UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays on the economics of environmental regulation Najjar, Nouri 2018

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2018_september_najjar_nouri.pdf [ 4.36MB ]
Metadata
JSON: 24-1.0371161.json
JSON-LD: 24-1.0371161-ld.json
RDF/XML (Pretty): 24-1.0371161-rdf.xml
RDF/JSON: 24-1.0371161-rdf.json
Turtle: 24-1.0371161-turtle.txt
N-Triples: 24-1.0371161-rdf-ntriples.txt
Original Record: 24-1.0371161-source.json
Full Text
24-1.0371161-fulltext.txt
Citation
24-1.0371161.ris

Full Text

Essays on the Economics of Environmental RegulationbyNouri NajjarB.Sc., University of Victoria, 2010M.A., University of British Columbia, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Economics)The University of British Columbia(Vancouver)August 2018c© Nouri Najjar, 2018The following individuals certify that they have read, and recommend to the Faculty of Grad-uate and Postdoctoral Studies for acceptance, the dissertation entitled:Essays on the Economics of Environmental Regulationsubmitted by Nouri Najjar in partial fulfillment of the requirements forthe degree of Doctor of Philosophyin EconomicsExamining Committee:Prof. Brian Copeland, Economics, University of British ColumbiaCo-supervisorProf. Carol McAusland, Food and Resource Economics, University of British ColumbiaCo-supervisorProf. Thomas Lemieux, Economics, University of British ColumbiaSupervisory Committee MemberProf. James Brander, Sauder School of Business, University of British ColumbiaUniversity ExaminerProf. Nicole Fortin, Economics, University of British ColumbiaUniversity ExaminerAdditional Supervisory Committee Members:Supervisory Committee MemberSupervisory Committee MemberiiAbstractEnvironmental regulations targeting producers are in place around much of the world. Yet,there is limited evidence of how firms are affected by these policies. This thesis provides newempirical and theoretical evidence on the effects of environmental regulation on producers.The first two chapters of this thesis explore a trend underway in much of the industrializedworld: pollution from manufacturing has been falling despite increased output. In the firstchapter, we develop a theoretical model to show the channels through which regulation cancontribute to an industry’s “clean-up”. This model highlights the role that fixed costs producersmust pay to adopt cleaner production processes play in dictating these channels. We show thatif these fixed costs are relatively low, the adoption of cleaner processes will be the primaryregulatory channel of an industry’s clean-up. However, if these fixed costs are relatively high,then plant exit and reductions in output from regulated plants will be the primary channels.The second chapter provides the first estimates of the regulatory channels of the manufac-turing clean-up. We estimate the share of the Canadian manufacturing clean-up explained bythe adoption of cleaner production processes, the reallocation of output across producers, andproducer entry and exit. To do this, we examine a major revision to Canadian environmentalpolicy using a novel, confidential dataset containing information on the production decisionsand pollution emissions of Canadian manufacturing plants. We find regulation explains, atmost, 61% of the Canadian clean-up, but the underlying channels differ strikingly across pol-lutants.A concern in debates over environmental regulation is a potential loss of international com-petitiveness among domestic producers. Despite its pervasiveness in policy discussions, evi-dence of these losses remains scarce. The third chapter of this thesis provides the first plant-level estimates of the effect of air pollution regulation on exporting. We study the effects of theCanada-Wide Standards for Particulate Matter and Ozone on the decision to export and exportvolumes of Canadian manufacturing facilities. We find evidence that environmental regula-tion caused relatively low-productivity exporters to leave the export market, and reduced theamount surviving exporters sold abroad.iiiLay SummaryThis thesis is intended to show some important dimensions through which environmental reg-ulation affects industry, and the firms and facilities therein. The first two chapters considerthe causes of an important trend in Canada: the manufacturing industry has become cleaner,in terms of air pollution, in recent decades. The first chapter develops an economic modelof how environmental regulation affects firms, and how these effects can cause an industry tobecome cleaner. In the second chapter, we then ask how a major increase in the stringency ofenvironmental regulation has contributed to this trend. We find regulation had considerable ef-fects on how manufacturing firms operate, which contributed to the manufacturing industry’sclean-up. The third chapter documents an additional margin through which environmentalregulation affects firms: exports. We find regulation reduced the amount of exports from regu-lated manufacturing facilities, and reduces the likelihood a relatively low-productivity facilitychooses to export.ivPrefaceAll chapters in this thesis were coauthored with Dr. Jevan Cherniwchan (University of Al-berta). I was the lead author of both of the first two chapters. In the first chapter, I developedthe economic model, which is the core contribution of that chapter. In the second chapter, Idiscovered the policy experiment that we study, which forms the basis of the empirical strat-egy we use to identify the effect of regulation. Both Jevan and I evenly contributed to the thirdchapter.In addition, I was a primary contributor to all additional stages of the research, includingdeveloping the research questions, preparing data, carrying out estimation, and organizing andpresenting results.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiIntroduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 A Theory of Industry Clean-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Channels of the Clean-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Theory of a Clean-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.2 The No-Regulation Equilibrium . . . . . . . . . . . . . . . . . . . . 81.3.3 The Partial Equilibrium Effects of Regulation . . . . . . . . . . . . . 121.3.4 The Equilibrium Effects of Regulation . . . . . . . . . . . . . . . . . 141.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 Estimating the Regulatory Channels of the Manufacturing Clean-Up . . . . . 252.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25vi2.2 The Clean-Up of Canadian Manufacturing . . . . . . . . . . . . . . . . . . . 302.3 Air Quality Regulation in Canada . . . . . . . . . . . . . . . . . . . . . . . . 332.3.1 Air Quality Improvements and the CWS . . . . . . . . . . . . . . . . 362.4 Empirics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.4.1 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422.4.2 Data and Measurement . . . . . . . . . . . . . . . . . . . . . . . . . 442.4.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713 Environmental Regulation and the Pollution Haven Effect . . . . . . . . . . . . 733.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733.2 The Pollution Haven Effect in a Small Open Economy . . . . . . . . . . . . . 783.2.1 Model Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783.2.2 Empirical Predictions of the Pollution Haven Effect . . . . . . . . . . 833.3 Data and Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903.3.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 903.3.2 Canadian Environmental Regulations . . . . . . . . . . . . . . . . . 923.4 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 943.4.1 Empirical Specification . . . . . . . . . . . . . . . . . . . . . . . . . 953.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963.5.1 Plant Revenue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963.5.2 Export Revenue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1003.5.3 Export Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1033.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109A Supporting Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117A.1 Chapter 1 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117A.1.1 Proof of Proposition 3 . . . . . . . . . . . . . . . . . . . . . . . . . 117A.1.2 Proof of Corollary 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 119A.1.3 Proof of Proposition 5 . . . . . . . . . . . . . . . . . . . . . . . . . 120A.1.4 Proof of Corollary 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 121A.1.5 Technology Upgrading . . . . . . . . . . . . . . . . . . . . . . . . . 121viiA.2 Chapter 2 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122A.2.1 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122A.2.2 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125A.2.3 Additional Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 132A.2.4 CWS Counterfactuals . . . . . . . . . . . . . . . . . . . . . . . . . . 136A.2.5 Policy Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138viiiList of TablesTable 2.1 Canadian Manufacturing Emission Intensity Decomposition: 92-15 . . . . 32Table 2.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46Table 2.3 The Effects of the CWS on Plant Pollution Emissions . . . . . . . . . . . 49Table 2.4 The Effects of the CWS on Plant Emission Intensity . . . . . . . . . . . . 51Table 2.5 The Effects of the CWS on Plant Output . . . . . . . . . . . . . . . . . . 52Table 2.6 The Effects of the CWS on Plant Exit . . . . . . . . . . . . . . . . . . . . 53Table 2.7 Counterfactual Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 57Table 2.8 The Effects of the CWS by Plant Productivity Level . . . . . . . . . . . . 59Table 2.9 Other Margins of Plant Adjustment . . . . . . . . . . . . . . . . . . . . . 62Table 2.10 CWS Mechanisms by Plant Productivity Level . . . . . . . . . . . . . . . 64Table 2.11 CWS Placebo Tests - PM2.5 Emissions . . . . . . . . . . . . . . . . . . . 66Table 2.12 CWS Placebo Tests - NOX Emissions . . . . . . . . . . . . . . . . . . . . 66Table 3.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91Table 3.2 Environmental Regulations and Plant Revenue . . . . . . . . . . . . . . . 98Table 3.3 Environmental Regulations and Revenue from Continuing Exporters . . . 101Table 3.4 Environmental Regulations and Export Revenue from Continuing Exporters 102Table 3.5 Environmental Regulations and Export Status . . . . . . . . . . . . . . . . 104Table A.1 Mean Emissions in Matched Dataset . . . . . . . . . . . . . . . . . . . . 126Table A.2 Regulation Cohorts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127Table A.3 CWS Effect on Emissions for Initial Treatment Cohort . . . . . . . . . . . 127Table A.4 CWS Effect on Emissions Dropping Large Emitters . . . . . . . . . . . . 128Table A.5 CWS Effect on Emissions with Large Emitter Trends . . . . . . . . . . . . 130Table A.6 CWS Effect on Emissions - Multi-Plant Firms . . . . . . . . . . . . . . . 134Table A.7 The Effects of the CWS on Plant Emissions of Unregulated Pollutants . . . 135Table A.8 Cross-Pollutant Elasticities . . . . . . . . . . . . . . . . . . . . . . . . . 137ixList of FiguresFigure 1.1 Technology Choices without Environmental Regulation . . . . . . . . . 11Figure 1.2 Technology Choices with Targeted Environmental Regulation . . . . . . 13Figure 1.3 Revenues for Surviving Firms with Targeted Environmental Regulation . 14Figure 1.4 Pollution Intensity for Surviving Firms with Targeted Environmental Reg-ulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15Figure 2.1 Output and Pollution from Canadian Manufacturing: 1992-2015 . . . . . 31Figure 2.2 Mean PM2.5 concentration by year with 95% confidence intervals. Panel A shows citiesnever above the PM2.5 standard. Panel B shows cities above the PM2.5 standard at leastonce. The red line represents the threshold for the PM2.5 Standard. The air quality metricused is the 98th percentile of each city’s 24-hour concentration. . . . . . . . . . . . 38Figure 2.3 Mean O3 concentration by year with 95% confidence intervals. Panel A shows citiesnever above the O3 standard. Panel B shows cities above the O3 standard at least once.The red line represents the threshold for the O3 Standard. The air quality metric used iseach city’s 4th highest 8-hour concentration. . . . . . . . . . . . . . . . . . . . 38Figure 2.4 Kernel density estimates of the distribution of PM2.5 concentrations across CMAs in thefirst half (2000-2005) and second half (2006-2011) of the CWS phase in period. Thepanel on the left displays pollution concentrations for CMAs that never exceeded thePM2.5 standard. The right panel displays the pollution concentrations for CMAs thatexceeded the PM2.5 standard at least once. The vertical red lines represents the thresholdused for the PM2.5 standard. . . . . . . . . . . . . . . . . . . . . . . . . . . . 39Figure 2.5 Kernel density estimates of the distribution of O3 concentrations across CMAs in thefirst half (2000-2005) and second half (2006-2011) of the CWS phase in period. Thepanel on the left displays pollution concentrations for CMAs that never exceeded the O3standard. The right panel displays the pollution concentrations for CMAs that exceededthe O3 standard at least once. The vertical red lines represents the threshold used for theO3 standard. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40xFigure 2.6 Mean Pollution Concentrations by Year for Large CMAs . . . . . . . . . 41Figure 2.7 Regulatory Status Changes under the CWS . . . . . . . . . . . . . . . . 47Figure 2.8 The Effect of PM2.5 Regulation on PM2.5 Emissions, by CMA Air Quality 68Figure 2.9 The Effect of O3 Regulation on NOX Emissions, by CMA Air Quality . . 68Figure 2.10 The Effect of PM2.5 Regulation on PM2.5 by Years Pre/Post Regulation . 70Figure 3.1 Export and Technology Choices without Environmental Regulation . . . 89Figure 3.2 Export and Technology Choices with Environmental Regulation . . . . . 90Figure 3.3 Regulatory Status Changes under the CWS . . . . . . . . . . . . . . . . 93xiAcknowledgmentsI wish to thank Brian Copeland and Carol McAusland for their guidance and mentorshipthroughout my time at UBC. I am deeply grateful for the many hours they contributed to mydevelopment as an economist, and will be forever indebted to them. Thank you to my othercommittee members, Thomas Lemieux and David Green, for invaluable advice, critiques, andcontributions. I would also like to thank my coauthor, Jevan Cherniwchan, for the depth ofhis commitment to our collaboration. To my colleagues at UBC – Alastair, Brad, Iain, Tom,Jacob, Jose, Joao, Timea, Hugo, Gaelle, Oscar, and Nathan, to name just a few – thank youfor your friendship and insights over the years.For helpful comments and suggestions, Jevan Cherniwchan and I wish to thank seminarparticipants at UBC, Western University, the University of Carlos the III, the University ofMannheim, the CREE annual meeting, the EEPRN symposium, the CEA annual conference,the Young Economists Symposium at Yale, the CU Environmental and Resource EconomicsWorkshop, and the Pacific Northwest Labor Day. We are especially thankful to Arik Levinsonand Joe Shapiro for insightful comments on an earlier version of the second chapter. Weare indebted to Michael Willox and Danny Leung at Statistics Canada’s Economic AnalysisDivision for their efforts to make this project possible. We thank members of Environment andClimate Change Canada’s (ECCC) Legislative and Regulatory Affairs Division for helpfuldiscussion on the design and implementation of Canadian air quality regulation. Fundingfrom the EEPRN, the Alberta School of Business Nova Faculty Fellowship, and the Centrefor Applied Business Research in Energy and the Environment at the University of Albertais gratefully acknowledged. The data used in this thesis was created through a collaborationbetween ECCC, Statistics Canada, and the Productivity Partnership, which is supported by theSocial Sciences and Humanities Research Council of Canada.I thank my parents for their unending support, advice, and love throughout my life. To mysisters and the rest of my family, thank you as well for all your support.Lastly, I owe the deepest debt of gratitude to Rebecca. Thank you for years of discussions,all manner of support, and for sharing your intelligence and drive with me each day.xiiIntroductionAir pollution has numerous negative effects on society. To name a few, certain pollutants havebeen shown to increase rates of mortality and morbidity, particularly among young childrenand adults with respiratory illness (Chay and Greenstone, 2003; Currie and Neidell, 2005;Schlenker and Walker, 2016). Moreover, exposure to air pollution can cause a reduction inworker productivity and labour supply (Chang et al., 2016; Graff Zivin and Neidell, 2012;Hanna and Oliva, 2015). As a consequence of these negative health and economic effects,in recent decades governments around much of the world have enacted increasingly stringentair pollution regulation. Despite their prevalence, relatively little is known about the way inwhich firms respond to and are affected by these policies.This thesis provides new evidence of the effects of air pollution regulation on manufactur-ing facilities. In the first two chapters, we determine the channels through which environmen-tal regulation can cause an industry to become less emission intensive (that is, to emit lesspollution per unit of output produced). In the first of these chapters, we present a theoreticalframework to capture these channels. We show regulation can cause an industry to “cleanup” through three channels: a selection channel, a reallocation channel, and a process chan-nel. The selection channel captures changes in an industry’s emission intensity that occursthrough facility entry and exit. If, for example, regulation causes relatively dirty facilities toshut down, then this selection channel will cause the industry to clean up. The reallocationchannel captures changes in an industry’s emission intensity that occurs due to reductions inoutput at surviving regulated facilities. If these facilities are relatively dirty, then this chan-nel will contribute to the industry’s clean-up. Lastly, the process channel occurs if regulationcauses surviving facilities to adopt cleaner production processes.In the first chapter we also show that the relative magnitude of each of the three channelswill depend on the size of the fixed cost plants need to pay to adopt cleaner processes. If thesefixed process costs are very small, then the process channel will drive an industry’s clean-up.Otherwise, the selection and reallocation channels will play an important role in an industry’sclean-up.1In the second chapter, we use a novel dataset to estimate the regulatory channels involvedin the clean up of the Canadian manufacturing sector. To determine causal estimates of regu-lation’s effect on manufacturing plants, we exploit a change in Canadian federal environmen-tal policy that has never before been studied. This policy implemented regional air qualitystandards for two pollutants in every major town and city of the country. In addition, onlyfacilities in a select group of industries were regulated. We exploit this variation in regula-tory stringency across cities, industries, and time using a triple difference research design. Inessence, this research design compares changes in outcomes for plants in targeted industriesand regions violating an air quality standard to changes in outcomes at other plants.We show robust evidence that this policy contributed to the Canadian manufacturing clean-up through each of the selection, reallocation, and process channels. Summing across allthree channels, we find this policy is responsible for 60% of the reduction in manufacturingnitrogen oxide emission intensity from 2004 to 2010. The policy is also responsible for 20%of the reduction in particulate matter emission intensity over this period. In addition, we showthat the relative magnitude of these channels varied across the two pollutants regulated by thispolicy. We argue these results are consistent with the theory presented in the first chapter,and in particular with our claim that the size of fixed process costs drives the channels of theclean-up. Lastly, we show additional evidence consistent with this hypothesis and theory.In the third chapter, we ask whether environmental regulation affects a firm’s incentive toexport. Debates over environmental regulation often center on their potential negative effect onthe competitiveness of domestic firms in international markets (see, e.g, Levinson and Taylor(2008)).1 Yet, to date, there have been no micro-level estimates of the effect of regulation onexport behaviour. We exploit the same policy change studied in Chapter 2 to estimate the effectof regulation on a facility’s decision to enter (or exit) export markets, and their export volumesconditional on exporting. Our results suggest regulation negatively affected the internationalcompetitiveness of Canadian manufacturing facilities. We find evidence this policy causeda negative effect on the international competitiveness of affected manufacturing plants. Forthe average affected plant, this policy caused a 22% reduction in export volumes. Moreover,regulation caused a 10% increase in the probability of exiting the export market for relativelylow-productivity exporters.1The concern in these debates is that regulation in one country may increase the cost of production fordomestic producers relative to their unregulated counterparts in foreign countries.2Chapter 1A Theory of Industry Clean-Up1.1 IntroductionThe past thirty years have witnessed a marked improvement in manufacturing pollutionlevels across much of the world despite large increases in manufacturing activity. In the UnitedStates, for example, manufacturing emissions of most air pollutants fell by between 52%-69%from 1990 to 2008, while total real shipments from the sector rose by 35% (Levinson, 2015).In Europe, manufacturing air pollution fell by between 23-59% from 1995 to 2008, whilereal shipments rose by 37% (Brunel, 2016). These patterns appear to extend outside of theUnited States and Europe; sulphur dioxide emissions from manufacturing have been fallingin a number of countries despite increases in shipments (Grether et al., 2009). These broadtrends imply that, for much of the industrialized world, manufacturing is becoming cleaner.Recent evidence suggests environmental regulation has played a large role in this “clean-up” of manufacturing (Shapiro and Walker, 2015). In addition, this clean-up appears to bedue to reductions in the emission intensity of individual industries (Brunel, 2016; Levinson,2015), rather than changes in the composition of the manufacturing sector. Yet, at present,little theoretical work has been done to directly assess how regulations change an industry’semission intensity. In this chapter we present a theoretical model to capture the channelsthrough which a particular type of environmental regulation can cause an industry to clean up– what we call a two-part regulatory structure. In this type of regulation, firms must eithermake technological changes to meet industry best practices, or are subject to a regulatorypenalty. The theoretical work has yet to assess this type of policy, despite its common use inregulating air pollutants. For example, the National Ambient Air Quality Standards used aspart of the US Clean Air Act require regulated facilities to adopt state-of-the-art abatement3technology, and fines those that fail to do so (Greenstone, 2002). Similarly, the Canada-WideStandards for Particulate Matter and Ozone require regulated facilities to use clean productionprocesses, and imposes production constraints on those that fail to do so (see Chapter 2 fordetails).Before presenting our model, we start by following a similar approach to Cherniwchan et al.(2017) to show that a change in an industry’s emission intensity can be decomposed into threechannels: a selection channel, a reallocation channel, and a process channel. The selectionchannel reflects the exit of plants in response to regulation. The reallocation channel reflectsthe reduction in output at surviving regulated plants in response to regulation. Lastly, theprocess channel captures the adoption of cleaner production processes at surviving regulatedplants.We next present the theoretical model we use to show the mechanisms driving these chan-nels. This model is based on a closed-economy variant of the Melitz (2003) model in whichpollution from heterogeneous firms is regulated. This model has three key features. First, itallows for firm productivity differences, which have been highlighted as a key determinant ofthe effects of environmental regulation in the existing theoretical literature (see, e.g., Konishiand Tarui (2015) or Anoulies (2017)). Second, it allows for endogenous technology adoptionby firms to capture the fact that leading technologies are often used as a benchmark for thetechnical changes required under regulation. Third, it allows for differences across pollutantsin the cost of adopting less-polluting production processes, which we call process costs. Thisfeature is important because pollutants often feature different process costs.1This chapter relates most closely to the recent body of theoretical work studying environ-mental regulations and pollution in the presence of firm heterogeneity. Our model makes threeprimary contributions to this literature. First, as we discussed above, we consider a regulationthat features a two-part regulatory structure. This is in contrast to the existing literature, whichfocuses on uniform pollution taxes (Andersen, 2018; Cao et al., 2016; Forslid et al., 2014; Liand Shi, 2017; Li and Sun, 2015), pollution permit trading (Anoulies, 2017; Cui et al., 2015;Konishi and Tarui, 2015), or pollution intensity standards (Li and Shi, 2017; Li and Sun, 2015).We make this departure to capture a common feature of environmental policy that has not yetbeen highlighted in the theoretical literature.Our second contribution is to make explicit the connection between environmental regula-tion and the channels through which an industry becomes less pollution intensive. While notexplicitly stated as such, the majority of the existing literature has focused on the selection and1For example, nitrogen oxide (NOX) process costs are relatively low, while PM2.5 process costs are typicallyrelatively high (Canadian Council of Ministers of the Environment, 1998b; Environment Canada, 2002).4reallocation effects induced by policy (e.g. Konishi and Tarui (2015) or Anoulies (2017)). Byallowing for process effects, our work is more closely related to that of Cao et al. (2016), whostudy the effects of a uniform pollution tax on investment in abatement technology. UnlikeCao et al., we rely on constant elasticity of substitution preferences, rather than quasi-linearpreferences, and allow for a different form of technology adoption (discussed below).Our third contribution is to focus on discrete technology choices, following an approachused by Bustos (2011), rather than consider continuous abatement investments. This is in con-trast to two alternative approaches. The first is the canonical approach used in environmentaleconomics in which facilities can make incremental process changes, which increase theirvariable production costs, but do not require additional fixed costs (see, e.g., Antweiler et al.(2001); Shapiro and Walker (2015)). The second is to allow for smooth changes in a fixed-costtechnology (see, e.g., Cao et al. (2016)). While a number of papers also present models basedon Bustos (2011) featuring heterogeneous polluting firms that make endogenous abatementdecisions (Batrakova and Davies, 2012; Cui et al., 2012; Forslid et al., 2014), these studiesfocus on the effects of international trade. Instead, we show how environmental regulationinfluences firms to make discrete changes in technology. As we show, the fixed costs requiredto make these changes play an important role in dictating the channels involved in a clean-up, particularly if facilities can effectively avoid regulation by adopting process changes. Inthis case, the size of the process cost directly affects the relative magnitudes of the channelsinvolved in an industry’s clean-up.The remainder of this chapter is as follows. In Section 1.2 we present the results of ourdecomposition exercise showing the potential channels of a clean-up. In Section 1.3, wepresent our theoretical model. In Section 1.4, we discuss the model’s main implications forthe channels of a clean-up. Lastly, a short conclusion summarizes.1.2 Channels of the Clean-UpBefore presenting our theoretical model, we first use a decomposition exercise to derive thechannels through which plant-level changes in response to environmental regulation can causean industry to clean-up. For this exercise, we follow an approach similar to that introducedby Cherniwchan et al. (2017), which itself extends the decomposition presented by Levinson(2009).To that end, let output and pollution from manufacturing industry i be given by Xi and Zi,respectively. We define an industry’s pollution intensity as the amount of pollution emittedper unit of output produced, and let this be given by Ei = Zi/Xi. In addition, suppose eachindustry is composed of a continuum of plants and let xi(n), zi(n), and ei(n) denote output,5pollution, and pollution intensity from plant n. Lastly, let λi(n) = xi(n)/Xi be plant n’s share ofproduction in industry i and ni denote the marginal plant that is endogenously determined bythe industry’s profitability.2 In this case, the emission intensity of industry i can be expressedas a weighted average of the plant emission intensitiesEi =∫ ni0ei(n)λi(n)dn. (1.1)Totally differentiating Equation (1.1) gives the change in emission intensity of any industry iasdEi =∫ ni0dei(n)λi(n)dn+∫ ni0ei(n)dλi(n)dn+ ei(ni)λi(ni)dni. (1.2)We call the first term on the right-hand side of equation (1.2) the “process effect”. Thiscaptures the change in industry emission intensity due to changes in plant emission intensityresulting from the adoption of new production processes. As such, this term captures thedirect effects of a shock; all else equal, industry emission intensity will fall if a shock suchas environmental regulation induces plants to lower their emission intensities. The remainingtwo terms capture indirect changes in industry emission intensity. The first of these, given inthe second term on the right-hand side of equation (1.2), captures the effects of the shock onthe relative size of plants within an industry. This “reallocation effect” would arise if the shockdoes not affect plants uniformly. If the shock only affects a subset of plants in an industry, asis common with many environmental regulations, this may cause a reduction in the relativeoutput of affected plants. This would cause a change in industry emission intensity, even inthe absence of direct changes in plant emission intensity. Finally, the “selection effect” givenby the third term captures the change in emission intensity created by a change in the set ofplants operating within the industry owing to plant entry and exit.Equation (1.2) shows that regulation may cause an industry’s emission intensity to fallby causing plant-level reductions in emission intensity (the process effect), changes in therelative output of dirty and clean plants (the reallocation effect), or a change in the plantsthat comprise the industry (the selection effect). In what follows, we present a theoreticalframework to capture the regulatory process, reallocation, and selection effects induced by aregulatory change.2As in Cherniwchan et al. (2017), we assume plants are ranked in reverse order of productivity. Conse-quently, selection removes the least productive plants.61.3 Theory of a Clean-Up1.3.1 Model SetupTo capture the regulatory channels of an industry’s clean-up, consider an economy com-prised of L identical consumers, each endowed with a single unit of labor. Labor is suppliedinelastically and used to produce differentiated products in a single industry. Production alsocreates pollution as a byproduct, and this harms consumers, lowering their utility. For conve-nience, in what follows, we let wages be the numeraire.The representative consumer derives utility from the consumption of goods and disutilityfrom aggregate pollution according to U = [∫M0 q(ω)ρdω]1/ρ − h(Z), where q(ω) denotesconsumption of good ω , and M denotes the measure of varieties available in the economy.It is assumed consumers ignore pollution when making their consumption decisions. As aresult, the demand for variety ω is given by q(ω) = IPσ−1 p(ω)−σ , where I denotes consumerincome, P = [∫M0 p(ω)1−σdω]1/(1−σ) is the economy’s price index, and σ = 1/[1−ρ]> 1 isthe elasticity of substitution between goods.The supply side of the economy features monopolistic competition and free entry, meaningeach firm in the economy produces a unique variety. To enter, firms pay a fixed entry cost fε ,and upon entry, draw a productivity level ϕ from a common distribution G(ϕ).3 Based on therealization of ϕ , firms decide whether to exit or stay in the market, and conditional on staying,how much to produce and what technology to use in production.Upon entering, firms are able to produce output x using a business-as-usual technology(labeled with subscript b) that features increasing returns to scale. With this technology, thetotal costs of production are given by Cb = clb(ϕ)x+ f , where clb(ϕ) is the marginal cost ofproducing x with technology b under regulatory regime l, which we describe further below.Moreover, the business-as-usual technology has an emission intensity of eb = κ/ϕ , meaningthe production of x creates zb(ϕ) = [κx]/ϕ units of pollution.While firms are endowed with the business-as-usual technology, they can choose to up-grade their technology along one of two dimensions. First, they can adopt a state-of-the-arttechnology (labeled with subscript s) that boosts labor productivity, lowering marginal costsby a factor 1/α . The state-of-the-art technology also produces fewer emissions per unit of out-put. In this case, the emission intensity of production is given by es = κ/[γϕ], where γ > 1, sototal pollution from production is zs(ϕ) = [κx]/[γϕ]. Adopting the state-of-the-art technologyrequires that firms pay an additional fixed cost fs, meaning total production costs with the3For simplicity, G(ϕ) is assumed to be a type-I Pareto distribution such that G(ϕ) = 1−ϕ−k.7state-of-the-art technology are given by Cs = cls(ϕ)x+ f + fs, where cls(ϕ) is the marginal costof producing x with the state-of-the-art technology in regime l.We view state-of-the-art technology as reflecting a suite of industry-leading processes andtechnologies. These reduce labor costs and pollution intensity jointly for at least two potentialreasons. First, some of these processes may make production cleaner by reducing the intensityof certain dirty inputs, such as fossil fuels.4 Second, this serves as a simple method to reflectthat some firms may be forward-looking, and believe that future regulation will be tightened.Consequently, when building a facility, they may choose to do so with clean processes inanticipation of future regulation, even if the precise nature of that regulation is unknown.5Instead of becoming state-of-the-art, firms may retrofit their business-as-usual technologyso that it has the same emission intensity as the state-of-the-art technology. As such, theemission intensity of a retrofitted plant (er) is also κ/[γϕ], meaning the total level of pollutiongenerated by production is zr(ϕ) = [κx]/[γϕ]. Retrofitting also requires firms to pay a fixedcost ( fr). However, retrofitting does not affect labor productivity, meaning it is less costlythan adopting the state-of-the-art technology, so fr < fs. The total costs of production fora retrofitted plant are given by Cr = clr(ϕ)x+ f + fr, where clr(ϕ) is the marginal cost ofproducing x in regulatory regime l with the retrofitted technology.1.3.2 The No-Regulation EquilibriumOur interest is in understanding the effects of environmental regulation that requires firmsto use clean production processes, and penalizes those that fail to do so. We call this targetedregulation. We first consider a no regulation regime (labeled with superscript no) in whichpollution is not regulated. This means labor costs are the only variable costs of production, socnob (ϕ) = cnor (ϕ) = 1/ϕ and cnos (ϕ) = 1/[αϕ].A firm that has drawn a productivity level maximizes profits by deciding whether to stay inthe market, and if they stay, choosing how much to produce and what technology to use. Giventhe structure of consumer preferences, this implies that producing firms set prices at a constantmark-up over marginal costs. Hence, in the absence of regulation, firms that employ business-as-usual and retrofitted technologies charge the same price: pnob (ϕ) = pnor (ϕ) = 1/[ρϕ]. If,instead, a firm employs the state-of-the-art technology, it charges pnos (ϕ) = 1/[ραϕ].Firms choose between the three available technologies to maximize profits. If firms em-ploy the business-as-usual technology, profits are given by pinob =1σ I [Pρ]σ−1ϕσ−1− f . Profits4For example, by increasing thermal efficiency, as can be the case with low-NOX burners (Applied Tech-nologies of New York Inc., 2018). In general, burning less fossil fuels on-site, all else equal, has the potential toreduces a firm’s emissions of many air pollutants.5As our model is static, we abstract away from the intertemporal nature of this decision.8from employing the retrofitted technology are pinor = 1σ I [Pρ]σ−1ϕσ−1− [ f + fr]. Finally, prof-its from choosing the state-of-the-art technology are given by pinos = 1σ I [Pρ]σ−1ϕσ−1ασ−1−[ f + fs].It is worth noting that firms never choose the retrofitted technology in the absence of reg-ulation. If firms adopt the retrofitted technology, the emission intensity of production falls,but this has no effect on the variable costs of production because pollution is not costly tothe firm if it is not regulated. As a result, retrofitting simply lowers firm profits below whatcan be obtained using the business-as-usual technology, by increasing the average costs ofproduction.In addition, we assume fs > [ασ−1−1] f , so that the marginal surviving firm uses business-as-usual technology. In this case, the marginal producer’s productivity cutoff ϕnoε can bedetermined by noting that pinob (ϕnoε ) = 0. Substituting for pinob (ϕnoε ) and rearranging yieldsϕnoε =[σ fI] 1σ−1 1ρPno. (1.3)That is, firms that draw a productivity level below ϕnoε exit the market, as they would not beprofitable enough to pay the fixed cost of production ( f ).Similarly, firms upgrade to the state-of-the-art technology s when it is profit maximizingto do so. Firms with low productivity levels choose to use the business-as-usual technol-ogy, while relatively productive firms will adopt state-of-the-art technology. The productivitycutoff for technology-upgrading, ϕnos is defined by pinob (ϕnos ) = pinos (ϕnos ). Substituting andrearranging yieldsϕnos =[σ fs∆1I] 1σ−1 1ρPno, (1.4)where ∆1 = ασ−1−1 > 1.By combining equations (1.3) and (1.4), we can can write the technology upgrading cutoffas a function of the exit cutoffϕnos = ϕnoε[fs∆1 f] 1σ−1. (1.5)Hence, all equilibrium values can be defined by the exit cutoff.To solve for the industry equilibrium, as in Melitz (2003), we exploit the fact that becauseof free entry, in expectation, firms earn zero profits. Using this condition solves for all endoge-nous variables, including industry prices, firm exit, and technology choices. In equilibrium,9the fixed entry cost fε , must equal expected profits. Allowing for some exogenous probabilityof exit, given by δ , then this free entry condition isfε =1−G(ϕnoε )δp¯ino, (1.6)where p¯ino are a firm’s expected profits conditional on surviving, given byp¯ino =∫ ϕnosϕnoεpinob (ϕ)g(ϕ)1−G(ϕnoε )dϕ+∫ ∞ϕnospinos (ϕ)g(ϕ)1−G(ϕnoε )dϕ. (1.7)After substituting for pinob (ϕ) and pinos (ϕ) and exploiting equations (1.3) and (1.5), it is possibleto showp¯ino =[σ −1] fk−σ +1Ψ1, (1.8)where k is the shape parameter of the Pareto productivity distribution, andΨ1 =[1+∆kσ−11[ffs] k−σ+1σ−1]. (1.9)Hence, the exit cutoff ϕnoε can be obtained by substituting equation (1.8) into equationequation (1.6) and noting 1−G(ϕ) = ϕ−k. Doing so yields the following expression for theexit cutoffϕnoε =[[σ −1k−σ +1][fδ fε]Ψ1] 1k. (1.10)To ensure expected profits are positive we impose the restriction k > σ −1.With an expression for ϕnoε it is possible to solve for the technology upgrading cutoff andthe aggregate price index. The technology upgrading cutoff ϕnos can be determined by substi-tuting equation (1.10) into equation (1.5)ϕnos =[fs∆1 f] 1σ−1[[σ −1k−σ +1][fδ fε]Ψ1] 1k. (1.11)The price index can be obtained by first noting that I = L, and then substituting equa-10ϕσ−1pinobpinorpinos− f−[ f + fr]−[ f + fs]ϕno[σ−1]ε ϕno[σ−1]sExit Business-as-Usual State-of-the-ArtFigure 1.1: Technology Choices without Environmental Regulationtion (1.10) into equation (1.3) and rearranging to obtainPno =[σ fL] 1σ−1[ρ[[σ −1k−σ +1][fδ fε]Ψ1] 1k]−1. (1.12)The exit and technology choices made by firms are highlighted in Figure 3.1, which depictsthe profits associated with adopting each technology as a function of firm productivity.6 Asthe figure shows, for productivity levels below ϕnoε it is unprofitable for a firm to operate usingany technology. Hence if a firm has a ϕ less than ϕnoε , it exits the market. If firms stay inthe market, they choose the technology that yields the highest profit. This means that if afirm has a productivity level ϕ ∈ {ϕnoε ,ϕnos }, then it will produce using the business-as-usualtechnology. However, if a firm has a productivity level ϕ > ϕnos , then the reduction in variablecost created by adopting the state-of-the-art technology is great enough to justify the fixed costof adoption, meaning that these firms adopt the state-of-the-art technology.6To linearize this figure, we show profits as a function of ϕσ−1, not ϕ .111.3.3 The Partial Equilibrium Effects of RegulationWe now consider the effects of adopting a targeted form of environmental regulation inwhich firms are penalized for failing to use a clean production process. This regime can bethought of as a weak technology (or process) standard in which firms need to either adopt aspecific form of production technology (or process), or face a constraint that limits the prof-itability of producing.7 We first examine this policy in a partial equilibrium context in whichindustry prices are held fixed at the no-regulation level. This is a useful exercise as it under-scores the intuition behind the selection, reallocation, and process effects of such a policy.In this regime (labeled with superscript tar), the government regulates pollution using atwo-part regulatory rule. If a firm uses a clean production process (either the state-of-the-arttechnology or the retrofitted technology), it is not subject to regulation because it is operat-ing with the lowest emission intensity currently available. As a result, the marginal costs ofproduction for these firms are unaffected by regulation. To make this explicit, firms that usethe retrofitted technology, which we label with subscript r, face the same marginal costs withregulation (labeled with superscript tar) and without regulation (labeled with superscript no):ctarr (ϕ) = cnor (ϕ) = 1/ϕ . Similarly, the same is true for firms using state-of-the-art technology(labeled with subscript s): ctars (ϕ) = cnos (ϕ) = 1/[αϕ]. In contrast, a firm that employs a dirtyproduction process (the business-as-usual technology) is subject to a regulatory constraint inthe form of a tax τ on each unit of pollution emitted.8 Hence, regulation raises the marginalcosts of production for these firms. Labeling business-as-usual production technology withsubscript b, this means cnob (ϕ)< ctarb (ϕ) = [1+κτ]/ϕ . To abstract away from the redistribu-tive aspects of environmental taxation, we assume tax revenue is not returned to consumersand is spent outside the model.Given that firm prices feature a constant markup, this increase in marginal costs raises theprice of output for firms producing with the business-as-usual technology. That is, pnob (ϕ) <ptarb (ϕ) = [1+κτ]/[ρϕ], and profits are pitarb =1σ I [Pnoρ]σ−1ϕσ−1[ 11+κτ ]σ−1− f . This means,holding industry prices fixed, the profit from using the business-as-usual technology falls forany level of productivity ϕ .This partial equilibrium outcome is depicted in Figure 3.2, which displays the technologicalchoices made by firms when faced with targeted regulation holding industry prices (P) fixed.As the figure shows, a reduction in the profitability of using the business-as-usual technology7The reduction in profitability could occur either because of increased production costs, say as the result ofa tax, fine, or a production constraint.8Alternatively, we could impose a production cap, without substantively affecting the intuition behind theresults. We use a tax for analytic tractability.12ϕσ−1pinobpitarbpinor = pitarrpinos = pitars− f−[ f + fr]−[ f + fs]ϕno[σ−1]ε ϕtar[σ−1]ε ϕtar[σ−1]r ϕtar[σ−1]s ϕno[σ−1]sExit Bus.-as-Usual Retrofit State-of-the-ArtFigure 1.2: Technology Choices with Targeted Environmental Regulationincreases the productivity level for which it is unprofitable to enter the market from ϕnoε toϕ tarε . As such, firms with ϕ ∈ {ϕnoε ,ϕ tarε } exit in response to regulation. Moreover, giventhe design of regulation, profits from using the retrofitted or state-of-the-art technology donot change. This means the increase in the variable cost of the business-as-usual technologymakes technology upgrading a profitable alternative for some firms. As depicted, it is profitmaximizing for firms with productivity ϕ ∈ {ϕ tarr ,ϕnos } to retrofit their technology in responseto regulation. For these firms, the benefit of avoided tax payments outweighs the increase infixed production costs. Similarly, firms with productivity ϕ ∈ {ϕ tarc ,ϕnos } adopt the state-of-the-art technology in response to regulation because it is now profit maximizing to do so.While Figure 3.2 clearly highlights how environmental regulations create selection effectsby causing firms to exit in response to regulation, the reallocation and process effects are notreadily apparent from the figure. As such, we further explore how regulations affect firmrevenues and emission intensities to make these additional effects clear.These effects for firms that survive regulation (those with ϕ > ϕ tarε ) are displayed in Fig-ure 1.3 and Figure 1.4. These figures depict the effects of environmental regulation on firmrevenues (Figure 1.3) and emission intensity (Figure 1.4) holding industry prices fixed. Bothfigures show that the most productive firms, with productivity ϕ > ϕnos , are unaffected by13ϕσ−1Rns = RtarsRnb = RtarrRtarbϕ tar[σ−1]ε ϕtar[σ−1]r ϕn[σ−1]sϕtar[σ−1]sExit Bus.-as-Usual Retrofit State-of-ArtFigure 1.3: Revenues for Surviving Firms with Targeted Environmental Regulationregulation, as they use a clean production technology in either regime. In contrast, regula-tion causes the least productive firms, with productivity ϕ < ϕ tarr , to produce less, but withthe same pollution intensity. This is because they use the business-as-usual technology undereither regime, and variable costs rise under regulation. Lastly, pollution intensity falls for thefirms in the middle of the productivity distribution, with productivity ϕ ∈ {ϕ tarr ,ϕnos }. Thisoccurs because they either retrofit or adopt state-of-the-art technology. The retrofitting firmsexperience no change in output, as their variable costs do not change relative to business-as-usual. However, output increases for the new state-of-the-art adopters, as both their pollutionintensity and variable costs fall.1.3.4 The Equilibrium Effects of RegulationThe discussion in the preceding section illustrated how targeted environmental regulationcauses an industry to clean up through selection, reallocation, and process effects, but did so ina partial equilibrium setting. In this section, we show that similar results hold in equilibrium.In particular, we show that in equilibrium:1. Regulation causes some firms to exit.2. The effects of regulation vary across the productivity distribution. Revenues fall for the14ϕeber = esϕ tarε ϕ tarr ϕnosϕ tarsExit Bus.-as-Usual Retrofit State-of-ArtFigure 1.4: Pollution Intensity for Surviving Firms with Targeted Environmental Regu-lationleast productive surviving firms using the business-as-usual technology, while emissionintensity falls for firms in the middle of the productivity distribution that retrofit orupgrade to the state-of-the-art technology.3. The fixed cost of retrofitting ( fr) plays an important role in determining the channelsthrough which regulation causes an industry to clean-up. Specifically, decreasing frincreases the measure of firms that adopt a clean production process in response toregulation and reduces the measure of firms that exit in response to regulation.Closing the Model Under Targeted Environmental RegulationTo show these equilibrium effects, we first solve the model under the targeted environmen-tal regulation regime, and then show the relevant comparative statics. As in Section 1.3.2, weagain solve the model by first deriving productivity cutoffs; however, in this case, we nowalso consider the cut-off for retrofitting in addition to the cut-offs for firm exit, and technologyupgrading.Again, the marginal firm uses the business-as-usual technology b, meaning the regulationexit cutoff ϕnoε can be determined by noting that pitarb (ϕtarε ) = 0. Substituting for pitarb (ϕtarε )15and rearranging yieldsϕ tarε =[σ fI] 1σ−1[1+ τκρPtar]. (1.13)While firms are endowed with the business-as-usual technology, they will retrofit or up-grade to the state-of-the-art technology if it is profitable to do so. The productivity cutofffor retrofitting with regulation, ϕ tarr is defined by pitarb (ϕtarb ) = pitarr (ϕ tarr ). Substituting andrearranging yieldsϕ tarr =[σ frI∆2] 1σ−1 1ρPtar, (1.14)where ∆2 = 1− 1[1+τκ]σ−1 > 0. Similarly, the productivity cutoff for technology upgrading withregulation, ϕ tars is defined by pitarr (ϕ tarr ) = pitars (ϕ tars ). Substituting and rearranging yieldsϕ tars =[σ [ fs− fr]∆1I] 1σ−1 1ρPtar. (1.15)Both ϕ tarr and ϕ tars can be expressed as functions of the exit cutoff productivity ϕ tarb . Theexpression for the retrofitting cutoff can be obtained by combining equation (1.14) with (1.13)to getϕ tarr =ϕ tarε1+ τκ[fr∆2 f] 1σ−1. (1.16)Similarly, an expression for the upgrading cutoff can be obtained by combining equation (1.15)with (1.13) to getϕ tars =ϕ tarε1+ τκ[fs− fr∆1 f] 1σ−1. (1.17)We impose the additional assumptions: fr > [[1+ τκ]σ−1 − 1] f , and fs > ∆1+∆2∆2 fr. Thisensures the technology-upgrading cutoff is always greater than the retrofitting cutoff, meaningall three technologies (b, s, and r) are used in the regulated equilibrium.To solve for the industry equilibrium, we again exploit the fact that in expectation, firmsearn zero profits due to free entry. Hence, in the regulated equilibrium, the fixed entry cost fεmust equal expected profitsfε =1−G(ϕ tarε )δp¯itar, (1.18)16where p¯itar are a firm’s expected profits conditional on surviving, given byp¯itar =∫ ϕtarrϕtarεpitarb (ϕ)g(ϕ)1−G(ϕ tarε )dϕ+∫ ϕtarsϕtarrpitarr (ϕ)g(ϕ)1−G(ϕ tarε )dϕ+∫ ∞ϕtarspitars (ϕ)g(ϕ)1−G(ϕ tarε )dϕ. (1.19)Substituting for pitarb (ϕ), pitarr (ϕ), and pitars (ϕ) and utilizing equations (1.13), (1.14) and (1.15),it is possible to showp¯itar =[σ −1] fk−σ +1Ψ2. (1.20)where:Ψ2 =[1+[1+ τκ]k[∆kσ−12[ffr] k−σ+1σ−1+∆kσ−11[ffs− fr] k−σ+1σ−1]]> 0. (1.21)The exit cutoff ϕ tarε can be obtained by substituting (1.20) into (1.18) and using 1−G(ϕ) =ϕ−kϕ tarε =[[σ −1k−σ +1][fδ fε]Ψ2] 1k. (1.22)Having determined ϕ tarε , it is again possible to obtain expressions for ϕ tarr and ϕ tars , and theprice index, Ptar. The retrofitting cutoff can be obtained by substituting equation (1.22) intoequation (1.16)ϕ tarr =11+ τκ[fr∆2][[σ −1k−σ +1][fδ fε]Ψ2] 1k. (1.23)The technology cutoff, on the other hand, can be obtained by substituting equation (1.22) intoequation (1.17)ϕ tars =11+ τκ[fs− fr∆1 f] 1σ−1[[σ −1k−σ +1][fδ fε]Ψ2] 1k. (1.24)An expression for the price index can be obtained by substituting equation (1.22) into17equation (1.13) and noting that given our assumption that environmental tax revenues are notreturned to consumers I = L. This yieldsPtar =[σ fL] 1σ−1[ρ[[σ −1k−σ +1][fδ fε]Ψ2] 1k]−1. (1.25)Comparative Statics of Targeted Environmental RegulationWe begin our comparative statics exercise by examining how regulation affects firm exit.In all comparative static exercises that follow, we assume fr > 0, as well as the parameterrestrictions imposed in previous sections.Proposition 1. Targeted environmental regulation causes firms to exit.Proof. This claim can be proved by comparing ϕnoε and ϕ tarε . Doing so yieldsϕ tarεϕnoε=1+[1+ τκ]k[∆kσ−12[ffr] k−σ+1σ−1+∆kσ−11[ffs− fr] k−σ+1σ−1]1+∆kσ−11[ffs] k−σ+1σ−1k. (1.26)A sufficient condition to ensure ϕ tarε /ϕnoε > 1 is[1+ τκ]k∆kσ−11[ffs− fr] k−σ+1σ−1> ∆kσ−11[ffs] k−σ+1σ−1, (1.27)which can be re-expressed as [1+ τκ]k [ fs/[ fs− fr]]k−σ+1σ−1 > 1. This is satisfied given the as-sumptions of the model.Proposition 1 shows that if more productive firms are less pollution intensive, then theselection channel is present for targeted regulation. In this case, targeted regulation causes theleast productive firms to exit, which contributes to an industry’s clean-up.Targeted regulation results in firms exiting because the industry’s price index does not riseenough to offset the increased cost of production for the least productive firms. Note that thiseffect need not always occur as a result of regulation. For example, Anoulies (2017) finds that,for a given initial permit allocation rule, tightening the emissions cap used in a permit tradingsystem does not cause firms to exit. This is because the marginal producer is perfectly offset18by the increase in industry prices.9,10We now turn to examine the effects of targeted regulation on firm revenues. Before doingso, it is useful to note that in either regulatory regime, revenues for any firm (and thus profits)can be written as a monotonic function of the exit cut-off. To see this, note that in the no-regulation equilibrium, revenues at a firm using the business-as-usual technology are given byrnob (ϕ) = [1/ρϕ]1−σ IPσ−1, and revenues at a firm using the state-of-the-art technology aregiven by rnos (ϕ) = [1/ραϕ]1−σ IPσ−1. Using the fact that free entry implies rnob (ϕnoε ) = σ f ,we havernob (ϕ) =[ϕϕnoε]σ−1σ f (1.28)rnos (ϕ) =[αϕϕnoε]σ−1σ f . (1.29)Similarly, in the regulated equilibrium, revenues at a firm using the business-as-usual tech-nology are given by rtarb (ϕ) = [[1+κτ]/ρϕ]1−σ IPσ−1, revenues at a firm using a retrofittedtechnology are given by rtarr (ϕ) = [1/ρϕ]1−σ IPσ−1, and revenues at a firm using the state-of-the-art technology are given by rtars (ϕ) = [1/ραϕ]1−σ IPσ−1. Again using the fact that freeentry implies rtarb (ϕtarε ) = σ f , we havertarb (ϕ) =[ϕϕ tarε]σ−1σ f , (1.30)rtarr (ϕ) =[ϕϕ tarε]σ−1[1+κτ]σ−1σ f (1.31)rtars (ϕ) =[αϕϕ tarε]σ−1[1+κτ]σ−1σ f . (1.32)Comparing revenues from one regime to another leads to the next result.Proposition 2. Targeted regulation reduces revenues for surviving firms with ϕ < ϕ tarr .Proof. By definition, surviving firms with ϕ < ϕ tarr use the business-as-usual technology inboth the no regulation and regulation regimes. For these firms, the effect of regulation can be9Anoulies, however, also shows that the initial permit allocation rule does affect exit.10In contrast, Andersen (2018) finds that tightening an emission tax causes firms to exit. This result is drivenby Andersen’s assumption on the abatement process used by firms, as this results in the pollution tax affectingthe fixed cost of production.19determined by comparing equations (1.28) with (1.30), yieldingrtarb (ϕ)rnob (ϕ)=[ϕnoεϕ tarε]σ−1. (1.33)Given ϕ tarε /ϕnε > 1 (from Proposition 1), it follows that rtarb (ϕ)< rnob (ϕ).Proposition 3. Targeted regulation has an ambiguous effect on revenues for surviving firmswith ϕ ≥ ϕ tarr .Proof. See Appendix A.1.Corollary 1. If [1+τκ]> α , then targeted regulation causes the largest reduction in revenuefor firms with ϕ < ϕ tarr .Proof. See Appendix A.1.Proposition 2 and Proposition 3 show that targeted regulation has differential effects onrevenues for firms of different productivity levels. Corollary 1 makes it clear that, with the ex-ception of one special case11, the largest reduction in revenues occurs for the least productivesurviving firms in an industry. This is intuitive, as the policy’s design results in increased vari-able costs only for those firms that fail to adopt cleaner processes. As the adoption of cleanerprocesses requires paying a fixed cost, it is the least productive firms in an industry that facean increase in variable costs directly as a result of regulation. Revenues for these firms mustfall as a result of regulation.We next turn to examine the effects of targeted regulation on the emission intensity of afirm’s production. Recall that the effect of regulation on firm emission intensities is determinedby the adoption of new technology. As the adoption of new technology requires paying a fixedcost, similar to targeted regulation’s effect on firm revenues, its effect on a firm’s emissionintensity also depends on the firm’s productivity.Proposition 4. Targeted regulation reduces the emission intensity of a firm with ϕ ∈ [ϕ tarr ,ϕnos ],and does not change the emission intensity of other firms.Proof. In the no-regulation regime, no firm retrofits. However, in the regulation regime, a firmwith ϕ ∈ [ϕ tarr ,ϕ tars } adopts the retrofitted technology. By construction, the emission intensityof these firms falls, as etarr (ϕ)enb(ϕ)=etars (ϕ)enb(ϕ)= 1γ < 1. If fs >∆1+∆2∆2 fr and fr >[[1+ τκ]σ−1−1]f ,then a positive measure of firms retrofit in response to environmental regulation. These11This special case is discussed in the appendix.20inequalities are satisfied by assumption, meaning targeted regulation causes some firms toretrofit in response to regulation.Given our setup, the emission intensity of firms that use the business-as-usual technology inboth regimes is unaffected. Similarly, there is no change in emission intensity form firms thatdowngrade from state-of-the-art to retrofitted technology, or use state-of-the-art technology inboth regimes.Proposition 4 shows the process effect of regulation. For a targeted regulation, this processeffect is driven by firms in the middle of the productivity distribution, as they respond toregulation by adopting process changes. The least productive firms, in contrast, may contributeto a clean-up through the reallocation effect, as is made clear in Proposition 2.12We are not the first to show regulation may have differential effects on firms of differentproductivity levels. Cao et al. (2016), for example, find low productivity firms respond to auniform pollution tax by allocating inputs that could have been used in production towardsabatement. This relates to Corollary 1 above, as given the same amount of inputs, this reallo-cation of inputs would serve to reduce output. Cao et al. also find that more productive firmsadopt less polluting technology, which is similar to our result presented in Proposition 4.13Our results differ from Cao et al. in two important ways. First, Cao et al.’s results require im-posing a quasi-linear preference system, and does not hold if preferences are of the constantelasticity of substitution form, as is commonly assumed in the heterogeneous firms literature.Second, we show that under targeted regulation, the most productive firms in an industry arenot directly affected by regulation because they use relatively clean processes in the absenceof regulation. This is an important distinction between targeted regulation and regulation thatdirectly affects the variable costs of all producers, such as a uniform tax or a cap-and-tradesystem.Given the connection between technology adoption and the clean-up channels, a naturalquestion is to ask how changing the fixed costs of retrofitting affects these channels?Proposition 5. Reducing fr lowers ϕ tarr , and if fs is not too large, also lowers ϕ tarε .Proof. See Appendix A.1.Corollary 2. Reducing fr lowers the measure of firms that use the business-as-usual technol-ogy.12If more productive firms are less pollution intensive, then Proposition 2 and Corollary 1 imply a reallocationeffect.13For details, see Proposition 5 in Cao et al. (2016).21Proof. See Appendix A.1.Proposition 5 implies that a regulated equilibrium with a lower fixed cost of retrofittingwill have a larger mass of firms using clean processes (either the retrofitted or state-of-the-art technology), relative to a regulated equilibrium with a relatively high retrofitting cost. Inaddition, the low retrofitting cost equilibrium will also have less exit, if the fixed cost of state-of-the-art technology is not too large. Consequently, lowering the fixed cost of retrofittingincreases the magnitude of the process effect relative to the selection effect. Corollary 2 showsthat, in addition, lowering the retrofitting cost leads to a smaller measure of firms that usethe business-as-usual technology following regulation, who are the drivers of the reallocationeffect.1.4 DiscussionThis model makes a number of clear predictions about the channels through which targetedregulation may cause an industry to clean-up. First, regulation will cause the least productiveplants in an industry to exit. This occurs because in equilibrium, relatively low-productivityfirms become less profitable, which reduces the measure of firms for whom it is worth payingthe fixed cost of production. If, as we have assumed, more productive plants are less pollutionintensive, then these plants exiting will reduce the industry’s pollution intensity. That is, aselection effect induced by regulation will contribute to the industry’s clean-up.Second, the process effect arises from firms choosing to adopt cleaner production processesin response to regulation, which makes them less pollution intensive. However, only firms inthe middle of the industry’s productivity distribution should be induced by regulation to adoptcleaner processes. As the adoption of cleaner processes require paying a fixed cost, relativelylow-productivity firms will not be profitable enough to pay these fixed costs. Moreover, themost productive firms in an industry use industry-leading processes and technologies evenin the absence of regulation. That is, regulation does not affect their incentive to use theindustry-leading technologies, and does not cause them to become cleaner. Firms in the middleof the productivity distribution, however, are productive enough to warrant adopting cleanerprocesses, but not productive enough to do so absent regulation.Finally, the equilibrium firm size falls for the least productive firms that remain in theindustry. This occurs because adopting a cleaner production process, which allows the firmto avoid the pollution tax, carries a fixed cost. As a result, some low-productivity plants willfind it optimal to produce, but will not be profitable enough to pay the fixed cost to changeprocesses. Hence, in equilibrium, variable production costs rise for the least productive plants,22and their optimal size shrinks. Again, if more productive plants are less pollution intensive,then the reduction in output from these plants will reduce the industry’s pollution intensity.This is the reallocation effect through which regulation contributes to an industry’s clean-up.In addition to clarifying the channels involved in a clean-up, this model shows the impor-tant role the fixed cost of adopting cleaner processes play in a clean-up. These fixed costsdetermine the number of firms (or measure of firms, to be more precise in this context) thatadopt cleaner processes in response to regulation. If these fixed costs are very low, then a largenumber of firms should choose to adopt cleaner processes in response to regulation. This, inturn, reduces the number of firms that face higher production costs from regulation, and leadsto less exit and a smaller reduction in output from surviving low-productivity firms. As aresult, the fixed cost of process changes directly affects the magnitude of the process effectrelative to the reallocation and selection effects.1.5 ConclusionIn this chapter, we present a new theoretical framework to capture the channels throughwhich targeted environmental regulation – regulation that requires firms to adopt clean pro-cesses, and penalizes those that fail to do so – causes an industry to clean up. We start byreplicating a decomposition exercise by Cherniwchan et al. (2017) to show there are three dis-tinct channels through which environmental regulation can cause a reduction in an industry’spollution intensity. The first channel, which we call the selection channel, captures the exitof plants in response to regulation. The second channel, which we call the reallocation chan-nel, captures the reduction in output at surviving regulated plants in response to regulation.Lastly, the process channel captures the adoption of cleaner production processes at survivingregulated plants.We then present a theoretical model to show how these channels may arise in response totargeted environmental regulation. We study this type of regulation because it is a commonform of air pollutant regulation (for example, the US Clean Air Act contains this type of regu-lation), and has not yet been studied in the theoretical literature on the economic consequencesof environmental regulation.Our model is based on a closed-economy variant of the Melitz (2003) model in which pol-lution from heterogeneous firms is regulated. In this model, firms choose whether to produce,how much to produce, and what type of production process or technology to use. Unlike themajority of the relevant environmental economics literature, we assume firms upgrade to acleaner production technology by paying a fixed cost, rather than facing higher variable costs.We show that targeted regulation causes the least productive firms in an industry to exit,23as well as relatively low-productivity surviving firms to shrink. If these firms are the mostpollution intensive in the industry, then these results imply that targeted regulation will createboth selection and reallocation effects that serve to clean up the industry. In addition, wefind moderately productive surviving firms adopt cleaner processes in response to targetedregulation, which means the process effect also contributes to a clean-up. Finally, the mostproductive firms in an industry are not directly affected by targeted regulation, in terms ofpollution intensity or production, as they use industry-leading technology with or withoutregulation.24Chapter 2Estimating the Regulatory Channels ofthe Manufacturing Clean-Up2.1 IntroductionManufacturing pollution intensity – the amount of pollution emitted from the manufactur-ing sector per dollar of output shipped – has plummeted across much of the industrializedworld in recent decades. From 1990 to 2008, for example, the pollution intensity of the USmanufacturing sector fell by up to 77% for some pollutants (Levinson, 2015). Similar reduc-tions have been shown for the European manufacturing sector (Brunel, 2016), and in severalcountries for sulphur dioxide emissions (Grether et al., 2009). Taken together, it appears theproduction of goods in much of the industrialized world is becoming cleaner.Work on this manufacturing “clean-up”, spurred by Levinson’s (2009) seminal paper, hasconcluded that the source of this clean-up appears to be reductions in industry- (Brunel, 2016;Levinson, 2009), and even product-level (Shapiro and Walker, 2015) pollution intensity. Thatis, changes in the composition of industries and products produced in these countries do littleto explain the clean-up. One obvious hypothesis as to why industries have become cleaner isthat environmental regulations may have changed the way in which industries operate, therebypushing them to become less pollution-intensive. Given the primary goal of regulation is toaddress pollution problems, this hypothesis seems plausible. Yet, there is little direct evidenceof how regulation causes an industry to clean-up. The goal of this chapter is to contribute tothis debate by asking how plant-level responses to environmental regulation have contributedto the manufacturing clean-up.While there is little direct evidence of regulation’s role in the clean-up, there is considerable25indirect evidence that regulation is a potential, if not driving, cause. For example, Shapiro andWalker (2015) use a structural model to ask whether the clean-up of the U.S. manufacturingsector is consistent with changes in regulation, trade, productivity growth, or other economicfactors. They conclude the clean-up would require a doubling of their model’s shadow-priceof emissions, which is consistent with regulation playing a large role in the clean-up. Whilethis presents compelling evidence for the importance of regulation, it is not causal.There is also indirect evidence on regulation’s role in the clean-up from the literature fo-cused on estimating the causal effects of regulation. This literature has shown, for example,that regulation may cause plants to exit (e.g. Becker and Henderson (2000); Henderson (1996);List et al. (2003)), and surviving plants to shrink (e.g. Greenstone (2002)). If these displacedand contracted producers were relatively pollution intensive, then these changes would serveto clean up an industry. Other work in this area has shown regulation may affect the produc-tion processes at regulated plants, by altering input-use and productivity (e.g. Berman andBui (2001b), Greenstone et al. (2012), and Walker (2013)), for example. However, withoutconnecting these changes to a plant’s pollution intensity, it is unclear whether these would beviable mechanisms through which regulation could lead to a cleaner industry.In this chapter, we present causal estimates of the clean-up caused by a particular regu-latory change. To do this, we estimate the effect of a policy change on each of the threeplant-level determinants of an industry’s pollution intensity: the number of plants operatingin an industry (which we call the “selection” effect), changes in output at regulated plants(the “reallocation” effect), and changes in plant emission intensity (the “process” effect). Wethen develop a simple framework to translate these causal micro-level estimates to estimatesof the aggregate channels of the clean-up. Taken together, these estimates provide a completecharacterization of how plant-level responses to regulation have contributed to the aggregateclean-up of manufacturing.To obtain causal estimates of the regulatory channels of the clean-up requires observinglongitudinal information on both pollution emissions and productive activities of individualmanufacturing plants. We obtain this information from a newly created confidential Canadiandataset. This dataset is one of just a few such datasets in the world, and to the best of ourknowledge, provides the most complete coverage of a nation’s manufacturing sector of any ofthese datasets. We use this data to estimate the effect of a a major revision to Canadian airquality regulation - the implementation of the Canada Wide Standards for Particulate Matterand Ozone (CWS).In addition to providing, potentially, the best data to answer our research question, Canadais a good setting for this study because the resulting estimates should be informative for un-26derstanding the mechanisms through which environmental regulations have contributed to themanufacturing clean-ups elsewhere. First, the policy we study is similar in design to the mainair pollution regulation in the US, the Clean Air Act (CAA), and shares features of the airquality standards in place in Europe. While there is much evidence of the CAA’s effects onAmerican producers’, we are the first to exploit the CWS to identify the effects of environ-mental regulation.1 As well as having similar policy, Canada’s clean-up appears to be verysimilar to those that have been documented in other countries. As we show below using theindustry decomposition developed by Levinson (2009, 2015), total manufacturing emissionsof most air pollutants in Canada have fallen substantially, primarily because of reductions inthe emission intensity of individual industries.Our goal is to estimate the effects of the CWS on the pollution intensity, output, and entryand exit decisions of affected Canadian manufacturing plants. To do so, we exploit variationin regulatory stringency created by the design of the CWS. The CWS was designed to ensureeach region met a minimum level of air quality by establishing thresholds for the ambientconcentration of PM2.5 and O3. Regions in which the ambient concentrations of either pollu-tant exceeded the relevant threshold in a given year were subject to more stringent regulationrelative to other regions. In addition, these regulations were explicitly focused on plants in“targeted industries” that were viewed as primary contributors to poor air quality. As a result,plants in targeted industries and regions violating one of the CWS standards were subject tomore stringent environmental regulation.2 We identify the effects of regulation on these plantsusing a triple-difference research design that exploits the variation in regulatory stringencyacross time, region and industry. This allows us to control for factors such as localized reces-sions or industry demand shocks that would otherwise confound the effects of environmentalregulation.We find robust evidence that the CWS reduced pollution emissions from affected manu-facturing plants. For the average PM2.5 emitting plant, the CWS is associated with a 15%reduction in PM2.5 emissions. Furthermore, the CWS is associated with a 33% reduction inNOX emissions from the average NOX emitting plant.The theory outlined in Chapter 1 predicts that these reductions will be driven by differentmechanisms. If the fixed costs of process changes are high on average, as we argue is the casefor PM2.5, only relatively productive plants will adopt cleaner production processes followingregulation. As a result, in this case, the CWS should have little to no effect on the emission1The CWS has been discussed previously, however, in the environmental policy literature (see, e.g., Angle(2014)).2The annual permits required by plants to operate in each province were used to impose these regulations.We describe the CWS in more detail in Section 2.3.27intensity of the average plant. If the fixed costs of process changes are low on average, as weargue is the case with NOX, then even less productive plants should adopt cleaner productionprocesses.3 In this case, the emission intensity of the average plant should fall in response tothe CWS. Our empirical estimates support these predictions; we find little evidence that theCWS affected the emission intensity of the average regulated PM2.5 emitting plant, but findevidence it is associated with a 29% reduction in the emission intensity of the average affectedNOX emitting plant. Our estimates of the effects of the CWS on plant output also fit withthe predictions of our model; we find that the CWS was associated with a 11% reduction inoutput from the average affected PM2.5 emitting plant, but had little to no effect on the outputof the average NOX emitter. As predicted, we also find that the CWS was associated with asignificant reduction in the number of plants that emit PM2.5, but had no significant effect onthe entry and exit of plants that emit NOX.Taken together, these estimates suggest that environmental regulations contributed consid-erably to the clean-up of the Canadian manufacturing sector. To make this contribution ex-plicit, we develop an empirical analogue of the industry decomposition suggested by Cherni-wchan et al. (2017). This approach allows us to translate our point estimates into estimatesof the process, reallocation and selection effects induced by environmental regulation. Theseestimates suggest that the effects of the CWS explain close to 21% of the reduction in thePM2.5 intensity of Canadian manufacturing, but nearly 61% of the reduction in aggregateNOX intensity. Moreover, the mechanisms driving these responses vary starkly across pollu-tants; the PM2.5 clean-up was primarily driven by reallocation and selection effects, whereasthe clean-up of NOX was primarily due to process effects induced by regulation.The model presented in Chapter 1 suggests these differential responses to regulation aredue to differences in the fixed cost of adopting cleaner production processes across pollutants.While we have focused on this channel given the available evidence documenting the substan-tial differences in the average costs of becoming less PM2.5 and NOX intensive, we do notobserve these costs directly. Hence, to provide further evidence that our estimates are consis-tent with this mechanism, we also examine the heterogeneity in plant responses to regulation.The model suggests the effects of the CWS should only vary across plants of different produc-tivity levels if the fixed costs of adopting cleaner processes are high. We test this predictionby allowing the estimated effects of the CWS to differ across plants on the basis of their initiallabor productivity level.The resulting estimates match our model’s predictions. We find pollution from relatively3We present details on the relative costs associated with process changes for these two pollutants in Sec-tion 2.3.28low-productivity regulated PM2.5 plants fell primarily due to reductions in output, whereaspollution emissions from the moderately-productive PM2.5 plants fell due to a reduction inemission intensity. In contrast, NOX pollution intensity fell for both mid- and low-productivityplants. These results further suggest that our findings are driven by differences in the fixedcosts of adopting cleaner production processes.4Altogether, our findings contribute to a burgeoning literature examining the sources of theclean-up of the manufacturing sector. This research stems from the work of Levinson (2009)who examined how trade-induced changes in industrial composition have contributed to theclean-up of US manufacturing. Levinson finds that these changes played a small role; theclean-up is primarily due to reductions in industry emission intensity.5 Our work adds tothis body of research in two ways. First, we show that the aggregate trends that have beendocumented in the US (Levinson, 2009, 2015) and Europe (Brunel, 2016) extend to Canada.Second, we provide causal evidence of how air quality regulations have affected these trends.6Our work is also closely related to that of Martin et al. (2014) who estimate the effects ofa carbon tax on the exit, sales, and energy intensity of UK manufacturing plants. While thepotential clean-up of UK manufacturing is not the focus of their work, these factors serve asthe determinants of industry carbon dioxide intensity and, as such, their estimates could beused to understand how carbon taxes have affected the carbon intensity of the UK manufac-turing sector. Despite this, our work differs along three key dimensions. First, we observeplant-level pollution emissions, rather than energy use. Second, we show how plant level re-sponses such as those estimated by Martin et al. can be used to obtain estimates of the process,reallocation and selection effects. Third, we use insights from a stylized model to understandthe mechanisms that could be driving these effects.This chapter also relates to the set of empirical studies examining the effects of air qualityregulation on the emissions of manufacturing plants. Fowlie et al. (2012), for example, findSouthern California’s RECLAIM cap-and-trade program reduced NOX emissions from man-ufacturing plants. In addition, the U.S. Clean Air Act appears to have reduced both the growth4We also examine the effects of the CWS on several additional margins via which plants could respond toregulation, including changes in primary factor use, intermediate input use, and productivity. This allows us totest alternative explanations for why we observe different responses to the CWS across pollutants. As we showbelow, we find little evidence to support these explanations.5Others have argued trade may have caused changes to how plants produce their goods (by, for example,outsourcing some production or adopting new technologies), leading to a reduction in plant emission intensity(see Martin (2012) or Cherniwchan (2017)).6Our work is also related to that of Barrows and Ollivier (2018), who study how a potential mechanism thatcould be driving the process effect responds to changes in industry competition: within-plant changes in productmix. We are unable to explore this channel as our data does not contain information on pollution emitted byproduct line. Instead, we show direct evidence of a process effect in response to a regulatory change.29(Greenstone, 2003) and level (Gibson, 2016) of air pollutant emissions from manufacturingplants. This chapter complements this work by determining whether changes in plant pollu-tion in response to regulation are due to changes in the level of output produced, or changesin the emission intensity of production.Lastly, our work also relates to a large literature examining the effects of air quality regula-tion on various aspects of manufacturing plant operations. Our work is most closely related tothe papers that have provided preliminary evidence of the importance of selection and reallo-cation effects by either examining the effects of regulation on plant entry and exit (e.g. Beckerand Henderson (2000); Henderson (1996); List et al. (2003)) or plant output (e.g. Greenstone(2002)). We build on these earlier studies by also estimating the effects of air quality regula-tion on plant emission intensity, which allows us to provide the first estimates of the processeffects induced by regulation.7 We also build on this earlier work by showing that the effectsof environmental regulation may vary across plants of different productivity levels.The remainder of this chapter proceeds as follows. In Section 2.2, we document the clean-up of the Canadian manufacturing sector. Section 2.3 provides a brief overview of the CWS.Section 2.4 presents our data, outlines our research design and empirical specification, andpresents our empirical results. Finally, Section 2.5 concludes.2.2 The Clean-Up of Canadian ManufacturingOur goal in this chapter is to determine how the effects of environmental regulation on indi-vidual plants have contributed to the clean-up of manufacturing. While the clean-up has beendocumented in several countries, including the United States (e.g. Levinson (2009, 2015))and the European Union (e.g. Brunel (2016)), it has yet to be documented in Canada. Hence,we first examine whether the changes in the pollution emitted by the Canadian manufacturingsector mirror those that have occurred elsewhere.These trends, relative to 1992 levels, are illustrated in Figure 2.1. The figure depictschanges in the aggregate emissions of four common pollutants from the Canadian manufac-turing sector, as well as changes in aggregate manufacturing output. As it shows, the emissionintensity of Canada’s manufacturing sector has fallen since 1992. Overall, from 1992 to 2015real manufacturing output rose approximately 39%, while emissions fell by between 41% and70%, depending on pollutant. These estimates imply that, on average, the emission intensityof the Canadian manufacturing sector fell by 3.5-4.7% annually.This suggests the clean-up of Canadian manufacturing was similar to those that occurred7Other related work considers regulation’s effect on input use and productivity (e.g. Berman and Bui (2001b),Greenstone et al. (2012), and Walker (2013)), which are dimensions potentially related to the process effect.300.511.51992 1999 2007 2015YearProduction: SalesPollution: PM2.5 NOX VOC COFigure 2.1: Output and Pollution from Canadian Manufacturing: 1992-2015Notes: Figure depicts trends from 1992 to 2015 in real manufacturing sales and aggregate emissions of fine scale particulate mat-ter (PM2.5), nitrogen oxide (NOX), volatile organic compounds (VOCs), and carbon monoxide (CO). Aggregate pollution is fromEnvironment and Climate Change Canada’s Air Pollutant Emission Inventory. Aggregate output is measured as the real value of man-ufacturing shipments, constructed by deflating data on industry-level nominal shipment values from Statistics Canada’s CANSIMtable 304-0014 using the industry price data given in Statistics Canada’s CANSIM table 329-0077. All series are expressed relativeto their 1992 levels.in the U.S. and Europe. For example, Levinson (2015) finds the emission intensity of USmanufacturing fell by 3.6-4.3% annually from 1990 to 2008. Similarly, Brunel (2016) showsthe emission intensity of European manufacturing fell by 3.4-5.5% annually over the period1995-2008.While this evidence shows the magnitudes of the clean-ups in Canada, the US, and Europewere similar, it reveals little as to whether the potential sources were the same. As such, weadopt a simple decomposition exercise first used by Levinson (2009) to study the potentialsources of the clean-up. This approach allows us to determine if the observed reductions inaggregate emission intensity are driven by a “composition effect” created by a reallocationof economic activity from dirty emission-intensive industries to clean industries with rela-tively low emission intensities or by a “technique effect” created by reductions in the emissionintensity of individual industries.To make this decomposition explicit, let Z, X , and E = Z/X denote the pollution emissions,31Table 2.1: Canadian Manufacturing Emission Intensity Decomposition: 92-15∆ Emission Technique Composition TechniqueIntensity Effect Effect SharePollutant (1) (2) (3) (4)PM2.5 -79 -78 -1 0.99NOX -58 -52 -6 0.90VOCs -71 -67 -4 0.94CO -74 -73 -1 0.99Notes: Table reports estimates from a decomposition of the change in emission intensity of the Canadian manu-facturing sector from 1992 to 2015 into composition and technique effects. Estimates are from a Laspeyre’s-typeindex following Levinson (2015). Each row reports estimates for a different pollutant. The first column reportsthe percentage change in emission intensity from the manufacturing sector. The second and third columns reportthe reduction in aggregate emission intensity due the technique and composition effects, respectively. The fourthcolumn shows the fraction of column (1) attributable to changes in the technique effect, calculated as (column(2)/column (1)).output, and pollution intensity of the manufacturing sector, respectively. Let Zi, Xi, and Ei de-note the same for individual manufacturing industries8, indexed by i. Manufacturing emissionintensity can then be written as E = ∑iθiEi, where θi = Xi/X denotes industry i’s share ofoutput from the manufacturing sector. Totally differentiating yieldsdE =∑iEidθi+∑iθidEi. (2.1)The first term of equation (2.1) is the aforementioned composition effect, while the secondterm is the technique effect.We follow the approach taken by Levinson (2015) and take equation (2.1) directly to thedata. This gives us estimates of the reduction in manufacturing emission intensity attributableto both the composition and technique effects for PM2.5, NOX, VOCs, and CO over the period1992-2015. These estimates are reported in Table 2.1. The first column reports the change inemission intensity that occurred for manufacturing as a whole. The second and third columnsreport the change in aggregate emission intensity attributable to the technique effect and com-position effects, respectively.9 The final column reports the share of the emission intensitychange due to the technique effect.The estimates reported in Table 2.1 suggest that the clean-up of the Canadian manufac-8Due to constraints from the pollution data, our industry definitions correspond to either the three- or four-digit NAICS code.9The technique effect is calculated by taking the percentage change in a Laspeyre’s-type index of ∑i θidEi.The composition effect is calculated as the difference between the change in manufacturing emission intensityand the technique effect.32turing sector can primarily be attributed to the technique effect. For example, the estimatereported in the first row indicates that during the 1992-2015 period, changes in industry emis-sion intensity accounted for 99% of the reduction in manufacturing PM2.5 intensity. This isfurther evidence that the Canadian clean-up is similar to those observed elsewhere; as shownby Levinson (2009, 2015) and Brunel (2016), the clean-ups of US and European manufac-turing are also primarily due to the technique effect.10 As a reference, the technique effect’sshare in Canada is generally higher than in the US, but lower than in Europe.2.3 Air Quality Regulation in CanadaIn order to understand how environmental regulations contributed to the clean-up of Cana-dian manufacturing, we examine the effects of the Canada Wide Standards for ParticulateMatter and Ozone (CWS). The CWS was the primary policy targeting particulate matter andozone pollution throughout Canada over the period 2000-2012.11 Moreover, the design of theCWS makes it an attractive setting for studying the effects of environmental regulation.First signed in 2000, the CWS was an agreement between the federal government ofCanada and the various provincial environment ministries. The intent of the CWS was toimprove air quality across the country by the end of 2010 by implementing two air qualitystandards – one for PM2.5 and one for O3 – that applied to each major town or city in Canada(we call these Census Metropolitan Areas or CMAs).12 Much like the National Ambient AirQuality Standards at the centre of the U.S. Clean Air Act Amendments (CAAAs), these stan-dards created a target level of air quality that were to be achieved by each CMA in Canada.These standards were common across all CMAs, and each CMA was required to meet thestandards by the end of 2010. The standard for particulate matter required each CMA’s 24-hour PM2.5 concentration lie below 30µg/m3. Achievement of the PM2.5 standard was basedon the 98th percentile of each region’s 24-hour ambient concentration in a given year. TheO3 standard was applied as an 8-hour standard that required each CMA’s O3 concentrationlie below 65 parts per billion (ppb). Achievement of the O3 standard was based on the 4thhighest 8-hour concentration reported in a given year. Both the O3 and PM2.5 standards were10In addition, Shapiro and Walker (2015) perform a product-level decomposition, and find the clean-up in theUS is primarily due to within-product reductions in pollution intensity.11It was replaced with the Canadian Ambient Air Quality Standards for Fine Particulate Matter and Ozone in2012. We end our study period in 2010 to avoid any potential contamination by this regulatory change, as theplanning for this transition began in 2011.12The agreement defines a major town or city as a Census Agglomeration (CA) or Census Metropolitan Area(CMA). A CMA must have a total population of at least 100,000, while a CA must have a core population of atleast 10,000. We use the term CMA for both.33determined over the calendar year.13,14 These standards were intended to be distinct. Thatis, regions violating the PM2.5 standard were required to address their PM2.5 problem, andwere not required to improve O3 concentrations (unless they were also in violation of the O3standard).In addition to differentiating between regions on the basis of air quality, the CWS explicitlydesignated a set of “targeted industries” that were to be the focus of more stringent regulation.These targeted industries were pulp and paper, lumber and wood product manufacturing, elec-tric power generation, iron and steel manufacturing, base metal smelting, and the concrete andasphalt industries (Canadian Council of Ministers of the Environment, 2000b). These indus-tries were chosen because they were viewed as major contributors to the air quality problemsthat motivated the CWS, and were common across all CMAs.15The key regulatory approach used by the CWS was to subject regulated plants – those intargeted industries and CMAs with ambient concentrations of either PM2.5 or O3 in excess ofthe relevant standard’s threshold – to more stringent environmental regulation than other plantsin the country. While in principal the agreement allowed provinces to choose from a numberof different regulatory approaches, in practice provinces primarily used their annual operationpermit systems to regulate plants. In general, these provincial systems require plants to provecompliance with certain environmental regulations in order to operate in any year (see, e.g.Canadian Council of Ministers of the Environment (2006); Environment Canada (2002); En-vironment Canada and Forest Products Association of Canada (2004)). To address the CWS,facilities could effectively follow one of two paths to meet the permitting requirements: eitheradopt technical changes recommended to their industry in the CWS (Government of Canada,2003), or reduce activities contributing to the problematic pollutant. When local air qualitywas relatively clean (i.e. regions were in compliance with the CWS), the permitting constraintswere laxer than when air quality was poor. Consequently, regulatory stringency facing a plantvaried over time according to its region’s air quality.In essence, these regional air quality standards were used to trigger technology standardson regulated facilities, with the important caveat that facilities could choose to simply reducetheir polluting activity rather than adopt cleaner technology.16 This same approach was used13For details of the CWS, see Canadian Council of Ministers of the Environment (2000a).14For comparison, the National Ambient Air Quality Standards in the United States currently contain a 24-hour PM2.5 standard set at 35µg/m3, and an 8-hour O3 standard set at 70 pbb (Environmental Protection Agency,2016).15Some non-targeted industries were subject to other, more limited, forms of regulation. However, these werenot explicitly part of the CWS and did not feature the same regional variation as the CWS air quality standards.We describe the other relevant regulations in Appendix A.2.5.16In some instances explicit production constraints limiting the amount of polluting activity in a given calen-34for both the PM2.5 and O3 standards.While the same regulatory approach was in place for both pollutants, due to technicalconstraints the options available for plants to adopt cleaner processes appear to have differedconsiderably across pollutants. While a full exploration of the myriad of process changesavailable to plants is beyond the scope of this paper, we present evidence of these differencesas they pertain to manufacturing facilities, particularly in Canada. We present this evidence asboth our theory and empirical results suggest the size of these process costs play an importantrole in dictating the channels through which an industry cleans up.We first consider NOX, the main ozone precursor targeted by the O3 regulations, whichis primarily caused by the combustion of fossil fuels. Industrial facilities can reduce NOXemissions at a relatively low fixed-cost by adopting efficient combustion processes17 or byadopting low-NOx emissions burners (see, e.g. Environment Canada (2002), Canadian Coun-cil of Ministers of the Environment (1998b), or Environmental Protection Agency (1999a)).Alternatively, post-combustion processes, such as selective non-catalytic reduction and selec-tive catalytic reduction, can be installed at a relatively high fixed-cost.18 In the CWS context,the low cost process changes appear to have been sufficient to satisfy the policy’s constraints.This is explicitly noted in Canada’s federal emissions guidelines for both industrial boilersand heaters, and cement kilns, both of which were intended to provide the basis for processchanges mandated by the CWS. The former, for example, states the guidelines for industrialboilers “are based on proven compatibility with efficient combustion operation and the use ofcost-effective technology such as low-NOX burners” (Canadian Council of Ministers of theEnvironment, 1998b). In addition, the document further claims that a post-combustion controltechnology would be required “only in isolated cases” (Canadian Council of Ministers of theEnvironment, 1998b). The cement kiln guidelines also note that combustion modificationscan be achieved at lower costs than low-NOX burners, and both of these are considerably lesscostly than post-combustion processes. While combustion modification costs are not provided,annualized costs for installing post-combustion processes are listed as four to sixty times thecost of low-NOX burners (Canadian Council of Ministers of the Environment, 1998a).In contrast, industrial PM2.5 emissions are caused by a number of processes, several ofwhich could potentially occur at the same facility. These processes include, for example, thecombustion of fossil fuels, chemical reactions, wear and tear on machinery, and the process-ing of lumber. Similar to NOX, there are both low- and high-cost approaches to reducingdar year were imposed on facilities. These appear to have primarily required percentage reductions relative tobase-year production levels, and as a consequence likely varied across facilities of different sizes.17This may entail changing the temperature or fuel-oxygen ratio of combustion.18Note that these typically produce larger emissions reductions than other alternatives.35PM2.5 emissions. The low-cost approaches include fuel-switching, inertial separators, or wetscrubbers. These low-cost methods, however, can have limited applicability in industrial uses(World Bank Group, 1998a). Fuel switching primarily pertains to coal, and only affects theemissions from fuel use. Inertial separators are primarily intended for medium and coarseparticulate matter, and are not particularly effective for PM2.5. While wet scrubbers can ac-commodate PM2.5 emissions, they are intended for production that involves a wet process.Instead, for the typical industrial facility, reducing PM2.5 emissions requires installing a largefiltration system, such as a baghouse or electrostatic precipitator, that carries a relatively largefixed-cost. This was also explicitly noted in the context of the CWS (see, e.g., EnvironmentCanada (2002) and Environmental Protection Agency (1998, 1999b, 2002)).The available evidence, then, suggests that Canadian industrial facilities could accomplishNOX process changes at a much lower cost than the PM2.5 process changes. As a reference,engineering abatement cost estimates for relevant process changes, adjusted to a per-ton basis,are between $1,000 to $20,000 per ton of PM2.5 using an electrostatic precipitator, between$2,000 to $100,000 per ton of PM2.5 using a baghouse, and between $200 to $1,000 per tonNOX using a low-NOX burner (Environmental Protection Agency, 2006).19,202.3.1 Air Quality Improvements and the CWSBefore turning to our main empirical analysis, we first present descriptive evidence thatthe improvements in ambient air quality in Canada over the CWS period were consistent withthe design of the CWS. In particular, as the CWS regulated facilities in the dirtiest regions ofthe country, air quality improvements should have been most pronounced in those particularregions.To examine the changes in Canadian air quality over this period, we use ambient air qual-ity data from Environment and Climate Change Canada’s National Air Pollution SurveillanceProgram (NAPS). The NAPS is a network of 286 air quality monitoring stations located acrossCanada, and is Canada’s main source for air quality data. Each monitoring station is operatedby a provincial authority, and the federal environment ministry oversees the network. Hourlymonitor-level pollution concentration measures are available for ozone, most Criteria Air Con-taminants, and some heavy metals (for data, see: Environment and Climate Change Canada19No cost estimates are available for combustion modifications to reduce NOX emissions, but these should belower than that of the low-NOX burners (Canadian Council of Ministers of the Environment, 1998a).20It is worth noting that the process changes available to extremely large emitters, such as electric utilities,can differ considerably to the changes available to a typical industrial facility. Most notably, NOX abatementis primarily achieved through high-cost post-combustion methods at these large utilities (World Bank Group,1998b).36(2013)).We use this data to construct the regional air quality measures used by the CWS. For PM2.5,we construct the 98th percentile of each CMA’s 24-hour concentration in a given year.21 ForO3, we construct the 4th highest 8-hour concentration reported in a CMA in a given year.22 Forany CMA that contains more than one monitor, we follow the rule defined by the CWS andcompute the average pollution concentration across all monitors for the PM2.5 measurementsand the maximum concentration for the O3 measurements (Canadian Council of Ministers ofthe Environment, 2002, p. 12).With this data, we then sort each of the CMAs into one of two groups for each standard:“clean” CMAs that never violated the relevant standard, and “dirty” CMAs that violated thestandard at least once over the phase-in period. Doing this allows us to assess whether thechanges in air quality across Canada matched with the design of the CWS.Over the period 2000 to 2011, there was no significant change in mean PM2.5 concen-trations among the CMAs that never exceeded the PM2.5 standard. Similarly, there was nosignificant changes in mean O3 concentrations among the CMAs that never exceeded the O3standard. Mean PM2.5 concentrations in the clean CMAs was approximately 15 µg/m3 ineach year; mean O3 concentrations were between 55 and 58 ppb in each year. In contrast,mean PM2.5 in the dirty CMAs fell from approximately 30 µg/m3 in the beginning of thedecade to approximately 22 µg/m3 at the end. Similarly, mean O3 in the dirty CMAs fellfrom approximately 80 ppb in the beginning of the decade to approximately 68 ppb at the endof the phase-in.23 These changes in regional air quality are shown in Figure 2.4 and Figure 2.5,which plot the yearly mean pollution concentrations for the clean and dirty cities from 2000to 2011, the entire period of the CWS.21The 24-hour concentration is the 24-hour average taken from midnight to midnight for each day. Thiscalculation collapses the hourly data to the daily frequency.22For each monitor, running eight-hour averages are computed for each hour, and reported as the value asso-ciated with the last hour used in the calculation. That is, for January 1st, 2000, there is no reported value frommidnight to 7am, the 8am value is the average from midnight to 8am, the 9am value is the average from 1am to9am, etc.23Both changes are statistically significant at the 95% confidence level.3705101520253035Mean PM2.5 Concentration (µg/m3 )2000 2002 2004 2006 2008 2010YearCities Never Above PM2.5 Standard05101520253035Mean PM2.5 Concentration (µg/m3 )2000 2002 2004 2006 2008 2010YearCities Above PM2.5 StandardMean PM2.5 Concentration by YearFigure 2.2: Mean PM2.5 concentration by year with 95% confidence intervals. Panel A shows citiesnever above the PM2.5 standard. Panel B shows cities above the PM2.5 standard at leastonce. The red line represents the threshold for the PM2.5 Standard. The air quality metricused is the 98th percentile of each city’s 24-hour concentration.020406080Mean O3 Concentration (ppb)2000 2002 2004 2006 2008 2010YearCities Never Above O3 Standard020406080Mean O3 Concentration (ppb)2000 2002 2004 2006 2008 2010YearCities Above O3 StandardMean O3 Concentration by YearFigure 2.3: Mean O3 concentration by year with 95% confidence intervals. Panel A shows cities neverabove the O3 standard. Panel B shows cities above the O3 standard at least once. The redline represents the threshold for the O3 Standard. The air quality metric used is each city’s4th highest 8-hour concentration.38As a further check, we also examine how the distributions of CMA air quality changedfrom the first half (2000-2005) to the second half (2006-2011) of the decade. As the intent ofthe CWS was to reduce extreme measures of air pollution, the largest change in air quality overthis period should occur in the top of the air quality distribution. For this exercise, we againseparate CMAs into two groups: (i) clean CMAs where ambient pollution concentrations werenever above the CWS, and (ii) dirty CMAs that exceeded the CWS at least once. We estimatethese distributions using kernel density estimation with a Gaussian kernel.These distributions are depicted in Figure 2.4 and Figure 2.5. As the figures show, therewas almost no change in either of the PM2.5 or O3 distributions for clean CMAs over theentire phase-in period. The same, however, cannot be said for the dirty CMAs. The PM2.5distribution shifted drastically from the beginning to the end of the phase-in, with almost allof the CMA-year observations lying below the CWS threshold in the second half of the phase-in. The right tail of the O3 distribution shifted leftward, and the mass of CMA-years nearthe CWS threshold increased substantially. By the end of the phase-in period, most CMAs inCanada had met the PM2.5 standard, and the dirty O3 cities were moving towards compliance.This provides further evidence of changes in air quality consistent with the CWS.PM2.5 Standard0.02.04.06.08.1Density10 20 30PM2.5 Concentration (µg/m3)2000 to 20052006 to 2011Cities Never Above PM2.5 ThresholdPM2.5 Standard0.02.04.06.08Density0 10 20 30 40 50PM2.5 Concentration (µg/m3)2000 to 20052006 to 2011Cities Above PM2.5 ThresholdPDF of PM2.5 ConcentrationsFigure 2.4: Kernel density estimates of the distribution of PM2.5 concentrations across CMAs in the firsthalf (2000-2005) and second half (2006-2011) of the CWS phase in period. The panel onthe left displays pollution concentrations for CMAs that never exceeded the PM2.5 standard.The right panel displays the pollution concentrations for CMAs that exceeded the PM2.5standard at least once. The vertical red lines represents the threshold used for the PM2.5standard.39O3 Standard0.02.04.06Density40 45 50 55 60 65O3 Concentration (ppb)2000 to 20052006 to 2011Cities Never Above O3 ThresholdO3 Standard0.01.02.03Density20 40 60 80 100 120O3 Concentration (ppb)2000 to 20052006 to 2011Cities Above O3 ThresholdPDF of O3 ConcentrationsFigure 2.5: Kernel density estimates of the distribution of O3 concentrations across CMAs in the firsthalf (2000-2005) and second half (2006-2011) of the CWS phase in period. The panel on theleft displays pollution concentrations for CMAs that never exceeded the O3 standard. Theright panel displays the pollution concentrations for CMAs that exceeded the O3 standard atleast once. The vertical red lines represents the threshold used for the O3 standard.A potential concern with these figures is that they could merely reflect different trendsacross regions owing to other factors beyond the CWS. A primary concern is that the CMAsexceeding one of these thresholds may be more heavily populated or industrialized than thosebelow the threshold. To show this is not the driver of the documented change in air quality, inFigure 2.6 we show the change in mean PM2.5 (panel (a)) and O3 (panel (b)) concentrations forCMAs with a population of at least 300,000 people. As above, the panel on the left displayspollution concentrations of CMAs that never exceed the relevant standard, and the right showsconcentrations of the dirty CMAs. The figure shows a pronounced drop in both PM2.5 andO3 air quality for the heavily populated regions that exceed the respective thresholds, and nochange in air quality for the clean CMAs.400102030Mean PM2.5 Concentration (µg/m3 )Never Above PM2.5 Above PM2.598−02 03−07 08−12 98−02 03−07 08−12CMAs With Pop. >300kPM2.5 Concentrations Over Time(a) PM2.5 Concentrations020406080Mean O3 Concentration (ppb)Never Above O3 Above O398−02 03−07 08−12 98−02 03−07 08−12CMAs With Pop. >300kO3 Concentrations Over Time(b) O3 ConcentrationsFigure 2.6: Mean Pollution Concentrations by Year for Large CMAsNotes: Figure depicts mean PM2.5 (panel (a)) and O3 (panel (b)) concentrations by year for CMAs with apopulation of at least 300,000 people. For each pollutant, the panel on the left displays pollution concentrationsof CMAs that never violated the relevant standard. The right panel displays the pollution concentrations ofCMAs that violated the relevant standard at least once.While there are potentially other factors that may explain the changes in air quality shownabove, a proper treatment of the effect of the CWS on air quality is beyond the scope of thischapter. As our goal is to estimate the effects of the CWS on manufacturing plants, such aregional trend would only represent a potential identification problem if it were specific tothe industries targeted by the CWS. This is because the cross-industry variation in regulatorystringency allows us to flexibly control for CMA trends.2.4 EmpiricsIn Chapter 1, we presented a theoretical model of how a targeted environmental regulation,such as the CWS, may affect plants. This model provided a number of clear predictions as tohow facilities would respond to the CWS. Taken together, these results imply that when thefixed costs of abatement are high, environmental regulations should primarily reduce industryemission intensity via reallocation and selection effects. In contrast, when the fixed costs ofabatement are low, the industry clean-up should be driven by process effects. In this section,we explore those plant-level predictions empirically by estimating the CWS’ effect on plantpollution intensity, production, and exit. We use the resulting estimates to determine howthe process, reallocation, and selection effects created by the CWS have contributed to the41clean-up of Canadian manufacturing.2.4.1 Research DesignGiven that certain industries and regions were the primary focus of regulation, we identifythe causal effects of the CWS by measuring its effects on manufacturing plants that wereboth located in dirty CMAs and operating in a targeted industry. We do so by using a triple-difference research design that exploits the variation in CWS regulation across time, industriesand regions.24Our design begins by comparing the average outcomes of plants in regulated CMAs whileregulated (i.e. while violating one of the standards) to their average outcomes while unreg-ulated. This allows us to control for any unobserved time-invariant industry, CMA or plantcharacteristics that would affect plant pollution emissions. Moreover, in the absence of anyother shocks, this comparison would identify the average causal effect of the CWS on pollu-tion emissions. Yet, such absence is unlikely; there is strong reason to believe that a simplebefore-and-after comparison of affected plants could also capture the effects of regional, in-dustry, or aggregate economic shocks.25 We discuss each in turn.To address possible confounding regional shocks, we exploit the fact that each CMA con-tains manufacturing plants in both regulated and unregulated industries. This allows us to uti-lize the unregulated plants in a given CMA as a counterfactual for regulated plants in the samelocation. This will capture the effects of any unobserved time-varying provincial or CMA-level heterogeneity, such as changes in regional economic conditions or concurrent changes inprovincial policy that would otherwise confound the effects of the CWS.The simple before-and-after comparison could also be contaminated as a result of economicshocks that affect individual industries, which could arise due to the effects of increased for-eign competition created by international trade, or by revisions to federal policies that targetcertain sectors. To address these issues we exploit cross-CMA variation in regulation, and24It is worth mentioning that, while plants in dirty CMAs that were operating in a targeted industry weresubject to more strict regulation and enforcement, it is possible that other plants in the country were regulatedto some degree as a result of the CWS. If this is the case, then our research design produces estimates that givea lower bound on the CWS’ effects on the manufacturing sector. We describe these potential policies in theappendix (see appendix A.2.5).25Note that this raises an issue with identifying the effects of any provincial environmental regulation inCanada: who gets regulated and when are unlikely to be randomly assigned if left entirely up to regional au-thorities. The CWS allows us to overcome this concern by providing within-province variation in regulatorystringency. As a result, the CWS can be thought of as an instrument that allows us to identify the effects ofenvironmental regulation on a select group of plants: those that are regulated because they are in a CMA withair quality above one of the CWS standards. Adopting the language used in the treatment effect literature, theseplants are called compliers, and the CWS provides a local average treatment effect of environmental regulationfor these plants.42utilize the fact that in any particular industry, only plants in areas with poor air quality weresubject to stringent environmental policy. This allows us to use the average outcomes fromplants in a targeted industry in an unregulated CMA as a counterfactual for the average out-comes of plants from that industry that are located in a regulated CMA. This captures theeffects of any industry specific shocks.The cross-industry and cross-CMA variation in the stringency of environmental regula-tion also allows us to compare the average outcomes from regulated plants with the averageoutcomes from plants in non-targeted industries located in unregulated CMAs. These non-targeted plants in unregulated CMAs are not regulated under the CWS, and as such, capturethe underlying aggregate trend in pollution emissions. This allows us to control for country-wide shocks, such as aggregate technological change, changes in national policy, or changesin aggregate expenditure due to the 2008 recession.We estimate the effect of regulation on plant outcomes using the following equationypict = βPMT PMict +βO3TO3ict +ρp+ξct +λit + εpict , (2.2)where ypict is the natural log of the dependent variable of interest (pollution, sales, etc), at plantp, in industry i, located in CMA c, at time t.26 T jict is an indicator of treatment for standardj, and takes a value of one for plants that are in industries targeted by the CWS for years inwhich their CMA exceeds threshold j.Equation (2.2) also includes plant (ρp), CMA-year (ξct), industry-year27 (λit) fixed effectsand an error term (εpict). The plant fixed effects account for any unobserved plant-specificheterogeneity, as well as time-invariant industry and CMA characteristics. The CMA-yearfixed effects capture any region specific shocks. The industry-year fixed effects account forany industry-wide events. Finally, the error term captures idiosyncratic changes in outcomesacross plants.The coefficients of interest in Equation (2.2) are βPM and βO3. βPM measures the averagepercentage change in outcomes for plants affected by the particulate matter standard relativeto those that are not. Similarly, βO3 measures the average percentage change in outcomesfor plants affected by the ozone standard relative to those that are not. These coefficients are26We employ the natural log transformation to address the skewness in the distribution of each variable.27The CWS defined the targeted industries at the 3- or 4-digit North American Industry Classification System(NAICS) level.28 We create an industry indicator that corresponds to either the 3- or 4-digit NAICS level. All3-digit industries that contain targeted industries defined at the 4-digit level are grouped at the 4-digit level. Theremaining industries are grouped at the 3-digit level.43identified from within plant comparisons over time.29,30Changes in plant regulatory status must be plausibly exogenous for this research designto credibly identify the effects of the CWS. There are two reasons to believe this is the case.First, as with the CAAAs in the US, regulations are determined by a nationally set air qualitythreshold that does not vary over time. As a result, these standards are unrelated to differencesin local tastes, characteristics or economic conditions (Greenstone, 2002), an issue that couldarise with any region-specific policy (Besley and Case, 2000). Second, PM2.5 and O3 are ca-pable of being transported long distances by prevailing wind patterns, meaning that ambientpollution levels in Canada do not solely reflect local economic activity.31 As variation in re-gional air quality determines assignment to treatment, this means it is unlikely a plant couldmanipulate their regulatory status. Indeed, transboundary pollution from the US appears tohave been a concern to the federal government over this period. Shortly after the CWS wasdeveloped, Canada and the US signed an air quality agreement to address transboundary pollu-tion, Canada’s contribution to which involved ensuring the CWS was met (International JointCommission, 2002). The transboundary nature of pollution means it is unlikely a single plantcan directly manipulate their treatment status.2.4.2 Data and MeasurementOur analysis relies on a unique confidential micro-dataset that contains information on thePM2.5 and NOX emission intensity of Canadian manufacturing plants. This dataset was cre-ated by merging data from two existing sources: the National Pollutant Release Inventory(NPRI) and the Annual Survey of Manufactures (ASM).32 The NPRI contains information onthe emissions of various pollutants from Canadian manufacturing plants. By law, any facilitythat emits one of the covered pollutants above a minimum threshold must report to the NPRI.The ASM was used as Statistics Canada’s manufacturing census until 2012, and it provides29It is worth noting that regulatory enforcement is applied more stringently to plants that are in regions thatcurrently violate a standard, and that if a region’s air quality improves sufficiently, regulation will become lessstrict. As a result, the variation we are using is from plants in regions that cross one of the CWS thresholdsover our sample period. Over our sample, some of these plants move from regulated to unregulated status.This means if plants make changes to production processes that result in permanently lower emissions, then ourresearch design will underestimate the effects of the CWS. As our goal is to be conservative in assessing theeffects of the CWS, we view this as an acceptable trade-off.30We are able to separately estimate the effect of both standards because there are cities that exceed one, both,or none of the standards. Of all treated CMA-years in our sample, approximately 80% violated one (and onlyone) standard, while the remaining 20% violated both standards.31For evidence of how wind patterns shape ambient pollution concentrations in Canada, see, for example,Brankov et al. (2003) or Johnson et al. (2007).32This dataset was created through a collaboration between the Economics and Environmental Policy Re-search Network, Environment and Climate Change Canada, and Statistics Canada.44longitudinal information on plant sales, production costs, employment, and other plant char-acteristics for the majority of manufacturing plants in Canada.33 Plants in these two datasetswere linked by Statistics Canada, allowing us to create a longitudinal dataset containing in-formation on PM2.5 and NOX emission intensity as well as other plant characteristics over theperiod 2004-2010. Additional details on each data source and the construction of the datasetused in our analysis are given in Appendix A.2.1.While there are several additional datasets that contain linked plant-level pollution andproduction information, the NPRI-ASM has a number of advantages relative to these othersources. A commonly used approach to compile this type of data is to match publicly avail-able plant-level pollution data produced by the Environmental Protection Agency with theNational Establishment Time Series (NETS), a proprietary dataset with plant employmentand sales. For example, Cherniwchan (2017), Holladay (2016), and Cui et al. (2015) all usedthis type of data in their work on trade and pollution.34 A clear advantage of our data relativeto this proprietary dataset is that the NPRI-ASM is produced using Canada’s official manu-facturing census, which derives its core information from tax files and representative surveys.In contrast, concerns have been raised over how representative the NETS is of the actual uni-verse of U.S. manufacturing facilities (see, e.g., Barnatchez et al. (2017); Haltiwanger et al.(2013)).35Descriptive statistics for the key variables that we employ are reported in Table 2.2. Eachcolumn in Table 2.2 presents averages and standard deviations for a different sample corre-sponding to emitters of each pollutant. The first column corresponds to the set of plants thatemit PM2.5, the second column shows statistics for plants that emit NOX, and the final col-umn of the table reports summary statistics for the entire sample of plants in the ASM. Thestatistics in columns one and two are weighted to account for potential sample bias inducedby the linking procedure used to match plants across datasets (see Appendix A.2.1 for furtherdetails). Each sample is an unbalanced panel; the sample for PM2.5 contains 6501 plant-yearobservations, the sample for NOX contains 3012 plant-year observations, and the full samplecontains 309,541 plant-year observations.The summary statistics reported in Table 2.2 suggests that there are systematic differencesin plants that emit different types of pollutants. For example, on average, the NOX sample33The ASM was discontinued in 2012 and was replaced with a repeated cross-section survey.34In addition, (Tang et al., 2015) utilize a dataset produced by the National Bureau of Statistics of China thatcontains firm-, but not facility-, level pollution and production data, although this data is cross-sectional.35Others have used datasets that contain plant-level production and fuel-use data (Barrows and Ollivier, 2018;Martin et al., 2014). While fuel-use is an important input into pollution, having this data alone does not allowone to fully capture the process effect.45Table 2.2: Summary StatisticsPM2.5 NOX Full ASM(1) (2) (3)Emissions (tonnes) 25.83 262.14(103.43) (646.14)Sales ($1 mill.) 194.62 342.15 11.12(890.55) (1,305.95) (123.56)Value Added ($1 mill.) 62.46 102.11 4.29(241.82) (346.27) (34.34)Employment 280.11 382.03 35.69(634.85) (868.68) (125.27)VA/Worker ($1,000) 200.18 265.41 84.78(243.63) (297.06) (166.11)N 6501 3012 309541Notes: Table reports averages and standard deviations of key variables examined in the main analysis. Eachcolumn reports the summary statistics for a different sample. Column (1) is the sample of PM2.5 polluters,column (2) is the sample of NOX polluters, and the final column reports plant characteristics for the entiremanufacturing sector. Statistics in columns 1 and 2 are weighted to account for potential sample bias induced bythe match of the NPRI and ASM. All monetary values are reported in 2007 Canadian dollars.emitted more pollution, produced more output, had higher employment levels, and had higherlabour-productivity levels than the PM2.5 sample. This potentially reflects substantial differ-ences in how pollution is produced and abated, given that pollutants are typically producedby a few industries (Greenstone, 2002), and there are substantial differences in the fixed costsof abatement across pollutants (Canadian Council of Ministers of the Environment, 1998b;Environment Canada, 2002).Table 2.2 also shows that polluters represent the largest plants in the manufacturing sector.Relative to the full manufacturing sector, the sample of plants that emit either PM2.5 or NOXsell more goods (15 to 30 times on average), employ more workers (7 to 10 times), and havehigher value added per worker (2 to 3 times) than the average manufacturing plant.36 Thisis, in part, due to the reporting requirements for the NPRI; by law, plants only report if theyemit at least one covered pollutant above a minimum threshold level and employ at least 10individuals or operate an on-site generator (Environment and Climate Change Canada, 2016c).While this means we systematically exclude small facilities, our analysis covers plants thataccount for the majority of manufacturing pollution in Canada.3736This is still true when we consider medians instead of averages.37In addition, the majority of PM2.5 and NOX emitters use an on-site generator or boiler, which means the theemployment thresholds are likely not relevant for most of these plants.46Figure 2.7: Regulatory Status Changes under the CWSNotes: Figure depicts PM2.5 and O3 standard status changes for each CMA from 2000 to 2010. Red CMAs changed status under boththe PM2.5 and O3 standards. Orange CMAs only changed status for the PM2.5 standard. Yellow CMAs only changed status for theO3 standard. Green CMAs didn’t change status under either standard. The mainland United States is shown in light gray. Part of thenorthern Canadian Territories are trimmed for scale. The inset shows detail on the most densely populated area of Canada, colored inlight red on the main map.Determining Regulatory Status under the CWSOur analysis also requires determining which CMAs were affected by the CWS. To doso, we use the local air quality information described in Section 2.3.1. We use this data toconstruct CMA-level pollution concentration measures for each year in our sample, where themeasures computed are those associated with each standard.The variation in regulatory status created by changes in ambient air quality is illustratedin Figure 2.7, which shows the CMAs that changed regulatory status for the PM2.5 and O3standards. In Figure 2.7, the red CMAs changed status under both the PM2.5 and O3 standards,the orange CMAs only changed status for the PM2.5 standard, the yellow CMAs only changedstatus for the O3 standard, and the green CMAs didn’t change status under either standard.As the figure shows, there was substantial variation in which CMAs changed their regulatorystatus over the 2000-2010 period. Of the 149 CMAs in our sample, 23% changed status underthe PM2.5 standard, 26% changed status under the O3 standard, 11% changed status under47both standards, and 60% never changed regulatory status.2.4.3 Empirical ResultsThe CWS and Plant Pollution EmissionsWe begin our analysis by estimating the effects of the CWS on the level of pollution emittedby affected Canadian manufacturing plants.38 We start here for two reasons. First, it providessome indication as to the effectiveness of the CWS; if the regulations were responsible for thereduction in pollution levels documented in Section 2.2, then we should observe reductionsin the emissions of targeted pollutants as a result of the CWS. Second, this also providesus with a means to assess the external validity of our results. As we discussed above, thereis little evidence as to the effects of environmental regulation on the emission intensity ofmanufacturing plants. Focusing on pollution levels allows us to directly compare the effectsof the CWS with the effects of other environmental policies.Table 2.3 reports our estimates of the effects of the CWS on plant pollution emissions. Weestimate Equation (2.2) for two samples of plants. The first sample (in Panel A) are plants thatemit PM2.5, which is the main contributor to PM2.5 pollution. The second sample (in Panel B)are plants that emit NOX, which is the main contributor to O3 pollution.39 The first column ofeach panel reports estimates from a version of Equation (2.2) that only includes the particulatematter standard. Similarly, the second column reports estimates from a specification that onlyincludes the ozone standard. Finally, column (3) in each panel reports estimates from thespecification given in Equation (2.2). The first row in each panel reports the effect of thePM2.5 standard (βPM in Equation (2.2)); the second row shows the effect of the O3 standard(βO3 in Equation (2.2)). The dependent variable in each of these regressions is the natural logof plant pollution emissions for the relevant pollutant. Each regression is weighted to correctfor potential sample bias introduced by the procedure used to match plants in the NPRI withplants in the ASM.40 In all cases, standard errors clustered at the CMA-industry level arereported in parentheses.38A related, but distinct question, is to ask what the CWS did to regional air quality. While this is beyond thescope of this chapter, in the appendix we provide descriptive evidence that air quality improved in Canada overthis period (see Section 2.3.1).39There are other pollutants that may also contribute to PM2.5 and O3 pollution, including volatile organiccompounds and carbon monoxide. In the appendix we examine the CWS’ effects on the emissions of a numberof other pollutants (see Appendix A.2.3).40In brief, the potential bias happens because the probability of a successful match is positively correlatedwith a plant’s size. If the effects of the CWS vary by plant-size, then relying on the matched data would producebias estimates. Details on the weighting procedure used to address this can be found in Appendix A.2.1.48Table 2.3: The Effects of the CWS on Plant Pollution EmissionsPanel A: PM2.5 Panel B: NOX(1) (2) (3) (4) (5) (6)PM2.5 Standard -0.149∗∗ -0.151∗∗ 0.107 0.106(0.076) (0.076) (0.070) (0.069)O3 Standard -0.105 -0.113 -0.327∗ -0.325∗(0.164) (0.164) (0.183) (0.179)R2 0.175 0.175 0.175 0.310 0.311 0.311N 6501 6501 6501 3012 3012 3012Notes: Table reports estimates of the effects of the CWS on plant pollution emissions. Each panel reports results for a different sampleof emitters. Each column displays estimates from a different regression. In all cases, the dependent variable is the natural log of pollutionemissions. The first row reports the effects of the PM2.5 standard, and the second row reports the effects of the O3 standard. All regressionsinclude plant, industry-year and CMA-year fixed effects, and are weighted by the inverse of the match probability to control for potentialmatch-induced sample bias. Standard errors are clustered by CMA-industry. Asterisks denote significance at the 1% (***), 5% (**), and10% (*) levels, respectively.The estimates reported in Table 2.3 indicate that the CWS regulations led to statisticallysignificant reductions in the emissions of both particulate matter and nitrogen oxide from af-fected plants. Our baseline estimates for PM2.5, reported in column (3) of Panel A, indicatethat the CWS particulate matter regulations are associated with a 15.1% reduction in emis-sions from affected plants. Our baseline estimates for NOX, reported in column (6) of PanelB indicate that the ozone regulations are associated with 32.5% decrease in emissions fromaffected plants. The estimates reported in Panels A and B also show no statistically significantcross-effects of either standard. That is, O3 regulation did not significantly affect particulatematter emissions and PM regulation did not significantly affect NOX emissions.We view the results in Table 2.3 as an exploratory analysis of the CWS’ effects on plants.While the effect of O3 on NOX emitters is only marginally significant, we call attention tothese estimates because, as we show later in this section, the average effects of the CWS maskconsiderable heterogeneity across plants (see Section 2.4.3). Taken together, our evidencesuggests O3 regulation had a meaningful effect on manufacturing plants. Moreover, the theorypresented in Chapter 1 suggests an average treatment effect is not very illustrative of howplants respond to a policy such as the CWS.These results are consistent with the few existing estimates of the effects of air quality reg-ulation on pollution emissions from manufacturing plants. For example, Fowlie et al. (2012)find California’s NOX trading program reduced NOX emissions from regulated plants by be-tween 10% and 30% over the period 1990-2005. Similarly, Gibson (2016) finds that CleanAir Act regulation reduced PM emissions from regulated plants by 38% between 1987 and492014.41 This suggests that the CWS had similar effects on pollution levels as the environmen-tal policies enacted elsewhere.It is also worth noting that in Section 2.4.3 we present evidence that the estimates reportedin Table 2.3 are robust to a number of potential identification concerns, including the potentialfor pre-trends and the effects of a negative relationship between a CMA’s air quality and theproduction choices of the plants therein.The CWS and the Clean-up of ManufacturingHaving determined the CWS significantly affected plant pollution levels, we now turn toestimating the process, reallocation, and selection effects caused by the CWS. To do this, westart by estimating the effect of the CWS on the emission intensity, output, and exit of affectedmanufacturing plants. We then use these estimates to determine the implied contribution ofthe CWS to the clean-up of Canadian manufacturing.Plant-Level EstimatesIn Table 2.4 we report our estimates of the CWS’ effect on the emission intensity of man-ufacturing plants. As in Table 2.3, panel A shows estimates of Equation (2.2) for the sampleof plants that emit PM2.5 and panel B shows estimates for the NOX emitters. In each panel,we report estimates from two separate regressions each with a different measure of emissionintensity, as well as reproducing our baseline estimates of the CWS’ effects on plant pollutionlevels. The first column shows the CWS effect on pollution levels. In the second column, weshow the CWS’ effects on emission intensity, measured as the ratio of emissions to total plantshipments (sales), given this is the measure of output used previously in the literature docu-menting the manufacturing clean-up. In the third column, we measure emission intensity asthe ratio of emissions to value-added. Value added may provide a more accurate reflection ofthe level of productive activity that occurs in each plant (Cherniwchan et al., 2017). However,we focus on the estimates in the second column of each panel, as our goal is to contribute to aliterature that uses shipments as its measure of output.42 In both cases, the dependent variableis the natural log of emission intensity. The first row in each panel reports the effect of thePM2.5 regulation (βPM in Equation (2.2)) and the second row reports the effect of the O3 reg-ulation (βO3 in Equation (2.2)). As before, each regression is weighted to correct for potential41Greenstone (2003) also finds the US Clean Air Act regulation reduced the growth of particulate matter,lead, and VOC emissions from regulated plants by between 4% and 7% over the period 1987-1997.42In addition, value added may be less precisely reported in our context. This occurs because Statistics Canadais able to use corporate tax filings to check annual shipment amounts reported by plants, but cannot do so forvalue added.50Table 2.4: The Effects of the CWS on Plant Emission IntensityPanel A: PM2.5 Panel B: NOX(1) (2) (3) (4) (5) (6)PM2.5 PM2.5/Sales PM2.5/VA NOX NOX/Sales NOX/VAPM2.5 Std. -0.151∗∗ -0.043 -0.013 0.106 0.127 0.333∗∗∗(0.076) (0.096) (0.110) (0.069) (0.080) (0.098)O3 Std. -0.113 -0.169 -0.224 -0.325∗ -0.286∗ -0.200(0.164) (0.169) (0.189) (0.179) (0.153) (0.157)R2 0.175 0.161 0.156 0.311 0.281 0.260N 6501 6501 6501 3012 3012 3012Notes: Table reports estimates of the effects of the CWS on plant emission intensity for PM2.5 (panel A) and NOX (panel B) emitting plants.For each group of emitters, the first column reports estimates from a regression of the CWS regulations on the natural log of plant emissions.The second column shows the CWS’ effects on the plant emissions-sales ratio, while the third reports estimates from a regression of theregulations on the natural log of the emissions-value added ratio. In all cases, the first row reports the effects of PM2.5 regulations, andthe second row reports the effects of the O3 regulations. All regressions include plant, industry-year, and CMA-year fixed effects, and areweighted by the inverse of the NPRI-ASM match probability to control for potential sample bias. Standard errors clustered by CMA-industryare reported in parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.bias from matching the NPRI and ASM, while standard errors clustered by CMA-industry arereported in parentheses.The estimates reported in Panel A of Table 2.4 indicate PM2.5 regulation had little-to-no-effect on the emission intensity of plants that emitted PM2.5, with an estimated coefficientin column (2) that is relatively small and statistically insignificant. In contrast, the CWS O3regulations appear to have caused a significant reduction in NOX pollution intensity. Theestimate reported in column (5) of Table 2.4 indicate that the CWS ozone regulations areassociated with a 28.6% decrease in the level of NOX emitted per unit of output.43In addition, PM2.5 regulation caused a significant increase in NOX intensity measured invalue added terms. These results are driven by a very small number of plants that are regulatedby the PM2.5 standard and emit NOX, but not PM2.5. For these plants, PM2.5 regulation causeda large increase in NOX emissions and decrease in value added. We do not probe these findingsfurther, as they are driven by fewer than ten plants.44PM2.5 regulation caused a sizable reduction in plant PM2.5 emissions, but had no significanteffect on plant emission intensities. On the other hand, the O3 standard caused a large reduc-tion in NOX emissions in both levels and pollution intensity. This implies the PM2.5 standardmust have led to large decreases in output from affected plants, whereas the ozone standard43Though there are no existing estimates to which we can directly compare, Martin et al. (2014) show a carbontax levied in the United Kingdom led to an 18% drop in energy intensity at affected manufacturing plants.44Dropping these plants yields a point estimate of the PM2.5 regulation’s effect on NOX emissions of 0.052with a standard error of 0.073.51Table 2.5: The Effects of the CWS on Plant OutputPanel A: PM2.5 Panel B: NOX(1) (2) (3) (4)Sales Value Added Sales Value AddedPM2.5 Standard -0.108∗∗ -0.138∗∗ -0.022 -0.227∗∗∗(0.050) (0.065) (0.059) (0.083)O3 Standard 0.056 0.111 -0.039 -0.125(0.060) (0.070) (0.161) (0.188)R2 0.224 0.221 0.265 0.294N 6501 6501 3012 3012Notes: Table reports estimates of the effects of the CWS on plant output for PM2.5 and NOX emitting plants. For each panel, each columnreports the results of a different regression. In the first column, the dependent variable is the natural log of plant sales. In the second, thedependent variable is the natural log of plant value added. In each panel, the first row reports the the effects of PM2.5 standard, and thesecond row reports the effects of the O3 standard. All regressions include plant, industry-year and CMA-year fixed effects are weighted bythe inverse of the NPRI-ASM match probability to control for potential sample bias. Standard errors clustered by city-industry are reportedin parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.had relatively minor affects on output. We confirm these conclusions by directly estimatingEquation (2.2) on both sales and value added.Estimates of the effects of the CWS on plant output are given in Table 2.5 for both PM2.5(Panel A) and NOX (Panel B) emitters, with each panel reporting estimates from two separateregressions. In the first, we measure output as the value of total plant shipments (sales), and inthe second as value added. In both cases, the dependent variable is the natural log of output.The first row in each panel reports the effect of the PM2.5 regulation (βPM in Equation (2.2))and the second row reports the effect of the O3 regulation (βO3 in Equation (2.2)). As be-fore, each regression is weighted to correct for potential bias from the NPRI-ASM matchingprocedure, and standard errors clustered by CMA-industry are reported in parentheses.The estimates reported in Panel A of Table 2.5 confirm the PM2.5 standard led to a largedecrease in output from affected plants that emitted particulate matter. The estimate in column(1) of Panel A indicate the CWS particulate matter regulation is associated with a 10.8%decrease in sales from plants that emitted PM2.5. Conversely, the estimates in panel B showthe O3 standard had no statistically significant effects on output.45Lastly, we estimate a variant of our main specification (Equation (2.2)) in which we com-pare the number of plants operating in a treated industry-CMA-year cell to the number oper-45Note that PM2.5 regulation also caused a significant reduction in value-added from affected NOX emitters.As we discuss above, this is driven by a very small number of plants. Thus, we pay little attention to this result.52Table 2.6: The Effects of the CWS on Plant ExitPanel A: Emit PM Panel B: Emit NOXOLS Poisson OLS PoissonPM2.5 Std. -1.134∗∗ -0.347∗∗ -0.188 -0.031(0.626) (0.169) (0.293) (0.119)O3 Std. 0.726 0.142 -0.457 -0.135(0.547) (0.147) (0.489) (0.221)R2 0.481 0.365 0.443 0.207N 2776 3023 1252 1582Notes: Table reports estimates of the effects of the CWS on the number of plants operating in an industry-CMA-year. Panel A showsestimates using plants that emit particulate matter only, and Panel B shows estimates using plants that emit nitrogen oxide only. In eachpanel, the first column shows the results using OLS estimation and the second column shows results using Poisson estimation. In all cases,the first row reports the effects of PM2.5 regulations, and the second row reports the effects of the O3 regulations. All regressions includeindustry-year and CMA-year fixed effects. Standard errors clustered by CMA are reported in parentheses. Asterisks denote significance atthe 1% (***), 5% (**), and 10% (*) levels.ating in an untreated industry-CMA-year cell. That is, we estimate the following regressionNict = βPMT PMict +βO3TO3ict +αI(CWS)ic+ξct +λit + εict , (2.3)where Nict is the number of active plants in industry i in CMA c, Tjict is the treatment indicatorfor standard j (which takes a value of one for industries targeted by the CWS for years inwhich their CMA exceeds threshold j), I(CWS)ic is an indicator for whether the industry-CMA was ever regulated by the CWS, λit are industry-year fixed effects, ξct are CMA-yearfixed effects, and εict is an error term that captures idiosyncratic changes in outcomes acrossindustry-regions. The main coefficients of interest (βPM and βO3) show the net exit (or en-trance) of plants in an industry-CMA due to the CWS.As the dependent variable is a count variable, we estimate Equation (2.3) using both ordi-nary least squares and Poisson regression. As above, we report estimates for two groups ofplants: those that emit PM2.5 (Panel A) and those that emit NOX (Panel B). These results arepresented in Table 2.6, which includes standard errors clustered by CMA’s in parentheses.We find a significant reduction in the number of plants operating in an industry-region inresponse to particulate matter regulation. For example, the estimates in column (1) of PanelA show that PM2.5 regulation reduced the number of operating plants in the average affectedindustry-CMA by 1.134 plants. In contrast, O3 regulation had no significant effect on plantexit. This is consistent with the predictions of the theoretical model presented in Chapter 1, asabatement caries a high fixed cost for PM2.5 and a low fixed cost for NOX.53Aggregate ImplicationsThe main implication of the results presented in Table 2.4, Table 2.5, and Table 2.6 is thatthe CWS contributed to the manufacturing clean-up through different channels for differentpollutants. The particulate matter standard primarily caused a reduction in output at regulatedplants and plants to exit. In contrast, the ozone standard caused regulated plants to adoptcleaner processes. To quantify the total contribution of the CWS to the manufacturing clean-up we present a simple counterfactual exercise in which we ask how much of the clean-upcan be attributed to the process, reallocation and selection effects induced by the CWS. We dothis by using our estimates, paired with a decomposition of an industry’s emission intensity,to compute the implied change in manufacturing pollution intensity over our sample that oc-curred because of each of the CWS channels. We then compare these estimates to the observedchange in manufacturing pollution intensity.46For our decomposition, we follow an approach used in much of the labor literature andconsider total changes in emission intensity over time (for a relevant review, see Foster et al.(2001)). This exercise is, in effect, an extension of the decomposition presented in Section 2.2that documented how industry-level changes contribute to a change in a sector’s pollutionintensity. In contrast, this decomposition documents how plant-level changes contribute to achange in an industry’s pollution intensity. This decomposition can also be thought of as anempirical analogue to the decomposition in Section 1.2 of Chapter 1.To that end, let output and pollution from manufacturing industry i in year t be givenby Xit and Zit , respectively. We define an industry’s pollution intensity as the amount ofpollution emitted per unit of output produced, and let this be given by Eit = Zit/Xit . In addition,suppose each industry is composed of a continuum of plants and let xit(n), zit(n), and eit(n)denote output, pollution, and pollution intensity from plant n. Lastly, let λit(n) = xit(n)/Xit beplant n’s share of production in industry i and year t, and ni denote the marginal plant that isendogenously determined by the industry’s profitability.47 In this case, the emission intensityof industry i at time t can be expressed as Eit =∫ nit0 eit(n)λit(n)dn. Assuming, for convenience,that plants only exit the industry over time and never enter, then the change in an industry’semission intensity from t−1 to t is given byEit−Eit−1 =∫ nit0eit(n)λit(n)dn−∫ nit0eit−1(n)λit−1(n)dn−∫ nit−1niteit−1(n)λit−1(n)dn.46For simplicity, we will focus on the direct effects of each standard and ignore any cross-pollutant effects.That is, we ignore the PM standard’s effect on NOX emitters and the O3 standard’s effect on PM emitters.47As in Cherniwchan et al. (2017), we assume plants are ranked in reverse order of productivity. Conse-quently, selection removes the least productive plants.54In Appendix A.2.4, we show that the percentage change in an industry’s emission intensity,E˙it =Eit−Eit−1Eit−1 , can then be expressed asE˙it =∫ nit0szit−1(n)e˙it(n)dn+∫ nit0szit−1(n)λ˙it(n)dn−∫ nit−1nitszit−1(n)dn+∫ nit0szit−1(n)e˙it(n)λ˙it(n)dn,(2.4)where szit−1(n) is plant n’s share of industry i’s pollution at time t − 1, and dot notation isused to denote percentage changes. The first term on the right-hand side of equation (2.4) isthe “process effect”. This captures the change in industry emission intensity due to changesin plant emission intensity over time resulting from the adoption of new production processes.The second term on the right-hand side of equation (1.2), captures the effects changes in therelative size of plants within an industry over time, which we call the “reallocation effect”.The “selection effect”, given by the third term, captures the change in emission intensity cre-ated by a change in the set of plants operating within the industry over time owing to plantexit. The final term is an interaction effect created by the interaction between the process andreallocation effects.48We use our estimates presented above in Section 2.4.3 to construct the four terms on theleft-hand side of Equation (2.4). As such, let βˆe, βˆx, and βˆn denote our estimates of theeffects of the CWS on plant pollution intensity (from Table 2.4), plant output (from Table 2.5),and selection (from Table 2.6), respectively. Moreover, recall that, given our identificationassumptions, βˆe captures the average change in emission intensity due to the CWS, meaningthat we can writee˙it(n) =βˆe, if n is treated0, otherwise. (2.5)In addition, an estimate of λ˙it(n) and∫ nit−1nit szit−1(n)dn can be constructed from βˆx and βˆn,respectively. In the appendix (see Appendix A.2.4), we show thatλ˙it(n) =βˆx(1−sTreatxit−1)+sExitxit−11−sExitxit−1+βˆxsTreatxit−1, if n is treatedsExitxit−1−βˆxsTreatxit−11−sExitxit−1+βˆxsTreatxit−1, otherwise,(2.6)48Note that this can be thought of as the approximation error in the decomposition presented in Section 1.2 ofChapter 1 caused by focusing on small, rather than potentially large, changes.55where sTreatxit−1 and sExitxit−1 are the fraction of output in time t−1 from treated and exiting plants,respectively. Substituting Equation (2.5) and Equation (2.6) into Equation (2.4) gives es-timates of the process, reallocation, and interaction effects. Letting sTreatzit−1 be the share ofindustry i’s pollution in time t−1 from treated plants, then the process effect isP̂E = βˆesTreatzit−1. (2.7)Similarly, the reallocation effect is given byR̂E =sExitxit−1+ βˆx(sTreatzit−1− sTreatxit−1)1− sExitxit−1+ βˆxsTreatxit−1, (2.8)and the interaction effect is given byÎE = βˆesTreatzit−1[βˆx(1− sTreatxit−1)+ sExitxit−11− sExitxit−1+ βˆxsTreatxit−1](2.9)To construct an estimate of the selection effect, recall our estimate of βˆn tells us the numberof facilities that closed in an industry-CMA cell because of the CWS. Letting NTreat be thenumber of regulated industry-CMA cells, then the selection effect isŜE = βˆnNTreat s¯Exitzit−1, (2.10)where s¯Exitzit−1 is the average exiting plant’s share of industry i’s pollution in time t−1.In Table 2.7 we present our estimates of each of the CWS channels relative to the observedchange in manufacturing pollution intensity from 2004 to 2010. The first row shows the frac-tion of the PM2.5 clean-up due to the CWS and the second shows the fraction of the NOXclean-up due to the CWS. Our estimates of the process effect, reallocation effect, selectioneffect and interaction effect for each pollutant are reported in columns (1)-(4), respectively.Column (5) reports the implied change in manufacturing pollution intensity that can be ex-plained by the CWS.The results of this exercise show that both the PM2.5 and O3 standards enacted under theCWS played a considerable role in the clean-up of Canadian manufacturing. The estimatesin column (5) show that, from 2004 to 2010, the O3 standard is responsible for 61% of thereduction in manufacturing NOX intensity and the PM2.5 standard is responsible for 21% ofthe reduction in manufacturing PM2.5 intensity. However, the channels responsible variedconsiderably across pollutants. The process effect, for example, associated with NOX regula-56Table 2.7: Counterfactual EstimatesProcess Reallocation Selection Interaction TotalEffect Effect Effect Effect(1) (2) (3) (4) (5)PM2.5 0.034 0.109 0.073 -0.004 0.212NOX 0.409 0.140 0.085 -0.025 0.610Notes: Table reports the share of the total change in manufacturing pollution intensity from 2004 to 2010attributable to each CWS channel. The first row shows estimates for PM2.5 and the second row for NOX. Columns(1) through (4) show the estimates of each channel. Column (5) shows the total across all channels.tion accounts for almost 41% of the clean-up. In contrast, the process effect accounts for justover 3% of the clean-up for PM2.5. Instead, the PM2.5 regulation primarily reduced aggregateemission intensity through a combination of reallocation and selection effects.Explaining How Industries Clean UpThe results presented above show that the channels through which the CWS caused themanufacturing sector to clean up varied across pollutants. The theoretical model presentedin Chapter 1 provides a potential explanation for this: differences in the fixed costs of adopt-ing cleaner production processes across pollutants. Indeed, as we discussed in Section 2.3,engineering assessments of these pollutants argue abatement of NOX can be accomplished atlow-cost, while abatement of PM2.5 pollution typically requires large fixed costs (CanadianCouncil of Ministers of the Environment, 1998b; Environment Canada, 2002; EnvironmentalProtection Agency, 1999a). We now turn to assess this mechanism further and examine otherpotential explanations for our findings.Differential Effects by Plant Productivity LevelWe begin by testing the theory’s prediction that there should be large differences acrossplants in how they respond to regulation when abatement fixed costs are high, but that theresponses should be relatively uniform when fixed costs are low. As we cannot observe thefixed costs of abatement directly, this is the most direct test of our hypothesized mechanism.To test this prediction, we use an approach similar in spirit to that of Bustos (2011) andallow the effects of the CWS to differ across plants on the basis of their initial productivitylevel. An obvious limitation with our data is that we do not observe a plant’s capital stockinformation, thereby preventing us from being able to estimate plant total factor productiv-ity. Instead, we use value added per worker (that is, labor productivity) as our productivitymeasure. As productivity is potentially affected by regulation, we use a plant’s initial labor57productivity in the first year they enter our dataset to avoid contamination with the CWS. Toaccount for potential differences in average productivity levels across industries and time, weregress these plant initial productivity levels on entry-year and industry fixed effects, and usethe residuals from this regression as our measure of plant productivity. For consistency, weuse the same industry definitions we employed in constructing our industry fixed-effects.49To avoid functional form assumptions, we use a non-parametric approach and assign plantsinto one of three productivity bins, according to where they lie on the productivity distribution.We then interact these bins with the CWS treatment indicators. The bins are constructed byfirst sorting plants in each sample into terciles based on their residual initial labor productivity.These bins are then used to create indicators Qq, which indicate whether a plant is in thebottom- (Q1), middle- (Q2), or top- (Q3) third of the productivity distribution. Note thatbecause we construct these bins separately for PM2.5 and NOX emitters, the composition ofplants in each tercile may vary across each pollutant sample.We accomplish this by estimating the following regressionYpict =3∑q=1βQqPM[TPMict ×Qq]+3∑q=1βQqO3 [TO3ict ×Qq]+ρp+ξct +λit + εpict , (2.11)where Qq is the indicator that takes the value one if plants that are in productivity tercile q, Tjicttakes a value of one for all plants in targeted industries for years in which their CMA violatesstandard j, βQqj is the treatment effect of standard j on plants in productivity tercile q, and theremaining variables are as defined for Equation (2.2).We use Equation (2.11) to examine the CWS’ effects on plant pollution levels, emissionintensity, and sales. Examining pollution levels allows us to assess whether the CWS affectedemissions from plants of all productivity levels, whereas examining emission intensity andsales allows us to quantify the channels by which regulation affected each plant.These results are shown in Table 2.8. Panel A reports our estimates for PM2.5 emitters;Panel B for NOX emitters. In each panel, we report estimates from three separate regressions.The first column in each panel shows our estimates from Equation (2.11) on plant emissions,the second column shows the effects on plant emissions per dollar of sales, and the third onplant sales. Natural logarithms are taken of all dependent variables. The first three rows ineach panel report the effects of the PM2.5 regulation (the βQqPM coefficients in Equation (2.11)).The first row shows the effect on plants in the lowest productivity tercile, the second row theeffect on plants in the middle tercile, and the third row the effects on plants in the highest49See Section 2.4.1 for more details on the industry definitions employed.58Table 2.8: The Effects of the CWS by Plant Productivity LevelPanel A: PM2.5 Panel B: NOX(1) (2) (3) (4) (5) (6)PM2.5 PM2.5/ Sales NOX NOX / SalesSales SalesPM2.5 Std.x Q1 -0.163∗∗ 0.038 -0.201∗∗∗ 0.079 0.084 -0.005(0.083) (0.102) (0.073) (0.091) (0.118) (0.084)x Q2 -0.279∗∗ -0.251∗ -0.028 0.155 0.188 -0.032(0.134) (0.143) (0.056) (0.120) (0.126) (0.096)x Q3 -0.023 -0.016 -0.007 0.079 0.109 -0.030(0.100) (0.101) (0.057) (0.134) (0.134) (0.056)O3 Std.x Q1 -0.281 -0.353 0.072 -0.457∗∗ -0.412∗∗ -0.045(0.210) (0.222) (0.074) (0.207) (0.207) (0.205)x Q2 0.076 0.065 0.011 -0.340∗∗ -0.277∗ -0.063(0.195) (0.227) (0.130) (0.173) (0.160) (0.056)x Q3 -0.093 -0.150 0.057 -0.183 -0.182 -0.001(0.237) (0.232) (0.071) (0.177) (0.180) (0.167)R2 0.176 0.162 0.226 0.312 0.282 0.266N 6501 6501 6501 3012 3012 3012Notes: Table reports estimates of the effects of the CWS where the estimated treatment effects are allowed to vary by plant initial productivitylevel. Panel A shows the effects on PM2.5 emitters and Panel B on NOX emitters. For each panel, the first column reports estimates froma regression of the CWS regulations on the natural log of plant emissions, the second column shows estimates on the natural logarithmof the emissions-sales ratio, and the third shows estimates on the natural logarithm of plant sales. In all cases, the first row reports theeffects of PM2.5 regulations for plants in the bottom tercile of their industry’s productivity distribution. The second row shows the effectsof PM2.5 regulations for plants in the middle tercile of their industry’s productivity distribution. The third row shows the effects of PM2.5regulations for plants in the top tercile of their industry’s productivity distribution. Rows four through six show similar estimates for the O3regulations. All regressions include plant, industry-year, and CMA-year fixed effects, and are weighted by the inverse of the NPRI-ASMmatch probability to control for potential sample bias. Standard errors clustered by CMA-industry are reported in parentheses. Asterisksdenote significance at the 1% (***), 5% (**), and 10% (*) levels.tercile. Similarly, the final three rows report the effects of the O3 regulation (βQqO3 in Equa-tion (2.11)). The fourth row shows the effect on plants in the lowest productivity tercile, thefifth row the effect on plants in the middle tercile, and the sixth row the effects on plants in thehighest tercile. As before, each regression is weighted to correct for potential bias from theNPRI-ASM matching procedure. In all cases, standard errors clustered by CMA-industry arereported in parentheses.The results for the PM2.5 standard show stark differences across PM2.5 plants of differentproductivity levels. PM2.5 regulation caused a drop in emissions among the bottom two-thirdsof the productivity distribution, with a reduction in emissions of 16.3% for low productivityplants and 27.9% for middle productivity PM2.5 emitters. In contrast, PM2.5 regulation had59no significant effect on the most productive PM2.5 plants, suggesting they were unaffected byregulation.The results in columns (2) and (3) indicate how the affected PM2.5 polluters reduced theiremissions varied considerably across the productivity distribution. The drop in emissionsamong the middle-productivity plants was almost entirely driven by a drop in plant emis-sion intensity, with pollution intensity falling by 25.1%. The drop in emissions from low-productivity plants was driven by a reduction in output, with no significant change in pollu-tion intensity and a 20.1% drop in output. These findings suggest changes in plant pollutionintensity driven by regulation played a role in the particulate matter clean-up, however, onlyamong relatively productive plants.In contrast to the effects of PM2.5 regulation, the O3 standard had relatively uniform effectsacross NOX emitters. NOX emissions fell considerably across the entire productivity dis-tribution, with estimated reductions of between 18-46%, though not significant for the mostproductive plants. The NOX clean-up in response to the CWS was primarily driven by changesin plant production techniques, as plant-level changes in emission intensities explain 80-100%of the reduction in emissions.The results in Table 2.8 are consistent with our theory, and the hypothesis that the channelsof the CWS clean-ups varied across pollutants because of differences in abatement costs.50 Asthe abatement of PM2.5 requires paying a relatively high fixed cost, only relatively productiveplants should choose to do so. These highly productive plants, in turn, experience a reductionin pollution intensity with a relatively small change in output and production inputs. The lessproductive plants, on the other hand, experience an increase in production costs, leading to areduction in input use, output, and productivity. In contrast, as NOX can be abated at a rela-tively low cost, there are smaller differences across plants of different productivity levels. Forboth pollutants, the most productive plants in an industry use state of the art technology, andare thus unaffected by the CWS.Other Margins of Plant AdjustmentLastly, we examine the effects of the CWS on several additional margins of plant adjust-ment, including changes in primary inputs, intermediate inputs, and productivity. Doing soallows us to examine a number of alternative explanations as to why the PM2.5 and O3 stan-dards caused the manufacturing sector to clean up through different channels.50This conclusion holds even if we consider alternative specifications in which we split the productivitydistribution into quartiles or quintiles, or use a quadratic interaction of plant productivity with the treatmentindicators.60Thus far, the hypothesis we have focused on is that PM2.5 and NOX have different abate-ment costs, which affects a plant’s willingness to adopt cleaner production processes. Analternative hypothesis is that the opportunities for input substitution may vary across pollu-tants. For example, there could be readily available alternatives to the inputs that create NOXpollution, but not for the inputs that create PM2.5 pollution. If this were the case, then regula-tion would reduce NOX intensity but not PM2.5 intensity.Examining the effect of the CWS on input use allows us to asses the above hypothesis. Ifthis hypothesis were true, then the CWS should have caused an increase in spending on inputsfor NOX emitters.51 In addition, examining the effect of the CWS on input use for plants ofdifferent productivity levels allows us to indirectly test our main hypothesis. While our modeldoes not contain intermediate inputs, their use should be positively correlated with output.As our model predicts a reduction in output only for the least productive PM2.5 emitters, thisshould also be accompanied by a reduction in spending on intermediate inputs for these less-productive plants.The literature on the Porter Hypothesis provides an additional alternative hypothesis. Thisliterature posits environmental regulation could cause an increase in innovative activities andproductivity among regulated firms.52 If the average plant became less productive in responseto PM2.5 regulation, but more productive in response to NOX regulation, then this could gen-erate the findings reported in Section 2.4.3. Examining the effect of the CWS on plant pro-ductivity allows us to test this hypothesis.We examine these alternative hypotheses using data on the total number of plant employ-ees53, spending on both production materials and fuel and energy, value added per worker, andthe probability a plant is involved in research and development.Estimates of the effects of the CWS on productivity and input use for the average manu-facturing plant are shown in Table 2.9. Panel A shows estimates of Equation (2.2) for PM2.5emitters and Panel B shows estimates for NOX emitters. In each panel, we report estimatesfrom five separate regressions corresponding to the different mechanisms of interest. Natu-ral logarithms are taken of the dependent variables in columns one to four. The first columnshows the CWS’ effects on employment, the second spending on materials, the third spendingon energy, and the fourth labour productivity. The final column estimates the CWS effect on51Here we have assumed plants would use the cheapest input in the absence of regulation.52For a recent review of this literature, see Ambec et al. (2013)53Although we do not observe plant capital stock information, given our relatively short period of study weexpect capital adjustment to play a minor role in this context. While capital adjustment could play an importantrole over larger time horizons, the existing literature seems to find limited evidence of capital stock adjustmentsin response to environmental regulation. See, e.g., Greenstone (2002) and Levinson (1996).61Table 2.9: Other Margins of Plant AdjustmentPanel A: PM2.5Prim. Inputs Inter. Inputs Productivity(1) (2) (3) (4) (5)Employment Materials Energy VA/Worker Pr(R&D)PM 2.5 Standard -0.040 -0.119∗ -0.086 -0.098 0.033(0.064) (0.064) (0.056) (0.073) (0.040)O3 Standard 0.071 -0.008 0.224∗∗ 0.039 -0.086(0.068) (0.071) (0.108) (0.060) (0.060)R2 0.188 0.218 0.151 0.185 0.155N 6501 6499 6478 6501 6501Panel B: NOXPrim. Inputs Inter. Inputs Productivity(1) (2) (3) (4) (5)Employment Materials Energy VA/Worker Pr(R&D)PM 2.5 Standard 0.003 0.039 -0.094 -0.231∗∗∗ 0.061(0.069) (0.077) (0.093) (0.085) (0.060)O3 Standard -0.064 -0.069 0.085 -0.062 -0.143(0.157) (0.154) (0.264) (0.117) (0.119)R2 0.285 0.276 0.218 0.242 0.248N 3012 3012 3009 3012 3012Notes: Table reports estimates of the effects of the CWS on additional margins of adjustment for plants that emit either PM2.5 or NOX . Foreach group of emitters, each column shows the results of a different regression. The first column reports estimates from a regression of theCWS regulations on the natural log of the number of workers employed at the plant. The second and third columns report estimates of theCWS’ effects on the natural log of spending on production materials and fuel and energy, respectively. The fourth column reports estimatesof the CWS’ effects on the natural log of value added per worker. The final column reports estimates of the CWS’ effects on an indicatorfor whether the plant spends money on research and development, using a linear probability model. In all cases, the first row reports theeffects of PM2.5 regulations, and the second row reports the effects of the O3 regulations. All regressions include plant, industry-year, andCMA-year fixed effects, and are weighted by the inverse of the NPRI-ASM match probability to control for potential sample bias. Standarderrors clustered by CMA-industry are reported in parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.an indicator for whether the plant is involved in research and development using a linear prob-ability model. In each specification, the first row reports the effect of the PM2.5 regulation andthe second row reports the effect of the O3 regulation. As before, each regression is weightedto correct for potential bias from the NPRI-ASM matching procedure. In all cases, standarderrors clustered by CMA-industry are reported in parentheses.We also examine if the effects of the CWS on productivity and input use differ across theinitial plant productivity distribution. These estimates are reported in Table 2.10.54 Panel54The effects on R&D are omitted, but are available upon request.62A shows the results for PM2.5 emitters and Panel B for NOX emitters. Each column in eachpanel corresponds to a different dependent variable, each measured in natural logarithms. Eachregression is weighted to correct for potential bias from the NPRI-ASM matching procedure.In all cases, standard errors clustered by CMA-industry are reported in parentheses.As the estimates reported in Table 2.9 and Table 2.10 show, the main channels by whichthe average PM2.5 emitting plant responded to PM2.5 regulation appears to be through changesin intermediate input use and labor productivity. PM2.5 regulation decreased spending onproduction materials by 11.9%, caused a drop in energy spending (although not significantat conventional levels), and reduced labor productivity (also not significant at conventionallevels). PM2.5 regulation also caused a significant reduction in labor productivity among NOXemitters. There is no evidence of a change in employment or R&D propensity in response tothe PM2.5 standard.The estimates of the effects of the PM2.5 standard by productivity level are also consistentwith our main hypothesis. These results show that the reductions in materials, energy inputs,and labor productivity in response to the PM2.5 standard were driven by the least productiveplants. In response to PM2.5 regulation, the least productive plants reduced spending on mate-rial inputs by 19.4% and energy inputs by 12.5%, and value added per worker fell by 24.7%.PM2.5 regulation had no significant effect on these mechanisms at relatively more productiveplants. Interestingly, PM2.5 regulation had no significant effect on employment for the leastproductive plants, but reduced employment among the middle-productivity plants. Thoughoutput did not fall for the middle-productivity plants, regulation appears to have made themless labor-intensive, in addition to causing them to adopt cleaner production processes. A po-tential explanation for this is that the PM2.5 process changes may have required new capitalinvestments, thereby changing the plants’ capital-labor ratio. Finally, the drop in productiv-ity among NOX emitters in response to the PM2.5 standard appears to be driven by relativelyless-productive plants.The estimates reported in Table 2.9 and Table 2.10 also suggest O3 regulation did not havea significant effect on input use, employment, labor productivity, or R&D propensity at theaverage affected plant. The exception to this is an increase in energy spending among PM2.5emitters. Allowing the effects of the CWS to vary across plant productivity levels, we stillfind no significant effect on NOX emitter employment, input spending, or labor productivity.These results are inconsistent with the two additional hypotheses described above, as neitherproductivity nor input spending rise in response to regulation, which further suggests ourresults are driven by the fixed costs of abatement.63Table 2.10: CWS Mechanisms by Plant Productivity LevelPanel A: PM2.5 Panel B: NOX(1) (2) (3) (4) (5) (6) (7) (8)Emp. Materials Energy VA/ Emp. Materials Energy VA/Worker WorkerPM2.5 Std.x Q1 0.003 -0.194∗∗ -0.125∗ -0.247∗∗ 0.141 0.165∗ -0.080 -0.418∗∗∗(0.113) (0.094) (0.074) (0.119) (0.097) (0.099) (0.122) (0.124)x Q2 -0.093∗ -0.044 -0.051 0.058 -0.031 -0.007 -0.189 -0.188∗∗(0.055) (0.070) (0.073) (0.061) (0.079) (0.094) (0.146) (0.092)x Q3 -0.065 -0.049 -0.041 0.027 -0.116 -0.049 -0.044 -0.076(0.072) (0.095) (0.098) (0.089) (0.100) (0.122) (0.102) (0.111)O3 Std.x Q1 0.131 0.004 0.311∗∗ -0.058 -0.079 -0.004 -0.181 -0.177(0.086) (0.092) (0.128) (0.078) (0.200) (0.194) (0.304) (0.156)x Q2 0.024 -0.031 0.252 0.014 -0.057 -0.108 0.190 0.011(0.128) (0.149) (0.243) (0.099) (0.163) (0.166) (0.257) (0.157)x Q3 0.047 -0.016 0.109 0.136 -0.010 -0.053 0.237 -0.085(0.076) (0.082) (0.142) (0.092) (0.170) (0.159) (0.279) (0.120)R2 0.189 0.219 0.152 0.188 0.288 0.277 0.220 0.245N 6501 6499 6478 6501 3012 3012 3009 3012Notes: Table reports estimates of the effects of the CWS where the estimated treatment effects are allowed to vary by plant initial productivity level. Panel A shows theeffects on PM2.5 emitters and Panel B on NOX emitters. For each group of emitters, each column shows the results of a different regression. The first column reportsestimates from a regression of the CWS regulations on the natural log of the number of workers employed at the plant. The second and third columns report estimates ofthe CWS’ effects on the natural log of spending on production materials and energy, respectively. The final column reports estimates of the CWS’ effects on the naturallogarithm of value added per worker. In all cases, the first row reports the effects of PM2.5 regulations for plants in the bottom tercile of their industry’s productivitydistribution. The second row shows the effects of PM2.5 regulations for plants in the middle tercile of their industry’s productivity distribution. The third row shows theeffects of PM2.5 regulations for plants in the top tercile of their industry’s productivity distribution. Rows four through six show similar estimates for the O3 regulations.All regressions include plant, industry-year, and CMA-year fixed effects, and are weighted by the inverse of the NPRI-ASM match probability to control for potentialsample bias. Standard errors clustered by CMA-industry are reported in parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.64Robustness of the CWSIn this section we present results of a several exercises we perform to examine the robust-ness of our main findings. In the interest of space, we only provide the estimation results forthe average effects of the CWS on emissions.55We begin assessing the robustness of these results with a series of placebo tests. Theintent of these tests are to ensure our research design does not produce significant differencesin plant emissions when plants are “randomly assigned” into (potentially) false treated andcontrol groups. As our research design exploits three margins of variation – across industries,regions, and time – we present two separate series of placebo tests. In the first series of tests,we keep the same variation in regulatory exposure across regions and time, and randomize theplants that are in targeted and non-targeted industries. This allows us to ensure that the effectof a plant being in a violating region only matters for plants that are in targeted industries. Inthe second series of tests, we keep the same variation in regulatory exposure across industries,and randomize the plant-years that are in violating and non-violating regions. This allowsus to ensure that the effect of a plant being in a targeted industry only matters for plants inviolating regions.The results of these placebo tests are presented in Table 2.11 and Table 2.12, for PM2.5and NOX emitters, respectively. For PM2.5 emitters we only show the effect of (placebo)PM2.5 regulation, and for NOX emitters we only show the effect of (placebo) O3 regulation.The dependent variable in each regression is the natural log of plant emissions. Panel A ofeach table reports estimates from our first series of placebo tests, randomizing the plants thatare assigned to targeted industries. Panel B of each table reports estimates from our secondseries of placebo tests, randomizing plant-years that are assigned to violating regions. Withineach panel, we display estimates of specifications with different fractions of plants assignedto targeted industries or violating regions. All regressions include plant, industry-year, andCMA-year fixed effects, and standard errors are clustered by CMA-industry.The results in Table 2.11 and Table 2.12 show no significant effects of these placebo reg-ulations on plant emissions. This suggests the estimates presented in Section 2.4.3 are notsimply driven by the structure of the research design.Next, we examine whether our results are simply capturing the effects of a non-linear rela-tionship between CMA air quality and the production choices of plants therein.56 We do this55Each robustness check we also performed for the CWS’ effects on output and by plant-productivity level.56Such a relationship could arise if plants select into regions based on unobserved regional characteristicsthat are correlated with air quality. For example, if the most productive polluters select into clean regions toavoid future regulation, then comparing outcomes in dirty regions to clean regions may simply reflect differentialtrends between high-productivity and low-productivity plants.65Table 2.11: CWS Placebo Tests - PM2.5 EmissionsPanel A: Within-CMA Placebo Panel B: Within-Industry Placebo(1) (2) (3) (4) (5) (6)Placebo PM2.5 Reg. -0.015 0.010 -0.033 0.024 0.015 0.045(0.036) (0.046) (0.042) (0.033) (0.042) (0.037)R2 0.938 0.938 0.938 0.938 0.938 0.938N 7058 7058 7058 7058 7058 7058Fraction of plants intargeted industry 0.50 0.80 0.20Fraction of plants inviolating region 0.50 0.80 0.20Notes: Table reports estimates of placebo tests of PM2.5 regulation’s effect on plant PM2.5 emissions. Panel A reports estimates from placebotests that randomize the plants that are assigned to targeted industries, but preserves the actual variation in the CMA-years that violate thePM2.5 regulation. Each column within Panel A assigns a different fraction of plants into the targeted industries. The Actual fraction of plantsin targeted industries is 0.49. Panel B reports estimates from placebo tests that randomize the plants that are assigned to violating regions,but preserves the actual variation in the industries that are targeted. Each column within Panel B assigns a different fraction of plants intothe violating regions. The Actual fraction of plants in violating regions is 0.18. All regressions include plant, industry-year and CMA-yearfixed effects. Standard errors are clustered by CMA-industry. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels,respectively.Table 2.12: CWS Placebo Tests - NOX EmissionsPanel A: Within-CMA Placebo Panel B: Within-Industry Placebo(1) (2) (3) (4) (5) (6)Placebo O3 Reg. -0.010 -0.006 -0.041 -0.023 -0.002 0.024(0.029) (0.031) (0.032) (0.039) (0.049) (0.046)R2 0.978 0.978 0.978 0.978 0.978 0.978N 2779 2779 2779 2779 2779 2779Fraction of plants intargeted industry 0.50 0.80 0.20Fraction of plants inviolating region 0.50 0.80 0.20Notes: Table reports estimates of placebo tests of O3 regulation’s effect on plant NOX emissions. Panel A reports estimates from placebotests that randomize the plants that are assigned to targeted industries, but preserves the actual variation in the CMA-years that violate theO3 regulation. Each column within Panel A assigns a different fraction of plants into the targeted industries. The Actual fraction of plantsin targeted industries is 0.54. Panel B reports estimates from placebo tests that randomize the plants that are assigned to violating regions,but preserves the actual variation in the industries that are targeted. Each column within Panel B assigns a different fraction of plants intothe violating regions. The Actual fraction of plants in violating regions is 0.42. All regressions include plant, industry-year and CMA-yearfixed effects. Standard errors are clustered by CMA-industry. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels,respectively.by estimating a flexible triple-difference regression in which we allow the potential effect oftreatment to vary by the air quality of the CMA in which the plant is located. If, as we haveclaimed, being above a CWS threshold results in greater regulatory stringency, then flexibly66estimating our triple-difference regression should produce estimates that are insignificant be-low the policy’s threshold, but significant (and negative) above the threshold. In effect, thisallows us to test, rather than assert, that the CWS air quality thresholds matter.To accomplish this, we assign each plant-year observation into a bin according to the rel-evant CMA’s air quality in that year, and then estimate a version of our main specification inwhich the target industry indicators are interacted with these air quality bins. This amountsto estimating a number of difference-in-difference regressions that, for a given year, compareoutcomes for plants in targeted industries to those in non-targeted industries within CMAswith a given range of air quality, and then comparing this to the same difference in an omit-ted group of CMAs. Every year in the sample is pooled, and the coefficient on each bin isidentified from regions changing air quality bins over time.This specification is given by:Ypict =∑bβ bPM[Ki× I(APMb ≤ aPMct < APMb )]+∑bβ bO3[Ki× I(AO3b ≤ aO3ct < AO3b )]+ρp+ξct +λit + εpict ,(2.12)where b indexes air quality bin numbers, Ki selects all industries targeted by the CWS, ajctis the air quality measured in CMA c for pollutant j in year t, A jb is the air quality lowerbound for bin b for pollutant j, A jb is the air quality upper bound for bin b for pollutant j, andI(A jb ≤ a jc < A jb) is an indicator for all CMA-years with air quality that corresponds to bin bfor pollutant j.57 The coefficient β bj gives the effects of standard j in air quality bin b.In estimating Equation (2.12), we omit the “cleanest” air quality bin for each standard. Forthe PM2.5 standard, we break the air quality distribution into seven equal-sized bins from 18to 36 µg/m3. For the O3 standard, we break the air quality distribution into six equal-sizedbins from 57 to 77 ppb.58The results of the estimating of Equation (2.12) using the full sample of polluters from theNPRI are displayed in Figure 2.8 and Figure 2.9 for PM2.5 and NOX emitters, respectively.Only the coefficients for the PM2.5 standard are shown for PM2.5 emissions, and the O3 stan-57For example, suppose PM2.5 air quality ranged from 20 to 40 µg/m3, and we split this into two equal-sizedbins. The upper and lower bounds for bin one would be APM1 = 30 and APM1 = 20, respectively. The upper andlower bounds for bin two would be APM1 = 40 and APM1 = 30, respectively. Bin one would select all plants inCMAs with air quality below 30 µg/m3, and bin two would select all plants in CMAs with air quality above 30µg/m3.58For the PM2.5 regulation we include all CMA-years with air quality above 36 µg/m3 in the top bin. For theO3 regulation we include all CMA-years with air quality above 77 ppb in the top bin.67dard for NOX emissions. Each figure also displays the fraction of observations in each bin thatare treated over the sample, to show that there are treated plants over the entire distribution ofair quality. The dependent variable in each regression is the natural log of plant emissions andstandard errors are clustered at the CMA-industry level.Figure 2.8: The Effect of PM2.5 Regulation on PM2.5 Emissions, by CMA Air QualityPM2.5 Standard0.040.33 0.32 0.31 0.31 0.34 0.33 0.46−.4−.20.2ln(PM 2.5 Emissions)15 18 21 24 27 30 33 36CMA’s PM 2.5 Concentration (µg/m3)PM2.5 Standard 90% Confidence IntervalFraction Ever TreatedNotes: Figure displays estimates from a flexible DDD estimation of the PM2.5 standard’s effect on PM2.5 emissions allowing theeffects of regulation to vary by CMA air quality. Diamonds reflect the point estimates for each CMA air quality bin, while the dashedline displays the associated 90% confidence interval. These coefficients are measured relative to the excluded group (air quality below18 µg/m3 for PM2.5 and below 57 ppb for O3). Standard errors are clustered by industry-CMA. The histogram shows the fraction ofobservations in each bin treated by the respective standard at some point over the sample.Figure 2.9: The Effect of O3 Regulation on NOX Emissions, by CMA Air QualityO3 Standard0.020.33 0.45 0.55 0.45 0.42 0.33−1−.50.5ln(NO X Emissions)53 57 61 65 69 73 77CMA’s O3 Concentration (µg/m3)O3 Standard 90% Confidence IntervalFraction Ever TreatedNotes: Figure displays estimates from a flexible DDD estimation of the O3 standard’s effects on NOX emissions allowing the effectsof regulation to vary by CMA air quality. Diamonds reflect the point estimates for each CMA air quality bin, while the dashed linedisplays the associated 90% confidence interval. These coefficients are measured relative to the excluded group (air quality below 18µg/m3 for PM2.5 and below 57 ppb for O3). Standard errors are clustered by industry-CMA. The histogram shows the fraction ofobservations in each bin treated by the respective standard at some point over the sample.68The results presented in Figure 2.8 Figure 2.9 show that a break that occurs just below thePM2.5 standard’s threshold for PM2.5 emissions and at the precise level of the O3 standard’sthreshold for NOX emissions. This suggests that there are no significant differences in thetrends of treated and control plants until a CMA’s air quality reaches that of the standard’sthreshold. The observed effect of the CWS appears to be coming from a break in trend for theplants in CMA-years above the standard’s thresholds. As these thresholds were not used forany other policy, this suggests the results in the main body of this chapter reflect the effectsof increased regulation driven by violation of the CWS thresholds, rather than some otherrelationship between a CMA’s air quality and the emissions of manufacturing plants therein.Finally, we adopt a common approach in the program evaluation literature and perform anevent-study analysis in which the effect of treatment is allowed to vary over time. This typeof robustness check is useful for two reasons. First, it allows us to test whether there is asignificant difference in outcomes between our treatment and control groups before treatmentoccurs. If we’ve constructed a valid control group, there should be no significant pre-treatmentdifferences. Secondly, it allows us to determine if the effects of treatment persist into thefuture.This is particularly demanding in this setting because the majority of treated CMAs beginthe sample period under treatment, particularly for the O3 standard. As a result, we must relyon a relatively small group of treated plants for the event-study analysis and are only able toperform this robustness check for the PM2.5 standard.We implement the event-study approach by determining the first year a plant exceeds thePM2.5 standard’s threshold, then comparing treated plants to untreated plants in each of theyears before a plant is treated and each of the years after a plant is treated (for which theyare are still treated). This regression is estimated by fitting the following generalized triple-difference estimator to the dataYpict = ∑k=−3β kPMTPMick +βO3TO3ict +ρp+ξct +λit + εpict , (2.13)where T PMick is an indicator for the years before (k < 0) or after (k ≥ 0) a plant is treated forstandard j, and T O3ict captures the average effect of the O3 standard. We exclude the yearprior to treatment for the PM2.5 standard (k =−1), so the coefficients of interest (β kPM) reportthe semi-elasticity of treatment k years before or after treatment relative to the year beforetreatment. In other words, β kPMis the triple-difference coefficients relative to the year before aplant is first treated by the standard.5959Note that in our main specification the triple-difference coefficient compares the average over all years69Figure 2.10: The Effect of PM2.5 Regulation on PM2.5 by Years Pre/Post Regulation42 47 64197 192 156236−.30.3.6ln(PM 2.5 Emissions)T−3 T−2 T−1 T T+1 T+2 ≥  T+3Years Pre/Post TreatmentPM2.5 Standard 90% Confidence IntervalNumber of Treated Obs.Notes: Figure shows the results of a flexible DDD estimation of the PM2.5 standard for PM2.5 emissions allowing the treatment effectto vary by years pre/post regulation. Diamonds show the triple-difference estimation coefficients by years before and after treatment,with a 90% confidence interval in light blue. Treated plants with no pre-treatment data are omitted. All coefficients are relative to theyear before treatment (T-1), indicated by a vertical red line. Standard errors are clustered by industry-CMA. The histogram shows thenumber of observations in each bin treated by the respective standard at some point over the sample.We estimate Equation (2.13) from three periods before a plant is treated onward. Separatecoefficients are estimated up to three periods post treatment, and all periods greater than threeyears after treatment are pooled. We drop all observations that occur prior to three periodsbefore a plant is treated. All plants in CMAs that began the sample period under treatment aredropped from the regression.The results of the effects of the PM2.5 standard on PM2.5 emitters are shown in Figure 2.10.The dependent variable is the natural logarithm of PM2.5 emissions and standard errors areclustered by CMA-industry.Figure 2.10 shows strong evidence that there was no significant difference in pre-regulationtrends for our treatment and control groups for the PM standard, with the pre-regulation coef-ficients hovering tightly around zero. In addition, there was a clear break in PM2.5 emissionsstarting in the year of regulation and persisting following treatment.We can also show these estimates are robust to accounting for preemptive changes by reg-ulated plants to avoid regulation, plants that account for a significant fraction of their CMA’sair pollution, differential trends across large and small emitters, and differences in firm own-ership structure. For the sake of space, we relegate these results to the appendix (see Ap-pendix A.2.2).during which a plant is treated to the average over all years during which a plant is not treated.702.5 ConclusionIn this chapter, we estimate the channels through which a change in environmental regula-tion contributed to the “clean-up” of the Canadian manufacturing sector. We start by showingthe Canadian manufacturing sector has cleaned-up considerably in recent decades, both interms of aggregate pollution emissions, and pollution emissions per dollar of output (emissionintensity). We then perform a decomposition exercise, first used in this literature by Levinson(2009), to show this clean-up was primarily driven by reductions in industry emission inten-sity. This result suggests the sources of the Canadian clean-up were similar to the sources ofthe clean-ups observed in the U.S. and Europe.Next, we examine how Canadian manufacturing plants responded to a major revision to en-vironmental policy, the Canada-Wide Standards for Particulate Matter and Ozone, and use theresulting empirical estimates to quantify the channels through which environmental regula-tions have contributed to the manufacturing clean-up. These estimates represent the first com-plete characterization of the regulatory channels driving the manufacturing clean-up. Whilethese estimates are specific to Canada, given the similarity between the clean-ups and regula-tory structures in Canada, the US, and Europe, we believe our results provide insights relevantfor all three regions.Our estimates imply that this policy explains approximately 60% of the reduction in nitro-gen oxide emission intensity of the Canadian manufacturing sector from 2004 to 2010, andapproximately 20% of the drop in particulate matter emission intensity. However, how thispolicy caused manufacturing to clean up varied considerably across pollutants. Over two-thirds of the nitrogen oxide clean-up caused by this policy was due to the adoption of cleanerproduction processes by surviving plants (the process effect). In contrast, over 80% of theparticulate matter clean-up caused by this policy was due to plant exit (the selection effect)and the reallocation of output from regulated to unregulated plants (the reallocation effect).These results suggests that transitioning to a less-pollution intensive economy may requirelarge changes in an industry’s composition. However, the degree to which an industry’s com-position will need to change likely depends on the costs of adopting cleaner production pro-cesses. When these costs are low, as we argue is the case for nitrogen oxide process im-provements, process improvements may yield considerable reductions in industry pollutionintensity, even in the absence of plant exit or reallocation across plants.This work also highlights the importance of linked pollution and production data in as-sessing the effects of environmental regulation. The mechanisms by which plants respondto regulation appears to vary considerably across emitters of different pollutants, and across71plants that emit a common pollutant. Accounting for this heterogeneity is likely important inboth the design and assessment of environmental policy, and doing so requires rich informa-tion on firm economic and environmental performance.72Chapter 3Environmental Regulation and thePollution Haven Effect3.1 IntroductionDebates over environmental regulations often centre on how these policies will impedeinternational competitiveness by restricting the ability of domestic manufacturers to competewith foreign producers. The intuition underlying these concerns is simple: environmentalpolicies raise production costs for domestic firms, making it more difficult to compete withforeign manufacturers who do not face similar policies, both at home and abroad. As a result,regulation may lead to what is known as a “pollution haven effect” (PHE), whereby domesticregulation reduces exports and increases imports (Copeland and Taylor, 2004). These out-comes are often seen as a problem by policy makers for two main reasons. First, there isa concern that regulations may overly disadvantage domestic producers, decreasing employ-ment. Second, the change in relative costs may lead to increased foreign production. If thepollutant targeted by regulation is transboundary, this could increase foreign emissions andserve to undercut the effectiveness of domestic policy.1While there is a growing empirical literature on the PHE, this literature has yet to directlytest the PHE at the firm-level. Given recent innovations in the trade literature highlighting theheterogeneous nature of trade (Bernard et al., 2003; Bustos, 2011; Melitz, 2003), however,there is reason to believe the PHE may also be driven by heterogeneous effects across firms orfacilities. For example, it is typically only a select group of firms within a given industry that1This issue is particularly salient in the context of greenhouse gases, where it is termed “carbon leakage.”For an overview of carbon leakage, see McAusland and Najjar (2015) or Fowlie et al. (2016).73are active in export markets. If an aspect of the PHE is the exit of firms from the export mar-ket, then this will clearly only be present for the firms that export absent regulation. Moreover,regulations themselves are often designed to differentially treat firms based on their charac-teristics.2 Despite the real potential for heterogeneity in the PHE, the empirical work in thisarea has yet to push forward on this dimension. In this chapter, we assess the potential hetero-geneity in the PHE by asking how environmental regulation affects the export behaviours ofmanufacturing facilities, both on average and across plants of different productivity levels.Determining whether the PHE includes heterogeneous effects across producers may beimportant for understanding the discourse between industry and policy-makers. This is partic-ularly evident in cases where policy makers believe that a particular policy will have a smalleffect on industry. While this may be true for the industry on average, if the effects of regu-lation on production costs vary across plants, this may not be true for the marginal exporter.As a consequence, even if a particular regulation were to produce a small PHE, measured inaggregate or average terms, it could still face substantial industry opposition.In this chapter, we examine the effects of a major revision to Canadian environmentalpolicy, the Canada Wide Standards for Particulate Matter and Ozone (CWS), on export vol-umes and the selection into (and out of) exporting of affected Canadian manufacturing plants.We estimate the effects of the CWS on plant export decisions using a unique longitudinaldataset that contains information on both the pollution emissions and production decisions ofCanadian manufacturing plants over the period 2004-2010. This dataset allows us to clearlyestablish which plants in each industry were subject to regulation, making it possible for us toidentify the effects of environmental policy on exports from individual plants.Our empirical approach borrows heavily from our previous work examining how plant-level responses to the particulate matter and ozone regulations enacted under the CWS con-tributed to the clean-up of the Canadian manufacturing sector (see Chapter 2). The CWSimplemented a pair of regional air quality standards across Canada, and targeted plants in se-lect industries for regulation. The regulations used imposed a two-part regulatory constraint:requiring plants to either use clean production processes or face production constraints. Toidentify the effect of the CWS, we adopt a triple-difference research design, which exploitsthe variation in these regulations across industries, regions, and time. This approach allows usto flexibly control for factors such as regional demand shocks or industry-wide policy changes2Regulations may explicitly do this, as is the case with intensity standards that do not affect the cleanestproducers in an industry (e.g. former sulphur dioxide regulations in the US electricity sector (Lemoine, 2017)),and regulations that grandfather old firms based on past performance (e.g. the European Union’s EmissionTrading System (Knight, 2013)). Regulations may also do this implicitly, as can be the case of policies thatpenalize firms that fail to adopt clean technologies, such as the Clean Air Act (Greenstone, 2002).74that would otherwise confound the effects of environmental regulation.Before discussing our empirical analysis, we present an overview of a simple theoreticalframework of the heterogeneous nature of the PHE. As the nature of our contribution is pri-marily empirical, we do not explicitly solve the model. Instead, we use the framework toderive a set of empirical predictions that would hold under plausible circumstances. We usethese predictions to provide intuition behind the PHE, and in particular highlight the potentialheterogeneity of the PHE. Our model is similar in nature to that of Cherniwchan et al. (2017),in that we allow firm heterogeneity, assume firms compete via monopolistic competition, andallow firms to endogenously upgrade their technology. In contrast, however, we explicitlyimpose a small open economy assumption for the domestic market, and assume firms areregulated via a two-part regulatory rule, rather than a uniform pollution tax.Our empirical predictions suggest regulation should cause some firms to exit the exportmarket, and others to reduce their export volumes.3 In addition, we show that under a two-part regulatory rule, the PHE should be most strongly felt by firms toward the middle of theirindustry’s productivity distribution. The intuition for this particular form of heterogeneity isthat, due to the fixed cost of exporting, the least productive firms in an industry will not partici-pate in foreign markets. In contrast, while the most productive firms in an industry will export,they adopt technological improvements in response to regulation that allow them to avoid in-creases in their variable production costs. As a consequence, the firms that exit exporting andthe surviving exporters that reduce export volumes should be moderately productive.Next, we take our empirical approach to the data. We use a standard triple-differenceresearch design to test for the average effects of the CWS. To test for the heterogeneity impliedby our model, we follow the approach used by Bustos (2011) to study the differential effectsof trade liberalization across plants and examine whether the effects of the CWS differ acrossplants on the basis of their initial productivity level.4We find the CWS had a large negative effect on export volumes, causing a 22% reductionin export volumes for the average affected exporter. The policy, however, had no significanteffect on the selection into or out of exporting for the average plant. In addition, as our modelpredicts, the effects of regulation varied considerably across plants of different productivitylevels. In particular, arranging plants from least- to most-productive, the effects of regulationare most pronounced for plants in the second quintile of their industry’s productivity distribu-tion. For these plants, the CWS caused a 42% reduction in exports and reduced the probability3These predictions require restricting the degree to which domestic wages and aggregate prices change inresponse to policy.4As we do not observe capital information in our dataset, we define a firm’s productivity as value added perworker in the first year they enter the dataset, after removing industry means.75of exporting by 10%.In addition, we show that our results are robust to a number of potential identification prob-lems. Specifically, we show that our baseline results are robust to controlling for the effects offoreign ownership, omitting the set of plants that do not sell domestically, and persistence inoutcomes due to the presence of sunk costs.This chapter relates to the large empirical literature on the PHE, for which there havebeen, essentially, three waves of literature. The earliest wave, prior to the late 1990s, triedto estimate the “competitiveness effects” of environmental regulations by studying changesin aggregate trade flows in response to policy. This early work typically found a paucity ofevidence for the PHE (see, for example, Jaffe et al. (1995)). The lack of evidence for the PHEin these early papers was typically attributed to difficulties inherent in credibly identifying theeffects of environmental policy on international trade (Cherniwchan et al., 2017). Levinsonand Taylor (2008), for example, argue this difficulty is due in part to the potential endogeneityof environmental policy, as well as potential unobserved heterogeneity and aggregation issues.The second wave of the empirical PHE literature attempted to address these methodologicalissues by either using an instrumental variable framework, or adopting a more nuanced viewof the PHE. Work in this second wave has found evidence consistent with the PHE. As anexample of the former, Levinson and Taylor (2008), examine the effects of environmentalregulation on bilateral trade between the US and their main trading partners. They adopt asimilar approach to much of the early literature, and rely on industry abatement spending as anindirect measure of the stringency of regulation facing an industry. Unlike the early literature,however, they develop an instrumental variable approach to address the potential endogeneityof this indirect measure. They find an increase in instrumented-abatement spending leads toan increase in net imports to the US from both Mexico and Canada. Other papers, such asKellenberg (2009), Millimet and Roy (2016), and Broner et al. (2012), also use instrumentalvariable frameworks to address the endogeneity of environmental policy, finding evidencefor the PHE. An example of the latter approach, Ederington et al. (2005) argue that work inthe first wave failed to account for the geographic mobility of certain industries, as well asimportant characteristics of an industry’s trade partners. Doing so, they find evidence for thePHE: industry abatement spending has a positive effect on imports from developing countries.One potential issue with the second wave of the empirical PHE literature, as has been dis-cussed by Cherniwchan et al. (2017), is that these approaches typically rely on model-drivenarguments for the validity of their instruments. To address this limitation, the third waveof the empirical PHE literature has shifted from using an instrumental variable approach toquasi-experimental empirical strategies, such as difference-in-difference estimation, that ex-76ploit policy changes to identify the effect of environmental regulation. The advantage of thequasi-experimental approach is that the identification arguments typically require less struc-ture, in terms of the underlying economic model, relative to the instrumental variable ap-proach.There are several examples of such work. Shi and Xu (2018), for example, exploit varia-tion in regulatory stringency across industries and provinces in China stemming from China’seleventh five-year plan. They find that relatively pollution-intensive industries in highly regu-lated provinces experience a reduction in exports relative to counterfactual industry-provinces.In addition, Aichele and Felbermayr (2015) exploit variation in the stringency of environmen-tal policy across countries stemming from the Kyoto Protocol, an international agreement onclimate change policy. They find the Kyoto Protocol caused an increase in imports to relativelyhighly regulated countries. 5The work in these three waves of literature on the PHE all share at least one common fea-ture: they all assess the PHE using industry or national-level data. In addition to missing thepotential heterogeneity in the PHE, relying on industry- or national-level data is potentiallyproblematic, as the presence of plant- or firm-level heterogeneity may have implications forthe ability to identify the effect of regulation on international competitiveness. Identificationis a potential issue, as it is harder to argue for the exogeneity of policy with respect to ag-gregate values than with respect to firms. For example, in the case of air quality standards,arguments that a single firm would be unable to influence the stringency of regulation rely onan assumption of that firm representing a relatively small amount of production in its region.Clearly, this same argument cannot be made for aggregate production in the region, particu-larly if regulators care about regional employment or firms are represented by regional tradeassociations.6In addition to contributing to the literature on the PHE, this chapter also contributes toa large literature examining the effects of environmental policy on manufacturing facilities.Most of this work has focused on domestic outcomes, such as output (e.g. Greenstone (2002)),productivity (e.g. Berman and Bui (2001b), Greenstone et al. (2012)), employment (e.g.Berman and Bui (2001a), Walker (2013)), pollution (e.g. Greenstone (2003), Fowlie et al.5In work related to the PHE, Hanna (2010) exploits variation in regulation across industries and regionsstemming from the US Clean Air Act to ask whether regulation affects the foreign output of multinational firms.Hanna finds that multinationals with locations in the US responded to increased domestic regulation by increasingoutput in foreign markets. While not directly on the international flow of goods, this is suggestive evidence forthe PHE.6While Hanna’s work avoids this pitfall, her focus on multinational firms means her results are potentiallynot representative of regulation’s effect on small- and medium-sized firms, or large firms that only operate in asingle market.77(2012)), and plant entry and exit (e.g. Becker and Henderson (2000), List et al. (2003)). Ourresearch complements this body of work by showing how regulation affects a facility’s partic-ipation in foreign markets.Finally, our results contribute to a burgeoning empirical literature examining the micro-foundations of the relationship between international trade and the environment. To date themajority of these studies have focused on how international trade affects environmental out-comes at individual manufacturing plants (e.g. Martin (2012), Cherniwchan (2017)). Ourstudy contributes to this line of research by providing evidence of how environmental regula-tions affect manufacturing plants’ participation in international trade.The remainder of this chapter proceeds as follows. Section 2 presents our model. Section 3discusses our data. Section 4 discusses our research design and presents our baseline empiricalspecification. Section 4 presents our results. Section 5 concludes.3.2 The Pollution Haven Effect in a Small Open EconomyIn this section, we present an overview of a simple model of how firms that differ on thebasis of their productivity respond to environmental regulation in a small open economy. Asthe goal of this paper is primarily empirical in nature, we do not fully solve the model, nordo we dive into the myriad of important theoretical questions about the PHE in an economywith heterogeneous firms. Instead, we present a framework we believe adequately capturesour empirical setting, and use this framework to present a series of empirical predictions. Weuse these empirical predictions to show outcomes that would hold under what we believe arerealistic, although not necessary, industry-level responses (such as average domestic pricesrising following regulation).We begin by presenting the model’s set-up and discussing how it would be solved. Wethen move onto our empirical predictions, and discuss what these predictions mean for ourempirical setting.3.2.1 Model Set-UpWe adopt a framework similar to that of Melitz (2003), in which heterogeneous firmscompete via monopolistic competition in both domestic and foreign markets. Unlike Melitz(2003), however, we explicitly impose a small open economy assumption for the domesticmarket, following an approach used by Demidova and Rodrı´guez-Clare (2009). In addition,we assume firms emit pollution as a by-product of production and are regulated via a two-78part regulatory constraint intended to capture a common feature of air pollutant regulation.7Our two-part regulatory constraint, which we describe further below, requires firms to adopt aclean production process, and levies a pollution tax against those that fail to do so.Consider an economy with a single industry, which itself is comprised of a continuumof firms that compete via monopolistic competition. Firms produce a single good with asingle input, labor (denoted l), the use of which creates pollution. As mentioned above, pol-lution emitted by domestic producers is regulated via a two-part regulatory constraint. Un-der this constraint, domestic firms that use a clean-production process, which we refer to asretrofitted technology (and label r), are unregulated. In contrast, domestic firms that use adirty-production process must pay a penalty, τ , levied on each unit of pollution emitted. Thepenalty is meant to reflect both a “compliance cost” of dealing with increased regulatory over-sight, as well as a direct penalty levied on dirty producers.8 We adopt this type of regulationbecause it is a common form of regulation used to address air pollution.9 In addition, whilenot a direct representation of our empirical setting, it is a good approximation of the policy westudy.10To enter the market, firms must employ fε worth of labor. This makes the fixed entry cost,w fε , where w is the wage rate. Entering the market allows a firm to draw a productivity level,ϕ , from some known distribution, denoted by H(ϕ). The productivity draw dictates the firm’sinput requirement, such that a firm with a higher ϕ has a lower unit-labor requirement, givenby lϕ .As ϕ affects the unit-labor requirement, it also affects a firm’s pollution. Firms that usethe dirty-production technology have a pollution-labor ratio given by z(ϕ)l(ϕ) = κ . Firms that usethe retrofitted production technology have a pollution-labor ratio given by zr(ϕ)lr(ϕ) =κγ . As aresult, the pollution emitted by a firm that uses the dirty-technology and has productivity ϕis z(ϕ) = κx(ϕ)ϕ , while pollution from a retrofitted producer is zr(ϕ) =κx(ϕ)γϕ . Moreover, toadopt the retrofitted technology, a firm must pay a fixed cost of fr units of labor. Notice thatthe assumptions on the retrofitted technology make it similar to the productivity-enhancingtechnology upgrading in Bustos (2011). Here, retrofitting affects a firm’s pollution intensity,whereas in Bustos (2011), technology upgrading reduces labor intensity.7For a Melitz-style model of heterogeneous firms that face a uniform pollution tax, see Cherniwchan et al.(2017).8Compliance costs may arise as a result of regular meetings between managers and regulators, or informationreporting, for example.9For example, the National Ambient Air Quality Standards used as part of the US Clean Air Act feature thistype of policy (Greenstone, 2002).10The policy we study, the CWS, imposed production constraints on plants that failed to adopt clean processes,as well as required more stringent oversight of these plants.79In what follows, we use the superscripts no and reg to represent outcomes in a regimewithout regulation and with regulation, respectively. With this set-up, absent regulation, thefirm’s unit cost is given by cno(ϕ) = wϕ , under either the dirty or retrofitted technology. Whenthe domestic market is regulated, the firm’s unit cost is creg(ϕ) = w+τκϕ and cregr (ϕ) = wϕ , fordirty and retrofitted producers, respectively.Domestic firms can sell in the domestic market or export goods to a foreign market. Serv-ing the domestic market requires paying a fixed market access cost of w f , while serving theforeign market requires paying an additional fixed cost w fx. Similarly, foreign producers canimport goods into the domestic market as well as sell in their own market. Import fixed-costsare given by fm, which is independent of the domestic wage.Preferences of domestic consumers are given byU =[∫v∈Ωq(v)ρdv+∫v′∈Ωmq(v′)ρdv′]1/ρ, (3.1)where v and v′ denote domestic and foreign varieties, and Ω and Ωm denote the set of allavailable domestic and imported varieties, respectively. The elasticity of substitution is givenby σ = 11−ρ , where 0 < ρ < 1.The domestic market is comprised of L identical consumers, each endowed with a unit oflabor that is inelastically supplied to the domestic market. By the small open economy as-sumption, changes in the domestic labor market do not affect wages in the foreign market.Hence, domestic wages are given by w and foreign wages are given by wm. Domestic con-sumers exhaust all income, which means demand for domestic good, v, and imported good,v′, are given byq(v) = IPσ−1 p(v)−σ and qm(v′) = IPσ−1 pm(v′)−σ , (3.2)where I denotes consumer income and the domestic price index is given by P, such thatP1−σ =∫v∈Ωp(v)1−σdv+∫v′∈Ωmpm(v′)1−σdv′.Given this demand system, a domestic producer charges a mark-up over their unit (ormarginal) costs. Absent regulation, a firm with productivity ϕ charges pno(ϕ) = wρϕ undereither technology. As a result, domestic revenues absent regulation under either technology80are given byrno(ϕ) =I [Pρ]σ−1ϕσ−1wσ−1. (3.3)Domestic profits using the dirty technology are pino(ϕ)= r(ϕ)σ −w f , and domestic profits usingretrofitted technology are pinor (ϕ) =r(ϕ)σ −w [ f + fr].To make the small open economy assumption explicit, we adopt the approach of Demidovaand Rodrı´guez-Clare (2009) and assume foreign demand for a domestic variety, v, is givenby Apx(v)−σ , where A is exogenous and px(v) is the price charged by a domestic exporter.11Notice this means that even though the domestic market is small, domestic exporters havesome price setting power. However, unlike in the domestic market, changes in their price onlyaffect the demand of their goods, and do not affect the demand for other varieties in the foreignmarket. Using either technology, without regulation, an exporter with productivity ϕ chargespx(ϕ) = wρϕ , and obtains revenues given byrnox (ϕ) =A [ρ]σ−1ϕσ−1wσ−1. (3.4)Export profits under either technology, absent regulation, are pinox (ϕ) =rx(ϕ)σ −w fx.Without regulation, adopting the retrofitted technology requires paying a fixed cost, butdoes not change the firm’s revenues. As a result, no firms retrofit absent regulation.Under regulation, domestic producers also set prices as a mark-up over their unit costs.Producers that use the dirty technology charge preg(ϕ) = w+τκρϕ for their output, whereas pro-ducers that use the retrofitted technology charge pregr (ϕ) = wρϕ . As a result, domestic revenuesfor the dirty technology are given byrreg(ϕ) =I [Pρ]σ−1ϕσ−1[w+ τκ]σ−1, (3.5)and domestic revenues for the retrofitted technology arerregr (ϕ) =I [Pρ]σ−1ϕσ−1wσ−1. (3.6)In addition, profits from production are pireg(ϕ) = r(ϕ)σ −w f and pireg(ϕ) = r(ϕ)σ −w [ f + fr]11As Demidova and Rodrı´guez-Clare (2013) show, this form of foreign demand is the limiting case of a twolarge economy model, where one economy becomes infinitesimally small.81for the dirty and retrofitted technology, respectively.Similarly, it can be shown that when the domestic market is regulated, export revenue for afirm using the dirty technology isrregx (ϕ) =A [ρ]σ−1ϕσ−1[w+ τκ]σ−1. (3.7)In addition, export profits under the dirty technology are piregx (ϕ) = rregx (ϕ)σ −w fx. In contrast,export revenue for a firm using the retrofitted technology, when regulated, isrregx,r (ϕ) =A [ρ]σ−1ϕσ−1wσ−1, (3.8)and profits are piregx,r (ϕ) =rregx,r (ϕ)σ −w fxNotice that under regulation, a firm chooses to adopt the retrofitted technology if doing soincreases revenues enough to cover the fixed cost of retrofitting. Assuming the fixed-cost ofretrofitting is large enough relative to the fixed-cost of exporting12, then a firm is indifferentbetween using the dirty and retrofitted technology ifpireg(ϕ)+piregx (ϕ) = pireg(ϕ)+piregx,r (ϕ).Denoting the productivity level that would make a firm indifferent between producing usingthe dirty or retrofitted technology as ϕr, then this productivity cut-off is given byϕσ−1r =σwσ frρσ−1[wL [Preg]σ−1+A][1−G(w,τ)], (3.9)where G(w,τ) =[ ww+τκ]σ−1 and we have assumed the tax revenue is destroyed, which meansdomestic income is I = wL. This means that, under regulation, any firm that draws a produc-tivity level greater than ϕr would choose to adopt the retrofitted technology. Any firm thatdraws a productivity level below ϕr would use the dirty technology.Three market clearing conditions are required to close the model: a labor market clearingcondition, a free entry condition, and a trade balance condition. Labor market clearing requirestotal domestic labor supply to equal total domestic labor demand. Labor market clearing under12The parameter restriction required for this to be a valid equilibrium is frfx >[wL[Preg]σ−1+AA][[wreg+τκwreg]σ−1−1]> 1.82regulatory regime j can be written asL = M j[fε + L¯jd + L¯jx+[1−H(ϕ jε )]f +[1−H(ϕ jx )]fx+[1−H(ϕ jr )]fr],where M j is the measure of entrants, ϕ jε is the productivity cut-off below which firms chooseto exit the market, L¯ jd is average labor demand for production to serve the domestic market,and L¯ jx is average labor demand for production to serve the foreign market. Free entry requiresfirms to earn zero expected profits from entering production. Letting δ be an exogenous exitprobability, and p¯i j be average profits for domestic producers, then free entry givesp¯i j = δw fε .Finally, the trade balance condition requires that the total value of exports equal the total valueof imports. Trade balance requiresM jR¯ jx = MFORR¯jm,where MFOR is the measure of foreign firms that enter the domestic market, R¯jx are averageexport revenues for domestic producers, and R¯ jm are average import revenues for foreign pro-ducers that sell in the domestic market.3.2.2 Empirical Predictions of the Pollution Haven EffectWith this set-up, we now present a series of empirical predictions relevant for the PollutionHaven Effect. We present each prediction, and use our theoretical framework to explain whythe prediction would hold under the specified conditions.Recall that the PHE arises because domestic regulation increases the cost of producing inthe domestic market relative to the cost of producing in the unregulated foreign market. Ourfirst two predictions make this argument clear by examining the effect imposing environmentalregulation has on a firm’s total and export revenues, respectively.The first of these predictions is intended to show the conditions under which a reduction intotal revenues can be used to infer that variable production costs have increased for domesticproducers. As we discuss in the empirical section, we cannot observe variable costs. This isproblematic because changes in variable costs are the cornerstone of the PHE. However, asis made clear by Empirical Prediction 1, by observing changes in a plant’s total revenue inresponse to regulation, we can conclude variable costs likely rose as a result of said policy.To introduce this prediction, notice that we can express total revenues for a firm that uses83production technology t in regulatory regime j astr jt (ϕ) =[I j[P j]σ−1]ρσ−1c jt (ϕ)if ϕ jε ≥ ϕ < ϕ jx ,[I j[P j]σ−1+A]ρσ−1c jt (ϕ)if ϕ ≥ ϕ jx ,(3.10)where c jt (ϕ) is the variable cost of producing with technology t in regime j, ϕjε is the produc-tivity cut-off below which firms exit the domestic market, and ϕ jx is the productivity cut-offbelow which firms exit the foreign market. Clearly, firms that export, those with ϕ ≥ ϕ jx , re-ceive revenues from both the domestic and foreign markets, while firms that do no export onlyreceive domestic revenues.For a firm with a given productivity level, dividing total revenues under regulation by theirtotal revenues without regulation gives the proportional change in revenues due to regulation.To make this explicit by way of example, taking this ratio for a non-exporting firm, usingI j = w jL, and simplifying, gives13trregt (ϕ)trregt (ϕ)=[wreg [Preg]σ−1][wno [Pno]σ−1] cno(ϕ)creg(ϕ). (3.11)Regulation affects revenues because of its direct effect on production costs, and its indirecteffects on wages and average industry prices. From Equation (3.11), it is straightforward toshow that revenues fall for a firm that uses technology t and does not export under eitherregime if the increase in production costs satisfiescregt (ϕ)cnot (ϕ)>wreg [Preg]σ−1wno [Pno]σ−1. (3.12)Similarly, restrictions on production cost changes that deliver a reduction in total revenuescan be shown for firms that export in both regimes, or that drop out of exporting as a resultof regulation.14 This is all to say that total revenues will fall for any firm that experiences alarge enough increase in production costs as a result of regulation (either directly, or indirectlythrough changes in wages).13As discussed above, we assume tax revenues are destroyed.14If regulation causes industry prices to rise relative to equilibrium wages, then it can be shown that Equa-tion (3.12) is also a sufficient condition to ensure total revenues fall for any surviving firm that uses technologyt. This claim follows because if[PregPno]σ−1> wregwno , thenIreg[Preg]σ−1Ino[Pno]σ−1> Ireg[Preg]σ−1+AIno[Pno]σ−1+A> Ireg[Preg]σ−1Ino[Pno]σ−1+A.84The discussion of the effects of regulation, so far, has ignored any potential heterogeneityinvolved in the effects of regulation on revenues. There are, however, two margins throughwhich heterogeneity arises. First, recall that due to the nature of this policy instrument, vari-able production costs rise more for firms that do not retrofit than for the firms that retrofit.As we showed in the preceding section, these retrofitting firms will be relatively productive.Thus, the most productive firms in an industry will retrofit, and will experience a relativelysmall reduction in revenues as a result.15Second, as we will discuss in greater detail for the second empirical prediction, exportrevenues may fall by more than domestic revenues following regulation. This would occurif average domestic prices rise following regulation, thereby insulating domestic revenuesfrom the cost increase. As the least productive firms only serve the domestic market, anddomestic sales are insulated due to domestic price adjustments, then regulation will cause arelatively small proportional reduction in total revenues for the least productive firms. Theconsequence of these two forms of heterogeneity is that the reduction in total revenues islargest, in percentage terms, in the middle of the productivity distribution.We summarize these results in the following empirical prediction.Empirical Prediction 1. If the increase in variable production costs for a given firm is suffi-ciently large, then(a) imposing environmental regulation will cause a reduction in the firm’s total revenues;(b) in percentage terms, total revenues will fall the most for firms toward the middle of theproductivity distribution if average domestic prices rise in the industry.By increasing the variable production costs for domestic producers, regulation should de-crease a firm’s optimal scale, provided the increase in production costs is sufficiently large.Our second empirical prediction explicitly connects this increase in production costs to thePHE.Empirical Prediction 2. If the domestic price index rises, then for a given domestic firm, im-posing environmental regulation will cause a larger proportional reduction in export revenuesthan domestic revenues.To see how Empirical Prediction 2 arises, it suffices to show how domestic and exportrevenues change in response to regulation for a firm with a given productivity level. Recallthat absent regulation, firm revenues from domestic sales under either technology are given15Note that if wages fall due to regulation, then revenues may even rise for these firms.85by Equation (3.3). In contrast, when regulation is imposed, a firm’s domestic revenues whenusing the retrofitted technology is given by Equation (3.6), and their domestic revenues whenusing the dirty technology is given by Equation (3.5). In addition, as discussed above, onlythe most productive firms in the industry adopt the retrofitted technology.Comparing the domestic revenues across the two regulatory regimes for a firm that uses theretrofitted technology under regulation givesrregr (ϕ)rno(ϕ)=[PregPno]σ−1[ wnowreg]σ. (3.13)That is, domestic revenues only change for firms that use the retrofitted technology because ofthe change in equilibrium industry prices and wages. In comparison, the change in domesticrevenues for a firm that does not retrofit under regulation givesrreg(ϕ)rno(ϕ)=[PregPno]σ−1[ wnowreg]σ [ 11+ τκwreg]σ−1. (3.14)This means domestic revenues change for the non-retrofitting producers because of increasedproduction costs, as well as the change in equilibrium industry prices and wages.A similar comparison gives the change in export revenues for the producers that choose toretrofit their technology asrregx,r (ϕ)rnox (ϕ)=[wnowreg]σ−1, (3.15)and the change in export revenues for non-retrofitters asrregx (ϕ)rnox (ϕ)=[wnowreg]σ−1[ 11+ τκwreg]σ−1. (3.16)Comparing Equation (3.16) to Equation (3.14) shows that if domestic prices rise as a result ofregulation, then the reduction in export revenues for non-retrofitting producers – those whosevariable costs are directly affected by regulation – is larger than the reduction in domesticrevenues. The comparison of Equation (3.15) to Equation (3.13) shows the same holds forproducers that adopt the retrofitted technology when regulated.Empirical Prediction 2 clarifies the intuition underlying the PHE in a small open economy.In equilibrium, domestic prices and wages adjust to domestic policy, which may insulate do-mestic producers who sell in the domestic market. However, because the domestic market is86assumed to be small relative to the foreign market, this policy change does not affect foreignprices. This leads to a relatively large reduction in export revenues compared to domesticrevenues.Our third empirical prediction highlights the heterogeneous nature of the PHE under thisregulatory environment. Under a two-part regulatory constraint, firms can avoid paying higherproduction costs by adopting clean (or retrofitted, as we have called it) technology. By avoid-ing increased production costs, these firms experience smaller reductions in both domesticand export revenues relative to the firms that use the dirty technology. This leads to our thirdprediction.Empirical Prediction 3. In percentage terms, the reduction in export revenues as a resultof environmental regulation will be largest for the firms that do not retrofit their productiontechnology. These firms will be the least productive exporters in the industry.Empirical Prediction 3 follows immediately by comparing Equation (3.15) to Equation (3.16).Firms that do not retrofit while regulated, which are the least productive in the industry, experi-ence an increase in production costs not felt by the firms that retrofit. As a result, the reductionin export revenues is largest for these low-productivity exporters. Empirical Prediction 3 saysthat the PHE, at least under a two-part regulatory constraint, will be most pronounced for theleast productive exporters. It is also worth noting that, as exporting requires paying a fixed-cost, only relatively productive firms choose to export. This implies that, while the heavilyaffected exporters will not be the most productive firms in their industry, they will not be theleast productive either, as the least productive firms will not export.The fixed-cost of exporting produces one final empirical prediction worth discussing. Dueto this export fixed-cost, not all domestic producing firms within a given industry serve theforeign market. Moreover, the marginal exporter will have a relatively low productivity level,and as a result will use the non-retrofitted production technology. This leads to the followingempirical prediction.Empirical Prediction 4. If the reduction in wages following regulation is sufficiently small,then imposing environmental regulation causes some firms to leave the export market. Thesefirms will be the least productive exporters in the industry.To see how Empirical Prediction 4 arises, recall that a firm’s size is a monotone transforma-tion of productivity. This means there is a productivity cut-off above which all firms export,and below which no firms export. Absent regulation, this export productivity cut-off, labeled87ϕx, is given byϕnox =[σ [wno]σ fxAρσ−1] 1σ−1. (3.17)Under regulation, the export cut-off isϕregx =[σ [wreg]σ fxAρσ−1] 1σ−1 [1+τκwreg]. (3.18)Comparing the two export productivity cut-offs gives[ϕregxϕnox]σ−1=[wregwno]σ [1+τκwreg]σ−1, (3.19)which is greater than one if, and only if,[wnowreg]σ<[1+ τκwreg]σ−1. Moreover, as exit fromexporting is caused by an increase in the exporting productivity cut-off, then by definition, theexiting firms will be the least productive exporters.Notice that Empirical Prediction 4 also has implications for identifying the Pollution HavenEffect. In particular, it implies that an observed reduction in firm revenues need not be evi-dence of a Pollution Haven Effect. The intuition behind this is that, as only relatively largefirms choose to export, observing a change in firm revenues may reflect changes at firms thatare not present in foreign markets. This could, in principal, reflect general equilibrium effectsof policy, rather than changes in the relative cost of production between domestic and foreignproducers.16To emphasize the intuition underlying the heterogeneous nature of the PHE we presenta series of graphs showing how regulation affects a firm’s export decision and technologychoices. Figure 3.1 plots a firm’s profit absent regulation as a non-exporter (pino(ϕ)), as anexporter (pinox (ϕ)), and as an exporter that uses the retrofitted technology (pinox,r(ϕ)).17 Asthe figure shows, firms with a productivity draw above ϕnox choose to pay the fixed cost toexport, and sell in both the domestic and foreign markets. The remaining producing firms,with ϕ ∈ (ϕnoε ,ϕnox ), only serve the domestic market. No firm chooses to retrofit, as doing soraises the fixed costs of production, and does not affect revenues.Figure 3.2 shows how regulation affects these decisions.18 In this figure, we plot firms’16Note that it is straightforward to show that if A < IPσ−1, then some firms only serve the domestic market.17To simplify the graph, we do not plot profits for firms that do not export but use the retrofitted technology.18For simplicity, we hold industry prices and wages fixed.88pi(ϕ)ϕσ−1pino(ϕ)pinox (ϕ)pinox,r(ϕ)0− f−[ f + fx]−[ f + fx+ fr]ϕno[σ−1]ε ϕno[σ−1]xExportFigure 3.1: Export and Technology Choices without Environmental Regulationprofits both with and without regulation for non-exporters, exporters, and exporters that usethe retrofitted technology. Profits under regulation are labeled with superscript reg, while theprofits without regulation are labeled no. Regulation reduces the profitability of operatingusing the dirty technology for both exporters and non-exporters alike. This causes the leastproductive exporters, those with ϕ < ϕregx , to exit the export market, as we discussed in Em-pirical Prediction 4. In addition, the most productive exporters, those with ϕ ≥ ϕregr , now findit worthwhile to retrofit. The remaining firms, however, remain in the export market, but donot retrofit. As we showed in Empirical Prediction 3, these low-productivity exporters facea large reduction in export sales because their production costs increase, and foreign pricescannot adjust due to the small open economy assumption.In the empirical analysis that follows, we test these predictions by examining the effect of achange in Canadian environmental policy on the export behaviours of Canadian manufacturingplants. Our tests of Empirical Prediction 1 and Empirical Prediction 2 come from examiningregulation’s effect on a plant’s total sales and comparing this to regulation’s effect on sales thatoccur in foreign markets. We test Empirical Prediction 3 by examining how regulation’s effecton export sales varies across plants of different productivity levels. Finally, we test EmpiricalPrediction 4 by estimating regulation’s effect on a plant’s decision to select into and out ofexporting, in particular, across plants of different productivity levels.89pi(ϕ)ϕσ−1pino(ϕ)pinox (ϕ)pinox,r(ϕ) = piregx,r (ϕ)pireg(ϕ)piregx (ϕ)0− f−[ f + fx]−[ f + fx+ fr]ϕno[σ−1]x ϕreg[σ−1]x ϕreg[σ−1]rExit Export RetrofitFigure 3.2: Export and Technology Choices with Environmental Regulation3.3 Data and MeasurementOur goal in this chapter is to determine the effect of environmental regulation on the exportdecisions of Canadian manufacturing plants. To do so, we utilize a unique micro dataset thatcontains information on both the pollution emissions and export decisions of Canadian manu-facturing plants over the period 2004-2010. This dataset was created by linking the data fromthe National Pollution Release Inventory (NPRI), a publicly available dataset containing infor-mation on the pollution emissions of Canadian manufacturing facilities, with the confidentialdata on plant characteristics from the Annual Survey of Manufacturers (ASM).19 Together,these data sources allow us to create a longitudinal dataset containing information on the ex-port decisions of plants that emit fine-scale particulate matter (PM2.5), a pollutant regulated inCanada over our period of study as part of the suite of of environmental regulations called theCanada-Wide Standards.3.3.1 Descriptive StatisticsIn order to understand how environmental regulations affect a plant’s participation in in-ternational markets, we examine three outcomes: the likelihood of exporting, total sales, and19These data were linked by Statistics Canada. For further details on the data and its construction, see Chap-ter 2.90Table 3.1: Summary StatisticsPM2.5 Full ASM(1) (2)Sales ($1 mill.) 194.62 11.12(890.55) (123.56)Exports ($1 mill.) 97.88 6.661(709.67) (89.74)Pr(Export) 0.76 0.36(0.43) (0.48)N 6501 309541Notes: Table reports averages and standard deviations of key variables examined in the mainanalysis. Each column reports the summary statistics for a different sample. Column (1) isthe sample of plants that emit PM2.5 and column (2) reports plant characteristics for the entiremanufacturing sector. Statistics in column (1) are weighted to account for potential sample biasinduced by the match of the NPRI and ASM. All monetary values are reported in 2007 Canadiandollars.export sales. Together, these variables allow us to determine if environmental regulations af-fect international competitiveness along the intensive (total sales, and exports) or extensive(likelihood of exporting) margins.Summary statistics for total sales, total exports and the probability of exporting are reportedin Table 3.1. Column (1) of the table reports statistics for our main dataset, which comprises anunbalanced panel of manufacturing plants that emit PM2.5 pollution. The summary statisticsin column (1) are weighted to account for any possible sample bias created by the procedureused to link the NPRI and ASM.20 For comparison, column (2) of the table reports summarystatistics for the entire sample of plants in the ASM. Although we do not use the full ASMdataset in our analysis, we present these statistics to highlight the difference between oursample of polluters and the average Canadian manufacturing facility.The descriptive statistics reported in Table 3.1 suggest that, on average, the manufacturingplants that emit PM2.5 are substantially larger, are much more likely to export, and export morethan the average plant in the Canadian manufacturing sector. This is driven both by reportingrequirements for the NPRI, as plants typically only report to the NPRI database if they have atleast ten employees, and structural differences between polluters and non-polluters.2120For details, see Chapter 2.21It is worth noting that the NPRI requires any plant that operates a boiler or generator on-site to report theirPM2.5 emissions, regardless of their number of employees. As many industrial PM2.5 emitters use an on-siteboiler or generator, it is unlikely that the employment threshold is the main cause of the differences reported inTable 3.1.913.3.2 Canadian Environmental RegulationsWe supplement the data from the NPRI-ASM dataset with data from Chapter 2 on whethereach plant faced regulation under the Canada-Wide Standards for Particulate Matter and Ozone(CWS).22 The CWS was a major revision to Canadian environmental policy that occurred inthe year 2000 as a result of an agreement between the federal government of Canada and theprovinces. The policy was intended to improve air quality across the country by creating airquality standards for fine-scale particulate matter (PM2.5) and ground-level ozone (O3) thatapplied to each major town or city in Canada.23 These standards required each CMA to meetan air quality target; those cities with poor ambient air quality were required to adopt stringentenvironmental regulations, while the remaining CMAs had to ensure that their air quality didnot deteriorate. In addition, the CWS designated a set of “targeted” industries that were to bethe focus of regulation given that they were viewed as key determinants of poor air quality.24Provincial authorities regulated plants that were in targeted industries and violating regionsusing two-part regulatory constraints. As part of the annual provincial operation permittingsystem, regulated plants had to either show they were operating using clean production pro-cesses, or face a production constraint.While the CWS regulated emissions of both PM2.5 and various O3 pre-cursors, in thischapter we only focus on its effect on plants that emit PM2.5. We make this choice because,due to differences in technical constraints facing emitters of PM2.5 and O3 pre-cursors, onlythe PM2.5 standard appears to have had a meaningful impact on variable production costs.As we show in Chapter 2, emitters of O3 pre-cursors appear to have responded to the CWSby adopting process changes that produced pollution reductions without increasing variableproduction costs. Moreover, the existing evidence suggests the fixed costs associated withthese process changes were relatively small.25 As the PHE is based on regulation’s effect onproduction costs, either variable or fixed, we focus on the pollutant where this effect appearsto be meaningful. We leave an examination of the O3 standard’s effect on the export decisions22The Canada-Wide Standards for Particulate Matter and Ozone were two of the many environmental stan-dards enacted under the Canada-Wide Standard system. Canada-Wide Standards were created for benzene,mercury, and dioxins and furans, among others.23Under the terms of the agreement, an urban area’s status as a major town or city was determined using Statis-tics Canada’s definitions of Census Agglomeration (CA) or Census Metropolitan Area (CMA). For convenience,we use the terms CMA and city to refer to both CAs and CMAs.24The targeted industries were pulp and paper, lumber and wood product manufacturing, electric power gen-eration, iron and steel manufacturing, base metal smelting, and the concrete and asphalt industries (CanadianCouncil of Ministers of the Environment, 2000b).25In Chapter 2, we present a detailed discussion on the technical differences in process changes available toemitters of these two pollutants.92Figure 3.3: Regulatory Status Changes under the CWSNotes: Figure depicts PM2.5 and O3 standard status changes for each CMA from 2000 to 2010. Red CMAschanged status under both the PM2.5 and O3 standards. Orange CMAs only changed status for the PM2.5standard. Yellow CMAs only changed status for the O3 standard. Green CMAs didn’t change status under eitherstandard. The mainland United States is shown in light gray. Part of the northern Canadian Territories aretrimmed for scale. The inset shows detail on the most densely populated area of Canada, colored in light red onthe main map. Source: Chapter 2.of the emitters of O3 pre-cursors for future work.26The variation in environmental regulation created by the CWS is shown in Figure 3.3,which depicts which CMAs were forced to adopt more stringent environmental regulationsto address ambient PM2.5 and O3 problems at least once over the period 2000-2010. In thefigure, CMAs that adopted more stringent environmental regulations due to ambient pollutionconcentrations exceeding the relevant air quality standard are depicted in red, orange and26It is worth noting that in our model, only changes in variable costs as a result of regulation directly affectexport revenues and a firm’s choice to enter the foreign market. However, in principle, general equilibriumchanges in wages resulting from regulation could also affect a firm’s exports. This means even if regulation onlydirectly affected the fixed-cost of producing, and not variable production costs, changes in equilibrium wagesresulting from general equilibrium effects would change exports.93yellow. The red CMAs were required to adopt more stringent policy under both the PM2.5 andO3 standards, while the orange and yellow CMAs were only required to adopt more stringentpolicy under the PM2.5 standard or the O3 standard, respectively. CMAs depicted in greenwere not required to adopt more stringent policy under either standard.As the figure shows, the CWS created substantial variation in environmental regulationsacross CMAs. Of the 149 CMAs in our sample, 23% adopted new regulations under only thePM2.5 standard, 26% adopted new regulations under only the O3 standard, and 11% adoptednew regulations under both standards. We exploit this variation, and the fact that the CWStargeted a subset of industries, to identify the effects of environmental regulation on the exportdecisions of Canadian manufacturing plants.3.4 Research DesignTo determine the effects of environmental regulation on the export decisions of Canadianmanufacturing plants, we adopt the research design we developed in our previous work tostudy the effects of the CWS on the clean-up of the Canadian manufacturing sector (see Chap-ter 2). As such, we exploit the variation in regulation created by the design and implemen-tation of the CWS. As shown in Figure 3.3, ambient air quality changes led to variation inthe environmental regulations faced by different CMAs over time. Moreover, these regula-tions targeted a subset of industries in the manufacturing sector, meaning that the regulationsvaried across industries as well. We use these three sources of variation – across industries,regions, and time – in a triple difference research design to isolate the causal effects of thePM2.5 standard on the export decisions of affected plants.27Our approach starts by exploiting the variation in PM2.5 regulation over time. To thatend, our research design compares the average outcomes from plants in targeted industrieslocated in regulated CMAs (the plants “treated” by PM2.5 regulation) while regulated to theiroutcomes while unregulated. This comparison allows us to control for any unobserved time-invariant plant, industry or CMA-specific heterogeneity that would otherwise confound theeffects of regulation. We then exploit the variation in regulation across industries, by com-paring the average outcomes from plants in targeted industries to the average outcomes fromplants in non-targeted industries located in the same CMAs in the same year. This allowsus to control for any unobserved time-varying CMA specific heterogeneity, such as localizedrecessions, that might affect the decision to export. We then exploit the variation in regulationacross regions by taking plants in the same industry in the same year, and comparing the av-27As we noted in Section 3.3, while the CWS regulated both PM2.5 and O3 pre-cursors, we focus on the PM2.5standard for our analysis.94erage outcomes from plants in regions that violate the PM2.5 standard, at some point in time,to the average outcomes from plants in non-violating regions. This allows us to control forany time-varying industry heterogeneity, such as foreign demand shocks, that would otherwiseconfound identification. Finally, our approach compares the average outcomes from treatedplants to the average outcomes from plants from non-targeted industries located in CMAsthat did not experience a change in PM2.5 regulation. These plants serve as a counterfactualthat allow us to capture the effects of any unobserved aggregate shocks, such as changes intechnology or exchange rate fluctuations, common across all facilities in the country.3.4.1 Empirical SpecificationWe implement this research design by estimating several variants of the following equation:ypi jt = βT PM2.5i jt +ρp+µ jt +λit + εpi jt (3.20)where ypi jt is the outcome of interest (either an indicator of export status, total shipments, orexport sales) at plant p, in industry i, located in CMA j at time t. T PM2.5i jt is a treatment indi-cator for the particulate matter (PM) standards implemented under the CWS. This indicatortakes the value one for plants that are in industries targeted by the CWS for years in whichtheir CMA exceeds the relevant pollution threshold. The ρp are plant fixed effects that captureany time-invariant plant specific heterogeneity. The µ jt are CMA×year fixed effects that cap-ture any time-varying region specific heterogeneity, such as localized recessions. The λit areindustry×year fixed effects that capture time-varying industry heterogeneity, such as demandshocks. Finally, εpi jt captures idiosyncratic changes in outcomes across plants.The coefficient of interest in (3.20) is β . This coefficient capture the average percentagedifference in outcomes across plants that were affected by the PM2.5 standard relative to thosethat were not, and is identified from within-plant comparisons over time. These comparisonswill identify the causal effect of environmental regulations if there are no other factors asidefrom the CWS particulate matter standard driving differences in export behaviours acrossplants over time. As we discuss in Chapter 2, there are two reasons to believe this is the case.First, the CWS air quality standards were set federally, meaning that they are unrelated to localtastes, characteristics and economic conditions. Second, ambient pollution levels in a CMAdo not necessarily reflect local economic activity due to the fact that particulate matter can betransported long distances via wind patterns. These facts suggest that treatment is exogenous.While equation (3.20) will produce estimates of the average effect of the CWS particulatematter standard, our model predicts that the effects of regulation will differ across plants on95the basis of their productivity level. To investigate this heterogeneity, we adopt the approachfirst used by Bustos (2011) to study the differential effects of trade liberalization across plantson the basis of their initial productivity. Specifically, we estimate:ypi jt =5∑q=1βBq[T PM2.5i jt ×Bq]+ρp+µ jt +λit + εpi jt (3.21)where Bq is an indicator variable equal to one if plant p is in productivity bin q, βBq is the effectof the CWS particulate matter standard on plants in productivity bin q, and all other variablesare defined as in equation (3.20). Given that the ASM does not include information on plantcapital stocks, making it impossible to calculate TFP measures using standard methods, weproxy for initial plant productivity using value added per worker in the first year a plant entersour sample, and construct Bq by dividing plants into productivity bins.28We assign plants into productivity bins according to where they lie on their industry’sproductivity distribution, using the entire set of plants that emit PM2.5. Consequently, B1, forexample, corresponds to the first quintile of the productivity distribution of PM2.5 emitters.We refer to these as bins, rather than quintiles, because in some regressions we restrict oursample to plants that are continuing exporters. For these regressions, we do not redefine theproductivity bins. As a result, a bin may have less than 20% of the observations for the sampleused in these regressions.3.5 Results3.5.1 Plant RevenueWe begin our empirical analysis by examining the effects of the CWS particulate matterregulations on total plant revenue (Empirical Prediction 1). While our ultimate interest is inunderstanding how environmental regulation affected participation in export markets and ex-port revenue, examining total revenue is a useful first step because it provides indirect evidenceas to the effects on production costs of complying with the CWS particulate matter regulations.Although the available evidence suggests that these costs are large, we do not observe themdirectly in our data. However, as stated in Empirical Prediction 1, if the compliance costs aresufficiently high, then the total revenues of affected plants should fall in response to regula-tion. If the CWS does not affect the revenue of affected plants, then it is unlikely that the costs28To address the possibility that the plant productivity measures capture differences inherent across industriesor time, we first demean all productivity measures by regressing initial productivity levels on entry-year andindustry fixed effects. We then use the residuals from this regression as our measure of plant productivity.96of the regulations are substantial enough to have affected export decisions.These results are presented in Table 3.2. The table reports estimates from seven separateregressions; in all cases, the dependent variable is the natural log of total revenue, and eachregression is weighted to correct for potential sample bias induced by the procedure used tomatch the NPRI to the ASM.29 Throughout, standard errors clustered by industry-CMA arereported in parentheses. The first six specifications reported in columns (1) to (6) are based onequation (3.20); as such, the estimated coefficient reports the average effect of the particulatematter standard on affected plants. Our baseline estimate, reported in column (1), includesplant, industry-year and CMA-year fixed effects. Columns (2) through (6) present evidenceof the robustness of PM regulation’s effect on total revenues. Column (2) adds an indicator ofwhether industry i in CMA j was also regulated under the CWS O3 standard in year t. Column(3) adds an indicator of whether plant p was owned by a foreign parent at time t. Column (4)restricts the sample to exclude the set of plants that only export and do not sell domestically.Finally, columns (5) and (6) include a lagged dependent variable. Given that estimating thesespecifications with a fixed effects estimator would yield inconsistent estimates, we follow theapproach of Arellano and Bond (1991) and adopt a GMM procedure with either one (column(5)) or two (column (6)) lags as instruments.The last specification, reported in column (7), is based on equation (3.21) and reports theestimated effects of the particulate matter standard by bins corresponding to a plant’s initialproductivity level. The estimates reported in column (7) allow us to test the heterogeneity inthe effects of environmental regulation implied in Empirical Prediction 1.The estimates reported in the first six columns of Table 3.2 suggest that the CWS partic-ulate matter regulations led to a significant reduction in the total revenue of affected plants.For example, the estimate reported in column (1) indicates that the CWS particulate matterstandard is associated with a 10.8% reduction in total revenue at the average affected plant.In addition to being statistically significant, this effect is also economically meaningful; giventhat the average plant in our sample had revenues of close to $195 million CAD, this estimateimplies that the CWS particulate matter reduced the revenues at the average plant by just over$21 million CAD. Our preferred estimate, reported in column (2), shows that this effect isrobust to controlling for the effects of the other regulation imposed under the CWS. Indeed,adding an indicator of treatment status under the CWS O3 regulations appears to have no ef-fect on either the point estimate or standard error, suggesting that our estimates of the effectof the PM2.5 standard are not capturing the effects of other CWS regulations.Columns (3) and (4) provide further evidence that the CWS particulate matter standard29Details of the match procedure are available in Chapter 2.97Table 3.2: Environmental Regulations and Plant Revenue(1) (2) (3) (4) (5) (6) (7)PM2.5 Std. -0.108b -0.108b -0.108b -0.110b -0.076b -0.077b(0.050) (0.050) (0.050) (0.051) (0.038) (0.039)PM2.5 Std.×B1 -0.097c(0.058)PM2.5 Std.×B2 -0.286a(0.092)PM2.5 Std.×B3 -0.043(0.078)PM2.5 Std.×B4 0.048(0.059)PM2.5 Std.×B5 -0.045(0.065)O3 Std. X X X X X XForeign Owner XRest. Sample XR2 0.224 0.224 0.225 0.225 0.227AR1 -3.358 -3.402AR2 -0.802 0.828N 6501 6501 6501 6149 3694 3694 6501Notes: Table reports estimates of the effects of the CWS on the natural log of manufacturing plant revenue. All regressions include plant,industry-year and CMA-year fixed effects, and are weighted by the inverse of the match probability to control for potential match-inducedsample bias. Column (3) includes an indicator of whether the plant is owned by a foreign company. Column (4) restricts the sample toexclude plants that only sell abroad. Columns (5) and (6) report Arellano-Bond estimates with one and two lags, respectively. In all cases,standard errors are clustered by CMA-industry. c, b, and a denote significance at the 10%, 5%, and 1% level, respectively.negatively affected manufacturing plant revenues. One concern with our preferred estimateis that it is not just capturing the effects of CWS regulation, but also the effects of foreignownership, which is time-varying. For example, foreign owners may be able to help offsetthe effects of a negative shock such as the CWS, in ways not possible for domestically ownedplants, by exploiting unique knowledge of their home markets. This type of activity wouldlead to a downward bias in our estimates. The estimate reported in column (3) shows thatour preferred estimate is robust to accounting for this explanation; including an indicator ofwhether a plant is owned by a foreign entity has no discernible effect.A second concern with our preferred estimate is that it may be driven by plants that do notsell domestically. While our theoretical framework considers the case where plants either onlysell domestically or sell both domestically and export, over 5% of our sample is comprised ofplants that do not sell in the Canadian market. In principle, the CWS could have a larger effect98on the total revenue of these plants because they are potentially at the largest disadvantagewhen they are regulated. Unlike plants that sell domestically and potentially compete withother plants that are regulated under the CWS, plants that do not sell in the domestic marketonly compete with foreign plants that are unaffected by the CWS. As a result, our baselineestimates could be simply capturing the effects of the CWS on this set of plants. However, asthe estimate reported in column (4) shows, restricting our sample to exclude the set of plantsthat only sell abroad has little effect on our results.A final concern with our preferred estimate is the possibility that the estimating equation ismis-specified because we have failed to account for persistence in revenues due to sunk costs.This is a particularly important consideration in our context because previous work studyingexport behaviour has emphasized the role of sunk costs in determining the export decisionsof plants (i.e. Roberts and Tybout (1997) or Bernard and Jensen (2004)). As such, we adaptthe approach taken by Bernard and Jensen (2004) to add a lagged dependent variable to ourestimating equation and estimate the resulting specification following the approach of Arellanoand Bond (1991). As columns (5) and (6) show, these estimates are similar in magnitude andsignificance to our preferred estimate.While the estimates reported in columns (1)-(6) provide robust evidence that the CWSparticulate matter caused revenues of affected plants to fall, it is important to note that theestimated coefficients report the average effect of regulation on affected plants. Hence, whilethese results are broadly supportive of Empirical Prediction 1, they do not reveal if the effectsof the CWS regulations differed across plants with different productivity levels.To address this issue, in column (7) we report estimates from a version of our preferredspecification where we allow the effects of the CWS to differ across plants according to theirinitial productivity bin. These estimates suggest that the the effects of the particulate matterstandard are concentrated at plants in the two lowest productivity quintiles. Moreover, theeffects are larger for plants closer to the middle of the initial productivity distribution as pre-dicted in Empirical Prediction 1; the estimates reported in rows two and three of column (7)indicate that the CWS caused total revenues to fall by 9.7% at plants in the lowest productivitybin but by 28.6% for plants in the second productivity bin. Recall that the least productiveplants in an industry are less likely to export. As a result, as Empirical Prediction 1 describes,they should experience a relatively small reduction in total revenues, in percentage terms,compared to the least productive exporters, as exports fall proportionally more than domesticsales.993.5.2 Export RevenueNext, we turn to examine the effects of the CWS on the revenue of exporting plants (Em-pirical Predictions 2 and 3). As such, we restrict our attention to the set of plants that arecontinuing exporters and export in all years.Our estimates of the effects of the CWS on the revenue of exporting plants are reported inTable 3.3 and Table 3.4. For the sake of comparison, we start by first re-estimating the effectsof the CWS on the natural log of total revenues for the set of plants that are continuing ex-porters. These estimates are reported in Table 3.3. Next, we estimate the effects of the CWS onthe natural log of export revenues. These estimates are reported in Table 3.4. The regressionsreported in each column of both tables correspond to the same column in Table 3.2. As such,column (1)-(6) report estimates based on equation (3.20). Column (1) only includes plant,industry-year and CMA-year fixed effects, while column (2) adds an indicator of whether theplant was regulated under the CWS O3 standard in a given year, column (3) adds an indicatorof foreign ownership and column (4) again restricts the sample to exclude the set of plants thatonly export and do not sell domestically. Columns (5) and (6) both include a lagged dependentvariable and are again estimated using the approach of Arellano and Bond (1991), using eitherone (column (5)) or two (column (6)) lags as instruments.Finally, column (7) reports estimates from a specification based on equation (3.21) andreports the estimated effects of the particulate matter standard by initial productivity bin. Tomake our results comparable across all tables, we maintain the same productivity bins used inTable 3.2. We maintain the same bins, as the goal behind allowing the effects to vary acrossproductivity levels is to trace-out the heterogeneity in plant-responses across the industry’sproductivity distribution. This means that B1 in the following two tables, for example, containsall continuing exporting plants that have initial productivity levels low enough to put them inthe first quintile of the productivity distribution defined by the full set of all PM2.5 emittingplants. Notice that because the least-productive plants in an industry are less-likely to export,there are relatively few observations in the first productivity bin in the following tables. As aresult, there is relatively little power in column (7) of both Table 3.3 and Table 3.4. 30The estimates reported in the first six columns of Table 3.3 are very similar to those re-ported in Table 3.2. For example, our preferred estimate indicates that the CWS is associatedwith a 10.7% reduction in total revenue from affected continuing exporter plants. This resultalso appears to be very robust; as the estimates reported in columns (3)-(6) show, controllingfor foreign ownership, restricting the sample to exclude plants that do not sell domestically30All regressions are weighted to correct for potential sample bias induced by the procedure used to matchthe NPRI to the ASM. Standard errors clustered by industry-CMA are reported in parentheses.100Table 3.3: Environmental Regulations and Revenue from Continuing Exporters(1) (2) (3) (4) (5) (6) (7)PM2.5 Std. -0.107c -0.107c -0.107c -0.106c -0.082c -0.083c(0.055) (0.056) (0.056) (0.056) (0.047) (0.047)PM2.5 Std.×B1 -0.019(0.120)PM2.5 Std.×B2 -0.331b(0.130)PM2.5 Std.×B3 -0.122(0.081)PM2.5 Std.×B4 -0.014(0.065)PM2.5 Std.×B5 -0.094(0.074)O3 Std. X X X X X XForeign Owner XRest. Sample XR2 0.322 0.322 0.323 0.316 0.324AR1 -2.563 -2.571AR2 1.274 1.289Obs. 4093 4093 4093 3807 2367 2367 4093Notes: Table reports estimates of the effects of the CWS on the natural log of plant revenues for plants that are continuing exporters. Allregressions include plant, industry-year and CMA-year fixed effects, and are weighted by the inverse of the match probability to controlfor potential match-induced sample bias. Column (3) includes an indicator of whether the plant is owned by a foreign company. Column(4) restricts the sample to exclude plants that only sell abroad. Columns (5) and (6) report Arellano-Bond estimates with one and two lags,respectively. In all cases, standard errors are clustered by CMA-industry. c, b, and a denote significance at the 10%, 5%, and 1% level,respectively.and allowing for lagged revenues to account for the possibility of sunk costs has little effecton our point estimates. The main difference between the results in these two tables is that thepoint estimates in Table 3.3 are less precise, owing to the fact that by focusing on continuingexporters we have fewer observations.Despite the fact that the restricting our sample to continuing exporters has little effect onour estimates of the average effect of the CWS particulate matter regulations, it appears thatthe effect is driven by a different set of plants. This can be seen from the estimates reportedin column (7) of Table 3.3. The particulate matter standard now appears to have little effecton plants with the lowest productivity levels. Instead the effect appears to be driven by plantsin the second and third productivity quantiles of the initial productivity distribution. The esti-mate reported in row three indicates the CWS particulate matter standard is associated with a101Table 3.4: Environmental Regulations and Export Revenue from Continuing Exporters(1) (2) (3) (4) (5) (6) (7)PM2.5 Std. -0.219b -0.219b -0.219b -0.242b -0.158 -0.157(0.055) (0.056) (0.056) (0.056) (0.108) (0.108)PM2.5 Std.×B1 -0.128(0.131)PM2.5 Std.×B2 -0.423a(0.156)PM2.5 Std.×B3 -0.243(0.164)PM2.5 Std.×B4 -0.154(0.144)PM2.5 Std.×B5 -0.182(0.160)O3 Std. X X X X X XForeign Owner XRest. Sample XR2 0.288 0.289 0.289 0.297 0.289AR1 -3.952 -3.952AR2 1.474 1.461Obs. 4093 4093 4093 3807 2367 2367 4093Notes: Table reports estimates of the effects of the CWS on the natural log of export revenues for plants that are continuing exporters. Allregressions include plant, industry-year and CMA-year fixed effects, and are weighted by the inverse of the match probability to controlfor potential match-induced sample bias. Column (3) includes an indicator of whether the plant is owned by a foreign company. Column(4) restricts the sample to exclude plants that only sell abroad. Columns (5) and (6) report Arellano-Bond estimates with one and two lags,respectively. In all cases, standard errors are clustered by CMA-industry. c, b, and a denote significance at the 10%, 5%, and 1% level,respectively.33.1% reduction in revenues from plants in the second productivity quintile. Furthermore, theestimate reported in row four suggests the standard reduced revenues at plants in the third pro-ductivity quintile by 12.2%, although this effect is imprecisely estimated and not statisticallysignificant at conventional levels. These are intuitive findings, as our model suggests the leastproductive plants in an industry should not export. As a result, there should be relatively fewobservations in the first productivity bin when we restrict the sample to continuing exporters.31The estimates reported in columns (1)-(6) of Table 3.4 suggest that the CWS led to a sig-nificant reduction in export revenue from affected plants. For example, our preferred estimate,reported in column (2), indicates that the CWS particulate matter standard is associated witha 21.9% reduction in export revenue from affected manufacturing plants. This estimate also31Recall that we maintain the same productivity bin groupings, regardless of sample restriction.102appears to be quite robust; controlling for foreign ownership (column (3)) or excluding plantsthat only export (column (4)) has little effect on the estimated effect of regulation. The esti-mated coefficients reported in columns (5) and (6) also suggest that the CWS had a negativeeffect on exporting, however, these estimates are imprecisely estimated and not statisticallysignificant at conventional levels. The loss of precision is likely owing to the considerablereduction in sample size from this restriction.The estimates reported in column (7) suggest that the reduction in export revenue is drivenby the responses of plants in the second and third productivity quintiles. The estimate reportedin row three indicates that the CWS particulate matter standard reduced the export revenue ofplants in the second productivity quintile by 42.3%. The estimate reported in row four suggeststhat the standard reduced export revenues at plants in the third productivity quintile by 24.3%,but as is the case with total revenues, this effect is imprecisely estimated and not statisticallysignificant at conventional levels.Taken together, the estimates reported in Table 3.3 and Table 3.4 are supportive of Empir-ical Predictions 2 and 3. The estimates indicate that the CWS had a much larger effect onthe export revenue than the total revenue of affected plants, which is consistent with environ-mental regulation causing a larger reduction in export revenues than domestic revenues, aspredicted by our model. In addition, the effects of regulation are concentrated on plants in thesecond and third productivity quintiles, which is consistent with the model’s prediction thatthe effects of regulation will be largest at the least productive exporters.3.5.3 Export StatusFinally, we examine the effects of the CWS on plant export status (Empirical Prediction4). If the costs associated with environmental regulation are large enough to reduce the totalrevenues of affected plants, as we have shown above, then our model suggests that some plantsshould exit the export market. To test this prediction, we again turn to examine our full sampleof plants, that includes both exporting and non-exporting plants.The results of estimating a linear probability model of the effects of the CWS particu-late matter regulations on an indicator of plant export status are presented in Table 3.5. Weagain report estimates from seven specifications. The first six, reported in columns (1)-(6)respectively, are based on equation (3.20), and report the average effect of particulate matterregulation on export status at affected plants. The seventh specification, reported in column(7), is based on equation (3.21), and reports the effects of the regulation by initial productivityquintile. Again, column (1) includes plant, industry-year and CMA-year fixed effects only,while column (2) adds an indicator of whether the plant was regulated under the CWS O3103Table 3.5: Environmental Regulations and Export Status(1) (2) (3) (4) (5) (6) (7)PM2.5 Standard -0.018 -0.018 -0.019 -0.013 -0.055 -0.053(0.026) (0.026) (0.026) (0.026) (0.038) (0.038)PM2.5 Std.×B1 -0.008(0.037)PM2.5 Std.×B2 -0.102b(0.045)PM2.5 Std.×B3 0.020(0.033)PM2.5 Std.×B4 0.051(0.060)PM2.5 Std.×B5 -0.003(0.058)O3 Standard X X X X X XForeign Ownership XRestricted Sample XR2 0.129 0.129 0.130 0.140 0.131AR1 -2.545 -2.702AR2 1.225 1.325Observations 6501 6501 6501 6149 3694 3694 6501Notes: Table reports estimates of the effects of the CWS on an indicator of plant export status. All regressions include plant, industry-yearand CMA-year fixed effects, and are weighted by the inverse of the match probability to control for potential match-induced sample bias.Column (3) includes an indicator of whether the plant is owned by a foreign company. Column (4) restricts the sample to exclude plants thatonly sell abroad. Columns (5) and (6) report Arellano-Bond estimates with one and two lags, respectively. In all cases, standard errors areclustered by CMA-industry. c, b, and a denote significance at the 10%, 5%, and 1% level, respectively.standard in a given year, column (3) adds an indicator of foreign ownership and column (4)restricts the sample to exclude the set of plants that only export and do not sell domestically.Columns (5) and (6) both include a lagged dependent variable and are again estimated usingthe approach of Arellano and Bond (1991), using either one (column (5)) or two (column (6))lags as instruments.32The estimates reported in the first six columns of Table 3.5 suggest that the average effectof the CWS on affected plants is small and insignificant. For example, our preferred estimate,reported in column (2), shows that the CWS only reduced the likelihood of a plant exportingby less than 2%. Not only is this effect economically small, it is statistically insignificant. Thisis still true when we allow for differences in foreign ownership, restrict our sample to exclude32All regressions are weighted to correct for potential sample bias induced by the procedure used to matchthe NPRI to the ASM. Standard errors clustered by industry-CMA are reported in parentheses.104foreign exporters, and allow for the possibility of sunk costs in exporting using a dynamicpanel specification.While our estimates of the average effect of the CWS on affected plants are small andinsignificant, the estimates reported in column (7) suggest that they mask substantial hetero-geneity in how plants respond to regulation. Specifically, the estimates reported in column(7) show a 10% reduction in the likelihood of exporting for plants in the second quintile ofthe plant productivity distribution as a result of the CWS. This result is consistent with ourmodel’s prediction that the least productive exporting firms will exit the export market in re-sponse to regulation. The intuition behind this heterogeneity, as our model highlights, is thatthe marginal exporter should lie somewhere in the middle of an industry’s productivity distri-bution. As a result, exit from exporting should be restricted to this group of firms.3.6 ConclusionIn this chapter we present plant-level evidence of the heterogeneous effects of environ-mental regulation on the international competitiveness of domestic industry. The concern thatstringent environmental policy may impede the competitiveness of domestic producers, re-ferred to as the Pollution Haven Effect (PHE), has been present in debates on environmentalregulation for decades (Jaffe et al., 1995). The PHE stems from the observation that policyimposed unilaterally by one country should increase production costs for domestic producersrelative to their unregulated foreign counterparts.Thus far, the empirical literature on the PHE has focused exclusively on the industry-or region-level effects of regulation. We argue, however, that if producers have differingproductivity-levels, then the PHE should vary across firms or plants on the basis of their pro-ductivity. This heterogeneity has implications for both identifying the PHE, and understandingthe mechanisms through which regulation disadvantages domestic industry.33We use a simple model to clarify the logic underpinning the heterogeneous nature of thePHE. In this model, firms that differ on the basis of their productivity levels compete via mo-nopolistic competition and face a two-part regulatory constraint on their pollution emissions.The model shows that, for an exporting firm of a given productivity level, regulation shouldcause a larger reduction in exports relative to domestic sales because domestic prices adjust toalleviate the effects of regulation. Moreover, under the type of policy we study, there should33Note that this heterogeneity is particularly pronounced under the type of policy we study, which is a commonform of air pollutant regulation (for example, the US Clean Air Act features this type of policy). This policyimposes a two-part regulatory constraint, that requires firms to adopt cleaner production processes, and penalizesthose that fail to do so.105be a U-shaped relationship between the effect of regulation on firm revenues and the firm’sproductivity. This occurs because only relatively productive firms select into exporting, whichexposes them to larger loses from regulation, but the most productive of these firms adopt newtechnology to avoid increased production costs from regulation. Finally, we show that regula-tion causes some firms to leave the export market, and these firms should be in the middle ofthe industry’s productivity distribution.Using a unique dataset that contains plant-level production and pollution information, wetest for the PHE by examining the effects of a major Canadian environmental policy on theexport participation, export volumes, and total sales of Canadian manufacturing plants overthe period 2004-2010. This policy, called the Canada Wide Standards for Particulate Matterand Ozone (CWS), implemented regional air quality standards in every major town or city ofCanada. In addition, the CWS explicitly targeted plants in a select group of industries. As aresult, the CWS created variation in regulatory stringency across industries, regions, and time,which we exploit using a triple-difference research design.We find the CWS had a large negative effect on both total sales and export volumes, causingan 11% reduction in total sales for the average affected plant and a 22% reduction in exportvolumes for the average affected exporter. The policy, however, had no significant effect onthe selection into or out of exporting for the average plant. As our model predicts, the effectsof regulation varied considerably across plants of different productivity levels. Our resultsshow a U-shaped relationship between a plant’s productivity and the effect of regulation onboth total sales and export volumes. Finally, in line with our model’s prediction, the CWSreduced the probability of exporting by 10% for moderately productive plants.Taken together, these results suggest that environmental regulations that increase variableproduction costs may reduce the international competitiveness of affected plants. This find-ing is consistent with recent empirical work on the PHE effect, such as Levinson and Taylor(2008), Shi and Xu (2018), and Broner et al. (2012). However, unlike the previous literature,we show evidence that regulation affects trade flows using plant-level data, rather than ag-gregate data. Importantly, our results also show that, at least in the context of a CWS-typeregulation, there seems to be considerable heterogeneity in the PHE; exporters that feel thecompetitiveness effects of regulation are in the middle of their industry’s productivity distri-bution.106ConclusionThis thesis provides new theoretical and empirical evidence of how environmental regulationaffects manufacturing facilities. This new evidence is important for informing debates onenvironmental policy, both on the effects these policies have on the economy, and on thedesign of regulations targeting firms. In addition, this new evidence provides insights into theworkings of firms, and in particular on how firms respond to regulatory constraints.In the first chapter, we present a theoretical model to show the channels through whichenvironmental regulation causes a reduction in an industry’s pollution intensity (measured asthe amount of pollution emitted per dollar of output). We study regulation that imposes atwo-part regulatory constraint on firms: they must either adopt clean production processes, orface a penalty. While a common form of environmental policy, two-part regulatory constraintshave not been studied in the theoretical literature on environmental regulation.Our model shows that this type of regulation causes the least productive firms in an industryto exit, which we call the selection channel, and low-productivity surviving firms to produceless (a reallocation channel). If these affected firms are relatively pollution intensive, thenthese channels will serve to reduce an industry’s pollution intensity. In addition, this type ofregulation causes moderately-productive firms to adopt cleaner production processes, whichreduces an industry’s pollution intensity through what we call the process channel.In the second chapter, we estimate the regulatory channels of the manufacturing clean-up. While there is much indirect evidence that suggests regulation may affect an industry’spollution intensity through several channels, work has yet to directly estimate the magnitudeof these channels. In this chapter, we use a novel confidential dataset that contains plant-levelpollution and production information for major manufacturing polluters in Canada to estimatethe three plant-level channels through which regulation contributes to a clean-up. With thisdata, we estimate the effect of a major revision to environmental policy in Canada, called theCanada Wide Standards for Particulate Matter and Ozone, on plant pollution intensity, output,and entry and exit decisions. This policy created variation in regulatory stringency acrossindustries, regions, and time, which we use to identify the effect of environmental regulation.107We find the Canada Wide Standards played a sizeable role in the Canadian manufacturingclean-up. From 2004 to 2010, this policy explains approximately 60% of the observed reduc-tion in sector-level nitrogen oxide emissions and 20% of the reduction in particulate matteremissions. Moreover, the channels involved in the CWS clean-up varied starkly across pollu-tants. The clean up of nitrogen oxide was primarily caused by the process channel, while thechannels driving the particulate matter clean-up were primarily selection and reallocation. Weargue these differences arise because of differences across pollutants in the fixed costs plantsneed to pay to adopt cleaner production processes, and show additional empirical evidenceconsistent with this hypothesis.In the third chapter, we ask whether environmental regulation affects the international com-petitiveness of manufacturing plants. By raising the costs of production in domestic marketsrelative to unregulated foreign markets, environmental regulation may disadvantage domesticproducers relative to their foreign counterparts. This hypothesis, referred to as the PollutionHaven Effect (PHE), is typically tested by examining regulation’s effect on the internationalflow of goods. Thus far, the literature on this topic has assessed the PHE at the industry orregional level. We contribute to this literature by asking how regulation affects the exportdecisions of individual plants.We provide a theoretical framework to show how regulation affects a plant’s decision ofwhether to export, as well as their export volumes. Our theoretical framework shows thata regulation that imposes a two-part regulatory constraint on firms should have differentialeffects on firms of different productivity levels. In particular, the PHE should be strongest forthe least-productive exporters.Finally, we estimate the effect of the same policy-change studied in the second chapter,the Canada Wide Standards, on the export behaviours of Canadian manufacturing plants. Inline with our theory’s predictions, we find the PHE appears strongest for the least produc-tive exporting plants. Those low-productivity exporters experience a 10% reduction in theprobability of exporting, as well as a 40% reduction in export sales.108BibliographyAichele, R. and G. Felbermayr (2015). Kyoto and carbon leakage: an empirical analysis ofthe carbon content of bilateral trade. Review of Economics and Statistics 97(1), 104–115.Ambec, S., M. a. Cohen, S. Elgie, and P. Lanoie (2013). The Porter Hypothesis at 20: CanEnvironmental Regulation Enhance Innovation and Competitiveness? Review ofEnvironmental Economics and Policy 7(1), 2–22.Andersen, D. C. (2018). Accounting for loss of variety and factor reallocations in the welfarecost of regulations. Journal of Environmental Economics and Management 88, 69–94.Angle, R. (2014). Ambient air quality objectives. In E. Taylor and A. McMillan (Eds.), AirQuality Management Canadian Perspectives on a Global Issue, Chapter 14, pp. 289–301.Dordrecht Heidelberg New York London: Springer.Anoulies, L. (2017). Heterogeneous firms and the environment: A cap-and-trade program.Journal of Environmental Economics and Management 84, 84–101.Antweiler, W., B. R. Copeland, and M. S. Taylor (2001). Is free trade good for theenvironment? American Economic Review 91(4), 877–908.Applied Technologies of New York Inc. (2018). Case Study: East River Housing.http://www.atiofny.com/case-study-east-river-housing/. [Online; accessed 5-May-2018].Arellano, M. and S. Bond (1991). Some tests of specification for panel data: Monte carloevidence and an application to employment equations. The Review of EconomicStudies 58(2), 277–297.Barnatchez, K., L. D. Crane, and R. A. Decker (2017). An Assessment of the NationalEstablishment Time Series (NETS) Database.Barrows, G. and H. Ollivier (2018). Cleaner firms or cleaner products? how product mixshapes emission intensity from manufacturing. Journal of Environmental Economics andManagement 88, 134–158.Batrakova, S. and R. B. Davies (2012). Is there an environmental benefit to being anexporter? Evidence from firm-level data. Review of World Economics 148(3), 449–474.109Becker, R. and V. Henderson (2000). Effects of air quality regulation on polluting industries.Journal of Political Economy 108(2), 379–421.Berman, E. and L. Bui (2001a, feb). Environmental regulation and labor demand: evidencefrom the South Coast Air Basin. Journal of Public Economics 79(2), 265–295.Berman, E. and L. Bui (2001b). Environmental regulation and productivity: Evidence fromoil refinieries. Review of Economics and Statistics 83(3), 498–510.Bernard, A. B., J. Eaton, J. B. Jensen, and S. Kortum (2003). Plants and productivity ininternational trade. American Economic Review 93(4), 1268–1290.Bernard, A. B. and J. B. Jensen (2004). Why some firms export. Review of Economics andStatistics 86(2), 561–569.Besley, T. and A. Case (2000). Unnatural Experiments? Estimating the Incidence ofEndogenous Policies. The Economic Journal 110(467), 672–694.Brankov, E., R. F. Henry, K. L. Civerolo, W. Hao, S. Rao, P. Misra, R. Bloxam, and N. Reid(2003). Assesing the effects of transboundary ozone pollution between ontario, canada andnew york, usa. Environmental Pollution 123, 403–411.Broner, F., P. Bustos, and V. Carvalho (2012). Sources of comparative advantage in pollutingindustries. NBER Working Paper 18337.Brunel, C. (2016). Pollution offshoring and emissions reductions in e.u. and u.s.manufacturing. Environmental and Resource Economics.Bustos, P. (2011). Trade liberalization, exports, and technology upgrading: Evidence on theimpact of MERCOSUR on Argentinian firms. American Economic Review 101(1),304–340.Canadian Council of Ministers of the Environment (1998a). National Emission Guideline forCement Kilns. https://www.ccme.ca/files/Resources/air/emissions/pn 1284 e.pdf.Canadian Council of Ministers of the Environment (1998b). National Emission Guideline forCommercial/Industrial Boilers and Heaters.http://www.ccme.ca/files/Resources/air/emissions/pn 1286 e.pdf. [Online; accessed25-September-2017].Canadian Council of Ministers of the Environment (2000a). Canada-Wide Standards forParticulate Matter and Ozone.http://www.ccme.ca/files/Resources/air/pm ozone/pmozone standard e.pdf.Canadian Council of Ministers of the Environment (2000b). Joint Initial Actions to ReducePollutant Emissions That Contribute to Particulate Matter and Ground-level Ozone.http://www.ccme.ca/files/Resources/air/pm ozone/pmozone joint actions e.pdf/.110Canadian Council of Ministers of the Environment (2002). Guidance Document onAchievement Determination: Canada-wide Standards for Particulate Matter and Ozone.http://www.ccme.ca/files/Resources/air/pm ozone/pn 1330.pdf.Canadian Council of Ministers of the Environment (2006). Canada-wide Standards forParticulate Matter and Ozone: Five Year Report: 2000-2005.https://www.ccme.ca/files/Resources/air/pm ozone/pm oz 2000 2005 rpt e.pdf.Cao, J., L. D. Qiu, and M. Zhou (2016). Who invests more in advanced abatementtechnology? theory and evidence. Canadian Journal of Economics 49(2), 637–662.Chang, T., J. G. Zivin, T. Gross, and M. Neidell (2016). Particulate pollution and theproductivity of pear packers. American Economic Journal: Economic Policy 8(3),141–169.Chay, K. Y. and M. Greenstone (2003). The Impact of Air Pollution on Infant Mortality:Evidence from Geographic Variation in Pollution Shocks Induced by a Recession.Quarterly Journal of Economics 118, 1121–1167.Cherniwchan, J. (2017). Trade liberalization and the environment: Evidence from NAFTAand U.S. Manufacturing. Journal of International Economics 105, 130–149.Cherniwchan, J., B. R. Copeland, and M. S. Taylor (2017). Trade and the environment: Newmethods, measurements, and results. Annual Review of Economics 9, 59–85.Copeland, B. and M. Taylor (2004). Trade, growth, and the environment. Journal ofEconomic Literature XLII, 7–71.Cui, J., H. Lapan, and G. Moschini (2012). Are Exporters More Environmentally Friendlythan Non-Exporters? Theory and Evidence.Cui, J., H. Lapan, and G. Moschini (2015). Productivity, Export, and EnvironmentalPerformance: Air Pollutants in the United States. American Journal of AgriculturalEconomics 98.Currie, J. and M. Neidell (2005). Air Pollution and Infant Health: What Can We Learn FromCalifornia’s Recent Experience? Quarterly Journal of Economics 120(3), 1003–1030.Demidova, S. and A. Rodrı´guez-Clare (2009). Trade policy under firm-level heterogeneity ina small economy. Journal of International Economics 78(1), 100–112.Demidova, S. and A. Rodrı´guez-Clare (2013). The simple analytics of the Melitz model in asmall economy. Journal of International Economics 90(2), 266–272.Ederington, J., A. Levinson, and J. Minier (2005). Footloose and pollution-free. Review ofEconomics and Statistics 87(February), 92–99.111Environment and Climate Change Canada (2013). National Air Pollution SurveillanceProgram (NAPS). http://www.ec.gc.ca/rnspa-naps/. [Online; accessed 15-August-2015].Environment and Climate Change Canada (2016a). National Pollutant Release Inventory(NPRI) Toolbox. https://www.ec.gc.ca/inrp-npri/default.asp?lang=En&n=65A75CDF-1.[Online; accessed 28-June-2016].Environment and Climate Change Canada (2016b). Reported Facility Greenhouse Gas Data.http://www.ec.gc.ca/ges-ghg/default.asp?lang=En&n=8044859A-1. [Online; accessed28-June-2016].Environment and Climate Change Canada (2016c). Summary of National Pollutant ReleaseInventory Reporting requirements.https://www.ec.gc.ca/inrp-npri/default.asp?lang=en&n=629573FE-1. [Online; accessed28-June-2016].Environment Canada (2002). Multi-Pollutant Emission Reduction Analysis Foundation(MERAF) for the Iron and Steel Sector. Unpublished report.Environment Canada (2005). Great Lakes Basin Airshed Management Framework PilotProject. https://www.epa.gov/sites/production/files/2015-07/documents/great lakes basin airshed management framework.pdf. [Online; accessed25-September-2017].Environment Canada and Forest Products Association of Canada (2004). Towards MoreInnovative Air Quality Management: Proposal for a Pulp and Paper Air Quality Forum.http://publications.gc.ca/site/eng/274270/publication.html.Environmental Protection Agency (1998). Particulate Matter Controls. Baghouses andFilters. EPA/452/B-02-001. https://www3.epa.gov/ttncatc1/cica/files/cs6ch1.pdf. [Online;accessed 29-August-2017].Environmental Protection Agency (1999a). Nitrogen Oxides (NOx), Why and How they areControlled. EPA-456/F-99-006R. https://www3.epa.gov/ttncatc1/cica/files/fnoxdoc.pdf.[Online; accessed 29-August-2017].Environmental Protection Agency (1999b). Particulate Matter Controls. ElectrostaticPrecipitators. EPA/452/B-02-001. https://www3.epa.gov/ttncatc1/cica/files/cs6ch3.pdf.[Online; accessed 29-August-2017].Environmental Protection Agency (2002). Particulate Matter Controls. Wet Scrubbers forParticulate Matter. EPA/452/B-02-001.https://www3.epa.gov/ttncatc1/cica/files/cs6ch2.pdf. [Online; accessed 29-August-2017].Environmental Protection Agency (2006). Chapter 3. Regulatory Impact Analysis for theParticulate Matter NAAQS.https://www3.epa.gov/ttn/ecas/regdata/RIAs/Chapter{%}203--Controls.pdf. [Online;accessed 3-November-2017].112Environmental Protection Agency (2016). NAAQS Table.https://www.epa.gov/criteria-air-pollutants/naaqs-table/. [Online; accessed10-March-2016].Forslid, R., T. Okubo, and K. H. Ulltveit-Moe (2014). Why are Firms that Export Cleaner?International Trade and CO2 Emissions. CESifo Working Paper 4817.Foster, L., J. Haltiwanger, and C. J. Krizan (2001). Aggregate Productivity Growth: Lessonsfrom Microeconomic Evidence. In C. Hulten, E. Dean, and M. Harper (Eds.), NewDevelopments in Productivity Analysis. University of Chicago Press.Fowlie, M., S. P. Holland, and E. T. Mansur (2012). What do emissions market deliver and towhom? Evidence from Southern California NOx trading program. American EconomicReview 102, 965–993.Fowlie, M. L., M. Reguant, and S. P. Ryan (2016). Measuring Leakage Risk. Mimeo.Gibson, M. (2016). Regulation-induced pollution substitution. Working Paper 2016-04,Williams College.Government of Canada (2003). Clean Air in Canada: Progress Report on Particulate Matterand Ozone. http://publications.gc.ca/pub?id=9.664805&sl=0. [Online; accessed25-September-2017].Graff Zivin, J. and M. Neidell (2012). The impact of pollution on worker productivity.American Economic Review 102(7), 3652–3673.Greenstone, M. (2002). The impacts of environmental regulations on industrial activity:Evidence from the 1970 & 1977 Clean Air Act amendments and the census ofmanufacturers. Journal of Political Economy 110(6), 1175–1219.Greenstone, M. (2003). Estimating Regulation-Induced Substitution: The Effect of the CleanAir Act on Water and Ground Pollution. American Economic Review: Papers andProceedings 93(2).Greenstone, M., J. A. List, and C. Syverson (2012). The effects of environmental regulationon the competitiveness of U.S. manufacturing. Working Paper 18392, National Bureau ofEconomic Research.Grether, J.-M., N. A. Mathys, and J. de Melo (2009). Scale, technique and compositioneffects in manufacturing so2 emissions. Environmental and Resource Economics 43(2),257–274.Haltiwanger, J. C., R. S. Jarmin, and J. Miranda (2013). Who Creates Jobs? Small vs. Largevs. Young. Review of Economics and Statistics 95(2).113Hanna, R. (2010). U.S. environmental regulation and FDI: Evidence from a panel ofUS-based multinational firms. American Economic Journal: Applied Economics 2(3),158–89.Hanna, R. and P. Oliva (2015). The effect of pollution on labor supply: Evidence from anatural experiment in Mexico City. Journal of Public Economics 122, 68–79.Henderson, J. V. (1996). Effects of air quality regulation. The American EconomicReview 86(4), 789–813.Holladay, J. S. (2016). Exporters and the Environment. Canadian Journal of Economics 49.International Joint Commission (2002). Agreement Between the Government of Canada andthe Government of the United States of America on Air Quality.http://www.ijc.org/rel/agree/air.html.Jaffe, A. B., S. R. Peterson, P. R. Portney, and R. N. Stavins (1995). EnvironmentalRegulation and the Competitiveness of U.S. Manufacturing: What does the evidence tellus? Journal of Economic Literature 33(1), 132–163.Johnson, D., D. Mignacca, D. Herod, D. Jutzi, and H. Miller (2007). Characterization andidenification of trends in average ambient ozone and fine particulate matter levels throughtrajectory cluster analysis in eastern canada. Journal of the Air & Waste ManagementAssociation 58(8), 907–918.Kellenberg, D. K. (2009). An empirical investigation of the pollution haven effect withstrategic environment and trade policy. Journal of International Economics 78(2),242–255.Knight, C. (2013). What is grandfathering? Environmental Politics 22(3), 410–427.Konishi, Y. and N. Tarui (2015). Emissions trading, firm heterogeneity, and intra-industryreallocations in the long ru. Journal of the Association of Environmental and ResourceEconomists 2(1), 1–42.Lanoie, P. and A. Rochon-Fabien (2012). Promoting Pollution Prevention in SmallBusinesses: Costs and Benefits of the ”Enviroclub” Initiative. Canadian PublicPolicy 38(2), 217–232.Lemoine, D. (2017). Escape from Third-Best: Rating Emissions for Intensity Standards.Environmental and Resource Economics 67(4), 789–821.Levinson, A. (1996). Environmental regulations and industry location: International anddomestic evidence. In J. Bhagwati and R. Hudec (Eds.), Fair Trade and Harmonization:Prerequisite for Free Trade? MIT Press.Levinson, A. (2009). Technology, international trade, and pollution from u.s. manufacturing.American Economic Review 99(5), 2177–2192.114Levinson, A. (2015). A direct estimate of the technique effect: Changes in the pollutionintensity of u.s. manufacturing, 1990-2008. Journal of the Association of Environmentaland Resource Economists 2(1), 43–56.Levinson, A. and M. S. Taylor (2008, feb). Unmasking the Pollution Haven Effect.International Economic Review 49(1), 223–254.Li, Z. and S. Shi (2017). Emission taxes and standards in a general equilibrium withproductivity dispersion and abatement. Macroeconomic Dynamics 21(8), 1857–1886.Li, Z. and J. Sun (2015). Emission taxes and standards in general equilibrium with entry andexit. Journal of Economic Dynamics and Control 61, 34–60.List, J. A., D. L. Millimet, P. G. Fredriksson, and W. W. McHone (2003). Effects ofenvironmental regulations on manufacturing plant births: Evidence from a propensityscore matching estimator. Review of Economics and Statistics 85(4), 944–952.Martin, L. A. (2012). Energy efficiency gains from trade: Greenhouse gas emissions andindia’s manufacturing sector. Mimeo.Martin, R., L. B. de Preux, and U. J. Wagner (2014). The impact of a carbon tax onmanufacturing: Evidence from microdata. Journal of Public Economics 117, 1–14.McAusland, C. and N. Najjar (2015). Carbon Footprint Taxes. 61(1), 37–70.Melitz, M. J. (2003). The Impact of Trade on Intra-Industry Reallocations and AggregateIndustry Productivity. Econometrica 71(6), 1695–1725.Millimet, D. and J. Roy (2016). Empirical Tests of the Pollution Haven Hypothesis WhenEnvironmental Regulation is Endogenous. Journal of Applied Econometrics 31, 652–677.Roberts, M. J. and J. R. Tybout (1997). The decision to export in colombia: An empiricalmodel of entry with sunk costs. The American Economic Review 87(4), 545–564.Schlenker, W. and W. R. Walker (2016). Airports, air pollution, and contemporaneous health.Review of Economic Studies 83(2), 768–809.Shapiro, J. S. and R. Walker (2015). Why is Pollution from U.S. Manufacturing Declining?The Roles of Trade, Regulation, Productivity, and Preferences.Shi, X. and Z. Xu (2018). Environmental regulation and firm exports: Evidence from theeleventh Five-Year Plan in China. Journal of Environmental Economics andManagement 89, 187–200.Tang, E., J. Zhang, and Z. Haider (2015). Firm productivity, pollution, and output: theoryand empirical evidence from China. Environmental Science and Pollution Research 22,18040–18046.115Walker, R. (2013). The transitional costs of sectoral reallocation: Evidence from the CleanAir Act and the workforce. Quarterly Journal of Economics 128(4), 1787–1835.World Bank Group (1998a). Airborne Particulate Matter: Pollution Prevention and Control.https://www.ifc.org/wps/wcm/connect/ab848080488557a8bd44ff6a6515bb18/HandbookAirborneParticulateMatterPollutionPreventionAndControl.pdf?MOD=AJPERES.World Bank Group (1998b). Nitrogen Oxides: Pollution Prevention and Control.https://www.ifc.org/wps/wcm/connect/60add78048855368aeecfe6a6515bb18/HandbookNitrogenOxidesPollutionPreventionAndControl.pdf?MOD=AJPERES.Yujian, L. and L. Bo (2007). A normalized Levenshtein distance metric. IEEE Transactionson Pattern Analysis and Machine Intelligence 29(6), 1091–1095.116Appendix ASupporting MaterialsA.1 Chapter 1 AppendixA.1.1 Proof of Proposition 3Consider firms that retrofit from technology b to technology r. For these firms, the effectof regulation can be determined by comparing equations (1.28) with (1.31):rtarr (ϕ)rnb(ϕ)=[ϕnoε [1+ τκ]ϕ tarε]σ−1(A.1)Note thatϕnε [1+ τκ]ϕ tarε=[1+ τκ]1k[1+∆kσ−11[ffs] k−σ+1σ−1]1+[1+ τκ]k[∆kσ−12[ffr] k−σ+1σ−1+∆kσ−11[ffs− fr] k−σ+1σ−1]k,117which is less than one if, and only if, the fixed cost of production f is large enough. That is, fmust satisfyfk−σ+1σ−1 >[[1+ τκ]1k −1[1+ τκ]1k][[[1+ τκ]σ−1−1] kσ−1[1+ τκ]1k[ 1fr] k−σ+1σ−1+[[1+ τκ]k2−1k[fsfs− fr] k−σ+1σ−1−1]∆kσ−11fk−σ+1σ−1s]−1.Hence, regulation will reduce revenues at firms that retrofit the business-as-usual technologyif the fixed cost of production are sufficiently high.For facilities that use state-of-the-art technology in both regimes, the effects of regulationcan be determined by comparing equations (1.29) and (1.32). This yieldsrtars (ϕ)rnos (ϕ)=[ϕnoε [1+ τκ]ϕ tarε]σ−1, (A.2)which is the same as the condition for retrofitting facilities given above. Thus, regulationreduces revenues for these facilities if and only if the fixed cost of production, f , is sufficientlyhigh.If ϕ tars < ϕnos , then regulation causes some facilities to switch from the business-as-usualtechnology to the state-of-the-art technology. In this casertars (ϕ)rnob (ϕ)=[ϕnoε [1+ τκ]αϕ tarε]σ−1. (A.3)This is greater than one if and only iffk−σ+1σ−1 >[α1k [1+ τκ]1k −1α1k [1+ τκ]1k][[[1+ τκ]σ−1−1] kσ−1[1+ τκ]1k[1fr] k−σ+1σ−1 1α1k+[[1+ τκ]k2−1k[fsfs− fr] k−σ+1σ−1−α 1k]∆kσ−11fk−σ+1σ−1s1α1k]−1,(A.4)which must be satisfied if ϕ tars > ϕnos (the only condition under which this scenario is plau-sible). Hence, if ϕ tars > ϕnos , revenues rise at firms that switch from the business-as-usual118technology to the state-of-the-art technologyIf ϕ tars > ϕnos , then regulation causes some facilities to downgrade from state-of-the-art tothe retrofitted technology. For these firmsrtarr (ϕ)rns (ϕ)=[ϕnε [1+ τκ]ϕ tarε α]σ−1, (A.5)which is less than one if and only iffk−σ+1σ−1 >[[1+ τκ]1k −α 1k[1+ τκ]1k][[[1+ τκ]σ−1−1] kσ−1[1+ τκ]1k[ 1fr] k−σ+1σ−1α1k+[[1+ τκ]k2−1k[fsfs− fr] k−σ+1σ−1− 1α1k]∆kσ−11fk−σ+1σ−1sα1k]−1.(A.6)Notice that this cut-off value for f is lower than that required to ensure revenues for retrofittersfalls. As such, there is a range of values for f for which facilities that retrofit business-as-usualtechnology experience an increase in revenue while those that switch from state-of-the-arttechnology to the retrofitted technology experience a reduction in revenue. Note also thatimposing α > [1+ τκ] is sufficient to guarantee rtarr (ϕ)/rns (ϕ)< 1.A.1.2 Proof of Corollary 1Notice that Equation (1.33) implies[ϕnoεϕtarε]σ−1=rtarb (ϕ)rnob (ϕ). Substituting this into Equation (A.1)and Equation (A.2) givesrtarr (ϕ)rnb(ϕ)=rtars (ϕ)rnos (ϕ)=[rtarb (ϕ)rnob (ϕ)][1+ τκ]σ−1 ,which means the relative reduction in revenues for firms that always use business-as-usualtechnology is larger than that of the firms that retrofit or always use state-of-the-art technology.If ϕ tars < ϕnos , then some firms switch from business-as-usual technology to the state-of-the-art technology. Equation (1.33) shows that revenues rise for these firms.If ϕ tars > ϕnos , then regulation causes some facilities to downgrade from state-of-the-art to119the retrofitted technology. Substituting Equation (1.33) into Equation (A.5) givesrtarr (ϕ)rns (ϕ)=[rtarb (ϕ)rnob (ϕ)][1+ τκα]σ−1, (A.7)which is less than one by assumption. Notice that if we allowed α to be greater than 1+ τκ ,then firms with ϕ ∈ [ϕnos ,ϕ tars ] would experience the largest reduction in revenues.A.1.3 Proof of Proposition 5Differentiating the retrofitting cut-off with respect to fr gives∂ϕ tarr∂ fr=[11+ τκ][fr∆2 f] 1σ−1[∂ϕ tarε∂ fr+[1σ −1]ϕ tarεfr]=[11+ τκ][fr∆2 f] 1σ−1[1k[ϕ tarε ]k∂ [ϕ tarε ]k∂ fr+[1σ −1]ϕ tarεfr].(A.8)Thus, ∂ϕtarr∂ fr > 0 if and only if∂ [ϕ tarε ]k∂ fr>−[kσ −1][ϕ tarε ]kfr, (A.9)where ∂ [ϕtarε ]k∂ fr = [fδ f ][1+τκ]k fk−σ+1σ−1[[ ∆1fs− fr ]kσ−1 − [∆2fr ]kσ−1]. One can show that Equation (A.9)reduces to[kk−σ −1]1fr[1f] k−σ+1σ−1[11+ τκ]σ−1+[1+[kk−σ −1]1fr1fs− fr][∆1fs− fr] kσ−1>−[σ −1k−σ +1][∆2fr] kσ−1,(A.10)which is always satisfied.In addition, lowering fr lowers the exit cut-off under the regulation regime if fs isn’t too120large. Differentiating ϕ tarε with respect to fr gives∂ϕ tarε∂ fr=[σ −1k−σ +1] 1k[fδ fε] 1k [Λtar] 1−kk [1+ τκ]k[k−σ +1σ −1]fk−σ+1σ−1∆ k−σ+1σ−12 [ 1fs− fr] k−2[σ−1]σ−1−∆k−σ+1σ−11[1fr] k−2(σ−1)σ−1 , (A.11)which is greater than zero if and only if fs <[1+[∆1∆2 ]k−[σ−1]k−2[σ−1]]fr. Note that if k > 2[σ − 1]this means the model requires both a maximum and minimum constraint on fs to produce theabove result and maintain ϕ tarr < ϕ tars . If k < 2[σ −1], then imposing fs > [∆1+∆2∆2 ] fr ensuresboth results.A.1.4 Proof of Corollary 2Note that the ratio of ϕ tarr to ϕ tarε reflects the measure of firms that use business-as-usualtechnology in the regulated equilibrium. This ratio is given byϕ tarrϕ tarε=[11+ τκ][fr∆2 f] 1σ−1. (A.12)The derivative of Equation (A.12) with respect to fr is positive, which means lowering frlowers the measure of surviving firms using business-as-usual technology.A.1.5 Technology UpgradingIn addition to the results shown in Proposition 4, regulation also affects the adoption ofthe state-of-the-art technology, however, its effects are ambiguous. To see this, note that theratio of s technology adoption cut-offs under the regulation and no regulation regimes can bewritten asϕ tarhϕnh=1+[1+ τκ]k[∆kσ−12[ffr] k−σ+1σ−1+∆kσ−11 (ffs− fr )k−σ+1σ−1][1+ τκ]1k[fs− frfs] 1k[σ−1][1+∆kσ−11[ffs] k−σ+1σ−1] . (A.13)121It can be shown that ϕtarhϕnh> 1 if the fixed cost of production, f , is large enough to satisfyfk−σ+1σ−1 >[1− 1[1+ τκ]1k[fsfs− fr] 1k[σ−1]][[[1+ τκ]σ−1−1] kσ−1[1+ τκ]1k1fk−σ+1σ−1r[fsfs− fr] 1k[σ−1]+[1+ τκ] k2−1k[σ−1] [ fsfs− fr] k[k−σ+1]k[σ−1] −1 ∆ kσ−11fk−σ−1σ−1s].Hence, the effects of regulation on state-of-the-art technology adoption depend on f ; if f isrelatively small, then regulation increases the number of firms using the s technology, but f islarge enough, then regulation reduces the number of firms using the s technology.A.2 Chapter 2 AppendixA.2.1 Data AppendixMicro DataOur micro-data was created by merging two existing datasets: the National Pollutant Re-lease Inventory (NPRI) and the Annual Survey of Manufactures (ASM). We describe eachhere, and provide details on how these two sources were matched.The NPRI is Canada’s main source for pollution information, and the only source of airpollution micro-data in the country. It records plant-level pollution activities for over 300 pol-lutants, including criteria air contaminants, toxins, and heavy metals. All plants in Canada thatemit at least one covered pollutant (above that pollutant’s minimum emissions threshold) andemploy at least 10 individuals are required by law to report to the NPRI (Environment and Cli-mate Change Canada, 2016c). In addition, all plants that use stationary combustion equipmentmust report to the NPRI, regardless of their number of employees. Failure to report, or thesubmission of incorrect data, may result in a penalty of between $25,000 and $12,000,000.1The federal ministry of environment performs inspections to confirm the completeness of sub-mitted data. From 2000 to 2010, there were 2,198 NPRI inspections completed, resulting in1,270 written warnings.2.1For details, see sections 272 and 273 of the Canadian Environmental Protection Act.2These figures are from the authors’ calculations computed using data from the Canadian EnvironmentalProtection Act annual reports. These reports are available here: http://www.ec.gc.ca/lcpe-cepa/default.asp?lang=En&n=477203E8-1122For each pollutant, plants are required to report their releases by medium (to air, water,and land), quantities sent for disposal and recycling, methods used to compute releases, andabatement activities3. Detailed guidelines on how to compute emissions for each pollutant areprovided for each sector and production activity (for a detailed list by sector, see: Environmentand Climate Change Canada (2016a)). Each plant is also required to report a number ofcharacteristics, including plant name, business number, industry, and location.The ASM was used as Statistics Canada’s manufacturing census until 2012, and pro-vides longitudinal information for the majority of manufacturing plants in Canada.4 Before2004, every manufacturing plant in the country was sampled annually. The sampling strat-egy changed in 2004 so that a new random sample of the smallest plants was taken in eachyear, rather than collecting information for every plant annually. All large plants were sam-pled annually. For the plants that weren’t sampled yearly, where possible, administrative taxfiles were used to fill-in missing sales and expenditure data. We restrict our analysis to 2004onwards to avoid any issues with the methodological change.The ASM collects information on sales, production costs (including energy expendituresby fuel type), employment, the distribution of sales by province and country, and plant charac-teristics (including plant name, business number, industry, and location). Sales, value added,and cost variables are expressed in 2007 Canadian dollars using industry price deflators fromStatistic Canada’s Industry Multifactor Productivity Program.To match the two datasets, Statistics Canada developed a cross-walk file between themfollowing a multi-stage linking strategy. The majority of plants were linked using businessnumber, year, and location information. A second round of linking was done using two-variable combinations of the above three variables (business number and location, etc). Afinal round of linking was done using plant names. Approximately 80% of manufacturingplants in the NPRI were successfully linked to the ASM.There are two potential issues that arise from the imperfect link between the NPRI and theASM. The first issue is to do with the representativeness of the matched sample. If the proba-bility of a successful match is non-random, then the matched sample will not be representativeof the universe of polluters. This means descriptive statistics from the matched sample willnot be reflective of polluters in general. Rather they will be informative about the subset ofpolluters that were successfully matched.The second issue is more problematic, as it could lead to biased estimates of the CWS’effects. This issue arises if the match probability is correlated with the CWS’ treatment effect.3Reporting of abatement activities was discontinued in 2010.4The ASM was discontinued in 2012 and was replaced with a repeated cross-section survey.123Note that if the effect of the CWS is homogenous, then the match probability cannot be cor-related with treatment, and the estimated effect of the CWS from the matched data will be anunbiased estimate of the true effect of the CWS. That is, this issue only arises when the effectof treatment varies across plants.In the case of the CWS, there is substantial heterogeneity in the treatment effects. Aswe show in the main body of this chapter, the treatment effects vary by plant productivity.Moreover, plant productivity is correlated with plant size, and the probability of a successfulmatch also appears to be correlated with plant size. As a result, the match probability ispotentially correlated with the treatment effect. This sample bias induced by the imperfectmatch should be addressed so as to obtain unbiased estimates of the CWS’ effects. We correctfor this bias using a simple weighting strategy.To see how weighting corrects for this sample bias, consider the estimation of a treatmenteffect, β , that varies across two groups, g1 and g2. Let the treatment effect in g be given byβ g. The average treatment effect is a weighted average of the two groups’ treatment effectsβ = Pr(g1)β g1 +Pr(g2)β g2, (A.14)where Pr(g) is the probability an observation is in group g.The treatment effect in the matched sample is given byβmatch = Pr(g1|match)β g1 +Pr(g2|match)β g2=Pr(match|g1)Pr(g1)Pr(match)β g1 +Pr(match|g2)Pr(g2)Pr(match)β g2,(A.15)where the second equality follows by Bayes’ theorem, Pr(match) is the probability of asuccessful match, and Pr(match|g) is the probability an observation in group g is successfullymatched.If the probability of a successful match is random, then Pr(match|g1) = Pr(match|g2) =Pr(match), and βmatch = Pr(g1)β g1 +Pr(g2)β g2 = β . That is, there is no bias and the im-perfect match does not matter. If the probability of a successful match is non-random, thenPr(match|g1) 6= Pr(match|g2), and βmatch 6= β .Now, suppose the match probabilities (Pr(match|g)) were known for each group, and wereused to construct weights defined as the inverse of the probability an observation was suc-cessfully matched. In this case, the weight for group g would be ωg = Pr(match)Pr(match|g) . Clearly,124performing a simple weighted regression on the matched data using these weights would pro-duce an unbiased estimate of the true treatment effect. The weighted treatment effect from thematched data would beβmatch,weighted = ωg1Pr(g1|match)β g1 +ωg2Pr(g2|match)β g2= Pr(g1)β g1 +Pr(g2)β g2 ,(A.16)which is the true treatment effect, β .The real issue is that these match probabilities are generally not known. In our case, how-ever, we can recover a reasonable approximation of these probabilities because our concern isthat the match probabilities and treatment effects vary by plant size, and we observe a reason-able measure of size (pollution) for both the universe of polluters and the matched sample.We operationalize this weighting procedure by splitting the distribution of pollution intoten evenly spaced bins in both the full NPRI and the matched NPRI-ASM. We then computethe match probability in each bin as the number of plants in that bin in the matched sampledivided by the total number of plants in that bin in the full NPRI. The weights are taken asthe inverse of this ratio for each bin. We compute these weights for each of the four pollutantsamples.To show the effect of our weighting procedure, Table A.1 compares the average plantemissions of each of the CWS pollutants from the full NPRI, the unweighted matched sample,and the weighted matched sample. The first column shows the mean emissions for the universeof polluters, and the second the percentage differences between the mean emissions in thematched sample using our weighting procedure and the universe of polluters. The third columnshows the percentage differences between the mean emissions in the matched sample withoutweighting and the universe of polluters.The match problem appears most severe for particulate matter emissions, with unweightedaverage emissions approximately 25% higher in the NPRI-ASM matched data than in theuniverse of polluters. Weighting reduces this over-estimate considerable, to 12% for PM2.5.The match problem is relatively small for NOX emissions, and weighting has a relatively smalleffect on the average emissions of these pollutants.A.2.2 RobustnessThis section presents a series of additional robustness exercises described in Section 2.4.3of the main text. We first examine when plants are regulated by the CWS. If the majority of125Table A.1: Mean Emissions in Matched DatasetUniverse ofPollutersMatched SampleWeighted UnweightedPM2.5 Emissions 23.0 +12% +26%NOX Emissions 276.4 -5% +1%Notes: Table reports the mean emissions in tonnes from the universe of polluters inthe NPRI and the matched NPRI-ASM samples. Column 1 shows the mean emis-sions from the full NPRI. Column 2 shows the difference in mean emissions in thematched data with weighting. Column 3 shows the difference in mean emissions inthe matched data without weighting.plants are regulated near the end of the CWS period, then there is a strong possibility thatplants may have been able to respond pre-emptively in anticipation of future regulation.Table A.2 shows the fraction of treated plants that are treated early in the policy. Panel Ashows the plants treated in the first year of the sample, and Panel B shows the plants treatedby the middle of the CWS phase-in. For each standard and pollutant, over half of the treatedplants start the sample treated. That fraction increases to between 80% and 90% by 2005 forall standard-pollutant pairs with the exception of the PM2.5 emitters treated by the O3 standard,for which two-thirds are treated by 2005.Restricting treatment to plants that start the sample treated (dropping all plants treatedlater from the sample) leaves the results qualitatively unchanged, and actually increases themagnitude of the main effects (though not significantly). The results for the average effect ofthe CWS on emissions of each pollutant are shown in Table A.3. For this group, the PM2.5standard reduced emissions of PM2.5 by 17%, and the O3 standard reduced emissions of NOXby 56%.5 The average effect of the CWS on scale, and the effects on emissions and scale byplant productivity levels have the same sign and are similar magnitude to the main results.These results suggest that the baseline estimates presented in our main analysis are notdriven by preemptive changes to avoid regulation. Nevertheless, an identification problemcould still arise if our effects are primarily driven by large emitters for whom changes inemissions directly affect CMA air quality. This could be problematic for two reasons. Firstly,it would mean influential plants could have potentially manipulated the length of time theywere treated, meaning treatment is not exogenous. Secondly, our results could be spurious if5Note that we estimate all robustness checks using the publicly available NPRI data, rather than the matcheddata, so as to reduce the number of estimates requiring vetting by Statistics Canada. As a result, the number ofobservations differ between the robustness checks and the main analysis. The results using the matched sampleare very similar, and can be provided upon request.126Table A.2: Regulation CohortsPanel A: % Reg. in 1st Year Panel B: % Reg. by 2005(1) (2) (4) (5)PM2.5 NOX PM2.5 NOXPM2.5 Standard 50% 52% 84% 80%O3 Standard 56% 68% 63% 87%Notes: Table reports the regulation cohorts for each standard and group of emitters. Panel A shows the percent-age of treated plants treated in the first year of the sample. Panel B shows the percentage of treated plants treatedby 2005. The first column within each panel shows the results for PM2.5 emitting plants, the second columnfor NOX plants. Each cell shows the fraction of plants that are ever regulated by each standard by the year inquestion. The first row reports results for the PM2.5 standard and the second for the O3 standard.Table A.3: CWS Effect on Emissions for Initial Treatment Co-hort(1) (2)PM2.5 NOXPM2.5 Standard -0.169∗ 0.0132(0.087) (0.072)O3 Standard -0.059 -0.560∗(0.082) (0.330)R2 0.268 0.336N 6538 2881Notes: Table reports estimates of the effects of the CWS on plant pollutionemissions for the cohort of plants treated at the beginning of the sample. Allplants treated after the beginning of the sample are dropped. Each panel re-ports results for a different sample of emitters. In each regression, the depen-dent variable is the natural log of pollution emissions. The first row reportsthe effects of the PM2.5 standard, and the second row reports the effects ofthe O3 standard. All regressions include plant, industry-year and CMA-yearfixed effects. Standard errors are clustered by CMA-industry. Asterisks denotesignificance at the 1% (***), 5% (**), and 10% (*) levels, respectively.large emitters are on a different trend relative to small emitters owing to some other factorsbeyond regulation, and treatment is positively correlated with large emitter status.Fortunately, we can test for both of the above concerns. To address the first, we drop plantsthat emit a large fraction of their CMA’s emissions. Dropping large plants lowers the potentialfor bias by removing plants who are potential drivers of their city’s air quality problem. Asthere is no obvious size cut-off above which a plant becomes “influential”, we start by drop-ping plants that account for more than 20% of their CMA’s emissions and continue tightening127Table A.4: CWS Effect on Emissions Dropping Large EmittersDrop 20% Drop 10% Drop 5% Drop 1%Panel A: PM2.5(1) (2) (3) (4)PM2.5 Standard -0.164∗∗ -0.205∗∗∗ -0.203∗∗∗ -0.133∗(0.0651) (0.0698) (0.0750) (0.0749)R2 0.220 0.217 0.215 0.246N 6342 5905 5399 4052Panel B: NOX(1) (2) (3) (4)O3 Standard -0.273∗∗ -0.205 -0.219 -0.0696(0.115) (0.129) (0.134) (0.133)R2 0.334 0.345 0.357 0.468N 2433 2192 1978 1341Notes: Table reports estimates of the effects of the CWS on plant pollution emissions dropping largeemitters. Each panel reports results for a different sample of emitters. In each regression, the dependentvariable is the natural log of pollution emissions. Column one drops all plant-years that account for morethan 20% of their CMA’s emissions. Column two drops all plant-years that account for more than 10% oftheir CMA’s emissions. Column three drops all plant-years that account for more than 5% of their CMA’semissions. Column four drops all plant-years that account for more than 1% of their CMA’s emissions.The first row reports the effects of the PM2.5 standard, and the second row reports the effects of the O3standard. The effect of the PM2.5 standard is shown for PM emitters, and the O3 standard is shown for O3NOX emitters. All regressions include plant, industry-year and CMA-year fixed effects. Standard errors areclustered by CMA-industry. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels,respectively.until we reach a 1% threshold.6 We report the results for emissions in Table A.4. The effectof the PM2.5 standard is remarkably robust. For PM2.5 emitters, the effect is negative and sta-tistically significant in each specification, and there is no significant difference between eachof the results in Table A.4 and the effect in the full sample. The effects of the O3 standard arealso consistent with the main results in this chapter, although they are less robust than the PMstandard. The O3 standard is only significant in the first specification for the NOX emitters;however, the results are qualitatively unchanged and there is no significant difference betweenthe first three specifications and the effects in the full sample. The O3 regulation’s effect onNOX emissions, however, disappears if we drop plants that emit more than 1% of their CMA’semissions.Our estimates of the effects of the CWS by plant productivity level are also robust todropping large emitters. For PM2.5 emitters, the average effects on output and by plant6For reference, the average plant fraction of city emissions is: 7% for PM2.5 and 10% for NOX.128productivity-levels for emissions and output are qualitatively unchanged in each of the sizethresholds employed in Table A.4. The same is true of the O3 standard’s effects for NOXemitters, with the exception of the most stringent size threshold. As in Table A.4, droppingNOX emitters that account for more than 1% of their city’s emissions causes the effect of theO3 standard to disappear. The O3 standard’s effects appear to be largely driven by plants thatemit between 1% and 5% of their city’s emissions.To address the possibility of differential trends across large and small emitters, we estimatea version of our main specification that allows for separate CMA-year fixed effects for rela-tively large and relatively small emitters. We accomplish this by determining the fraction oftheir CMA’s annual emissions each plant accounts for, then placing each plant into one of threebins reflecting small, medium, and large emitters. Small emitters produce less than 1% of theirCMA’s emissions (for the respective pollutant). Medium emitters produce between 1-20% oftheir CMA’s emissions. Large emitters produce more than 20% of their CMA’s emissions. Wethen include a full set of emitter size-by-CMA-by-year fixed effects in our regressions. Weare able to do this because, while targeted industries are those that are relatively dirty, howdirty they are relative to other industries varies across the country. In some regions, plants innon-targeted industries are larger emitters than plants in targeted industries, which gives usvariation in treatment that is not perfectly correlated with how dirty a plant is relative to otherplants in their region.The results are presented in Table A.5. Flexibly controlling for emitter size-by-CMA fixedeffects produces similar results to our baseline specification, albeit with a minor attenuation inour estimates of the effects of the CWS. PM2.5 regulation significantly reduced PM emissionsfrom affected plants, and O3 regulation significantly reduced NOX emissions from affectedplants. Consequently, we conclude our results are unlikely to be reflective of differentialtrends across large and small emitters.Finally, we turn to address the possibility that our results are capturing the effects of firmownership. While we treat each plant in our analysis as an independent agent, approximately50% of the plants in our sample are directly owned by a firm that owns at least two plantsin the manufacturing sector. These multi-plant firms create a potential identification problembecause the treatment of one plant may alter the potential outcomes of another plant owned bythe same firm, leading to a violation of the Stable Unit Treatment Value Assumption (SUTVA)that is implicit in our analysis. We address this here by identifying the plants owned by thesemulti-plant firms, and then testing whether the treatment effects differ for plants owned by129Table A.5: CWS Effect on Emissions with Large Emitter Trends(1) (2)PM2.5 NOXPM2.5 Standard -0.128∗∗ 0.0573(0.0590) (0.0841)O3 Standard -0.0644 -0.277∗∗(0.0776) (0.134)R2 0.563 0.652N 6296 2243Notes: Table reports estimates of the effects of the CWS on plant pollutionemissions controlling for separate trends within each CMA for small, medium,and large emitters. Small emitters are those that account for less than 1% oftheir CMA’s pollution for a given pollutant. Medium emitters emit between1-20%, and large emitters are those that emit above 20%. Each panel reportsresults for a different sample of emitters. In each regression, the dependentvariable is the natural log of pollution emissions. The first row reports theeffects of the PM2.5 standard, and the second row reports the effects of theO3 standard. All regressions include plant, industry-year and emitter size-by-CMA-by-year fixed effects. Standard errors are clustered by CMA-industry.Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels,respectively.mutli- and single-plant firms.7,8We use the parent company name information reported in the NPRI to identify multi-plantfirms. This information is entered as a text string, which is imprecise. To improve our match-ing, we use a string-similarity algorithm called the Levenshtein Edit Distance. The Leven-shtein Distance measure, in essence, tracks the number of changes required to convert onestring to another. Two strings requiring few changes would have a relatively small distance.9We classify firms in two ways. In our first approach we classify firms as multi-plant if theyown more than one plant that emit the same pollutant (either PM2.5 or NOX). In our secondapproach we classify firms as multi-plant if they own more than one plant in our dataset (thatis, that emit any of 300 pollutants tracked in the NPRI). In both approaches we present resultsusing both coarse matching, which produces more matches but is open to more false positives,and fine matching, which is more conservative but more likely to miss correct matches.7An alternative approach is to simply drop all multi-plant firms. Doing this produces similar results.8Our data only allows us to identify the immediate parent of a plant, rather than the ultimate corporate parent.As such, our definition of a multi-plant firm is a firm that is the immediate parent of more than one plants, ratherthan the parent of another firm that owns another plant.9For details on the Levenshtein Distance measure, see Yujian and Bo (2007).130We estimate a version of our main specification in which we include a time-varying indi-cator that selects all plant-years owned by a multi-plant firm, and an interaction between themulti-plant indicator and our treatment indicators. For PM2.5 emitters we estimate the PM2.5standard’s effect on plant emissions, and for NOX emitters we estimate the O3 standard’s effecton plant emissions. These results are reported in Table A.6. As can be seen from the table, inall specifications there is no significant difference in the estimated effect of the CWS for plantsowned by single-plant firms and those owned by multi-plant firms. As a result, it appears thepotential failure of SUTVA through the common-ownership channel does not appear to be anissue for our analysis.Additional RobustnessRecall that an identification problem exists if there is an unobservable characteristic thatvaries by CMA-industry-year and is correlated with treatment under the CWS. We believethere are two potential identification problems of concern that our robustness checks may nothave addressed. The first is to do with differential trade shocks. If plants in targeted industriesare more likely to export, and the CMAs that eventually exceed the CWS are more connectedwith Canada’s major trading partners (i.e. the US), then exchange rate fluctuations would havea larger effect on the treated plants than the untreated plants. This is a potential issue becauseover the CWS phase-in period the Canadian dollar appreciated significantly with respect to theUS dollar (in 2000, one Canadian dollar was worth 67 cents US, but by 2010 one Canadiandollar was worth 97 cents US). This appreciation in the Canadian dollar made Canadian goodsmore expensive, which could have depressed relatively export-intensive manufacturing plants.Note, however, that this is only an identification problem if the treated plants (those in thetargeted industries in dirty regions) are more trade-exposed than the untreated plants. Aswe have plant-level data on exports to the US, we can test whether this is true. Testing fordifferences in trade exposure between our treated and untreated groups, we find plants that areeventually treated by the CWS are less export-intensive than those that are untreated.10 As aresult, differential trade shocks should be less costly to the treated plants, and would bias ourresults upwards, if at all.The second remaining potential identification problem is to do with local industrial policy.If local authorities enact policy to protect regulated plants, then this will create industry-by-CMA-by-year variation that is correlated with CWS assignment. However, the goal of thesepolicies would presumably be to support regulated plants, thereby biasing our results upward.As this type of local industrial policy would lead to attenuation bias, it is not a major concern.10These results are available on request131A.2.3 Additional ResultsThe CWS’ Effect on Other PollutantsIn this section we present our estimates of the effect of the CWS on plant-level emissions ofpollutants not directly regulated by the CWS. These results are useful for two reasons. Firstly,ambient PM2.5 and O3 pollution may be formed through chemical reactions in the atmospherebetween other pollutants besides PM2.5 and NOX , in particular other criteria air contaminants(CACs). Secondly, this allows us to assess whether there were positive or negative spillovers inresponse to the CWS. A positive spillover would occur if plants substitute toward unregulatedpollutants, whereas a negative spillover would occur if emissions were were correlated acrosspollutants. The former is typically referred to as regulation-induced substitution, and the latteras co-pollutant effects.We consider emissions of other important air pollutants collected in the NPRI, includ-ing other CACs and heavy metals, as well as greenhouse gas (GHG) emissions. For CACs,we consider emissions of large-scale particulate matter (PM10), volatile organic compounds(VOCs), sulphur dioxide (SO2), and carbon monoxide (CO). For heavy metals, we considerlead and zinc.GHG information is not available in the NPRI, however, Environment and Climate ChangeCanada collect GHG emission data for the largest plants in the country, and is publicly avail-able through the Greenhouse Gas Reporting Program (GHGRP).11 The GHGRP reports emis-sions for several GHGs (including carbon dioxide, methane, and nitrogen dioxide), total fa-cility GHGs, and provides a crosswalk file to match plants in the GHGRP with plants in theNPRI. We use this crosswalk file to merge the facility GHG data to the NPRI. Virtually allmanufacturing plants in the GHGRP over our sample were successfully matched to the NPRI.We report the effect of the CWS on total GHG emissions, carbon dioxide emissions, andnitrogen dioxide emissions.The results of these regression are shown in Table A.7. Each column reports our estimatesof the effect of the CWS on a different pollutant. The dependent variable in each of these re-gressions is the natural log of plant pollution emissions for the relevant pollutant, and standarderrors clustered at the CMA-industry level are reported in parentheses.The results in Table A.7 show the PM2.5 standard caused a significant drop in PM10 emis-sions. This, to some extent, is a mechanical result: by definition, reported PM10 emissionsinclude emissions of PM2.5. Nonetheless, these results provide added confidence to our mainresults, and indicate there was no significant substitution from fine to large scale particu-11See Environment and Climate Change Canada (2016b) for data.132late matter emissions in response to the CWS. PM2.5 regulation had only minor effects onemissions of the other air pollutants and greenhouse gases. PM2.5 regulation caused a small(insignificant) increase in SO2, CO, and lead, a small (insignificant) drop in VOCs, and hadvirtually no effect on GHGs. The only pollutant showing a sizeable response to PM regulationis zinc, which fell by 23%, although this is not significant at conventional levels.The O3 standard, however, had no effect on heavy metals or PM10, but caused a largereduction in emissions of other CACs and greenhouse gases. O3 regulation caused a 37%reduction in total GHG emissions, a 21% reduction in VOCs (which is a potential ozoneprecursor), a 44% reduction in CO emissions (which is also a potential ozone precursor), anda 51% reduction in SO2 emissions. The drop in GHGs was driven by reductions in both CO2(32%) and N2O (66%), the latter being both a GHG and an ozone precursor.133Table A.6: CWS Effect on Emissions - Multi-Plant FirmsPanel A: PM2.5Same Pollutant Any PollutantCoarseMatchingFineMatchingCoarseMatchingFineMatching(1) (2) (3) (4)PM2.5 Std. -0.200∗∗ -0.197∗∗ -0.236∗∗ -0.230∗∗(0.0937) (0.0949) (0.0996) (0.0998)PM2.5 Std. x Multi-Plant 0.0645 0.0594 0.116 0.107(0.0997) (0.103) (0.0960) (0.0971)Multi-Plant 0.0690 0.0634 -0.00161 0.00190(0.0767) (0.0767) (0.0825) (0.0813)R2 0.938 0.938 0.938 0.938N 7058 7058 7058 7058Panel B: NOXSame Pollutant Any PollutantCoarseMatchingFineMatchingCoarseMatchingFineMatching(5) (6) (7) (8)O3 Std. -0.434∗∗ -0.421∗∗ -0.386∗∗ -0.376∗∗(0.181) (0.182) (0.187) (0.187)O3 Std. x Multi-Plant 0.132 0.112 0.0206 -0.00447(0.0875) (0.0875) (0.101) (0.0990)Multi-Plant 0.0823 0.0840 0.129∗ 0.135∗(0.0666) (0.0657) (0.0744) (0.0725)R2 0.978 0.978 0.978 0.978N 2779 2779 2779 2779Notes: Table reports estimates of the effects of the CWS on plant pollution emissions allowing treatmentto vary by the number of plants owned by the plant’s parent firm.Each panel reports results for a differentsample of emitters. In each regression, the dependent variable is the natural log of pollution emissions.Column one drops all plant-years that account for more than 20% of their CMA’s emissions. Column twodrops all plant-years that account for more than 10% of their CMA’s emissions. Column three drops allplant-years that account for more than 5% of their CMA’s emissions. Column four drops all plant-years thataccount for more than 1% of their CMA’s emissions. The first row reports the effects of the PM2.5 standard,and the second row reports the effects of the O3 standard. The effect of the PM2.5 standard is shown for PMemitters, and the O3 standard is shown for O3 NOX emitters. All regressions include plant, industry-yearand CMA-year fixed effects. Standard errors are clustered by CMA-industry. Asterisks denote significanceat the 1% (***), 5% (**), and 10% (*) levels, respectively.134Table A.7: The Effects of the CWS on Plant Emissions of Unregulated PollutantsPanel A: CACs Panel B: Metals Panel C: GHGs(1) (2) (3) (4) (5) (6) (7) (8) (9)PM10 VOCs SO2 CO Lead Zinc GHGs CO2 N20PM2.5 Std. -0.17∗∗∗-0.07 0.05 0.07 0.05 -0.23 -0.02 0.02 0.04(0.06) (0.08) (0.21) (0.09) (0.32) (0.18) (0.19) (0.19) (0.15)O3 Std. 0.01 -0.22∗ -0.51∗∗ -0.44∗ -0.09 0.01 -0.37∗∗ -0.32∗∗ -0.66∗∗∗(0.09) (0.12) (0.21) (0.24) (0.56) (0.79) (0.17) (0.15) (0.19)R2 0.22 0.22 0.45 0.38 0.37 0.37 0.61 0.62 0.68N 8,003 7,045 2243 3352 1411 1496 701 701 613Notes: Table reports estimates of the effects of the CWS on plant emissions of pollutants not directly regulated by the CWS. Each column reports estimates from aregression of the CWS regulations on the natural log of the emissions a different pollutant. Panel A shows the effects on other criteria air contaminants (large scaleparticulate matter, volatile organic compounds, sulphur dioxide, and carbon monoxide). Panel B shows the effects on heavy metals (lead and zinc). Panel C showsthe effects on greenhouse gases (total emissions, carbon dioxide, and nitrous oxide). In all cases, the first row reports the effects of PM2.5 regulations, and the secondrow reports the effects of the O3 regulations. All regressions include plant, industry-year, and CMA-year fixed effects. Standard errors clustered by CMA-industry arereported in parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.135There is a clear explanation for the observed effects on these other pollutants: co-pollutanteffects. The correlation between changes in PM2.5 emissions and PM10 emissions is veryhigh, but very low for other pollutants. Whereas the correlation between NOX emissions andemissions of other CACs and GHGs is relatively high. We show this by estimating simpleco-pollutant elasticities between the regulated pollutants (PM2.5 and NOX ) and unregulatedpollutants. Our approach is to estimate the within-plant cross-pollutant elasticity for eachregulated-unregulated pollutant pair, by estimating the following equationln(zu,i,t) = αu,rln(zr,i,t)+λu,i+ εu,i,t ,where u indexes an unregulated pollutant included in Table A.7, r indexes either PM2.5or NOX , λu,i is a pollutant-plant fixed-effect, and αu,r is our estimate of the cross-pollutantelasticity between unregulated pollutant u and regulated pollutant r.We restrict our sample to years before 2006 to try to limit the potential interference ofthe CWS in changing these cross pollutant elasticities (recall, most of the CWS regulationswere implemented between 2005 and 2007). The results from these regressions are shown inTable A.8.The findings show an intuitive result: the cross pollutant elasticity between PM2.5 andPM10 is over 80%, whereas the cross pollutant elasticities between PM2.5 and other pollutantemissions are relatively low (between 0.05 and 0.33). In contrast, the cross-pollutant elastici-ties are much higher for NOX emissions (between 0.33 and 0.77). The one outlier is that wefind a relatively high correlation between PM10 and NOX , despite finding no significant effectof O3 regulation on PM10 emissions.A.2.4 CWS CounterfactualsFirst, we present details on the plant-level decomposition. Recall that the change in anindustry’s pollution intensity is given by∆Eit =∫ nit0eit(n)λit(n)dn−∫ nit0eit−1(n)λit−1(n)dn−∫ nit−1niteit−1(n)λit−1(n)dn.136Table A.8: Cross-Pollutant Elasticities(1) (2) (3) (4) (5) (6)PM10 VOCs SO2 CO Total Metals GHGsPM 2.5 0.821∗∗∗ 0.267∗∗∗ 0.211∗∗∗ 0.0568 0.339∗∗∗ 0.121∗∗∗(0.013) (0.032) (0.0485) (0.0356) (0.0851) (0.0342)R2 0.708 0.086 0.040 0.004 0.043 0.118N 2,613 1,207 739 1109 584 199(1) (2) (3) (4) (5) (6)PM10 VOCs SO2 CO Total Metals GHGsNOx 0.561∗∗∗ 0.533∗∗∗ 0.766∗∗∗ 0.700∗∗∗ 0.403 0.333∗∗∗(0.052) (0.057) (0.0658) (0.0494) (0.248) (0.0591)R2 0.156 0.143 0.234 0.246 0.012 0.245N 1,035 869 737 1008 341 209Notes: Table reports cross-pollutant elasticities between regulated pollutants (either PM2.5 or NOX ) and unregulated pollutants. Theestimates are computed by regressing the natural log of plant emissions for each unregulated pollutant on the natural log of plantemissions for each regulated pollutant, including a plant fixed effect. Only early years are use (before 2006). The top panel shows theelasticities for PM2.5 emissions; the bottom panel the elasticities for NOX emissions.This can be written as∆Eit =∫ nit0(λit(n)−λit−1(n))eit(n)dn−∫ nit0λit(n)eit−1(n)dn+∫ nit0(eit(n)− eit−1(n))λit−1(n)dn+∫ nit0λit(n)eit(n)dn−∫ nit0(eit(n)− eit−1(n))λit−1(n)dn−∫ nit−1niteit−1(n)λit−1(n)dn.With some algebra, this reduces to∆Eit =∫ nit0eit−1(n)∆λit(n)dn+∫ nit0λit−1(n)∆eit(n)dn+∫ nit0∆λit(n)∆eit(n)dn−∫ nit−1niteit−1(n)λit−1(n)dn.Dividing by Eit−1 gives the desired decomposition.To express λˆit(n) as a function of our estimates, note thatλˆit(n) =λ f t(n)λ f t−1(n)−1=x f t(n)x f t−1(n)Xit−1Xit−1.137By assumption, if n is untreated, then x f t(n) = x f t−1(n), and if n is treated, then x f t(n) =(1+βx)x f t−1(n). Plugging this into Xit =∫ nit0 xit(n)dn givesXit = (1+βx)∫treatedxit−1(n)dn+∫untreatedxit−1(n)dn= Xit−1−∫ nit−1nitxit−1(n)dn+βx∫treatedxit−1(n)dn.Rearranging gives XitXit−1 = 1− sExitxt−1+βxsTreatxt−1 . With some algebra it can be shown that λˆit(n)is as in the text.A.2.5 Policy DetailsNova ScotiaIn 2004, Nova Scotia adopted emissions taxes for particulate matter and ozone-precursorpollutants (nitrogen oxides and volatile organic chemicals) (N.S. Reg. 31/2005). The emis-sions taxes were tiered such that small emitters were exempt, mid-size emitters paid a flat fee,and large emitters paid a flat fee plus a tax of $2.70/tonne for emissions above a given thresh-old. In 2005, Nova Scotia also strengthened its Air Quality Regulations, which were firstpassed in 1995 (N.S. Reg. 28/2005). There were three substantive changes; the provincialsulphur dioxide cap was reduced from 189,000 tonnes to 141,750 tonnes, the sulphur dioxideemissions cap for the electricity generation sector was strengthened, and nitrogen oxide andmercury emission caps were added for the electricity generation sector.New BrunswickNew Brunswick amended their provincial air quality regulations (regulation 97-133 underthe New Brunswick Clean Air Act) in 2005 to increase the emissions fees assessed for particu-late matter (and sulphur dioxide) emitters. New Brunswick uses a staggered annual emissionsfee schedule, with the highest annual fees being levied against the largest emitters. The 2005amendment increased these fees by between 30%-900%, depending on the class of emitter.For example, the annual fees for the largest emitters rose from $42,000 to $60,000, for mid-range emitters from $15,000 to $28,00, and for the smallest emitters from zero to $500. Fordetails, see part five of the regulation.OntarioIn 2005, Ontario adopted site-specific air quality standards (regulation O Reg 419/05).These standards targeted many different pollutants, including ozone, ozone pre-cursors (in-138cluding nitrogen oxides and various volatile organic compounds), and particulate matter12.The regulation contained more stringent standards for a number of industries, including sev-eral of the industries targeted by the CWS.13 In addition to more stringent standards, plants inthese industries must submit annual emissions reports to the Ontario environment ministry.In 2006 Ontario introduced a limited NOX and SO2 trading program for the twenty largestemitters in four of the five CWS-targeted industries (regulation O Reg 194/05). While permittrading allowed flexibility in compliance with the policy, permits were allocated based on thepollution intensity of each facility, such that cleaner plants received relatively more permits.QuebecQuebec developed the Clean Air Regulations – which included local air quality standardsand site-specific emissions standards – during the phase-in period (regulation QLR Q-2, r4.1). Air quality standards were developed for a large number of pollutants, including ozoneand particulate matter (the PM2.5 standard was set at the level of the CWS and the ozonestandard was set slightly more stringent than the CWS at 62.5 ppb). Emissions standards weredeveloped for many different industries and industrial processes, including particle emissionsfrom a variety of sources (chapter II), VOCs from a variety of sources (chapter IV), pollutantsfrom combustion plants (chapter VI), and pollutants from incinerators (chapter VII). Althoughthe regulations were first published in 2005, it took six years before they were officially madelaw.Prince Edward IslandPrince Edward Island amended their Air Quality Regulations in 2004 to add particulatematter emissions fees for fuel-burning equipment (for details, see schedule D of http://www.gov.pe.ca/law/regulations/pdf/E&09-02.pdf and http://www.gov.pe.ca/photos/original/leg tableregs.pdf).14NewfoundlandIn 2004, Newfoundland amended its Air Pollution Control Regulations, which had beenin place since 1996 (NLR 39/04). The original regulations contained air quality standards forPM2.5 that were more stringent than the CWS and a one-hour ozone standard. The amend-12A standard was set for total particulate matter, but no standard was set for PM2.5. The Ontario Ministryof Environment’s rationale for omitting a PM2.5 standard was to avoid duplicating the existing CWS (point 8 inhttp://www.airqualityontario.com/downloads/AmbientAirQualityCriteria.pdf).13In particular, pulp and paper, electric power generation, iron and steel manufacturing, and base metal smelt-ing.14Two amendments were made: EC161/04 and EC423/04.139ments left the PM2.5 standard unchanged and added an eight-hour ozone standard of 43.5 ppb(more stringent than the CWS). The Newfoundland standards allow the province’s minister ofthe environment to regulate individual facilities should regional air quality exceed one of thestandards (see paragraph 3.(3) of the regulations). The amendments also added NOx emissionintensity standards for all new or modified fossil fuel fired boilers and heaters (paragraph 19).In 2014, the regulations were amended further to add an annual PM2.5 standard equal to thatunder the Canadian Ambient Air Quality Standards, which replaced the CWS (for details, see:http://www.assembly.nl.ca/legislation/sr/regulations/rc040039.htm#3 ).ManitobaManitoba uses objectives and guidelines to manage air quality, rather than provincial reg-ulations. In 2005, the province added the CWS’ ozone and PM2.5 standards to this list ofobjectives (for details, see: https://www.gov.mb.ca/conservation/envprograms/airquality/pdf/criteria table update july 2005.pdf).SaskatchewanSaskatchewan’s Clean Air Regulations have imposed ambient air quality standards in theprovince since 1989 (see: http://www.qp.gov.sk.ca/documents/English/Regulations/Repealed/C12-1R1.pdf). These standards remained in place until they were repealed in 2015 by theEnvironmental Management and Protection Regulations (regulation E-10.22 REG 2) (source:http://www.qp.gov.sk.ca/documents/English/Regulations/Regulations/E10-22R2.pdf). The newregulations imposed more stringent air quality standards, including adopting the CanadianAmbient Air Quality Standards for PM2.5 and ozone (see Table 20 of the Saskatchewan Envi-ronmental Quality Standard, https://envonline.gov.sk.ca/Pages/SEQS/Table20-SEQS-SAAQS.pdf).AlbertaAlberta primarily manages air quality using ambient air quality objectives and guidelines,that are enforced through the provincial permitting and licensing process. Industrial facilitiesmust be designed and operate so as to ensure the provinces ambient air quality objectives aremet; however, they are given relative freedom in deciding how to manage their pollution. Morestringent permitting regulations were passed in 2003 (Alberta Regulation 276/2003), underthe Environmental Protection and Enhancement Act. In 2007, the CWS’ PM2.5 standard wasadopted as an objective (for details, see: http://environment.gov.ab.ca/info/library/5726.pdf).An ozone objective has been in place since 1975, and was reviewed in 2007 but left unchanged.Firms can be fined for violating the conditions of an operation permit, such as failure to comply140with air pollutant-related constraints. For example, in 2012 a refinery was fined for failing toinstall proper air pollution control equipment (for details, see https://www.alberta.ca/release.cfm?xID=32232CC295887-C17E-3ABE-EC7823B5948337D0).British ColumbiaBritish Columbia manages air quality using a combination of air quality objectives, localairshed management plans, and industrial codes of practice. Air quality objectives are non-binding standards that set the air quality levels to which regulators should aim. Over the phase-in period, the province adopted the PM2.5 and O3 Canada-Wide Standards as air quality ob-jectives (for details, see: http://www.bcairquality.ca/regulatory/air-objectives-standards.html).Towards the end of the phase-in, the province adopted additional, more stringent, PM2.5 ob-jectives (see: http://www.bcairquality.ca/regulatory/pm25-objective.html). Provincial regula-tors achieve these objectives using mandatory codes of practice or other regulations (for de-tails, see paragraph 4.3.2 of http://www.bcairquality.ca/reports/pdfs/pm25-implement-guide.pdf). These provincial regulations can target specific industries, regions, or facilities15. In ad-dition, local regulators develop local airshed management plans to meet the air quality objec-tives. Over the CWS phase-in period, airshed management plans were developed for thirteenregions in the province (see http://www.bcairquality.ca/airsheds/bc-airsheds.html).SubsidiesOver the CWS period, some provinces provided subsidies to encourage plants to adoptthe cleaner production techniques suggested in the industry MERS. These subsidies were rel-atively small, and were intended to offset the costs of developing an abatement plan, butnot cover the capital and operating costs involved in abatement. Examples included the En-viroclubs initiative in Quebec (see (Lanoie and Rochon-Fabien, 2012) for details), and theBusiness Air Quality Program Pilot in Ontario (Environment Canada, 2005).15Regulations and codes of practice exist for the pulp and paper, wood product manufacturing, asphalt, andagricultural sectors. For details, see http://www.env.gov.bc.ca/epd/codes/index.htm141

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0371161/manifest

Comment

Related Items