rapidly over the same period.2 We examine the implications of these new transport and communication technologies by combining county data on employment by occupation and sector for 1880 and 1930 with newly collected county data on telephone use and the road network in the 1930s. We develop instruments for the geographical dissemination of both technologies. For the telephone, we use its network properties to construct an instrument based on proximity to nodes on the American Telegraph and Telephone’s (AT&T) company’s long distance trunk network, whose construction was influenced by the strategic objectives of connecting the nation as a whole. For roads, we use the 1922 “Pershing Map” of highway routes of military importance for coastal and border defense.3 We provide suggestive evidence connecting increases in the interactiveness of employment to the diffusion of these new technologies predicted by our instruments. Our paper is related to a number of literatures. We build on the wider literature on agglomeration economies, as surveyed by Duranton and Puga (2004, Chap. 48) and Rosenthal and Strange (2004, Chap. 49). One strand of this literature emphasizes differences in the composition of economic activity between urban and rural areas. Studies emphasizing the role of human capital and skills in promoting agglomeration include Glaeser and Saiz (2004), Glaeser and Resseger (2010), Bacolod, Blum, and Strange (2009a), Glaeser, Ponzetto, and Toblo (2014), and Moretti (2004, Chap. 51). Particular types of skills are highlighted by Bacolod, Blum, and Strange (2009b), which introduces the concept of soft skills that enable agents to interact in cities and industry clusters. More generally, the role of idea generation and exchange is emphasized by Davis and Dingel (2012), which develops a system of cities model in which costly idea exchange is the agglomeration force. Another line of research has distinguished different dimensions along which cities specialize. Duranton and Puga (2005) provides theory and evidence that in recent decades cities have shifted from specializing by sector—with integrated headquarters and plants—to specializing mainly by function—with headquarters and business services clustered in larger cities and plants clustered in smaller areas.4 Rossi-Hansberg, Sarte, and Owens (2009) develops a model in which firms choose locations of their headquarters and production facilities, and argues that the increased separation of these locations accounts for observed changes in patterns of residential and business activity. Ota and Fujita (1993) models the distinction between the front-unit (e.g., business office) and back-unit (e.g., plant or back-office) of firms and explores its implications for city structure. Helsley and Strange (2007) explicitly analyzes the vertical integration decision of the firm in conjunction with its location decision. Fujita and Tabuchi (1997) provides evidence that the increased separation of headquarters and production has contributed to observed changes in the distribution of economic activity across Japanese regions.5 Related research has examined the impact of roads and urban growth (e.g., Baum-Snow 2007; Duranton and Turner 2012) and the implications of innovations in communication technologies for cities (e.g., Pool 1977; Fischer 1992; Gaspar and Glaeser 1998; Leamer and Storper 2001). Our analysis is also related to the task-based approach to the labor market, including in particular Autor, Levy, and Murnane (2003, henceforth ALM), Acemoglu and Autor (2011, Chap. 12), Autor and Handel (2009), Autor (2013), Gray (2013), and Deming (2017). This research has developed measures of the task content of employment based on numerical scores from the Dictionary of Occupations (DOTs), such as “Direction, Control, and Planning (DCP)”. Prior to this research, the canonical model of the labor market in terms of skilled and unskilled labor had assumed that technological change is skill-biased. In contrast, this task-based approach recognizes that the direction of technological change is endogenous, as emphasized by Acemoglu (1998, 2002). Therefore, the extent to which new technologies complement or substitute for skills or tasks can change over time. In the labor literature, Autor, Katz, and Krueger (1998) argue that there was an acceleration in the skill-bias of technical change in the 1980s and 1990s. In the historical literature, several studies have argued that technical change often replaced rather than complemented skilled artisans in the 19th century, including Hounshell (1985), James and Skinner (1985), and Mokyr (1992).6 Relative to each of these lines of research, our main contribution is to develop a new approach for measuring individual production tasks that uses the verbs from occupational descriptions and their meanings. Using this new methodology, we are able to track the production tasks performed by workers at a much higher resolution than has hitherto been possible. We use our approach to provide evidence on changes in the task content of employment over a much longer historical time period than previously considered. We also apply this approach to the organization of economic activity in urban and rural areas separately. The remainder of the paper is structured as follows. Section 2 discusses the data. Section 3 introduces our methodology and presents our main empirical results on changes in the task content of employment in urban and rural areas over time. Section 4 outlines a theoretical framework for interpreting these empirical results that points to the role played by transportation and communication technologies. Section 5 provides further evidence on explanations for the observed changes in the task content of employment and their relationship with changes in transportation and communication technologies. Section 6 concludes. 2. Data Our empirical analysis uses two main sources of data. The first is individual-level records from the U.S. Population Census for 20-year intervals from 1880 to 2000 from Integrated Public Use Microdata Series (IPUMS): see Ruggles et al. (2010). These census micro data report individuals’ location, occupation, and sector, as well as other demographic information. We use these data to determine whether an individual is located in a metro area as well as the occupation and sector in which an individual is employed. A metro area is defined as a region consisting of a large urban core together with surrounding communities that have a high degree of economic and social integration with the urban core.7 We weight individuals by their person weights to ensure the representativeness of the sample. Our main data set is a panel from 1880 to 2000 that uses information on the share of employment within an occupation and sector in metro areas. To provide evidence on improvements in communication and transportation technologies, we also use long-differenced data from 1880 to 1930, aggregating our individual-level data to the county level. We use the standardized 1950 occupation classification from IPUMS, which distinguishes 11 two-digit occupations (e.g., “Clerical and Kindred”) and 281 three-digit occupations (e.g., “Opticians and Lens Grinders and Polishers”). We also use the standardized 1950 sector classification from IPUMS, which distinguishes twelve two-digit sectors (e.g., “Finance, Insurance and Real Estate”) and 158 three-digit sectors (e.g., “Motor Vehicles and Motor Vehicle Equipment”).8 Since we are concerned with employment structure, we omit workers who do not report an occupation and a sector (e.g., because they are unemployed or out of the labor force). We also exclude workers in agricultural occupations or sectors, because we compare task specialization in urban and rural areas over time, and agriculture is unsurprisingly overwhelmingly located in rural areas.9 We use time-varying definitions of metro areas to ensure that they correspond to meaningful economic units. However, we also report robustness tests, in which we hold the sample composition of metro areas constant over time, and in which we use administrative cities whose boundaries are more stable over time. Our second main data source is the Dictionary of Occupational Titles (U.S. Department of Labor 1991), which contains detailed descriptions of more than 12,000 occupations. Following Autor et al. (2003), previous research using DOTs typically uses the numerical scores that were constructed for each occupation by the Department of Labor (e.g., a Nonroutine Interactive measure based on the DCP numerical score). In contrast, we use verbs from the detailed occupational descriptions in DOTs to directly measure the tasks performed by workers in each occupation. We use a list of over 3,000 English verbs from “Writing English”, a company that offers English language consulting.10 This approach enables us to provide a rich analysis of the tasks undertaken in urban and rural areas using the verbs and occupational descriptions without being restricted to the numerical scores. Nonetheless, we also compare our measures of occupational characteristics to those from the numerical scores. We match the DOTs occupations to the three-digit occupations in our census data using the crosswalk developed by ALM. In our baseline specification, we use a time-invariant measure of tasks based on the occupational descriptions from the digital edition of the 1991 DOTs, which ensures that our results are not driven by changes in language use over time. In sensitivity checks, we also report results using digitized occupational descriptions from the first edition of the DOTs in 1939 (U.S. Department of Labor 1939). We complement these two main data sources with information from a variety of other sources. We use the standard reference for word usage in English (Roget’s Thesaurus) to quantify the meanings of verbs from the occupational descriptions.11 We use ArcGIS shapefiles from the National Historical Geographical Information System (NHGIS) to track the evolution of county boundaries over time. We also use measures of improvements in transport and communication technologies. We measure the length of roads in each county using a georeferenced 1931 road map (Gallup 1931).12 At the beginning of our sample in 1880, most U.S. roads were little more than dirt tracks (see, for example, Swift 2011) and widespread paved road construction only occurred following the Federal Aid Road Act of 1916 and the Federal Highway Act of 1921. Therefore, we use the 1931 map to construct a measure of the growth of the paved road network from 1880 to 1930. We measure the number of residence telephones in each county in 1935 using newly digitized data from American Telephone and Telegraph Company (AT&T 1935). The telephone was not patented until 1876 just before the beginning of our sample period and the telephone network developed rapidly from 1890 onward (see, e.g., Fischer 1992). Therefore, we use the data on telephones to construct a measure of the growth of telephones from 1880 to 1930. To address the concern that the road network could be influenced by changes in the interactiveness of economic activity, we use an instrument based on the “Pershing” map of highway routes of military importance for coastal and border defense. To address similar concerns for the telephone, we use an instrument based on proximity to primary and secondary outlets on AT&T’s long distance trunk network, whose construction was influenced by the strategic objective of connecting the nation as a whole. 3. Empirical Evidence on Task Specialization In this section, we present our main results on changes in the task content of employment in urban and rural areas over time. In Section 3.1, we begin by characterizing changes in specialization across occupations and sectors in metro areas relative to nonmetro areas over our long historical time period. In Section 3.2, we introduce our new methodology for measuring individual production tasks using the verbs from occupational descriptions. In Section 3.3, we explain how we quantify the meanings of these verbs using Roget’s thesaurus, and introduce a new measure of the interactiveness of the tasks undertaken by workers. In Section 3.4, we demonstrate the robustness of our results across a range of specifications. 3.1. Specialization Across Occupations and Sectors To provide some initial motivating evidence of changes in specialization across occupations and sectors in metro areas relative to nonmetro areas, we estimate the following regression for each year t separately using data across occupations o and sectors s: \begin{equation} \text{MetroShare}_{ost}=\mu _{ot}+\eta _{st}+\varepsilon _{ost}, \end{equation} (1) where MetroShareost is the share of employment in metro areas in occupation o, sector s and year t; observations are weighted by person weights; μot are occupation-year fixed effects; ηst are sector-year fixed effects; and εost is a stochastic error. We normalize the sector-year and occupation-year fixed effects so that they each sum to zero in each year, and hence they capture deviations from the overall mean in each year. Although we estimate the previous regression using a share as the left-hand side variable so that the estimated coefficients have a natural interpretation as frequencies, we find a very similar pattern of results in a robustness test in which we use a logistic transformation of the left-hand side variable: MetroShareost/(1 − MetroShareost). The occupation-year fixed effects (μot) capture the average probability of being in a metro area for workers in each occupation in each year, after controlling for differences across sectors in metro probabilities. Similarly, the sector-year fixed effects (ηst) capture the average probability of being located in a metro area for workers in each sector in each year, after controlling for differences across occupations in metro probabilities. The sector and occupation fixed effects are separately identified because there is substantial overlap in occupations and sectors, such that each sector contains multiple occupations and each occupation is employed in several sectors.13 We estimate this regression using both the aggregate (two-digit) and disaggregate (three-digit) definitions of occupations and sectors discussed previously. As reported in Table 1 for two-digit occupations and sectors, we find substantial changes in specialization across occupations and sectors in metro areas relative to nonmetro areas over time. From Panel A, in 1880, “Clerical and Kindred” workers were the most likely to be located in metro areas. In contrast, by 2000, “Clerical and Kindred” workers were ranked only fourth, and “Professional and Technical” workers were the most likely to be located in metro areas. From 1880 to 2000, declines in ranks were observed for “Craftsmen” (from 2 to 6) and “Operatives” (from 3 to 7), whereas increases in ranks were observed for “Professional and Technical” workers (from 7 to 1) and “Managers, Officials, and Proprietors” (from 6 to 3). As apparent from the first and fourth columns of the table, these changes in ranks reflect substantial changes in the probabilities of workers in individual occupations being located in metro areas over time. Table 1. Metro area specialization for aggregate occupations and sectors. Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Notes: Table reports the estimated coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on two-digit occupation-year and two-digit sector-year fixed effects (equation (1) in the paper). A separate regression is estimated for each year. Occupation and sector fixed effects are each normalized to sum to zero in each year. Standard errors are clustered by occupation. Occupations and sectors are sorted by the rank of their estimated coefficients for 1880. View Large Table 1. Metro area specialization for aggregate occupations and sectors. Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Notes: Table reports the estimated coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on two-digit occupation-year and two-digit sector-year fixed effects (equation (1) in the paper). A separate regression is estimated for each year. Occupation and sector fixed effects are each normalized to sum to zero in each year. Standard errors are clustered by occupation. Occupations and sectors are sorted by the rank of their estimated coefficients for 1880. View Large Since the regression (1) includes sector-year fixed effects, these changes in the metro probabilities for each occupation are not driven by changes in sector composition, but rather reflect changes in the organization of economic activity within sectors. Nonetheless, we also observe substantial changes in sector structure in metro areas relative to nonmetro areas over time. From Panel B, declines in ranks from 1880 to 2000 were observed for “Wholesale and Retail Trade” (from 2 to 6) and “Manufacturing” (from 4 to 10). In contrast, increases in ranks from 1880 to 2000 were observed for “Transportation, Communication and Other Utilities” (from 6 to 3) and “Business and Repair Services” (from 9 to 1). In Figures A.1 and A.2 of the Online Appendix, we show the evolution of the occupation and sector coefficients across each of the 20-year intervals in our data. Although “Professional and Technical” workers display an increased propensity to locate in metro areas from 1880 to 1960, the probability that “Managers, Officials and Proprietors” are located in urban areas increases particularly sharply from 1940 to 2000. In contrast, the likelihood that “Craftsmen” are found in metro areas declines throughout our sample period, whereas the probability for “Clerical and Kindred” workers declines from 1900 onward, and the probability for “Service” workers initially rises until 1920 and later declines until around 1960. Such changes in specialization are not limited to the aggregate categories considered so far, but are also found using more disaggregated measures of occupations and sectors. In Table A.1 of the Online Appendix, we report the results of estimating the regression (1) including three-digit-occupation-year and three-digit-sector-year fixed effects. Panels A and B report the 20 occupations within the largest increases and decreases respectively in the within-sector probability of being located in a metro area from 1880 to 2000. Both the top agglomerating occupations in Panel A and the top dispersing occupations in Panel B are diverse and span multiple sectors. For example, leading agglomerating occupations include “Editors and Reporters”, “Buyers and Department Heads”, and “Judges and Lawyers”, whereas prominent dispersing occupations are “Book Binders”, “Welders and Flame Cutters”, and “Upholsterers”. In our empirical analysis in what follows, we provide evidence on the systematic characteristics shared by occupations that agglomerate versus disperse over time. 3.2. Measuring the Tasks Undertaken by Occupations We now introduce our new methodology for measuring individual production tasks using the detailed descriptions from more than 12,000 disaggregated occupations included in the DOTs. We use the verbs from each occupation’s description to measure the tasks performed by workers within that occupation, because verbs capture an action (bring, read, walk, run, learn), an occurrence (happen, become), or a state of being (be, exist, stand), and hence capture the task being performed. To focus on persistent characteristics of occupations and abstract from changes in word use over time, our baseline analysis uses time-invariant occupational descriptions from the 1991 digital edition of the DOTs. Although the tasks undertaken within each occupation can change over time, the relative task content of occupations is likely to be more stable. To provide evidence on the extent to which this is the case, we have also digitized the occupational descriptions from the first edition of the DOTs in 1939. Although the descriptions of occupations are less detailed and the boundaries between occupations are less clear in the historical DOTs, we find a similar pattern of results using both sets of occupational descriptions, as discussed further in what follows. The first step of our procedure uses a list of over 3,000 English verbs from “Writing English”, a company that offers English language consulting. Using this list of verbs, we search each occupational description in the 1991 DOTs for occurrences of each verb in the first-person singular (e.g., (I) talk), third-person singular (e.g., (she) talks) or present participle (e.g., (he is) talking).14 For example, the occupational description for an Economist is given as follows: “ECONOMIST: Plans, designs, and conducts research to aid in interpretation of economic relationships and in solution of problems arising from production and distribution of goods and services: Studies economic and statistical data in area of specialization, such as finance, labor, or agriculture. Devises methods and procedures for collecting and processing data, utilizing knowledge of available sources of data and various econometric and sampling techniques. Compiles data relating to research area, such as employment, productivity, and wages and hours. Reviews and analyzes economic data in order to prepare reports detailing results of investigation, and to stay abreast of economic changes ...”, where the words detected by our procedure as capturing the tasks performed by an economist are italicized.15 Note that sometimes the first-person singular, third-person singular or present participle forms of a verb have the same spelling as the corresponding adjectives and nouns (e.g., “prepare reports”). In this case, our procedure treats these adjectives and nouns as verbs. To the extent that the use of the same word as an adjective or noun is closely related to its use as a verb, both uses are likely to capture the tasks performed. From this first step, we obtain the number of occurrences of each verb for each DOTs occupation. We next match the more than 12,000 DOTs occupations to IPUMS standardized 1950 occupations using the crosswalk developed by ALM. Finally, we calculate the frequency with which each verb v is used for each IPUMS occupation o: \begin{equation*} \text{VerbFreq}_{vo}=\frac{\text{Appearances of verb }v\text{ matched to }o}{\text{Appearances of all verbs matched to }o}, \end{equation*} where we focus on the frequency rather than the number of verb uses to capture the relative importance of tasks for an occupation and to control for potential variation in the length of the occupational descriptions matched to each IPUMS occupation.16 We provide evidence on changes in task specialization in metro areas relative to nonmetro areas over time by estimating the following regression for each verb v and year t separately using data across occupations o and sectors s: \begin{equation} \text{MetroShare}_{ost}=\alpha _{vt}\text{VerbFreq}_{vo}+\eta _{vst}+\varepsilon _{ost}, \end{equation} (2) where MetroShareost is again the share of employment in metro areas in occupation o, sector s, and year t; VerbFreqvo is defined previously for verb v and occupation o; ηvst are verb-sector-year fixed effects; and εost is a stochastic error. The coefficient of interest αvt captures a conditional correlation: the correlation between occupations’ shares of employment in metro areas and their frequency of use of verb v. The verb-sector-year fixed effects (ηvst) control for differences across sectors in the frequency of verb use and for differences across sectors and over time in the concentration of employment in metro areas. Since VerbFreqvo is time invariant, a rise in αvt over time implies that employment in occupations using that verb is increasingly concentrating in metro areas within sectors over time. In Panels A and B of Table 2, we report for each year the ten verbs with the highest and lowest standardized coefficient αvt (the estimated coefficient multiplied by the standard deviation of VerbFreqvo).17 As apparent from Panel A, we find substantial changes in the tasks most concentrated in metro areas within sectors over time. In 1880, the verbs with the highest metro employment shares typically involve physical tasks such as “Braid”, “Sew”, “Stretch”, and “Thread”. By 1920, the top ten verbs include an increased number of clerical tasks, such as “Bill”, “File”, “Notice”, and “Record”. By 1980 and 2000, the leading metro verbs include a proliferation of interactive tasks, such as “Analyze”, “Advise”, “Confer”, and “Report”. As shown in Panel B, we also find some changes in the tasks least concentrated in metro areas, although here the pattern is less clear cut (e.g., “Tread” appears from 1880 to 1960 and “Turn” appears from 1960 to 2000). Table 2. Verbs most and least strongly correlated with metro area employment shares. Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Notes: Table reports the ranks of standardized coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which a verb is used for an occupation and three-digit sector-year fixed effects (equation (2) in the paper). A separate regression is estimated for each verb and year. Sector fixed effects are normalized to sum to zero in each year. Estimated coefficients are normalized by the standard deviation for the verb frequency. Verbs are sorted by the rank of their standardized coefficients. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large Table 2. Verbs most and least strongly correlated with metro area employment shares. Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Notes: Table reports the ranks of standardized coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which a verb is used for an occupation and three-digit sector-year fixed effects (equation (2) in the paper). A separate regression is estimated for each verb and year. Sector fixed effects are normalized to sum to zero in each year. Estimated coefficients are normalized by the standard deviation for the verb frequency. Verbs are sorted by the rank of their standardized coefficients. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large 3.3. Quantifying Task Specialization The approach developed in the previous section allows us to provide a detailed characterization of the tasks performed in urban and rural areas using the full list of verbs and all occupational descriptions. In this section, we now develop a quantitative measure of task specialization based on the meanings of these verbs. To do so, we use the online computer-searchable version of Roget’s Thesaurus (1911), which has been the standard reference for English language use for more than a century, and explicitly classifies words according to their underlying concepts and meanings. Roget’s classification was inspired by natural history, with its hierarchy of Phyla, Classes, Orders, and Families. Therefore, words are grouped according to progressively more disaggregated classifications that capture ever more subtle variations in meaning. A key advantage of this classification is that it explicitly takes into account that words can have different meanings depending on context by including extensive cross-references to link related groups of words.18 Roget’s Thesaurus is organized into “Classes” that are further disaggregated into the progressively finer partitions of “Divisions”, “Sections”, and “Categories”. There are 6 classes, 10 divisions, 38 sections, and around 1,000 categories.19 The first three classes cover the external world: Class I (Abstract Relations) deals with ideas such as number, order and time; Class II (Space) is concerned with movement, shapes and sizes; and Class III (Matter) covers the physical world and humankind’s perception of it by means of the five senses. The last three classes relate to the internal world of human beings: the human mind (Class IV, Intellect), the human will (Class V, Volition), and the human heart and soul (Class VI, Emotion, Religion, and Morality). To characterize the meaning of each verb v, we use the frequency with which it appears in each partition k of Roget’s Thesaurus: \begin{equation} \text{ThesFreq}_{vk}=\frac{\text{Appearances of verb }v\text{ in category }k\text{ of thesaurus}}{\text{Total appearances of verb }v\text{ in thesaurus }}, \end{equation} (3) where the partition k could be a class, division, section, or category of the thesaurus; our use of a frequency takes into account that each verb can have multiple meanings and provides a measure of the relative importance of each meaning. In counting verb appearances, we make use of the thesaurus’s structure, in which words with similar meanings appear under each thesaurus Category in a list separated by commas or semicolons. Based on this structure, we count appearances of a verb that are followed by a comma or semicolon, which enables us to abstract from appearances of a word in idioms that do not reflect its common usage.20 Combining the frequency with which a verb appears in each occupation’s description (VerbFreqvo in the previous section) and the frequency with which the verb appears in each category of the thesaurus (ThesFreqvk), we construct a quantitative measure of the extent to which the tasks performed in an occupation involve the concepts from each thesaurus category,21 \begin{equation*} \text{TaskContent}_{ko}=\sum _{v\in V}\text{VerbFreq}_{vo}\times \text{ThesFreq}_{vk}. \end{equation*} We use this measure to examine changes in task specialization in metro areas relative to nonmetro areas over time by estimating an analogous regression for each thesaurus category k and year t as for each verb and year in the previous section: \begin{equation} \text{MetroShare}_{ost}=\beta _{kt}\text{TaskContent}_{ko}+\eta _{kst}+\varepsilon _{ost}, \end{equation} (4) where MetroShareost is the share of employment in metro areas in occupation o, sector s and year t; TaskContentko is defined previously for thesaurus partition k and occupation o; ηkst are thesaurus-category-year fixed effects; and εost is a stochastic error. The coefficient of interest βkt again captures a conditional correlation: the correlation between occupations’ shares of employment in metro areas and their frequency of use of verbs in thesaurus category k. The thesaurus-category-sector-year fixed effects (ηkst) control for differences across sectors in the frequency of use of thesaurus categories and differences across sectors and over time in the concentration of employment in metro areas. Since TaskContentko is time invariant, a rise in βkt over time implies that employment in occupations using that category of the thesaurus is increasingly concentrating in metro areas within sectors over time. In Table 3, we report the estimation results for the 38 sections of the thesaurus (denoted by S), organized by the 6 classes (denoted by C) and 10 divisions of the thesaurus (denoted by D). We calculate the standardized coefficient for each thesaurus section (the estimated coefficient βkt multiplied by the variable’s standard deviation) and report the ranking of these standardized coefficients in 1880 and 2000 as well the difference in rankings between these two years (1880 minus 2000).22 Since the thesaurus section with the highest standardized coefficient is assigned a rank of one, positive differences in rankings correspond to thesaurus sections that are becoming more concentrated in metro areas within sectors over time. Table 3. Ranking of thesaurus sections by concentration in metro areas in 1880 and 2000. Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Notes: Coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which the verbs used for an occupation are classified within thesaurus sections and three-digit sector-year fixed effects (equation (4) in the paper). A separate regression is estimated for each thesaurus section and year. Verbs are from the time invariant occupational descriptions from the 1991 Dictionary of Occupations. Thesaurus sections ranks in 1880 and 2000 based on their estimated coefficient normalized by the standard deviation for the thesaurus section frequency; the largest value is assigned a rank of one. The difference in ranks in the final column is defined such that a positive value corresponds to a thesaurus section that becomes more concentrated in metro areas from 1880 to 2000. View Large Table 3. Ranking of thesaurus sections by concentration in metro areas in 1880 and 2000. Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Notes: Coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which the verbs used for an occupation are classified within thesaurus sections and three-digit sector-year fixed effects (equation (4) in the paper). A separate regression is estimated for each thesaurus section and year. Verbs are from the time invariant occupational descriptions from the 1991 Dictionary of Occupations. Thesaurus sections ranks in 1880 and 2000 based on their estimated coefficient normalized by the standard deviation for the thesaurus section frequency; the largest value is assigned a rank of one. The difference in ranks in the final column is defined such that a positive value corresponds to a thesaurus section that becomes more concentrated in metro areas from 1880 to 2000. View Large The results in Table 3 reveal a sharp change the relative ranking of thesaurus sections involving the external world (Classes I–III) and those involving the internal world of human beings (Classes IV–VI). In 1880, the top-five thesaurus sections most concentrated in metro areas included: Quantity (Class I), Time (Class I), Dimensions (Class II), Matter in General (Class III), and Affections in General (Class VI). In contrast, in 2000, the top-five thesaurus sections were: Nature of Ideas Communicated (Class IV), Special Intersocial Volition (Class V), Extension of Thought (Class IV), Reasoning Processes (Class IV), and Results of Reasoning (Class IV). The correlation between the rankings of the thesaurus sections in 1880 and 2000 is negative and statistically significant (−0.63). Positive changes in ranks in Table 3 are typically concentrated in thesaurus Classes IV and V, which correspond to the human mind and the human will, respectively. These classes include Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition). We summarize this combination of tasks—thought, communication and intersocial activity—as “interactiveness”. Our interpretation is that interaction inherently involves each of these components: thinking for oneself, communication these thoughts, and communicating them to other people in a social environment. We exclude Class V, Division 1 (Individual Volition) from our definition of interactiveness, because it is more concerned with individual reflection and decision making (e.g., motive, habit, willingness, choice) rather than interaction between people. Although some of the categories in Class VI could be interpreted as interactive (e.g., Section III, “Sympathetic Affections”), the other categories within this class seem to point more to contemplation and introspection than interaction between people (e.g., “Affections in General”, “Personal Affections”, “Moral Affection”, and “Religious Affection”). Furthermore, the interpersonal relationships described in “Sympathetic Affections” seem to largely concern relationships outside of work. Therefore, we measure the interactiveness of an occupation using the frequency with which verbs appear in that occupation’s description and the frequency with which those verbs appear in Divisions 1 and 2 of Class IV and Division 2 of Class V of the thesaurus: \begin{equation} \text{Interactive}_{o}=\sum _{v\in V}\text{FreqVerb}_{vo}\times \text{FreqInteractive}_{v}, \end{equation} (5) where FreqVerbvo is the frequency with which verb v is used for occupation o from above; FreqInteractivev is the frequency with which verb v appears in these partitions of the thesaurus (computed as in (3)). We also report results in what follows for all four divisions of Classes IV and V of the thesaurus. In Panels A and B of Table 4, we report the top ten and bottom ten interactive occupations using our measure. Although any single quantitative measure of interactiveness is unlikely to capture the full meaning of this concept, the occupations identified by our procedure as having high and low levels of interactiveness appear intuitive. “Buyers and Department Heads”, “Clergymen”, and “Pharmacists” arguably perform more interactive tasks than “Blasters and Powdermen”, “Roofers and Slaters”, and “Welders and Flame Cutters”. Table 4. Most and least interactive occupations. Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Notes: The table reports the ten occupations with the lowest and highest interactiveness, as measured by the frequency of verb use in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of Roget’s Thesaurus. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large Table 4. Most and least interactive occupations. Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Notes: The table reports the ten occupations with the lowest and highest interactiveness, as measured by the frequency of verb use in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of Roget’s Thesaurus. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large In Figure 1, we measure the interactiveness of metro areas, nonmetro areas, and the economy as a whole using the employment-weighted average of interactiveness for each occupation. In this measure, interactiveness only differs between metro and nonmetro areas to the extent that they have different distributions of employment across occupations: \begin{equation} \text{Interactive}_{jt}=\sum _{o=1}^{O}\frac{E_{ojt}}{E_{jt}}\text{Interactive }_{o},\qquad j\in \left\lbrace M,N\right\rbrace, \end{equation} (6) where j indexes a type of location and we again denote metro areas by M and nonmetro areas by N;Eojt corresponds to employment in occupation o in location type j ∈ {M, N} in year t. Figure 1. View largeDownload slide Mean interactiveness in metro and nonmetro areas over time. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 1. View largeDownload slide Mean interactiveness in metro and nonmetro areas over time. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. In 1880, metro and nonmetro areas have similar levels of interactiveness, with if anything metro areas having lower interactiveness than nonmetro areas. Over time, interactiveness increases in both sets of locations, but this increase is greater in metro areas than in nonmetro areas. This increase in the relative interactiveness of metro areas is particularly sharp from 1900 to 1920, which coincides with the dissemination of improvements in communication and transport technologies in the form of the telephone and roads and the automobile. In our empirical analysis in what follows, we provide further evidence on the extent to which changes in interactiveness are related to these new communication and transport technologies. 3.4. Robustness Having presented our baseline evidence of an increase in the interactiveness of employment in metro areas relative to nonmetro areas over time, we now document the robustness of this finding across a range of different specifications. 3.4.1. 1939 DOTs Our baseline specification measures the task content of employment using time-invariant occupational descriptions from the 1991 DOTs. Although this approach ensures that our findings are not driven by changes in language use over time, it assumes that the relative task content of occupations is persistent over time. One concern is that the interactiveness of occupations could have changed over time and these changes in interactiveness could be correlated with occupations’ shares of employment in metro areas. To address this concern, we replicated our analysis using the first edition of the DOTs from 1939. We digitized the occupational descriptions in the 1939 DOTs and implemented our procedure of searching for verbs in each occupational description. The boundaries between occupations are less well defined and the occupational descriptions are less detailed in the 1939 DOTs, which implies that the resulting measures of the task content of employment are likely to be less precise than those using the 1991 DOTs. Nonetheless, as reported in Table A.2 of the Online Appendix, we find similar changes in task specialization in this robustness test. The verbs most correlated with metro employment shares in 1880 include physical tasks such as “Retouch”, “Trawl”, and “Lure”. In contrast, the verbs most correlated with metro employment shares in 2000 include interactive tasks such as “Advise”, “Question”, and “Appraise”. Using the verbs from the 1939 occupational descriptions and the frequency with which these verbs appear in Class IV and Division 2 of Class V of the thesaurus, we again find an increase in the interactiveness of employment over time that is more rapid in metro areas than in nonmetro areas, as shown in Figure A.3 in the Online Appendix. This similarity of the results using both the 1939 and 1991 occupational descriptions suggests that our findings are unlikely to be driven by changes in the relative interactiveness of occupations over time. Indeed, although the layout of the occupational descriptions implies that our measure of interactiveness using the 1939 DOTs is less precise than our baseline measure using the 1991 DOTs (which by itself would induce an imperfect correlation), we find that they are positively and statistically significantly correlated. As reported in Table A.3 of the Online Appendix, the unweighted correlation coefficient between the 1939 and 1991 measures across the sample of occupations in 2000 is 0.62. 3.4.2. Metro Areas and Administrative Cities Our analysis has so far used variation between metro and nonmetro areas. To provide further evidence of a relative increase in the interactiveness of employment in densely populated locations, we now present evidence using a different source of variation across metro areas of differing population densities. In the top-left and top-right panels of Figure 2, we display mean interactiveness for each metro area (as calculated using (6)) against log population density for 1880 and 2000, respectively, as well as the fitted values and confidence intervals from locally weighted linear least squares regressions. To make the panels more legible, we omit a few outliers on both ends of the distribution from the figure (but not from the locally weighted linear least squares regressions). We use time-varying definitions of metro areas to ensure that they correspond to meaningful economic units, which implies that the number of observations changes over time as new metro areas enter the sample, as can be seen from comparing the two panels. In 1880, we find little relationship between interactiveness and log population density across metro areas, which is reflected in a negative but statistically insignificant OLS coefficient (standard error) of −0.0002 (0.0013). In contrast, in 2000, we find a positive and statistically significant relationship between interactiveness and log population density, which is reflected in an OLS coefficient (standard error) of 0.0018 (0.0002). In the bottom-left panel of Figure 2, we show that even when we restrict the 2000 sample to metro areas that exist in 1880, we continue to find a positive relationship that is statistically significant at the 10% level, confirming that these findings are not driven by a change in the composition of metro areas. Therefore, the increase in the relative interactiveness of densely populated locations over time is observed not only comparing metro and nonmetro areas but also comparing metro areas of differing population densities. Metro areas with relatively high levels of interactiveness conditional on population density in 2000 include Boston (BOS, MA) and New York (NYC, CT/NY/NJ), whereas those with low levels of interactiveness conditional on population density include Anniston (ANN, AL) and Mansfield (MAN, OH). Figure 2. View largeDownload slide Mean interactiveness across metro areas in 1880 and 2000. X-axes are log population density. Y-axes are mean interactiveness. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Thick solid lines are the fitted values from locally weighted linear least squares regressions. Thin solid lines are 95% point confidence intervals. Figures (but not the regressions) are truncated at both ends for outliers. Figure 2. View largeDownload slide Mean interactiveness across metro areas in 1880 and 2000. X-axes are log population density. Y-axes are mean interactiveness. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Thick solid lines are the fitted values from locally weighted linear least squares regressions. Thin solid lines are 95% point confidence intervals. Figures (but not the regressions) are truncated at both ends for outliers. Although we use time-varying definitions of the boundaries of metro areas to ensure that they correspond to meaningful economic units, we find similar results if we instead define urban areas as administrative cities, which have much more stable geographical boundaries over time. Again we find an increase in the relative interactiveness of urban areas over time, whether we compare administrative cities to all other locations (Figure A.4 in the Online Appendix) or only to nonmetro areas (Figure A.5 in the Online Appendix). Therefore, the increase in the relative interactiveness of urban areas also occurs within existing geographical boundaries. 3.4.3. Other Occupational Characteristics Our approach of using verbs from the occupational descriptions enables us to measure individual production tasks at a much finer level of resolution than has hitherto been possible. We now compare aggregations of our individual task measures, such as interactiveness, to existing measures of tasks, including the numerical scores from the DOTs used by ALM. Since these numerical scores are not available in the first edition of the DOTs in 1939, we use their values from the 1991 digital edition of the DOTs. As a point of comparison, in Figure A.8 of the Online Appendix, we show the employment-weighted average of the five ALM measures of task inputs over our long historical time period (analogous to Figure 1 in ALM but for 1880–2000 instead of 1960–2000). In Figures A.9– A.10 of the Online Appendix, we show the employment-weighted average of our measures of task inputs based on the 38 thesaurus sections. As is clear from these figures, there is a lot more variation using our thesaurus measures than using the five numerical scores. Therefore, we obtain a much richer picture of changes in task inputs over time and in urban versus rural areas using our measures based on the meanings of verbs from occupational descriptions. In Table A.3 of the Online Appendix, we report the correlation coefficients between our interactiveness measure and other measures of occupation characteristics across the sample of occupations in 2000. We report both unweighted correlations and correlated weighted by occupation employment. The highest correlation coefficients are for the Nonroutine Interactive (DCP) and Nonroutine Analytic (MATH) used by ALM. Although both of these measures are related to the concepts of thought, communication and intersocial activity captured by our interactiveness measure, the correlations are around 0.5. Therefore, our interactiveness measure captures distinctive information about the tasks performed by workers within occupations. Although DCP is orientated toward top–down interactions between workers (e.g., between a manager and her subordinates), our measure captures all interactions between workers (e.g., between members of a product design team). Although MATH is orientated toward thought, our measure of interactiveness also captures communication and intersocial activity. As a further check on our measure of interactiveness, Table A.4 of the Online Appendix reports the top-five verbs concentrated in each thesaurus section. We find that the verbs most concentrated in those sections of the thesaurus included in our definition of interactiveness do indeed seem to involve interactive tasks. For example, the top-five verbs most concentrated in “Class IV, Division II, Section I, Nature of Ideas Communicated” are “Annotate”, “Decipher”, “Interpret”, “Fudge”, and “Clarify”. As a final check on our measure of interactiveness, Table A.5 in the Online Appendix compares it to a number of other measures of interactiveness from the Occupational Information Network (O*NET). In particular, we report the correlation across occupations between our measure of interactiveness and 17 subcategories of the work activity “Interacting with Others” from the O*NET. These measures were constructed by US Department of Labor/Employment and Training Administration (USDOL/ETA) based on questionnaires about detailed work activities issued to a random sample of businesses and workers. Panel A reports unweighted correlations, whereas Panel B reports correlations weighted by employment. The measures cover a wide range of forms of interaction, including “Assisting and caring for others” and “Resolving conflict and negotiating with others”. We find that the correlations with our measure of interactiveness are all positive and typically statistically significant. The five categories with the highest unweighted correlations are as follows: “Communicating with persons outside organization”, “Establishing and maintaining interpersonal relationships”, “Performing administrative activities”, “Resolving conflict and negotiating with others”, and “Provide consultation and advice to others”. These five categories appear to correspond to the concepts of thought, communication and intersocial activity captured in our measure of interactiveness. 4. Theoretical Model In this section, we outline a theoretical model that we use to interpret our empirical finding of an increased interactiveness of employment in urban areas relative to rural areas over time.23 The model explains the distribution of employment across occupations, sectors and locations. Despite allowing for a large number of locations and a rich geography of trade costs, the model remains tractable, because of the stochastic formulation of productivity differences across occupations, sectors and locations. The key predictions of the model are comparative statics with respect to the costs of trading the tasks produced by each occupation and the final goods produced by each sector. When these costs are large, all locations have similar employment structures across sectors, and all tasks within each sector are undertaken in the same location where the final good is produced. As the costs of trading final goods and tasks fall, locations specialize across sectors and across occupations within sectors according to their comparative advantage as determined by productivity differences. If densely-populated urban locations have a comparative advantage in interactive tasks relative to sparsely-populated rural locations, the model predicts that a fall in the costs of trading tasks leads to an increase in the interactiveness of employment within sectors in urban relative to rural areas. 4.1. Preferences and Endowments The economy consists of many locations indexed by n ∈ N. Each location n is endowed with an exogenous supply of land $$\bar{H}_{n}$$. The economy as a whole is endowed with a measure of workers $$\bar{L}$$, who are perfectly mobile across locations. Workers’ preferences are defined over a goods consumption index (Cn) and residential land use (Hn) and are assumed to take the Cobb–Douglas form24 \begin{equation} U_{n}=\left( \frac{C_{n}}{\alpha } \right)^{\alpha } \left( \frac{H_{n}}{1-\alpha } \right)^{1-\alpha },\qquad 0<\alpha <1. \end{equation} (7) The goods consumption index (Cn) is assumed to be a constant elasticity of substitution (CES) function of consumption indices for a number of sectors (e.g., Manufacturing, Services) indexed by s ∈ S: \begin{equation} C_{n}=\left[ \sum _{s \in S} C_{ns}^{\frac{\beta -1}{\beta }}\right] ^{\frac{\beta }{\beta -1}}, \end{equation} (8) where β is the elasticity of substitution between sectors. Sectors can be either substitutes (β > 1) or complements in goods consumption (0 < β < 1), where the standard assumption in the literature on structural transformation in macroeconomics is complements (e.g., Ngai and Pissarides 2007; Yi and Zhang 2013). The consumption index for each sector is in turn a CES function of consumption of a continuum of goods (e.g., Motor Vehicles, Drugs, and Medicines) indexed by j ∈ [0, 1]: \begin{equation} C_{ns}=\left[ \int _{0}^{1}c_{ns}(j)^{\frac{\sigma _{s}-1}{\sigma _{s}}}dj \right] ^{\frac{\sigma _{s}}{\sigma _{s}-1}}, \end{equation} (9) where the elasticity of substitution between goods σs varies across sectors. Although in the data we observe a finite number of goods within sectors, we adopt the theoretical assumption of a continuum of goods for reasons of tractability, because it enables us to make use of law of large numbers results in determining specialization at the sectoral level. Goods can be either substitutes (σs > 1) or complements (0 < σs < 1) and we can allow any ranking of the elasticities of substitution between goods and sectors, although the conventional assumption in such a nested CES structure is a higher elasticity of substitution at the more disaggregated level (σs > β). Expenditure on residential land in each location is assumed to be redistributed lump-sum to residents of that location, as in Helpman (1998). Therefore, total income in each location (vnLn) equals payments to labor used in production (wnLn) plus expenditure on residential land (rnHn = (1 − α)vnLn): \begin{equation} v_{n}L_{n}=w_{n}L_{n}+\left( 1-\alpha \right) v_{n}L_{n}=\frac{w_{n}L_{n}}{\alpha }, \end{equation} (10) where vn is income per worker; wn is the wage; rn is the land rent; Ln is the population of location n; and equilibrium land rents in each location are determined by land market clearing. 4.2. Production Goods are homogeneous in the sense that one unit of a given good is the same as any other unit of that good. Production occurs under conditions of perfect competition and constant returns to scale. The cost to a consumer in location n of purchasing one unit of good j within sector s from location i is therefore \begin{equation} p_{nis}(j)=\frac{d_{nis}G_{is}(j)}{z_{is}(j)}, \end{equation} (11) where dnis are iceberg goods trade costs, such that dnis > 1 must be shipped from location i to location n within sector s in order for one unit to arrive; zis(j) is productivity for good j within sector s in location i; and Gis(j) is the unit cost of the composite factor of production used for good j within sector s in location i, as determined in what follows. Final goods productivity is stochastic and modeled as in Eaton and Kortum (2002) and Costinot, Donaldson, and Komunjer (2012). Final goods productivity for each good, sector and location is assumed to be drawn independently from a Fréchet distribution:25 \begin{equation} F_{is}(z)=e^{-T_{is}L_{is}^{\eta _{s}}z^{\theta _{s}}}, \end{equation} (12) where θs > 1 is the Fréchet shape parameter that controls the dispersion of productivity across goods within each sector; Tis is a scale parameter that controls average productivity for each sector s and location i;Lis is employment in sector s and location i; and ηs parameterizes the strength of agglomeration forces in sector s. We model these agglomeration forces as external economies of scale (as in Ethier 1982), which imply that average productivity in sector s in location i is increasing in employment in that sector and location. We assume that the final good for each sector is produced using a number of stages of production, where each stage of production within a sector is supplied by a separate occupation indexed by o ∈ Os (e.g., Managers, Operatives). Output of good j within sector s in location i (yis(j)) is a CES function of the inputs of each occupation (Xiso(j)): \begin{equation} y_{is}(j)=\left[ \sum _{o \in O_{s}} X_{iso}(j)^{\frac{\mu _{s}-1}{\mu _{s}}} \right] ^{\frac{\mu _{s}}{\mu _{s}-1}}, \end{equation} (13) where μs is the elasticity of substitution between occupations and again we can allow occupations to be either substitutes (μs > 1) or complements (0 < μs < 1). We allow sectors to differ in terms of the set of occupations Os, and firms within each sector adjust the proportions with which workers in different occupations are employed depending their cost. Workers within each occupation perform a continuum of tasks t ∈ [0, 1] as in Grossman and Rossi-Hansberg (2008) (e.g., as captured by the verbs Advising, Typing, Stretching, Stamping in our empirical analysis). The input for occupation o and good j within sector s and location i (Xiso(j)) is a CES function of the inputs for these tasks (xiso(j, t)): \begin{equation} X_{iso}(j)=\left[ \int _{0}^{1}x_{iso}(j,t)^{\frac{\nu _{so}-1}{\nu _{so}}}dt \right] ^{\frac{\nu _{so}}{\nu _{so}-1}}, \end{equation} (14) where the elasticity of substitution between tasks νso varies across sectors and occupations. Although in the data we observe a finite number of tasks within occupations, we again adopt the theoretical assumption of a continuum of tasks for reasons of tractability, because it enables us to make use of law of large numbers results in determining specialization at the occupational level.26 We allow tasks within occupations to be either substitutes (νso > 1) or complements (0 < νso < 1), and we can consider any ranking of the elasticities of substitution between tasks and occupations, although the conventional assumption in such a nested CES structure is again a higher elasticity of substitution at the more disaggregated level (νso > μs).27 Tasks are performed by labor using a constant returns to scale technology and can be traded between locations. For example, product design can be undertaken in one location, whereas production and assembly occur in another location. The cost to a firm in location n of sourcing a task t from location i within occupation o and sector s is \begin{equation} g_{niso}(j,t)=\frac{\tau _{niso}w_{i}}{a_{iso}(j,t)}, \end{equation} (15) where wi is the wage; τniso are iceberg task trade costs, such that τniso > 1 units of the task must be performed in location i in order for one unit to be completed in location n for occupation o and sector s;aiso(j, t) is productivity for task t and good j within occupation o and sector s in location i. Input productivity for each task, occupation, sector, and location is also stochastic and is assumed to be drawnx independently from a Fréchet distribution: \begin{equation} \mathcal {F}_{iso} = e^{-U_{iso} L_{iso}^{\chi _{so}} a^{-\epsilon _{so}}}, \end{equation} (16) where εso > 1 is the Fréchet shape parameter that controls the dispersion of productivity across tasks within occupations; Uiso is a scale parameter that controls average productivity for each occupation o, sector s and location i;Liso is employment in occupation o, sector s, and location i; and χso parameterizes the strength of agglomeration forces in occupation o and sector s. We again model these agglomeration forces as external economies of scale (as in Grossman and Rossi-Hansberg 2012), which imply that average productivity in occupation o and sector s in location i is increasing in employment in that occupation, sector and location. 4.3. Trade in Tasks Firms within a given location n source each task t within an occupation o, good j, and sector s from the lowest cost source of supply for that task, \begin{equation*} g_{nso}(j,t)= \min \left\lbrace g_{niso}(j,t) ; i \in N \right\rbrace . \end{equation*} Given finite task trade costs, locations supply tasks for which they have high productivity draws themselves, and source other tasks for which they have low productivity draws from other locations. Using our assumption of a Fréchet distribution of input productivity, the share of firm costs in location n accounted for by tasks sourced from location i within occupation o and sector s (λniso) is equal to the fraction of tasks sourced from that location,28 \begin{equation} \lambda _{niso}=\frac{U_{iso} L_{iso}^{\chi _{so}} \left( \tau _{niso}w_{i}\right) ^{-\epsilon _{so}}}{\sum _{k \in N} U_{kso} L_{kso}^{\chi _{so}} \left( \tau _{nkso}w_{k}\right) ^{-\epsilon _{so}}}. \end{equation} (17) Intuitively, the share of firm costs accounted for by tasks from each location depends on production costs in each location (as determined by wages wi, the exogenous productivity parameter Uiso, and the endogenous component of productivity from agglomeration forces $$L_{iso}^{\chi _{so}}$$) and the bilateral costs of trading tasks (τniso). 4.4. Trade in Final Goods Consumers within a given location n source each final good j within a sector s from the lowest cost source of supply for that final good \begin{equation*} p_{ns}(j)=\min \left\lbrace p_{nis}(j);i\in N\right\rbrace . \end{equation*} For finite final goods trade costs, locations supply final goods for which they have low unit costs themselves, and source other final goods for which they have high units costs from other locations. These unit costs for final goods depend on input productivities and trade in tasks, as characterized in the previous section, as well on final goods productivities. Using our assumption of a Fréchet distribution of final goods productivity, the share of location n’s expenditure on final goods produced in location i within sector s (πnis) is equal to the fraction of final goods sourced from that location:29 \begin{equation} \pi _{nis}=\frac{T_{is}L_{is}^{\eta _{s}}\left( d_{nis}\Phi _{is}w_{i}\right) ^{-\theta _{s}}}{\sum _{k\in N}T_{ks}L_{ks}^{\eta _{s}}\left( d_{nks}\Phi _{ks}w_{k}\right) ^{-\theta _{s}}}, \end{equation} (18) where Φis is a summary statistic for the unit costs of sourcing the tasks for all occupations in sector s in location i, as derived in the Online Appendix. Intuitively, the share of expenditure on final goods from each location within a given sector depends on production costs in each location (as determined by wages wi, the unit cost summary statistic Φis, the exogenous productivity parameter Tis, and the endogenous component of productivity from agglomeration forces $$L_{is}^{\eta _{s}}$$) and the bilateral costs of trading final goods (dnis). 4.5. Population Mobility Population mobility implies that workers must receive the same indirect utility in all populated locations: \begin{equation} V_{n}=\frac{v_{n}}{P_{n}^{\alpha }r_{n}^{1-\alpha }}=\bar{V}, \end{equation} (19) where indirect utility depends on per capita income (vn), the consumption goods price index (Pn), and land rents (rn). Labor market clearing requires that the sum of total employment across all locations equals the economy’s labor supply \begin{equation} \sum _{n\in N}L_{n}=\bar{L}. \end{equation} (20) In the Online Appendix, we provide a further characterization of the general equilibrium of the model, including the distribution of population across locations (Ln) and each location’s own trade share for final goods (πnns) and tasks (λnnso). 4.6. Reductions in Transport and Communication Costs The distribution of employment across occupations, sectors and locations in the model is determined by two sets of forces: productivity differences (which depend on both an exogenous component and an endogenous component through agglomeration forces) and the costs of trading both tasks and final goods. Together these two sets of forces determine comparative advantages across occupations within sectors and across sectors. Patterns of comparative advantage across occupations within sectors can be characterized by a double difference for a given import market. The first difference computes the ratio of exports of tasks from two locations i and k in a third market n in a single occupation; the second difference compares this ratio of exports of tasks for two separate occupations o and m. Taking this double difference in the unit cost share in equation (17), we obtain \begin{eqnarray} \frac{\lambda _{niso}/\lambda _{nkso}}{\lambda _{nism}/\lambda _{nksm}}&=& \frac{\left[ U_{iso} L_{iso}^{\chi _{so}} \left( \tau _{niso}w_{i}\right) ^{-\epsilon _{so}}\right] /\left[ U_{kso} L_{kso}^{\chi _{so}} \left( \tau _{nkso}w_{k}\right) ^{-\epsilon _{so}}\right] }{\left[ U_{ism} L_{ism}^{\chi _{so}} \left( \tau _{nism}w_{i}\right) ^{-\epsilon _{sm}}\right] /\left[ U_{ksm} L_{ksm}^{\chi _{so}} \left( \tau _{nksm}w_{k}\right) ^{-\epsilon _{sm}}\right] }. \nonumber\\ \end{eqnarray} (21) Therefore, a location i specializes more in occupation o relative to occupation m compared to another location k when it has lower production costs (as determined by wages wi, the exogenous productivity parameter Uiso, and the endogenous component of productivity from agglomeration forces $$L_{iso}^{\chi _{so}}$$) and lower bilateral costs of trading tasks (as determined by τniso). Patterns of comparative advantage across sectors can be characterized by an analogous double difference for a given import market. The first difference computes the ratio of exports of final goods from two locations i and k in a third market n in a single sector; the second difference compares this ratio of exports of final goods for two separate sectors s and r. Taking this double difference in the expenditure share in equation (18), we obtain \begin{equation} \frac{\pi _{nis}/\pi _{nks}}{\pi _{nir}/\pi _{nkr}}=\frac{\left[ T_{is} L_{is}^{\eta _{s}} \left( d_{nis}\Phi _{is}w_{i}\right) ^{-\theta _{s}}\right] /\left[ T_{ks} L_{ks}^{\eta _{s}} \left( d_{nks}\Phi _{ks}w_{k}\right) ^{-\theta _{s}}\right] }{\left[ T_{ir} L_{ir}^{\eta _{r}} \left( d_{nir}\Phi _{ir}w_{i}\right) ^{-\theta _{r}}\right] /\left[ T_{kr} L_{kr}^{\eta _{r}} \left( d_{nkr}\Phi _{kr}w_{k}\right) ^{-\theta _{r}}\right] }. \end{equation} (22) Therefore, a location i specializes more in sector s relative to sector r compared to another location k when it has lower production costs (as determined by wages wi, the unit cost summary statistic Φis, the exogenous productivity parameter Tis, and the endogenous component of productivity from agglomeration forces $$L_{is}^{\eta _{s}}$$) and lower bilateral costs of trading final goods (as determined by dnis). When the costs of trading tasks and final goods are large, all locations have similar employment structures across sectors, and all tasks within each sector are undertaken in the same location where the final good is produced. As the costs of trading final goods and tasks fall, locations specialize across sectors and across occupations within sectors according to their comparative advantage as determined by productivity differences.30 If densely-populated urban locations have a comparative advantage in interactive tasks relative to sparsely-populated rural locations (e.g., as argued in Gaspar and Glaeser 1998), the model predicts that a fall in the costs of trading tasks leads to an increase in the interactiveness of employment within sectors in urban relative to rural areas. Therefore, the model highlights the role of reductions in communication costs (e.g., telephones) and improvements in transport technologies (e.g., roads and the automobile) in influencing the interactiveness of employment. In interpreting these predictions, several caveats are relevant. In particular, the model implies that relative productivities across occupations, sectors, and locations are an important determinant of comparative advantage across occupations within sectors and across sectors (and hence matter for the interactiveness of employment). Additionally, our long historical time period includes other changes that could have influenced the relative importance of different occupations, including the extent to which technological change was skill-biased. In the historical literature, several studies argue that technical change often replaced—rather than complemented—skilled artisans in the 19th century, including Hounshell (1985), James and Skinner (1985), and Mokyr (1992). However, there remains substantial debate about the extent to which this was the case. In their classic study of the race between technology and skills, Goldin and Katz (2008) present evidence that manufacturing technologies were skill complementary in the early-20th century, but may have been skill substituting prior to that time.31 In subsequent work, Katz and Margo (2014) report some evidence of deskilling in manufacturing during the 19th century, but find a reallocation of employment toward high-skill jobs for the aggregate economy as a whole.32 5. Explaining Increased Interactiveness Having established a robust increase in the interactiveness of employment in urban areas relative to rural areas, and having developed a model to interpret these results in terms of specialization according to comparative advantage, we now provide further evidence on explanations for the observed changes in interactiveness. First, we decompose the overall change in interactiveness into the contributions of individual occupations and sectors, which enables us to explore explanations that emphasize particular occupations and sectors. Second, we report regression specifications using variation in interactiveness between sectors, within sectors, and within sectors and occupations over time. Using these regressions, we explore the importance of the constituent components of interactiveness (thought, communication and intersocial) and present evidence on a number of potential explanations. Third, we provide evidence on the relationship between changes in interactiveness and the dissemination of new communication and transport technologies. 5.1. Decomposing Interactiveness We begin by decomposing the change in the overall interactiveness of metro and nonmetro areas into the contributions of each two-digit occupation and sector. Overall interactiveness for metro and nonmetro areas is the employment-weighted average of interactiveness for each two-digit-sector-occupation cell: \begin{equation} I_{jt}=\sum _{z\in \Omega }\sum _{o\in \Omega _{z}}\frac{E_{ojt}}{E_{jt}} I_{o},\qquad j\in \left\lbrace M,N\right\rbrace, \end{equation} (23) where z indexes two-digit-sector-occupation cells; o indexes disaggregated three-digit occupations within these cells; and t indexes time; Ω is the set of two-digit-sector-occupation cells; Ωz is the set of three-digit occupations within each cell z; the interactiveness of each three-digit occupation is measured using (5) based on the time-invariant occupational descriptions from the 1991 DOTs. Taking differences between times T and t > T, the change in the overall interactiveness of metro and nonmetro areas can be decomposed as follows: \begin{equation} \triangle I_{jt}=\sum _{z\in \Omega }\sum _{o\in \Omega _{z}}\left[ \triangle \left( \frac{E_{ojt}}{E_{jt}} \right) \right] I_{o},\qquad j\in \left\lbrace M,N\right\rbrace , \end{equation} (24) where ▵Ijt = Ijt − IjT; ▵(Eojt/Ejt) is the change in the employment share of occupation o in location j ∈ {M, N}; and we have used the fact that occupation interactiveness is constant over time. Taking differences again between metro and nonmetro areas, we obtain an analogous decomposition of the change in the relative interactiveness of metro and nonmetro areas: \begin{equation} \triangle I_{Mt}-\triangle I_{Nt}=\sum _{z\in \Omega }\sum _{o\in \Omega _{z}} \left[ \triangle \frac{E_{oMt}}{E_{Mt}}-\triangle \frac{E_{oNt}}{E_{Nt}} \right] I_{o}, \end{equation} (25) where the right-hand sides of the decompositions (24) and (25) are summations over the contributions from each two-digit-sector-occupation-cell. These contributions correspond to a matrix with two-digit sectors for rows and two-digit occupations for columns, where the right-hand side is a summation across both rows and columns. Metro areas display a larger increase in interactiveness than nonmetro areas to the extent that they experience a greater reallocation of employment shares toward high-interactiveness occupations. Figures 3 and 4 summarize the results from the decompositions of the change in the relative interactiveness of metro and nonmetro areas in equation (25). Figure 3 shows the contributions for each two-digit occupation (summing across sectors in the rows of the matrix of contributions) for each 20-year interval in our sample, whereas Figure 4 shows the corresponding contributions for each two-digit sector (summing across occupations in the columns of the matrix of contributions).33 Figures A.6 and A.7 in the Online Appendix report analogous results from the decompositions of the change in interactiveness for metro and nonmetro areas separately in equation (24). Figure 3. View largeDownload slide Decomposition of difference in change in interactiveness between metro and nonmetro areas, occupations. Decomposition of the difference between mean and nonmetro areas in the change in mean interactiveness over 20-year time periods (equation (25) in the paper) into the contributions of two-digit occupations. X-axes are 20-year differences. Y-axes are differences in the change in mean interactiveness between metro and nonmetro areas. Mean interactiveness is the employment-weighted average of interactiveness for each three-digit occupation. Interactiveness for each three-digit occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 3. View largeDownload slide Decomposition of difference in change in interactiveness between metro and nonmetro areas, occupations. Decomposition of the difference between mean and nonmetro areas in the change in mean interactiveness over 20-year time periods (equation (25) in the paper) into the contributions of two-digit occupations. X-axes are 20-year differences. Y-axes are differences in the change in mean interactiveness between metro and nonmetro areas. Mean interactiveness is the employment-weighted average of interactiveness for each three-digit occupation. Interactiveness for each three-digit occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 4. View largeDownload slide Decomposition of difference in change in interactiveness between metro and nonmetro areas, sectors. Decomposition of the difference between mean and nonmetro areas in the change in mean interactiveness over 20-year time periods (equation (25) in the paper) into the contributions of two-digit sectors. X-axes are 20-year differences. Y-axes are differences in the change in mean interactiveness between metro and nonmetro areas. Mean interactiveness is the employment-weighted average of interactiveness for each three-digit occupation. Interactiveness for each three-digit occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 4. View largeDownload slide Decomposition of difference in change in interactiveness between metro and nonmetro areas, sectors. Decomposition of the difference between mean and nonmetro areas in the change in mean interactiveness over 20-year time periods (equation (25) in the paper) into the contributions of two-digit sectors. X-axes are 20-year differences. Y-axes are differences in the change in mean interactiveness between metro and nonmetro areas. Mean interactiveness is the employment-weighted average of interactiveness for each three-digit occupation. Interactiveness for each three-digit occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 3 shows that the sharp increase in the relative interactiveness of metro areas from 1880 to 1920 is largely driven by positive contributions from Clerks, with Operatives, Sales Workers, and Managers all making negative contributions. From 1920 to 1960, Professionals (and to a lesser but growing extent Managers) make the largest positive contributions, whereas Craftsmen and Operatives make negative contributions. From 1960 to 2000, Professionals and Managers have the largest positive contributions, whereas Clerks have the largest negative contribution. Figure 4 shows that Professional and Business services are the two sectors that make the largest contributions to the increase in the relative interactiveness of metro areas over the sample period as a whole. Business Services make positive contributions toward the beginning and end of the sample period, whereas Professional Services become more important later on. Although the contribution from manufacturing is initially positive, it becomes negative later in the sample period. Taking these decomposition results together, the increase in the relative interactiveness of metro areas is not driven by any one occupation or sector. Our results are not solely explained by Managers (whose contribution only becomes positive toward the end of our sample period). Clerks and Professionals make notable positive contributions toward the beginning and end of our sample period respectively. Our results are also not simply driven by a decline of manufacturing in urban areas (indeed manufacturing makes a positive contribution in the early decades of our sample when some of the largest changes in interactiveness were observed). Similarly, our findings are not simply attributable to an expansion of services in urban areas (indeed services was a relatively small share of employment in the early decades of our sample when some of the largest changes in interactiveness were observed). Furthermore, our regression specifications stated previously include sector-year fixed effects, which control for common changes across all occupations within each sector. Therefore, our results cannot simply be explained by reallocation across sectors, and instead reflective a pervasive reallocation of employment toward more interactive occupations within sectors. 5.2. Variation within and between Sectors We now provide further evidence against explanations based on reallocation across sectors (such as the movement of manufacturing jobs out of cities or the concentration of service jobs in cities) by exploiting variation within versus between sectors. We begin by examining between-sector variation. We define sector interactiveness as the employment-weighted mean of the interactiveness of each occupation \begin{equation} \text{Interactive}_{st}=\sum _{o}\frac{E_{ost}}{E_{st}}\text{Interactive}_{o}. \end{equation} (26) Using this measure, we run a regression across sectors of the share of a sector’s employment in metro areas (MetroSharest) on its interactiveness (Interactivest) for each year separately: \begin{equation} \text{MetroShare}_{st}=\alpha _{t}\text{Interactive}_{st}+\varepsilon _{st}, \end{equation} (27) where εst is a stochastic error; αt captures the correlation between sectors’ shares of employment in metro areas and their interactiveness in each year. Although we estimate the previous regression and the remaining regressions in this section using a share as the left-hand side variable so that the estimated coefficients have a natural interpretation as frequencies, we again find a very similar pattern of results in a robustness test in which we use a logistic transformation of the left-hand side variable: MetroSharest/(1 − MetroSharest). Panel A of Table 5 reports the results, where each cell in the table corresponds to a separate regression. In 1880, there is a negative but statistically insignificant correlation between a sector’s metro employment share and its interactiveness. Starting in 1900, there is an increase in the correlation between a sector’s metro employment share and its interactiveness, which is particularly sharp from 1900 to 1940, and becomes positive and statistically significant at conventional critical values in 1960. Therefore, more interactive sectors become increasingly concentrated in metro areas over time. Although the estimated coefficients on the interactiveness measure do not become significant until 1960, there is a relatively constant increase in the value of these coefficients across decades, as also shown in Figure 1. Table 5. Metro employment and wagebill shares and interactiveness. LHS Measure 1880 1900 1920 1940 1960 1980 2000 Panel A: Between sectors Employment Interactiveness − 0.130 − 0.132 0.258 0.556 0.728*** 0.901*** 0.814*** (0.267) (0.239) (0.419) (0.405) (0.267) (0.200) (0.182) Employment Thought − 0.722*** − 1.293*** − 1.806*** − 0.622 0.190 0.788*** 1.202*** (0.260) (0.261) (0.357) (0.487) (0.310) (0.274) (0.237) Employment Communication − 0.459*** − 0.582*** − 0.645*** − 0.220 0.210 0.360* 0.530** (0.146) (0.151) (0.186) (0.266) (0.193) (0.208) (0.233) Employment Intersocial − 0.351** − 0.481*** − 0.599*** − 0.117 0.101 0.268** 0.342*** (0.135) (0.135) (0.165) (0.209) (0.133) (0.122) (0.109) Employment Individual volition − 0.157*** − 0.195*** − 0.268*** − 0.212*** − 0.115** 0.019 0.085 (0.051) (0.054) (0.079) (0.059) (0.054) (0.054) (0.062) Wagebill Interactiveness 0.557 0.557* 0.814*** 0.733*** (0.366) (0.283) (0.215) (0.201) Panel B: Within sectors Employment Interactiveness − 0.410*** − 0.261** − 0.104 − 0.036 0.190*** 0.274*** 0.317*** (0.120) (0.119) (0.119) (0.119) (0.064) (0.051) (0.040) Employment Thought − 0.340** − 0.411*** − 0.299*** − 0.145 0.153*** 0.227*** 0.246*** (0.134) (0.132) (0.093) (0.095) (0.049) (0.037) (0.039) Employment Communication − 0.041 − 0.042 0.025 0.118 0.183*** 0.168*** 0.140*** (0.144) (0.118) (0.098) (0.079) (0.036) (0.032) (0.038) Employment Intersocial − 0.030 − 0.081 − 0.017 0.0197 0.105*** 0.0652* 0.046 (0.130) (0.078) (0.058) (0.049) (0.032) (0.034) (0.048) Employment Individual volition − 0.095* − 0.058 − 0.021 − 0.016 0.006 0.015 0.027** (0.056) (0.070) (0.054) (0.039) (0.025) (0.016) (0.013) Wagebill Interactiveness 0.0430 0.207*** 0.281*** 0.311*** (0.0874) (0.0529) (0.0433) (0.0374) Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes LHS Measure 1880 1900 1920 1940 1960 1980 2000 Panel A: Between sectors Employment Interactiveness − 0.130 − 0.132 0.258 0.556 0.728*** 0.901*** 0.814*** (0.267) (0.239) (0.419) (0.405) (0.267) (0.200) (0.182) Employment Thought − 0.722*** − 1.293*** − 1.806*** − 0.622 0.190 0.788*** 1.202*** (0.260) (0.261) (0.357) (0.487) (0.310) (0.274) (0.237) Employment Communication − 0.459*** − 0.582*** − 0.645*** − 0.220 0.210 0.360* 0.530** (0.146) (0.151) (0.186) (0.266) (0.193) (0.208) (0.233) Employment Intersocial − 0.351** − 0.481*** − 0.599*** − 0.117 0.101 0.268** 0.342*** (0.135) (0.135) (0.165) (0.209) (0.133) (0.122) (0.109) Employment Individual volition − 0.157*** − 0.195*** − 0.268*** − 0.212*** − 0.115** 0.019 0.085 (0.051) (0.054) (0.079) (0.059) (0.054) (0.054) (0.062) Wagebill Interactiveness 0.557 0.557* 0.814*** 0.733*** (0.366) (0.283) (0.215) (0.201) Panel B: Within sectors Employment Interactiveness − 0.410*** − 0.261** − 0.104 − 0.036 0.190*** 0.274*** 0.317*** (0.120) (0.119) (0.119) (0.119) (0.064) (0.051) (0.040) Employment Thought − 0.340** − 0.411*** − 0.299*** − 0.145 0.153*** 0.227*** 0.246*** (0.134) (0.132) (0.093) (0.095) (0.049) (0.037) (0.039) Employment Communication − 0.041 − 0.042 0.025 0.118 0.183*** 0.168*** 0.140*** (0.144) (0.118) (0.098) (0.079) (0.036) (0.032) (0.038) Employment Intersocial − 0.030 − 0.081 − 0.017 0.0197 0.105*** 0.0652* 0.046 (0.130) (0.078) (0.058) (0.049) (0.032) (0.034) (0.048) Employment Individual volition − 0.095* − 0.058 − 0.021 − 0.016 0.006 0.015 0.027** (0.056) (0.070) (0.054) (0.039) (0.025) (0.016) (0.013) Wagebill Interactiveness 0.0430 0.207*** 0.281*** 0.311*** (0.0874) (0.0529) (0.0433) (0.0374) Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Notes: Each cell of each panel of the table corresponds to a separate regression. Coefficients estimated from a regression of the share of either employment or the wagebill in metro areas on the frequency with which the verbs from occupational descriptions appear in a thesaurus section; the wagebill data are only available from 1940 onward; the frequency with which verbs appear in a thesaurus section is measured using time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs); Interactiveness is the frequency with which verbs from occupational descriptions appear in Class IV, Division 1, Class IV, Division 2 and Class V, Division 2 of the thesaurus; thought is the frequency with which verbs appear in Class IV (Division 1) of the thesaurus; Communication is the frequency with which verbs appear in Class IV (Division 2) of the thesaurus; Intersocial is the frequency with which verbs appear in Class V (Division 2) of the thesaurus; Individual volition is the frequency with which verbs appear in Class V (Division 1). In Panel A, observations are three-digit sectors for each year, the frequency of verb use for each sector is the employment-weighted average of the frequency for occupations within that sector, and standard errors are heteroskedasticity robust (equation (27) in the paper). In Panel B, observations are three-digit sectors and occupations for each year, three-digit sector-year fixed effects are included, and standard errors in Panel B are heteroskedasticity robust and clustered on occupation (equation (28) in the paper). *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 5. Metro employment and wagebill shares and interactiveness. LHS Measure 1880 1900 1920 1940 1960 1980 2000 Panel A: Between sectors Employment Interactiveness − 0.130 − 0.132 0.258 0.556 0.728*** 0.901*** 0.814*** (0.267) (0.239) (0.419) (0.405) (0.267) (0.200) (0.182) Employment Thought − 0.722*** − 1.293*** − 1.806*** − 0.622 0.190 0.788*** 1.202*** (0.260) (0.261) (0.357) (0.487) (0.310) (0.274) (0.237) Employment Communication − 0.459*** − 0.582*** − 0.645*** − 0.220 0.210 0.360* 0.530** (0.146) (0.151) (0.186) (0.266) (0.193) (0.208) (0.233) Employment Intersocial − 0.351** − 0.481*** − 0.599*** − 0.117 0.101 0.268** 0.342*** (0.135) (0.135) (0.165) (0.209) (0.133) (0.122) (0.109) Employment Individual volition − 0.157*** − 0.195*** − 0.268*** − 0.212*** − 0.115** 0.019 0.085 (0.051) (0.054) (0.079) (0.059) (0.054) (0.054) (0.062) Wagebill Interactiveness 0.557 0.557* 0.814*** 0.733*** (0.366) (0.283) (0.215) (0.201) Panel B: Within sectors Employment Interactiveness − 0.410*** − 0.261** − 0.104 − 0.036 0.190*** 0.274*** 0.317*** (0.120) (0.119) (0.119) (0.119) (0.064) (0.051) (0.040) Employment Thought − 0.340** − 0.411*** − 0.299*** − 0.145 0.153*** 0.227*** 0.246*** (0.134) (0.132) (0.093) (0.095) (0.049) (0.037) (0.039) Employment Communication − 0.041 − 0.042 0.025 0.118 0.183*** 0.168*** 0.140*** (0.144) (0.118) (0.098) (0.079) (0.036) (0.032) (0.038) Employment Intersocial − 0.030 − 0.081 − 0.017 0.0197 0.105*** 0.0652* 0.046 (0.130) (0.078) (0.058) (0.049) (0.032) (0.034) (0.048) Employment Individual volition − 0.095* − 0.058 − 0.021 − 0.016 0.006 0.015 0.027** (0.056) (0.070) (0.054) (0.039) (0.025) (0.016) (0.013) Wagebill Interactiveness 0.0430 0.207*** 0.281*** 0.311*** (0.0874) (0.0529) (0.0433) (0.0374) Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes LHS Measure 1880 1900 1920 1940 1960 1980 2000 Panel A: Between sectors Employment Interactiveness − 0.130 − 0.132 0.258 0.556 0.728*** 0.901*** 0.814*** (0.267) (0.239) (0.419) (0.405) (0.267) (0.200) (0.182) Employment Thought − 0.722*** − 1.293*** − 1.806*** − 0.622 0.190 0.788*** 1.202*** (0.260) (0.261) (0.357) (0.487) (0.310) (0.274) (0.237) Employment Communication − 0.459*** − 0.582*** − 0.645*** − 0.220 0.210 0.360* 0.530** (0.146) (0.151) (0.186) (0.266) (0.193) (0.208) (0.233) Employment Intersocial − 0.351** − 0.481*** − 0.599*** − 0.117 0.101 0.268** 0.342*** (0.135) (0.135) (0.165) (0.209) (0.133) (0.122) (0.109) Employment Individual volition − 0.157*** − 0.195*** − 0.268*** − 0.212*** − 0.115** 0.019 0.085 (0.051) (0.054) (0.079) (0.059) (0.054) (0.054) (0.062) Wagebill Interactiveness 0.557 0.557* 0.814*** 0.733*** (0.366) (0.283) (0.215) (0.201) Panel B: Within sectors Employment Interactiveness − 0.410*** − 0.261** − 0.104 − 0.036 0.190*** 0.274*** 0.317*** (0.120) (0.119) (0.119) (0.119) (0.064) (0.051) (0.040) Employment Thought − 0.340** − 0.411*** − 0.299*** − 0.145 0.153*** 0.227*** 0.246*** (0.134) (0.132) (0.093) (0.095) (0.049) (0.037) (0.039) Employment Communication − 0.041 − 0.042 0.025 0.118 0.183*** 0.168*** 0.140*** (0.144) (0.118) (0.098) (0.079) (0.036) (0.032) (0.038) Employment Intersocial − 0.030 − 0.081 − 0.017 0.0197 0.105*** 0.0652* 0.046 (0.130) (0.078) (0.058) (0.049) (0.032) (0.034) (0.048) Employment Individual volition − 0.095* − 0.058 − 0.021 − 0.016 0.006 0.015 0.027** (0.056) (0.070) (0.054) (0.039) (0.025) (0.016) (0.013) Wagebill Interactiveness 0.0430 0.207*** 0.281*** 0.311*** (0.0874) (0.0529) (0.0433) (0.0374) Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Notes: Each cell of each panel of the table corresponds to a separate regression. Coefficients estimated from a regression of the share of either employment or the wagebill in metro areas on the frequency with which the verbs from occupational descriptions appear in a thesaurus section; the wagebill data are only available from 1940 onward; the frequency with which verbs appear in a thesaurus section is measured using time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs); Interactiveness is the frequency with which verbs from occupational descriptions appear in Class IV, Division 1, Class IV, Division 2 and Class V, Division 2 of the thesaurus; thought is the frequency with which verbs appear in Class IV (Division 1) of the thesaurus; Communication is the frequency with which verbs appear in Class IV (Division 2) of the thesaurus; Intersocial is the frequency with which verbs appear in Class V (Division 2) of the thesaurus; Individual volition is the frequency with which verbs appear in Class V (Division 1). In Panel A, observations are three-digit sectors for each year, the frequency of verb use for each sector is the employment-weighted average of the frequency for occupations within that sector, and standard errors are heteroskedasticity robust (equation (27) in the paper). In Panel B, observations are three-digit sectors and occupations for each year, three-digit sector-year fixed effects are included, and standard errors in Panel B are heteroskedasticity robust and clustered on occupation (equation (28) in the paper). *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Panel A of Table 5 also breaks out overall interactiveness into its three components of thought (Class IV, Division 1), communication (Class IV, Division 2), and intersocial (Class V, Division 2). We also include for comparison individual volition (Class V, Division 1). Interpreting the magnitudes of the coefficients for these measures in Table 5 involves taking into account both the estimated coefficients and the standard deviations of each thesaurus measure. In Table A.6 of the Online Appendix, we report the corresponding beta coefficients, which scale the estimated coefficients using the standard deviations of the dependent and independent variables. We find a broad increase in the concentration of each of the components of interactiveness in urban areas, which is larger for thought and communication and smaller for intersocial volition. The increase in the beta coefficients for our overall measure of interactiveness can be larger or smaller than for its individual components, depending on the correlation in verb use across these components and the relative standard deviations of the verbs in each component. In principle, these changes in patterns of specialization in metro versus nonmetro areas could be explained in terms of relative demand or relative supply. An increase in the relative demand for an occupation raises both its employment and its wage (and hence raises its wagebill). In contrast, an increase in the relative supply of an occupation raises its employment but reduces its wage (and hence reduces its wagebill if the demand for occupations is inelastic). To assess the relative importance of these two explanations, Panel A of Table 5 also reports the results of regressions in which we measure sector interactiveness using wagebill shares instead of employment shares in equation (26). Although the wage data are available for a much shorter time period than the employment data, we find a similar pattern of results using this alternative left-hand side variable, which is consistent with relative demand moving relative wagebills and employment in the same direction. Having established these relationships between sectors, we next examine within-sector variation. We run a regression across sectors and occupations of the share of a sector-occupation’s employment in metro areas (MetroShareost) on occupation interactiveness (Interactiveo) for each year separately: \begin{equation} \text{Metro}_{ost}=\alpha _{t}\text{Interactive}_{o}+\eta _{st}+\varepsilon _{ost}, \end{equation} (28) where ηst are sector-year fixed effects and εost is a stochastic error. The sector-year fixed effects (ηst) control for changes in sector composition over time, so that the coefficient αt is identified solely from variation within sectors. The coefficient αt captures the within-sector correlation between the share of employment in metro areas and the interactiveness of occupations. Panel B of Table 5 reports the results, where each cell in the table again corresponds to a separate regression. In line with our previous results, the correlation between metro employment shares and interactiveness is negative in 1880. Over time, this correlation becomes more positive and becomes statistically significant by 1960. Therefore, within sectors, more interactive occupations become increasingly concentrated in metro areas over time. This finding of the same pattern of reallocation across occupations both between and within sectors is consistent with a wide-ranging secular process favoring specialization in interactive occupations in metro areas. Panel B of Table 5 also breaks out overall interactiveness into its three components of thought, communication and intersocial volition. We again also include individual volition for comparison. We find a similar pattern of results as in Panel A, with an increase in coefficient for all three components of interactiveness. Panel B of Table 5 also reports the results of regressions in which we measure sector interactiveness using wagebill shares instead of employment shares in equation (26). For the shorter period over which we have the wage data, we again find a similar pattern of results using this alternative left-hand side variable, which is consistent with relative demand moving relative wagebills and employment in the same direction. Finally, to use variation within sectors and occupations, we pool our sector-occupation data over time and estimate a panel data regression that facilitates the inclusion of sector, occupation and year fixed effects. We regress the share of a sector-occupation’s employment in metro areas on these fixed effects and interaction terms between time dummies and our measure of occupation interactiveness: \begin{equation} \text{MetroShare}_{ost}=\alpha _{t}\left[ \text{Interactive}_{o}\times \text{Year}_{t}\right] +\mu _{o}+\eta _{s}+\delta _{t}+\varepsilon _{ost}, \end{equation} (29) where εost is a stochastic error; we choose 1880 as the excluded year from the interaction terms. The occupation fixed effects (μo) control for time-invariant differences between metro and nonmetro areas in the share of an occupation in employment and capture the main effect of occupation interactiveness. The sector fixed effects (ηs) control for time-invariant differences between metro and nonmetro areas in the share of a sector in employment. The year fixed effects (δt) control for changes in the shares of metro areas in employment across all occupations and sectors. The coefficients αt capture the change in the correlation between metro employment shares and interactiveness relative to 1880. Table 6 reports the estimation results. Column (1) confirms our previous findings of an increasing correlation between metro employment shares and occupation interactiveness over time, which becomes positive and statistically significant by 1960. As shown in column (2), this increasing correlation between metro employment shares and occupation interactiveness is robust to replacing the sector and year fixed effects with sector-year fixed effects to control for changes in sector composition over time. These sector-year fixed effects also control for changes in the metropolitan area status of counties because of changes in sectoral specialization. Table 6. Metro area employment shares and interactiveness, within-sector and within-occupation. Metro employment share (1) (2) (3) (4) (5) (6) (7) (8) Interactiveness × 1900 − 0.261** 0.104 − 0.001 0.260 0.076 0.126 − 0.032 0.119 (0.117) (0.162) (0.121) (0.224) (0.159) (0.198) (0.092) (0.158) Interactiveness × 1920 − 0.104 0.187 0.019 0.428 0.328 0.214 − 0.012 0.218 (0.118) (0.218) (0.198) (0.273) (0.198) (0.234) (0.102) (0.203) Interactiveness × 1940 − 0.04 0.321 0.177 0.534* 0.424** 0.405 0.012 0.409 (0.119) (0.235) (0.231) (0.286) (0.210) (0.251) (0.124) (0.157) Interactiveness × 1960 0.190** 0.485*** 0.331* 0.756*** 0.578*** 0.563*** (0.064) (0.185) (0.180) (0.243) (0.200) (0.215) Interactiveness × 1980 0.274*** 0.560*** 0.449*** 0.777*** 0.658*** 0.651*** 0.295*** 0.634*** (0.052) (0.174) (0.168) (0.231) (0.191) (0.210) (0.090) (0.160) Interactiveness × 2000 0.317*** 0.596*** 0.478*** 0.798*** 0.794*** 0.697*** 0.339*** 0.684*** (0.040) (0.174) (0.169) (0.233) (0.196) (0.227) (0.091) (0.157) Observations 56,760 56,760 50,180 42,460 23,189 31,133 38,647 42,653 Occupation fixed effects Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Married only sample Yes Single only sample Yes Manufacturing only Yes Services only Yes No more skilled MSAs Yes No less skilled MSAs Yes Metro employment share (1) (2) (3) (4) (5) (6) (7) (8) Interactiveness × 1900 − 0.261** 0.104 − 0.001 0.260 0.076 0.126 − 0.032 0.119 (0.117) (0.162) (0.121) (0.224) (0.159) (0.198) (0.092) (0.158) Interactiveness × 1920 − 0.104 0.187 0.019 0.428 0.328 0.214 − 0.012 0.218 (0.118) (0.218) (0.198) (0.273) (0.198) (0.234) (0.102) (0.203) Interactiveness × 1940 − 0.04 0.321 0.177 0.534* 0.424** 0.405 0.012 0.409 (0.119) (0.235) (0.231) (0.286) (0.210) (0.251) (0.124) (0.157) Interactiveness × 1960 0.190** 0.485*** 0.331* 0.756*** 0.578*** 0.563*** (0.064) (0.185) (0.180) (0.243) (0.200) (0.215) Interactiveness × 1980 0.274*** 0.560*** 0.449*** 0.777*** 0.658*** 0.651*** 0.295*** 0.634*** (0.052) (0.174) (0.168) (0.231) (0.191) (0.210) (0.090) (0.160) Interactiveness × 2000 0.317*** 0.596*** 0.478*** 0.798*** 0.794*** 0.697*** 0.339*** 0.684*** (0.040) (0.174) (0.169) (0.233) (0.196) (0.227) (0.091) (0.157) Observations 56,760 56,760 50,180 42,460 23,189 31,133 38,647 42,653 Occupation fixed effects Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Married only sample Yes Single only sample Yes Manufacturing only Yes Services only Yes No more skilled MSAs Yes No less skilled MSAs Yes Notes: Each column of the table corresponds to a separate regression (equation (29) in the paper). Estimation sample is a panel of observations on three-digit occupations, three-digit sectors and years for 20-year time periods from 1880–2000; 1880 is the excluded year from the interaction terms; Interactiveness is the frequency with which verbs from 1991 occupational descriptions appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Married only sample includes married workers only. Single only sample excludes married workers. Manufacturing only sample includes workers in manufacturing only. Services only sample includes workers in services only. More and less-skilled metro areas are defined as in Glaeser and Resseger (2010), based on whether the share of adults with a college degree in a Metropolitan Statistical Area (MSA) is greater than or less than 25.025% in 2006. The year 1960 is omitted in columns (7) and (8) because the IPUMS 1960 data do not contain the identifiers for individual MSAs. Standard errors are heteroskedasticity robust and clustered on occupation. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 6. Metro area employment shares and interactiveness, within-sector and within-occupation. Metro employment share (1) (2) (3) (4) (5) (6) (7) (8) Interactiveness × 1900 − 0.261** 0.104 − 0.001 0.260 0.076 0.126 − 0.032 0.119 (0.117) (0.162) (0.121) (0.224) (0.159) (0.198) (0.092) (0.158) Interactiveness × 1920 − 0.104 0.187 0.019 0.428 0.328 0.214 − 0.012 0.218 (0.118) (0.218) (0.198) (0.273) (0.198) (0.234) (0.102) (0.203) Interactiveness × 1940 − 0.04 0.321 0.177 0.534* 0.424** 0.405 0.012 0.409 (0.119) (0.235) (0.231) (0.286) (0.210) (0.251) (0.124) (0.157) Interactiveness × 1960 0.190** 0.485*** 0.331* 0.756*** 0.578*** 0.563*** (0.064) (0.185) (0.180) (0.243) (0.200) (0.215) Interactiveness × 1980 0.274*** 0.560*** 0.449*** 0.777*** 0.658*** 0.651*** 0.295*** 0.634*** (0.052) (0.174) (0.168) (0.231) (0.191) (0.210) (0.090) (0.160) Interactiveness × 2000 0.317*** 0.596*** 0.478*** 0.798*** 0.794*** 0.697*** 0.339*** 0.684*** (0.040) (0.174) (0.169) (0.233) (0.196) (0.227) (0.091) (0.157) Observations 56,760 56,760 50,180 42,460 23,189 31,133 38,647 42,653 Occupation fixed effects Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Married only sample Yes Single only sample Yes Manufacturing only Yes Services only Yes No more skilled MSAs Yes No less skilled MSAs Yes Metro employment share (1) (2) (3) (4) (5) (6) (7) (8) Interactiveness × 1900 − 0.261** 0.104 − 0.001 0.260 0.076 0.126 − 0.032 0.119 (0.117) (0.162) (0.121) (0.224) (0.159) (0.198) (0.092) (0.158) Interactiveness × 1920 − 0.104 0.187 0.019 0.428 0.328 0.214 − 0.012 0.218 (0.118) (0.218) (0.198) (0.273) (0.198) (0.234) (0.102) (0.203) Interactiveness × 1940 − 0.04 0.321 0.177 0.534* 0.424** 0.405 0.012 0.409 (0.119) (0.235) (0.231) (0.286) (0.210) (0.251) (0.124) (0.157) Interactiveness × 1960 0.190** 0.485*** 0.331* 0.756*** 0.578*** 0.563*** (0.064) (0.185) (0.180) (0.243) (0.200) (0.215) Interactiveness × 1980 0.274*** 0.560*** 0.449*** 0.777*** 0.658*** 0.651*** 0.295*** 0.634*** (0.052) (0.174) (0.168) (0.231) (0.191) (0.210) (0.090) (0.160) Interactiveness × 2000 0.317*** 0.596*** 0.478*** 0.798*** 0.794*** 0.697*** 0.339*** 0.684*** (0.040) (0.174) (0.169) (0.233) (0.196) (0.227) (0.091) (0.157) Observations 56,760 56,760 50,180 42,460 23,189 31,133 38,647 42,653 Occupation fixed effects Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Married only sample Yes Single only sample Yes Manufacturing only Yes Services only Yes No more skilled MSAs Yes No less skilled MSAs Yes Notes: Each column of the table corresponds to a separate regression (equation (29) in the paper). Estimation sample is a panel of observations on three-digit occupations, three-digit sectors and years for 20-year time periods from 1880–2000; 1880 is the excluded year from the interaction terms; Interactiveness is the frequency with which verbs from 1991 occupational descriptions appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Married only sample includes married workers only. Single only sample excludes married workers. Manufacturing only sample includes workers in manufacturing only. Services only sample includes workers in services only. More and less-skilled metro areas are defined as in Glaeser and Resseger (2010), based on whether the share of adults with a college degree in a Metropolitan Statistical Area (MSA) is greater than or less than 25.025% in 2006. The year 1960 is omitted in columns (7) and (8) because the IPUMS 1960 data do not contain the identifiers for individual MSAs. Standard errors are heteroskedasticity robust and clustered on occupation. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large In columns (3) and (4), we examine an alternative potential explanation for our findings based on changes in female labor force participation. Over our long historical time period, female labor force participation increased substantially, which implies that more and more couples face a colocation problem where both partners are looking for work in a common location (e.g., Costa and Kahn 2000). Since solving such a colocation problem is likely to be easier in more densely-populated locations, one concern is that the movement of such “power couples” into densely populated locations could be driving the increase in the relative concentration of employment in interactive occupations in metro areas. Although it is not necessarily the case that power couples work in interactive occupations, columns (3) and (4) provide evidence against this concern by estimating the specification in column (2) separately for single and married people. Comparing the two columns, we find a similar pattern of results irrespective of marital status, which suggests that our findings are not being driven by the location decisions of power couples. In columns (5) and (6), we provide further evidence against explanations based on individual sectors. In column (5), we include only workers in the manufacturing sector and demonstrate a similar pattern of results, which corroborates that our findings are not simply being driven by the rise of the services sector in urban areas. In column (6), we include only workers in the services sector, which confirms that our findings are not simply being driven by a decline in manufacturing in urban areas. Therefore, the increase in the interactiveness of urban areas over time does not simply reflect manufacturing jobs moving out of cities or service jobs moving into cities. In columns (7) and (8), we examine the role of differences in human capital across cities. Glaeser and Resseger (2010) find that the positive average relationship between productivity and metro area population is driven by a strong positive relationship for more-skilled metro areas, whereas this relationship is almost nonexistent for less-skilled metro areas. Using Glaeser and Resseger’s (2010) classification of metro areas by skill, columns (7) and (8) re-estimate the specification in Column (2) excluding more and less-skilled metro areas respectively.34 In both samples, we find a positive and statistically significant increase in the relative concentration of employment in interactive occupations in metro areas over time. Therefore, although the size of this increase is larger in the sample excluding less-skilled metro areas, even in the sample excluding more-skilled metro areas we find the same reallocation of employment toward interactive occupations in metro areas. Therefore, our findings of an increased interactiveness of urban areas are not simply explained by differences in levels of human capital across cities. 5.3. Improvements in Transport and Communication Technologies We now provide some suggestive evidence on the role of improvements in transport and communication technologies in influencing the interactiveness of employment. We combine data on employment by occupation, sector and county for 1880 and 1930 with information on the spatial diffusion of the telephone and road network in the opening decades of the 20th century. We focus on this period because both the telephone and paved highways were virtually nonexistent in 1880 and diffused rapidly from 1880 to 1930; we observe the largest increase in the relative interactiveness of metro areas over these decades; and county identifiers are reported in the individual-level population census data for these years. Our baseline specification regresses the change in interactiveness in each county from 1880 to 1930 (▵Interactivec) on log telephones per capita (Phonepcc) and highways per kilometer (Highwaypac) in the 1930s: \begin{equation} \triangle \text{Interactive}_{c} = \alpha _{P} \ln \left( \text{Phonepc}_{c} \right) + \alpha _{H} \text{Highwaypa}_{c} + X_{c} \alpha _{X} + u_{c}, \end{equation} (30) where Phonepcc is residence telephones in 1935 divided by population in 1930; Highwaypac is the length of highways from the Gallup (1931) map in each county divided by county area; Xc are controls for other county characteristics; uc is a stochastic error; since telephones and paved highways were both essentially nonexistent in 1880, the values of these variables in the 1930s capture their growth from 1880 to 1930. Telephones and highways are unlikely to be randomly assigned to counties. Therefore, a concern is that changes in interactiveness and the diffusion of these technologies both could be influenced by omitted third factors that enter the error term uc and hence induce a correlation between the diffusion of these technologies and the error term. In particular, we have already shown that more densely populated locations experienced an increase in their relative interactiveness over time, and telephones and highways may have also diffused more rapidly to more densely-populated locations. For this reason, we include among our controls Xc each county’s initial log population in 1880 and its log area. To further address the concern that telephones and roads are nonrandomly assigned, we develop instruments based on institutional features of the development of the telephone and highway network. We include these instruments alongside our controls in the following first-stage regressions: \begin{equation} \ln \left( \text{Phonepc}_{c} \right) = \beta _{P} Z_{Pc} + \beta _{H} Z_{Hc} + X_{c} \beta _{X} + \varepsilon _{c}, \end{equation} (31) \begin{equation} \text{Highwaypa}_{c} = \gamma _{P} Z_{Pc} + \gamma _{H} Z_{Hc} + X_{c} \gamma _{X} + \omega _{c}, \end{equation} (32) where ZPc is our instrument for telephones (P is mnemonic for phones) and ZHc is our instrument for highways (H is mnemonic for highways); εc and ωc are stochastic errors. To develop an instrument for log telephones per capita, we exploit the network structure of telephone communication. Following Alexander Graham Bell’s successful filing for a patent in 1876, the Bell Telephone Company was incorporated in 1877, and the first telephone exchange was opened under license from Bell Telephone in New Haven, CT in 1878. As local telephone exchanges began to emerge in major U.S. cities, the AT&T was formed in 1885 as a subsidiary of American Bell Telephone to build and operate a long distance telephone network. In these early years, there was considerable debate within American Bell Telephone about the strategic rationale for developing a long distance network and whether such a network would be profitable given that much of the initial demand for telephones appeared to be local (see, e.g., John 2010). By the end of 1885, the first long distance line was completed between New York and Philadelphia with an initial capacity of one telephone call, and it was not until 1892 that a long distance line to Chicago was finished again with an initial capacity of one call. Following Theodore Vail’s accession to the Presidency of AT&T in 1907, the company aggressively pursued the development of its long distance network, with the strategic goals of connecting the nation as a whole (e.g., Osbourne 1930) and pressing for nationwide monopoly powers under Vail’s slogan of “One System, One Policy, Universal Service”. Ultimately this goal was achieved in 1913 with the issuance of the Kingsbury Commitment, which established AT&T as a government-sponsored monopoly, in return for it divesting its interests in the manufacture of telephone and telegraph equipment and allowing independent telephone companies to connect with its long distance network. By 1915, the first transcontinental long distance line to San Francisco was completed. As our instrument for county log telephones per capita, we use county proximity to AT&T’s long distance network (see Map A.1 in the Online Appendix). We measure proximity using the log of the sum of the distances from each county’s centroid to the nearest primary and secondary outlets on this network, which captures the centrality of each county relative to the network. This instrument uses the fact that AT&T’s long distance network was developed with the strategic objective of connecting the nation rather than based on interactiveness in individual counties. Our identifying assumption is that conditional on our controls for initial population and area there is no direct effect of proximity to long distance outlets on county interactiveness other than through log telephones per capita. The locations of these long distance outlets have predictive power for log telephones per capita, because they facilitated the connection of local telephone companies to the long distance network, which increased the value of a telephone connection to local subscribers, and hence increased telephone diffusion. In this way, we exploit the network properties of the telephone, which it shares with for example distribution networks, as in Holmes (2011). Our instrument for highways per kilometer uses the institutional development of the U.S. highways network. In 1880, paved roads were the exception and were concentrated in the immediate vicinity of central business districts.35 Demand for road improvements grew following the production of the first American gasoline-powered automobile in Chicopee, Massachusetts in 1893 and the rapid growth in car registrations, which reached 8,000 in 1900, nearly 33,000 in 1903, and over 10 million by 1921 (U.S. Department of Transport 1976; Lewis 1997; Swift 2011). The federal government’s involvement in the road network dates back to the formation of the Office of Road Inquiry in 1893, which became the Office of Public Roads in 1905 and the Bureau of Public Roads in 1915. Federal government participation was stimulated in part by its responsibility for the postal service, which was a department of the federal government from 1792 to 1971. Thus the Federal Aid Road Act of 1916 provided federal funding for rural post roads on the condition that these roads were open to public at no charge and that states submitted plans, surveys, and estimates for the approval of the Secretary of Agriculture. The scale of federal government participation grew with the Federal Aid Highway Act of 1921, which provided 50–50 matching funds for state highway building. Each state was required to propose a system of roads for federal aid that did not exceed 7% of its highway mileage, and the Department of Agriculture was authorized to publish a map of the network on which federal aid would be spent by November 1923. As part of the planning process for this network, the Bureau of Public Roads commissioned General John J. Pershing to draw up a map of roads of military importance in the event of war. This “Pershing Map” identified 75,000 miles of road as strategically important for reasons of coastal and border defense (see Map A.2 in the Online Appendix).36 More than 10,000 miles of Federal Aid Highways were laid down in 1922 and by 1929 more than 90% of the Federal Aid Highways (around 170,000 miles) had been improved. We instrument the length of highways per kilometer from the Gallup (1931) map using the length of Pershing highways per kilometer within each county from the Pershing Map. Our identifying assumption is that conditional on our controls for initial population and area there is no direct effect of Pershing highways per kilometer on county interactiveness other than through actual highways per kilometer. Pershing highways per kilometer have predictive power for actual highways per kilometer, because these highways of military importance were incorporated into the final network of Federal Aid Highways in the Department of Agriculture’s 1923 map. In column (1) of Table 7, we begin by running an OLS regression of the change in county interactiveness from 1880 to 1930 on county log telephones per capita in 1935, highways per kilometer in 1931 and our controls (equation (30)). We find a positive and statistically significant coefficient for telephones and a positive but statistically insignificant coefficient for highways.37 In column (2), we report our instrumental variables estimates of equations (30)–(32). We find positive and statistically significant coefficients for both telephones and highways. Therefore, both the diffusion of telephones induced by AT&T’s long distance network and the development of highways for military reasons raise county interactiveness. The increase in the estimated coefficients on highways between the OLS and IV specifications is consistent with the view that conditional on our controls for population density highways are disproportionately assigned to locations with lower growth in interactiveness. This finding is in line with Duranton and Turner’s (2012) results for the later interstate highway system, in which conditional on their controls highways also appear to be disproportionately assigned to relatively less-developed locations. Although these specifications control for population density through the inclusion of log population and log area, we also find a very similar pattern of results if we also include a (0,1) dummy for whether a county is located within a metro area. Table 7. Interactiveness and improvements in communication and transport technologies. (1) (2) (3) (4) (5) Change in interactiveness 1880–1930 Change in interactiveness 1880–1930 Log phones per capita 1935 Highways per km 1931 Change in interactiveness 1880–1930 Highways per km 0.007 0.086*** (0.004) (0.028) Log phones per capita 0.022*** 0.083*** (0.002) (0.019) Log area 0.007*** 0.010*** − 0.013** − 0.030*** 0.007*** (0.001) (0.001) (0.005) (0.003) (0.001) Log population 1880 0.004*** 0.002* 0.006* 0.016*** 0.004*** (0.001) (0.001) (0.003) (0.002) (0.007) Pershing highways per km − 0.113** 0.274*** 0.015** (0.055) (0.032) (0.005) Log remoteness from long distance outlet − 0.063*** 0.008** − 0.005*** (0.009) (0.004) (0.001) Observations 2467 2467 2467 2509 2509 R-squared 0.12 0.19 0.02 0.19 0.09 Estimation OLS 2SLS OLS OLS OLS Specification Second-stage Second-stage First-stage First-stage Reduced-form F-statistic instruments 26.35 38.4 14.05 Underidentification test (Kleibergen–Paap LM statistic) 35.63 Weak identification test (Kleibergen–Paap F-statistic) 18.61 (1) (2) (3) (4) (5) Change in interactiveness 1880–1930 Change in interactiveness 1880–1930 Log phones per capita 1935 Highways per km 1931 Change in interactiveness 1880–1930 Highways per km 0.007 0.086*** (0.004) (0.028) Log phones per capita 0.022*** 0.083*** (0.002) (0.019) Log area 0.007*** 0.010*** − 0.013** − 0.030*** 0.007*** (0.001) (0.001) (0.005) (0.003) (0.001) Log population 1880 0.004*** 0.002* 0.006* 0.016*** 0.004*** (0.001) (0.001) (0.003) (0.002) (0.007) Pershing highways per km − 0.113** 0.274*** 0.015** (0.055) (0.032) (0.005) Log remoteness from long distance outlet − 0.063*** 0.008** − 0.005*** (0.009) (0.004) (0.001) Observations 2467 2467 2467 2509 2509 R-squared 0.12 0.19 0.02 0.19 0.09 Estimation OLS 2SLS OLS OLS OLS Specification Second-stage Second-stage First-stage First-stage Reduced-form F-statistic instruments 26.35 38.4 14.05 Underidentification test (Kleibergen–Paap LM statistic) 35.63 Weak identification test (Kleibergen–Paap F-statistic) 18.61 Notes: Each column of the table corresponds to a separate regression; observations are a cross-section of counties; Interactiveness is the frequency with which verbs from 1991 occupational descriptions appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus; Highways per km is length of highways within a county in the Gallup 1931 map divided by county area; Log phones per capita is log number of residence telephones in a county in 1935 divided by county population in 1930; Log area is log county area; Log Population 1880 is log county population in 1880; Pershing highways per km is the length of highways proposed for military reasons within a county in the Pershing 1922 map divided by county area; Log remoteness from long distance outlet is the log of the sum of the distances to primary and secondary outlets on the AT&T long distance telephone network. Heteroskedasticity robust standard errors in parentheses. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large Table 7. Interactiveness and improvements in communication and transport technologies. (1) (2) (3) (4) (5) Change in interactiveness 1880–1930 Change in interactiveness 1880–1930 Log phones per capita 1935 Highways per km 1931 Change in interactiveness 1880–1930 Highways per km 0.007 0.086*** (0.004) (0.028) Log phones per capita 0.022*** 0.083*** (0.002) (0.019) Log area 0.007*** 0.010*** − 0.013** − 0.030*** 0.007*** (0.001) (0.001) (0.005) (0.003) (0.001) Log population 1880 0.004*** 0.002* 0.006* 0.016*** 0.004*** (0.001) (0.001) (0.003) (0.002) (0.007) Pershing highways per km − 0.113** 0.274*** 0.015** (0.055) (0.032) (0.005) Log remoteness from long distance outlet − 0.063*** 0.008** − 0.005*** (0.009) (0.004) (0.001) Observations 2467 2467 2467 2509 2509 R-squared 0.12 0.19 0.02 0.19 0.09 Estimation OLS 2SLS OLS OLS OLS Specification Second-stage Second-stage First-stage First-stage Reduced-form F-statistic instruments 26.35 38.4 14.05 Underidentification test (Kleibergen–Paap LM statistic) 35.63 Weak identification test (Kleibergen–Paap F-statistic) 18.61 (1) (2) (3) (4) (5) Change in interactiveness 1880–1930 Change in interactiveness 1880–1930 Log phones per capita 1935 Highways per km 1931 Change in interactiveness 1880–1930 Highways per km 0.007 0.086*** (0.004) (0.028) Log phones per capita 0.022*** 0.083*** (0.002) (0.019) Log area 0.007*** 0.010*** − 0.013** − 0.030*** 0.007*** (0.001) (0.001) (0.005) (0.003) (0.001) Log population 1880 0.004*** 0.002* 0.006* 0.016*** 0.004*** (0.001) (0.001) (0.003) (0.002) (0.007) Pershing highways per km − 0.113** 0.274*** 0.015** (0.055) (0.032) (0.005) Log remoteness from long distance outlet − 0.063*** 0.008** − 0.005*** (0.009) (0.004) (0.001) Observations 2467 2467 2467 2509 2509 R-squared 0.12 0.19 0.02 0.19 0.09 Estimation OLS 2SLS OLS OLS OLS Specification Second-stage Second-stage First-stage First-stage Reduced-form F-statistic instruments 26.35 38.4 14.05 Underidentification test (Kleibergen–Paap LM statistic) 35.63 Weak identification test (Kleibergen–Paap F-statistic) 18.61 Notes: Each column of the table corresponds to a separate regression; observations are a cross-section of counties; Interactiveness is the frequency with which verbs from 1991 occupational descriptions appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas) and Class V, Division 2 (Intersocial Volition) of the thesaurus; Highways per km is length of highways within a county in the Gallup 1931 map divided by county area; Log phones per capita is log number of residence telephones in a county in 1935 divided by county population in 1930; Log area is log county area; Log Population 1880 is log county population in 1880; Pershing highways per km is the length of highways proposed for military reasons within a county in the Pershing 1922 map divided by county area; Log remoteness from long distance outlet is the log of the sum of the distances to primary and secondary outlets on the AT&T long distance telephone network. Heteroskedasticity robust standard errors in parentheses. *Significant at 10%; **significant at 5%; ***significant at 1%. View Large In columns (3) and (4) of Table 7, we report the first-stage regressions for phones and highways, respectively, whereas column (5) reports the reduced-form regression. We find that proximity to the AT&T long distance network and Pershing highways have predictive power for the endogenous variables, with first-stage F-statistics on the excluded exogenous variables of 38.40 and 26.35 in columns (3) and (4), respectively. Consistent with this, we reject the null hypotheses of underidentification and weak identification in the Kleibergen–Paap test statistics reported in column (2). We acknowledge several caveats to these results. We cannot definitively rule out that areas on the Pershing map and further from ATT trunk lines could otherwise have changed differently even in the absence of improvements in transportation and communication technologies. Our sample also includes counties in cities, where the network connections may have been designed to connect these urban areas, and hence our empirical approach differs from studies that focus solely on rural areas on the route between cities. Nonetheless, these results provide some suggestive evidence that the change in the interactiveness of employment is related to reductions in transportation and communication costs. 6. Conclusions Although there is a large literature on agglomeration, there is relatively little evidence on the tasks undertaken within cities and how these have changed over time. We develop a new methodology for quantifying the tasks undertaken in urban and rural areas that uses over 3,000 verbs from more than 12,000 occupational descriptions over a period of more than a century. We first characterize changes in the task content of employment using verbs and their meanings in a way that imposes little prior structure on the data. Guided by these findings, we construct a quantitative measure of the interactiveness of occupations based on the frequency with which verbs from the occupational descriptions appear in thesaurus categories involving thought, communication and intersocial activity. Using this measure, we find an increase in the employment share of interactive occupations within sectors over time that is larger in metro areas than nonmetro areas. These findings highlight a change in the nature of agglomeration over time toward an increased emphasis on human interaction. We interpret these findings using a model of trade in the tasks produced by each occupation and in the final goods produced by each sector. When the costs of trading tasks and final goods are large, all locations have similar employment structures across sectors, and all tasks within each sector are undertaken in the same location where the final good is produced. As the costs of trading final goods and tasks fall, locations specialize across sectors and across occupations within sectors according to their comparative advantage as determined by productivity differences. If densely-populated urban locations have a comparative advantage in interactive tasks relative to sparsely populated rural locations, a fall in trade costs leads to an increase in the interactiveness of employment within sectors in urban relative to rural areas. Guided by these predictions, we examine alternative potential explanations for the increase in the interactiveness of employment in urban areas relative to rural areas over time. Our results are not simply explained by reallocation across sectors, because our regressions control for sector-year fixed effects. Furthermore, we find a similar pattern of results when we restrict attention to workers within manufacturing or workers within services. Although we find that the movement of manufacturing jobs out of cities contributes toward the observed changes in interactiveness, it is only part of the story, because we also find important contributions from Business Services, Personal Services and Professional Services. We establish similar results using both employment and wagebill shares, consistent with an explanation for changes in interactiveness in terms of movements in relative demand. Finally, we provide suggestive evidence that the increase in interactiveness is related to improvements in communication and transport technologies, using the diffusion of telephones and highways in the late-19th and early-20th centuries. Notes The editor in charge of this paper was Paola Giuliano. Acknowledgments We are grateful to the Centre for Economic Performance and Princeton University for research support. We are also grateful to Gilles Duranton and Matt Turner for sharing data. We thank William D. Caughlin and George Kupczak for their help with the telephone data. We also thank the editor, four referees, Don Davis, Gilles Duranton, Gene Grossman, Gordon Hanson, Marco Manacorda, Alan Manning, Nathan Nunn, Vernon Henderson, Rick Hornbeck, Gianmarco Ottaviano, Steve Pischke, Esteban Rossi-Hansberg, Yona Rubinstein, Will Strange, Nancy Qian, Tony Venables and seminar and conference participants at Copenhagen, the London School of Economics, MIT, NARSC, Oxford, Sorbonne and Sussex for helpful comments and suggestions. The usual disclaimer applies. Michaels is a Research Associate at CEP and a Research Fellow at CEPR; Rauch is a Visiting Research Affiliate at CEP; and Redding is an International Research Associate at CEP and a Research Fellow at CEPR. Footnotes 1 See, for example, John (2010). The electric telegraph was patented much earlier in 1837 by Samuel Morse and the U.S. telegraph network was largely complete by 1880 (see Standage 2007). 2 See, for example, Swift (2011). 3 Following the Federal Aid Highway Act of 1921, the Bureau of Public Roads commissioned General John J. Pershing to draw up a map of roads of military importance in the event of war, as discussed further in what follows. 4 See Henderson (1974) for the classic analysis of industry specialization and the size distribution of cities. 5 Other dimensions of specialization in urban areas examined in existing research include localization versus urbanization externalities (e.g., Jacobs 1969; Henderson 2003); the division of labor (e.g., Duranton and Jayet 2011); diversified versus specialized cities (e.g., Duranton and Puga 2001); employment in newly-created occupations (e.g., Lin 2011); and agglomeration versus firm selection (in particular Combes et al. 2012). 6 In their classic study of the race between technology and skills, Goldin and Katz (2008) present evidence that manufacturing technologies were skill complementary in the early-20th century, but may have been skill substituting prior to that time. In subsequent work, Katz and Margo (2014) report some evidence of deskilling in manufacturing during the 19th century, but find a reallocation of employment toward high-skill jobs for the aggregate economy as a whole. 7 From 1950 onward, the definitions of metropolitan areas are those constructed by the Census Bureau. Before 1950, IPUMS applies to the historical population data the same criteria as used by the Census Bureau from 1950 onward. 8 See IPUMS for the full concordance between two-digit and three-digit occupations and sectors. Although both occupation and sector classifications are standardized by IPUMS, there are a small number of occupations and sectors that enter and exit the sample over time. All our results are robust to restricting attention to occupations and sectors that are present in all years. 9 Our key findings, however, are robust to the inclusion of these agricultural workers. For further analysis of the relationship between urbanization and structural transformation away from agriculture, see Michaels, Rauch, and Redding (2012). 10 See http://www.writingenglish.com/englishverbs.htm. 11 We use a computer-searchable edition of Roget (1911) from the University of Chicago. 12 Recent economics research on the U.S. road network has largely concentrated on the later development of the interstate highway system, as in Baum-Snow (2007), Michaels (2008), and Duranton and Turner (2012). 13 The average three-digit sector employs workers from 111 three-digit occupations, whereas the average three-digit occupation contains workers employed in 81 sectors. 14 An emerging literature in economics and the social sciences uses textual search as the basis for quantitative analysis: see, for example, Gentzkow, Shapiro, and Sinkinson (2014) on political influence and Michel et al. (2011) on culture. 15 As an indication of the wide coverage of our list of over 3,000 verbs, only 1,830 appear in the 1991 DOTs occupational descriptions. 16 A data set with the verb frequencies for each verb and occupation is available as part of the replication data for this paper. 17 We find a similar pattern of results just using the estimated coefficients instead of the estimated coefficients times the standard deviation of VerbFreqvo. 18 For further discussion of the genesis of Roget’s Thesaurus, see, for example, Hüllen (2003). 19 Division is an intermediate level in between class and section, which is only used in the thesaurus to disaggregate Classes IV and V. In particular, “Class IV Words Relating to the Intellectual Faculties” is split into “Division I Formation of Ideas” and “Division II Communication of Ideas”. Additionally, “Class V Words Relating to the Voluntary Powers” is split into “Division I Individual Volition” and “Division II Intersocial Volition”. Classes I–III and VI are not disaggregated into divisions. 20 For example, the verb “Consult” appears in six thesaurus Categories. The entry followed by a comma is 695 Advice, which captures the word’s meaning. Entries not followed by a comma correspond to idiomatic uses not closely related to the word’s meaning: 133 Lateness (“consult one’s pillow”); 463 Experiment (“consult the barometer”); 707 Aid (“consult the wishes of”); 943 Selfishness (“consult one’s own pleasure”); 968 Lawyer (“juris consult [Latin]”). We do not require that a verb is preceded by a comma or semi-colon, because this would exclude the first verb listed under each thesaurus category. 21 A data set containing the task content measures for each thesaurus category and occupation is also available as part of the replication data for this paper. 22 Again we find a similar pattern of results using just the estimated coefficient instead of the estimated coefficient times the standard deviation of TaskContentko. 23 A more detailed exposition of the model including the technical derivations of relationships is contained in the Online Appendix. 24 For empirical evidence using U.S. data in support of the constant expenditure share implied by the Cobb–Douglas functional form, see Davis and Ortalo-Magne (2011). 25 To simplify the exposition, we use i to denote locations of production and n to denote locations of consumption, except where otherwise indicated. 26 To reduce the notational burden, we assume the same [0, 1] interval of tasks for all occupations, but it is straightforward to allow this interval to vary across occupations. 27 Although we interpret production as being undertaken by workers in occupations that perform many tasks, an equivalent interpretation is that each occupation corresponds to a stage of production and each task corresponds to an intermediate input within that stage of production. 28 Since the Fréchet distribution is unbounded from above, each location draws an arbitrarily high input productivity for a positive measure of tasks. To allow for the possibility that a location may not have positive employment in an occupation o and sector s, we take lim Uiso → 0, in which case the location’s employment in that occupation and sector converges to zero. Similarly, to allow for the possibility that an occupation o may not be traded, we take lim dnis → ∞, in which case trade in that occupation converges to zero. 29 Since the Fréchet distribution is unbounded from above, each location draws an arbitrarily high final goods productivity for a positive measure of final goods. To allow for the possibility that a location may not have positive employment in a sector s, we take lim Tis → 0, in which case the location’s employment in that sector converges to zero. Similarly, to allow for the possibility that a sector s may not be traded, we take lim dnis → ∞, in which case trade in that sector converges to zero. 30 For further discussion of trade in tasks and the unbundling of production, see, for example, Baldwin (2014, Chap. 5). 31 Using data from the early 20th century, Gray (2013) finds that electrification led to a polarization of the employment distribution, increasing the demand for nonroutine and routine cognitive tasks, whereas simultaneously reducing relative demand for the nonroutine manual jobs that comprised the middle of the skill distribution. 32 In a study of the merchant shipping industry in the late-19th and early-20th centuries, Chin, Juhn, and Thompson (2006) find that the adoption of the steam engine raised skill premia. Using data on manufacturing plants in the late-19th century, Atack, Bateman, and Margo (2004) find that plant wages are decreasing in size, but are increasing in the use of steam power, which is consistent with technology-skill complementarity. 33 Since the change in overall interactiveness is the sum across all elements in the matrix, adding the sums for occupations and the sums for sectors would result in double-counting (since each element would be counted twice). 34 In Glaeser and Resseger’s (2010) classification, more-skilled Metropolitan Statistical Areas (MSAs) have a share of adults with college degrees of greater than 25.025% in 2006. The year 1960 is omitted in columns (7) and (8) because the IPUMS 1960 data do not contain the identifiers for individual MSAs. 35 At the end of 1909, concrete accounted for only nine miles of state and county roads (Macdonald 1928). 36 Consistent with these objectives, the Pershing Map excluded parts of the Deep South and Florida that were considered to be sufficiently swampy as to render foreign invasion impractical. 37 As discussed in Section 2, our telephones data for 1935 are for residence telephones. Separate data for business and residence telephones are available for 1945 and we find a strong correlation between them. Regressing log residence telephones on log business telephones across counties in 1945, we find an estimated coefficient (standard error) of 0.8950 (0.0090) and a regression R-squared of 0.87. References Acemoglu Daron ( 1998 ). “Why do New Technologies Complement Skills? Directed Technical Change and Wage Inequality.” Quarterly Journal of Economics , 113 , 1055 – 1090 . Google Scholar CrossRef Search ADS Acemoglu Daron ( 2002 ). “Directed Technical Change.” Review of Economic Studies , 69 , 781 – 810 . Google Scholar CrossRef Search ADS Acemoglu Daron Autor David ( 2011 ). “Skills, Tasks and Technologies: Implications for Employment and Earnings.” In Handbook of Labor Economics , Vol. 4B , edited by Ashenfelter Orley Card David . Elsevier , Amsterdam , pp. 1043 – 1171 . American Telephone and Telegraph Company ( 1935 ). Residence Telephones by Counties . American Telephone and Telegraph Company , New York . Atack Jeremy Bateman Fred Margo Robert ( 2004 ). “Skill Intensity and Rising Wage Dispersion in Nineteenth-Century American Manufacturing.” Journal of Economic History , 64 , 172 – 192 . Google Scholar CrossRef Search ADS Autor David ( 2013 ). “The ‘Task Approach’ to Labor Markets: An Overview.” Journal for Labour Market Research , 46 , 185 – 199 . Google Scholar CrossRef Search ADS Autor David Handel Michael ( 2009 ). “Putting Tasks to the Test: Human Capital, Job Tasks and Wages.” Journal of Labor Economics , 31 , S59 – S96 . Google Scholar CrossRef Search ADS Autor David Katz Lawrence Krueger Alan ( 1998 ). “Computing Inequality: Have Computers Changed the Labor Market?” Quarterly Journal of Economics , 113 , 1169 – 1214 . Google Scholar CrossRef Search ADS Autor David Levy Frank Murnane Richard J. ( 2003 ). “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics , 4 , 1279 – 1333 . Google Scholar CrossRef Search ADS Bacolod Marigee Blum Bernardo Strange William C. ( 2009a ). “Skills in the City.” Journal of Urban Economics , 65 , 136 – 153 . Google Scholar CrossRef Search ADS Bacolod Marigee Blum Bernardo Strange William C. ( 2009b ). “Urban Interactions: Soft Skills Versus Specialization.” Journal of Economic Geography , 9 , 227 – 262 . Google Scholar CrossRef Search ADS Baldwin Richard ( 2014 ). “Trade and Industrialisation after Globalisation’s 2nd Unbundling: How Building and Joining a Supply Chain are Different and Why it Matters.” In Globalization in an Age of Crisis: Multilateral Economic Cooperation in the Twenty-First Century , edited by Feenstra Robert Taylor Alan . University of Chicago Press , Chicago , pp. 165 – 212 . Baum-Snow Nathaniel ( 2007 ). “Did Highways Cause Suburbanization?” Quarterly Journal of Economics , 122 , 775 – 805 . Google Scholar CrossRef Search ADS Chin Aimee Juhn Chinhui Thompson Peter ( 2006 ). “Technical Change and the Demand for Skills during the Second Industrial Revolution: Evidence from the Merchant Marine, 1891–1912.” Review of Economics and Statistics , 88 , 572 – 578 . Google Scholar CrossRef Search ADS Combes Pierre-Philippe Duranton Gilles Gobillon Laurent Puga Diego Roux Sebastien ( 2012 ). “The Productivity Advantages of Large Cities: Distinguishing Agglomeration from Firm Selection.” Econometrica , 80 , 2543 – 2594 . Google Scholar CrossRef Search ADS Costa Dora Kahn Matthew ( 2000 ). “Power Couples: Changes in the Locational Choice of the College Educated, 1940–1990.” Quarterly Journal of Economics , 115 , 1287 – 1315 . Google Scholar CrossRef Search ADS Costinot Arnaud Donaldson Dave Komunjer Ivana ( 2012 ). “What Goods Do Countries Trade? A Quantitative Exploration of Ricardo’s Ideas.” Review of Economic Studies , 79 , 581 – 608 . Google Scholar CrossRef Search ADS Davis Donald Dingel Jonathan ( 2012 ). “A Spatial Knowledge Economy.” NBER Working Paper No. 18188, National Bureau of Economic Research, Cambridge, MA . Davis Morris Ortalo-Magne François ( 2011 ). “Household Expenditures, Wages, Rents.” Review of Economic Dynamics , 14 , 248 – 261 . Google Scholar CrossRef Search ADS Deming David ( 2017 ). “The Growing Importance of Social Skills in the Labor Market.” Quarterly Journal of Economics , 4 , 1593 – 1640 . Google Scholar CrossRef Search ADS Duranton Gilles Puga Diego ( 2001 ). “Nursery Cities: Urban Diversity, Process Innovation, and the Life Cycle of Products.” American Economic Review , 91 (5), 1454 – 1477 . Google Scholar CrossRef Search ADS Duranton Gilles Puga Diego ( 2004 ). “Micro-foundations of Urban Agglomeration Economies.” In Handbook of Regional and Urban Economics , Vol. 4 , edited by Henderson J. Vernon Thisse Jacques-François . Elsevier , Amsterdam , pp. 2063 – 2117 . Duranton Gilles Puga Diego ( 2005 ). “From Sectoral to Functional Urban Specialization.” Journal of Urban Economics , 57, 343 – 370 . Duranton Gilles Jayet Hubert ( 2011 ). “Is the Division of Labour Limited by the Extent of the Market? Evidence from French Cities.” Journal of Urban Economics , 69 , 56 – 71 . Google Scholar CrossRef Search ADS Duranton Gilles Turner Matthew ( 2012 ). “Urban Growth and Transportation.” Review of Economic Studies , 79 , 1407 – 1440 . Google Scholar CrossRef Search ADS Eaton Jonathan Kortum Samuel ( 2002 ). “Technology, Geography, and Trade.” Econometrica , 70 , 1741 – 1779 . Google Scholar CrossRef Search ADS Ethier Wilfred ( 1982 ). “Decreasing Costs in International Trade and Frank Graham’s Argument for Protection.” Econometrica , 50 , 1243 – 1268 . Google Scholar CrossRef Search ADS Fischer Claude ( 1992 ). America Calling: A Social History of the Telephone to 1940 . University of California Press , Los Angeles . Fujita Masahisa Tabuchi Takatoshi ( 1997 ). “Regional Growth in Postwar Japan.” Regional Science and Urban Economics , 27 , 643 – 670 . Google Scholar CrossRef Search ADS Gallup ( 1931 ). Motor Trails Map of the United States . Gallup Map Company , Kansas City . Gaspar Jess Glaeser Edward ( 1998 ). “Information Technology and the Future of Cities.” Journal of Urban Economics , 43 , 136 – 156 . Google Scholar CrossRef Search ADS Gentzkow Matthew Shapiro Jesse Sinkinson Michael ( 2014 ). “Competition and Ideological Diversity: Historical Evidence from U.S. Newspapers.” American Economic Review , 104 (10), 3073 – 3114 . Google Scholar CrossRef Search ADS Glaeser Edward Saiz Albert ( 2004 ). “The Rise of the Skilled City.” Brookings-Wharton Papers on Urban Affairs , 5 , 47 – 94 . Google Scholar CrossRef Search ADS Glaeser Edward Ponzetto Giacomo Tobio Kristina ( 2014 ). “Cities, Skills and Regional Change.” Regional Studies , 48 , 7 – 43 . Google Scholar CrossRef Search ADS Glaeser Edward Resseger Matthew ( 2010 ). “The Complementarity Between Cities and Skills.” Journal of Regional Science , 50 , 221 – 244 . Google Scholar CrossRef Search ADS Goldin Claudia Katz Lawrence ( 2008 ). The Race between Education and Technology . Harvard University Press , Cambridge . Gray Rowena ( 2013 ). “Taking Technology to Task: The Skill Content of Technological Change in Early Twentieth Century U.S.” Explorations in Economic History , 50 , 351 – 367 . Google Scholar CrossRef Search ADS Grossman Gene Rossi-Hansberg Esteban ( 2008 ). “Trading Tasks: A Simple Theory of Offshoring.” American Economic Review , 98 (5), 1978 – 1997 . Google Scholar CrossRef Search ADS Grossman Gene Rossi-Hansberg Esteban ( 2012 ). “Task Trade between Similar Countries.” Econometrica , 80 , 593 – 629 . Google Scholar CrossRef Search ADS Helpman Elhanan ( 1998 ). “The Size of Regions.” In Topics in Public Economics: Theoretical and Applied Analysis , edited by Pines David Sadka Efraim Zilcha Itzhak . Cambridge University Press , Cambridge. Helsley Robert Strange William C. ( 2007 ). “Agglomeration, Opportunism and the Organization of Production.” Journal of Urban Economics , 62 , 55 – 75 . Google Scholar CrossRef Search ADS Henderson J. Vernon ( 1974 ). “The Sizes and Types of Cities.” American Economic Review , 64 (4), 640 – 656 . Henderson J. Vernon ( 2003 ). “Marshall’s Scale Economies.” Journal of Urban Economics , 53 , 1 – 28 . Google Scholar CrossRef Search ADS Holmes Thomas ( 2011 ). “The Diffusion of Wal-Mart and Economies of Density.” Econometrica , 79 , 253 – 302 . Google Scholar CrossRef Search ADS Hounshell David ( 1985 ). From the American System to Mass Production, 1800–1932: The Development of Manufacturing Technology in the United States . Johns Hopkins University Press , Baltimore . Hüllen Werner ( 2003 ). A History of Roget’s Thesaurus: Origins, Development, and Design . Oxford University Press , Oxford . Google Scholar CrossRef Search ADS Jacobs Jane ( 1969 ). The Economy of Cities . Random House , London . James John Skinner Jonathan ( 1985 ). “The Resolution of the Labor-scarcity Paradox.” Journal of Economic History , 45 , 513 – 540 . Google Scholar CrossRef Search ADS John Richard ( 2010 ). Network Nation: Inventing American Telecommunications . Bellknap Press , Cambridge . Google Scholar CrossRef Search ADS Katz Lawrence Margo Robert ( 2014 ). “Technical Change and the Relative Demand for Skilled Labor: The United States in Historical Perspective.” In Human Capital in History: The American Record , edited by Platt-Boustan Leah Frydman Carola Margo Robert . University of Chicago Press , Chicago , pp. 15 – 58 . Leamer Edward Storper Michael ( 2001 ). “The Economic Geography of the Internet Age.” Journal of International Business Studies , 32 , 641 – 665 . Google Scholar CrossRef Search ADS Lewis Tom ( 1997 ). Divided Highways: Building the Interstate Highways, Transforming American Life . Viking , New York . Lin Jeffrey ( 2011 ). “Technological Adaptation, Cities and New Work.” Review of Economics and Statistics , 93 , 554 – 574 . Google Scholar CrossRef Search ADS Macdonald Thomas ( 1928 ). “The History and Development of Road Building in the United States.” Transactions of the American Society of Civil Engineers , 92 , 1181 – 1206 . Michaels Guy ( 2008 ). “The Effect of Trade on the Demand for Skill—Evidence from the Interstate Highway System.” Review of Economics and Statistics , 90 , 683 – 701 . Google Scholar CrossRef Search ADS Michaels Guy Rauch Ferdinand Redding Stephen ( 2012 ). “Urbanization and Structural Transformation.” Quarterly Journal of Economics , 127 , 535 – 586 . Google Scholar CrossRef Search ADS Michel Jean-Baptiste Shen Yuan Aiden Aviva Veres Adrian Gray Matthew , The Google Books Team , Pickett Joseph Hoiberg Dale Clancy Dan Norvig Peter Orwant Jon Pinker Steven Nowak Martin Aiden Erez ( 2011 ). “Quantitative Analysis of Culture Using Millions of Digitized Books.” Science , 331 , 176 – 182 . Google Scholar CrossRef Search ADS PubMed Mokyr Joel ( 1992 ). The Lever of Riches: Technological Creativity and Economic Progress . Oxford University Press , Oxford . Google Scholar CrossRef Search ADS Moretti Enrico ( 2004 ). “Human Capital Externalities in Cities.” In Handbook of Urban and Regional Economics , Vol. 4 , edited by Henderson J. Vernon Thisse Jacques-François . Elsevier , Amsterdam , pp. 2243 – 2292 . Ngai Rachel Pissarides Chris ( 2007 ). “Structural Change in a Multi-sector Model of Growth.” American Economic Review , 97 (1), 429 – 443 . Google Scholar CrossRef Search ADS Osbourne H. ( 1930 ). “A General Switching Plan for Telephone Toll Service.” Bell System Technical Journal , 9 , 429 – 447 . Google Scholar CrossRef Search ADS Ota Mitsuru Fujita Masahisa ( 1993 ). “Communication Technologies and Spatial Organization of Multi-unit Firms in Metropolitan Areas.” Regional Science and Urban Economics , 23 , 695 – 729 . Google Scholar CrossRef Search ADS Pool Ithiel ( 1977 ). The Social Impact of the Telephone . MIT Press , Cambridge . Roget Peter ( 1911 ). Thesaurus of English Words and Phrases . Computer Readable Version of 1911 Edition , Evergreen Review Incorporated , Thomas Y. Crowell Company , New York . Rosenthal Stuart Strange William ( 2004 ). “Evidence on the Nature and Sources of Agglomeration Economics.” In Handbook of Regional and Urban Economics , Vol. 4 , edited by Henderson J. Vernon Thisse Jacques-François . Elsevier , Amsterdam , pp. 2119 – 2171 . Rossi-Hansberg Esteban Sarte Pierre-Daniel Owens III Raymond ( 2009 ). “Firm Fragmentation and Urban Patterns.” International Economic Review , 50 , 143 – 186 . Google Scholar CrossRef Search ADS Ruggles Steven Alexander J. Trent Genadek Katie Goeken Ronald Schroeder Matthew Sobek Matthew ( 2010 ). Integrated Public Use Microdata Series, Version 5.0 [Machine-readable database] . University of Minnesota , Minneapolis . Standage Tom ( 2007 ). The Victorian Internet: The Remarkable Story of the Telegraph and the Nineteenth Century’s On-line Pioneers . Walker & Company , New York . Swift Earl ( 2011 ). The Big Roads: The Untold Story of the Engineers, Visionaries, and Trailblazers Who Created the American Superhighways . Mariner Books , Houghton Mifflin Harcourt , Boston . U.S. Department of Labor ( 1939 ). Dictionary of Occupational Titles . U.S. Department of Labor , Washington, DC . U.S. Department of Labor ( 1991 ). Dictionary of Occupational Titles . U.S. Department of Labor , Washington, DC . U.S. Department of Transportation ( 1976 ). America’s Highways 1776–1976: A History of the Federal Aid Program . U.S. Government Printing Office, Washington, DC. Yi Kei-Mu , Zhang Jing ( 2013 ). “Structural Change in an Open Economy.” Journal of Monetary Economics , 60 , 667 – 682 . Google Scholar CrossRef Search ADS Supplementary Data Supplementary data are available at JEEA online. © The Author(s) 2018. Published by Oxford University Press on behalf of European Economic Association. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of the European Economic Association Oxford University Press

Task Specialization in U.S. Cities from 1880 to 2000

Loading next page...
 
/lp/ou_press/task-specialization-in-u-s-cities-from-1880-to-2000-38LplPWLQm
Publisher
European Economic Association
Copyright
© The Author(s) 2018. Published by Oxford University Press on behalf of European Economic Association. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
ISSN
1542-4766
eISSN
1542-4774
D.O.I.
10.1093/jeea/jvy007
Publisher site
See Article on Publisher Site

Abstract

Abstract We develop a new methodology for quantifying the tasks undertaken within occupations using over 3,000 verbs from more than 12,000 occupational descriptions in the Dictionary of Occupational Titles (DOTs). Using micro data from the United States from 1880 to 2000, we find an increase in the employment share of interactive occupations within sectors over time that is larger in metro areas than nonmetro areas. We interpret these findings using a model in which reductions in transport and communication costs induce urban areas to specialize according to their comparative advantage in interactive tasks. We present suggestive evidence relating increases in employment in interactive occupations to improvements in transport and communication technologies. Our findings highlight a change in the nature of agglomeration over time toward an increased emphasis on human interaction. 1. Introduction Agglomeration forces are widely understood to play a central role in sustaining the dense concentrations of population observed in urban areas. Much less is known about the detailed tasks undertaken in urban areas and how these have changed over time. Yet understanding the task content of employment in urban and rural areas is central to evaluating alternative theories of agglomeration and assessing the likely impact of improvements in transport and communication technologies on spatial concentrations of economic activity. In this paper, we provide new evidence on the detailed tasks undertaken by workers in urban and rural areas over a long historical time period in the United States. We develop a new methodology for measuring the individual tasks undertaken within occupations using the verbs from occupational descriptions in the Dictionary of Occupational Titles (DOTs). We implement this methodology using micro data on employment in disaggregated occupations and sectors in metro and nonmetro areas from 1880 to 2000. To measure the individual tasks undertaken within each occupation, we use over 3,000 verbs from more than 12,000 occupational descriptions in both historical and contemporary editions of the DOTs. Using these verbs, we find a systematic change in the composition of employment across tasks in urban versus rural areas over time. We quantify this change in task content using the meaning of verbs from Roget’s Thesaurus, as the standard reference for English usage. In both metro and nonmetro areas, we find a systematic reallocation of employment over time toward interactive occupations, which involve tasks described by verbs that appear in thesaurus categories concerned with thought, communication and intersocial activity. At the beginning of our sample period, metro areas actually have lower shares of employment in interactive occupations than nonmetro areas. Over time, employment growth in interactive occupations is much higher in metro areas, so that by the end of our sample period the initial pattern of specialization is reversed, and metro areas are more interactive than nonmetro areas. This increasing interactiveness of employment at higher population densities is observed not only between metro and nonmetro areas but also across metro areas of different population densities. Although in 1880 there is little relationship between specialization in interactive occupations and population density, by 2000 this relationship is positive, strong, and statistically significant. Taken together, these results suggest that human interaction has become increasingly important in agglomerations of economic activity over time. To interpret these empirical results, we develop a model of the spatial distribution of employment across occupations, sectors, and locations. The model features trade in the final good produced by each sector and in the tasks undertaken by the occupations within each sector. When the costs of trading tasks between locations are prohibitively high, all tasks are performed where the final good is produced. As the costs of trading tasks fall, it becomes feasible to unbundle production across locations and trade tasks between them. If agglomeration forces are stronger for interactive tasks, densely populated urban locations have a comparative advantage in interactive tasks, which implies that reductions in task trade costs induce them to specialize in more-interactive occupations, whereas more sparsely populated rural locations specialize in less-interactive occupations. We provide empirical evidence on the relationship between changes in the interactiveness of employment and the major innovations in transport and communication technology that occurred during our sample period. Following the award of Alexander Graham Bell’s patent for the telephone in 1886, the U.S. telephone network grew rapidly in the opening decades of the 20th century.1 After the award of Karl Benz’s patent for the internal combustion engine in 1879 and after the passage of the Federal Aid Road Act of 1916 and the Federal Highway Act of 1921, the U.S. road network and automobile use expanded rapidly over the same period.2 We examine the implications of these new transport and communication technologies by combining county data on employment by occupation and sector for 1880 and 1930 with newly collected county data on telephone use and the road network in the 1930s. We develop instruments for the geographical dissemination of both technologies. For the telephone, we use its network properties to construct an instrument based on proximity to nodes on the American Telegraph and Telephone’s (AT&T) company’s long distance trunk network, whose construction was influenced by the strategic objectives of connecting the nation as a whole. For roads, we use the 1922 “Pershing Map” of highway routes of military importance for coastal and border defense.3 We provide suggestive evidence connecting increases in the interactiveness of employment to the diffusion of these new technologies predicted by our instruments. Our paper is related to a number of literatures. We build on the wider literature on agglomeration economies, as surveyed by Duranton and Puga (2004, Chap. 48) and Rosenthal and Strange (2004, Chap. 49). One strand of this literature emphasizes differences in the composition of economic activity between urban and rural areas. Studies emphasizing the role of human capital and skills in promoting agglomeration include Glaeser and Saiz (2004), Glaeser and Resseger (2010), Bacolod, Blum, and Strange (2009a), Glaeser, Ponzetto, and Toblo (2014), and Moretti (2004, Chap. 51). Particular types of skills are highlighted by Bacolod, Blum, and Strange (2009b), which introduces the concept of soft skills that enable agents to interact in cities and industry clusters. More generally, the role of idea generation and exchange is emphasized by Davis and Dingel (2012), which develops a system of cities model in which costly idea exchange is the agglomeration force. Another line of research has distinguished different dimensions along which cities specialize. Duranton and Puga (2005) provides theory and evidence that in recent decades cities have shifted from specializing by sector—with integrated headquarters and plants—to specializing mainly by function—with headquarters and business services clustered in larger cities and plants clustered in smaller areas.4 Rossi-Hansberg, Sarte, and Owens (2009) develops a model in which firms choose locations of their headquarters and production facilities, and argues that the increased separation of these locations accounts for observed changes in patterns of residential and business activity. Ota and Fujita (1993) models the distinction between the front-unit (e.g., business office) and back-unit (e.g., plant or back-office) of firms and explores its implications for city structure. Helsley and Strange (2007) explicitly analyzes the vertical integration decision of the firm in conjunction with its location decision. Fujita and Tabuchi (1997) provides evidence that the increased separation of headquarters and production has contributed to observed changes in the distribution of economic activity across Japanese regions.5 Related research has examined the impact of roads and urban growth (e.g., Baum-Snow 2007; Duranton and Turner 2012) and the implications of innovations in communication technologies for cities (e.g., Pool 1977; Fischer 1992; Gaspar and Glaeser 1998; Leamer and Storper 2001). Our analysis is also related to the task-based approach to the labor market, including in particular Autor, Levy, and Murnane (2003, henceforth ALM), Acemoglu and Autor (2011, Chap. 12), Autor and Handel (2009), Autor (2013), Gray (2013), and Deming (2017). This research has developed measures of the task content of employment based on numerical scores from the Dictionary of Occupations (DOTs), such as “Direction, Control, and Planning (DCP)”. Prior to this research, the canonical model of the labor market in terms of skilled and unskilled labor had assumed that technological change is skill-biased. In contrast, this task-based approach recognizes that the direction of technological change is endogenous, as emphasized by Acemoglu (1998, 2002). Therefore, the extent to which new technologies complement or substitute for skills or tasks can change over time. In the labor literature, Autor, Katz, and Krueger (1998) argue that there was an acceleration in the skill-bias of technical change in the 1980s and 1990s. In the historical literature, several studies have argued that technical change often replaced rather than complemented skilled artisans in the 19th century, including Hounshell (1985), James and Skinner (1985), and Mokyr (1992).6 Relative to each of these lines of research, our main contribution is to develop a new approach for measuring individual production tasks that uses the verbs from occupational descriptions and their meanings. Using this new methodology, we are able to track the production tasks performed by workers at a much higher resolution than has hitherto been possible. We use our approach to provide evidence on changes in the task content of employment over a much longer historical time period than previously considered. We also apply this approach to the organization of economic activity in urban and rural areas separately. The remainder of the paper is structured as follows. Section 2 discusses the data. Section 3 introduces our methodology and presents our main empirical results on changes in the task content of employment in urban and rural areas over time. Section 4 outlines a theoretical framework for interpreting these empirical results that points to the role played by transportation and communication technologies. Section 5 provides further evidence on explanations for the observed changes in the task content of employment and their relationship with changes in transportation and communication technologies. Section 6 concludes. 2. Data Our empirical analysis uses two main sources of data. The first is individual-level records from the U.S. Population Census for 20-year intervals from 1880 to 2000 from Integrated Public Use Microdata Series (IPUMS): see Ruggles et al. (2010). These census micro data report individuals’ location, occupation, and sector, as well as other demographic information. We use these data to determine whether an individual is located in a metro area as well as the occupation and sector in which an individual is employed. A metro area is defined as a region consisting of a large urban core together with surrounding communities that have a high degree of economic and social integration with the urban core.7 We weight individuals by their person weights to ensure the representativeness of the sample. Our main data set is a panel from 1880 to 2000 that uses information on the share of employment within an occupation and sector in metro areas. To provide evidence on improvements in communication and transportation technologies, we also use long-differenced data from 1880 to 1930, aggregating our individual-level data to the county level. We use the standardized 1950 occupation classification from IPUMS, which distinguishes 11 two-digit occupations (e.g., “Clerical and Kindred”) and 281 three-digit occupations (e.g., “Opticians and Lens Grinders and Polishers”). We also use the standardized 1950 sector classification from IPUMS, which distinguishes twelve two-digit sectors (e.g., “Finance, Insurance and Real Estate”) and 158 three-digit sectors (e.g., “Motor Vehicles and Motor Vehicle Equipment”).8 Since we are concerned with employment structure, we omit workers who do not report an occupation and a sector (e.g., because they are unemployed or out of the labor force). We also exclude workers in agricultural occupations or sectors, because we compare task specialization in urban and rural areas over time, and agriculture is unsurprisingly overwhelmingly located in rural areas.9 We use time-varying definitions of metro areas to ensure that they correspond to meaningful economic units. However, we also report robustness tests, in which we hold the sample composition of metro areas constant over time, and in which we use administrative cities whose boundaries are more stable over time. Our second main data source is the Dictionary of Occupational Titles (U.S. Department of Labor 1991), which contains detailed descriptions of more than 12,000 occupations. Following Autor et al. (2003), previous research using DOTs typically uses the numerical scores that were constructed for each occupation by the Department of Labor (e.g., a Nonroutine Interactive measure based on the DCP numerical score). In contrast, we use verbs from the detailed occupational descriptions in DOTs to directly measure the tasks performed by workers in each occupation. We use a list of over 3,000 English verbs from “Writing English”, a company that offers English language consulting.10 This approach enables us to provide a rich analysis of the tasks undertaken in urban and rural areas using the verbs and occupational descriptions without being restricted to the numerical scores. Nonetheless, we also compare our measures of occupational characteristics to those from the numerical scores. We match the DOTs occupations to the three-digit occupations in our census data using the crosswalk developed by ALM. In our baseline specification, we use a time-invariant measure of tasks based on the occupational descriptions from the digital edition of the 1991 DOTs, which ensures that our results are not driven by changes in language use over time. In sensitivity checks, we also report results using digitized occupational descriptions from the first edition of the DOTs in 1939 (U.S. Department of Labor 1939). We complement these two main data sources with information from a variety of other sources. We use the standard reference for word usage in English (Roget’s Thesaurus) to quantify the meanings of verbs from the occupational descriptions.11 We use ArcGIS shapefiles from the National Historical Geographical Information System (NHGIS) to track the evolution of county boundaries over time. We also use measures of improvements in transport and communication technologies. We measure the length of roads in each county using a georeferenced 1931 road map (Gallup 1931).12 At the beginning of our sample in 1880, most U.S. roads were little more than dirt tracks (see, for example, Swift 2011) and widespread paved road construction only occurred following the Federal Aid Road Act of 1916 and the Federal Highway Act of 1921. Therefore, we use the 1931 map to construct a measure of the growth of the paved road network from 1880 to 1930. We measure the number of residence telephones in each county in 1935 using newly digitized data from American Telephone and Telegraph Company (AT&T 1935). The telephone was not patented until 1876 just before the beginning of our sample period and the telephone network developed rapidly from 1890 onward (see, e.g., Fischer 1992). Therefore, we use the data on telephones to construct a measure of the growth of telephones from 1880 to 1930. To address the concern that the road network could be influenced by changes in the interactiveness of economic activity, we use an instrument based on the “Pershing” map of highway routes of military importance for coastal and border defense. To address similar concerns for the telephone, we use an instrument based on proximity to primary and secondary outlets on AT&T’s long distance trunk network, whose construction was influenced by the strategic objective of connecting the nation as a whole. 3. Empirical Evidence on Task Specialization In this section, we present our main results on changes in the task content of employment in urban and rural areas over time. In Section 3.1, we begin by characterizing changes in specialization across occupations and sectors in metro areas relative to nonmetro areas over our long historical time period. In Section 3.2, we introduce our new methodology for measuring individual production tasks using the verbs from occupational descriptions. In Section 3.3, we explain how we quantify the meanings of these verbs using Roget’s thesaurus, and introduce a new measure of the interactiveness of the tasks undertaken by workers. In Section 3.4, we demonstrate the robustness of our results across a range of specifications. 3.1. Specialization Across Occupations and Sectors To provide some initial motivating evidence of changes in specialization across occupations and sectors in metro areas relative to nonmetro areas, we estimate the following regression for each year t separately using data across occupations o and sectors s: \begin{equation} \text{MetroShare}_{ost}=\mu _{ot}+\eta _{st}+\varepsilon _{ost}, \end{equation} (1) where MetroShareost is the share of employment in metro areas in occupation o, sector s and year t; observations are weighted by person weights; μot are occupation-year fixed effects; ηst are sector-year fixed effects; and εost is a stochastic error. We normalize the sector-year and occupation-year fixed effects so that they each sum to zero in each year, and hence they capture deviations from the overall mean in each year. Although we estimate the previous regression using a share as the left-hand side variable so that the estimated coefficients have a natural interpretation as frequencies, we find a very similar pattern of results in a robustness test in which we use a logistic transformation of the left-hand side variable: MetroShareost/(1 − MetroShareost). The occupation-year fixed effects (μot) capture the average probability of being in a metro area for workers in each occupation in each year, after controlling for differences across sectors in metro probabilities. Similarly, the sector-year fixed effects (ηst) capture the average probability of being located in a metro area for workers in each sector in each year, after controlling for differences across occupations in metro probabilities. The sector and occupation fixed effects are separately identified because there is substantial overlap in occupations and sectors, such that each sector contains multiple occupations and each occupation is employed in several sectors.13 We estimate this regression using both the aggregate (two-digit) and disaggregate (three-digit) definitions of occupations and sectors discussed previously. As reported in Table 1 for two-digit occupations and sectors, we find substantial changes in specialization across occupations and sectors in metro areas relative to nonmetro areas over time. From Panel A, in 1880, “Clerical and Kindred” workers were the most likely to be located in metro areas. In contrast, by 2000, “Clerical and Kindred” workers were ranked only fourth, and “Professional and Technical” workers were the most likely to be located in metro areas. From 1880 to 2000, declines in ranks were observed for “Craftsmen” (from 2 to 6) and “Operatives” (from 3 to 7), whereas increases in ranks were observed for “Professional and Technical” workers (from 7 to 1) and “Managers, Officials, and Proprietors” (from 6 to 3). As apparent from the first and fourth columns of the table, these changes in ranks reflect substantial changes in the probabilities of workers in individual occupations being located in metro areas over time. Table 1. Metro area specialization for aggregate occupations and sectors. Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Notes: Table reports the estimated coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on two-digit occupation-year and two-digit sector-year fixed effects (equation (1) in the paper). A separate regression is estimated for each year. Occupation and sector fixed effects are each normalized to sum to zero in each year. Standard errors are clustered by occupation. Occupations and sectors are sorted by the rank of their estimated coefficients for 1880. View Large Table 1. Metro area specialization for aggregate occupations and sectors. Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Coefficient 1880 Standard Error 1880 Rank 1880 Coefficient 2000 Standard Error 2000 Rank 2000 Panel A: Two-digit occupation Clerical and Kindred 0.15 0.08 1 0.04 0.01 4 Craftsmen 0.09 0.06 2 − 0.01 0.01 6 Operatives 0.06 0.07 3 − 0.05 0.01 7 Sales workers 0.01 0.07 4 0.05 0.01 2 Service Workers 0.00 0.08 5 0.00 0.01 5 Managers, Officials, and Proprietors − 0.03 0.08 6 0.05 0.01 3 Professional, Technical − 0.07 0.08 7 0.07 0.01 1 Laborers − 0.2 0.18 8 − 0.15 0.07 8 Panel B: Two-digit sector Entertainment and Recreation Services 0.29 0.08 1 0.04 0.01 4 Wholesale and Retail Trade 0.13 0.05 2 0.02 0.01 6 Finance, Insurance, and Real Estate 0.13 0.06 3 0.06 0.01 2 Manufacturing 0.06 0.05 4 − 0.01 0.01 10 Personal Services 0.01 0.06 5 0.03 0.01 5 Transportation, Communication, and Other Utilities 0.01 0.04 6 0.05 0.01 3 Public Administration − 0.03 0.07 7 0.01 0.01 7 Professional and Related Services − 0.03 0.06 8 0.00 0.01 9 Business and Repair Services − 0.12 0.08 9 0.08 0.01 1 Construction − 0.14 0.08 10 0.00 0.01 8 Mining − 0.31 0.05 11 − 0.27 0.03 11 Notes: Table reports the estimated coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on two-digit occupation-year and two-digit sector-year fixed effects (equation (1) in the paper). A separate regression is estimated for each year. Occupation and sector fixed effects are each normalized to sum to zero in each year. Standard errors are clustered by occupation. Occupations and sectors are sorted by the rank of their estimated coefficients for 1880. View Large Since the regression (1) includes sector-year fixed effects, these changes in the metro probabilities for each occupation are not driven by changes in sector composition, but rather reflect changes in the organization of economic activity within sectors. Nonetheless, we also observe substantial changes in sector structure in metro areas relative to nonmetro areas over time. From Panel B, declines in ranks from 1880 to 2000 were observed for “Wholesale and Retail Trade” (from 2 to 6) and “Manufacturing” (from 4 to 10). In contrast, increases in ranks from 1880 to 2000 were observed for “Transportation, Communication and Other Utilities” (from 6 to 3) and “Business and Repair Services” (from 9 to 1). In Figures A.1 and A.2 of the Online Appendix, we show the evolution of the occupation and sector coefficients across each of the 20-year intervals in our data. Although “Professional and Technical” workers display an increased propensity to locate in metro areas from 1880 to 1960, the probability that “Managers, Officials and Proprietors” are located in urban areas increases particularly sharply from 1940 to 2000. In contrast, the likelihood that “Craftsmen” are found in metro areas declines throughout our sample period, whereas the probability for “Clerical and Kindred” workers declines from 1900 onward, and the probability for “Service” workers initially rises until 1920 and later declines until around 1960. Such changes in specialization are not limited to the aggregate categories considered so far, but are also found using more disaggregated measures of occupations and sectors. In Table A.1 of the Online Appendix, we report the results of estimating the regression (1) including three-digit-occupation-year and three-digit-sector-year fixed effects. Panels A and B report the 20 occupations within the largest increases and decreases respectively in the within-sector probability of being located in a metro area from 1880 to 2000. Both the top agglomerating occupations in Panel A and the top dispersing occupations in Panel B are diverse and span multiple sectors. For example, leading agglomerating occupations include “Editors and Reporters”, “Buyers and Department Heads”, and “Judges and Lawyers”, whereas prominent dispersing occupations are “Book Binders”, “Welders and Flame Cutters”, and “Upholsterers”. In our empirical analysis in what follows, we provide evidence on the systematic characteristics shared by occupations that agglomerate versus disperse over time. 3.2. Measuring the Tasks Undertaken by Occupations We now introduce our new methodology for measuring individual production tasks using the detailed descriptions from more than 12,000 disaggregated occupations included in the DOTs. We use the verbs from each occupation’s description to measure the tasks performed by workers within that occupation, because verbs capture an action (bring, read, walk, run, learn), an occurrence (happen, become), or a state of being (be, exist, stand), and hence capture the task being performed. To focus on persistent characteristics of occupations and abstract from changes in word use over time, our baseline analysis uses time-invariant occupational descriptions from the 1991 digital edition of the DOTs. Although the tasks undertaken within each occupation can change over time, the relative task content of occupations is likely to be more stable. To provide evidence on the extent to which this is the case, we have also digitized the occupational descriptions from the first edition of the DOTs in 1939. Although the descriptions of occupations are less detailed and the boundaries between occupations are less clear in the historical DOTs, we find a similar pattern of results using both sets of occupational descriptions, as discussed further in what follows. The first step of our procedure uses a list of over 3,000 English verbs from “Writing English”, a company that offers English language consulting. Using this list of verbs, we search each occupational description in the 1991 DOTs for occurrences of each verb in the first-person singular (e.g., (I) talk), third-person singular (e.g., (she) talks) or present participle (e.g., (he is) talking).14 For example, the occupational description for an Economist is given as follows: “ECONOMIST: Plans, designs, and conducts research to aid in interpretation of economic relationships and in solution of problems arising from production and distribution of goods and services: Studies economic and statistical data in area of specialization, such as finance, labor, or agriculture. Devises methods and procedures for collecting and processing data, utilizing knowledge of available sources of data and various econometric and sampling techniques. Compiles data relating to research area, such as employment, productivity, and wages and hours. Reviews and analyzes economic data in order to prepare reports detailing results of investigation, and to stay abreast of economic changes ...”, where the words detected by our procedure as capturing the tasks performed by an economist are italicized.15 Note that sometimes the first-person singular, third-person singular or present participle forms of a verb have the same spelling as the corresponding adjectives and nouns (e.g., “prepare reports”). In this case, our procedure treats these adjectives and nouns as verbs. To the extent that the use of the same word as an adjective or noun is closely related to its use as a verb, both uses are likely to capture the tasks performed. From this first step, we obtain the number of occurrences of each verb for each DOTs occupation. We next match the more than 12,000 DOTs occupations to IPUMS standardized 1950 occupations using the crosswalk developed by ALM. Finally, we calculate the frequency with which each verb v is used for each IPUMS occupation o: \begin{equation*} \text{VerbFreq}_{vo}=\frac{\text{Appearances of verb }v\text{ matched to }o}{\text{Appearances of all verbs matched to }o}, \end{equation*} where we focus on the frequency rather than the number of verb uses to capture the relative importance of tasks for an occupation and to control for potential variation in the length of the occupational descriptions matched to each IPUMS occupation.16 We provide evidence on changes in task specialization in metro areas relative to nonmetro areas over time by estimating the following regression for each verb v and year t separately using data across occupations o and sectors s: \begin{equation} \text{MetroShare}_{ost}=\alpha _{vt}\text{VerbFreq}_{vo}+\eta _{vst}+\varepsilon _{ost}, \end{equation} (2) where MetroShareost is again the share of employment in metro areas in occupation o, sector s, and year t; VerbFreqvo is defined previously for verb v and occupation o; ηvst are verb-sector-year fixed effects; and εost is a stochastic error. The coefficient of interest αvt captures a conditional correlation: the correlation between occupations’ shares of employment in metro areas and their frequency of use of verb v. The verb-sector-year fixed effects (ηvst) control for differences across sectors in the frequency of verb use and for differences across sectors and over time in the concentration of employment in metro areas. Since VerbFreqvo is time invariant, a rise in αvt over time implies that employment in occupations using that verb is increasingly concentrating in metro areas within sectors over time. In Panels A and B of Table 2, we report for each year the ten verbs with the highest and lowest standardized coefficient αvt (the estimated coefficient multiplied by the standard deviation of VerbFreqvo).17 As apparent from Panel A, we find substantial changes in the tasks most concentrated in metro areas within sectors over time. In 1880, the verbs with the highest metro employment shares typically involve physical tasks such as “Braid”, “Sew”, “Stretch”, and “Thread”. By 1920, the top ten verbs include an increased number of clerical tasks, such as “Bill”, “File”, “Notice”, and “Record”. By 1980 and 2000, the leading metro verbs include a proliferation of interactive tasks, such as “Analyze”, “Advise”, “Confer”, and “Report”. As shown in Panel B, we also find some changes in the tasks least concentrated in metro areas, although here the pattern is less clear cut (e.g., “Tread” appears from 1880 to 1960 and “Turn” appears from 1960 to 2000). Table 2. Verbs most and least strongly correlated with metro area employment shares. Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Notes: Table reports the ranks of standardized coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which a verb is used for an occupation and three-digit sector-year fixed effects (equation (2) in the paper). A separate regression is estimated for each verb and year. Sector fixed effects are normalized to sum to zero in each year. Estimated coefficients are normalized by the standard deviation for the verb frequency. Verbs are sorted by the rank of their standardized coefficients. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large Table 2. Verbs most and least strongly correlated with metro area employment shares. Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Rank 1880 1900 1920 1940 1960 1980 2000 Panel A: Verbs most strongly correlated with metro area employment shares 1 Thread Thread File File Document Identify Develop 2 Stretch Stitch Distribute Bill Schedule Document Determine 3 Interfere Telephone Record Take File Advise Analyze 4 Hand Sew Notice Compile Record Concern Factor 5 Ravel Hand Telephone Distribute Distribute Report Review 6 Sew Assist Bill Pay Compile Schedule Confer 7 Braid Visit Envelope Letter Notice Develop Advise 8 Visit Describe Document Notice Identify Analyze Report 9 Receive Number Learn Record Send Determine Concern 10 Sack Stamp Number Send Notify Notify Plan Panel B: Verbs least strongly correlated with metro area employment shares 1821 Conduct Abstract Counsel Recur Accord Power Restrain 1822 Teach Tread Discuss Enlist Feed Pour Cut 1823 Channel Pinch Hear Labor Escape Erect Power 1824 Sound Assign Assign Tread Hook Clean Massage 1825 Rule Settle Teach Assign Traverse Massage Remove 1826 Matter Matter Matter Approve Tread Pump Feed 1827 Drill Tunnel Consolidate Extract Loosen Cut Clean 1828 Tread Sound Rule Tunnel Range Feed Pump 1829 Tunnel Rule Tunnel Malt Activate Move Move 1830 Pinch Sole Sound Establish Turn Turn Turn Notes: Table reports the ranks of standardized coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which a verb is used for an occupation and three-digit sector-year fixed effects (equation (2) in the paper). A separate regression is estimated for each verb and year. Sector fixed effects are normalized to sum to zero in each year. Estimated coefficients are normalized by the standard deviation for the verb frequency. Verbs are sorted by the rank of their standardized coefficients. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large 3.3. Quantifying Task Specialization The approach developed in the previous section allows us to provide a detailed characterization of the tasks performed in urban and rural areas using the full list of verbs and all occupational descriptions. In this section, we now develop a quantitative measure of task specialization based on the meanings of these verbs. To do so, we use the online computer-searchable version of Roget’s Thesaurus (1911), which has been the standard reference for English language use for more than a century, and explicitly classifies words according to their underlying concepts and meanings. Roget’s classification was inspired by natural history, with its hierarchy of Phyla, Classes, Orders, and Families. Therefore, words are grouped according to progressively more disaggregated classifications that capture ever more subtle variations in meaning. A key advantage of this classification is that it explicitly takes into account that words can have different meanings depending on context by including extensive cross-references to link related groups of words.18 Roget’s Thesaurus is organized into “Classes” that are further disaggregated into the progressively finer partitions of “Divisions”, “Sections”, and “Categories”. There are 6 classes, 10 divisions, 38 sections, and around 1,000 categories.19 The first three classes cover the external world: Class I (Abstract Relations) deals with ideas such as number, order and time; Class II (Space) is concerned with movement, shapes and sizes; and Class III (Matter) covers the physical world and humankind’s perception of it by means of the five senses. The last three classes relate to the internal world of human beings: the human mind (Class IV, Intellect), the human will (Class V, Volition), and the human heart and soul (Class VI, Emotion, Religion, and Morality). To characterize the meaning of each verb v, we use the frequency with which it appears in each partition k of Roget’s Thesaurus: \begin{equation} \text{ThesFreq}_{vk}=\frac{\text{Appearances of verb }v\text{ in category }k\text{ of thesaurus}}{\text{Total appearances of verb }v\text{ in thesaurus }}, \end{equation} (3) where the partition k could be a class, division, section, or category of the thesaurus; our use of a frequency takes into account that each verb can have multiple meanings and provides a measure of the relative importance of each meaning. In counting verb appearances, we make use of the thesaurus’s structure, in which words with similar meanings appear under each thesaurus Category in a list separated by commas or semicolons. Based on this structure, we count appearances of a verb that are followed by a comma or semicolon, which enables us to abstract from appearances of a word in idioms that do not reflect its common usage.20 Combining the frequency with which a verb appears in each occupation’s description (VerbFreqvo in the previous section) and the frequency with which the verb appears in each category of the thesaurus (ThesFreqvk), we construct a quantitative measure of the extent to which the tasks performed in an occupation involve the concepts from each thesaurus category,21 \begin{equation*} \text{TaskContent}_{ko}=\sum _{v\in V}\text{VerbFreq}_{vo}\times \text{ThesFreq}_{vk}. \end{equation*} We use this measure to examine changes in task specialization in metro areas relative to nonmetro areas over time by estimating an analogous regression for each thesaurus category k and year t as for each verb and year in the previous section: \begin{equation} \text{MetroShare}_{ost}=\beta _{kt}\text{TaskContent}_{ko}+\eta _{kst}+\varepsilon _{ost}, \end{equation} (4) where MetroShareost is the share of employment in metro areas in occupation o, sector s and year t; TaskContentko is defined previously for thesaurus partition k and occupation o; ηkst are thesaurus-category-year fixed effects; and εost is a stochastic error. The coefficient of interest βkt again captures a conditional correlation: the correlation between occupations’ shares of employment in metro areas and their frequency of use of verbs in thesaurus category k. The thesaurus-category-sector-year fixed effects (ηkst) control for differences across sectors in the frequency of use of thesaurus categories and differences across sectors and over time in the concentration of employment in metro areas. Since TaskContentko is time invariant, a rise in βkt over time implies that employment in occupations using that category of the thesaurus is increasingly concentrating in metro areas within sectors over time. In Table 3, we report the estimation results for the 38 sections of the thesaurus (denoted by S), organized by the 6 classes (denoted by C) and 10 divisions of the thesaurus (denoted by D). We calculate the standardized coefficient for each thesaurus section (the estimated coefficient βkt multiplied by the variable’s standard deviation) and report the ranking of these standardized coefficients in 1880 and 2000 as well the difference in rankings between these two years (1880 minus 2000).22 Since the thesaurus section with the highest standardized coefficient is assigned a rank of one, positive differences in rankings correspond to thesaurus sections that are becoming more concentrated in metro areas within sectors over time. Table 3. Ranking of thesaurus sections by concentration in metro areas in 1880 and 2000. Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Notes: Coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which the verbs used for an occupation are classified within thesaurus sections and three-digit sector-year fixed effects (equation (4) in the paper). A separate regression is estimated for each thesaurus section and year. Verbs are from the time invariant occupational descriptions from the 1991 Dictionary of Occupations. Thesaurus sections ranks in 1880 and 2000 based on their estimated coefficient normalized by the standard deviation for the thesaurus section frequency; the largest value is assigned a rank of one. The difference in ranks in the final column is defined such that a positive value corresponds to a thesaurus section that becomes more concentrated in metro areas from 1880 to 2000. View Large Table 3. Ranking of thesaurus sections by concentration in metro areas in 1880 and 2000. Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Thesaurus Class (C), Division (D), and Section (S) Rank Section 1880 Rank Section 2000 Difference C 1, Abstract relations, S I. EXISTENCE 15 12 3 C 1, Abstract relations, S II. RELATION 6 15 − 9 C 1, Abstract relations, S III. QUANTITY 1 34 − 33 C 1, Abstract relations, S IV. ORDER 23 9 14 C 1, Abstract relations, S V. NUMBER 24 10 14 C 1, Abstract relations, S VI. TIME 3 23 − 20 C 1, Abstract relations, S VII. CHANGE 34 11 23 C 1, Abstract relations, S VIII. CAUSATION 26 22 4 C 2, Space, S I. SPACE IN GENERAL 10 32 − 22 C 2, Space, S II. DIMENSIONS 4 36 − 32 C 2, Space, S IV. MOTION 19 27 − 8 C 3, Matter, S I. MATTER IN GENERAL 2 31 − 29 C 3, Matter, S II. INORGANIC MATTER 7 37 − 30 C 3, Matter, S III. ORGANIC MATTER 11 38 − 27 C 4, Intellect, D I, S I. OPERATIONS OF INTELLECT IN GENERAL 21 14 7 C 4, Intellect, D I, S II. PRECURSORY CONDITIONS & OPERATIONS 16 19 − 3 C 4, Intellect, D I, S III. MATERIALS FOR REASONING 25 7 18 C 4, Intellect, D I, S IV. REASONING PROCESSES 35 4 31 C 4, Intellect, D I, S V. RESULTS OF REASONING 33 5 28 C 4, Intellect, D I, S VI. EXTENSION OF THOUGHT 8 3 5 C 4, Intellect, D I, S VII. CREATIVE THOUGHT 38 21 17 C 4, Intellect, D II, S I. NATURE OF IDEAS COMMUNICATED. 27 1 26 C 4, Intellect, D II, S II. MODES OF COMMUNICATION 28 17 11 C 4, Intellect, D II, S III. MEANS OF COMMUNICATING IDEAS 32 18 14 C 5, Will, D I, S I. VOLITION IN GENERAL 14 29 − 15 C 5, Will, D I, S II. Prospective Volition 1 29 20 9 C 5, Will, D I, S III. VOLUNTARY ACTION 20 33 − 13 C 5, Will, D I, S IV. ANTAGONISM 30 16 14 C 5, Will, D II, S I. GENERAL INTERSOCIAL VOLITION 31 13 18 C 5, Will, D II, S II. SPECIAL INTERSOCIAL VOLITION 37 2 35 C 5, Will, D II, S III. CONDITIONAL INTERSOCIAL VOLITION 9 30 − 21 C 5, Will, D II, S IV. POSSESSIVE RELATIONS 13 8 5 C 5, Will, S V. RESULTS OF VOLUNTARY ACTION 22 25 − 3 C 6, Emotion, Religion, Morality, S I. AFFECTIONS IN GENERAL 5 35 − 30 C 6, Emotion, Religion, Morality, S II. PERSONAL AFFECTIONS 12 28 − 16 C 6, Emotion, Religion, Morality, S III. SYMPATHETIC AFFECTIONS 18 26 − 8 C 6, Emotion, Religion, Morality, S IV. MORAL AFFECTIONS 36 6 30 C 6, Emotion, Religion, Morality, S V. RELIGIOUS AFFECTIONS 17 24 − 7 Notes: Coefficients from a regression of the share of employment in metro areas within an occupation, sector, and year on the frequency with which the verbs used for an occupation are classified within thesaurus sections and three-digit sector-year fixed effects (equation (4) in the paper). A separate regression is estimated for each thesaurus section and year. Verbs are from the time invariant occupational descriptions from the 1991 Dictionary of Occupations. Thesaurus sections ranks in 1880 and 2000 based on their estimated coefficient normalized by the standard deviation for the thesaurus section frequency; the largest value is assigned a rank of one. The difference in ranks in the final column is defined such that a positive value corresponds to a thesaurus section that becomes more concentrated in metro areas from 1880 to 2000. View Large The results in Table 3 reveal a sharp change the relative ranking of thesaurus sections involving the external world (Classes I–III) and those involving the internal world of human beings (Classes IV–VI). In 1880, the top-five thesaurus sections most concentrated in metro areas included: Quantity (Class I), Time (Class I), Dimensions (Class II), Matter in General (Class III), and Affections in General (Class VI). In contrast, in 2000, the top-five thesaurus sections were: Nature of Ideas Communicated (Class IV), Special Intersocial Volition (Class V), Extension of Thought (Class IV), Reasoning Processes (Class IV), and Results of Reasoning (Class IV). The correlation between the rankings of the thesaurus sections in 1880 and 2000 is negative and statistically significant (−0.63). Positive changes in ranks in Table 3 are typically concentrated in thesaurus Classes IV and V, which correspond to the human mind and the human will, respectively. These classes include Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition). We summarize this combination of tasks—thought, communication and intersocial activity—as “interactiveness”. Our interpretation is that interaction inherently involves each of these components: thinking for oneself, communication these thoughts, and communicating them to other people in a social environment. We exclude Class V, Division 1 (Individual Volition) from our definition of interactiveness, because it is more concerned with individual reflection and decision making (e.g., motive, habit, willingness, choice) rather than interaction between people. Although some of the categories in Class VI could be interpreted as interactive (e.g., Section III, “Sympathetic Affections”), the other categories within this class seem to point more to contemplation and introspection than interaction between people (e.g., “Affections in General”, “Personal Affections”, “Moral Affection”, and “Religious Affection”). Furthermore, the interpersonal relationships described in “Sympathetic Affections” seem to largely concern relationships outside of work. Therefore, we measure the interactiveness of an occupation using the frequency with which verbs appear in that occupation’s description and the frequency with which those verbs appear in Divisions 1 and 2 of Class IV and Division 2 of Class V of the thesaurus: \begin{equation} \text{Interactive}_{o}=\sum _{v\in V}\text{FreqVerb}_{vo}\times \text{FreqInteractive}_{v}, \end{equation} (5) where FreqVerbvo is the frequency with which verb v is used for occupation o from above; FreqInteractivev is the frequency with which verb v appears in these partitions of the thesaurus (computed as in (3)). We also report results in what follows for all four divisions of Classes IV and V of the thesaurus. In Panels A and B of Table 4, we report the top ten and bottom ten interactive occupations using our measure. Although any single quantitative measure of interactiveness is unlikely to capture the full meaning of this concept, the occupations identified by our procedure as having high and low levels of interactiveness appear intuitive. “Buyers and Department Heads”, “Clergymen”, and “Pharmacists” arguably perform more interactive tasks than “Blasters and Powdermen”, “Roofers and Slaters”, and “Welders and Flame Cutters”. Table 4. Most and least interactive occupations. Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Notes: The table reports the ten occupations with the lowest and highest interactiveness, as measured by the frequency of verb use in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of Roget’s Thesaurus. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large Table 4. Most and least interactive occupations. Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Panel A: Top ten interactive occupations  Economists  Nurses, professional  Pharmacists  Clergymen  Religious workers  Accountants and auditors  Postmasters  Buyers and dept heads, store  Aeronautical-Engineers  Statisticians and actuaries Panel B: Bottom ten interactive occupations  Brickmasons, stonemasons, and tile setters  Attendants, auto service, and parking  Painters, except construction or maintenance  Plumbers and pipe fitters  Upholsterers  Asbestos and insulation workers  Welders and flame cutters  Blasters and powdermen  Dressmakers and seamstresses except factory  Roofers and slaters Notes: The table reports the ten occupations with the lowest and highest interactiveness, as measured by the frequency of verb use in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of Roget’s Thesaurus. Verbs are from the time-invariant occupational descriptions from the 1991 Dictionary of Occupations (DOTs). View Large In Figure 1, we measure the interactiveness of metro areas, nonmetro areas, and the economy as a whole using the employment-weighted average of interactiveness for each occupation. In this measure, interactiveness only differs between metro and nonmetro areas to the extent that they have different distributions of employment across occupations: \begin{equation} \text{Interactive}_{jt}=\sum _{o=1}^{O}\frac{E_{ojt}}{E_{jt}}\text{Interactive }_{o},\qquad j\in \left\lbrace M,N\right\rbrace, \end{equation} (6) where j indexes a type of location and we again denote metro areas by M and nonmetro areas by N;Eojt corresponds to employment in occupation o in location type j ∈ {M, N} in year t. Figure 1. View largeDownload slide Mean interactiveness in metro and nonmetro areas over time. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Figure 1. View largeDownload slide Mean interactiveness in metro and nonmetro areas over time. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. In 1880, metro and nonmetro areas have similar levels of interactiveness, with if anything metro areas having lower interactiveness than nonmetro areas. Over time, interactiveness increases in both sets of locations, but this increase is greater in metro areas than in nonmetro areas. This increase in the relative interactiveness of metro areas is particularly sharp from 1900 to 1920, which coincides with the dissemination of improvements in communication and transport technologies in the form of the telephone and roads and the automobile. In our empirical analysis in what follows, we provide further evidence on the extent to which changes in interactiveness are related to these new communication and transport technologies. 3.4. Robustness Having presented our baseline evidence of an increase in the interactiveness of employment in metro areas relative to nonmetro areas over time, we now document the robustness of this finding across a range of different specifications. 3.4.1. 1939 DOTs Our baseline specification measures the task content of employment using time-invariant occupational descriptions from the 1991 DOTs. Although this approach ensures that our findings are not driven by changes in language use over time, it assumes that the relative task content of occupations is persistent over time. One concern is that the interactiveness of occupations could have changed over time and these changes in interactiveness could be correlated with occupations’ shares of employment in metro areas. To address this concern, we replicated our analysis using the first edition of the DOTs from 1939. We digitized the occupational descriptions in the 1939 DOTs and implemented our procedure of searching for verbs in each occupational description. The boundaries between occupations are less well defined and the occupational descriptions are less detailed in the 1939 DOTs, which implies that the resulting measures of the task content of employment are likely to be less precise than those using the 1991 DOTs. Nonetheless, as reported in Table A.2 of the Online Appendix, we find similar changes in task specialization in this robustness test. The verbs most correlated with metro employment shares in 1880 include physical tasks such as “Retouch”, “Trawl”, and “Lure”. In contrast, the verbs most correlated with metro employment shares in 2000 include interactive tasks such as “Advise”, “Question”, and “Appraise”. Using the verbs from the 1939 occupational descriptions and the frequency with which these verbs appear in Class IV and Division 2 of Class V of the thesaurus, we again find an increase in the interactiveness of employment over time that is more rapid in metro areas than in nonmetro areas, as shown in Figure A.3 in the Online Appendix. This similarity of the results using both the 1939 and 1991 occupational descriptions suggests that our findings are unlikely to be driven by changes in the relative interactiveness of occupations over time. Indeed, although the layout of the occupational descriptions implies that our measure of interactiveness using the 1939 DOTs is less precise than our baseline measure using the 1991 DOTs (which by itself would induce an imperfect correlation), we find that they are positively and statistically significantly correlated. As reported in Table A.3 of the Online Appendix, the unweighted correlation coefficient between the 1939 and 1991 measures across the sample of occupations in 2000 is 0.62. 3.4.2. Metro Areas and Administrative Cities Our analysis has so far used variation between metro and nonmetro areas. To provide further evidence of a relative increase in the interactiveness of employment in densely populated locations, we now present evidence using a different source of variation across metro areas of differing population densities. In the top-left and top-right panels of Figure 2, we display mean interactiveness for each metro area (as calculated using (6)) against log population density for 1880 and 2000, respectively, as well as the fitted values and confidence intervals from locally weighted linear least squares regressions. To make the panels more legible, we omit a few outliers on both ends of the distribution from the figure (but not from the locally weighted linear least squares regressions). We use time-varying definitions of metro areas to ensure that they correspond to meaningful economic units, which implies that the number of observations changes over time as new metro areas enter the sample, as can be seen from comparing the two panels. In 1880, we find little relationship between interactiveness and log population density across metro areas, which is reflected in a negative but statistically insignificant OLS coefficient (standard error) of −0.0002 (0.0013). In contrast, in 2000, we find a positive and statistically significant relationship between interactiveness and log population density, which is reflected in an OLS coefficient (standard error) of 0.0018 (0.0002). In the bottom-left panel of Figure 2, we show that even when we restrict the 2000 sample to metro areas that exist in 1880, we continue to find a positive relationship that is statistically significant at the 10% level, confirming that these findings are not driven by a change in the composition of metro areas. Therefore, the increase in the relative interactiveness of densely populated locations over time is observed not only comparing metro and nonmetro areas but also comparing metro areas of differing population densities. Metro areas with relatively high levels of interactiveness conditional on population density in 2000 include Boston (BOS, MA) and New York (NYC, CT/NY/NJ), whereas those with low levels of interactiveness conditional on population density include Anniston (ANN, AL) and Mansfield (MAN, OH). Figure 2. View largeDownload slide Mean interactiveness across metro areas in 1880 and 2000. X-axes are log population density. Y-axes are mean interactiveness. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Thick solid lines are the fitted values from locally weighted linear least squares regressions. Thin solid lines are 95% point confidence intervals. Figures (but not the regressions) are truncated at both ends for outliers. Figure 2. View largeDownload slide Mean interactiveness across metro areas in 1880 and 2000. X-axes are log population density. Y-axes are mean interactiveness. Mean interactiveness is the employment-weighted average of interactiveness for each occupation. Interactiveness for each occupation is measured using the frequency with which verbs from time-invariant occupational descriptions from the 1991 DOTs appear in Class IV, Division 1 (Formation of Ideas), Class IV, Division 2 (Communication of Ideas), and Class V, Division 2 (Intersocial Volition) of the thesaurus. Thick solid lines are the fitted values from locally weighted linear least squares regressions. Thin solid lines are 95% point confidence intervals. Figures (but not the regressions) are truncated at both ends for outliers. Although we use time-varying definitions of the boundaries of metro areas to ensure that they correspond to meaningful economic units, we find similar results if we instead define urban areas as administrative cities, which have much more stable geographical boundaries over time. Again we find an increase in the relative interactiveness of urban areas over time, whether we compare administrative cities to all other locations (Figure A.4 in the Online Appendix) or only to nonmetro areas (Figure A.5 in the Online Appendix). Therefore, the increase in the relative interactiveness of urban areas also occurs within existing geographical boundaries. 3.4.3. Other Occupational Characteristics Our approach of using verbs from the occupational descriptions enables us to measure individual production tasks at a much finer level of resolution than has hitherto been possible. We now compare aggregations of our individual task measures, such as interactiveness, to existing measures of tasks, including the numerical scores from the DOTs used by ALM. Since these numerical scores are not available in the first edition of the DOTs in 1939, we use their values from the 1991 digital edition of the DOTs. As a point of comparison, in Figure A.8 of the Online Appendix, we show the employment-weighted average of the five ALM measures of task inputs over our long historical time period (analogous to Figure 1 in ALM but for 1880–2000 instead of 1960–2000). In Figures A.9– A.10 of the Online Appendix, we show the employment-weighted average of our measures of task inputs based on the 38 thesaurus sections. As is clear from these figures, there is a lot more variation using our thesaurus measures than using the five numerical scores. Therefore, we obtain a much richer picture of changes in task inputs over time and in urban versus rural areas using our measures based on the meanings of verbs from occupational descriptions. In Table A.3 of the Online Appendix, we report the correlation coefficients between our interactiveness measure and other measures of occupation characteristics across the sample of occupations in 2000. We report both unweighted correlations and correlated weighted by occupation employment. The highest correlation coefficients are for the Nonroutine Interactive (DCP) and Nonroutine Analytic (MATH) used by ALM. Although both of these measures are related to the concepts of thought, communication and intersocial activity captured by our interactiveness measure, the correlations are around 0.5. Therefore, our interactiveness measure captures distinctive information about the tasks performed by workers within occupations. Although DCP is orientated toward top–down interactions between workers (e.g., between a manager and her subordinates), our measure captures all interactions between workers (e.g., between members of a product design team). Although MATH is orientated toward thought, our measure of interactiveness also captures communication and intersocial activity. As a further check on our measure of interactiveness, Table A.4 of the Online Appendix reports the top-five verbs concentrated in each thesaurus section. We find that the verbs most concentrated in those sections of the thesaurus