PAPERS DISCUSSION IN ECONOMICS
Working Paper No. 08-07
Are Exporters Mother Nature's Best Friends?
Scott Holladay
University of Colorado
October 2008
Center for Economic Analysis Department of Economics
University of Colorado at Boulder Boulder, Colorado 80309
October 2008 Scott Holladay
Are Exporters Mother Natures Best Friends?
Scott Holladay
This draft: August 2008
Abstract This paper empirically analyzes the relationship between international trade and plant level pollution emissions for U.S. manufacturers. I develop a theoretical framework to study emissions in a heterogenous rm international trade model. The results suggest that exporters should pollute less per unit of output than non-exporters in the same industry. Industries that face import competition should have fewer plants that generate high levels of emissions per unit of output. Their average emissions per output level should also be lower than other industries that are sheltered from international competition. These implications are tested against a unique dataset built by combining plant level emissions data from the EPA and plant characteristic data from the National Establishment Time Series. The data set consists of 15,000 plants observed over 12 years and includes 8-digit SIC industry denitions. The empirical results conrm that exporters pollute around 8% less per unit of output than non-exporters. These results are consistent with a nearest-neighbor matching procedure used to address potential unobserved variable bias. I use this framework to examine the channels through which productivity impacts emissions. Industries that face import competition pollute 0.75% less than more sheltered industries. This dierence is due to the exit of small rms with high levels of emissions per output. I nd no evidence that this eect is related to polluting plants relocating in search of lower levels of environmental regulation.
Keywords: Trade and environment, Firm heterogeneity, Plant-level emissions JEL Codes: F1, Q5
Department of Economics, [e-mail].
256
UCB
Boulder
CO
80309-0256,
E-mail:
1
Introduction
Deepening cross border links have brought increased attention to the impacts of international trade on the environment. There are signicant economic literatures analyzing the eect of trade liberalization on pollution, the pollution haven hypothesis, the Environmental Kuznets Curve and the impact of environmental regulation on trade. Despite all this attention, surprisingly little is known about how international trade aects individual polluting plants. This work focuses on the impact of international trade on plant polluting behavior and the impact of this behavior on aggregate pollution emissions at the industry level. The theoretical literature has produced conicting results for the inuence of international trade on pollution levels. For example, Copeland and Taylor (1995) nd that trade liberalization can lead to an increase or decrease in pollution depending on how incomes dier across countries that liberalize. Cole and Elliott (2003) suggest that models which use dierences in environmental policy to generate trade between countries nd an increase in emissions after liberalization. Models that use dierences in endowments to generate trade typically nd a decrease in emissions post-liberalization. These conicting results suggest the need for empirical studies of the impact of trade on pollution emissions. Much of the empirical work analyzing the impact of globalization on the environment relies on cross-country variation in pollution levels and trade
1
behavior. Antweiler et al (2001) compare levels of openness to pollution concentrations and nd that greater openness is associated with small, but signicant decreases in pollution. Frankel and Rose (2005) employ instruments to control for possible endogeneity in trade policy, environmental policy and income levels. They also nd openness associated with decreases in pollution levels, though the results are not statistically signicant for some pollutants. This literature separates the impact of trade liberalization on the environment into three parts: the impact generated by increased economic activity (the scale eect), the changing industry mix (the composition eect) and the impact of increased income on environmental regulation (the technique eect). Unfortunately these eects do not explain how polluting establishments or industries respond to changes in trade levels. There is an extensive international trade literature examining rm level heterogeneitys impact on international trade behavior. This research has found that rms that serve foreign markets through exports tend to dier substantially from rms that only enter domestic markets. Exporters tend to be larger, more productive and pay workers more than their competitors1 . Melitz (2003) introduce heterogenous rms to an international trade framework. Potential entrepreneurs draw a productivity at random then decide whether to set up business and enter foreign markets. Fixed costs to enter the market and additional xed costs to export ensure that only the most productive rms export. In addition to its role in determining international trade outcomes, productivity also plays an important role in determining
See Bernard and Jensen (2004) and Bernard, Eaton, Jensen and Kortum (2003) among many others.
1
2
an establishments pollution prole. Cole et al (2005) examines the impact of rm level characteristics and environmental regulation on industry level emissions for manufacturing plants in the UK. They nd emissions to be positively related to capital intensity and negatively related to rm size and productivity. Earnhart (2006) nds that better managed rms (measured by return on sales) have higher levels of environmental management in the US chemical manufacturing industry. This study takes advantage of the relationship between international trade, environmental performance and productivity to analyze the impact of trade on individual polluting plants through productivity dierences. The next section introduces heterogenous polluting rms to a trade model. The model generates several testable implications about the relationship between trade status and pollution levels. The third section describes the unique dataset that has been collected to test these implications and works through the empirical analysis of the relationship. The nal section draws conclusions and suggests avenues for future research.
Theoretical Framework
This section develops a simple framework that adds pollution emissions as a by-product of production to a trade model developed in Melitz (2003). This framework is used to explore the behavior of polluting plants in conjunction with international trade status and policy. Plants dier in productivity, which is exogenously determined. Pollution emissions are a function
3
of output and productivity. There are xed costs to enter the market and additional xed costs to serve foreign markets. These costs ensure that the entrepreneurs with the lowest productivity exit and only the most productive plants are able to serve foreign markets. The relationship between exporting behavior and productivity, coupled with the relationship between productivity and emissions generates a channel through which international trade can impact pollution level. In the Melitz model there is only one factor (labor) that is used by a continuum of establishments to produce a unique variety. Potential entrants pay a xed cost (fe ) and then draw a productivity level () at random. Labor is a linear function of output q: l=fe +
q .
Higher productivity is modeled
as a reduction in marginal cost. Entrants who receive a low productivity draw expect to earn negative prots and will choose to exit without producing. The remaining establishments compete in the domestic market and can choose to pay an additional xed cost (fx ) to serve foreign markets by exporting. Because each plant receives the same market price, only those with the lowest marginal costs can aord to endure the additional xed costs required to enter export markets. This framework produces two cut-o productivity levels: , the cuto productivity for entry and , the cuto x productivity for exporting. Because preferences are C.E.S., prices are a constant markup above marginal cost and the ratio of establishments sales and revenue simplify to the ratio of those plants productivities. Taken together this shows that exporters enjoy higher productivity, sales and revenue than establishments that do not serve foreign markets. Pollution emissions are modeled as a by-product of production. Emis4
sions are a function of productivity, output (which itself is a function of productivity) and an industry specic emissions intensity (zj ): Ei j=f(q(i ), i , zj ), where j is an industry subscript. Emissions increase with output, but it is not clear what impact productivity has on emissions holding output constant. Investment in recycling and other waste treatment programs is nonproductive, and may make it appear that lower productivity is associated with decreased emissions. On the other hand, more productive plants are able to produce more output from the same quantity of input. That may mean they can generate their output with fewer toxic inputs that must later be emitted. This issue has been addressed empirically by a number of studies of the determinants of rm level emissions. These studies consistently nd that high levels of productivity are associated with lower levels of emissions after controlling for output level2 . For this reason, emissions are modeled as a decreasing function of productivity. Possible explanations for this relationship are explored in the empirical section below. This study seeks to explain the impact of international trade on both polluting establishments and industries. The most straightforward implication of this analysis is that exporters pollute less than non-exporters after controlling for output3 and industry dierences. Exporters are more productive than non-exporters, but they also have greater output so the relationship between export status and total emissions depends on the relative strength
These results tend to be a by-product of the literature that analyzes the productivity impacts of environmental regulation. See Gray and Shadbegian (2003), Shadbegian and Gray (2005), Earnhart (2006) and Cole et al (2005) for examples. 3 In this framework output is completely determined by productivity so it is impossible to compare two rms (in the same industry) with the same output and dierent productivity levels
2
5
of these eects. If the productivity eect outweighs the output eect then exporters should pollute strictly less than non-exporters despite dierences in output. On the other hand, if the output eect is strong, exporters will pollute more than non-exporters. Perhaps more interesting is the impact of increased imports on emissions. While most countries strongly encourage exports, imports tend to be less popular politically. The impact of import competition on job growth and wages has been studied extensively, but there has been comparably little focus on the impact of imports on individual polluting establishments. Melitz (2003) has shown that trade liberalization leads to a smaller number of rms and increase in the average productivity compared to autarky. The least productive rms exit and those rms that export are able to take advantage of the resources freed up by this exit to increase total sales. The change in total output is not clear. Less productive rms that do not exit see a decrease in output while the more productive exporting rms experience an increase in output. This makes predicting the total change in emissions after a trade liberalization dicult. There are several straightforward implications for the distribution of establishment level emissions within an industry. Import competition should force the least productive rms to exit the market. Those plants will tend to pollute more per unit of output than their more productive competitors. Industries facing import competition should have the left side of their emissions per unit of output distribution truncated by the exit of less productive rms. The variance of emissions per unit of output within industries that face import competition should be smaller than industries that do not face international competi6
tion. The relationship between productivity and sales suggest that similar distributional results should hold for sales as well. This framework can also be used to analyze concerns about the pollution haven hypothesis, which argues that countries will attempt to attract polluting rms and industries by lowering environmental regulation. This appears to be at odds with the prediction of the model described above. If polluting rms were leaving countries with strict environmental regulation for more enticing locales, the rms that pollute the most (without regard for output) would be the ones that would be most likely to move. Those establishments would have the most to gain from reduced environmental compliance cost. If environmental regulation were driving location decisions for polluting plants the distribution of plant emissions would likely be truncated on the right as large polluters moved abroad and served the local market through exports.
Empirical Analysis
The model described in the previous section generated several straightforward implications for the relationship between international trade and pollution emissions. The rst is that exporters within an industry should pollute less than non-exporters after controlling for dierences in output. The second set of implications involve the impact of import competition on polluting rms and industries. Industries facing foreign competition should produce less emissions after controlling for output. These industries should also have a smaller variance in emissions than non-import competing industries. The
7
smaller variance should come from a truncated right side of the emissions per unit of output distribution as a result of the exit of the most polluting rms. This section will seek to test these implications against the data, which requires measures of pollution emissions, total output, industry and exporting status.
Data
This paper relies on a unique dataset to test implications of the model outlined in the previous section. The data are constructed by merging the National Establishment Time Series (NETS) with the EPAs Risk-Screening Environmental Indicators (RSEI). The NETS is complied from Dunn and Bradstreet data on creditworthiness by Walls and Associates. Dunn and Bradstreet collects establishment level information that is used to generate credit scores. These scores are required to receive government contracts and are used to make decisions about accepting payment, leasing equipment or oce space and setting nancing terms. The data is collected by surveying establishments, tracking payment histories with other establishments and through research in trade publications and news archives. Neumark et al (2004) analyzed the NETS data and compared it to data collected by the Current Population Survey and the Current Employment Statistics Payroll Survey. They nd that the NETS data on employment is comparable to that reported in the CPS and CES. They also use a media search to nd stories about plant relocation. The NETS reected around three-quarters of the moves that crossed a county or city line. That rate is similar to the rates 8
found in Lexis-Nexis and Hoovers.com company location datasets. The data is annual from 1988-2006 with observations on the number of employees, value of sales, exporter status, information on corporate parents, children, siblings, SIC codes at the 8-digit level and credit rating among many other variables. The NETS contains no information on capital making estimating productivity using a production function approach impossible. The data set acquired for this study contains about 35,000 manufacturing establishments that have been listed in the RSEI at one time. The RSEI is an establishment level record of toxic pollution emission collected by the EPA. Manufacturing establishments that release more than a thresh-hold level of toxic chemicals must report how those chemicals are disposed of to the EPA. That information is used to build an annual report on emissions called the Toxic Release Inventory. This data is cross referenced with measures of the toxicity of each pollutant to build a measure of the hazard from pollution generated by each polluting establishment. That data is then combined with information on population density and age structure to create a measure of the risk of emissions to the nearby population. These measures, along with the total quantity of emissions, are reported annually for each establishment from 1988-2002. The data also contains a DUNS number eld, which is the identier used by Dunn and Bradstreet to index establishments, along with a variety of location information. This makes it possible to match NETS data to the emissions data in RSEI. Due to incomplete data on location (and DUNS numbers) in the RSEI dataset, matching every polluting establishment is impossible. 74.7% of the establishments identied by the EPA match with observations in the 9
NETS each year. The RSEI observations that were not matched produce more pounds of emissions and have a higher hazard score, but there is no signicant dierences in the risk generated by those emissions. While there are dierences in the level of emissions between the two groups, there are no dierences in the ratios of any measure of emissions4 . The merged dataset is an unbalanced panel of between 14,000 and 16,000 annual establishment level observations between 1990-2002. The matched variables are summarized in Table 1. To control for the price ination the values of sales was divided by the manufacturing PPI deator provided by the BLS5 .
Exporters Environmental Performance
The model described in the previous section predicts that rms that draw productivity levels above the export cuto should pollute less per unit of output than those who do not export. The merged NETS and TRI data will be used to compare the emissions of exporting and non-exporting rms after controlling for output and industry. Table 2 summarizes the dierences between exporters and non-exporters across the observable variables. Exporters are larger as measured by both sales and employees, and the dierences are signicant at the 5% and 1% levels respectively. The exporters average nearly 21,000 fewer pounds of toxic emissions than their non-exporting competitors, but they do not fare as well in the broader meaThe dierence between the matched and unmatched groups in pounds and hazard are signicant at the 1% level. The dierences between the groups risk scores and the ratio of pounds to hazard, pounds to risk and risk to hazard are not signicant at the 10% level 5 See Levinson 2007 for a description of the trade o between using industry specic and economy wide deators.
4
10
sures of the damages from emissions. Exporters have signicantly higher hazard and risk scores. The hazard score suggests that while exporters generate a smaller quantity of emissions those pollutants are more toxic. The nal line compares the emissions per unit of output for each industry. Exporters generate far less pollution per dollar of sales than non-exporters on average, but the huge variance of both groups makes this dierence insignificant at conventional levels. These results are generally consistent with the theoretical framework described above, but they do not control for dierence in industry emissions intensity. If the United States has a comparative advantage in the production of less polluting goods, or the pollution haven eect has driven polluters to foreign countries the same pattern would emerge. The estimation equation is:
Eijt = + W Wijt + Xijt + j + t +
ijt ,
where i references a plant, j indicates an industry and t indexes years. The outcome variable, Eijt is a plant-level measure of pollution from the RSEI such as pounds of emissions, hazard score or risk level. j is a set of industry xed eects that control for the diering levels of emissions intensity of production across industries and t are year xed eects.
ijt
is the
stochastic error term. W is a vector of plant-level characteristics such as sales, employees and credit ratings. X is an indicator variable that equals 1 if the plant exported any amount of its production and is the parameter of interest. It measures the dierence in plant level emissions between exporters and non-exporters conditional on all the plant-level characteristics 11
and indicator variables. The combination of xed-eects guarantee that is identied from variation between exporters and non-exporters in the same industry during the same year. The model does not include establishment xed eects because there is such a high degree of persistence in exporting. Fewer than 1% of establishments in the sample switch their export status during the sample period. This is consistent with Bernard and Jensens (2004)ndings on export behavior over time. The regression results are described in Table 3. They examine the relationship between exporting status and pollution emissions after controlling for industry type. Pollution emissions are measured in pounds of emissions as reported by the RSEI. Industry classications are at the 6-digit SIC level as reported in the NETS and conrmed (at the four-digit level) in the RSEI. In regression 1 the impact of exporter status is measured without controlling for sales. This tests the relative size of the productivity eect on emissions (negative) and the output eect (positive). The regression includes industry xed eects at the SIC 6-digit level to control for the industry specic emissions intensity (zi ) which is not observed. The results indicate that an exporter pollutes around 7% more than a non-exporter in the same industry. This implies that output eect outweighs the productivity eect. It also implies that the United States has a comparative advantage in polluting goods. The stronger implication of the theoretical framework was that an exporter should pollute less than a non-exporter in the same industry after controlling for output. This is accomplished by including establishment sales as reported in the NETS in regression 2. The results suggest that 12
exporters pollute around 8% less than non-exporters after controlling for output. This regression also makes it possible to estimate the magnitude of the output eect. A 1% increase in sales is associated with a 0.6% increase in emissions. Environmental regulations have been strengthening with time so if newer plants are more likely than average to be exporters, this may bias the exporter coecient downward. To address this problem year xed effects are included. A similar issue arises with plant location. Certain states have stricter environmental regulations. If establishments in those states are more likely to export, the export coecient may be biased downward. To address this concern state xed eects have been added to the regression6 . The results of the regression including these xed eects have been included in regression 3. In this specication is identied from variation between exporters and non-exporters in the same industry, during the same year that are located in the same state. The additional controls cause a small change in both the sales and export coecients and the signicance of the export coecient drops from the 5% to 10% level. The additional controls do not change the conclusion that exporters pollute less than non-exporters after controlling for output. The nal regression includes additional controls that may be related to export status and emissions. The number of employees can be thought of as proxying for the capital intensity of production, which is closely related to pollution emissions. The more employees it takes to produce a given amount of output the more capital the plant likely possesses. By including
These regressions have also been calculated using county xed eects and two-digit SIC xed eects and the results were similar. Due to computational restrictions six-digit SIC xed eects and county xed eects cannot be run at the same time
6
13
the number of employees and controlling for output, the issue of capital intensity is partially addressed. Establishments that relocate often may be moving to take advantage of changes in environmental regulation and/or exporting infrastructure. To control for that possibility the number of times a rm has changed location during the time period is included as a control in this regression. The additional controls reduce the output eect of sales on emissions slightly, but it has no impact on the export indicator variable. Regressions 1-4 conrm that, after controlling for output, exporters generated fewer emissions than non-exporters. While, to my knowledge, this fact has never been documented, it is related to a debate in the environmental economics literature about the determinants of rm level emissions. There are several hypothesized channels through which productivity may aect emissions. Larger rms tend to be the most productive and have a higher public prole and therefore seek to limit pollution. It is also possible that productive rms are better able to control the long-term liability of emitting pollution. Less productive rms may be more worried about the companys survival than minimizing a potential liability which may not appear for many years. Some authors have argued that the most productive rms locate in the regions with the strictest environmental regulations and are therefore compelled to pollute less. A nal hypothesis suggests that the same management skills that generate frequent innovation and high productivity can be applied to preventing pollution emissions. While there has been research indicating that highly productive rms pollute less, there is no consensus on why this may be the case. Konar and Cohen (1997) argue that more productive establishments may 14
pollute less because they are more concerned with the long term liability that toxic emissions may generate. More productive rms have a larger incentive to reduce their long term liability, since they are more likely to survive to see claims made against them. Less productive rms are concerned with the day-to-day struggle of staying in business and do not worry about the long term liability that toxic emissions will bring. If this were, the case we would expect the most productive rms to reduce their liability by reducing the level of emissions and the toxicity of their emissions. This suggests exporters should have hazard and risk scores substantially lower than non-exporters. Regressions 5-7 show that this is not the case. In each regression the dependent variable is a dierent measure of plant level emissions. Regression 5 is similar to regression 2 above, except the dependent variable has not been logged to make it comparable to the other regressions in this table7 . Exporters produce around 62,300 pounds fewer emissions than a non-exporter in the same industry after controlling for sales. That point estimate is 21% of the average plants emissions. Recall that hazard is a measure of the toxicity of an establishments emissions. Exporters have higher hazard scores than non-exporters despite their productivity advantage, however the coecient is insignicant. Risk measures the toxicity of emissions weighted by location. Again exporters have insignicantly higher scores than non-exporters. Liability is a function of the toxicity and the location of emissions more than the quantity. Hazard and risk scores are a better proxy for liability than pounds. Taken together they suggest that exporters actually emit pollutants
Hazard and risk scores may be zero for small quantities of relatively non-toxic pollutants.
7
15
that are more toxic than their competitors in the same industry. They do produce fewer of those emissions. Arora and Cason (1996) argue that large rms have higher public proles and therefore must pollute less than their smaller competitors. Larger rms may receive more attention from regulators, watchdog groups and environmentally conscious consumers. In this framework exporters are larger than their competitors due to their productivity advantages. To test the impact of plant size on emissions the sample was stratied into 5 quintiles based on establishment sales8 and regression 2 was run on each quintile. If rm size is the primary channel through which productivity impacts emissions, then export status (and the increased productivity it signals) should not have a negative impact on emissions among the smallest rms. Table 5 lists the coecients and t-statistics for the log sales and the export indicator variables for each of the 5 regressions. In four of the ve regressions the exporter coecient negative, is in three of the regressions its is negative and signicant. In no case was the exporter coecient positive and signicant. Among the smallest plants exporters may pollute more than non-exporters, but there is not the monotonic relationship predicted by the proponents of the highprole polluter explanation. Figure 1 illustrates the same point. The left vertical axis shows the fraction of rms that are exporters (represented by the bars) and the right vertical axis represents the exporter coecients. The bars are display the 95% condence intervals and the squares represent the point estimates. The horizontal line is at 0 on the exporter coecient axis,
The results are robust to dividing the sample into quintiles by employees and using 10 and 20 groups.
8
16
so for the percentile groups whose coecient condence interval entirely below the line exporters pollute signicantly less than non-exporters after controlling for output. Among the 20-40 and 40-60 percentile groupings exporters pollute signicantly less than non-exporters. This eect would not be predicted if rm size was the main driver of pollution behavior. The elasticity of sales on pollution emissions is between 0.6 and 1.0 for every group, with the exception of the smallest (for which the point estimate is -0.07). Firm size does not appear to have much impact on polluting behavior within dierent percentiles. These results are robust to dening size by the number of plant employees. It is clear that exporters pollute less than non-exporters in the same industry after controlling for output. This is consistent with the model described in the previous section. It appears that this phenomena is a function of plant productivity and not rm size or liability concerns as proposed by previous authors. These results seem to support the view that more productive rms generate fewer emissions because they are better able to manage their production process to minimize waste of all sorts. This management expertise explanation was advanced by Earnhart (2007). The exact channel through which productivity impacts emissions is a subject for future research.
Matching Estimators The previous section attempted to control for dierences between exporters and non-exporters using available data. There are still some variables which 17
may impact both emissions level and export status that remain unobserved. In an eort to control for these unobserved variables this paper implements a nearest neighbor matching estimator to examine the impact of exporting on measures of emissions after controlling for dierences between the two types of establishments. Matching estimators are used to examine the impact of treatment on an outcome variable of interest. They work by matching two observations that are similar across the observable variables, but dier in the exposure to treatment. The dierent values of the outcome variable for those observations are used to identify the impact of treatment9 . The export indicator is the treatment variable and the pounds of emissions is the outcome variable of interest. The other explanatory variables in the NETS serve as controls for exporting status. The matching estimator is consistent if two conditions are met. First the level of pollution must be independent of export status after controlling for dierences in observable establishment characteristics. Secondly there must be some overlap in observable variables between those that export and those that do not. If the observable variables do not share similar values, then it will be impossible to nd similar establishments to compare for estimates of treatment eect. Nearest neighbor matching is extremely computationally intensive. The distance from each observation to every other observation must be calculated based on the matching variables. The procedure nds two observations that are nearest neighbors to each observation in the data set. The nearest neighbors are required to match exactly on industry (at the 6 digit level) and state. Groups with dierent exposure to the treatment
9
See Abadie et al for a full description of matching estimators
18
variable (export status) were used to identify the impact of exporting on the pounds of emissions variable. In the theoretical framework described above, there can be no exact matching of establishments that are exporters with those that are nonexporters in the same industry. Exporters are larger by denition. The nearest neighbor procedure will nd close matches between the smallest exporters and the largest non-exporters giving this procedure a regressiondiscontinuity-type estimate. Unfortunately, even at the 6 digit SIC level, industry denitions are not specic enough to allow for true regression discontinuity estimation. These industry denitions produce the overlap necessary for the matching estimator to be consistent. Table 6 describes the average eect of export status on two plants that are the same across the matched variables. The results consistently show that exporters pollute less than similar non-exporters though the dierence is not statistically signicant. Each establishment was matched over its sales, number of employees, relocations and credit rating. In the rst matching procedure establishments almost be in the same SIC 6-level industry. Plants that export had a sample average treatment eect (SATE) of 37,269 fewer pounds of emissions though this dierence is not signicant at standard levels. That amounts to around 12.6% of the average establishment emissions. This estimate is broadly similar to the one produced by the regression estimations above. Restricting matches by forcing establishments to be in the same state and from the same year reduced the signicance of this impact. Estimating the impact on logged sales improves the signicance of the estimated impact. The matching was conducted separately for each year to study the im19
pact of exporting status over time. The annual SATEs appear in table 7. Nine of the twelve yearly coecients are negative. The Z-scores vary between insignicant and marginally signicant. Splitting the sample size into twelve groups reduces the signicance somewhat, but exporting appears to have a negative impact on emissions overall. The impact of exporting on emissions appears to be growing over time. The last three years of data (1999-2001) show the strongest treatment eect. This generates further evidence that exporters pollute less than non-exporters, which is consistent with the theoretical framework described above.
Import Competitions Impact on Emissions
The previous empirical analysis has examined the relationship between export status and pollution. Import competitions impact on plant and industry pollution dynamics is likely to be more interesting. Melitz (2003) nds that import competition will force the least productive rms to exit in a given industry. The model described above suggests that those plants should generate the most emissions per unit of output. Their exit should result in a truncated distribution of emissions per unit of output and a cleaner industry on average after controlling for output. Following Pavcnik (2002) a variable is created using the ratio of imports(mj ) in a given industry to that industrys total output j ): (y 1 if mj > 0.1, yj ImportCompetitionj = 0 otherwise,
20
which is an indicator for industries that face sti import competition. Several dierent thresholds for exposure to import competition are tested.This variable is created using data from the NBERs collection of bilateral international trade date described in Feenstra et al (2002) and the productivity and output data described in Bartelsman and Gray (1996). The trade and productivity data are reported annually at the four-digit SIC level. Data for 1990 was used to construct this variable. This eect is estimated using the following equation:
Eijt = + W Wijt + Mjt + j + t +
ijt ,
where, as above, Eijt is a plant level measure of of emissions, Wijt is a vector of plant characteristics that serve as controls and t is set of year xed eects. Mjt is an industry level indicator variable that takes a value of 1 if the industry faces import competition and a value of 0 if the industry does not. This variable is calculated at the four-digit SIC industry level, for this reason is a set of industry xed eects at the two-digit SIC level in this specication. is the parameter of interest in this set of regressions. It is identied from dierences in emissions levels between plants in the same two-digit industry, whose four digit industries dier in exposure to import competition. The results of this specication are described in table 8. Regressions 811 test the various thresholds for dening import competition. Regression 8 uses a 15% ratio of imports to output, which is the same level used in Pavcnik (2002). At this level the impact of import competition is negative, but it 21
is not statistically signicant. Regression 9 reports a 25% threshold and regression 11 reports a 10% threshold. As the threshold increases the impact of import competition increases and becomes more statistically signicant. Industries with imports greater than 25% of output pollute nearly 77% less per unit of output than other industries that do not face this high level of competition. The import competition variables are dened at the four-digit SIC level meaning that the industry xed eects for these regressions must be at the two-digit SIC level. Regressions 12 and 13 examine the impact of exporting on regressions taking import competition into account. Regression 12 indicates that exporters pollute around 10% less than non-exporters in the same industry. The impact for exports in an import competing industry is even lager. Regression 13 uses an import competition-exporter xed eect to estimate the impact of export status on plant level emissions in industries classied as import competing. The indicator takes the value 1 if rm is an exporter and in an import competing industry and 0 otherwise. Those rms pollute 57% less than other rms after controlling for output and the dierence is signicant at the 1% level. This suggests that the productivity eect described in the theoretical framework outweighs any output eect. The theoretical framework also suggests that the establishments that exit an import competing industry should be the smallest, least productive ones with the highest level of emissions per unit of output. A straightforward test of this implication is to compare the size and emissions per unit of output level for rms facing import competition to those that are not. Table 9 compares the rst percentile of rms that face import competition compared to those that do not for a few example two-digit SIC industries. These 22
industries have the highest concentrations of rms in four-digit industries that face high levels of import competition. Import competition 15 equals to 1 indicates the size measures for establishments that face imports equal to at least 15% of total domestic output. In each case the smallest establishments that are subject to import competition are larger than those that are not. This is consistent with the smallest rms exiting the market due to import competition. The theoretical framework further suggested that import competing industries should have a smaller variance of emissions per unit of output. Table 10 reports a series of industry level regressions to test these implications. In each regression the independent variable is the level of import competition for the four-digit SIC level industry. Regression 13 tests the impact of import competition on average industry pollution. As expected there is a small negative impact, but the coecient is not signicant. The average industry emissions drop when import competition is sti. The results are similar if industry output is included as an explanatory variable. Regression 14 and 15 test the impact of import competition on the variance of industry sales and emissions. In both cases import competition reduces the variance, but the coecients are not signicant. Regressions 16 and 17 test the impact of import competition on the size of the smallest rms in an industry by using the 1st percentile of the distribution as the dependent variable. Industries that face import competition have larger smallest rms, suggesting that the smallest rms have been forced exit (or never enter) the market. The fact that the smallest rms that face import competition pollute more than the smallest rms that do not suggests that the output eect outweighs the 23
productivity eect. The results described in this section are consistent with the theoretical framework described in the previous section. Taken together they suggest that import competition leads to the exit of the smallest rms. Those establishments tend to pollute more per unit of output than their competitors. Their departure reduces the industry output per unit of emissions. These results are not consistent with the pollution haven hypothesis, but further evidence is considered below.
Source of Imports
The results in the previous section have described the impact of import competition on pollution emissions. The results are consistent with the model described at the beginning of this study. The distribution of emissions per unit of output does not appear to be consistent with the pollution haven hypothesis, but because of the policy import of this issue, this study will further analyze the relationship between imports and the pollution haven hypothesis by considering the source of those imports and its impact on plant level emissions. This can be done taking advantage of the bilateral trade data described above. This data was used to pinpoint the source of imports. Those sources were then matched with their per capita GDP and measures of their environmental stringency from the Environmental Performance Index (EPI). The EPI compares countries across more than 20 measures of environmental outcomes and policies. This data was used to create a weighted average of environmental measures and income for each 24
industrys imports. The higher the measure the better the environmental performance of the countries that import this sectors goods to the U.S. The measures of environmental performance and income embodied in U.S. imports are highly correlated, which reects the strong relationship between environmental regulation and income. If the reduction in emissions in import competing industries is due to pollution-haven-type eects then industries which receive the majority of their imports from countries with the lowest EPI scores should see the the biggest drops. This would be consistent with establishments relocating to take advantage of the lower levels of regulation. In fact, the results suggest that the source of imports has little impact on the reductions in emissions. Table 11 reports the regressions describing the relationship between the source of the imports and plant-level environmental performance. The EPI competition variable is the average of the environmental performance score for each country that exports goods to the U.S. weighted by the quantity of exports. The GDP competition variable is a similar variable measured for the GDP of the exporting country and the environmental competition variable uses a subset of EPI data to calculate a pollution score10 . Regression 18 reports the impact of the EPI competition score. The higher the level of EPI competition embodied in imports the higher the level of emissions produced by plants in that industry. This relationship is neither particularly strong, nor statistically signicant. Per capita GDP is highly correlated with environmental regulation levels. For that reason regressions 20 reports the
In addition to pollution measures the EPI includes measure of ecosystem health. There is a high degree of correlation between ecosystem, pollution and aggregate EPI scores. The results described here are robust to the inclusion of other measures.
10
25
impact of the per capita GDP embodied in imports. Again the relationship is positive, but statistically insignicant. The nal measure of environmental competition is the EPIs pollution prevention specic score, reported in regression 21. Again the relationship with establishment emissions is positive and statistically insignicant. Regression 19 tests the impact of the source of imports on the import competition variable described earlier. Introducing a measure of the environmental competition embodied in imports does not eect the conclusion that import competition drives down plant emissions. The nal regression conrms that exporting plants pollute less than their competitors even after import competition and import source are taken into account. Further analysis of the relationship between import sources and plant level emissions is ongoing.
Conclusion
This study has sought to analyze the relationship between international trade and plant level environmental performance. The empirical results are largely consistent with a model of heterogenous plants that vary in productivity. The relationship between productivity and output per unit of emissions is strong and consistent. Exporters consistently pollute less than non-exporters after controlling for a laundry-list of other variables. Industries that face import competition tend to pollute less on average than those who do not. The higher the level of import competition the further the reduction in average emissions. The reduction in average emissions is a result
26
of the exit of the smallest, least productive plants. These plants tend to generated more emissions per unit of output than their more productive competitors. This is inconsistent with plants relocating to take advantage of lax environmental regulation in other countries. To conrm that this result is not a function of the pollution haven hypothesis, this study takes advantage of the variation in sources of imports across industries. There source of imports seems to have little impact on the behavior of plant level emissions. This study bring up a host of interesting questions about the impact of trade policy on pollution emissions. Most countries actively promote exporters. To the extent that exporting increases productivity, this should lead to a reduction in rm level emissions per unit of output and likely a reduction in overall emissions. Import competition is more sensitive politically, but the results of this study suggest that improvements in productivity generated by import competition should reduce plant level emissions in addition to broader economic eciency gains. This leaves a host of interesting questions about the impact of trade policy (tari rates, non-tari barriers and antidumping cases to name a few) on plant emissions unanswered. Any trade policy behavior that protects low productivity plants is likely to have negative environmental consequences.
Citations
Abadie, Alberto, David Drukker, Jane Leber Herr and Guido W. Imbens Implementing Matching Estimators for Average Treatment Eects in Stata. Stata Journal. Janurary, 2001. 1(1):1-18.
27
Antweiler, Werner, Brian Copeland and M. Scott Taylor, Is free trade good for the environment? American Economic Review. 91 (2001) 877-908. Arora, Seema and Timothy N. Cason. Why Do Firms Volunteer to Exceed Environmental Regulations? Understanding Participation in EPAs 33/50 Program. Land Economics. November, 1996 72(4):413-432. Bartelsman, Eric and Wayne Gray. The NBER Manufacturing Productivity Database. NBER Technical Working Paper 205. October, 2006. Bernard, Andrew B and J Bradford Jensen. Why do Some Firms Export? Review of Economics and Statistics. May 2004, 86(2):561-569. Bernard, Andrew B, Jonathan Eaton, J Bradford Jensen and Samuel Kortum. Plants, Productivity and International Trade. American Economic Review. November 2003, 93(4):1268-1290. Cole, Matthew A, Robert J R Elliott and Kenichi Shimamoto. Industrial Characteristics, Environmental Regulations and Air Pollution: An Analysis of the UK Manufacturing Sector. Journal of Environmental Economics and Management. July 2005, 50(1):121-143. Collins, Alan and Richard I. D. Harris. Does Plant Ownership Aect the Level of Pollution Abatement Expenditure? Land Economics. May 2002, 78(2):171-189. Copeland, Brian and M. Scott Taylor, Trade and Transboundary Pollution. American Economic Review. 85 (1995) 716-737. Earnhart, Dietrich. The Eects of Financial Status on Corporate Environmental Performance: Liquidity, Solvency and Bottom Line Success. University of Kansas Working Paper. August 2006. Feenstra, Robert, John Romalis and Peter Schott. U.S. Imports, Exports and Tari Data 1989-2001. NBER Working Paper w9387. December, 2002. Frankel, Jeery and Andrew Rose. Is Trade Good or Bad for the Environment? Sorting Out the Causality. The Review of Economics and Statistics. February, 2005. 85(1) 85-91. 28
Gray, Wayne and Ronald Shadbegian. Plant vintage, technology, and environmental regulation. Journal of Environmental Economics and Management. November, 2003. 46: 384-402. Konar, Shameek and Mark A Cohen. Information as Regulation: The Effect of Community Right to Know Laws on Toxic Emissions. Journal of Environmental Economics and Management. January 1997, 32(1):109-124. Levinson, Arik. Technology, International Trade and Pollution from U.S. Manufacturing. NBER Working Paper 13616. November 2007. Melitz, Marc J. The Impact of Trade on Intra-industry Reallocations and Aggregate Industry Productivity. Econometrica. November 2003. 71(6):16951725. Neumark, David, Junfu Zhang, and Brandon Wall. Employment Dynamics and Business Relocation: New Evidence from the National Establishment Time Series. Public Policy Institute of California Working Paper. November 2005. Pavcnik, Nina. Trade Liberalization, Exit, and Productivity Improvements: Evidence from Chilean Plants. The Review of Economic Studies. January, 2002 69:245-276. Shadbegian, Ronald and Wayne Gray. Pollution Abatement Expenditures and Plant-Level Productivity: A Production Function Approach. Ecological Economics. August, 2005. 54 196-208.
29
Variable Sales Employees Pounds Hazard Risk
N 148,085 148,085 148,085 148,085 148,085
Mean 307,931 294 294,091 2,007 2,790
Std. Dev. 885,846 692 1,976,998 23,836 61,115
Min 0.0374532 1 0.000172 0 0
Max 72,200,000 27,000 250,000,000 2,730,232 8,273,306
Table 1: Summary Statistics Variable Sales Employees Pounds of Emissions Hazard Score Risk Score Emissions Per Sale Exporter 313,006 307 280,841 2,194 3,434 4.3 Non-exporter 304,934 286 301,745 1,899 2,415 67.1 Dierence 8,071 21 -20,904 295 1,019 -62.8 T-stat 1.69 ** 5.53 *** -1.96 ** 2.30 ** 3.09 *** -0.93
Table 2: Comparing Exporters to Non-exporters Note: Dierence in means between exporters and non-exporters for selected variables. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 10% level.
30
Regression Num Dep Var Log Sales Employees (in 100s) Relocations Export Constant R2 N FE
1 Log Pounds . . . 0.07 (1.96)* 9.84 (477.32)*** 0.0006 148,085 SIC6
2 Log Pounds 0.60 (43.77)*** . . -0.08 (-2.26)** 0.08 (0.34) 0.08 148,085 SIC6
3 Log Pounds 0.58 (42.57)*** . . -0.06 (-1.79)* 3.25 (2.41)** 0.09 148,085 SIC6, State, Year
4 Log Pounds 0.54 (33.90)*** 0.02 (5.60)*** 0.12 (2.95)*** -0.06 (-1.76)* 3.92 (2.91)*** 0.10 148,085 SIC6, State, Year
Table 3: Exporters Pollution Emissions Note: All standard errors clustered at the establishment level. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 10% level.
31
Regression Num Dep Var Log Sales Export Constant R2 N Dep var avg Fixed Eects
5 Pounds 176,100 (11.65)*** -62,285 (-2.08)** -2,574,420 (-10.73)*** 0.01 148,085 294,060 SIC 6
6 Hazard 913.92 (5.52)*** 122.43 (0.36) -1,3043.38 (-4.83)*** 0.004 148,085 2007.17 SIC 6
7 Risk 1,268.88 (4.81)*** 526.75 (0.77) -18237.54 (-4.28)*** 0.006 148,085 2789.32 SIC 6
Table 4: Exporters Emissions Measures Note: Pounds are the quantity of emissions, hazard is a score that measures the quantity and toxicity of emissions and risk measures the quantity, toxicity and location of emissions. All standard errors clustered at the establishment level. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 10% level.
Percentile 0-20 20-40 40-60 60-80 80-100
Fraction Exporters 25% 37% 41% 39% 41%
Export Coe. 0.015 -0.134 -0.134 -0.013 -0.197
Export T-stat 0.41 -4.24 -4.29 -0.42 -6.45
Table 5: Regression Coecients From Firm Size Regressions Note: The export coecient and t-statistic are taken from the baseline regression with year and state xed eects.
32
0.1
70 0.05
0 60
Fra actionofFirm msExporting
50 0.1
40
0.15
0.2 30 0.25
20 020 2040 4060 6080 80100
0.3
Figure 1: Firm size regressions and condence intervals Note: The bar indicates the fraction of rms in a given quintile (read o the left axis). The line is the 95% condence interval for the exporter coecient in the baseline regression (read o the right axis). The square represents the point estimate.
33
ExporterCo oefficient
0.05
Dep Var Pounds Pounds Pounds Log Pounds
SATE -37269 -44559 615 -0.10798
Z Score -1.24 -1.44 0 -1.69
P-value 0.215 0.151 0.997 0.092
Exact Match SIC 6 SIC 6 State SIC 6 State Year SIC 6
Table 6: Nearest Neighbor Matching Estimators Note: The Sample Average Treatment Eect (SATE) measures the impact of treatment (in this case exporting) on emissions by comparing matched treated and untreated establishments that are similar across observable variables. Matching variables were sales, employees, relocations and credit ratings for each match.
Year 90 91 92 93 94 95 96 97 98 99 00 01
SATE 23,321 -10,212 -61,882 38,174 -42,659 -30,836 -28,579 -69,647 13,375 -47,641 -26,004 -69,996
Z-score 0.39 -0.15 -0.83 0.46 -1.54 -1.12 -1.01 -1.66 0.18 -1.28 -1.60 -1.67
Table 7: Average Treatment Eect By Year Note: The Sample Average Treatment Eect (SATE) measures the impact of treatment (in this case exporting) on emissions by comparing matched treated and untreated establishments that are similar across observable variables. The SATE is measured in pounds of emissions.
34
Dep Var Log Sales Import Comp 15 Import Comp 25 Import Comp 10 Export Ex-Im Interact Constant R2 N FE
8 Log Pounds 0.64 131.08*** -0.11 -1.1 . . . . -0.59 -7.34*** 0.1 148,133 SIC 2
9 Log Pounds 0.64 131.05*** . -0.77 -3.91***
10 Log Pounds 0.64 131.07*** . . 0.03 0.44 . . -0.59 -7.35*** 0.1 148,133 SIC 2
11 Log Pounds 0.64 131.16*** . -0.78 -3.94*** . -0.1 -7.17*** . -0.6 -7.52*** 0.1 148,133 SIC 2
12 Log Pounds 0.64 131.18*** . -0.56 -2.73*** . -0.1 -6.93*** -0.57 -3.70*** -0.6 -7.54*** 0.1 148,133 SIC 2
. . -0.58 -7.31*** 0.1 148,133 SIC 2
Table 8: Import Competition Note: Import Competition variables are industry-level dummies that indicate if more than X% of the sales in a particular industry come from imports. Those industries are dened as import competing. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 1% level
35
SIC Code 25 25 30 30 36 36 39 39
Sales $2,968 $20,810 $2,646 $1,569 $2,244 $6,999 $966 $4,280
Employees 8 41 3 3 3 6 2 8
Import Comp 15 0 1 0 1 0 1 0 1
Table 9: Smallest Firms in Selected Industries Note: This table displays the 1 percentile of rms by their exposure to import competition. Import competition 15=1 implies that more than 15% of industry sales occur through imports.
36
Dep Var Import Comp Constant R2 N
13 Avg Pounds -131,808 -0.2 241749 6.10*** 0 442
14 Var Pounds -437,020 -0.3 586610 6.49*** 0 441
15 Var Sales -466,358 -0.91 436647 13.93*** 0 441
16 Small Pounds 23,737 4.75*** 458 1.5 0.05 442
17 Small Sales 61,083 1 23564 6.30*** 0 442
Table 10: Import Competitions Eects Note: These regressions test the impact of import competition on moments of the distribution of establishment sales and pollution emissions. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 1% level.
37
Log Sales EPI Competition GDP Competition Pollution Competition Import Competition 25 Export Constant R2 N FE
18 Log Pounds 0.62 113.77*** 1.69 0.60
19 Log Pounds 0.62 113.71*** 1.54 0.68
20 Log Pounds 0.62 115.18***
21 Log Pounds 0.62 113.74***
22 Log Pounds 0.62 113.78*** 1.04 0.51
1.25 0.59 0.89 0.60 -0.82 -4.13***
-0.82 -4.13***
-0.82 -4.14***
-0.33 -3.63*** 0.1 122397 SIC 2
-0.32 -3.57*** 0.1 122397 SIC 2
-0.39 -4.41*** 0.1 122397 SIC 2
-0.32 -3.58*** 0.1 122397 SIC 2
-0.82 -4.15*** -0.08 -4.95*** -0.32 -3.59*** 0.1 122397 SIC 2
Table 11: The pollution haven hypothesis and Environmental Competition Note: The EPI, GDP and Pollution competition variables measure the average of EPI, GDP and Pollution Prevention Index across countries that send imports to the United States. The average is weighted by the value of imports. The import competition variable is an indicator that equals one if more than 25% of an industrys sales are from imports. *** signicant at the 1% level, ** signicant at the 5% level, * signicant at the 1% level.
38