This article provides the first rigorous estimates of how industrial air pollution from coal burning affects long-run city growth. I introduce a new theoretically grounded strategy for estimating this relationship and apply it to data from highly polluted British cities from 1851 to 1911. I show that local industrial coal use substantially reduced long-run city employment and population growth. Moreover, a counterfactual analysis suggests that plausible improvements in coal-use efficiency would have led to a higher urbanisation rate in Britain by 1911. These findings contribute to our understanding of the effects of air pollution and the environmental costs of industrialisation.
From the mill towns of nineteenth-century England to the mega-cities of modern China and India, urbanisation has often gone hand in hand with pollution. Much of this pollution comes from industry, a by-product of the job-creating engines that drive city growth. This pollution, in turn, represents a disamenity that can act as a drag on urban growth. As a result, policymakers face a trade-off between encouraging the growth of industry and increasing the costs associated with local pollution. Yet, despite substantial research into the effects of air pollution, when it comes to understanding how air pollution may impact long-run local economic growth we have virtually no rigorous evidence to rely on. This matters: local policymakers in developing countries regularly face important choices about whether to encourage the growth of polluting industries in their area. High among their concerns are how these decisions will affect job growth over the following years or decades.
A classic line of research in urban economics examines the impact of industrial structure on city growth through local external effects (Glaeser et al., 1992; Henderson et al., 1995; Glaeser et al., 1995). Amenities and other public goods, including environmental quality, are also thought to play a central role in city success. This study offers a link between these two lines of research, by showing how local industrial structure can influence city amenities, specifically environmental quality, and offering a new, theoretically grounded, analysis strategy that can be used to measure these effects. This study contributes to a growing literature examining endogenous changes to local amenity levels, such as Diamond ( 2016), but differs from previous studies by focusing on pollution and the link to local industrial structure. As a consequence, it sheds light on one important mechanism through which industrial structure influences long-run city growth.
In order to study how pollution affects long-run city growth, three challenges must be overcome. First, air pollution is just one of many factors that influences city growth, and its effects may take years to develop. Thus, identifying the relationship between air pollution and city growth requires a setting in which one can observe a number of industrial cities that experience high, and highly variable, levels of pollution over multiple decades. This essentially rules out studying cities in modern developed countries, where air pollution levels are relatively low. This raises a second challenge. In highly polluted industrial cities, including both modern developing cities and historical industrial areas such as Britain, data tracking air pollution over long periods are typically unavailable. Thus, one needs a method for inferring pollution levels that does not rely on direct pollution measures. Third, new analytical methods are needed in order to separate the positive direct effect that growth in local industry can have on city employment from the negative indirect effects of any pollution that the industry generates.
This study overcomes these challenges in order to offer the first rigorous evidence documenting the impact of industrial air pollution on long-run city growth. To do so, I turn to a historical setting: British cities in the late nineteenth and early twentieth centuries. This setting was characterised by very high levels of air pollution due, in large part, to industrial coal burning. Not only were pollution levels high; they varied substantially—an important feature that allows me to separate pollution effects from other factors that impacted city growth. This setting also offers a sufficient number of industrial cities for statistical power, as well as rich data on employment broken down by city and industry. In addition, I am able to infer industrial emissions of coal smoke—the most important pollutant, based on coal consumption by industry—which allows me to get around the lack of direct pollution measures.
I begin by offering a new analytical framework for estimating the effect of industrial pollution on long-run city employment growth. This framework extends a standard Rosen-Roback model to accommodate many industries that are heterogeneous in their use of a polluting input, coal. The theory delivers a new estimation approach that allows me to separate the positive effect of industry growth on local employment growth, through job creation, from the negative effects that are generated when this growth occurs in heavily polluting industries. These negative pollution effects, which operate on all industries in a city, can occur either because pollution makes a location less attractive (the amenities channel), or because pollution makes workers and firms less productive (the productivity channel). My estimation strategy will capture the impact of pollution occurring through either of these channels. In addition, this strategy can be implemented without the need for local wage and rent data, which are largely unavailable during the period that I study. Instead, the model shows how data on quantities, in this case the quantity of employed workers, can be used in place of the more scarce data on prices (real wages in this case). As a result, my approach requires only panel data on city-industry employment, which I have constructed for every decade from 1851 to 1911 for 31 English cities.
My results show that industrial coal use substantially reduced long-run employment growth in English cities during this period. Specifically, in English cities that experienced rapidly rising industrial coal use, employment growth was systematically lower relative to the growth that we would have expected given the initial mix of industries in each city and national industry growth rates. The magnitude of these effects was large: based on my estimates, over a two-decade period, a city in which local industrial coal use grew at a rate that was one standard deviation above the national average would, as a consequence, have experience a reduction in employment growth of 21–26 percentage points, equal to about one-half of the average growth in employment across two-decade periods. These estimates reflect the external effect that coal use in some industries exerted on other sectors of the local economy. These findings are robust to the inclusion of a wide range of control variables reflecting factors that urban economists most commonly cite as influencing city growth.
To quantify the effect on overall urbanisation levels, I conduct a simple counterfactual looking at the impact of more efficient coal use. This counterfactual is motivated by the 1871 Coal Commission Report—a detailed 1,300-page study of coal use in Britain commissioned by Parliament. The report highlights a number of inefficiencies in industrial coal use and describes how simple low-cost improvements could have substantially reduced industrial coal use, and thus coal-based pollution. However, these improvements were not adopted due to the combination of low coal prices, weak pollution regulation, and the fact that most of the impacts of pollution were external to firms. Guided by this report, I consider a counterfactual in which the growth of coal use from 1851 to 1911 was reduced by 10%. My results suggest that the 31 analysis cities would have had an additional 1.5 million residents by 1911 and that their share of the English population would have been higher by four percentage points. Thus, my results suggest that had Britain adopted regulations to improve coal use efficiency the nation would have been substantially more urbanised by the early twentieth century.
To my knowledge this is the first study to document the effects of industrial pollution on local economic development over the long run, though I build on previous work such as Kahn ( 1999). 1 This is possible, in part, because of the unique features offered by the historical setting that I consider. Among the important features of this setting are the high variation in the level of local pollution, the high level of population mobility, which meant that city population and employment could respond to the effects of pollution, and the fact that regulation, including both pollution regulation and urban regulations such as zoning, were extremely limited. 2
This study highlights the fact that city employment growth can be impacted by pollution either through the effect on local amenities, which affects the supply of workers, or because pollution makes workers less productive, affecting the demand for workers. The model makes it clear that regardless of whether coal use affects consumer amenities or firm productivity, the implications for employment are the same. Thus, focusing on employment as the outcome of interest allows me to capture the combined effect of both of these channels. This contrasts with previous work on this topic, such as Williamson ( 1981b), which has focused only on the amenity channel by looking at the wage premium in more polluted cities. However, a growing body of literature suggests that air pollution can have important effects on productivity. 3 The amenity and productivity channels have opposing effects on the urban wage premium, so if the productivity channel is important then a small urban wage premium can still be consistent with large pollution costs. Thus, the model makes it clear that when pollution affects productivity the costs of urban pollution cannot be inferred from the urban wage premium alone. Using a cross-section of local wage, rent and price data from 1905, I provide tentative evidence that the productivity effects of coal use were particularly important during the period that I study, which suggests that approaches that ignore the impact of pollution on worker productivity may be missing much of the effect of local pollution on employment growth.
This study is connected to existing historical studies on pollution effects, including Troesken and Clay ( 2011), Barreca et al. ( 2014), Clay et al. ( 2016; 2018), Heblich et al. ( 2016), Beach and Hanlon ( 2018) and Hanlon ( 2018). However, none of these studies looks at the impact of air pollution on long-run local development, nor am I aware of any modern studies that estimate such effects. My results also have implications for a long-running debate over the cost of the disamenities generated by industrial growth in nineteenth-century Britain. 4 In a series of articles, Jeffrey Williamson argued that the lack of a large urban wage premium implied that conditions in nineteenth-century British cities were not as bad as contemporary reports suggest. While I replicate Williamson’s results, I also show that his analytical approach missed the large negative effect of pollution on productivity, which led him to conclude, incorrectly, that industrial pollution did not have substantial negative consequences. Along the way, my results reconcile the quantitative estimates of the costs of industrial pollution during the Industrial Revolution with the qualitative historical evidence describing the severity of the pollution problem during this period as well as with our current understanding of the substantial impacts that air pollution can have, even at the much lower concentrations experienced in modern developed economies. 5
In the next section, I describe the empirical setting. Data and measurement are discussed in Section 2, followed by the theory, in Section 3. The analysis is presented in Section 4, while Section 5 concludes.
Landes ( 1998) describes the Industrial Revolution as being composed of three elements: the replacement of human skill by machines, the introduction of engines to convert heat into work, and the substitution of mineral power sources—chiefly in the form of coal—for other power sources. One consequence of these changes was rapid growth in coal use by industry, particularly in the second half of the nineteenth century. British coal consumption averaged 65 million tons annually in 1852–62 and rose to 181 million tons in the 1903–12 period. 6 This amounted to 4.3 tons per person in 1911. 7 Most of this coal (60–65%) was burned by industry, and coal remained the dominant power source, by far, throughout this period. 8 While electricity use was growing during the latter portion of the study period, even by 1907 electricity powered only one-ninth of the motor power used in manufacturing (Hannah, 1979), and essentially none of the most coal-intensive processes, like blast furnaces. Even where electricity was used, it was typically generated on site at factories by burning coal (Hannah, 1979). Because some industries were particularly intensive users of coal, and these industries tended to agglomerate, industrial coal use could be highly geographically concentrated. 9 Also, before long-distance electricity transmission, power had to be generated on site at factories, which were located in urban areas where they could be reached on foot by workers, thereby increasing pollution exposure.
The pollution released by coal burning factories in nineteenth-century Britain was widely recognised and discussed. For example, The Times 10 wrote,
(The Times, 7 Feb. 1882, p. 10)‘There was nothing more irritating than the unburnt carbon floating in the air; it fell on the air tubes of the human system, and formed a dark expectoration which was so injurious to the constitution; it gathered on the lungs and there accumulated.’
While pollution in London was more likely to be experienced by visitors and noted by the press, coal smoke pollution was particularly severe in the industrial cities of England. For example, describing a visit to north-west England in 1890, Cannon Hardwicke Drummond Rawnsley wrote,
(Quoted from Mosley, 2001, p. 24)‘ . chimneys, solid and square, were belching forth clouds of Erebean darkness and dirt . The heavens were black with smoke, and the smother of the mills, to one whose lungs were unaccustomed to breathing sulphurised air, made itself felt.’
Figure 1 provides an illustration of the impact of industrial pollution in Sheffield, perhaps the most polluted of the northern industrial cities. These images come from 1920, after the end of the study period, but are likely to be similar to the conditions experienced during the late nineteenth and early twentieth centuries. The left-hand image was taken on Sunday morning, when the factories were at rest, while the right-hand image was taken from the same vantage point on Monday at noon, when the factories were at work. Residential pollution would have been present at both times, so the contrast between these images illustrates the impact that industrial pollution had in the industrial cities of England.
An Illustration from Sheffield in 1920.
Notes: The pictures above were taken from the same vantage point in Sheffield in 1920. While this is after the study period, the levels of pollution that it reveals are likely similar to those experienced during the period I study. From William Blake Richmond (1921), ‘The Smoke Plague of London’, in (Sir Aston Webb, ed.), London of the Future, London: The London Society. Reproduced from Peter Thorsheim ( 2006),Inventing Pollution, Athens, OH: Ohio University Press.
While the health effects of air pollution were not fully understood by contemporaries, there was some appreciation for the link between coal-based air pollution and poor health. 11 Today we know that burning coal releases a variety of pollutants into the atmosphere, including suspended particles of soot and other matter, sulphur dioxide, and carbon dioxide. The release of suspended particles is particularly severe when combustion is inefficient, as it often was in the nineteenth century. These pollutants have a variety of negative effects on the human system, which have been documented in a large literature. 12
Several recent studies have documented the impact of pollution exposure on worker productivity. For example, Graff Zivin and Neidell ( 2012) show that ozone exposure reduced the productivity of agricultural workers. Using data from Mexico City, Hanna and Oliva ( 2015) show that air pollution can impact hours worked. He et al. ( 2019) documents the impact on manufacturing workers in China. Chang et al. ( 2016a) shows that day-to-day variation in particulate pollution exposure lowered the productivity of pear packers. Their estimates suggest that the relatively small reductions in PM2.5 particulates achieved in the United States from 1999 to 2008 generated $16.5 billion in labour cost savings. Chang et al. ( 2016b) uses evidence from call-centre workers in China to show that the productivity effects of air pollution exposure extend even to white-collar jobs. Lichter et al. ( 2017) show effects on German soccer players. In addition, early-life pollution exposure has been linked to a range of negative outcomes, including on cognitive ability and human capital formation (Ebenstein et al., 2016; Bharadwaj et al., 2017) and adult earnings (Isen et al., 2017).
An important feature of this empirical setting is that Britain was a ‘highly mobile’ (Long and Ferrie, 2003) society during this period, with large flows of population from rural areas as well as Ireland and Scotland into English cities. 13 This means that, when considering factors that influence city employment or population growth, the marginal mover that we should have in mind was someone outside the cities who was faced with a decision about where to migrate. The search for work was the primary driver of these migration flows, though there is also some evidence that pollution levels affected location decisions, both within and across cities. 14
Another important feature of this setting was the limited level of government regulation, including both pollution regulation and other regulations that would have affected city growth. While some steps were taken to regulate industrial pollution, these efforts often ran up against the laissez-faire ideology that dominated British policy during this period, as well as the political power of mill owners. New pollution regulations were passed, including the Sanitary Act of 1866, the Public Health Act of 1875 and the Public Health (London) Act of 1891. However, these acts allowed for substantial interpretation, contained important loopholes and imposed relatively small fines. 15 As a result, historical evidence suggests that their effectiveness was limited, though they may have had more impact toward the end of the nineteenth century. 16 Other regulations affecting city growth, such as zoning laws, were also largely absent from this setting, which provides a particular clean opportunity for investigating the impact of pollution on city growth. 17
The first key piece of data for this study is a measure of local industrial composition. These data come from the Census of Population, which reports the occupation of each person at each ten-year census interval from 1851 to 1911 for 31 of the largest cities in England. 18 The occupational categories reported in these data generally closely correspond to industries, such as cotton spinner or steel manufacturer. 19 To construct consistent series for 1851–1911, I combine the many occupational categories available in each census into a set of 26 broad industries, spanning nearly the entire private-sector economy. 20
Because I am working with fairly aggregated industry categories, almost all industries are present in all cities. 21 However, the spread of industries across cities was far from even. For example, textile producers agglomerated in cities in Lancashire and Yorkshire, where they could account for as much as half of all private-sector employment. Cities such as Sheffield, Birmingham and Wolverhampton had a disproportionate share of metals industries, while ports such as Bristol and Liverpool had high shares of transportation and services.
The second necessary piece of information for this study is a measure of the coal intensity of each industry. This information is drawn from the first Census of Production, which was completed in 1907. 22 This Census collected detailed information on the amount of coal used in each industry, as well as industry employment, allowing me to construct a measure of coal use per worker in each industry. 23
These data show that coal use intensity varied enormously across industries—a feature that plays a key role in this study. A table describing coal use intensity by industry is available in Online Appendix A.2.5. The most intensive industrial coal users, such as metal and machinery, or earthenware and bricks, used coal to heat material up to high temperatures. These industries used more than 40 tons per worker per year. Textiles, a moderate coal-using industry that consumed around 10 tons per worker per year, generally used coal to power steam engines. Other industries, such as apparel or tobacco products, used very little coal (less than 2 tons per worker per year). This large variation in coal use intensity at the industry level, together with the tendency of industries to agglomerate in particular locations, resulted in substantial variation in the amount of industrial coal use at the city level.
I model industrial coal use in cities as determined by city-industry employment (Lict), the coal use intensity of each industry (θi) and the national efficiency of coal use per worker, ρt:
$$\begin \textit _= \rho _t \sum _ \left( L_\textit \times\theta _i \right). \end$$Estimates of θi for manufacturing industries are provided by the 1907 Census of Production, while Census of Population data provide city-industry employment. The ρt term can be calculated by comparing data on industrial coal use at the national level to the values obtained using data on θi and Lit. 24 The ρt term is included here mainly for completeness and is not crucial to main estimates, though it will matter for counterfactuals. In general, other industries, such as services, were not likely to be major coal users, so this measure should capture most industrial coal use. 25
One assumption implicit in this approach is that relative coal use per worker across industries did not vary too much over time. Another important assumption is that industry coal use does not vary too much across locations in response to variation in the relative level of wages or coal prices. Put another way, it is important that variation in city coal use due to local industry composition and differences in industry coal use intensity resulting from technological factors is substantially more important than the variation due to differences in the local prices of coal or other inputs. The enormous variation in coal use intensity across industries is important for making this a reasonable assumption.
One way to check both of these assumptions is to compare estimated levels of coal use calculated using the method described here to data on local coal use levels. While such data are generally unavailable, there is information on county-level coal use in the 1871 Coal Commission report. Comparing estimates of industrial coal use at the county level for 1871, based on the approach I have just described, to county-level coal use data from the 1871 report shows that my approach does a good job of replicating industrial coal use at the county level (the correlation is 0.912), particularly for more industrial and urbanised locations. The full analysis is available in Online Appendix A.2.7.
It is also possible to check the extent to which industry coal use varied over time by comparing the 1907 data to data from the 1924 Census of Production, the next full production census. This analysis, described in Online Appendix A.2.6, shows that the relative coal use intensity across industries was quite stable over time, even across a period in which the British economy was hit by enormous shocks. 26 The fact that relative industry coal use intensity remains quite stable over time suggests that variation in coal use is largely due to industry fundamentals, rather than being a response to more fleeting industry-specific conditions. The relatively fixed nature of industry coal use intensity strengthens my identification strategy, by reducing concerns that this variable might be endogenous to current economic conditions. Also, comparing 1907 and 1924 coal use per worker suggests that there was broad improvement in coal use efficiency over time, which occurred relatively evenly across industries. This type of efficiency improvement will be captured in the ρt term.
Estimates of industrial coal use per worker at the city level are described in Table A1 in Online Appendix A.2.4. These data show that there was substantial variation across cities in the expected level of coal use per worker, even among similarly sized cities. Sheffield, often cited as the prototypical polluted industrial city, emerges as the most intensive user of coal in the database, followed by other cities specialising in metals such as Birmingham and Wolverhampton. Textile manufacturing towns, such as Manchester and Leeds, show moderate levels, near the average. Commercial and trading cities, such as Liverpool and Bristol, as well as London, use industrial coal less intensively. Bath, a resort town, is the least polluted city in the database.
This section presents a spatial equilibrium model in the Rosen-Roback tradition, but modified in a few important ways in order to fit the empirical setting. The economy is made up of a fixed number of cities, indexed by c. These cities are small open economies that take goods prices as given. As is standard in spatial equilibrium models, workers and firms can move freely across cities and goods are freely traded. I begin by modeling the demand for labour in cities.
The economy is composed of many industries, indexed by i, each of which produce a homogeneous good. Each industry is composed of many perfectly competitive firms, indexed by f. Firms produce output using labour, a polluting input (coal), and a fixed local industry-specific resource. 27 The production function is,
$$\begin y_\textit = \, a_ \textit L_\textit^ C_\textit^ R_\textit^, \end$$where Lfict is labour, Cfict is coal, Rfict is a local resource, and aict is the local productivity level in industry i. Let αi, βi ∈ [0, 1) for all i, and αi + βi < 1 for all i. Note that the production function parameters are allowed to vary at the industry level. This will result in industries employing different input mixes, with some using coal more intensively than others.
Local resources are fixed within each city and are industry specific, with an available supply given by |$\bar_$| . 28 These resources can be thought of as natural features or local endowments of entrepreneurial ability in a particular sector. They play an important role in the model; by introducing decreasing returns at the city-industry level, they allow multiple cities to be active in an industry even when productivity varies across cities, trade is costless, and markets are perfectly competitive.
Firms maximise profit subject to output prices pit, the coal price ϕt, a city wage |$w_$| , and the price of local resources χict. The firm’s maximisation problem in any particular period is,
Using the first order conditions from this problem, I obtain the following expression for the relationship between employment and coal use in each industry,
$$\begin \frac>> = \left(\frac\right)\left(\frac<\phi _t>\right) w_\textit. \end$$This expression tells us that variation in the use of polluting inputs across industries will be governed in part by the industry-specific production function parameters αi and βi. The empirical analysis exploits the exogenous variation due to the βi |$/$| αi parameters, reflected by the θi term in ( 1), while abstracting from the variation due to the endogenous wct term. The (1 |$/$| ϕt) term in ( 2) implies that coal use per worker can vary over time in a way that is common to all industries: a feature that is reflected in the ρt term in ( 1).
It is worth emphasising that the expression in ( 2) maps directly into the coal use values calculated using ( 1). The fact that those coal use values do a good job of reproducing observed coal use levels in 1871 (see Online Appendix A.2.7), suggests that it is reasonable to apply the functional form used in the model across the study period. Put another way, if the model were a poor approximation of the world, then we would not expect coal use estimates based on the structure of the model to do a reasonable job of matching the observed data. Furthermore, the results in Online Appendix A.2.6 suggest that the patterns of change observed from 1907 to 1924 are also consistent with ( 2).
Using the first order conditions from the firm’s maximisation problem, and summing across all firms within an industry, I obtain the industry labour demand equation:
$$\begin L_ = \alpha _i^> \left(a_ p_\right)^> \left(\beta _i /\phi _\right)^> w_^> \bar_. \end$$Note that, in equilibrium, the sum of firm resource use must equal total city-industry resources, which are fixed at |$\bar
One congestion force in the model is the limited supply of housing. The housing market itself is not a central focus of this paper, so I model housing in a reduced-form way,
$$\begin \ln (r_) = \lambda \ln (L_) + \ln (\eta _c), \end$$where rct is the rental rate, Lct is total city population, ηc represents fixed city-specific factors that influence construction costs, and λ > 0 is a parameter that determines the impact of increasing population on the housing price. 29
Now, we turn to the supply of labour in a city. The model is populated by a continuum of homogeneous workers, each of which supply one unit of labor to the market. Workers consume a basket of goods with price Pt and housing. They also benefit from local amenities. The indirect utility function is,
where wct is the wage, Act is the amenity value, and the γ ∈ (0, 1) parameter determines the relative expenditure shares of housing and goods.
Workers are freely mobile across cities and have an outside option utility |$\ln (v_t^*)$| in each period. In the empirical setting that I consider, this can be thought of as either the utility of emigrating or the utility of living in the rural areas of the country. Given this, and using ( 4), the inverse labour supply equation for city c is,
$$\begin w_\textit = P_t^ <\gamma >\, L_\textit^ <(1-\gamma )\lambda >\, \eta _c^ <1-\gamma >\, A_\textit^ \, v_t^*. \end$$In addition to workers, the model is also populated by capitalists who receive the rent from land and local resources. For simplicity, I assume that capitalists live and spend their income outside the city.
Next, I want to incorporate the impact of local industrial pollution into the model. Coal pollution can impact the city by affecting both workers and firms. Focusing first on residents, I express the local amenity value as |$A_ = \delta _c \, C_^ <-\psi >\epsilon ^A_$| , where Cct is city coal use, δc represents a fixed city amenity, the ψ parameter determines the impact of local coal use on the amenity level, and |$\epsilon ^A_$| represents an idiosyncratic shock to the local amenity level.
Coal use can also affect the productivity of local firms. To build this channel into the model, I assume that local industry productivity can be separated into the impact of national changes in industry productivity, ait, the impact of city-level coal use on firm productivity, |$C_^$| , where the parameter ν ≥ 0 determines the impact of local coal use on firm productivity, and an idiosyncratic shock to city-industry productivity, |$\epsilon ^P_$| . Thus, I have |$a_ = a_ C_^ \epsilon ^P_$| .
Given the outside option utility, the national coal price, a set of national industry output prices, technology levels, and city industry resources, equilibrium in a city is defined as the set of local wages, resource prices, housing rent and population, and a set of industry employment and coal use levels, such that firms maximise profits, the local markets for resources clear, the housing market clears in each city, and city labour supply equals city labour demand.
For the empirical analysis, I need an expression that relates the growth in local industry employment to changes in local industrial pollution. The starting point for this derivation is the industry labour demand expression given in ( 3) and the city labour supply expression in ( 5). Differencing these expressions over time, taking logs, and substituting out the wage terms, I obtain,
$$\begin
Equation ( 6) forms the basis for the main empirical specifications used in this article. The Δln (Lct) and Δln (Cct) terms on the right-hand side of this equation capture, respectively, the impact of city congestion and of city coal use. The model suggests that both of these will negatively impact city-industry employment growth, though it is worth noting that the impact of city size may be positive if a city-size agglomeration force is included in the model. 30 In the middle row of ( 6) is a set of terms that vary only over time, but not across space. These will be absorbed by year effects in the empirical analysis. On the bottom row of ( 6), the first term reflects national industry-level demand or productivity shocks, the building blocks of the Bartik instrument. These can be absorbed by industry-time effects in the main analysis. The final two terms on the bottom row of ( 6) are the error terms. The structure of these terms makes it clear that I should allow for correlated errors across industries within the same location and time period in the empirical analysis.
The focus of the empirical analysis will be estimating the coefficient on the coal use and city-size terms in ( 6). As ( 6) shows, the impact of either coal use or congestion is determined by a combination of several model parameters. In the empirical analysis, I will estimate a single coefficient reflecting how, together, these parameters govern the relationship between either congestion or coal use and city growth, but I will not be able to identify the component parameters individually. For further discussion of this expression and its link to the coefficients estimated in the empirical analysis, see Online Appendix A.3.2. Online Appendix A.3.1 relates the estimation approach suggested by ( 6) to the larger Bartik instrumentation literature.
This section begins with an analysis of the impact of coal use on local employment growth, first at the level of city-industries and then at the city level. These are the central results of the article. Following that, I present a simple counterfactual that can help us think about the implications of coal use for overall urbanisation levels. Finally, I provide some tentative evidence on the channels through which coal use may have affected city growth.
$$\begin \Delta \ln (L_\textit) = b_0 + b_1 \Delta \ln (C_\textit) + b_2 \Delta \ln (L_\textit) + \xi _\textit + e_\textit, \end$$
where the ξit is a set of industry-time effects which absorb the national-level factors in ( 6) as well as the industry-specific productivity and demand shocks, while eict incorporates the idiosyncratic shocks to city amenities and city-industry productivity.
It is clear that a regression implementing ( 7) will suffer from serious identification issues. In particular, both the change in overall city employment and the change in city coal use will be endogenously affected by city-industry employment growth. To deal with this, I replace these terms with predicted values. For overall city employment, let,
where GRi − ct, t − τ is the growth rate of industry i in all cities other than c from t − τ to t. In this expression, τ determines the size of the time period over which differences are taken. 31 Thus, Δln (PrCityEMPct) represents the expected growth in employment in all other local industries, given national industry growth rates and the initial industrial composition of the city. Note that, when studying industry i, that industry is dropped when constructing Δln (PrCityEMPct). 32 This helps avoid endogeneity concerns, but ultimately it does not have a substantial impact on the results.
Next, to reflect the predicted change in city coal use, I define,
$$\begin \Delta \ln (\textit _\textit) = \ln \left(\sum _ L_ \times GR_ \times \theta _j\right) - \ln \left(\sum _ L_ \times \theta _j\right). \end$$where θj is coal use per worker in industry j. It is important to note that the difference between Δln (PredCoalct) and Δln (PrCityEMPct) is due only to variation in the coal intensity of industries, represented by θj. 33
Before introducing the regression specification, it is useful to use the variables introduced above to provide some preliminary evidence on the impact of changes in coal use on employment growth at the city level. Let,
$$\begin \textit = \Delta \ln (\textit _\textit)-\Delta \ln (\textit _\textit), \end$$where Δln (CityEmp) is the change in actual city employment from t − τ to t and Δln (PrCityEMPct) is defined above. Thus, DEVIATION can be interpreted as the difference between the actual change in log city employment in a particular period and the change that we would have expected the city to achieve given the city’s industrial structure at the beginning of the period and the industry growth rates in all other cities observed across that period. In other words, this reflects the extent to which employment growth in a city over or under-performs relative to what we would expect given national industry growth rates. In Figure 2 I plot this against the predicted change in coal use in the city over the same period (Δln (PredCoalct)) for each city over each two-decade period. What this figure shows us is that, in locations where we expect rising coal use, city employment growth is systematically underperforming what we would have expected given the city’s industrial composition at the beginning of each period and national industry growth rates.
Deviation versus Predicted Change in City Coal Use.
Notes: The y axis is the difference between actual city employment growth over each two-decade period in city c and the predicted employment growth in that city industry based on each city’s initial employment by industry and employment growth in each industry in all other cities, summed across industries. The x axis is the predicted change in city-level industrial coal use over the period, which is generated using the initial composition of city industries interacted with national industry growth rates and measures of industry coal use per worker. The trend line is based on a third-order polynomial.
While Figure 2 provides some preliminary evidence at the city level, the main analysis focuses on regressions at the city-industry level, consistent with the underlying theory. The main regression specification is,
$$\begin \Delta \ln (L_\textit) = b_0 + b_1 \Delta \ln (\textit _\textit) + b_2 \Delta \ln (\textit _\textit) + \xi _ + e_. \end$$This specification addresses the most important identification concerns in ( 7), i.e., the endogenous effect of city-industry employment growth on city-level congestion and coal use. Note that the inclusion of the Δln (PrCityEMPct) term in this expression is vital, because it picks up the direct effect of employment growth in other industries in city c on the employment growth of industry i, which may operate through channels such as congestion or agglomeration forces. This allows the b1 coefficient to pick up the additional impact that is generated when this employment growth occurs in more coal-intensive industries.
Identification in this estimation approach relies on assumptions that are standard in articles following Bartik ( 1991), particularly those that rely on variation in industry characteristics such as Diamond ( 2016). The main threat to identification in this approach is that there could be some other industry feature that is both correlated with industry coal use intensity and affects local employment growth. After presenting the main regression results, I present a variety of additional results including controls for the most likely channels through which the identification assumption might be violated. These additional checks allow me to strengthen identification beyond what is typical within the literature following Bartik ( 1991).
An alternative to the reduced-form approach represented by equation ( 8) is to use the predicted coal use to instrument for the actual change in coal use. In the main results I prefer the reduced-form approach because it is easier to work with and because the advantages of the IV approach are limited since the variable that one would ideally want to instrument for, the local pollution level, is unobserved. Nevertheless, I have also estimated IV regressions and these deliver similar results (see Online Appendix A.4.6).
The specification in ( 8) includes an assumption that the impact of coal use is linear in logs. There are two available pieces of evidence supporting this functional form. First, this functional form is consistent with the scatterplot shown in Figure 2. Second, Beach and Hanlon ( 2018) provides evidence that the impact of coal use on mortality is linear in logs. To the extent that the mortality rate is a good indicator of the impact of coal use this suggests that the specification used here is reasonable.
Note that ( 8) abstracts from heterogeneous industry responses to changing levels of city pollution or city congestion forces—a feature suggested by the theory. While I begin the analysis by abstracting from heterogeneity in the response to coal use across industries, later I will also present results that explore these heterogeneous responses.
In relation to the theory, the estimated b1 coefficient from ( 8) will reflect the impact of changes in local industrial coal use on city-industry employment growth, which will depend on how coal use affects the city amenity level, how coal use affects firm productivity, as well as the extent to which industries can respond to these effects by shifting employment away from polluted locations. 34 The theory suggests that this coefficient should be negative. Note that, because Δln (PrCityEMPct) is also included in the regression specification, the b1 coefficient should be interpreted as the impact of a rise in local industrial coal use holding constant the overall local employment level, i.e., as an increase in the pollution intensity of local industry. Similarly, the b2 coefficient should be interpreted as reflecting the impact of an increase in local employment holding fixed the level of local industrial coal use, i.e., a rise in completely clean employment. 35
This estimation approach abstracts from variation in industry coal use intensity across cities. This is driven in part by data constraints, since city-specific industry coal use intensities are not observed. However, even if city-level industry coal use intensity was observed, I would probably not want to incorporate this into the explanatory variable because, as suggested by the theory, this value will be endogenous and dependent on local wage levels. Abstracting from spatial variation in industry coal use intensity avoids this endogeneity concern.
Estimation is done using pooled cross-sections of data (after taking differences), an approach that allows me to exploit as much of the available data as possible. This is vital because the key variation in this study occurs at the city level and only 31 cities are observed in the data. We may be concerned about spatial and serial correlation in this setting. To deal with these potential issues, I allow correlated standard errors across industries within the same city, following Conley ( 1999) and across time within the same city-industry, as in Newey and West ( 1987). 36
I begin the analysis, in Table 1, by exploring results with differences taken over time periods ranging from one to three decades. The table includes results for all industries, in Columns 1–3, and for a set of manufacturing industries only, in Columns 4–6. I provide separate results for manufacturing industries only because these produce more tradable products and so are a better fit for the model, and also because some of the control variables that I will introduce later are available for only this set of industries.
Baseline City Industry Regression Results.