Journal of International Economics 105 (2017) 130–149

Contents lists available at ScienceDirect

Journal of International Economics

journal homepage: www.elsevier.com/locate/jie

Trade liberalization and the environment: Evidence from NAFTA and

U.S. manufacturing夽
Jevan Cherniwchan
Department of Marketing, Business Economics & Law, Alberta School of Business, University of Alberta, 3-23 Business Building, Edmonton, Alberta T6G 2R6, Canada

A R T I C L E I N F O A B S T R A C T

Article history: The unobserved responses of individual polluters are often used to rationalize the aggregate effects of inter-
Received 31 May 2015 national trade on the environment. In this paper, I provide the first evidence of these responses. I estimate
Received in revised form 15 November 2016 the effects of NAFTA on the emissions of particulate matter (PM10 ) and sulfur dioxide (SO2 ) from manufac-
Accepted 3 January 2017 turing plants in the United States. My findings suggest that trade liberalization led to significant reductions
Available online 11 January 2017
of these pollutants at affected plants. On average, nearly two-thirds of the reductions in PM10 and SO2 emis-
sions from the U.S. manufacturing sector between 1994 and 1998 can be attributed to trade liberalization
JEL classification: following NAFTA.
F18
© 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
F14
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Q56

Keywords:
NAFTA
Trade liberalization
Pollution emissions

1. Introduction primarily relied on cross-country variation in pollution levels and

Over the past twenty-five years, one of the most widely debated
aspects of globalization has been the environmental consequences of
trade liberalization.1 While this debate has typically been framed in
terms of industry responses to trade liberalization, these responses
hinge on the behavior of individual polluters within each industry.
Yet, surprisingly little is known about how trade liberalization affects
the pollution emitted by individual plants.
This is largely due to the lack of disaggregate, plant-level data
on emissions and other plant characteristics.2 To date, research has
夽 I would like to thank two anonymous referees, M. Scott Taylor, Eugene Beaulieu,
Arvind Magesan, Branko Bošković, Joule Bergerson, Karen Chang, Brian Copeland,
Cecilia Garcia, Scott Holladay, Andrew Leach, Runjuan Liu, Carol McAusland, Juan
Moreno-Cruz, Irving Rosales, Barry Scholnick, Trevor Tombe and various seminar par-
ticipants for their comments and suggestions. I would also like to thank John Romalis
for sharing his tariff data. Funding for this research was provided by SSHRC and the
Centre for Applied Business Research in Energy and the Environment at the University
of Alberta. The usual disclaimer applies.
E-mail address: jevan@ualberta.ca.
1
For overviews of the literature on trade and the environment, see Copeland and
Taylor (2004) and Cherniwchan et al. (2017).
2
While plant-level emissions data from emissions inventories such as the Toxic
Release Inventory in the United States and the National Pollutant Release Inventory
in Canada have been publicly available for close to 20 years, it is only recently that
these data have started to be matched to data on plant characteristics for plants from
multiple industries (e.g. Holladay (2016)).
trade flows to examine the link between international trade and
the environment.3 Consequently, existing studies have focused on
the relationship between trade and aggregate pollution levels; these
studies find that trade is not necessarily bad for the environment.4
Even so, the literature often points to the unobserved responses of
individual polluters, such as the adoption of new technologies or
the exit of dirty plants, to explain the mechanisms underlying this
finding.
In this paper, I provide the first empirical evidence of how trade
liberalization affects the pollution emitted by individual manufac-
turing plants. To do so, I rely on two unique longitudinal datasets
constructed from two main sources: the Toxic Release Inventory
and the National Establishment Time Series. Each dataset contains
information on the emissions of a common pollutant (either par-
ticulate matter or sulfur dioxide) and other characteristics of U.S.
manufacturing plants that were in operation during the 1991 to
1998 period. I employ these data to examine how one of the most
politically contentious episodes of trade liberalization in U.S. history,
3
This is not to say that all studies rely on cross-country variation. Some research,
such as that of Dean (2002) or Chintrakarn and Millimet (2006), relies on within-
country variation in pollution and trade.
4
See, for example, Antweiler et al. (2001) or Frankel and Rose (2005).

http://dx.doi.org/10.1016/j.jinteco.2017.01.005
0022-1996/© 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
the North American Free Trade Agreement (NAFTA), affected the
pollution emitted by U.S. manufacturing plants.
As the origin of the debate over the environmental consequences
of globalization, NAFTA is a compelling setting for examining the
effects of trade liberalization on the environment. The agreement
was a passionately debated policy issue because it liberalized trade
between a developed country (the U.S.) and a developing country
with weaker environmental policy (Mexico).5 This difference in envi-
ronmental policy led to predictions that dirty, pollution intensive
production would relocate to Mexico after the agreement to take
advantage of lax environmental regulations, a phenomenon now
termed the pollution haven hypothesis.6 Yet, while the evidence
suggests that the agreement significantly increased North Ameri-
can trade (Romalis, 2007; Caliendo and Parro, 2015), there is little
evidence confirming what effect, if any, NAFTA had on pollution
emissions from manufacturing in the U.S. or Mexico.7
The structure of the agreement also makes it possible to examine
the effects of changes in both domestic (U.S.) and foreign (Mexico)
trade policy. There are three potential channels through which these
policy changes could affect plant pollution emissions: 1) decreases
in the cost of exporting to the Mexican market due to lower Mexican
tariffs, 2) decreases in protection from imports of products produced
in Mexico due to lower U.S. tariffs on final goods, and 3) decreases
in the cost of obtaining intermediate goods from Mexican producers
from lower U.S. tariffs on manufactured inputs. Previous research has
shown that international trade affects various aspects of plant activ-
ity through these channels, but to date, there is no evidence of their
effects on plant pollution emissions.8
The empirical challenge in this paper is credibly identifying the
effects of increased foreign market access, decreased protection from
import competition and increased access to imported intermediate
inputs following NAFTA given the possibility of concurrent changes
in environmental policies and other shocks. To do so, I utilize a triple-
difference research design that exploits temporal and cross-industry
variation due to changes in U.S. and Mexican trade policy, as well
as variation due to differences in geographically determined trade
frictions arising from the dispersion of manufacturing activity across
the U.S. prior to NAFTA.
I begin my analysis by examining NAFTA’s effects on the lev-
els of particulate matter (PM10 ) and sulfur dioxide (SO2 ) emitted by
U.S. manufacturing plants. I find robust evidence that NAFTA signif-
icantly reduced the emissions of both PM10 and SO2 from affected
plants.9 These reductions are due to responses to two aspects of
the liberalization: (i) increased access to the Mexican market, and
(ii) increased access to imported intermediate inputs from Mexico.
My preferred estimates indicate that, for the average plant, a 1%
increase in Mexican tariff preferences for U.S. goods reduced PM10
5
The agreement also liberalized trade between Canada and Mexico, but did not
affect trade between Canada and the U.S. which was previously liberalized as a result
of the Canada-U.S. Free Trade Agreement.
6
For example, during the 1992 presidential debates Ross Perot predicted that
NAFTA would lead to a “giant sucking sound” as U.S. production relocated to Mexico.
7
Research on the environmental effects of NAFTA follows from the influential
work of Grossman and Krueger (1991), who suggested that Mexico would not
necessarily become a pollution haven following the agreement. Subsequent studies
that have examined NAFTA’s effect on the environment, such as Cole (2004) and
Gamper-Rabindran (2006), examine changes in trade flows from clean and dirty
industries rather than examining changes in emissions from production directly.
Other research, such as Davis and Kahn (2010), has examined the effects of NAFTA on
the pollution from consumption.
8
Examples of this research include the work of Trefler (2004), Amiti and Konings
(2007), and Bustos (2011).
9
My baseline estimates include controls for the effects of the Mexican Peso Crisis,
changes in environmental regulations under the Clean Air Act, and ongoing trade
liberalization due to the Canada-U.S. free trade agreement. In the appendix, I also
show that my baseline estimates are robust to accounting for anticipatory responses
by plants and are not simply capturing the effects of pre-existing trends.
131
emissions by close to 1.30% and reduced SO2 emissions by just under
1.46%. Given that Mexican tariff preferences affect the decision of
an American plant to export to Mexico, these estimates suggest that
exporting can improve environmental quality by reducing pollution
emissions from existing plants. Yet, I find that the effects of liberal-
ized trade in intermediate inputs are much larger; for the average
plant, a 1% increase in U.S. tariff preferences for Mexican intermedi-
ate inputs reduced PM10 and SO2 emissions by approximately 3.25%
and 13.19% respectively. Given that tariff preferences for intermedi-
ates alter the cost of sourcing inputs from abroad, these estimates
suggest that importing can affect environmental quality by affecting
the inputs available to plants.10
Together, these estimates suggest that NAFTA played an impor-
tant role in the clean-up of U.S. manufacturing since the early 1990s.
Recent research by Shapiro and Walker (2016) indicates that, on
average, emissions of criteria pollutants from U.S. manufacturing
fell by roughly 60% between 1990 and 2008, despite a 35% increase
in output. Similarly, Levinson (2015) shows that aggregate PM10
and SO2 emissions from U.S. manufacturing fell by 3.55% and 3.61%
annually over this period, with nearly all of the decrease driven by
within-industry changes in pollution emissions. My estimates imply
that the effects of NAFTA account for nearly two-thirds of these
reductions on average; for the average affected plant, NAFTA reduced
PM10 emissions by 1.69% per year and reduced SO2 emissions by
3.06% per year.
This finding stands in contrast to the results presented by Shapiro
and Walker (2016), who suggest that the clean-up is primarily due to
the effects of changing environmental regulation. As such, one con-
cern with my baseline estimates is that they are simply capturing
the effects of ongoing revisions to environmental policy. To ensure
that this is not the case, I adopt two strategies that allow the effects
of environmental regulation to vary across plants on the basis of
how long they have been regulated. The resulting estimates provide
additional evidence that my baseline estimates are not capturing
the effects of environmental regulation, which further suggests that
trade liberalization following NAFTA played an important role in the
clean-up of the U.S. manufacturing sector.
The next step in my analysis is to examine whether the changes
in emissions levels that I observe are due to changes in the physi-
cal quantity of output produced, or changes in the level of pollution
emitted per unit of output (the emission intensity of production).
By definition, trade liberalization can effect pollution emissions via
either one of these two channels; to distinguish between the two,
I analyze the effect of trade liberalization on the emission intensity of
production at each plant. These estimates indicate that the changes
in emission levels that I observe are primarily due to changes in
emission intensity, not changes in the levels of output at affected
plants.
Finally, I turn to examine three possible explanations for the
reductions in emission intensity. First, I investigate whether the
reductions result from extensive margin changes due to plant entry
and exit. Given recent evidence that more productive firms have
lower emission intensities (e.g. Bloom et al. (2010), Shapiro and
Walker (2016)), standard models of international trade featuring
firm heterogeneity (e.g. Melitz (2003), Melitz and Ottaviano (2008))
would predict that trade liberalization could lead to lower average
plant emission intensity as a result of entry and exit.11 Moreover,
previous studies have emphasized that the decline in pollution from
10
Although the estimates for PM10 and SO2 suggest that trade liberalization will
lower plant pollution emissions necessarily, such reductions are not ubiquitous. In the
appendix, I show that NAFTA had little to no effect on the emissions of volatile organic
compounds from U.S. manufacturing plants.
11
For such a model see, for example, Baldwin and Ravetti (2014) or Cherniwchan
et al. (2017).
132 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
U.S. manufacturing may be due to plant entry and exit as produc-
tion shifts between clean and dirty industries (e.g. Levinson (2009,
2015)) or between plants within an industry (e.g. Shapiro and Walker
(2016)). As such, I re-estimate the effects of NAFTA on the emission
intensity of both PM10 and SO2 using sub-samples of plants that are
in operation throughout my period of study. These estimates indicate
that the NAFTA’s effects are primarily driven by the intensive margin;
the emission intensity reductions that I observe are predominantly a
result of within-plant changes rather than plant entry and exit.
Second, I examine whether the emission intensity reductions are
a product of changes in the mix of imported intermediate inputs. As
mentioned above, in the lead up to NAFTA it was often suggested
that Mexico would become a pollution haven as dirty U.S. production
relocated to Mexico to take advantage of differences in environmen-
tal regulations across the two countries. I find evidence consistent
with this motive; my estimates indicate that the intermediate input
driven reductions in PM10 and SO2 emission intensity are primarily
due to increased access to relatively dirty intermediate inputs. These
findings suggest that after NAFTA, affected plants began to source
relatively dirty intermediate inputs from Mexico rather than pro-
duce them in house due to differences in the costs of environmental
regulation.
Third, I ask whether the reductions in emission intensity that I
observe in response to increased foreign market access are consistent
with trade-induced changes in process innovation or technologi-
cal upgrading. At present, the relationship between foreign market
access and plant emission intensity is poorly understood, but there
are two reasons to focus on this channel. First, a number of stud-
ies have indicated that increased foreign market access can increase
firm productivity by creating incentives for firms to engage in pro-
cess innovation or adopt new technologies (e.g. Costantini and Melitz
(2008), Atkeson and Burstein (2010), Bustos (2011)); the existing
empirical evidence suggests that productivity and emission inten-
sity are negatively related, meaning that such innovation or adoption
could reduce plant emission intensity. Second, a burgeoning liter-
ature has adapted the theoretical framework developed in Bustos
(2011) to suggest that increased foreign market access can lower
emission intensity directly by creating incentives for firms to alter
their pollution abatement technologies (e.g. Cui et al. (2012), Forslid
et al. (2014)).
Testing this channel is challenging as I do not observe the
technology choices or productivity of plants in my data.12 As such,
I exploit a theoretical prediction from the Bustos (2011) model of
trade and technological upgrading as it has been used to formalize
the effects of trade on both forms of technology adoption. Her frame-
work suggests that increased foreign market access will only affect
the subset of plants that are induced to adopt new technology; the
effects of trade will vary across plants based on their initial produc-
tivity. To test this prediction, I follow Bustos by proxying for initial
productivity using initial plant size, and then estimating whether the
effects of increased foreign market access on plant emission inten-
sity vary across quartiles of the initial plant size distribution. I find
that the observed reductions in emission intensity due to increased
Mexican market access are heterogeneous and concentrated in the
upper quartiles of the initial plant size distribution, which is con-
sistent with previous evidence on technological upgrading following
trade liberalization. These estimates provide suggestive evidence
that emission intensity is falling due to trade-induced technological
change following NAFTA.
Altogether, my findings contribute to a large literature examining
the effects of international trade on the environment by providing
12
Unfortunately, the NETS data does not include information on plant capital stocks
or material use, making it impossible to estimate productivity estimation using
standard methods.
robust evidence of the micro-foundations underlying the aggregate
relationship between international trade and environmental quality.
These micro-foundations are the topic of a burgeoning literature, but
given the dearth of data linking plant pollution emissions to other
plant characteristics, most studies have typically inferred the effects
of trade from changes in fuel use (e.g. Martin (2012), Forslid et al.
(2014), Barrows and Ollivier (2014)) or ambient pollution concentra-
tions (e.g. Gutierrez and Teshima (2011)). Furthermore, while trade
can affect the activity of plants by decreasing the cost of export-
ing, increasing competition from imports or decreasing the cost of
obtaining intermediate inputs, studies typically focus on a single
channel. Indeed, the few studies that utilize emissions data have
focused on the relationship between export status and pollution
emissions (e.g. Holladay (2016), and Cui et al. (2016)). In contrast,
this paper provides evidence of the relationship between trade lib-
eralization and the pollution emitted by individual manufacturing
plants through three distinct channels, and examines several possi-
ble mechanisms for the observed changes in pollution emissions.
This paper also contributes to a literature examining the effects
of NAFTA. Most of these studies have examined how the agree-
ment affected trade creation and diversion (Romalis, 2007; Caliendo
and Parro, 2015) or labor market outcomes (Burfisher et al., 2001;
McLaren and Hakobyan, 2016). This study is the first to examine the
effects of NAFTA on the pollution emitted by manufacturing plants in
the U.S.
The remainder of this paper proceeds as follows. Section 2
outlines my research design. Section 3 discusses the data, while
Section 4 presents my empirical specification. Section 5 presents my
results. Finally, Section 6 concludes.
2. Identifying NAFTA’s effects on the environment
The goal of this paper is to estimate the effects of trade liberal-
ization between the U.S. and Mexico following the North American
Free Trade Agreement (NAFTA) on the pollution emitted by affected
U.S. manufacturing plants. NAFTA came into effect on January 1st,
1994, after several years of negotiations and debate. The agreement
liberalized bilateral trade between the United States and Mexico and
between Canada and Mexico, leading to the removal of all tariff barri-
ers between member countries.13 While there were no exceptions to
the requirement that bilateral tariffs be reduced to zero, the structure
of the agreement meant that speed of tariff reductions varied across
industries once NAFTA was enacted (Kowalczyk and Davis, 1998).
In some industries, such as Concrete Block and Brick Manufacturing
(SIC 3271), tariffs were eliminated immediately. In other cases, such
as Electronic Capacitors (SIC 3675), tariffs were eliminated gradually
following a set schedule of reductions.14 However, NAFTA did not
affect the tariffs of all industries; some, such as Gaskets, Packing, and
Sealing Devices (SIC 3053), were already traded freely prior to the
agreement.15
Identifying NAFTA’s effects is complicated by the fact that these
reductions were the result of trilateral negotiations between Canada,
Mexico and the U.S. Although differences in tariff rates across
industries and time are potentially useful sources of variation, the
13
Canada-U.S. trade was previously liberalized as a result of the Canada-U.S. Free
Trade Agreement.
14
Most tariffs were to be eliminated immediately, or in equal-sized annual tariff
reductions spaced over a period five, seven or ten years. A few tariffs were phased
out over a longer period; roughly 1% of U.S. imports from Mexico and 1.5% of Mexican
imports from the U.S. obtained a 15-year phase out (Kowalczyk and Davis, 1998). The
length of each phase-out period was negotiated prior to the agreement.
15
As reported by Kowalczyk and Davis (1998), a substantial amount of trade
between the U.S. and Mexico was duty free prior to NAFTA. In particular, 14% of
U.S. imports from Mexico and 18% of Mexican imports from the U.S. were unaffected
because they were already freely traded prior to the agreement.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
speed of tariff reductions could reflect the effects of unobserved
industry-specific political economy factors. There are strong theo-
retical reasons to believe this is the case (Grossman and Helpman,
1994), and existing empirical evidence indicates political economy
motives are an important determinant of trade barriers in the U.S.
(Trefler, 1993; Goldberg and Maggi, 1999). Indeed, Kowalczyk and
Davis (1998) suggest that the cross-industry variation in the speed
of tariff reductions following NAFTA can be explained as a product
of industry-specific motives for protection that existed prior to the
agreement; industries with high initial tariffs typically obtained the
longest tariff phase-out periods.
I address this issue by exploiting additional variation in the trade
barriers faced by manufacturing plants within the same industry. The
starting point for my approach is the observation that tariffs are only
one component of the trade barriers that affect trade flows. Trade
flows are also impeded by other trade barriers created by geogra-
phy, such as transportation costs, information costs, and distribution
costs. Moreover, geographically determined trade barriers are large
in magnitude relative to tariffs; as Anderson and van Wincoop (2004)
indicate, trade policy accounts for less than 5% of total trade costs
for the representative industrialized nation. These geographic fric-
tions create an additional source of variation that can be employed
to identify NAFTA’s effects.
This variation arises because the geographically determined costs
of trading with Mexico differ across the continental U.S. As I show
in Section 3.2 below, the cost of trading with Mexico varies signifi-
cantly across states. As such, even in the absence of any tariff policy,
producers in some states face large trade costs while producers in
other states do not. Such cost differences mean that, prior to NAFTA,
there was significant variation in the level of exposure to the Mexican
market across the U.S.16
I combine the variation in trade barriers created by cross-
state differences in geographically determined trade costs with the
cross-industry and temporal variation in trade barriers arising from
changes in U.S. and Mexican tariff policy to identify the effects of
NAFTA. Specifically, I exploit the fact that the geographic variation
in trade costs means that the effects of a tariff reduction will dif-
fer across states. If, for example, a plant faced prohibitively high
geographic frictions prior to NAFTA, a reduction in domestic tariffs
would not increase competition from Mexican imports because the
plant remained protected by its location. On the other hand, if the
plant was located in a state where trade was completely friction-
less, it would be fully exposed to the change in trade policy. Thus,
for any industry, only a fraction of plants could be affected by an
increase in market access, an increase in access to imported inputs
or an increase in import competition depending on their location.
Given the cross-industry variation in bilateral tariff reductions, this
means the subset of plants in liberalized industries that are located
in states with low frictions for trade with Mexico will be affected
by the changes in domestic and foreign trade policy. These are the
plants that are “treated” by the different reductions in tariffs. Identi-
fying NAFTA’s effects amounts to measuring the effects of each tariff
reduction for these plants. I do so using a triple difference approach.
For each policy change, I start by comparing the average out-
comes of treated plants before and after NAFTA. This comparison
allows me to control for any unobserved time-invariant, industry,
state or plant characteristics that affect pollution emissions. More
importantly, this difference would measure the average causal effect
of trade liberalization in the absence of concurrent policy changes or
16
Indeed, geographic variation in exposure to the Mexican market has been at the
center of much of the popular debate about the effects of the agreement. Prior to the
agreement it was believed that NAFTA would primarily benefit those states, such as
California or Texas, with the closest ties to Mexico. Since the agreement, NAFTA has
been blamed for Ohio’s decline (Leonhardt, 2008), Texas’s success (Kumar, 2006), as
well as various changes in state employment throughout the U.S. (Bolle, 2000).
133
secular changes in outcomes. However, it is unlikely that this com-
parison captures the average causal effect in the present setting.
As mentioned above, there is strong reason to believe that the tar-
iff changes that I observe are driven by political economy factors,
meaning the treatment group is treated because of an underlying
industry trend. Moreover, there were contemporaneous changes in
policy at the time of NAFTA. Isolating the causal effects of the agree-
ment requires controlling for these shocks. To do so, I exploit the
cross-industry and cross-location variation in protection.
The variation in protection across locations allows me to com-
pare the difference in average outcomes for plants in low trade
costs states with the corresponding difference for plants from the
same industry in states with high trade costs. Thus, for a liberal-
ized industry, I treat the high trade cost plants as a counterfactual
for what would have happened to pollution in the industry in the
absence of liberalization. This allows me to control for any unob-
served time-varying, industry-specific political economy factors that
affected tariff reductions, as well as any industry-specific changes in
environmental regulation during my period of study.
The cross-industry variation in the trade policies faced by plants
within a state allows me to control for unobserved time-varying state
heterogeneity in a similar manner. Hence, in a low trade cost state,
I treat the difference in average outcomes from unliberalized plants
as the counterfactual for the average change for liberalized plants.
Employing such a counterfactual is important because there is also
some evidence that state characteristics determine the trade policy
preferences of a constituency (see for example, Scheve and Slaughter
(2001) or Beaulieu (2002)), meaning that trade policy may be corre-
lated with unobserved state features. Moreover, state legislators may
enact policies to offset the effects of trade liberalization within their
domain, meaning there may be confounding state-specific policy
changes, such as revisions to state-wide environmental policy.
The cross-industry and cross-location variation in protection also
allows me to compare the difference in average outcomes from
treated plants with the difference in average outcomes from plants
facing high trade costs from an unliberalized industry. Given that
these plants are not in the same industry or state as the treated
plants, and are not faced by the same changes in policy, they cap-
ture the underlying aggregate trend in pollution. This allows me to
control for nationwide shocks, such as changes in aggregate expen-
ditures due to the recession at the beginning of the 1990’s, aggregate
technological change, country-wide policy changes, or the increased
use of global supply chains by U.S. firms.17
For this research design to identify the effects of NAFTA, the vari-
ation in geographically determined trade costs across states must be
plausibly exogenous. There is strong reason to believe that this is the
case. The available evidence suggests that U.S. plants primarily serve
local domestic markets (Hillberry and Hummels, 2008; Holmes and
Stevens, 2012) and the location of these markets is driven by long
run trends (Kim, 1995). This suggests that plants were not systemat-
ically sorting across locations in anticipation of NAFTA.18 Moreover,
the uncertainty over the passage of the agreement minimizes the
possibility that plant locations were chosen in anticipation of NAFTA.
Despite its prominence as a policy issue at the time, the uncer-
tainty over NAFTA’s passage means that the agreement functioned
as a de facto shock for U.S. manufacturers. Trilateral negotiations
began in 1991 and the initial agreement was signed 18 months later,
nearly one year before NAFTA came into effect. In principle, this
means that producers had close to 12 months to adjust their pro-
duction decisions prior to the agreement. However, there was a great
17
Harrison and McMillan (2011) examine the increased use of offshoring by U.S.
firms during this time.
18
Indeed, Hanson (1998) indicates that the agreement had little effect on the
location of U.S. manufacturing facilities.
134 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

deal of uncertainty about the future of the NAFTA throughout this Table 1
period, particularly after the election of President Bill Clinton in 1992. Descriptive statistics: 1991–1998.

Clinton required additional side deals on labor and the environment Pollutant Variable Obs. Mean Std. dev. Median
before sending the agreement to Congress.19 The negotiation of these PM10 Emissions (lb) 4162 22889 309314 176
side agreements created uncertainty over when NAFTA would be Employees 4162 587 1247 240
ratified (if at all).20 This uncertainty persisted even after the side Sales (1,000,000s) 4162 78 147 26
agreements were signed in September of 1993, due to political oppo- Emissions/worker 4162 48 453 1
Emissions/sales 4162 367 3258 7
sition in Congress, and continued until NAFTA was finally ratified
SO2 Emissions (lb) 1645 46497 227460 2000
on November 20, 1993. Hence, the uncertainty was resolved mere Employees 1645 726 1519 350
weeks before NAFTA came into effect, leaving producers little time Sales (1,000,000s) 1645 165 637 56
to respond beforehand. Emissions/worker 1645 187 1119 6
Emissions/sales 1645 1903 9374 46

Notes: Table reports summary statistics of key variables for the two data sets examined
in the main analysis. Each panel reports statistics for a different dataset.
3. Data and measurement

3.1. Data overview
To conduct my analysis, I rely on two unique data sets com-
piled from a variety of sources. My primary data source is the Toxic
Release Inventory (TRI) maintained by the U.S. Environmental Pro-
tection Agency (EPA). When combined with the concordance used by
the EPA to classify TRI chemicals into the chemical groups reported in
the National Emission Inventory (Environmental Protection Agency,
2013), this dataset gives me information on the emissions of particu-
late matter (PM10 ) and sulfur dioxide (SO2 ) from U.S. manufacturing
plants. I use this data to construct a dataset containing information
on the plant-level emissions for each pollutant. I then match these
two datasets to proprietary data on plant employment and sales from
the National Establishment Time Series (NETS), tariff data for the
U.S. from Feenstra et al. (2002) and for Mexico from Romalis (2007),
trade data from Schott (2008) and the UN Comtrade database, and
data on state exports and imports for 1993 from Feenstra (1997) and
the North American Transborder Freight Data database maintained
by the U.S. Bureau of Transportation Statistics. After matching, each
dataset is an unbalanced panel containing information on pollution
emissions and other plant characteristics covering the years 1991 to
1998. Additional details on the data sources and the construction of
the datasets are provided in the appendix.
and emit less emissions per worker and per million dollars of sales
than plants that emit SO2 .22
The differences in plant characteristics across datasets reflect dif-
ferences in how each pollutant is produced and abated both within
and across industries. In the U.S., only a few industries produce sig-
nificant quantities of PM10 , or SO2 (Greenstone, 2002), and there
is significant variation in the pollution elasticity of production for
each pollutant across industries (Shapiro and Walker, 2016). This
suggests that the differences in plant characteristics that I observe
reflect differences in the characteristics and production processes of
the industries that emit each pollutant. These differences may lead to
different responses to trade liberalization, a possibility I investigate
below.
For each dataset there is substantial variation in the location
of plants across the continental U.S. This variation is depicted in
Fig. 1, which illustrates the number of plants in each state that pro-
duce each pollutant. As the figure shows, in each dataset, plants are
primarily located in states that are associated with the manufactur-
ing sector: Texas and California in the southwest and the Rust-Belt
states of the northeast. Given that distance to market is an important
determinant of trade costs (Anderson and van Wincoop, 2004), the
variation in location depicted in the figure provides suggestive evi-
dence that there is considerable variation in the trade frictions faced
by plants in each dataset.

3.1.1. Descriptive statistics 3.2. Measuring trade frictions

Descriptive statistics of key variables for both datasets that I
employ are reported in Table 1. The dataset for PM10 contains
4162 plant-year observations, and the dataset for SO2 contains 1645
plant-year observations.21
The summary statistics reported in Table 1 indicate that there are
substantial differences in the characteristics of plants that emit PM10
and SO2 . Plants that emit PM10 have fewer employees, lower sales,
19
The environmental side agreement created an institution known as the North
American Commission for Environmental Cooperation (CEC) to address environmental
issues related to NAFTA. The CEC is meant to address the three core issues of the side
agreement (Hufbauer et al., 2000): (i) the improvement of environmental conditions
through cooperative initiatives, (ii) ensuing the implementation of environmental leg-
islation, and (iii) mediation of environmental disputes. In principle, the creation of
the CEC could influence the estimates I obtain below. However, this is unlikely to
be the case in practice. The CEC has been largely ineffective, primarily because of
budget shortcomings and the inability to impose sanctions on governments that fail
to address environmental problems (Hufbauer et al., 2000; Markell and Knox, 2003).
20
For example, shortly after Clinton’s election, when Democratic campaigner Barry
Carter was asked when Clinton would send the NAFTA legislation to Congress, Carter
replied, “It will depend on how long it takes to negotiate additional agreements on
labor and the environment.” (Cameron and Tomlin, 2000).
21
Table A1 in the appendix presents additional descriptive statistics comparing
plants in each dataset to plants from the Census of Manufacturers.
My research design relies on variation in trade frictions arising
from changes in trade policy and the geographic dispersion of man-
ufacturing plants across the U.S. to identify NAFTA’s effects. The
following describes how I measure these trade frictions.
3.2.1. Trade policy
A key feature of NAFTA is that it is a preferential trade agree-
ment, meaning that it reduced barriers to trade between member
countries without changing the barriers that they extended to other
nations around the world. This is illustrated in Fig. 2, which depicts
the import-weighted average tariff rates on all imports for the U.S.
and Mexico during my sample period.23 Panel (a) plots Mexico’s
average tariff on imports from the United States, labeled USA, and
the rest of the world, labeled ROW. Panel (b) plots the average U.S.
22
Although there is substantial variation in employment and sales across both
datasets, the average plant in each is larger than the average manufacturing plant
in the U.S. owing to the reporting requirements for the TRI. Plants are only required
to report to the TRI if they employ at least ten employees, meaning that small
manufacturing plants are systematically excluded from each data set. For further
details, see the appendix.
23
I construct the import-weighted average tariff measures using the tariff data from
Feenstra et al. (2002) and Romalis (2007), and the trade data from Schott (2008) and
the UN Comtrade database. For further details, see the appendix.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149 135

Number of Plants Number of Plants

15 - 50 9 - 21
7 - 15 5-9
3-7 2-5
1-3 1-2
No data No data

Fig. 1. Plant locations.

.16

.06

.03
.14

.025
.05
.12

.04

.02
.1
Tariff

Tariff

Tariff
.015
.08

.03
.06

.02

.01
.04

.005
.01
.02
0

0
1991 1992 1993 1994 1995 1996 1997 1998 1991 1992 1993 1994 1995 1996 1997 1998 1991 1992 1993 1994 1995 1996 1997 1998
Year Year Year

USA ROW MEX GSP ROW MEX GSP ROW

Fig. 2. Bilateral tariff reductions.

tariff on imports of final goods from Mexico, labeled MEX, as well
as the average tariff rates the U.S. extended to the rest of the world,
labeled GSP and ROW. Finally, panel (c) plots average U.S. tariffs on
imports of intermediate inputs.24 As in panel (b), the average tariff
rate on imports of intermediates from Mexico is labeled MEX, while
the average tariff rates on imports of intermediates from the rest of
world are labeled GSP and ROW.
The preferential nature of NAFTA is evident from the figure. For
example, consider U.S. tariffs on imports of final goods shown in
panel (b). Prior to NAFTA, most imports from Mexico fell under the
Generalized System of Preferences (GSP) that extended lower tariff
rates to developing countries. As such, the average tariff on imports
from Mexico was approximately 3%, while the average tariff the U.S.
extended to developed countries (the ROW tariff) was approximately
5.5%. After NAFTA was implemented in 1994 (denoted by the vertical
line), the average tariff on Mexican imports fell to close to 0.5%, while
the average GSP tariff remained near its pre-NAFTA level. Hence, fol-
lowing NAFTA, U.S. trade policy favored imports from Mexico relative
to imports from the rest of the world.
I utilize this feature of NAFTA to measure the magnitude of trade
liberalization between the U.S. and Mexico. Specifically, I measure
bilateral trade liberalization between the U.S. and Mexico using
the changes in foreign and domestic tariff preferences for output
24
Given that tariffs on intermediate inputs are not available in the data from
Feenstra et al. (2002), I follow the approach of Amiti and Konings (2007) and construct
input tariffs for each industry using data on output tariffs and cost shares. Doing so, the

input tariff for industry j at any time t is given by tjtInt = h hhj × tjtUS , where tjtInt and tjtUS
denote the U.S. ad-valorem tariffs on inputs and final goods in industry j at time t, and
hhj is the cost share of input h in the total cost of output j from the Bureau of Economic
Analysis’s (BEA) benchmark input-output table for 1992. While this approach allows
me to examine the effects of liberalized trade in intermediate inputs, it is not without
costs. In particular, the resulting input tariffs, tjtInt , could be measured with error. This
would create an attenuation bias, meaning the estimates presented below understate
the effects of liberalized trade in intermediate inputs.
from each industry.25 Letting tjtMex−d denote Mexican ad-valorem
tariff rates on imports from industry j in region d, Mexican tariff
preferences for U.S. imports are given by:
tjtMex = tjtMex−ROW − tjtMex−US (1)
This value corresponds to the vertical distance between the lines
labeled USA and ROW in panel (a) of Fig. 2. Similarly, U.S. tariff prefer-
ences for imports of final goods and intermediate inputs from Mexico
are given by:
tjtUS = tjtUS−GSP − tjtUS−Mex and tjtInt = tjtInt−GSP − tjtInt−Mex (2)
where tjtUS−d and tjtInt−d denote U.S. ad-valorem tariff rates on import
of final goods and intermediate inputs from industry j in region
d, respectively. These values correspond to the vertical distance
between the lines labeled MEX and GSP in panels (b) and (c).
The variation in tariff preferences for each dataset is summarized
in Table 2, which reports the mean and standard deviation of the
level of tariff preference for all industries before and after NAFTA,
as well as the mean and standard deviation of the change in tar-
iff preferences due to the agreement. As the table shows, neither
the U.S. nor Mexico had a preferential tariff for imports from each
other prior to NAFTA. In fact, U.S. trade policy was slightly biased
against imports from Mexico; for each dataset, U.S. tariff preference
25
This approach exploits the fact that the preferential nature of the agreement cre-
ates an implicit counterfactual for the level of protection that would have occurred
in the absence of trade liberalization. This makes it possible to control for ongoing
trade liberalization due to the Uruguay Round of the General Agreement on Tariffs
and Trade and changes to the Generalized System of Preferences (GSP) that would not
be captured by the changes in bilateral tariffs between the U.S. and Mexico. A similar
strategy was employed by Trefler (2004) to analyze the effects of the Canada-U.S. Free
Trade Agreement.
136 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

Table 2
Summary of U.S. and Mexican tariff preferences pre- and post-NAFTA.

Pollutant Variable Mean Mean Mean

pre-NAFTA post-NAFTA change in
preference

PM10 Mexico (TjtMex ) −0.0007 0.0644 0.0651

(0.0037) (0.0244) (0.0253)
U.S. - All Imports (TjtUS ) −0.0009 0.0041 0.0050
Normalized HRI
(0.0030) (0.0121) (0.0122) (.0048874,.0887409]
(.0028056,.0048874]
U.S. - Intermediates (TjtInt ) −0.0008 0.0047 0.0056 (.0008039,.0028056]
[0,.0008039]
(0.0016) (0.0085) (0.0086) No data

SO2 Mexico (TjtMex ) −0.0001 0.0641 0.0606

(0.0034) (0.0618) (0.0583) Fig. 3. Normalized Head-Ries Index by State.
U.S. - All Imports (TjtUS ) −0.0004 0.0061 0.0066
(0.0017) (0.0114) (0.0111)
U.S. - Intermediates (TjtInt ) −0.0002 0.0046 0.0049
and imports for 1993 from Feenstra (1997) and the North American
(0.0003) (0.0052) (0.0052)
Transborder Freight Data database.27
Notes: Table reports summary of changes in tariff preferences for each dataset. The HRI measures all trade costs associated with shipping goods
Each panel reports statistics for a different dataset. Standard deviations reported in
between state s and Mexico. As such, it not only captures state
parentheses.
specific trade costs; it also captures the effects of trade policy which
is common across states. To isolate the state-specific component of
on all imports and imported intermediate inputs is small and nega- trade costs, I normalize the HRI by the average HRI for all continental
tive, indicating that the tariffs the U.S. placed on Mexican products U.S. states to obtain:28
were slightly higher than tariffs placed on products from elsewhere
in the world. On average, NAFTA increased Mexican tariff preferences 0Mex
s
Vs ≡ 1 Mex
(4)
by 6.06 to 6.51 percentage points, U.S. tariff preferences on final good s s∈S 0s
imports by 0.50 to 0.66 percentage points, and U.S. tariff preferences
on imports of intermediate inputs by 0.49 to 0.56 percentage points. The normalized HRI, Vs ∈ [0, ∞), measures the state-specific trade
While these changes are small relative to the changes in trade policy frictions associated with bilateral trade between state s and Mexico.
that occurred during other recent episodes of trade liberalization, the This measure varies inversely with the size of trade frictions, mean-
work of Romalis (2007) and Caliendo and Parro (2015) has shown ing Vs increases as the costs of trade fall. Thus, based on this measure,
that these policy changes had a large effect on trade. Moreover, as trade with Mexico is easier for states with high values of Vs .
the final column of Table 2 demonstrates, the changes in tariff pref- The variation in Vs across states is displayed in Fig. 3. In the figure,
erences vary across industries in each data set. As discussed above, I darker values correspond to higher values of Vs , meaning that darker
exploit this variation to identify the effects of the agreement. states have lower trade costs than lighter states. As the figure shows,
there is substantial geographic variation in trade costs across states.
Finally, low trade costs states are not concentrated in the south
3.2.2. Geographic trade frictions
and southwest regions of the U.S.; states in the northeast are also
Implementing my research design also requires a method of
highly exposed. This feature reflects elements of trade costs, such as
quantifying differences in the ease of trading with Mexico across
transport infrastructure and distribution networks, that would not
locations. Ideally I would be able to employ data on geographic fric-
be captured by a naive measure (such as distance from the Mexican
tions for this purpose, but unfortunately, such data are currently
border).
unavailable. As such, I construct a measure of trade frictions using the
most disaggregate trade data available: state-level data on bilateral
trade with Mexico. 4. Empirical specification
I begin by constructing the Head-Ries Index (HRI), a measure of
trade costs which takes the form: I implement my research design using the following empirical
specification:

 Mex s      
X XMex
0Mex = ss (3) ln zijst = b1 tjtMex ×Vs + b2 tjtUS ×Vs + b3 tjtInt ×Vs
s Mex
Xs XMex
+ qi + ljt + dst + eijst (5)

s
where XsMex denotes total exports from state s to Mexico, XMex
s
denotes Mexico’s exports to state s, Xs denotes total production from
Mex
state s that is consumed in state, and XMex denotes the value of
Mexican production that is consumed domestically. First proposed
by Head and Ries (2001), this index provides a theoretically con-
sistent measure of trade costs; it can be derived from a large class
of trade models (Head and Mayer, 2013; Novy, 2013). Moreover,
it largely captures the effect of geography (Novy, 2013), and can
be readily calculated using observable data.26 I construct the HRI
for each state in the continental U.S. using data on state exports
26 Mex
I calculate Xss and XMex as the difference between the value of total production
minus total exports.
where lnzijst is the natural log of pollution emissions from plant i
in industry j and state s at time t.29 tjtMex , tjtUS and tjtInt are the tariff
preferences defined in Eqs. (1)–(2), and Vs is the normalized HRI
index defined in Eq. (4). qi is a plant fixed effect that captures
time-invariant plant heterogeneity, such as plant-level preferences
for trade liberalization, as well as constant industry and geographic
27
Employing data from prior to NAFTA is important as it ensures that the measured
difference in exposure is not the result of an endogenous response to the agreement.
I employ data from 1993 because data on state imports from Mexico are not available
for prior years.
28
This normalization is also convenient for the interpretation of the estimates
presented below.
29
I employ the natural log of pollution emissions as the dependent variable to
remove the skewness in the distribution of emissions for each pollutant.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149 137

characteristics that affect pollution emissions. l jt are industry×year Table 3

fixed effects that capture temporal variation in industry outcomes, NAFTA and pollution emissions.

such as industry-specific demand shocks or changes in industry- (1) (2) (3) (4) (5)
wide environmental policy. dst are state×year fixed effects that Panel A: Particulate Matter (PM10 )
capture state specific shocks such as changes in state-wide envi- TjtMex × Vs −1.267b −1.403b −1.175b −1.268b −1.298b
ronmental policy. eijst is an error term that captures idiosyncratic (0.541) (0.559) (0.548) (0.540) (0.566)
changes in emissions. TjtUS × Vs 1.141 1.020 1.286 1.139 1.169
The coefficients of interest in Eq. (5) are b1 , b2 and b3 . The (2.064) (2.069) (2.113) (2.065) (2.116)
TjtInt × Vs −3.352 −3.227 −3.381 −3.349 −3.247
coefficient b1 measures the average percentage change in pollution
(3.842) (3.889) (3.897) (3.847) (3.944)
emissions for plants that experience an increase in access to the Peso Crisis X X
Mexican market following trade liberalization relative to the change CUSFTA X X
in pollution for plants in the same industry or the same state that do 1990 CAAAs X X
not. Similarly, b2 measures the average percentage change in pollu- Observations 4162 4162 4162 4162 4162
Adjusted R2 0.19 0.19 0.20 0.19 0.20
tion from plants that experience an increase in Mexican competition
following trade liberalization relative to the change in pollution for Panel B: Sulfur Dioxide (SO2 )
plants in the same industry or same state that do not. b3 measures TjtMex ×Vs −1.457c −1.492c −1.429c −1.452b −1.459c
the average percentage change in pollution from plants that expe- (0.742) (0.759) (0.738) (0.733) (0.746)
rience a decrease in the cost of obtaining intermediate inputs from TjtUS ×Vs −3.792 −3.845 −3.918 −3.695 −3.871
(3.180) (3.182) (3.184) (3.142) (3.150)
Mexico following trade liberalization relative to the change in pollu-
TjtInt ×Vs −13.143a −13.118a −13.174a −13.177a −13.185a
tion for plants in the same industry or same state that do not. These
(5.033) (5.031) (5.035) (4.996) (4.996)
coefficients are identified from within industry×state comparisons Peso Crisis X X
over time. CUSFTA X X
Eq. (5) is a triple difference estimator with continuous measures 1990 CAAAs X X
of treatment. As such, identification requires that there be no other Observations 1645 1645 1645 1645 1645
Adjusted R2 0.25 0.25 0.25 0.25 0.25
factors aside from the changes in trade policy due to NAFTA gen-
erating a difference in outcomes for plants that were affected and Notes: Table reports the estimated effects of NAFTA on the emissions of PM10 and
SO2 from U.S. manufacturing plants. Each panel contains estimates for a different
unaffected by the agreement. However, my period of study overlaps
pollutant and reports regression coefficients from five separate regressions based
with three additional events that might have generated such dif- on Eq. (5). In all cases, the dependent variable is the log of pollution emissions. All
ferences: the Mexican Peso Crisis, ongoing trade liberalization due specifications include plant, industry×year and state×year fixed effects. See text for
to the Canada-U.S. Free Trade Agreement (CUSFTA) and the imple- details. Standard errors clustered by industry and state are reported in parentheses.
mentation of the 1990 Clean Air Act Amendments (CAAAs). In what Significance: a 1%; b 5%; c 10%.
follows, I also estimate versions of Eq. (5) that account for the effects
of these shocks.
for any ongoing changes in environmental policy due to the amend-
ments. My preferred specification, reported in column (5), controls
5. Results for all of these shocks simultaneously. In all cases, standard errors
are clustered by industry×state.34
5.1. NAFTA and pollution levels Panel A of Table 3 reports estimates of NAFTA’s effects on the
emissions of PM10 . The parameter estimates reported in row one
Table 3 reports estimates of five specifications based on Eq. (5) for of the panel indicate that there were significant reductions in PM10
PM10 and SO2 . The first specification, reported in column (1) of each emissions from affected plants due increased Mexican market access
panel, includes plant, industry×year, and state×year fixed effects. following NAFTA; the coefficient estimate from my preferred spec-
Column (2) of each panel adds an interaction of the Mexican lib- ification suggests that for a plant facing average trade frictions
eralization measure with an indicator for the Peso Crisis.30 This is (that is, a plant with Vs = 1) a one percentage point increase
done to account for the significant contraction of the Mexican econ- in Mexican tariff preferences for U.S. goods reduced emissions by
omy following the Peso Crisis of 1995, which may potentially offset 1.298%. Although imprecise, the estimates reported in row two sug-
the effects of increased Mexican market access due to NAFTA.31 In gest that increased import competition following NAFTA may have
column (3) of each panel, I account for ongoing trade liberalization increased PM10 emissions. My preferred estimate indicates that for a
between the U.S. and Canada plant facing average trade frictions, a one percentage point increase
 dueto CUSFTA  using measures that
are analogous to tjtMex × Vs and tjtUS × Vs .32 Column (4) of each in U.S. tariff preferences for final goods from Mexico increased emis-
sions by 1.169%. Finally, the parameter estimates reported in row
panel includes an indicator reflecting the environmental regulations three of the panel indicates that reductions in the cost of accessing
faced by each plant under the 1990 CAAAs.33 This is done to account intermediate inputs from Mexico also decreased emissions of PM10 .
The estimate from my preferred specification indicates that a one
percentage point increase in U.S. tariff preferences for Mexican inter-
mediates reduced PM10 emissions from a plant facing average trade
30
The Peso Crisis indicator is equal to one for the year 1995. frictions by 3.247%. While it is imprecisely estimated, this decrease
31
The available evidence suggests the Peso Crisis decreased exports from the US, but
did not affect US imports from Mexico (Burfisher et al., 2001).
is large and suggests that plants change their pollution emissions
32
Specifically, I add measures of the bilateral tariff preferences for Canada and the by altering their input choices. Altogether, these estimates provide
U.S. (tjtCan−US and tjtUS−Can respectively) interacted with measures of bilateral trade fric- robust evidence that NAFTA had significant effects on the level of

tions between each state s and Canada VCan s . Further details of the construction of PM10 emitted by individual U.S. manufacturing plants.
these variables is given in the appendix.
33
This measure reflects the attainment status of the county in which each plant is
located. Under the CAAA, “non-attainment” counties where ambient pollution con-
centrations exceed the National Ambient Air Quality Standards (NAAQS) have stricter
34
regulations than “attainment” counties where ambient pollution concentrations fall In the appendix, I present additional results showing why baseline estimates are
below air quality standards. The indicator is equal to one if the plant is located in a robust to accounting for anticipatory responses by plants, and are not simply capturing
non-attainment county for the pollutant under consideration in a given year. the effects of pre-existing differences in industry×state trends.
138 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

The estimates reported in Panel B indicate that NAFTA also sig- Table 4
nificantly affected the level of SO2 emitted by affected plants. The Average annual percentage change in pollution emissions due to NAFTA.

parameter estimates reported in row one of the panel indicate that PM10 SO2
increased Mexican market access significantly reduced SO2 emis- D Mexico Tariff Preferences × Average Vs −1.69 b
−1.77c
sions. The estimate from my preferred specification in column (5) D US Tariff Preferences (All Imports) × Average Vs 0.12 −0.51
indicates that a one percentage point increase in Mexican tariff pref- D US Tariff Preferences (Intermediate Inputs) × Average Vs −0.36 −1.29a
erences for U.S. goods reduced emissions from a plant facing aver- Notes: Table reports the average annual changes in emissions due to NAFTA for a plant
age export frictions by 1.459%. Although imprecise, the estimates with average exposure to the Mexican market. Changes calculated using tariff prefer-
reported in row two suggest that increased import competition fol- ence changes from Table 2 and estimates from column (5) of each panel of Table 3. See
lowing NAFTA also decreased SO2 emissions; my preferred estimate text for details. Significance: a 1%; b 5%; c 10%.

indicates that for a plant facing average trade frictions, a one per-
centage point increase in Mexican tariff preferences for U.S. goods
reduced emissions by 3.871%. Finally, as the estimates reported in
average plant, NAFTA reduced PM10 emissions by 1.69% per year and
row three indicate, SO2 emissions also fell as a result of reductions
reduced SO2 emissions by 3.06% per year.
in the cost of accessing intermediate inputs from Mexico. The esti-
The estimates reported in Tables 3 and 4 also provide new evi-
mate from my preferred specification indicates that for a plant facing
dence on the micro-foundations underlying the aggregate relation-
average import frictions, a one percentage point increase in U.S. tar-
ship between international trade and environmental quality. These
iff preferences for Mexican intermediates reduced SO2 emissions by
micro-foundations are the subject of several recent studies, but data
13.185%.35
limitations have prevented the direct study of trade’s effects on pol-
To further quantify the magnitude of NAFTA’s effects on pollution
lution from individual producers. Instead, researchers have typically
emissions from affected manufacturing plants, I calculate the annual
inferred the effects of trade from changes in fuel use (e.g. Martin
percentage change in pollution emissions implied by my baseline
(2012), Forslid et al. (2014), Barrows and Ollivier (2014)) or ambi-
estimates given the actual change in tariff preferences. Specifically,
ent pollution concentrations (e.g. Gutierrez and Teshima (2011)). In
for each pollutant, I calculate the annual percentage change in emis-
addition, while trade can affect the activity of plants through mul-
sions implied by each estimated coefficient reported in column (5)
tiple channels, previous studies typically focus on a single margin
of Table 3 for a hypothetical plant facing average trade frictions and
of adjustment. Indeed, the few studies that utilize emissions data
the average change in U.S. and Mexican tariff preferences reported in
have examined the relationship between export status and pollution
Table 2. These changes are reported in Table 4.36
emissions (e.g. Holladay (2016), and Cui et al. (2016)). In contrast,
While the coefficient estimates presented in Table 3 suggest that
the estimates presented above provide evidence of how changes in
changes in the cost of importing intermediate inputs have a larger
foreign market access, import competition and access to imported
effect on pollution emissions for a given change in tariff preferences,
intermediate inputs affect the pollution emitted by individual man-
the numbers reported in Table 4 suggest that, in practice, increased
ufacturing plants.
access to the Mexican market following NAFTA had a larger effect
Finally, the results presented in Tables 3 and 4 add to the debate
on emissions. This is perhaps unsurprising given the nature of the
over the environmental effects of NAFTA. NAFTA’s possible effects
agreement; as Fig. 1 shows, the U.S. set tariffs at a much lower level
on the environment were a highly contentious policy issue prior to
than Mexico prior to NAFTA. As such, the biggest policy change came
the agreement’s signing particularly over concerns that emissions
from Mexico opening up its market to U.S. producers.
would fall as U.S. producers relocated dirty production to Mexico
Together, the estimates reported in Table 4 imply that NAFTA
to take advantage of weaker environmental regulation. Yet, to date
played an important role in the clean-up of U.S. manufacturing
there has been little empirical confirmation of how the agreement
that has occurred since the late 1980s. Recent research by Levinson
affected pollution emissions.37 My results indicate that emissions
(2009) indicates that between 1987 and 2001, pollution from U.S.
did fall, but largely for reasons not predicted prior to the agree-
manufacturing fell by 25% despite a 24% increase in output. In related
ment; plants primarily reduced emissions in response to new export
work, Levinson (2015) shows that aggregate emissions of PM10 and
opportunities.
SO2 fell by 3.55% and 3.61% annually between 1990 and 2008, with
Two additional points bear mention. First, while NAFTA is a com-
nearly all of these decreases driven by within-industry changes in
pelling setting for examining the effects of trade on the environment
pollution emissions. My estimates imply that the effects of NAFTA
because of the importance of environmental issues in debates over
account for nearly two-thirds of these reductions on average; for the
the agreement, it is worth emphasizing that the present exercise
examines the effects of an episode of North-South trade liberal-
ization on the pollution emitted by manufacturing plants in the
35
It is worth noting that the estimates presented in Panels A and B of Table 3 may North (the U.S.). As such, the results presented here likely apply to
potentially understate NAFTA’s effects due to plant-level political economy concerns. other Northern countries in similar settings, such as the agreements
While the timing of NAFTA negotiations means that the speed of tariff reductions
between India and Japan or Canada and Colombia, where there are
was determined before the agreement, affected plants may have responded to the
liberalization by lobbying state regulators to offset the potential effects of increased substantial developmental differences between the two trading part-
competitive pressure following NAFTA by relaxing the state-specific environmental ners. The results presented here may not generalize to the South
regulations faced by their industry. This would violate my identifying assumption. (Mexico) due to differences in many factors, such as the enforcement
However, it would increase pollution emissions from firms, meaning my estimates are of environmental policy. Similarly, episodes of North-North trade or
biased towards zero.
36
For example, for PM10 the estimated effect of increased access to the Mexican mar-
South-South trade may have different effects on pollution emissions
ket (b3 = −1.298) implies that for a plant facing the average level of trade frictions
(Vs = 1), the average increase in Mexican tariff preferences (6.51 percentage points)
reduced the level of pollution emitted by 1.69% annually, which is the value reported
in the table. Similarly, the estimated effect of reducing protection from Mexican com-
petition (b2 = 1.169) implies that the average increase in U.S. tariff preferences (0.50
percentage points) increased PM10 emissions by 0.12% annually for plants facing the
37
average level of trade frictions. Finally, the estimated effect of increased access to Those studies that have investigated NAFTA’s effects on the environment, such as
intermediate inputs from Mexico (b3 = −3.247) implies that the average increase Cole (2004) and Gamper-Rabindran (2006), typically examine changes in trade flows
in U.S. tariff preferences on intermediate inputs (0.56 percentage points), decreased from clean and dirty industries rather than examining changes in emissions levels
emissions by 0.36% annually for plants facing the average level of trade frictions. directly.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149 139

that are not captured in the present context.38 Second, the estimates Table 5
NAFTA and pollution levels: Alternative controls for the effects of the Clean Air Act.
presented above suggest that trade liberalization will always lower
plant pollution emissions. This is not necessarily the case; in the (1) (2) (3) (4) (5)
appendix, I present additional results that show NAFTA had little to Panel A: Particulate Matter (PM10 )
no effect on the emissions of volatile organic compounds from U.S. TjtMex × Vs −1.298b −1.552a −1.571a −1.479a −1.440b
manufacturing plants.39 (0.566) (0.539) (0.606) (0.562) (0.602)
TjtUS × Vs 1.169 0.386 0.670 0.996 1.297
(2.116) (2.081) (2.195) (2.106) (2.147)
5.2. NAFTA and the clean-up of U.S. manufacturing TjtInt × Vs −3.247 −4.122 −4.959 −3.150 −4.166
(3.944) (3.949) (3.979) (3.993) (3.901)
As indicated above, the estimates reported in Tables 3 and 4 Observations 4162 4162 4162 4162 4162
suggest that the effects of trade liberalization following NAFTA are Adjusted R2 0.20 0.22 0.28 0.21 0.26
responsible for a significant fraction of the clean-up of the U.S. man-
Panel B: Sulfur Dioxide (SO2 )
ufacturing sector during the 1990s. This finding contrasts with the
TjtMex × Vs −1.459c −1.414c −1.130 −1.458c −1.251
results presented by Shapiro and Walker (2016), who attribute the (0.746) (0.777) (0.753) (0.762) (0.784)
majority of the clean-up of U.S. manufacturing during this period to TjtUS × Vs −3.871 −2.976 −2.236 −3.894 −1.952
the effects of environmental regulation. While my baseline results (3.150) (3.222) (3.299) (3.276) (3.304)
include an indicator to control for the effects of environmental regu- TjtInt × Vs −13.185a −13.494a −13.893b −12.978b −13.949b
lations due to the Clean Air Act in any given year, it is possible that (4.996) (5.055) (5.470) (5.233) (5.461)
Observations 1645 1645 1645 1645 1645
the effects of such regulation are not immediate. This means that my 2
Adjusted R 0.25 0.25 0.26 0.25 0.26
baseline estimates may be potentially capturing both the effects of
trade liberalization due to NAFTA and the effects of ongoing changes Notes: Table reports the estimated effects of NAFTA on the emissions of PM10 and
SO2 from U.S. manufacturing plants. Each panel contains estimates for a different
to environmental regulations. I adopt two strategies to examine this
pollutant and reports regression coefficients from five separate regressions based on
possibility.40 Eq. (5) that include different controls for the effects of environmental regulation. In
First, I modify my baseline estimating equation and replace the all cases, the dependent variable is the log of pollution emissions. All specifications
indicator of a county’s non-attainment status under the Clean Air Act include plant, industry×year and state×year fixed effects, as well as controls for the
at time t with a set of fixed effects indicating the number of years Mexican Peso Crisis and CUSTFA. Column (1) of each panel includes an indicator the
non-attainment status of the county where the plant is located in each year. Column
each county was in non-attainment in the post-NAFTA period inter- (2) includes an indicator of the number of years the county was in non-attainment
acted with year fixed effects. This allows the effect of regulation to interacted with year fixed effects. Column (3) includes an indicator of the number of
vary across counties based on how long they have been regulated. As years the county was in non-attainment interacted with industry dummies and year
a further robustness check, I also estimate a specification that allows fixed effects. Column (4) includes an indicator of the year the county was first in non-
attainment interacted with year fixed effects. Column (5) includes an indicator of the
the effects of these interactions to vary across industries.
year the county was first in non-attainment interacted with industry dummies and
Second, I replace the indicator of a county’s non-attainment sta- year fixed effects. See text for further details. Standard errors clustered by industry
tus under the Clean Air Act at time t with an indicator of the first year and state are reported in parentheses. Significance: a 1%; b 5%; c 10%.
the county was in non-attainment status interacted with year fixed
effects. This approach also allows the effect of regulation to vary flex-
ibly across counties over time. As with my first strategy, I also allow The estimates reported in Table 5 provide further evidence that
these effects to vary by industry as a further robustness check. my estimates are not simply capturing the effects of environmen-
The results of these exercises are reported in Panels A and B of tal regulation. For the most part, replacing my baseline controls for
Table 5. For ease of comparison, column (1) reports my baseline the effects of the CAAAs with alternative measures that allow the
estimates previously reported in column (5) of the corresponding effects of regulation to vary over time has little effect on my baseline
panel in Table 3. Columns (2) and (3) report estimates from my results. For example, as the estimates reported in row one of Panel A
first strategy: column (2) includes an indicator of the number of show, allowing the effect of environmental regulations to vary flexi-
years the county was in non-attainment interacted with year fixed bly across plants on the basis of the number of non-attainment years
effects, while the specification in column (3) includes an indicator or on the basis of the first year the county was in non-attainment
of the number of years the county was in non-attainment inter- lead to modest increases in the magnitude of the estimated effect of
acted with industry dummies and year fixed effects. The last two increased Mexican market access on PM10 emissions. Similarly, the
columns of the table report estimates from my second strategy: estimates in row three of Panel B indicate that flexibly controlling for
column (4) includes an indicator of the year the county was first in the effects of environmental policy has little effect on the magnitude
non-attainment interacted with year fixed effects, and the specifica- of the estimated effect of increased access to imported intermediate
tion in column (5) includes an indicator of the year the county was inputs for plants that emit SO2 .
first in non-attainment interacted with industry dummies and year For plants that emit SO2 , adopting alternative measures of regu-
fixed effects. lation leads to modest reductions in the magnitude of the estimated
effect of increased access to the Mexican market for two out of four
specifications. As shown in row one of columns (2) and (4), allow-
38 ing the effect of environmental regulations to vary on the basis
It is worth noting, however, that the few existing studies examining the effects of
trade liberalization on plant pollution in different contexts, such as the examination of the number of years in non-attainment or on the basis of the
of trade liberalization in India by Martin (2012), find effects of a similar magnitude to first year in non-attainment has little effect on the estimated coef-
the results presented here. ficients. However, as shown in columns (3) and (5), the estimated
39
The fact that NAFTA had different effects on VOCs than PM10 and SO2 likely reflects
coefficients decrease in magnitude after I allow the effects of regu-
differences in how these pollutants are produced. Different industries produce differ-
ent quantities of each pollutant due to differences in production processes. However, lation presented in the preceding column to also vary by industry.
the available evidence indicates that sectors that emit substantial amounts of PM10 The estimates reported in columns (3) and (5) are just above typi-
also usually emit significant amounts of SO2 (Greenstone, 2002), and have similar pol- cal thresholds for statistical significance (with p-values of 0.13 and
lution elasticities of production (Shapiro and Walker, 2016), but both pollutants have 0.11, respectively) and, while smaller in magnitude, they are not
little overlap with sectors that produce significant amounts of VOCs. This suggests that
PM10 and SO2 are produced by similar production production processes, while VOCs
statistically different than my baseline estimates. More importantly,
are not. the estimates are still economically significant; they suggest that
40
I am indebted to an anonymous referee for suggesting these robustness checks. increased Mexican market access decreased SO2 emissions from
140 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
the average manufacturing plant by 1.37–1.52% per year, which is
still larger than the estimated annual effect of increased access to
imported intermediate inputs.
Altogether, the estimates presented in Table 5 indicate that my
baseline results are not simply capturing the effects of ongoing
changes to environmental policy. This provides further support for
my finding that trade liberalization following NAFTA played a mate-
rial role in the clean up of the U.S. manufacturing sector.
As indicated above, this result stands in contrast to the findings
of Shapiro and Walker (2016), who use a quantitive model to sug-
gest that the clean up can be attributed to increases in the stringency
of environmental regulation. There are two potential explanations
for this disparity. First, as my baseline estimates show, reductions in
the cost of importing intermediate inputs from Mexico had a signif-
icant effect on the pollution emitted by affected U.S. manufacturing
plants. As I discuss further below, such changes can affect emissions
by altering the mix of activities that occur at a manufacturing plant
as a result of offshoring. This mechanism is not present in the model
developed by Shapiro and Walker.41 Second, as I show below, the
decreases in PM10 and SO2 that I observe in response to increased
Mexican market access are consistent with trade-induced technolog-
ical upgrading. Recent research has suggested that increased foreign
market access can lead to emissions reductions by creating incen-
tives for firms to alter their technology choices (see Cherniwchan
et al. (2017) for an overview), but this mechanism is also absent
from the model employed by Shapiro and Walker. The absence of
these two mechanisms may potentially account for the different
findings as to the role of international trade in the clean-up of U.S.
manufacturing.
41
Indeed, Shapiro and Walker argue that offshoring is unlikely to play an important
role in the clean-up of U.S. manufacturing. To do so, they rely on a statistical decom-
position exercise in the vein of Levinson (2009), in which they break down changes in
U.S. emissions into changes due to changes in the scale, composition and techniques
of production. The results of this exercise suggest that changes in the composition of
final products produced by manufacturing plants explain little of the observed reduc-
tion in U.S. emissions. Instead, the reductions are due to within-product changes in
emission intensity. Shapiro and Walker interpret these results as evidence against the
offshoring of intermediate inputs; they claim that offshoring should lead to a shift in
the composition of production towards clean goods. However, this does not neces-
sarily rule out the effects of changes in intermediate input use. This can be seen by
extending the decomposition used by Shapiro and Walker to allow for within product
changes in emission intensity using an approach similar to that of Cherniwchan et al.
(2017). Cherniwchan et al. develop a firm-level decomposition that can be reinter-
preted as a decomposition of a single product line, for which production is composed
of a series of j tasks that are embodied in intermediate goods delivered during the pro-
duction process. Their approach yields the following (see Cherniwchan et al. (2017)
for details):
kdij
k∗ij one percentage-point increase in Mexican tariff preferences for U.S.
êi = hij êij + [hij − vij ]ŝij − I
[hij − vij ]k̂dij − [hij − vij ]k̂∗ij − vij l̂i
j j j
k ij j
kIij
where hij is the fraction of product i s emissions generated by task j, eij is the emission
intensity of task j, vij is the share of in-house production of task j in value added, s ij
is the share of task j in the total cost of tasks, kIij , kdij and k∗ij are the fractions of each
task that are done at the plant, outsourced to other domestic producers or offshored
to foreign producers, respectively, and l i is the product’s markup. This decomposition
highlights five potential channels that may drive changes in the emission intensity of
individual products. The first term is a technique effect that reflects changes in the
emission intensities of the tasks used in production. The second term is a composition
effect that reflects changes in the relative importance of various tasks in production.
The third and fourth terms reflect further changes in the mix of activities due to the
outsourcing of tasks to other domestic producers or to foreign firms. The final term
captures changes in emission intensity through the effects of changes in markups on
value-added. As the decomposition shows, changes in the set of in-house tasks, such
as those resulting from the offshoring of intermediate goods, may affect the emission
intensity of an individual product.
5.3. NAFTA and emission intensity
The estimates presented above suggest that NAFTA significantly
reduced the emissions of PM10 and SO2 from U.S. manufacturers.
I now examine the extent to which these reductions are due to
changes in the emission intensity of production (the emissions per
unit of output).
Examining emission intensity is a useful first step in understand-
ing what is driving the changes in pollution emissions that I observe.
By definition, trade liberalization can affect the level of pollution
emitted by a plant through two channels: by altering the physical
quantity of output produced (the scale of production), or by alter-
ing emission intensity through changes in either the mix of activities
done at the plant, or the production techniques used in those activ-
ities. The relative importance of changes in scale and intensity in
determining the effects of trade liberalization cannot be separately
identified through the examination of changes in pollution levels. To
address this, I exploit the fact that each of the estimated coefficients
reported in Table 3 can be decomposed into the change in emissions
due to a change in emission intensity and the change in emissions
due to a change in the scale of output.42 Both of these components
can be estimated directly using my empirical approach. Hence, to
distinguish between the two effects, I estimate the effects of NAFTA
on emission intensity using a specification similar to Eq. (5):
     
ln eijst = c1 tjtMex × Vs + c2 tjtUS × Vs + c3 tjtInt × Vs
+ X  g + qi + ljt + dst + eijst (6)
where eijst is a measure plant i s emission intensity and c1 , c2 and c3
are the coefficients of interest. The relative importance of changes in
the level of output and changes in production methods can be recov-
ered by comparing the magnitudes of the estimated coefficients from
Eq. (6) with the corresponding estimates from Eq. (5).
Estimates of Eq. (6) for PM10 and SO2 are reported in Table 6. For
ease of comparison, the first column for each pollutant reports the
corresponding baseline estimates reported in column (5) of Table 3.
Given that I do not observe physical output in my data, I use plant
employment and total sales to construct my measures of emission
intensity. Hence, in columns (2) and (5), the dependent variable is the
log of emissions per worker, while in columns (3) and (6), the depen-
dent variable is the log of the emissions-sales ratio. In all cases, the
estimated specifications include controls for the Peso Crisis, CAAA
and CUSFTA, plant, industry×year and state×year fixed effects and
reports standard errors that are clustered by industry×state.
The estimates reported in row one of the table indicate that the
effects of increased access to the Mexican market on PM10 and SO2
are primarily due to changes in the emission intensity of produc-
tion. For example, the coefficient estimate reported in row one of
column (2) indicates that, for a plant facing average trade frictions, a
exports decreased the level of PM10 emitted per worker by close to
1.250%. This effect is very similar to the baseline effect for pollution
levels reported in column (1), suggesting that the entire decease in
PM10 levels is due to changes in the emission intensity of production
rather than changes in the scale of production. Similarly, the coef-
ficient estimate reported in row one of column (5) indicates that a
one percentage-point increase in Mexican tariff preferences for U.S.
goods decreased the level of SO2 emissions per worker from a plant
42
To see this, let xijst denote output from plant i at time t. Then, by definition emis-
sions can be written as zijst = eijst × xijst , where eijst = zijst /xijst is the emission
intensity of production. Partially differentiating log emissions with respect to any vari-
able y yields: (∂ zijst /∂ y)/zijst = (∂ eijst /∂ y)/eijst +(∂ xijst /∂ y)/xijst , where (∂ zijst /∂ y)/zijst ,
(∂ eijst /∂ y)/eijst , and (∂ xijst /∂ y)/xijst are the percentage changes in emissions, pollution
intensity and output due to a change in y.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149 141

Table 6 changes in foreign market access and access to imported interme-

NAFTA and pollution intensity. diate inputs affect emission intensity, I now turn to explore three
Particulate Matter (PM10 ) Sulfur Dioxide (SO2 ) potential mechanisms: (i) entry and exit, (ii) changes in the mix of
imported intermediate inputs, and (iii) technological upgrading.
(1) (2) (3) (4) (5) (6)

tjtMex × Vs−1.298 b
−1.250 b
−1.216 b
−1.459 c
−1.386 c
−1.347c
(0.566) (0.555) (0.564) (0.746) (0.757) (0.764) 5.4.1. Entry and exit
tjtUS × Vs 1.169 1.800 2.152 −3.871 −3.581 −3.401 I begin by examining whether the reductions in average emission
(2.116) (2.133) (2.275) (3.150) (3.225) (3.111) intensity I observe result from extensive margin changes due to plant
tjtInt × Vs −3.247 −3.491 −4.083 −13.185a −13.054a −12.424b
entry and exit. There are two reasons to do so. First, recent empirical
(3.944) (3.863) (4.155) (4.996) (4.952) (4.902)
Observations 4162 4162 4162 1645 1645 1645 research, such as Bloom et al. (2010) and Shapiro and Walker (2016),
Adjusted R2 0.20 0.22 0.23 0.25 0.23 0.30 has documented an inverse relationship between plant productivity
and emission intensity. Incorporating this fact into a standard model
Notes: Table reports the estimated effects of NAFTA on the emissions and emission
intensity of PM10 and SO2 based on Eqs. (5) and (6) respectively. For each pollu- of international trade featuring firm heterogeneity (e.g. Melitz (2003)
tant, the dependent variable in the first column is the log of pollution emissions, the or Melitz and Ottaviano (2008)) yields a prediction that trade could
dependent variable in the second column is the log of emissions per worker, and the lead to lower average plant emission intensity as a result of entry
dependent variable in the third column is the log of the emissions-sales ratio. All spec- and exit, meaning the reductions I observe are simply a product of
ifications include plant, industry×year and state×year fixed effects, and controls for
the Mexican Peso Crisis, CUSTFA and the CAAAs. See text for details. Standard errors
reallocations of activity across plants. Second, the clean-up of U.S.
clustered by industry and state are reported in parentheses. Significance: a 1%; b 5%; manufacturing between 1990 and 2008 is often explained as prod-
c
10%. uct of changes in plant entry and exit as production shifts between
industries (Levinson, 2009, 2015), or across plants within an industry
(Shapiro and Walker, 2016).
facing average trade frictions by 1.386%. This effect is very close in To check whether my findings are being driven by entry and exit,
magnitude to corresponding estimate reported in column (4), which I re-estimate Eq. (6) using the sub-samples of plants that were in
means that the observed change in SO2 levels can also be attributed operation throughout my period of study. If my estimates are entirely
to changes in emissions intensity rather than changes in the scale of due to reallocations along the extensive margin, then NAFTA should
production. have no effect on plants that were in operation throughout my period
The estimates reported in Table 6 also suggest that emission of study.
intensity was also the primary channel through which decreases These estimates are reported in Table 7. The first four columns
in the cost of importing intermediate inputs from Mexico affected of the table report estimates for PM10 , while the last four columns
pollution emissions. For example, the estimate reported in col- of the table report estimates for SO2 . For each pollutant, the depen-
umn (5) indicates that for a plant facing average import frictions a dent variable in the first two columns is the log of emissions per
one percentage-point increase in U.S. tariff preferences for Mexican worker and the dependent variable in second two columns is the
exports of intermediate inputs decreased the level of SO2 emitted log of the emissions-sales ratio. In both cases, the even columns
per worker by approximately 13.054%. This number is very similar report the effects of NAFTA for the full sample previously reported
in magnitude to the effect on emission levels reported in column (4), in Table 6, and the odd columns report estimates from regressions
which suggests that the change in levels is primarily due to changes using the sub-samples of plants that were in operation throughout
in emission intensity. This is also true for PM10 ; although imprecise, my period of study. All specifications include controls for the Peso
the estimates reported in row three of columns (2) and (3) suggest Crisis, the CAAAs and CUSTFA, plant, industry×year, and state× year
that the effects of reductions in the costs of importing intermediate fixed effects. All standard errors are clustered by industry×state.
inputs are also driven by changes in emission intensity. The estimates reported in Table 7 indicate that restricting my
sample to plants that were in operation in continuously over the
5.4. Mechanisms period 1991–1998 increases the magnitudes of my estimates with-
out a substantial change in precision. For example, the estimates
Altogether, the estimates presented in Table 6 indicate that reported in row one of columns (1) and (2) indicate that for PM10 ,
NAFTA primarily affected pollution emissions through reductions restricting my sample increases the size of the estimated effect of
in the average emission intensity of plants, rather than changes a one percentage-point increase in Mexican tariff preferences from
in the scale of production. Given that little is known about how −1.250% to −1.521%. Similarly, the estimates reported in rows one

Table 7
NAFTA, pollution intensity and entry and exit.

Particulate Matter (PM10 ) Sulfur Dioxide (SO2 )

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

tjtMex × Vs −1.250 b
−1.521 b
−1.216 b
−1.481 c
−1.386 c
−1.671c
−1.347c
−1.617c
(0.555) (0.750) (0.564) (0.765) (0.757) (0.861) (0.764) (0.866)
tjtUS × Vs 1.800 2.278 2.152 2.613 −3.581 −2.041 −3.401 −1.935
(2.133) (2.264) (2.275) (2.440) (3.225) (3.044) (3.111) (2.940)
tjtInt × Vs −3.491 −4.440 −4.083 −5.047 −13.054a −16.452a −12.424b −15.605a
(3.863) (4.378) (4.155) (4.717) (4.952) (5.284) (4.902) (5.226)
Observations 4162 3799 4162 3799 1645 1517 1645 1517
Adjusted R2 0.22 0.25 0.23 0.26 0.23 0.24 0.30 0.31

Notes: Table reports the estimated effects of NAFTA on the emission intensity of PM10 and SO2 based on Eq. (6) . For each pollutant, the dependent variable in the first two columns
is the log of emissions per worker, while the dependent variable in the second two columns is the log of the emissions-sales ratio. For both pollutants, odd-number columns
report estimates for the full sample, while the even numbered columns report estimates for a subsample of plants that are in operation in all years. All specifications include plant,
industry×year and state×year fixed effects, and controls for the Mexican Peso Crisis, CUSTFA and the CAAAs. See text for details. Standard errors clustered by industry and state
are reported in parentheses. Significance: a 1%; b 5%; c 10%.
142 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

and three of columns (5) and (6) indicate that for SO2 , restricting Table 8
my sample increases the magnitude of the estimated effect of a one Intermediate input mix and the pollution haven hypothesis.

percentage-point increase in Mexican tariff preferences −1.386% to (1) (2) (3) (4)
−1.671%, and increases the magnitude of the estimated effect of a Panel A: Particulate Matter (PM10 )
one percentage-point increase in U.S. preferences for imported inter- tjtMex × Vs −1.250b −0.974c −1.216b −0.927c
mediate inputs from −13.054% to 16.452%. Together, these results (0.555) (0.525) (0.564) (0.531)
indicate that the effects I observe are primarily driven by intensive tjtUS × Vs 1.800 0.764 2.152 1.067
margin responses. In the context of the discussion presented above, (2.133) (2.168) (2.275) (2.333)
tjtInt × Vs −3.491 −4.083
these estimates suggest that NAFTA’s effect on the clean up of U.S.
(3.863) (4.155)
manufacturing was primarily due to within-plant responses rather Int−Dirty
tjt × Vs −5.833b −5.968b
than plant entry and exit. (2.337) (2.457)
tjtInt−Clean × Vs 2.857 2.234
5.4.2. Imported intermediate inputs and the pollution haven hypothesis (3.706) (4.006)
Next, I examine if the mix of imported intermediate inputs played Observations 4162 4162 4162 4162
a significant role in how NAFTA affected plant emission intensities in Adjusted R2 0.22 0.22 0.23 0.23

both of my samples. Changes in access to imported intermediates can

Panel B: Sulfur Dioxide (SO2 )
alter the mix of activities that occur within a manufacturing facility, tjtMex × Vs −1.386c −1.594b −1.347c −1.556b
potentially leading to a change in emission intensity. For example, (0.757) (0.750) (0.764) (0.749)
following trade liberalization, a plant’s emission intensity could fall tjtUS × Vs −3.581 −5.294 −3.401 −5.143
as relatively clean imported intermediate inputs are substituted for (3.225) (3.467) (3.111) (3.339)
tjtInt × Vs −13.054a −12.424b
relatively dirty intermediate inputs that are produced at the plant.
(4.952) (4.902)
Alternatively, the emission intensity of a plant could be reduced by Int−Dirty
tjt × Vs −14.671a −14.846a
fragmenting its production process on the basis of “cleanliness” and (4.888) (4.885)
sourcing dirty intermediate inputs from abroad instead of producing tjtInt−Clean × Vs −1.798 −1.222
them domestically.43 Given that I do not observe plant input choices (5.156) (5.084)
directly in my data, I examine if the changes in emission intensity Observations 1645 1645 1645 1645
documented above are driven by changes in access to relatively clean Adjusted R2 0.23 0.23 0.30 0.31
or relatively dirty intermediate inputs. Notes: Table reports the estimated effects of NAFTA on emission intensity for each
To do so, I modify Eq. (6) to include measures  of U.S. tariff of the pollutants under study. Each panel contains estimates for a different pollu-
preferences on relatively clean intermediate inputs tjtInt−Clean and tant and reports estimated coefficients from four separate regressions. In columns
  (1) and (2), the dependent variable is the log of pollution emissions per worker. In
Int−Dirty
relatively dirty intermediate inputs tjt interacted with the columns (3) and (4) the dependent variable is the log of the emissions-sales ratio. All
specifications include plant, industry×year and state×year fixed effects. See text for
normalized Head-Ries Index, Vs .44 This specification is given by:
details. Standard errors clustered by industry and state are reported in parentheses.
      Significance: a 1%; b 5%; c 10%.
Int−Dirty
ln eijst = a1 tjtMex × Vs + a2 tjtUS × Vs + a3 tjt × Vs
 
+ a4 tjtInt−Clean × Vs + X  g + qi + ljt + dst + eijst (7) and began sourcing clean intermediate inputs from Mexico, making
the mix of activities done at the plant relatively dirty. If â4 is nega-
tive, it means that the emission intensity of affected plants is falling
The coefficients a 3 and a 4 capture the changes in emission intensity
due to increased access to relatively clean intermediate inputs. This
due to changes in access to relatively dirty and clean intermediate
suggests that affected plants are importing relatively clean inputs
inputs respectively.
to utilize at the plant, making the mix of activities relatively less
The estimated coefficients â3 and â4 provide insight into how
pollution intensive.
the mix of imported intermediate inputs affected emission inten-
Estimates of Eq. (7) are reported in Table 8. The top panel reports
sity for each pollutant. If â3 is positive, it means that the emission
estimate for PM10 , while the bottom panel reports estimates for
intensity of affected plants is rising due to increased access to rel-
SO2 . In the first two columns of each panel, the dependent vari-
atively dirty intermediate inputs. This suggests that affected plants
able is the natural log of emissions per worker. In the last two
are importing relatively dirty inputs to utilize at the plant, making
columns of each panel, the dependent variable is the natural log of
the mix of activities relatively pollution intensive. Conversely, if â3
the emissions-sales ratio. For ease of comparison, columns (1) and (3)
is negative, it means increased access to relatively dirty intermediate
of each panel report estimates of Eq. (6). Each specification reported
inputs reduced emission intensity. This suggests that affected plants
in Table 8 includes controls for the Peso Crisis, CAAA, and CUSTFA,
fragmented production and began sourcing dirty intermediate inputs
plant, industry×year and state×year fixed effects. In all cases, the
from Mexico instead of producing them domestically; such fragmen-
standard errors are clustered by industry×state.
tation would make the mix of activities done at the plant relatively
The estimates reported in Panel A of the table confirm that
clean. If â4 is positive, it means that the emission intensity of affected
increased access to imported intermediate inputs led to signifi-
plants is rising due to increased access to relatively clean intermedi-
cant reductions in PM10 through changes in emission intensity. The
ate inputs. This suggests that affected plants fragmented production
results from Sections 5.1 and 5.3 were suggestive of this channel,
but the reported coefficients were imprecise. The estimates in Panel
A indicate that this is because emission intensity was only affected
43
There is some evidence that fragmentation is an important mechanism through
by a change in access to dirty intermediate inputs. This can be seen
which trade affects the environment. For example, Dean and Lovely (2010) show that
the fragmentation of production helps explain reductions in the pollution intensity of clearly for columns (1) and (2), in which the dependent variable is
China’s trade between 1995 and 2004. the log of emissions per worker. The baseline estimate reported in
44
I construct these preference measures by labeling industries as clean or dirty on row three of column (1) indicates that for a plant facing average trade
the basis of their emission intensity. For each pollutant, I label an industry as relatively frictions, a 1% increase in the preference for intermediate inputs from
dirty if its emission intensity is in the top 10% of the emission intensity distribution,
and relatively clean otherwise. I then construct input tariffs for both groups of indus-
Mexico decreased emissions per worker by approximately 3.491%. In
Int−Dirty
tries, and use these tariffs to construct tjtInt−Clean and tjt . For further details on the contrast, the estimate reported in row four of column (2) indicates
construction of these measures, see the appendix. that a 1% increase in the preference for relatively dirty intermediate
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
inputs from Mexico decreased emissions per worker by approxi-
mately 5.833%. The coefficients reported in columns (3) and (4) paint
a similar picture. Hence, it appears that emission intensity fell as
plants fragmented production and began sourcing dirty intermedi-
ate inputs from Mexico instead of producing them domestically. This
type of fragmentation is consistent with the available evidence; pre-
vious research (e.g. Hummels et al. (2001), Yi (2003)) suggests that
vertical specialization was becoming more common in both the U.S.
and Mexico prior to NAFTA, meaning that producers were more likely
to fragment production and import intermediate inputs from abroad
around this time. The estimates reported in the panel also suggest
that PM10 emission intensity may have been affected by imports of
relatively clean inputs, but these effects are imprecise.
The estimates reported in Panel B indicate that increased access
to dirty intermediate inputs also led to significant reductions in
emission intensity for plants that emit SO2 . For example, the esti-
mate reported in the fourth row of column (2) indicates that, for a
plant facing average import frictions, a 1% increase in tariff prefer-
ences for relatively dirty intermediate inputs from Mexico decreased
emissions per worker by approximately 14.671%. This suggests that
affected plants fragmented production and began sourcing dirty
intermediate inputs from Mexico instead of producing them domes-
tically. The estimates reported in the fifth row of the panel suggest
that plants may have also sourced some clean intermediate inputs
from Mexico, but this effect is relatively small and imprecisely
estimated.
The finding of reductions in plant emission intensities in response
to changes in access to relatively dirty intermediate inputs provides
new evidence of how international trade affects the environment.
It suggests that individual producers may have shifted production
from a country with relatively stringent environmental regulations
to a country with relatively weak environmental regulations, a phe-
nomenon known as the pollution haven hypothesis.45 The pollution
haven hypothesis is a highly contentious prediction that has its ori-
gins in the debate over NAFTA’s effects on the environment; yet
there is scant evidence suggesting pollution havens exist.46 The
estimates presented above contribute to this body of research by
providing the first plant-level evidence that is consistent with a real-
location of activity across borders following trade liberalization due
to pollution-haven type motives.47
5.4.3. Technology upgrading
Finally, I examine whether the decreases in PM10 and SO2 inten-
sity that I observe in response to increased Mexican market access
are consistent with trade-induced changes in process innovation
or technological upgrading. While little is presently known about
the relationship between foreign market access and plant emission
intensity, there are two reasons to focus on technology. First, previ-
ous research has suggested that increased foreign market access can
increase firm productivity by creating incentives for firms to engage
in process innovation or adopt new technologies (e.g. Costantini
and Melitz (2008), Atkeson and Burstein (2010), Bustos (2011)).
Given the empirical evidence documenting a negative relationship
between productivity and emission intensity, trade-induced innova-
tion or technology adoption could lower plant emission intensity.
45
For an overview of the pollution haven hypothesis, see Taylor (2004).
46
Prior to the agreement, it was often predicted that trade liberalization following
NAFTA would lead to the relocation of dirty production from the U.S. to Mexico as
a result of differences in the stringency of environmental regulation across the two
countries.
47
There is reason to believe that cross-country differences in environmental reg-
ulations could have affected the distribution of activities in the U.S. and Mexico;
differences in environmental policy were an important determinant of trade flows
between the U.S. and Mexico prior to NAFTA (Levinson and Taylor, 2008).
143
Second, a burgeoning literature has adapted the theoretical frame-
work developed in Bustos (2011) to suggest that increased foreign
market access can reduce emission intensity directly by creating
incentives for firms to alter their pollution abatement technologies
(e.g. Cui et al. (2012), Forslid et al. (2014)).
Testing this channel is challenging as I do not observe the technol-
ogy choices or productivity of plants in my data.48 As such, I exploit
a theoretical prediction from the Bustos (2011) model of trade and
technological upgrading as it has been used to formalize the effects of
trade on both forms of technology adoption. Her framework suggests
that increased foreign market access will only affect innovation and
technology adoption by a subset of plants within an industry. This
adoption occurs because some plants enter the foreign market fol-
lowing trade liberalization; the resulting increase in revenues allows
the plant to cover the fixed costs of investing in a new technology.
Given that export market participation depends largely on produc-
tivity (e.g. Bernard and Jensen (1999)), this suggests that the effects
of increased foreign market access on technology adoption will vary
across plants based on their initial productivity.
To test this prediction, I follow Bustos by proxying for initial pro-
ductivity using initial plant size, and then estimating whether the
effects of increased foreign market access on plant emission inten-
sity vary across quartiles of the initial plant size distribution.49 That
is, I estimate the following based on Eq. (6):
   
4
ln eijst = f1,q tjtMex × Vs × Q q + f2 tjtUS × Vs
q=1
 
+ f3 tjtInt × Vs + X  g + qi + l jt + dst + eijst (8)
where Q q is an indicator variable that takes the value of one if plant
i is in quartile q of the initial size distribution.
The estimated coefficients f̂1,1 to f̂1,4 provide insight as to which
plants are altering their emission intensities in response to increased
foreign market access following NAFTA. For example, if the estimate
f̂1,3 is negative, it means that plants in the third quartile of the initial
size distribution are reducing their emission intensities in response
to increased foreign market access. Conversely, if f̂1,3 is positive,
plants in the third quartile of the initial size distribution are becom-
ing relatively emission intensive in response to increased foreign
market access. I am primarily interested in the pattern the estimates
display across quartiles. If the reductions in emission intensity are
consistent with trade-induced technological upgrading or process
innovation as predicted by the Bustos (2011) model, then the effects
should be concentrated at plants in the upper quartiles of the ini-
tial size distribution; these plants are the most likely to enter the
export market in response to trade liberalization. If the reductions
are due to some other mechanism, the estimates should exhibit no
clear pattern.
Estimates of Eq. (8) are presented in Table 9. The first two columns
report estimates for PM10 , while the last two columns report esti-
mates for SO2 . For each pollutant, the dependent variable in the first
48
Unfortunately, the NETS data does not include information on plant capital stocks
or material use, making it impossible to estimate productivity estimation using
standard methods.
49
Because technology adoption is driven by export decisions in the framework dis-
cussed above, ideally I would also examine whether entry into the export market
varies across plants on the basis on their initial productivity. Unfortunately, I am not
able to test this channel due to data limitations; the NETS database does not include
information on plant export status during my period of study. However, Mexican
demand for U.S. exports grew substantially following NAFTA; Hillberry and McDaniel
(2002) report that U.S. exports to Mexico grew by 93% over the 1993–2001 period.
The work of Romalis (2007) and Caliendo and Parro (2015) also indicates that NAFTA
increased U.S. manufacturing exports to Mexico.
144 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

Table 9 My results suggest that NAFTA played an important role in the

NAFTA and pollution intensity by quartile of the initial plant size distribution. observed clean-up of the U.S. manufacturing sector. I find that trade
Particulate Matter (PM10 ) Sulfur Dioxide (SO2 ) liberalization following NAFTA decreased emissions of particulate
matter and sulfur dioxide from affected plants; emission of these pol-
(1) (2) (3) (4)
lutants fell in response to increased access to the Mexican market
tjtMex × V s × Q4 −0.713 −0.672 −1.517 b
−1.502c and increased access to imported intermediate inputs from Mexico.
(0.756) (0.750) (0.768) (0.770)
I also find that increased import competition had little effect on the
tjtMex × Vs × Q 3 −1.266c −1.241c −1.800c −1.736c
(0.681) (0.671) (0.951) (0.961)
emissions of any pollutant. Altogether, my estimates suggest that, on
tjtMex × Vs × Q 2 −1.541a −1.509b −0.095 0.049 average, nearly two-thirds of the aggregate reductions in the emis-
(0.581) (0.605) (1.483) (1.497) sions of both pollutants from U.S. manufacturing in the first five years
tjtMex × Vs × Q1 −0.681 −0.583 0.008 0.038 following NAFTA can be attributed to the effects of the agreement.
(0.625) (0.637) (1.183) (1.182) Given the current scarcity of research examining how plants alter
tjtUS × Vs 2.048 2.434 −4.137 −3.970
their emissions in response to trade, my analysis also provides evi-
(1.963) (2.075) (3.434) (3.332)
tjtInt × Vs −3.348 −4.020 −14.474a −13.734b
dence as to the mechanisms driving the emissions reductions that
(3.473) (3.707) (5.471) (5.359) I observe. To start, I show that the decreases in emission levels are
Observations 4162 4162 1645 1645 primarily driven by changes in the emission intensity of production
Adjusted R2 0.22 0.23 0.24 0.31 rather than changes in the level of output produced. I then inves-
Notes: Table reports the estimated effects of NAFTA on emission intensity by quar- tigate three potential channels that explain the change in emission
tile of the initial plant size distribution. For each pollutant, the dependent variable intensity. I find that the reductions in emission intensity are due
in the first column is the log of emissions per worker, while the dependent vari- to within-plant changes rather than plant entry and exit, suggest-
able is second column is the log of the emissions-sales ratio. All specifications include ing that much of NAFTA’s effects occurred through the intensive
plant, industry×year and state×year fixed effects. See text for details. Standard errors
clustered by industry and state are reported in parentheses. Significance: a 1%; b 5%;
margin rather than the extensive margin. I also find that the inter-
c
10%. mediate input induced reductions in emission intensity are driven by
changes in access to relatively dirty intermediate inputs, suggesting
that U.S. manufacturing plants altered their emission intensities by
column is the log of emissions per worker, while the dependent vari-
fragmenting production in response to differences in environmen-
able is second column is the log of the emissions-sales ratio. All
tal regulation across the U.S. and Mexico. Finally I provide evidence
specifications include controls for the Mexican Peso Crisis, CUSTFA
consistent with the hypothesis that changes in foreign market access
and the CAAAs, as well as plant, industry×year and state×year fixed
reduced emission intensity by creating incentives for plants to adopt
effects. Standard errors clustered by industry×state are reported in
new technologies. Altogether, these findings suggest that within-
parentheses.
plant responses are an important determinant of how international
The coefficients reported in Table 9 indicate that there is substan-
trade affects the environment.
tial heterogeneity in the effects of increased Mexican market access.
I find that the effects are concentrated around the upper middle of
the initial plant size distribution; for PM10 the effects of increased Appendix A. Data: Additional details
foreign market access are driven by plants in the second and third
quartiles, while for SO2 the effects of increased market access are This appendix describes additional details of the data sources and
due to plants in the third and fourth quartiles. These responses are of procedure used to construct and clean the datasets used in the
consistent with the empirical evidence presented in Bustos (2011), analysis. It also describes additional characteristics of the datasets
who finds that trade liberalization primarily induces technological used in the analysis.
and process innovation upgrading at plants in the upper quartiles
of the size distribution. As such, these estimates provide sugges- A.1. Main data sources and data construction
tive evidence that the reductions in emission intensity are due to
trade-induced technological change following NAFTA. A.1.1. Environmental data
The primary data source for this study is the Toxic Release Inven-
tory (TRI) maintained by the U.S. Environmental Protection Agency
6. Conclusion (EPA). The TRI is a publicly available database that includes detailed
information on the emissions of various toxic chemicals to air, water,
This paper contributes to a long-standing debate over the envi-
ground and other media from plants that are required to report their
ronmental consequences of trade liberalization. To date, research has
releases under the Emergence Planning and Community Right to
primarily focused on the relationship between trade and aggregate
Know Act (EPCRA) of 1986.50 Under the EPCRA, plants are required
pollution levels. While these studies find that trade is not necessar-
to report their emissions to the EPA if they: (i) employ at least ten
ily bad for the environment, they often appeal to the unobserved
employees, (ii) operate in a TRI-covered industry, and (iii) manufac-
responses of individual polluters to explain the mechanisms under-
ture, process or otherwise use a regulated substance in excess of a
lying their findings. Yet, there has been little evidence of how trade
designated threshold in a calendar year. While the TRI database con-
liberalization affects the pollution emitted by individual manufactur-
tains detailed information on pollution emissions, several steps must
ing plants.
be taken before the data can be utilized.
In this paper, I provide the first empirical evidence of these
I begin by creating a dataset of plant-level pollution emissions
responses. My analysis examines how one of the most politically
by linking the annual datasets available from the Toxic Release
contentious episodes of trade liberalization in U.S. history, the North
American Free Trade Agreement (NAFTA), affected pollution emis-
sions from U.S. manufacturing plants. In my analysis I rely on two
50
unique longitudinal datasets containing detailed information on the It is important to note that while these emissions are self-reported, facilities are
emissions of either particulate matter, and sulfur dioxide matched penalized if they fail to report or misreport their emissions, but do not face penal-
ties for the amount of emissions that they report. For example, the EPA fined Valimet
with other plant characteristics. I identify NAFTA’s effects using an Incorporated of Stockton California $193,996 for failing to report releases between
empirical approach that exploits variation in trade frictions across 2001 and 2005. The existence of these penalties creates an incentive for plants to
time, industries and states. accurately report their emissions.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
Table A1
Manufacturing plant characteristics.
1992 Census of manufacturers
This study All plants Large plants
Pollutant Mean Std. dev. Mean Std. dev. Mean Std. dev.
PM10
Employees 587 (1247) 142 (154) 225 (176)
Sales ($1,000,000s) 78 (147) 58 (110) 83 (128)
SO2
Employees 726 (1519) 192 (200) 275 (214)
Sales ($1,000,000s) 165 (637) 84 (130) 117 (148)
Notes: Table compares the characteristics of manufacturing plants used in this study
with plant characteristics from the 1992 Census of Manufacturers. Each panel of the
table corresponds to a different dataset. The first two columns of each panel report the
mean and standard deviation of plant sales and employment for the plants examined
in the analysis, while the remaining columns report the corresponding statistics from
the Census of Manufacturers for the set of industries covered by each dataset. The
category “large plants” contains plants with 20 or more employees.
Inventory. In its raw form, this dataset contains information on the
emissions of every chemical reported by each plant in every year it is
required to report. However, there have been numerous changes in
the list of covered chemicals since the TRI’s inception.51 This means
that changes in the total level of emissions reported by any one
plant may be due to changes in the list of chemicals that the plant
is required to report rather than a change in economic activity.52 To
address this, I exclude any chemicals that are not listed on the TRI in
each year of my sample period.
For each plant, I also exclude any chemicals that are reported
under the Alternate Threshold for Facilities With Low Annual
Reportable Amounts in any year. This threshold was enacted
on November 30, 1994, and established an alternative reporting
requirement for chemicals with annual releases of less than five hun-
dred pounds (under this threshold, plants are required to report if
they manufacture, process or otherwise use a regulated substance
in excess of one million pounds per year). In practice, this means
some plants were no longer required to report their emissions of
some chemicals in the middle of my period of study. Excluding these
chemicals for each plant ensures that I am not capturing decreases in
emissions due to the change in reporting threshold.
Lastly, I exclude any chemicals for which releases are reported
using range codes in any year. In the TRI reporting forms, plants can
use range codes to document releases of less than 1000 lb. In these
cases, emissions are reported using a range (either 1–10, 11–499 or
500–1000 lb) and then recorded in the TRI data using the midpoint
of the range. However, most plants do not use range codes, making
it impossible to use standard approaches for dealing with censoring.
As such, I drop these chemicals from the data.
After these restrictions, I am left with a dataset that contains
information on the annual emissions of various chemicals by plant.
These substances come from a variety of sources and production
methods, meaning the toxic chemicals produced at each plant may
not share similar characteristics. Given my research design, this cre-
ates the possibility that simply comparing changes in total emissions
across plants would reflect differences in the particular substances
51
These changes fall into one of three categories: (i) expansion in industry cover-
age to include sectors outside of manufacturing, (ii) revisions to the thresholds for
reporting, and (ii) modifications to the list of chemicals that facilities are required to
report.
52
For example, in 1990, facilities were required to report their emissions if they
manufacture or use of nearly 300 chemicals exceeded the reporting thresholds. By
1996, the list of required compounds had grown to include 588 chemicals. Hence, for
a plant that reports in both years, an increase in total emissions may be driven by
an increase in the quantities of emissions of base year pollutants, an increase in the
number of chemicals the facility is required to report, or changes in both factors.
145
produced at each plant as opposed to meaningful differences due
to changes in competitive pressures. To deal with this, I classify
these chemicals as particulate matter (PM10 ), and sulfur dioxide
(SO2 ) using the concordance used by the EPA to classify TRI chem-
icals into the chemical groups reported in the National Emission
Inventory (Environmental Protection Agency, 2013) and create a
separate dataset for both pollutants. I then restrict the data fur-
ther to only include plants with positive emissions to air, land, and
water.53,54 Moreover, given the focus of the NAFTA debate on pol-
lution from manufacturing, I only consider plants from this sector
(4-digit Standard Industrial Classification (SIC) categories 2000-
3999). Lastly, I exclude any unrealistically large changes in the data
by dropping observations for which the annual change in emissions
growth, plant sales growth or plant employment growth is in the top
1% or bottom 1% of values.
A.1.2. Plant characteristics
I match the emissions data for each pollutant to proprietary data
on plant characteristics from the National Establishment Time Series
(NETS). The NETS database is a longitudinal plant database compiled
by Walls and Associates from Dun & Bradstreet’s DUNS Marketing
Information archives.55 The NETS data contain information on plant
location, sales and employment.56 I match the NETS data to the
pollution data using the facility identification numbers reported in
the TRI using a correspondence created by Walls and Associates. I
deflate plant sales using the industry deflators from the NBER-CES
manufacturing database (Bartelsman and Gray, 1996).
A.1.3. International trade data
I supplement the data from the TRI and NETS with tariff data for
the U.S. from Feenstra et al. (2002) and for Mexico from Romalis
(2007). In both cases, the tariff data is classified using the Harmo-
nized System (HS) classification system while the matched TRI and
NETS data is classified according to the four digit Standard Industrial
Classification (SIC4). I aggregate the tariff data to the SIC4 classifi-
cation using the concordance from Pierce and Schott (2012) using
the trade data from Schott (2008) and the UN Comtrade database
as weights. This procedure yields import weighted average out-
put tariffs for the U.S. and Mexico. As described in the main text,
I follow the approach of Amiti and Konings (2007) and construct
intermediate input tariffs by industry using the constructed U.S.
output tariffs and the BEA input-output table from 1992. Finally,
I match the data from the TRI and NETS to data on state exports
and imports for 1993 from Feenstra (1997) and the North American
Transborder Freight Data database maintained by the U.S. Bureau of
Transportation Statistics.57 I use this data to construct the Head-Ries
Index.
53
The TRI also includes information on releases to the environment that occur at
off-site treatment facilities. I exclude these emissions, because they may include sub-
stances that were produced in previous years and stored onsite before being shipped
to an off-site treatment facility.
54
It is possible for plants to meet the reporting requirements and produce zero emis-
sions, however missing values are recorded as zeroes in the raw TRI data, making it
impossible to distinguish zero emissions from missing values. Hence, I am forced to
drop all plants with zero reported emissions from each dataset.
55
For a detailed discussion of the NETS database, see the appendix of Neumark et al.
(2011).
56
The NETS data also includes information on other plant characteristics, such as
foreign ownership and export status, but these characteristics are only reported for
the last year each plant is observed in the NETS data, which is typically well after of
my period of study.
57
I rely on trade data from 1993 for two reasons. First and foremost, 1993 is the
earliest year for which state import data is available from North American Transbor-
der Freight Data database. Data on state exports is also available for 1991 and 1992,
however as Feenstra (1997) documents, the method used by the U.S. Census Bureau
to determine the state of origin for exports during these years does not necessarily
reflect the true location of production, but rather the location of wholesalers and other
intermediaries, making the export data from 1993 better suited for my purposes.
146 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

A.2. Additional summary statistics tariffs are weighted averages constructed from data on output tar-
iffs and data on cost shares from the BEA’s benchmark input-output
Table A1 compares the characteristics of plants in each data set table for 1992. Here the averages are restricted to sets of industries
with plant characteristics from the 1992 Census of Manufacturers. that are relatively dirty or clean. I calculate these tariffs for imports
Each panel of the table displays statistics for a different dataset. The from Mexico as well as for imports that fall under the Generalized
first two columns of each panel report the average sales and employ- System of Preferences. Finally, I construct measures of the tariff pref-
ment statistics reported previously in Table 1, while the last four erences for relatively clean and dirty intermediates by industry and
columns report corresponding statistics constructed using data from year that are analogous to Eq. (2).
the Census of Manufacturers for the set of industries covered by each
dataset. Columns three and four of each panel report statistics for all
plants, while the last two columns report statistics for large plants Appendix B. Additional results
with twenty or more employees.
The information presented in Table A1 indicates that the manu- This appendix presents additional results references in the main
facturing plants in each of my samples are relatively large on average. text. The first subsection presents results of additional robustness
For example, the statistics reported in the first panel of the Table exercises, while the second presents additional estimates on the
indicate that the average PM10 emitting plant in my data has much effects of NAFTA on emissions of volatile organic compounds.
higher sales and employment levels than the average plant from sim-
ilar industries in the Census of Manufacturers. In fact, these statistics B.1. Additional robustness checks
suggest that the average PM10 emitting plant is even larger in size
than the average large plant. The statistics for SO2 exhibits a sim- The estimates presented in Table 3 indicate that my main find-
ilar pattern. This is likely due to the reporting requirements for ings are not simply capturing the effects of other major shocks that
the TRI. As mentioned above, plants are only required to report to occurred during my period of study. Here, I examine the possibility
the TRI if they employ at least ten employees, meaning that small that I am capturing anticipatory responses by plants, or differences
manufacturing plants are systematically excluded from my data. in pre-existing trends.
To start, I examine the possibility of anticipatory responses by
A.3. Additional variable construction plants prior to NAFTA. In my analysis, I treat NAFTA as an exogenous
shock that was unforeseen by U.S. manufacturers. This approach
A.3.1. Measures of trade liberalization due to CUSTFA exploits the inherent uncertainty surrounding the passage of the
In my analysis, to control for the ongoing effects of trade liber- agreement and the fact that it was ratified just over one month
alization due to the Canada-U.S. Free Trade Agreement (CUSTFA), I before it came into effect, leaving producers little time to respond to
incorporate measures of the bilateral tariff preferences for Canada any policy changes. However, some manufacturers may have viewed
and the U.S. (tjtCan−US and tjtUS−Can respectively) interacted with a mea- progress on NAFTA’s side agreements following President Clinton’s
sure of bilateral trade frictions between each state s and Canada inauguration at the beginning of 1993 as a signal that a deal was
(VCan imminent, leading them to respond to the expected changes in policy
s ). Bilateral tariff preferences between Canada and the U.S. are
given by: before they came into effect. If this is the case, my baseline estimates
will understate the true effects of NAFTA. To check for this, I re-
estimate my baseline version of Eq. (5) for PM10 and SO2 , but exclude
tjtCan−US = tjtCan−ROW − tjtCan−US and tjtUS−Can = tjtUS−ROW − tjtUS−Can observations from 1993.
These estimates are reported in the second column of each panel
where the tjtc−d are country c s ad-valorem tariff rates on imports in in Table B1. For ease of comparison, the first column of each panel
industry j from region d at time t. VCan is analogous to Vs from Eq. (4) in the table displays the estimated coefficients from the baseline
s
in the main text, and is given by: specification reported in column (5) of the corresponding panel
from Table 3. Both specifications include plant, industry×year and
⎡ Can s ⎤  ⎡

 Can s ⎤
state×year fixed effects, as well as controls for the Peso Crisis,
⎣
 Xs XCan
 ⎦ ⎣ 1
  Xs XCan ⎦ CUSTFA and the CAAAs.
VCan
s ≡
Xss XCan
Can s Xss XCan
Can
s∈S
Table B1
NAFTA and pollution levels: Additional results.
s
where XsCan are exports from state s to Canada, XCan are state s imports
Particulate Matter (PM10 ) Sulfur Dioxide (SO2 )
from Canada, Xss is production from state s that is consumed in
Can
state, and XCan denotes the value of Canadian production consumed (1) (2) (3) (4) (5) (6)
domestically. Like Vs , I construct VCan
s using data from 1993.
tjtMex × Vs −1.298b −1.521b −2.137a −1.459c −1.474c −1.279b
(0.566) (0.665) (0.647) (0.746) (0.855) (0.556)
A.3.2. Measures of tariff preferences for clean and dirty intermediate tjtUS × Vs 1.169 1.774 0.002 −3.871 −3.745 −2.050
inputs (2.116) (2.343) (0.765) (3.150) (3.678) (1.649)
tjtInt × Vs −3.247 −4.808 −1.499 −13.185a −12.648b −5.609c
Constructing measure of tariff preferences for clean and dirty
(3.944) (4.256) (2.191) (4.996) (5.772) (2.992)
intermediate inputs involves three steps. First, I construct a measure Observations 4162 3728 4162 1645 1472 1645
of emission intensity by industry using data from the EPA’s National Adjusted R2 0.20 0.21 0.20 0.25 0.26 0.24
Emission Inventory (NEI) for 1990 and the NBER-CES Manufactur-
Notes: Table reports the estimated effects of NAFTA on emissions of PM10 and SO2 . In
ing Database. I use data from 1990 to construct these measures as all cases, the dependent variable is the log of pollution emissions. Columns (1) and
it is the only year prior to NAFTA for which data from the NEI is (3) report estimates from the baseline specification previously reported in Table 3.
available. Using this data, I label an industry as relatively dirty if its In columns (2) and (4), observations from 1993 are dropped from the analysis. In
emission intensity is in the top 10% of the emission intensity dis- columns (3) and (6), the observed policy change is replaced with the interaction of
the maximum tariff preference and a post-NAFTA indicator. All specifications include
tribution and relatively clean otherwise. Second, for each industry, I plant, industry×year and state×year fixed effects, as well as controls for the Mexican
calculate the input tariffs for intermediate inputs sourced from rel- Peso Crisis, CUSTFA and the CAAAs. See text for details. Standard errors clustered by
atively clean and relatively dirty industries. As above, these input industry and state are reported in parentheses. Significance: a 1%; b 5%; c 10%.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149 147

Table B2 when NAFTA was ratified. There is some evidence that U.S man-
NAFTA and pollution levels: Flexible estimates. ufacturers reacted to anticipated changes in trade policy follow-
Particulate Matter (PM10 ) Sulfur Dioxide (SO2 ) ing NAFTA; McLaren and Hakobyan (2016) show that anticipatory
adjustments post-NAFTA had large effects on labor markets in the
(1) (2) (3) (4)
U.S. This also creates the possibility that my baseline results under-
t̄jMex × Vs × 1992 0.302 0.285 −1.083 −1.109 state NAFTA’s true effects. To check for this, I replace the treat-
(1.044) (1.040) (1.161) (1.140)
ment variables from my baseline specification with the anticipated
t̄jMex × Vs × 1993 −1.528 −1.615 −2.220c −2.255c
(1.139) (1.154) (1.192) (1.176)
tariff preference in 1993, interacted with the normalized  HRI and
t̄jMex × Vs × 1994 −2.450c −2.496c −2.337 −2.310 a post-NAFTA dummy. Formally, in Eq. (5), I replace tjtMex × Vs
     
(1.318) (1.337) (1.504) (1.478)
with t̄jMex × Vs × Postt , tjtUS × Vs with t̄jUS × Vs × Postt , and
t̄jMex × Vs × 1995 −2.415c −3.857b −4.522a −5.239a    
(1.390) (1.866) (1.549) (1.820) tjtInt × Vs with t̄jInt × Vs × Postt , where t̄jMex , t̄jUS , and t̄jInt are the
t̄jMex × Vs × 1996 −1.925 −1.835 −4.746a −4.685a
(1.351) (1.371) (1.651) (1.630)
level of tariffs for U.S. exports to Mexico, Mexican exports of final
t̄jMex × Vs × 1997 −3.224b −3.065b −4.643a −4.583a goods to the U.S., and Mexican exports of intermediate goods in
(1.421) (1.459) (1.632) (1.609) 1993, and Postt is an indicator that is equal to one during the post-
t̄jMex × Vs × 1998 −2.304 −2.165 −5.073a −5.020a NAFTA period. Given that tariffs decrease to zero following NAFTA,
(1.410) (1.448) (1.568) (1.546)
the tariff level in 1993 reflects the anticipated change in tariff pref-
t̄jUS × Vs × 1992 3.114 3.234 −1.082 −1.105
(2.400) (2.363) (2.076) (2.090)
erences following NAFTA. If plants react to the anticipated benefits
t̄jUS × Vs × 1993 1.933 1.748 −1.003 −0.957 of the agreement rather than the realized change in trade policy
(2.814) (2.815) (3.565) (3.598) in a given year, then re-estimating my baseline specification while
t̄jUS × Vs × 1994 1.187 0.842 0.390 0.358 controlling for anticipated policy changes instead of actual policy
(3.534) (3.583) (3.930) (3.937) changes should have a substantial effect on my conclusions.
t̄jUS × Vs × 1995 4.255 3.447 −0.285 −1.011
The estimates from this exercise are reported in the third column
(4.455) (4.425) (3.626) (3.787)
t̄jUS × Vs × 1996 3.208 2.763 −4.808 −5.068
of each panel in Table B1. As the table shows, including anticipated
(3.924) (3.890) (5.001) (5.025) policy changes rather than actual policy changes in the regressions
t̄jUS × Vs × 1997 2.772 2.085 −3.490 −3.649 alters the magnitude and significance of most of the estimated coef-
(3.774) (3.861) (4.596) (4.615) ficients. However, this is largely a byproduct of modifying how
t̄jUS × Vs × 1998 3.161 2.477 −1.297 −1.415 treatment is defined; altering the measure of treatment has little
(3.275) (3.318) (4.990) (4.950)
effect on the implied change in annual pollution emissions resulting
t̄j kInt × Vs × 1992 −9.089b −9.117b 1.614 2.008
(4.596) (4.554) (3.445) (3.483) from NAFTA in most cases. For example, the estimates for SO2 sug-
t̄j kInt × Vs × 1993 −12.275b −12.652b −1.424 −0.673 gest that fora plant facingthe average anticipated change in tariff
(5.147) (5.106) (4.713) (4.749) preferences t̄jtInt = 0.0093 , increased access to intermediate inputs
t̄j kInt × Vs × 1994 −10.766c −11.211c −2.905 −2.264
(5.771) (5.777) (5.293) (5.291) from Mexico decreased emissions by approximately 1.04% annually.
t̄j kInt × Vs × 1995 −12.196 −12.741c −13.997a −13.814a This decrease is comparable to the baseline estimate reported in
(7.542) (7.101) (4.473) (4.483) Table 4.
t̄j kInt × Vs × 1996 −12.759c −12.792c −18.190a −17.745a The key exceptions are the effects of increased Mexican market
(6.870) (6.743) (5.383) (5.426)
t̄j kInt × Vs × 1997 −17.087b −17.372b −15.789a −15.409a
access for PM10 and SO2 . In these cases, accounting for anticipated
(6.619) (6.756) (5.174) (5.177) policy changes leads to a substantial increase in the implied change
t̄j kInt × Vs × 1998 −10.958c −11.398c −10.787 −10.299 in annual pollution emissions. The estimated coefficients for these
(5.884) (5.948) (6.829) (6.847) pollutants indicate that the average anticipated changes in tariff
Peso Crisis X X
preferences (t̄jtMex = 0.1291 and t̄jtMex = 0.1097), decreased the
CUSFTA X X
1990 CAAAs X X emissions of PM10 and SO2 by 5.52% and 2.87%, respectively. Both
Observations 4162 4162 1645 1645 of these changes are larger than the changes implied by my base-
Adjusted R2 0.20 0.20 0.28 0.28 line estimates, suggesting my baseline estimates are understating
Notes: Table reports the estimated effects of NAFTA on emissions of PM10 and SO2 . the true effects of increased foreign market access.
Each column reports estimates from a regression based on Eq. (9). In all cases, the Finally, I turn to address the concern that my results may be
dependent variable is the log of pollution emissions. Columns (1) and (2) report capturing pre-existing differences in industry×state specific trends
estimates for PM10 , while columns (3) and (4) report estimates for SO2 . All spec-
ifications include plant, industry×year and state×year fixed effects. See text for
rather than the effects of trade liberalization due to NAFTA. To do so,
details. Standard errors clustered by industry and state are reported in parentheses. I estimate a flexible version of Eq. (5) that takes the form:
Significance: a 1%; b 5%; c 10%.

1998  

1998  
With the exception of making the point estimates less precisely ln zijst = b1t t̄jMex × Vs × kt + b2t t̄jUS × Vs × kt
estimated, excluding observations from 1993 appears to have lit- t=1992 t=1992
tle effect on most of the conclusions in my main analysis. These

decreases in precision can be attributed to the decrease in statistical
power arising from smaller sample sizes; for most of the point esti-
mates, the associated standard error increases substantially when
observations are excluded. The fact that dropping 1993 from has lit-
tle effect on the magnitude of the estimates suggests that, for the
most part, U.S. manufacturers did not respond in anticipation of the
agreement’s effects.
My baseline specification also implicitly assumes that plants do
not react to changes in tariff preferences until the year in which
they occur, although the schedule of changes were announced
1998  
+ b3t t̄jInt × Vs × kt + X  g + qi + ljt + dst + eijst (9)
t=1992
where t̄jMex , t̄jUS and t̄jInt are the anticipated changes in tariff prefer-
ences for each industry, kt are year fixed effects, X are the controls
for the Peso Crisis, CUSFTA, CAAAs and differences in pollution usage
described above, and all other variables are defined as in Eq. (5).
The interactions t̄jMex × Vs , t̄jUS × Vs and t̄jInt × Vs reflect differ-
ences in the magnitude of trade liberalization across plants following
148 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149

Table B3 Table B4
Descriptive statistics: VOCs. NAFTA and VOC emissions.

Pollutant Variable Obs. Mean Std. dev. Median (1) (2) (3) (4) (5)

VOCs Emissions (lb) 23495 109483 686099 22000 tjtMex × Vs 0.180 c

0.198 c
0.180 c
0.180 c
0.199c
Employees 23495 406 1008 165 (0.102) (0.110) (0.102) (0.102) (0.110)
Sales (1,000,000s) 23495 47 201 15 tjtUS × Vs −0.585 −0.580 −0.587 −0.578 −0.574
Emissions/worker 23495 534 2141 150 (0.498) (0.497) (0.498) (0.496) (0.497)
Emissions/sales 23495 6152 29178 1652 tjtInt × Vs 0.803 0.796 0.808 0.811 0.808
Notes: Table reports summary statistics of key variables for plants that emit volatile (0.907) (0.908) (0.911) (0.904) (0.910)
organic compounds. Peso Crisis X X
CUSFTA X X
1990 CAAAs X X
NAFTA, so the estimated vectors of b1 , b2 and b3 describe the cor- Observations 23495 23495 23495 23495 23495
relations between post-NAFTA treatment and pollution emissions in Adjusted R2 0.16 0.16 0.16 0.16 0.16
each year, relative to a baseline year, which I choose to be 1991.58 Notes: Table reports the estimated effects of NAFTA on the VOCs from U.S. manufactur-
Hence, the estimated coefficients reveal differences in the correlation ing plants. Each column reports estimated coefficients a different specification based
between treatment and pollution emissions in each year. As such, on Eq. (5). In all cases, the dependent variable is the log of pollution emissions. All
specifications include plant, industry×year and state×year fixed effects. See text for
I am interested in the pattern the coefficients display over time; if
details. Standard errors clustered by industry and state are reported in parentheses.
there are no pre-existing differences in trends in pollution emissions Significance: a 1%; b 5%; c 10%.
across plants, the estimated coefficients should not be statistically
different from zero prior to NAFTA.
Coefficient estimates from Eq. (9) are presented in Table B2. The estimates reported in Table B4 indicate that NAFTA had
The first two columns report estimates for PM10 , while the last little-to-no effect on the emissions of VOCs from affected plants.
two columns report estimates for SO2 . For each pollutant, the first For example, the estimate reported in row one of column (5) indi-
column reports estimates from a specification that includes plant, cates that, for a plant facing average trade frictions, a 1% increase in
industry×year and state×year fixed effects only, while the second Mexican market access increased VOC emissions by 0.199%. When
also includes controls for the Peso Crisis, CUSTFA and CAAAs. In each compared to the compared to the corresponding changes in PM10
case, standard errors clustered by industry×state are reported in and SO2 , it is very small in magnitude; given the actual increase
parentheses. in tariff preferences (0.0725), the estimate suggests that emis-
From the Table it appears that there are no clear trends in the sions increased by less than 0.3% annually. The remaining estimates
estimates for the years prior to NAFTA. This is an important valida- imply similarly small changes in emissions. These estimates suggest
tion of my research design; it indicates that my results are not simply that trade liberalization is not a sufficient condition for emissions
capturing pre-existing differences in trends across industry×states. reductions; it appears that the environmental effects of trade depend
the characteristics of production and abatement specific to different
B.2. NAFTA and emissions of volatile organic compounds pollutants.

The results reported in Section 5.1 indicate that NAFTA had sig-
nificant effects on the emissions of PM10 and SO2 . Here, I examine
the effects of the agreement on the emissions of another common
pollutant: volatile organic compounds (VOCs).59
Although it is a common pollutant, plants that emit VOCs are
much different than plants that emit PM10 and SO2 . This can be seen
by comparing the summary statistics for VOCs reported in Table B3
with the corresponding statistics from Table 1. Plants that emit VOCs
have fewer employees and lower sales, but emit substantially more
pollution than plants that produce PM10 or SO2 . Furthermore, VOCs
intensive industries produce little PM10 or SO2 ; based on estimates
reported in Shapiro and Walker (2016) the correlation between the
industry pollution elasticity of production for VOCs and PM10 is
0.048, while the correlation between the industry pollution elasticity
of production for VOCs and SO2 is 0.489. These differences may lead
to different responses to trade liberalization.
NAFTA’s effects on the emissions of VOCs are reported in Table B4.
As in my main analysis, this table reports estimates from five
specifications based on Eq. (5). The first, reported in column (1),
includes plant, industry×year, and state×year fixed effects only.
Columns (2)-(4) add controls for the Mexican Peso Crisis, the effects
of ongoing trade-liberalization due to CUSTFA, and the effects of
changes in environmental regulations due to the 1990 CAAAs. The
last specification, reported in column (5), controls for all of these
shocks simultaneously.
58
It should be noted that the choice of baseline year is arbitrary; choosing a different
year changes all of the estimated coefficients and the associated standard errors, but
does not affect conclusions.
59
I construct the dataset for VOCs using the same approach outlined in Appendix A.
References
Amiti, M., Konings, J., 2007. Trade liberalization, intermediate inputs, and productivity.
Am. Econ. Rev. 97 (05), 1611–1638.
Anderson, J.E., van Wincoop, E., 2004. Trade costs. J. Econ. Lit. 42 (3), 691–751.
Antweiler, W., Copeland, B.R., Taylor, M.S., 2001. Is free trade good for the environ-
ment? Am. Econ. Rev. 91 (4), 877–908.
Atkeson, A., Burstein, A., 2010. Innovation, firm dynamics and international trade. J.
Polit. Econ. 118 (3), 433–484.
Baldwin, R.E., Ravetti, C., 2014. Emissions, Exporters and Heterogeneity: Asymmetric
Trade Policy and Firms’ Selection. Working Paper 2014-2. Centre for Trade and
Economic Integration.
Barrows, G., Ollivier, H., 2014, November. Does trade make firms cleaner? Theory and
evidence from Indian manufacturing. Mimeo.
Bartelsman, E.J., Gray, W., 1996. The NBER Manufacturing Productivity Database.
Technical Working Paper 205. National Bureau of Economic Research. October.
Beaulieu, E., 2002. The Stolper-Samuelson theorem faces congress. Rev. Int. Econ. 10
(2), 343–360.
Bernard, A.B., Jensen, J.B., 1999. Exceptional exporter performance: cause, effect, of
both? J. Int. Econ. 47 (1), 1–25.
Bloom, N., Genakos, C., Martin, R., Sadun, R., 2010. Modern management: good for the
environment or just hot air? Econ. J. 120 (544), 551–572. (May).
Bolle, M.J., 2000. NAFTA: Estimated US Job Gains and Losses by State over 5-1/2 years.
Report 98-782. Congressional Research Service.
Burfisher, M.E., Robinson, S., Thierfelder, K., 2001. The impact of NAFTA on the United
States. J. Econ. Perspect. 15 (1), 125–144.
Bustos, P., 2011. Trade liberalization, exports and technology upgrading: evidence on
the impact of Mercosur on Argentinian firms. Am. Econ. Rev. 101 (1), 304–340.
Caliendo, L., Parro, F., 2015. Estimates of the trade and welfare effects of NAFTA. Rev.
Econ. Stud. 82 (1), 1–44.
Cameron, M.A., Tomlin, B.W., 2000. The Making of NAFTA: How the Deal was Done.
Cornell University Press.
Cherniwchan, J., Copeland, B.R., Taylor, M.S., 2017. Trade and the environment: new
methods, measurements, and results. Annu. Rev. Econ. 9.
Chintrakarn, P., Millimet, D.L., 2006. The environmental consequences of trade: evi-
dence from subnational trade flows. J. Environ. Econ. Manag. 52 (1), 430–453.
Cole, M.A., 2004. US environmental load displacement: examining consumption,
regulations and the role of NAFTA. Ecol. Econ. 48 (4), 439–450.
 J. Cherniwchan / Journal of International Economics 105 (2017) 130–149
Copeland, B.R., Taylor, M.S., 2004. Trade, growth and the environment. J. Econ. Lit. 42
(1), 7–71.
Costantini, J.A., Melitz, M.J., 2008. The Dynamics of Firm-level Adjustment to Trade
Liberalization. In: Helpman, E., Marin, D., Verdier, T. (Eds.), The Organization of
Firms in a Global Economy. Harvard University Press. pp. 107–141
Cui, J., Lapan, H., Moschini, G., 2012. Are Exporters more Environmentally Friendly
than Non-exporters? Theory and evidence. Working Paper 12022. Iowa State
University.
Cui, J., Lapan, H., Moschini, G., 2016. Productivity, export, and environmental perfor-
mance: air pollutants in the United States. Am. J. Agric. Econ. 98 (2), 447–467.
Davis, L.W., Kahn, M.E., 2010. International trade in used vehicles: the environmental
consequences of NAFTA. Am. Econ. J. Econ. Pol. 2 (4), 58–82.
Dean, J.M., 2002. Does trade liberalization harm the environment? A new test. Can. J.
Econ. 35 (4), 819–842.
Dean, J.M., Lovely, M.E., 2010. Trade Growth, Production Fragmentation, and China’s
Environment. In: Feenstra, R.C., Wei, S.-J. (Eds.), China’s Growing Role in World
Trade. University of Chicago Press. pp. 429–469
Agency, Environmental Protection, 2013. 2008 National Emissions Inventory. Techni-
cal Support Document version 3.
Feenstra, R.C., 1997. U.S. Exports, 1972–1994: With State exports and other U.S. Data.
Working Paper 5990. NBER. (April)
Feenstra, R.C., Romalis, J., Schott, P.K., 2002. U.S. Imports, Exports, and Tariff Data,
1989–2001. Working Paper 9387. NBER. (December).
Forslid, R., Okubo, T., Ulltveit-Moe, K.H., 2014. Why are Firms that Export Cleaner?
Working Paper 4817. CESifo.
Frankel, J.A., Rose, A.K., 2005. Is trade good or bad for the environment? Sorting out
the causality. Rev. Econ. Stat. 87 (1), 85–91.
Gamper-Rabindran, S., 2006. NAFTA and the environment: what can the data tell us?
J. Econ. Dev. Cult. Chang. 54 (3), 605–634.
Goldberg, P.K., Maggi, G., 1999. Protection for sale: an empirical investigation. Am.
Econ. Rev. 89 (5), 1135–1155.
Greenstone, M., 2002. The impacts of environmental regulations on industrial activity:
evidence from the 1970 and 1977 clean air act amendments and the census of
manufactures. J. Polit. Econ. 110 (6), 1175–1219.
Grossman, G., Helpman, E., 1994. Protection for sale. Am. Econ. Rev. 84 (4), 833–850.
Grossman, G., Krueger, A., 1991. Environmental Impacts of a North American Free
Trade Agreement. Working Paper 3914. NBER.
Gutierrez, E., Teshima, K., 2011. Import Competition and Environmental Performance:
Evidence from Mexican Plant-level and Satellite Imagery Data. Mimeo.
Hanson, G.H., 1998. North American economic integration and industry location. Oxf.
Rev. Econ. Policy 14 (2), 30–44.
Harrison, A., McMillan, M., 2011. Offshoring jobs? Multinationals and U.S. manufac-
turing employment. Rev. Econ. Stat. 93 (3), 857–875.
Head, K., Mayer, T., 2013. Gravity Equations: Workhorse, Toolkit, and Cookbook. In:
Gopinath, G., Helpman, E., Rogoff, K. (Eds.), Handbook of International Economics.
vol. 4. Elsevier.
Head, K., Ries, J., 2001. Increasing returns versus national product differentiation as an
explanation for the patter of U.S.-Canada trade. Am. Econ. Rev. 91 (4), 858–876.
Hillberry, R., Hummels, D., 2008. Trade responses to geographic frictions: a decompo-
sition using micro-data. Eur. Econ. Rev. 52 (3), 527–550.
Hillberry, R., McDaniel, C.A., 2002. A Decomposition of North American Trade Growth
since NAFTA. Office of Economics Working Paper 2002-12-A. United States Inter-
national Trade Commission. (December).
Holladay, J.S., 2016. Exporters and the environment. Can. J. Econ. 49 (1), 147–172.
Holmes, T.J., Stevens, J.J., 2012. Exports, borders, distance, and plant size. J. Int. Econ.
88 (1), 91–103.
149
Hufbauer, G.C., Esty, D.C., Orejas, D., Rubio, L., Schott, J.J., 2000. NAFTA and the Envi-
ronment: Seven Years Later. Institute for International Economics, Washington,
DC.
Hummels, D., Ishii, J., Yi, K.M., 2001. The nature and growth of vertical specialization
in world trade. J. Int. Econ. 54 (1), 75–96.
Kim, S., 1995. Expansion of markets and the geographic distribution of economic activ-
ities: the trends in U.S. regional manufacturing structure, 1860–1987. Q. J. Econ.
110 (4), 881–908. (November).
Kowalczyk, C., Davis, D., 1998. Tariff Phase-outs: Theory and Evidence from GATT
and NAFTA. In: Frankel, J.A. (Ed.), The Regionalization of the World Economy.
University of Chicago Press. pp. 227–258 (Chapter 8)
Kumar, A., 2006. Did NAFTA spur Texas exports? Southwest Econ. 3–7.
Leonhardt, D., 2008. The politics of trade in Ohio. The New York Times. (February 27).
Levinson, A., 2009. Technology, international trade, and pollution from US manufac-
turing. Am. Econ. Rev. 99 (5), 2177–2192.
Levinson, A., 2015. A direct estimate of the technique effect: changes in the pollution
intensity of US manufacturing, 1990–2008. J. Assoc. Environ. Resour. Econ. 2 (1),
43–56.
Levinson, A., Taylor, M.S., 2008. Unmasking the pollution haven effect. Int. Econ. Rev.
49 (1), 223–254.
Markell, D.L., Knox, J.H. (Eds.), 2003. Greening NAFTA: The North American Commission
for Environmental Cooperation. Stanford University Press, Stanford, CA.
Martin, L.A., 2012. Efficiency Gains from Trade: Greenhouse Gas Emissions and India’s
Manufacturing Firms. Mimeo.
McLaren, J., Hakobyan, S., 2016. Looking for local labor market effects of NAFTA. Rev.
Econ. Stat. 98 (4), 728–741.
Melitz, M.J., 2003. The impact of trade on intra-industry reallocations and aggregate
industry productivity. Econometrica 71 (6), 1695–1725.
Melitz, M.J., Ottaviano, G.I., 2008. Market size, trade, and productivity. Rev. Econ. Stud.
75, 295–316.
Neumark, D., Wall, B., Zhang, J., 2011. Do small businesses create more jobs? New
evidence for the United States from the national establishment time series. Rev.
Econ. Stat. 93 (1), 16–29.
Novy, D., 2013. Gravity redux: measuring international trade costs with panel data.
Econ. Inq. 51 (1), 101–121.
Pierce, J.R., Schott, P.K., 2012. A concordance between ten-digit U.S. harmonized system
codes and SIC/NAICS product classes and industries. J. Econ. Soc. Meas. 37 (1–2),
61–96.
Romalis, J., 2007. NAFTA’s and CUSTFA’s impact on international trade. Rev. Econ. Stat.
89 (3), 416–435.
Scheve,K.F.,Slaughter,M.J.,2001.Whatdeterminesindividualtrade-policypreferences?
J. Int. Econ. 54, 267–292.
Schott, P.K., 2008. The relative sophistication of Chinese exports. Econ. Policy 23 (53),
6–49.
Shapiro, J.S., Walker, R., 2016. Why is Pollution from U.S. Manufacturing Declining? The
Roles of Trade, Regulation, Productivity and Preferences. Mimeo.
Taylor, M.S., 2004. Unbundling the pollution haven hypothesis. Adv. Econ. Anal. Policy
4 (2), 1–26. (February).
Trefler, D., 1993. Trade liberalization and the theory of endogenous protection: an
econometric study of U.S. import policy. J. Polit. Econ. 101 (1), 138–160.
Trefler, D., 2004. The long and short of the Canada-U.S. free trade agreement. Am. Econ.
Rev. 94 (4), 870–895.
Yi, K.-M., 2003. Can vertical specialization explain the growth of world trade? J. Polit.
Econ. 111 (1), 52–102. (February).