Tragedy Of The Commons
Why is Pollution from U.S. Manufacturing Declining? The Roles of Trade, Regulation, Productivity, and Preferences∗
Joseph S. Shapiro Yale University and NBER
joseph.shapiro@yale.edu
Save your time - order a paper!
Get your paper written from scratch within the tight deadline. Our service is a reliable solution to all your troubles. Place an order on any task and we will take care of it. You won’t have to worry about the quality and deadlines
Order Paper NowReed Walker University of California, Berkeley
IZA and NBER rwalker@berkeley.edu
September 2015
Abstract
Between 1990 and 2008, air pollution emissions from U.S. manufacturing fell by 60 percent despite a substantial increase in manufacturing output. We show that these emissions reductions are primarily driven by within-product changes in emissions intensity rather than changes in output or in the compo- sition of products produced. We then develop and estimate a quantitative model linking trade with the environment to better understand the economic forces driving these changes. Our estimates suggest that the implicit pollution tax that manufacturers face doubled between 1990 and 2008. These changes in environmental regulation, rather than changes in trade, productivity, or preferences, explain most of the emissions reductions. JEL: F18, F64, H23, Q56
∗We thank Joe Altonji, Costas Arkolakis, Randy Becker, Jared Carbone, Brian Copeland, Arnaud Costinot, Sharat Ganapati, Michael Greenstone, Sam Kortum, Arik Levinson, Bill Nordhaus, Noam Yuchtman, and numerous seminar participants for useful comments and discussions. We also thank Adrian Fernandez, Vivek Sampathkumar, David Silver, and Yusuf Mercan for excellent research assistance. Shapiro acknowledges support from the Yale MacMillan Center and the Yale Program on Applied Economics and Policy Research. Walker acknowledges additional support from the Robert Wood Johnson Foundation and the UC office of the President through the University of California Center for Energy and Environmental Economics. All results have been reviewed by the U.S. Census Bureau to ensure that no confidential information is disclosed.
1 Introduction
Between 1990 and 2008, emissions of the most common air pollutants from U.S. manufacturing fell by 60
percent, even as real U.S. manufacturing output grew substantially. Figure 1 shows just how stark these
environmental improvements have been. Between 1990 and 2000, the real value of U.S. manufacturing output
grew by a third even as manufacturing’s emissions of major regulated air pollutants like nitrogen oxides,
particulate matter, sulfur dioxide, and volatile organic compounds fell on average by 35 percent. After 2000,
growth in real manufacturing output slowed, even while manufacturing pollution emissions fell another 25
percentage points relative to 1990 levels.
Research suggests at least four possible explanations of these substantial improvements in U.S. air qual-
ity. First, U.S. manufacturing trade has grown substantially (Autor, Dorn, and Hanson, 2013; Pierce and
Schott, 2012). When dirty industries like steel or cement move abroad, total U.S. pollution emissions may
fall. Second, federal and state agencies require firms to install increasingly stringent pollution abatement
technologies. Some research directly attributes national changes in air quality to the Clean Air Act and to
other environmental regulations (Henderson, 1996; Chay and Greenstone, 2005; Correia, III, Dockery, Wang,
Ezzati, and Dominici, 2013). The third explanation for these emissions changes stems from non-homothetic
tastes or secular change; Americans may gradually choose to spend less on heavy manufactured goods and
more on services and cleaner goods (Levinson and O’Brien, 2013). Finally, if manufacturers use fewer dirty
inputs each year to produce the same outputs, then annual productivity growth could improve air quality. In
support of this final explanation, Figure A1 shows a clear negative relationship between plant-level pollution
per unit of output and total factor productivity in the U.S. manufacturing sector; as total factor productivity
rises, pollution per unit of output falls.1
The goal of this paper is to better understand the underlying forces that have caused changes in pollu-
tion emissions from U.S. manufacturing. We do this in two complementary ways. We begin by decomposing
changes in total manufacturing emissions into changes that can be explained by the total scale of manufac-
turing output, the composition of products produced, and the pollution intensity of a given set of products.
This data driven exercise exploits newly available, administrative data on product-level emissions intensities
that affords additional granularity relative to the existing literature. While the statistical decomposition
delivers clear conclusions, it lacks the ability to uncover the primitive economic forces driving emissions
reductions.
We complement this decomposition with a quantitative model of pollution emissions in the U.S. manu-
facturing industry. The model consists of firms endogenously choosing investments in pollution abatement
to avoid a per-unit tax on pollution emissions. Production and pollution abatement choices depend on
trade costs, pollution taxes, productivity, and consumer preferences. The model weaves together elements of
workhorse models from the international (Melitz, 2003) and environmental literatures,(Copeland and Taylor,
2003). While the model is highly stylized, it has several attractive features — it has explicit and simple
theoretical microfoundations from which all the analysis is derived; it accounts for imperfect competition
and can accommodate various market structures; it incorporates firm entry, exit, and cross-firm realloca-
1See Appendix C.1 for further details on Figure A1.
1
tion; it requires few parameters that can be estimated using reduced-form regressions; it can analyze a wide
variety of counterfactuals; and it can account for many general equilibrium forces in settings where partial
equilibrium, program evaluation methods may be able to deliver less insight.
We then combine the model with administrative plant-level data from the U.S. Census Bureau and EPA
with two main objectives. First, we use intermediate results from the model, combined with actual pollution
abatement and emissions decisions, to back out the implicit tax per unit of pollution emissions that firms
face. U.S. federal, state, and local environmental regulations take many overlapping forms: command-and-
control technology standards, cap-and-trade programs, and many others.2 Our quantitative exercise yields a
closed form expression for the overall regulatory burden from these disparate and overlapping environmental
policies, allowing us to observe how this measure has changed over time. The second main objective of the
quantitative exercise is to evaluate a range of counterfactuals, such as how pollution emissions would have
evolved if air pollution regulation had remained unchanged after 1990. Many researchers in environmental
economics and international trade use quantitative models to forecast the future — they study untested
policies such as a global 10 percent decrease in all trade barriers or a national carbon tax. Unlike such
work, this paper uses a model to interpret the past — we quantify how different kinds of economic shocks
(e.g., trade costs, environmental regulation, productivity, and preferences) led to observed changes in actual
pollution emissions. Similar general equilibrium decompositions have been used to understand the causes of
the collapse of trade around the Great Recession and the changes in between-group wage inequality (Eaton,
Kortum, Neiman, and Romalis, 2011; Burstein, Morales, and Vogel, 2013).
Our results suggest that changes in the scale of manufacturing output or changes to the composition of
products produced cannot explain trends in pollution emissions from U.S. manufacturing between 1990 and
2008. Instead, decreases in pollution per unit of output within narrowly defined product categories explain
almost all of the changes in emissions over this time. We then show that the model-driven measure of the
pollution tax rate that rationalizes observed pollution emissions and abatement decisions – a scalar measure
of the stringency of environmental regulation – more than doubled between 1990 and 2008. We find similar
patterns in regulation across all the main pollutants the Clean Air Act regulates (“criteria pollutants”), but
we find no changes in unregulated pollutants, such as CO2 over this time period. Lastly, we find that this
increasing stringency of environmental regulation explains 75 percent or more of the 1990-to-2008 decrease
in pollution emissions from U.S. manufacturing. Despite the plant-level relationship between pollution and
productivity documented in Figure A1, and similar relationships found in related literature (Bloom, Genakos,
Martin, and Sadun, 2010; Holladay, 2011; Martin, 2011), we find that changes in U.S. productivity have had
small effects on U.S. pollution emissions at the economy-wide level.
This paper departs from the literature in four primary ways. First, we provide new evidence on why
pollution from U.S. manufacturing is declining. Some research relates national changes in pollution emissions
to three channels: changes in the aggregate level of manufacturing output, changes in the composition of
2Berman and Bui (2001a,b) describe the entire menu of local air quality regulations facing manufacturing firms around Los Angeles, finding 11 local air quality regulations for petroleum refining and 46 for manufacturing (a count which excludes state and federal regulations). Most of the manufacturing policies apply to only a few industries each. The analysis includes the years 1979 to 1993. Los Angeles has among the most stringent air quality regulations in the country. We thank Eli Berman and Linda Bui for sharing details of these regulations.
2
output across manufacturing industries, and changes in the pollution emitted per unit output within an
industry (Levinson, 2009). Research describes these channels as scale, composition, and technique. The
methodology behind our statistical decomposition resembles this work, but detailed administrative data
allow us to extend previous analyses to look within physical products and not merely within industries.
This added granularity helps to address previous concerns regarding the inability of industry level data to
distinguish between changes in the within-industry reallocation of production towards cleaner products and
industry-level reductions in emissions intensity (Koo, 1974; Gamper-Rabindran, 2006; Ederington, Levinson,
and Minier, 2008; Levinson, 2009). More importantly, the conclusion of previous research that pollution per
unit output within industries is falling (i.e., pollution is declining due to the ”technique” effect) is silent on
deeper economic causes. Pollution per unit output is an endogenous outcome of the global economy and
can be explained by numerous possible forces. We use a quantitative model to relate changes in pollution
to policy-relevant choices like trade costs and environmental regulation. Our analysis of this quantitative
model suggests that environmental regulation, rather than trade, productivity, or other forces, explains most
of the decline in pollution.
A second contribution of this study is to quantify the change in the overall regulatory burden, or shadow
price of pollution, that manufacturing firms face due to local and national air pollution regulations. We find
that this price more than doubled between 1990 and 2008 for all air pollutants we study except CO2, which
had a nearly-constant shadow price. While analyzing the overall regulatory burden affecting firms does not
prescribe an single law or policy lever, it does explain what all the hundreds of regulations have added up
to, which is a question of central importance.
Third, this paper estimates, for the first time, a parameter that has played a central role in environmental
economics models for at least 30 years – the elasticity governing a firm’s tradeoff between production and
pollution abatement (Siebert, Eichberger, Gronych, and Pethig, 1980; Copeland and Taylor, 2003). This
parameter has equivalent interpretations as the Cobb-Douglas cost share of pollution taxes in production
or, alternatively, as the elasticity of pollution emissions with respect to productivity.
Lastly, this paper develops a flexible and tractable approach to analyzing economy-wide changes in
pollution. Research studying changes in pollution typically uses quasi-experimental regressions, industrial
organization models, or macro-trade models (Deschenes, Greenstone, and Shapiro, 2013; Isen, Rossin-Slater,
and Walker, 2014; Fabra and Reguant, 2014; Copeland and Taylor, 1994).3 Quasi-experimental studies can
isolate the effect of individual policies one-at-a-time, but regulators have implemented dozens of overlapping
pollution regulations over the last 20 years, many of which have not been analyzed with policy evaluation
tools and have no natural comparison group. Industrial organization models have rich industry-specific detail
but typically do not study an entire industrial sector or account for general equilibrium forces. Theoretical,
macro-trade models have provided considerable insight but have generally resisted estimation. Our method-
ology builds on tools from a recent trade literature sometimes described as “structural gravity” (Costinot
and Rodriguez-Clare, 2014), which has become standard in the trade literature, though its application to
environment questions is nascent (Shapiro, 2013). We study the specific counterfactual of explaining his-
3Some research describes a model of the environment and trade and then estimates linear regressions where the explanatory variables proxy for important variables in the theory (Antweiler, Copeland, and Taylor, 2001). We take a literal interpretation of the model by estimating its primitive parameters and then solving for equilibrium outcomes given a vector of inputs.
3
toric changes in pollution. However, our approach is versatile enough to evaluate prospective environmental
policies or design optimal environmental policy. One of our goals is to make clear how similar approaches
could be used to study a range of environmental questions.
The rest of the paper proceeds as follows. Section 2 presents a statistical decomposition in order to
break down aggregate emissions trends in our data, while also highlighting the frontier of what we are able
to say with the data alone. Section 3 outlines our trade-environment model. Section 4 discusses the data,
and Section 5 discusses how we estimate the parameters. Section 6 presents the main results, and Section 7
discusses alternative explanations and additional robustness concerns. Section 8 concludes.
2 A Statistical Decomposition of U.S. Emissions 1990-2008
Much economic research interprets national changes in industrial air pollution via three pathways (Copeland
and Taylor, 1994; Grossman and Krueger, 1995). One is a change in the scale of real output. The second
is a change in the composition of production from relatively clean products like “household furniture” to
relatively dirty products like “carbon black.” The third is a change in the production technique used to
produce a single product, which could decrease a product’s pollution emissions per unit of output.
We begin by presenting a statistical decomposition of manufacturing pollution emissions using newly
developed administrative data on manufacturing plant-product-year production from 1990 to 2008. The
Census of Manufacturers and the Annual Survey of Manufacturers collect sub-industry, product-level output
data, at the plant-product-year level. We use this information to illustrate whether changes in the total scale
of output and/or changes in the composition of products produced is able to explain the observed reductions
in ambient air emissions. Our focus on products rather than industries is unique to the literature and is
meant to capture the fact that even within a fairly narrow industry code (e.g., 4-digit Standard Industrial
Classification (SIC) code), many products differ significantly in their emissions intensities. Previous research
has explored trends in manufacturing pollution emissions using industry level data. The previous literature
has acknowledged that a limitation of industry level production data is the inability to distinguish changes in
the reallocation of production towards cleaner products from industry-level “technique” based reductions in
emissions intensity (Koo, 1974; Gamper-Rabindran, 2006; Ederington, Levinson, and Minier, 2008; Levinson,
2009). For example, while the U.S. manufacturing sector contains 455 4-digit SIC codes, the product trailer
from the Census and Annual Survey of Manufacturers allows us to perform this decomposition using 1,440
products. This granularity allows us to quantify by how much the scale of output versus the types of products
produced can explain the observed reductions in manufacturing air emissions.
Consider the following representation of total manufacturing pollution, denoted Z:
Z = ∑ s
zs = ∑ s
xses = X ∑ s
κses (1)
Total manufacturing pollution Z equals the sum of pollution from each manufacturing product s, zs. A
manufacturing product in our setting can be thought of as a sub-industry classification, where for example,
SIC 3312 (blast furnaces and steel mills) is subdivided into 24 different products ranging from steel wire
4
(33125) to cold rolled sheets and strip (excluding metallic coated and electrical) (33127).4 Alternatively,
we can write manufacturing pollution as equal to the total output of a product xs multiplied by a product-
specific emissions factor es. We can also represent manufacturing pollution emissions as the total output
shipped by all manufacturing industries, X, multiplied by the sum of each product’s share of total output,
κs ≡ xs/X, times an emissions coefficient reflecting pollution per dollar of output shipped of that product (es ≡ zs/xs). In vector notation, we have
Z = Xκ′e
where κ and e are S × 1 vectors containing the market shares of each of the S products and their pollution intensities, respectively. Totally differentiating then dividing through by Z yields three terms representing
the scale, composition, and technique effects:
dZ
Z =
dX
X︸︷︷︸ scale
+ dκ
κ︸︷︷︸ composition
+ de
e︸︷︷︸ technique
(2)
Taking the decomposition in equation (2) to the data requires annual data on total pollution, total output,
each product’s contribution to output, and each product’s emissions intensity. Pollution and total output
come from the EPA’s National Emissions Inventory (NEI) and the NBER-CES Manufacturing Industry
Database, respectively. We construct product-level output shares in each year using the product trailer from
the Census and Annual Survey of Manufacturers. In order to construct product-level emissions factors we
match the National Emissions Inventory to the Annual Survey of Manufacturers in 1990 via name and address
string matching. Appendix C.2 describes the string matching process in more detail. Since the NEI reports
emissions at the plant-level and the Census product trailer reports output at the product level, we apportion
plant-level emissions to a plant-product-year using product shares produced at the plant in 1990.5,6 We take
the total emissions attributable to each product in 1990 and divide by the total product shipments in 1990
to construct emissions intensities.7 We then use these 1990 product-level emissions intensities to project the
scale and composition effects forward in time, holding technology (i.e. our emissions intensities) constant at
1990 emissions rates. The decomposition allows us to observe what emissions would have looked like in 2005
if firms still produced products with 1990 emissions intensities. Appendix C.3 provides additional details
about both the underlying data and the decomposition.
Figure 2 illustrates the resulting statistical decomposition for nitrogen oxide emissions (NOx). Appendix
Figure A2 presents results from other pollutants, which show similar patterns. Appendix Table shows
the numbers corresponding to these graphs. The top solid line in Figure 2 depicts the total real value of
manufacturing shipments, where each industry’s output is deflated by the NBER-CES industry specific price
4Output at the five-digit SIC level is the most disaggregate data available for all plants in the Census and Annual Survey of Manufacturers.
5We have also created emissions intensities by simply dividing total plant emissions by product shipments for each product produced at a plant. Conclusions are not sensitive to either method of constructing product-level emissions intensities.
6Previous research has used the World Bank’s Industrial Pollution Projection System (IPPS) for emissions intensities. The IPPS data provides a list of emissions intensities by four-digit Standard Industrial Classification (SIC) codes (Hettige, Martin, Singh, Wheeler, and Mundial, 1995; Levinson, 2009). Levinson (2014) constructs industry-level emissions intensities using the NBER-CES productivity database combined with raw NEI data.
7We deflated total product output by industry-year specific price indices, from the NBER-CES database, where 2008=1.
5
index and then totaled. We scale total output so it equals 100 in 1990. This line summarizes what emissions
would have been if we had kept the same emissions rates and the same product composition as in 1990. The
middle dashed line plots NOx emissions that would have occurred if emissions intensities had remained fixed
at 1990 levels but the composition of output across manufacturing products had equaled observed, historical
values. The bottom dotted line plots actual NOx emissions from manufacturing, as reported by the NEI.
The bottom line implicitly summarizes the joint result of changing the scale, composition, and technique of
manufacturing production over this time period.
The statistical decomposition leads to several conclusions. First, the dotted line shows that actual NOx
emissions fell by almost 50 percent. Second, the proximity of the solid and dashed lines shows that the
composition between clean and dirty manufacturing products has not changed much over time, although
between 2000 and 2007, manufacturing shifted to slightly cleaner products. Third, the solid and dashed
lines each show that if the pollution intensity of industries had not changed, NOx emissions would have risen
by about 30 percent. Finally, the gap between the solid line on top and dotted line at bottom shows that
changes in the pollution intensity of individual products (i.e. “technique) explains why NOx emissions fell
by 50 percent rather than rising by 30 percent.
This relatively clear conclusion, that most reductions in emissions are driven by within-product changes
in emissions intensity, echoes previous findings in the literature, albeit with more granular data. The data,
however, are relatively silent on what might be causing these changes. The rest of the paper investigates the
underlying economic forces driving these patterns in the data. If more productive plants emit less pollution
per unit output, then product-level productivity growth could explain these patterns. Alternatively, changes
in trade costs like the introduction of NAFTA or China’s WTO ascension may have caused a reallocation of
production away from unproductive and dirty firms toward more productive and perhaps cleaner firms that
produce the same product. Lastly, increases in environmental regulatory stringency may also explain these
reductions.
The quantitative model, which fills the remainder of this paper, makes different and arguably stronger
assumptions than this statistical decomposition. The advantage of these stronger assumptions is an ability to
explain how trade, regulation, productivity, and preferences contribute to the environmental improvements
documented in Figures 1 and 2. The disadvantage is that these assumptions only roughly approximate
reality. We discuss ways in which these assumptions can be relaxed in future research. The reader interested
in additional detail on specific assumptions of this model is referred to Costinot and Rodriguez-Clare (2014),
Melitz and Redding (2014), and to Copeland and Taylor (2003).
3 Model of Heterogeneous Firms with Endogenous Pollution Abatement
We describe a model of firm entry, production, trade, and pollution abatement, which is designed to reflect
a stylized description of polluting industries. In the model, firms differ in their productivity level, which
leads these firms to differ in their pollution abatement investments and ultimately pollution emissions. The
model accounts for endogenous changes in firm entry, exit, production, and export decisions in a tractable
way that yields analytical solutions and allow us to analyze counterfactuals. The model fits a variety of
6
key stylized facts—dirty industries invest more in pollution abatement than clean industries; productive
plants emit less pollution per unit output than unproductive plants; productive plants produce more than
unproductive plants; and only the most productive plants export. Like all models, this approach seeks
to reflect systematic patterns across firms while recognizing that some strict assumptions which enhance
tractability, like monopolistic competition and constant elasticity of substitution (CES) utility, are not
literally accurate descriptions of firms and consumers.
The model has a straightforward economic environment. We analyze a world of two countries (US and
Foreign), each with a representative agent. Each country has one productive factor, which we call labor,
and which is inelastically supplied. We present the model’s main results here and intermediate derivations
in Appendix B. This section explains the model’s four assumptions, shows how we use the model to analyze
counterfactuals, then discusses how we obtain empirical counterparts to the economic objects in the model.
3.1 Model Assumptions
A1. Preferences. The representative agent in destination country d has the following utility function:
Ud = ∏ s
[∑
o
∫ ω∈Ωo,s
qod,s (ω) σs−1 σs dω
] σs σs−1
βd,s Z−δd (3)
Equation (3) describes CES utility across product varieties within a sector, Cobb-Douglas preferences across
sectors, and multiplicative damages from pollution. The representative agent allocates expenditure across
varieties of goods ω from the measure Ωo,s of goods produced by industry s in origin county o. The parameter
βd,s represents the share of country d’s expenditure devoted to industry s, where ∑
s βd,s = 1. We refer to
βd,s as consumer preferences. The variable qod,s (ω) represents the quantity of variety ω goods in industry s
which are exported from origin country o to destination country d. The industry-specific parameter σs > 1
represents the elasticity of substitution across varieties. The parameter δ governs the disutility of pollution
Zd. We assume that pollution is a pure externality, so the representative agent ignores the term Z −δ d in
making expenditure choices.8
The assumption of CES utility, which is common in trade and macroeconomic research, implies that
consumers experience decreasing marginal utility from consuming a given variety and increasing utility in
the total measure of varieties. We assume this utility function because it provides a simple way to account
for different varieties within a sector while leading to parsimonious aggregate descriptions of production and
trade flows across countries and industries.9
A2. Firms and Market Structure.
8Because the empirical analysis of the paper does not use information on the marginal damages of air pollution, the choice of functional form for pollution damages does not affect the empirical section of the paper. The empirical results would be numerically equivalent if pollution damages were assumed to be multiplicative, additive, or any other form.
9Research finds that non-CES utility functions, such as the linear demand system, translog utility, and certain generalizations which can allow for endogenous markups, are all part of the same “gravity” family of models. While this implies that their measures of the gains from trade are closely related, these structures do not always obtain the kind of tractable closed-form relationships we use here (Melitz and Ottaviano, 2008; Feenstra and Weinstein, 2010; Arkolakis, Costinot, Donaldson, and Rodriguez-Clare, 2012).
7
A competitive fringe of entrepreneurs may choose to pay the sunk entry cost feo,s to draw a productivity
ϕ. The productivity is drawn from a Pareto distribution with shape parameter θs and location parameter
bo,s. We make the standard assumption that θs > σs − 1 so the entrants have finite expected profits.10
After observing the productivity draw, an entrepreneur who decides to produce must pay a separate fixed
cost. Firms engage in monopolistic competition so that, conditional on choosing to operate, an entrepreneur
chooses prices pod,s and abatement investments ξ to maximize profits:
πo,s (ϕ) = ∑ d
πod,s (ϕ) −wofeo,s, (4)
where πod,s (ϕ) = pod,s (ϕ) qod,s (ϕ) −wolod,s (ϕ) τod,s − to,szod,s (ϕ) τod,s −wdfod,s
The firm sells qod,s = (1 −ξ)lod,sϕ units. The profit function involves several terms. For simplicity, we drop the variety notation ω and index a firm by its productivity ϕ. A consumer in destination d pays price pod,s (ϕ)
for goods from firm ϕ. Each firm receives revenue pod,s (ϕ) qod,s (ϕ) and requires lod,s (ϕ) units of productive
labor at wage wo to produce goods for sending to destination d. A fraction of this labor 1 − ξ is used to produce output and the remaining fraction ξ to abating pollution. Each firm pays the pollution tax to,s per
ton of pollution emitted on zod,s (ϕ) tons of pollution emitted for producing goods shipped to destination
d. Firms face iceberg trade costs, so τod,s ≥ 1 units must be shipped for one unit to arrive (hence, the firm must produce τod,sqod,s in order to sell qod,s). A firm that chooses to enter the destination market d must pay
the fixed cost fod,s. 11 Domestic trade costs are normalized so τoo,s = foo,s = 1. We assume that pollution
tax revenues are lost to rent-seeking. While this is assumption is made for simplicity, the considerable sums
that corporations spend to lobby politicians on energy and environmental legislations gives one basis for this
assumption. Here bo,s describes a country’s productivity while θs describes the dispersion of productivity
draws within an industry s.
We assume this market structure for several reasons. Many dirty industries like cement and steel are
concentrated and have barriers to entry. By accounting for fixed entry costs and industry-specific markups,
our assumptions reflect a stylized version of polluting industries. At the same time, this formulation lets us
account for firm entry and exit and for reallocation of productive factors and output across firms. Finally, the
Pareto technology distribution has plausible theoretical microfoundations (Gabaix, 1999; Luttmer, 2007) and
provides a good fit to the empirical firm distribution, at least in the upper tail (Axtell, 2001; Eaton, Kortum,
and Kramarz, 2011). 12 While many model-based analyses do not need or use closed-form descriptions of
equilibrium, closed-forms are more important in our setting in order to “invert out” the changes in technology,
regulation, productivity, and preferences that will constrain the model to exactly match the data (allowing
10In this setting, the strict criterion for entrants to have finite expected profits is actually slightly weaker: θs > (σs−1)(1−α), where α is described below but remains between 0 and 1.
11The assumption that this fixed cost is paid in destination- and not origin-country labor simplifies counterfactual analysis. Related work makes the same assumption that fixed market entry costs are paid in destination labor (Hsieh and Ossa, 2011; Eaton, Kortum, and Kramarz, 2011). The equivalence between welfare calculations in some perfect and monopolistic competition models in Arkolakis, Costinot, and Rodriguez-Clare (2012) requires fixed exporting costs to be paid in destination country labor.
12 Most of the literature following Melitz (2003) and using parametric distributions assumes that technology has a Pareto distribution. A few studies explore other productivity distributions, including the lognormal Head, Mayer, and Thoenig (2013) and bounded Pareto Feenstra (2014). Some research suggests that the Pareto distribution provides a more accurate fit to the distribution of U.S. firms than the lognormal does (Axtell, 2001).
8
us to decompose historical trends).
An important implication of monopolistic competition is that markups are fixed by the elasticity of
substitution and do not change in response to market condition. This approach follows a long line of
macroeconomics and trade papers, though rules out potentially interesting analysis of how market power
affects pollution, which is a question we leave for future work.
A3. Pollution. Firms produce pollution emissions with the following technology:
zod,s = (1 − ξ) 1/αs ϕlod,s (5)
We assume pollution regulations are stringent enough that all firms engage in some abatement. Equation (5)
states that pollution is an increasing function of output and a decreasing function of pollution abatement
expenditures. In addition, this formulation also implies that more productive firms emit less pollution per
unit output produced. Equation (5) is essentially the pollution production technology adopted in Copeland
and Taylor (2003). Modeling emissions in this way is appealing because many sensible and seemingly different
ways of modeling pollution turn out to be equivalent to equation (5). As we show later, α represents the
elasticity of pollution emissions intensity with respect to pollution abatement intensity. Pollution emissions
intensity is measured as units of pollution emitted per unit of output, and pollution abatement intensity is
measured as abatement expenditures divided by total factor costs. We will also show that pollution emissions
in this model can be described as another factor of production in a Cobb-Douglas production technology,
and in this interpretation α is the Cobb-Douglas share for pollution emissions. Copeland and Taylor (2003)
discuss other equivalent interpretations of this model of abatement.
It may be useful to explain conceptually how firm and consumer decisions determine pollution and
abatement in this model. Equation (5) shows that for an operating firm, pollution emissions decline when
the firm reallocates productive factors to abatement investments. Those abatement investments affect profits
because they decrease required pollution tax payments but increase abatement costs. The endogenous
decisions in the model like firm entry, exit, production, and trade all respond to environmental regulations,
and all of these forces interact to determine pollution emissions. A firm that is not operating produces no
pollution; a firm that begins exporting increases its output and also its pollution emissions. So although
equation (5) only literally describes pollution abatement as a decision for operating firms, the model more
broadly accounts for a variety of ways in which firm and consumer behavior affects pollution emissions.
The pollution technology assumption implies constant returns to scale in pollution abatement. A model
with increasing returns to scale in abatement would have different structure (see e.g., Forslid, Okubo, and
Ultveit-Moe (2011)). We considered the implications of such a model but chose not to pursue it for two
reasons. First, the importance of fixed costs for abatement technologies is empirically unknown. Scale
economies could be positive for capital investments like scrubbers, zero for fuel-switching like low-sulfur
coal, and negative due to principal-agent issues for management innovations. Second, prices in such a model
depend directly on market size, and market size appears in the equilibrium conditions in ways that prevent
the expression of these conditions in terms of purely exogenous variables and shocks, and hence prevent us
from backing out shocks and undertaking the decomposition this paper reports.13
13An alternative approach to the paper would be to construct a synthetic dataset of firms obeying the model’s assumptions
9
A4. Competitive Equilibrium. Consumers maximize utility; firms maximize profits; and in each country,
labor supply equals labor demand:
Lo = L e o + L
m o + L
p o (6)
A country’s labor supply Lo is allocated to three activities: paying the fixed entry cost f e o,s for drawing
a productivity (Leo); paying country-specific entry costs fod,s, which can be interpreted as marketing costs
(Lmo ); and engaging in production, including pollution abatement (L p o).
Utility maximization implies that the share λod,s of country d’s expenditure on goods from industry s in
country o has the following “gravity” structure:14
λod,s = Meo,s
( wo bo,s
)−θs (τod,s)
− θs 1−αs (fod,s)
1− θs (σs−1)(1−αs) (to,s)
− αsθs 1−αs∑
i M e i,s
( wi bi,s
)−θs (τid,s)
− θs 1−αs (fid,s)
1− θs (1−αs)(σs−1) (ti,s)
− αsθs 1−αs
(7)
where Mo,s represents the mass of entrepreneurs who attempt entry into production in country o and sector
s. Ignoring some terms which cancel, the numerator of equation (7) represents the value of imports and
the denominator represents total expenditure, where the denominator is the sum over countries i of total
imports from each country.
The gravity equation (7) is an intermediate result of the model, but it summarizes nicely how environ-
mental regulation affects economic activity. It shows that a country with a high pollution tax to,s will have
lower exports and less production, particularly in dirty industries with high values of the pollution parame-
ter αs. An empirical literature measures how environmental regulation affects plant openings, employment,
shipments, and other outcomes (see discussion of literature in Walker (2013)). A voluminous trade literature
derives gravity equations expressing trade flows in a way that approximate actual bilateral trade patterns
(see discussion of literature in Anderson and van Wincoop (2003)). Equation (7) provides an appealing
link between these two literatures. In addition, equation (7) allows empirical tests of the hypothesis that
industry reallocates production to countries with weaker environmental regulation (known as the “pollution
havens hypothesis”; see Levinson and Taylor (2008)).
3.2 Comparative Statics
One key motivation for using a model of heterogeneous firms is the finding that more productive plants
emit less pollution per unit of output. This relationship is not immediately apparent from the pollution
technology assumption in equation (5). However, the first-order conditions of the firm’s problem show that
more productive firms invest a greater share of productive factors in pollution abatement (see equation (19)
and matching moments of the data we observe, then adding trade or regulation shocks to those firms and observing their change in emissions. We found the decomposition pursued in this paper with actual data more direct since it lets us write down and describe equations for equilibrium, shocks, and counterfactuals rather than reporting these numerically; however, it comes at the cost of needing analytical expressions.
14The “gravity” description reflects the fact that bilateral trade in this and many other models is proportional to two countries’ incomes (in an analogy to gravity in physics, their mass) and inversely proportional to their trade cost (analogously to physics, their distance). Most trade models with constant elasticity of substitution preferences and iceberg trade costs produce a gravity equation. This relationship also has empirical support—Leamer and Levinsohn (1995, p. 1384), for example, describe empirical estimates of gravity equations as “some of the clearest and most robust empirical findings in economics.”
10
in Appendix B). Substituting that first-order condition into the firm’s choice of pollution emissions gives the
following firm-level relationship:
zod,s(ϕ)
qod,s(ϕ) =
1
ϕ1−αs
( wo to,s
αs 1 −αs
)1−αs (8)
Because αs ∈ (0, 1) can be interpreted as the cost share of pollution taxes, this expression implies that more productive firms (greater ϕ) emit less pollution zod,s per unit of output qod,s. Because more productive firms
are larger, this also implies that larger firms emit relatively less pollution per unit of output. It also implies
that a regression of log(zod,s/qod,s) on log(ϕ) will give a coefficient of αs − 1. Since α is likely to be a small number, it is notable that the slopes of the trendlines depicted in Figure A1 are around minus one.
3.3 Equilibrium in Changes
By combining assumptions A1 through A4 in specific ways, we can use this model to analyze counter-
factuals. We first combine the model’s assumptions into two equilibrium conditions that summarize firm
behavior. One equilibrium condition is the labor market clearing assumption (6). This condition says that
any counterfactual must have total labor demanded equal total labor supplied in each country.
The other equilibrium condition combines two facts. First, the expected profit from operating a firm must
equal the fixed cost of forming a firm. Second, a cutoff productivity for each destination market determines
whether the expected profit from entering that market exceeds the fixed marketing cost required to sell
there (i.e., a free entry condition and zero cutoff profit). Equation (21) of Appendix B formally describes
the second equilibrium condition. These are equilibrium conditions because if a set of data satisfies these
conditions, then those data represent a competitive equilibrium in the sense of Assumption A4. We use
these conditions to analyze how counterfactuals affect welfare.
The two equilibrium conditions include numerous variables that are difficult to measure. Rather than
attempt to measure these variables, we rewrite each variable as a proportional change from a base year, as in
Dekle, Eaton, and Kortum (2008). The benefit of writing the model in changes is that we do not need data
on difficult-to-observe variables because many of them do not appear in changes. For example, constructing
an empirical analogue to the model’s equilibrium conditions in levels (equation (6) and equation (21)) would
require data on productivity levels for each country and industry, the fixed costs of entering each market,
and the measure of firms operating in each country and industry. But in changes these variables will not
appear in the equilibrium conditions, and so we do not need to measure them.
Our formal use of this technique to analyze the model proceeds as follows. Let x denote some variable
from the model, let x′ denote the value under a counterfactual scenario, and let x̂ ≡ x′/x denote the proportional change in x due to the counterfactual. Written in changes, the two equilibrium equations (6)
and (21) become the following:
1 = ψo
(∑ s M̂
e o,sRo,s
(σs−1)(θs−αs+1) σsθs
+ η′o∑ s Ro,s
(σs−1)(θs−αs+1) σsθs
+ ηo
) (9)
11
ŵo = ∑ d
ζod,s
( ŵo b̂o,s
)−θs (τ̂od,s)
− θs 1−αs
( f̂od,s
)1− θs (σs−1)(1−αs) (
t̂o,s )− αsθs
1−αs
∑ i λid,sM̂
e i,s
( ŵo b̂o,s
)−θs (τ̂od,s)
− θs 1−αs
( f̂od,s
)1− θs (σs−1)(1−αs) (
t̂o,s )− αsθs
1−αs
β̂d,sŵd (10)


