World Development Vol. 32, No. 8, pp. 1419–1439, 2004 2004 Elsevier Ltd. All rights reserved Printed in Great Britain www.elsevier.com/locate/worlddev 0305-750X/$ - see front matter doi:10.1016/j.worlddev.2004.03.004 The Rise and Fall of the Environmental Kuznets Curve

DAVID I. STERN * Rensselaer Polytechnic Institute, Troy, NY, USA Available online Summary. — This paper presents a critical history of the environmental Kuznets curve (EKC). The EKC proposes that indicators of environmental degradation first rise, and then fall with increasing income per capita. Recent evidence shows however, that developing countries are addressing environmental issues, sometimes adopting developed country standards with a short time lag and sometimes performing better than some wealthy countries, and that the EKC results have a very flimsy statistical foundation. A new generation of decomposition and efficient frontier models can help disentangle the true relations between development and the environment and may lead to the demise of the classic EKC. 2004 Elsevier Ltd. All rights reserved.

Key words — environmental Kuznets curve, pollution, economic development, econometrics, review, global

1. INTRODUCTION movement and associated scientists in the past (e.g., Meadows, Meadows, Randers, & Beh- The environmental Kuznets curve (EKC) is a rens, 1972), economic growth would be the hypothesized relationship between various means to eventual environmental improvement. indicators of environmental degradation and This change in thinking was already underway income per capita. In the early stages of eco- in the emerging idea of sustainable economic nomic growth degradation and pollution development promulgated by the World Com- increase, but beyond some level of income per mission on Environment and Development capita, which will vary for different indicators, (1987) in Our Common Future. The possibility the trend reverses, so that at high income levels of achieving sustainability without a significant economic growth leads to environmental deviation from business as usual was an obvi- improvement. This implies that the environ- ously enticing prospect for many––letting mental impact indicator is an inverted U-shaped function of income per capita. Typically, the logarithm of the indicator is modeled as a quadratic function of the logarithm of income. * I thank Cutler Cleveland, Quentin Grafton and two An example of an estimated EKC is shown in anonymous referees, for comments on drafts of this Figure 1. The EKC is named for Kuznets (1955) manuscript. Participants at seminars at the University of who hypothesized that income inequality first Wollongong, Clark University, University of York, rises and then falls as economic development University of Cambridge, Rensselaer Polytechnic Insti- proceeds. tute, and the Australian National University and at The EKC concept emerged in the early 1990s conferences organized by the Association of American with Grossman and Krueger’s (1991) path- Geographers in Los Angeles and the International breaking study of the potential impacts of Society for Ecological Economics in Sousse, Tunisia NAFTA and the concept’s popularization aided in developing my arguments. Additionally, I am through the 1992 World Bank Development grateful to my collaborators Mick Common, Roger Report (IBRD, 1992). If the EKC hypothesis Perman, and Kali Sanyal for their contributions dis- were true, then rather than being a threat to the cussed in this paper. Final revision accepted: 23 March environment, as claimed by the environmental 2004. 1419 1420 WORLD DEVELOPMENT

300

250

200

150 per Capita 2

100 Kg SO

50

0 0 5000 10000 15000 20000 25000 30000 $ GNP per Capita

Figure 1. Environmental Kuznets curve for sulfur emissions. Source: Panayotou (1993) and Stern, Common, and Barbier (1996). humankind ‘‘have our cake and eat it’’ (Rees, is, itself, lacking in wisdom. In rapidly growing 1990, p. 435). middle-income countries, however the scale The EKC is an essentially empirical phe- effect, which increases pollution and other nomenon, but most of the EKC literature is degradation, overwhelms the time effect. In econometrically weak. In particular, little or no wealthy countries, growth is slower, and pol- attention has been paid to the statistical prop- lution reduction efforts can overcome the scale erties of the data used––such as serial depen- effect. This is the origin of the apparent EKC dence or stochastic trends in time-series 1––and effect. The econometric results are supported by little consideration has been paid to issues of recent evidence that, in fact, pollution problems model adequacy such as the possibility of are being addressed and remedied in developing omitted variables bias. 2 Most studies assume economies (e.g., Dasgupta, Laplante, Wang, & that, if the regression coefficients are nominally Wheeler, 2002). individually or jointly significant and have the This paper follows the development of the expected signs, then an EKC relation exists. EKC concept in approximately chronological However, one of the main purposes of doing order. I do not attempt to review or cite all of econometrics is to test which apparent rela- the rapidly growing number of studies. The tionships, or ‘‘stylized facts,’’ are valid and next two sections of the paper review in more which are spurious correlations. detail the theory behind the EKC and the When we do take diagnostic statistics and econometric methods used in EKC studies. The specification tests into account and use appro- following sections review some EKC analyses priate techniques, we find that the EKC does and their critique. Sections 6 and 7 discuss the not exist (Perman & Stern, 2003). Instead, we more important recent developments that have get a more realistic view of the effect of eco- changed the picture that we have of the EKC. nomic growth and technological changes on The final sections discuss alternative approa- environmental quality. It seems that emissions ches––decomposition of emissions and efficient of most pollutants and flows of waste are frontiers––and summarize the findings. monotonically rising with income, though the ‘‘income elasticity’’ is less than one and is not a simple function of income alone. Income-inde- 2. THEORETICAL BACKGROUND pendent, time-related effects reduce environ- mental impacts in countries at all levels of The EKC concept emerged in the early 1990s income. The new (post-Brundtland) conven- with Grossman and Krueger’s (1991) path- tional wisdom that developing countries are breaking study of the potential impacts of ‘‘too poor to be green’’ (Martinez-Alier, 1995) NAFTA and Shafik and Bandyopadhyay’s THE ENVIRONMENTAL KUZNETS CURVE 1421

(1992) background study for the 1992 World structure or product mix, changes in techno- Development Report. The EKC theme was logy, and changes in input mix, as well as popularized by the World Bank’s World underlying causes such as environmental regu- Development Report 1992 (IBRD, 1992), which lation, awareness, and education, which can argued that: ‘‘The view that greater economic only have an effect via the proximate variables. activity inevitably hurts the environment is Let us look in more detail at the proximate based on static assumptions about technology, variables: tastes and environmental investments’’ (p. 38) (a) Scale of production implies expanding and that ‘‘As incomes rise, the demand for production at given factor-input ratios, out- improvements in environmental quality will put mix, and state of technology. It is nor- increase, as will the resources available for mally assumed that a 1% increase in scale investment’’ (p. 39). Others have expounded results in a 1% increase in emissions. This this position even more forcefully with Beck- is because if there is no change in the in- erman (1992) claiming that ‘‘there is clear evi- put–output ratio or in technique there has dence that, although economic growth usually to be a proportional increase in aggregate leads to environmental degradation in the early inputs. However, there could, in theory, be stages of the process, in the end the best––and scale economies or diseconomies of pollu- probably the only––way to attain a decent tion (Andreoni & Levinson, 2001). Some environment in most countries is to become pollution control techniques may not be rich.’’ (p. 491). In his highly publicized and practical at a small scale of production controversial book, The Skeptical Environmen- and vice versa or may operate more or less talist, Lomborg (2001) relies heavily on the effectively at different levels of output. 1992 World Development Report (Cole, 2003a) (b) Different industries have different pollu- to argue the same point, while many environ- tion intensities. Typically, over the course of mental economists take the EKC as a stylized economic development the output mix fact that needs to be explained by theory. All changes. In the earlier phases of develop- this is despite the fact that the EKC has never ment there is a shift away from agriculture been shown to apply to all pollutants or envi- toward heavy industry which increases ronmental impacts and recent evidence, 3 dis- emissions, while in the later stages of devel- cussed in this paper, challenges the notion of opment there is a shift from the more re- the EKC in general. The remainder of this source intensive extractive and heavy section discusses the economic factors that industrial sectors toward services and light- drive changes in environmental impacts and er manufacturing, which supposedly have may be responsible for rising or declining lower emissions per unit of output. 4 environmental degradation over the course of (c) Changes in input mix involve the substitu- economic development. tion of less environmentally damaging inputs If there were no change in the structure or for more damaging inputs and vice versa. technology of the economy, pure growth in the Examples include: substituting natural gas scale of the economy would result in growth in for coal and substituting low sulfur coal in pollution and other environmental impacts. place of high sulfur coal. As scale, output This is called the scale effect. The traditional mix, and technology are held constant, this view that economic development and environ- is equivalent to moving along the isoquants mental quality are conflicting goals reflects the of a neoclassical production function. scale effect alone. Proponents of the EKC (d) Improvements in the state of technology hypothesis argue that involve changes in both: • Productivity in terms of using less, ceteris at higher levels of development, structural change to- paribus of the polluting inputs per unit of wards information-intensive industries and services, output. A general increase in total factor coupled with increased environmental awareness, productivity will result in lower emis- enforcement of environmental regulations, better sions per unit of output even though this technology and higher environmental expenditures, is not necessarily an intended conse- result in leveling off and gradual decline of environ- quence. mental degradation (Panayotou, 1993, p. 1). • Emissions specific changes in process re- sult in lower emissions per unit of input. Thus there are both proximate causes of the These innovations are specifically in- EKC relationship––scale, changes in economic tended to reduce emissions. 5 1422 WORLD DEVELOPMENT

This framework applies most directly to to produce an inverted U-shaped curve is not a emissions of pollutants. For concentrations of particularly desirable property. pollutants, decentralization of economic activity with development is also important (Stern et al., 1996). Deforestation is also a flow of envi- 3. ECONOMETRIC FRAMEWORK ronmental degradation. Improved technology would imply more replanting, selective cutting, The earliest EKCs were simple quadratic wood recovery etc. that reduces deforestation functions of the levels of income. But, eco- per unit wood produced. Stock pollutants or nomic activity inevitably implies the use of impacts need a different, dynamic framework. resources and, by the laws of thermodynamics, Though any actual change in the level of use of resources inevitably implies the produc- pollution must be a result of change in one of tion of waste. Regressions that allow levels of the proximate variables, those variables may be indicators to become zero or negative are driven by changes in underlying variables that inappropriate except in the case of deforesta- also vary over the course of economic devel- tion where afforestation can occur. A logarith- opment. A number of papers have developed mic dependent variable will impose this theoretical models of how preferences and restriction. Some studies, including the original technology might interact to result in different Grossman and Krueger (1991) paper, used a time paths of environmental quality. The vari- cubic EKC in levels and found an N-shape ous studies make different simplifying assump- EKC. This might just be a polynomial tions about the economy. Most of these studies approximation to a logarithmic curve. The can generate an inverted U-shape curve of standard EKC regression model is, therefore: pollution intensity but there is no inevitability about this. The result depends on the assump- lnðE=PÞit ¼ ai þ ct þ b1 lnðGDP=PÞit tions made and the values of particular þ b ðlnðGDP=PÞÞ2 þ e ; ð1Þ parameters. Lopez (1994) and Selden and Song 2 it it (1995) assume infinitely lived agents, exogenous where E is emissions, P is population, and ln technological change and that pollution is indicates natural logarithms. The first two generated by production and not by consump- terms on the RHS are intercept parameters tion. John and Pecchenino (1994), John, Pec- which vary across countries or regions i and chenino, Schimmelpfennig, and Schreft (1995), years t. The assumption is that, though the level and McConnell (1997) develop models based of emissions per capita may differ over coun- on overlapping generations where pollution is tries at any particular income level, the income generated by consumption rather than by pro- elasticity is the same in all countries at a given duction activities. In addition, Stokey (1998) income level. The time specific intercepts allows endogenous technical change and Lieb account for time-varying omitted variables and (2001) generalizes Stokey’s (1998) model, stochastic shocks that are common to all arguing that satiation in consumption is needed countries. The ‘‘turning point’’ income, where to generate the EKC. Finally, Ansuategi and emissions or concentrations are at a maximum, Perrings (2000) incorporate transboundary is given by: externalities. Magnani (2001) discusses how s ¼ expðb =ð2b ÞÞ: ð2Þ individual preferences are converted into public 1 2 policy. Andreoni and Levinson (2001) argue Usually the model is estimated with panel that none of these special assumptions is data. Most studies attempt to estimate both the needed and economies of scale in abatement are fixed and random-effects models. The fixed- sufficient to generate the EKC. Most studies effects model treats the ai and ct as regression model the emission of pollutants. Lopez (1994) parameters. The random-effects model treats and Bulte and van Soest (2001), among others, the ai and ct as components of the random develop models for the depletion of natural disturbance. If the effects ai and ct and the resources such as forests or agricultural land explanatory variables are correlated, then the fertility. It seems easy to develop models that random-effects model cannot be estimated generate EKCs under appropriate assumptions. consistently (Hsiao, 1986). Only the fixed- None of these theoretical models has been effects model can be estimated consistently. A tested empirically. Furthermore, if, in fact, the Hausman (1978) test can be used to test for EKC for emissions is monotonic, as more inconsistency in the random-effects estimate by recent evidence suggests, the ability of a model comparing the fixed-effects and random-effects THE ENVIRONMENTAL KUZNETS CURVE 1423 slope parameters. A significant difference indi- explanatory variables have been estimated. But cates that the random-effects model is estimated the key features differentiating the models for inconsistently, due to correlation between the different pollutants, data etc. can be displayed explanatory variables and the error compo- by reviewing a few of the early studies and nents. Assuming that there are no other statis- examining a single impact in more detail. I tical problems, the fixed-effects model can be review the contributions of Grossman and estimated consistently, but the estimated Krueger (1991), Shafik (1994), and Selden and parameters are conditional on the country and Song (1994) and then look in more detail at time effects in the selected sample of data studies for sulfur pollution and emissions. (Hsiao, 1986). Therefore, they cannot be used Finally, I briefly discuss studies that estimate an to extrapolate to other samples of data. This EKC for energy use. means that an EKC estimated with fixed-effects Many EKC studies have also been published using only developed country data might say that include additional explanatory additional little about the future behavior of developing explanatory variables, intended to model countries. Many studies compute the Hausman underlying or proximate factors, such as statistic and, finding that the random-effects ‘‘political freedom’’ (e.g., Torras & Boyce, 1998) model cannot be consistently estimated, esti- or output structure (e.g., Panayotou, 1997), mate the fixed-effects model. But few have or trade (e.g., Suri & Chapman, 1998). Stern pondered the deeper implications of the failure (1998) reviews several of these. In general, the of this orthogonality test. included variables turn out to be significant at GDP may be an integrated variable. If the traditional significance levels. Testing different EKC regressions do not cointegrate, then the variables individually is however subject to the estimates will be spurious. Until recently, very problem of potential omitted variables bias. few studies have reported any diagnostic sta- Further, these studies do not report cointegra- tistics for integration of the variables or coin- tion or other statistics that might tell us if tegration of the regressions. Therefore, it is omitted variables bias is likely to be a problem unclear what we can infer from the majority of or not. Therefore, it is not clear what we can EKC studies. Testing for integration and coin- infer from this body of work. Given these tegration in panel data is a rapidly developing problems, I do not review these studies sys- field. Perman and Stern (2003) employ some of tematically here. these tests and find that sulfur emissions and To some (e.g., Lopez, 1994) the early EKC GDP per capita may be integrated variables. studies indicated that local pollutants were The unit root hypothesis could be rejected for more likely to display an inverted U-shape sulfur (but not GDP) using the Im, Pesaran, and relation with income, while global impacts such Shin (2003) (IPS) test when the alternative was as carbon dioxide did not. This picture fits trend stationarity. But alternative hypotheses environmental economics theory––local impacts and tests result in acceptance of the unit root are internalized within a single economy or hypothesis. Heil and Selden (1999) find the same region and are likely to give rise to environ- result for carbon dioxide emissions and GDP mental policies to correct the externalities on using the IPS test. But they prefer results that pollutees before such policies are applied to allow for a structural break 1974, in which globally externalized problems. But as we will allows them to strongly reject the unit root see, the picture is not quite so clear cut even in hypothesis for both GDP and carbon. Coondoo the early studies. Furthermore, the more recent and Dinda (2002) yield similar results to Per- evidence on sulfur and carbon dioxide emis- man and Stern (2003) for carbon dioxide emis- sions shows there may be no strong distinction sions. de Bruyn (2000) and Day and Grafton between the effect of income per capita on local (2003) carry out time-series unit root tests for and global pollutants. Stern et al. (1996) the Netherlands, the United Kingdom, the determined that higher turning points were United States, West Germany, and Canada for found for regressions that used purchasing a variety of pollutants with very similar results. power parity (PPP) adjusted income compared to those that used market exchange rates and for studies using emissions of pollutants rela- 4. RESULTS OF EKC STUDIES tive to studies using ambient concentrations in urban areas. In the initial stages of economic Many basic EKC models relating environ- development urban and industrial development mental impacts to income without additional tends to become more concentrated in a smaller 1424 WORLD DEVELOPMENT number of cities which also have rising central Selden and Song (1994) estimated EKCs for population densities. Many developing coun- four emissions series: SO2,NOx, SPM, and CO tries have a ‘‘primate city’’ that dominates a using longitudinal data primarily from devel- country’s urban hierarchy and contains much oped countries. The estimated turning points of its modern industry––Bangkok is one of the were all very high compared to the two earlier best such examples. In the later stages of eco- studies. For the fixed-effects version of their nomic development, urban and industrial model they were (converted to US$1,990 using development tends to decentralize. Moreover, the US GDP implicit price deflator): SO2, the high population densities of less-developed $10,391; NOx, $13,383; SPM, $12,275; and CO, cities are gradually reduced by suburbaniza- $7,114. This study showed that the turning tion. So, it is possible for peak ambient pollu- point for emissions was likely to be higher than tion concentrations to fall as income rises even that for ambient concentrations. if total national emissions are rising. Table 1 summarizes several studies of sulfur The first empirical EKC study was the emissions and concentrations, listed in order of NBER working paper by Grossman and Krue- estimated income turning point. Panayotou 6 ger (1991) that estimated EKCs for SO2, dark (1993) used cross-sectional data, nominal GDP, matter (fine smoke), and suspended particles and the assumption that the emission factor for (SPM) using the GEMS dataset as part of each fuel is the same in all countries––this study a study of the potential environmental impacts has the lowest estimated turning point of all. of NAFTA. The GEMS dataset is a panel With the exception of the Kaufmann et al. of ambient measurements from a number of (1998) estimate, all turning point estimates locations in cities around the world. Each using concentration data are less than $6,000. regression involved a cubic function in levels Kaufmann et al. (1998) used an unusual speci- (not logarithms) of PPP per capita GDP and fication that includes GDP per area and GDP various site-related variables, a time trend, and per area squared variables. a trade intensity variable. The turning points Among the emissions based estimates, both for SO2 and dark matter were at around Selden and Song (1994) and Cole et al. (1997) $4,000–5,000 while the concentration of sus- used databases dominated by, or consisting pended particles appeared to decline even at solely of, emissions from OECD countries. low income levels. At income levels over Their estimated turning points are $10,391 and $10,000–15,000, Grossman and Krueger’s esti- $8,232 respectively. List and Gallet (1999) used mates show increasing levels of all three pollu- data for the 50 US states from 1929–94. Their tants. estimated turning point is the second highest in The results of Shafik and Bandyopadhyay’s the table. Income per capita in their sample (1992) 7 study were used in the 1992 World ranges from $1,162 to $22,462 in 1987 US Development Report (IBRD, 1992) and were, dollars. This is a greater range of income levels therefore, particularly influential. They esti- than is found in the OECD-based panels for mated EKCs for 10 different indicators using recent decades. This suggests that including three different functional forms: log-linear, log- more low-income data points in the sample quadratic and, in the most general case, a log- might yield a higher turning point. Stern and arithmic cubic polynomial in PPP GDP per Common (2001) estimated the turning point at capita as well as a time-trend and site-related over $100,000. They used an emissions data- variables. In each case, the dependent variable base produced for the US Department of was untransformed. They found that lack of Energy by ASL (Lefohn, Husar, & Husar, clean water and lack of urban sanitation decline 1999) that covers a greater range of income uniformly with increasing income, and over levels and includes more data points than any time. Both measures of deforestation were of the other sulfur EKC studies. found to be insignificantly related to the income We see that the recent studies that used more terms. River quality tended to worsen with representative samples find that there is a increasing income. The two air pollutants, monotonic relation between sulfur emissions however, conform to the EKC hypothesis. The and income just as there is between carbon turning points for both pollutants were at dioxide and income. Interestingly, Dijkgraaf income levels of between $3,000 and $4,000. and Vollebergh (1998) estimate a carbon EKC Finally, both municipal waste and carbon for a panel data set of OECD countries find- emissions per capita increased unambiguously ing an inverted U-shape EKC in the sample with rising income. as a whole (as well as many signs of poor Table 1. Sulfur EKC studies Authors Turning point 1990 Emissions or PPP Additional Data source for Time period Countries/cities USD concentrations variables sulfur Panayotou (1993) $3,137 Emissions No – Own estimates 1987–88 55 developed and developing countries Shafik (1994) $4,379 Concentrations Yes Time trend, loca- GEMS 1972–88 47 cities in 31 tional dummies countries

Torras and Boyce $4,641 Concentrations Yes Income inequal- GEMS 1977–91 1425 Unknown num- CURVE KUZNETS ENVIRONMENTAL THE (1998) ity, literacy, ber of cities in 42 political and civil countries rights, urbaniza- tion, locational dummies Grossman and $4,772–5,965 Concentrations No Locational GEMS 1977, ‘82, ‘88 Up to 52 cities in Krueger (1991) dummies, popu- up to 32 coun- lation density, tries trend Panayotou (1997) $5,965 Concentrations No Population GEMS 1982–84 Cities in 30 deve- density, policy loped and deve- variables loping countries Cole, Rayner, and $8,232 Emissions Yes Country dummy, OECD 1970–92 11 OECD Bates (1997) technology level countries Selden and Song $10,391–10,620 Emissions Yes Population WRI––primarily 1979–87 22 OECD and 8 (1994) density OECD source developing countries Kaufmann, $14,730 Concentrations Yes GDP/Area, steel UN 1974–89 13 developed and Davidsdottir, exports/GDP 10 developing Garnham, and countries Pauly (1998) List and Gallet $22,675 Emissions N/A – US EPA 1929–94 US States (1999) Stern and $101,166 Emissions Yes Time and coun- ASL 1960–90 73 developed and Common (2001) try effects developing coun- tries 1426 WORLD DEVELOPMENT econometric behavior). The turning point is at The key criticism of Arrow et al. (1995) and only 54% of maximal GDP in the sample. A others was that the EKC model, as presented in study by Schmalensee, Stoker, and Judson the 1992 World Development Report and else- (1998) also finds a within sample turning point where, assumes that there is no feedback from for carbon. In this case, a 10-piece spline was environmental damage to economic production fitted to the data such that the coefficient esti- as income is assumed to be an exogenous var- mates for high-income countries are allowed to iable. The assumption is that environmental vary from those for low-income countries. All damage does not reduce economic activity these studies suggest that the differences in sufficiently to stop the growth process and that turning points that have been found for differ- any irreversibility is not so severe that it reduces ent pollutants may be due, at least partly, to the the level of income in the future. In other different samples used. I will discuss the words, there is an assumption that the economy econometric reasons for this sample-dependent is sustainable. But, if higher levels of economic behavior below. activity are not sustainable, attempting to grow In an attempt to capture all environmental fast in the early stages of development when impacts of whatever type, a number of environmental degradation is rising may prove researchers (e.g., Suri & Chapman, 1998; Cole counterproductive. 8 et al., 1997) have estimated EKCs for a proxy It is clear that emissions of many pollutants total environmental impact indicator––total per unit of output have declined over time in energy use. In each case, they found that energy developed countries with increasingly stringent use per capita increases monotonically with environmental regulations and technical inno- income per capita. This result does not preclude vations. But the mix of residuals has shifted the possibility that energy intensity––energy from sulfur and nitrogen oxides to carbon used per dollar of GDP produced––declines dioxide and solid waste so that aggregate waste with rising income or even follows an inverted is still high and per capita waste may not have U-shaped path (e.g., Galli, 1998). declined. 9 Economic activity is inevitably The only robust conclusions from the EKC environmentally disruptive in some way. Sat- literature appear to be that concentrations of isfying the material needs of people requires the pollutants may decline from middle income use and disturbance of energy flows and mate- levels, while emissions tend to be monotonic in rials stocks. Therefore, an effort to reduce some income. As we will see below, emissions may environmental impacts may just aggravate decline simultaneously over time in countries at other problems. 10 widely varying levels of development. Given the Both Arrow et al. (1995) and Stern et al. poor statistical properties of most EKC mod- (1996) argued that, if there was an EKC type els, it is hard to come to any conclusions about relationship, it might be partly or largely a the roles of other additional variables such as result of the effects of trade on the distribution trade. Too few quality studies have been done of polluting industries. The Hecksher-Ohlin of other indicators apart from air pollution to trade theory suggests that, under free trade, come to any firm conclusions about those developing countries would specialize in the impacts either. production of goods that are intensive in the factors that they are endowed with in relative abundance: labor and natural resources. The 5. THEORETICAL CRITIQUE OF THE developed countries would specialize in human EKC capital and manufactured capital intensive activities. Part of the reduction in environ- A number of critical surveys of the EKC mental degradation levels in the developed literature have been published (e.g., Ansuategi, countries and increases in environmental deg- Barbier, & Perrings, 1998; Arrow et al., 1995; radation in middle income countries may reflect Copeland & Taylor (2004); Dasgupta et al., this specialization (Hettige, Lucas, & Wheeler, 2002; Ekins, 1997; Pearson, 1994; Stern, 1998; 1992; Lucas, Wheeler, & Hettige, 1992; Suri & Stern et al., 1996). This section discusses the Chapman, 1998). Environmental regulation in criticisms raised against the EKC in the earlier developed countries might further encourage surveys on theoretical (rather than methodo- polluting activities to gravitate toward the logical) grounds. The more recent surveys raise developing countries (Lucas et al., 1992). These similar points but have more evidence to mar- effects would exaggerate any apparent decline shal. in pollution intensity with rising income along THE ENVIRONMENTAL KUZNETS CURVE 1427 the EKC. In our finite world the poor countries available econometric estimates might have of today would be unable to find further lead one to believe that, given likely future countries from which to import resource- levels of mean income per capita, environmen- intensive products as they, themselves, become tal degradation should decline from the present wealthy. When the poorer countries apply onward. This interpretation is evident in the similar levels of environmental regulation they 1992 World Development Report (IBRD, 1992). would face the more difficult task of abating Income is not however, normally distributed these activities rather than outsourcing them to but very skewed, with much larger numbers of other countries (Arrow et al., 1995; Stern et al., people below mean income per capita than 1996). Copeland and Taylor (2004) conclude above it. Therefore, it is median rather than that, in contrast to earlier work (e.g., Jaffe, mean income that is the relevant variable. Sel- Peterson, Portney, & Stavins, 1995), recent den and Song (1994) and Stern et al. (1996) research shows that increased regulation does performed simulations that, assuming that the tend to result in more decisions to locate in less EKC relationship is valid, showed that global regulated locations. On the other hand, there is environmental degradation was set to rise for a no clear evidence that trade liberalization long time to come. Figure 2 presents projected results in a shift in polluting activities to less- sulfur emissions using the EKC in Figure 1 and regulated countries. UN and World Bank forecasts of economic and Furthermore, Antweiler, Copeland, and population growth. Despite this and despite Taylor (2001) and Cole and Elliott (2003) argue recent estimates that indicate higher or nonex- that the capital-intensive activities that are istent turning points, the impression produced concentrated in the developed countries are by the early studies in the policy, academic, and more polluting and hence developed countries business communities seems slow to fade (e.g., have a natural comparative advantage in pol- Lomborg, 2001). luting goods in the absence of regulatory dif- ferences. There are no clear answers on the impact of trade on pollution from the empirical 6. RECENT DEVELOPMENTS EKC literature. Stern et al. (1996) argued that early EKC Significant developments, since my last gen- studies showed that a number of indicators: eral survey of the EKC in 1998, fall into three SO2 emissions, NOx, and deforestation, peak at classes: (a) Empirical case study evidence on income levels around the current world mean environmental performance and policy in per capita income. A cursory glance at the developing countries that is discussed in this

1.40E+09

1.20E+09

1.00E+09 2 8.00E+08

6.00E+08 Tons SO 4.00E+08

2.00E+08

0.00E+00 1990 1995 2000 2005 2010 2015 2020 2025 Year

Figure 2. Projected sulfur emissions. Source: Stern et al. (1996). 1428 WORLD DEVELOPMENT section; (b) improved econometric testing and review the theoretical literature and some of the estimates discussed in the following section; econometric specification issues. But their main and (c) a new wave in the investigation of contribution is presenting evidence that envi- environment-development relations using de- ronmental improvements are possible in devel- composition analysis and efficient frontier oping countries and that peak levels of methods, discussed in Section 8. environmental degradation will be lower than Dasgupta et al. (2002) wrote a critical review in countries that developed earlier. of the EKC literature and other evidence on the According to Dasgupta et al. (2002), regula- relation between environmental quality and tion of pollution and enforcement increase with economic development in the Journal of Eco- income but the greatest increases happen from nomic Perspectives. Figure 3 illustrates four low to middle income levels and increased alternative viewpoints discussed in the article regulation is expected to have diminishing regarding the nature of the emissions and returns. There is also informal or decentralized income relation. The conventional EKC needs regulation in developing countries––Coasian no further discussion. Two viewpoints argue bargaining. Further, liberalization of develop- that the EKC is monotonic. The ‘‘new toxics’’ ing economies over the last two decades has scenario claims that while some traditional encouraged more efficient use of inputs and less pollutants might have an inverted U-shape subsidization of environmentally damaging curve, the new pollutants that are replacing activities––globalization is in fact good for the them do not. These include carcinogenic environment. The evidence seems to contradict chemicals, carbon dioxide, etc. As the older the ‘‘race to the bottom’’ scenario. Multi- pollutants are cleaned up, new ones emerge, so national companies respond to investor and that overall environmental impact is not consumer pressure in their home countries and reduced. The ‘‘race to the bottom’’ scenario raise standards in the countries in which they posits that emissions were reduced in developed invest. Further, better methods of regulating countries by outsourcing dirty production to pollution such as market instruments are hav- developing countries. These countries will find ing an impact even in developing countries. it harder to reduce emissions. But the pressure Better information on pollution is available, of globalization may also preclude further encouraging government to regulate and tightening of environmental regulation in empowering local communities. Those that developed countries and may even result in its argue that there is no regulatory capacity in loosening in the name of competitiveness. developing countries seem to be wrong. The revised EKC scenario does not reject the Much of the Dasgupta et al. evidence is from inverted U-shape curve but suggests that it is China. Other researchers of environmental and shifting downward and to the left over time due economic developments in China come to to technological change. But this argument is similar conclusions. Gallagher (2003) finds that already present in the 1992 World Development China is adopting European Union standards Report (IBRD, 1992). Dasgupta et al. also for pollution emissions from cars with an

Figure 3. Environmental Kuznets curve: alternative views. Source: Dasgupta et al. (2002) and Perman and Stern (2003). THE ENVIRONMENTAL KUZNETS CURVE 1429 approximately 8–10-year lag. Clearly, China’s the traditional stationary econometric model income per capita is far more than 10 years (Perman & Stern, 2003). Coondoo and Dinda behind that of Western Europe. Streets et al. (2002) test for Granger Causality between CO2 (2001), Zhang (2000), Jiang and McKibbin emissions and income in various individual (2002), and Wang and Wheeler (2003) all report countries and regions. As the data are differ- on substantial reductions of pollution intensi- enced to ensure stationarity, this test can only ties and levels in recent years. address short-run effects. The overall pattern that emerges is that causality runs from income to emissions or there is no significant relation- 7. ECONOMETRIC CRITIQUE ship in developing countries, while in developed OF THE EKC countries causality runs from emissions to income. This suggests that simultaneity is not Econometric criticisms of the EKC fall into important. four main categories: heteroskedasticity, simul- Stern and Common (2001) use three lines of taneity, omitted variables bias, and cointegra- evidence to suggest that the EKC is an inade- tion issues. quate model and that estimates of the EKC in Stern et al. (1996) raised the issue of heter- levels can suffer from significant omitted vari- oskedasticity that may be important in the ables bias: (a) Differences between the para- context of regressions of grouped data (see meters of the random-effects and fixed-effects Maddala, 1977). Schmalensee et al. (1998) models, tested using the Hausman test; (b) found that regression residuals from OLS were differences between the estimated coefficients in heteroskedastic with smaller residuals associ- different subsamples, and (c) tests for serial ated with countries with higher total GDP and correlation. Table 2 presents the key results population as predicted by Stern et al. (1996). from an EKC model estimated with data from Stern (2002) estimated a decomposition model 74 countries (in the World sample) over 1960– using feasible GLS. Adjusting for hetero- 90. For the non-OECD and World samples, the skedasticity in the estimation significantly Hausman test shows a significant difference in improved the goodness of fit of globally the parameter estimates for the random-effects aggregated fitted emissions to actual emissions. and fixed-effects model. This indicates that the Cole et al. (1997) and Holtz-Eakin and Sel- regressors––the level and square of the loga- den (1995) used Hausman tests for regressor rithm of income per capita––are correlated with exogeneity to directly address the simultaneity the country effects and time effects. As these issue. They found no evidence of simultaneity. effects model the mean effects of omitted vari- In any case, simultaneity bias is less serious in ables that vary across countries or across time, models involving integrated variables than in this indicates that the regressors are likely

Table 2. Stern and Common (2001) key results Region Model Levels First differences Turning Hausman Chow F -test q Turning Mean points test points income elasticity OECD FE $9,239 0.9109 $55,481 0.67 RE $9,181 0.3146 0.9070 (0.8545) Non-OECD FE $908,178 0.8507 $18,039 0.50 RE $344,689 14.1904 0.8574 (0.0008) World FE $101,166 10.6587 0.8569 $33,290 (0.0156) RE $54,199 10.7873 4.0256 0.8624 (0.0045) (0.0399) All turning points in real 1990 purchasing power parity US dollars. 1430 WORLD DEVELOPMENT correlated with omitted variables and the use emissions data, individual time dummies, regression coefficients are biased. 11 The OECD and a static specification. Stern and Common’s results pass this Hausman test but this result (2001) first differences results (Table 2) are very turned out to be very sensitive to the exact similar to the Harbaugh et al.’s (2002) results, sample of countries included in the subsample. which suggests that the dynamic specification As expected, given the Hausman test results, could be important. the parameter estimates are dependent on the Millimet, List, and Stengos (2003) use a dif- sample used, with the non-OECD estimates ferent strategy to test the robustness of the showing a turning point at extremely high- parametric EKC––comparing it to semi-para- income levels and the OECD estimates a within metric curves estimated using the same dataset sample turning point (Table 2). As mentioned for US states used by List and Gallet (1999). above, these results exactly parallel those for But they claim that parametric models are too developed and developing country samples of pessimistic––finding high turning points––while carbon emissions. The Chow F-test tests their alternative semi-parametric models result whether the two subsamples can be pooled, and in U-shaped curves with lower turning points. therefore that there is a common regression In addition, they reject the parametric specifi- parameter vector, a hypothesis that is rejected. cation in favor of the semi-parametric. But The parameter q is the first order autore- neither parametric nor semi-parametric curves gressive coefficient of the regression residuals. seem to fit the observed data very well in the This level of serial correlation indicates mis- figures presented in the paper. Furthermore, specification either in terms of omitted vari- results for individual states are varied, with the ables or missing dynamics. nitrogen dioxide curves mostly rising through- Harbaugh et al. (2002) carry out a sensitivity out the income range and many of the sulfur analysis of the original Grossman and Krueger dioxide curves falling––the reverse of the (1995) results. They use an updated and larger national panel data results. These results could, version of the ambient pollution data set and therefore, be further evidence of the fragility of test a number of alternative specifications. the EKC rather than evidence for a low turning Using the new extended dataset with Grossman point semi-parametric specification. and Krueger’s original cubic specification In contrast to Harbaugh et al. (2002), Cole results in the coefficients changing sign and (2003b) claims that the EKC model is fairly peak and trough levels altering wildly. Altering robust. But his basic levels sulfur emissions the specification in various ways––adding EKC has a significant Hausman statistic for a explanatory variables, using time dummies test of whether the random and fixed-effects instead of a time trend, using logs, removing parameters differ. Adding trade variables to the outliers, and averaging the observations across model results in an insignificant Hausman sta- monitors in each country––changes the shape tistic and a somewhat higher turning point. of the curve. The final experiment they carry Using logarithms increases this turning point out is to include only countries with GDP per further (Stern, 2004). The sulfur series cannot capita above $8,000. In contrast to Stern and be tested for unit roots, but other series in the Common (2001), this results in a monotonic dataset he uses do show unit root behavior and curve. The authors comment: results using first differences indicate a higher turning point than the levels results. I conclude that the model with trade variables performs This may seem counterintuitive. SO2 concentrations in Canada and the United States have declined over time better but the basic EKC is misspecified and at ever decreasing rates...the regressions...include... appropriate econometric techniques appear to a linear time trend... after detrending the data with raise the turning point. the time function, pollution appears to increase as a Perman and Stern (2003) test Stern and function of GDP (p. 548). Common’s (2001) data and models for unit roots and cointegration respectively. Panel unit There are several differences between the root tests indicate that all three series––log Harbaugh et al. (2002) model and the Stern and sulfur emissions per capita, log GDP capita, Common (2001) model that may explain the and its square––have stochastic trends. Results different results obtained for high income for cointegration are less clear cut. Around half countries. Harbaugh et al. (2002) use concen- the individual country EKC regressions coin- trations data, a linear time trend and a dynamic tegrate, but many of these have parameters specification, while Stern and Common (2001) with ‘‘incorrect signs.’’ Some panel cointegra- THE ENVIRONMENTAL KUZNETS CURVE 1431 tion tests indicate cointegration in all countries countries saw rising emissions in the 1960s and and some accept the noncointegration hypoth- declining emissions since 1973, ceteris paribus. esis. But even when cointegration is found, the Similarly, Lindmark (2002) uses the Kalman form of the EKC relationship varies radically filter to extract a technological change trend for across countries with many countries having U- carbon dioxide emissions in Sweden from shaped EKCs. A common cointegrating vector 1870–1997. This trend had a positive growth for all countries is strongly rejected. Koop and rate until about 1970 and a negative growth Tole (1999) similarly found that random and rate since. 12 fixed-effects specifications of a deforestation Day and Grafton (2003) test for cointegra- EKC were strongly rejected in favor of a tion of the EKC relation using Canadian time- random coefficients model with widely varying series data on a number of pollutants using the coefficients and insignificant mean coefficients. Engle-Granger and Johansen methods. They In the presence of possible noncointegration, fail to reject the noncointegration hypothesis in we can estimate a model in first differences. The almost every case. de Bruyn’s (2000) time-series estimated turning points indicate a largely Engle-Granger tests for the Netherlands, the monotonic EKC relationship and are more United Kingdom, the United States, and West similar across subsamples, though the para- Germany for SOx,NOx and CO2 finds cointe- meters are still significantly different (Table 2). gration for the CO2 EKC in the Netherlands The estimated income elasticity is less than and West Germany, but not in any other case. one––there are factors that change with income Using various lines of evidence, the majority which offset the scale effect, but they are of studies have found the EKC to be a fragile insufficiently powerful to overcome fully the model suffering from severe econometric mis- scale effect. specification. Use of more appropriate methods Figure 4 presents the time effects from the tends to indicate higher turning points and first difference estimates. The OECD saw possibly a monotonic curve for emissions of declining emissions holding income constant major pollutants. A better model may result over the entire period, though the introduction from including additional variables to represent of the LRTAP agreement in the mid-1980s in either proximate or underlying causes of change Europe resulted in a larger decline. Developing in emissions. I next turn to a consideration of

0.3

World OECD 0.2 Non-OECD

0.1

0

-0.1

-0.2

-0.3 1960 1965 1970 1975 1980 1985 1990

Figure 4. Time effects: first differences sulfur EKC. Source: Stern and Common (2001). 1432 WORLD DEVELOPMENT

a new literature that goes beyond the EKC in tions of CO2 emissions do not however explic- this way. itly include fuel mix or output structure and so do not require industry level data. This decomposition is given by: 8. DECOMPOSING EMISSIONS AND MEASURING ENVIRONMENTAL Et FECt TECt GDPt Et Pt; ð4Þ EFFICIENCY FECt TECt GDPt Pt

Two alternative approaches are emerging where E is emissions, FEC and TEC are fossil that go beyond the EKC by analyzing the fuel and total energy consumption respectively, proximate factors driving changes in pollution and P is population. As carbon abatement emissions described in Section 2: index number technologies do not yet exist, the first term on decompositions of emissions and production the RHS reflects the impact of shifts in the mix frontiers estimated using linear programming of fossil fuel types. The second term reflects or econometrics. The index numbers decom- shifts between fossil and nonfossil fuels, while position approach requires detailed sectoral the remaining terms are energy intensity and information on fuel use, production, emissions two components of the scale effect. etc. The average effects of the different factors De Bruyn (1997), Viguier (1999), Selden, can be derived from the results for the indi- Forrest, and Lockhart (1999), and Bruvoll and vidual countries. No stochasticity is allowed Medin (2003) all calculate versions of (3) for a for. By contrast, the frontier models do not variety of European countries and the United usually require data on fuel use by sector; they States. A progressively larger number of pol- can estimate factors common to all countries as lutants have been investigated. de Bruyn (1997) well as idiosyncratic components for each only analyzes sulfur, while Bruvoll and Medin country, and the econometric versions allow for (2003) cover 10 pollutants. Output structure random error. On the other hand, the frontier effects are mostly small except for some of the models usually make stronger assumptions additional pollutants investigated by Bruvoll about production relations. There are also and Medin (2003) (CH4,N2O, and NH3). Fuel studies that are hybrids of the two approaches mix effects are varied depending on the pollu- (e.g., Hettige, Mani, & Wheeler, 2000; Hilton & tants and countries investigated––in some cases 13 Levinson, 1998; Judson et al., 1999). Gross- they act to increase emissions (for SOx,NOx, man (1995) and de Bruyn (1997) proposed the and CO2 in the US) and in others to reduce following decomposition: emissions (SOx and CO2 in Norway). In all cases, technique effects were dominant in off- Xn E Y I S ; ð3Þ setting the increase in scale. Energy intensity is it it ijt ijt most important for carbon and nitrogen but for j¼1 both these and sulfur it is insufficient to over- where Eit is emissions in country i in year t, Y is come scale. Therefore, for sulfur, which has GDP, Ij is the emissions intensity of sector j, declined substantially in these countries, emis- and Sj is the share of that sector in GDP. This sions specific technological change accounts for decomposition, therefore, attributes emissions the majority of the decline. For example, de to what Grossman calls the scale, composition Bruyn (1997) found a 55–60% reduction in (output mix), and technique effects. The latter sulfur emissions due to this factor. Similarly, includes the effects of both fuel mix and Hamilton and Turton (2002) find that the main ‘‘technological change’’ and the latter can be factor increasing carbon emissions in the further broken down into general productivity OECD during 1982–97 was income per capita improvements, where more output is derived (37%) and the second most important was from a unit of input, and emissions reducing population growth (12%). The main factor technological change, where less emissions are reducing emissions was energy intensity. Zhang produced per unit of input. A problem for the (2000) finds that the decline in energy intensity index number studies is that, usually, fuel use in China almost halved the increase in emis- data are collected on a different sectoral basis sions that would otherwise have occurred. than output. This makes the standard type of Other factors had minor effects. index number study impossible to implement in All these studies treat energy intensity as an most countries. Hamilton and Turton’s (2002) exogenous variable, whereas actually it is also and Zhang’s (2000) index number decomposi- partly determined by shifts between fuels of THE ENVIRONMENTAL KUZNETS CURVE 1433 different qualities and industries with different opment. Emissions rise up to around $7,000 energy intensities. Kander’s (2002) study of per capita and then are fairly constant at carbon emissions in Sweden over the two- higher income levels. Composition effects work century period from 1800 to 1998 also together with scale at lower income levels to decomposes changes in energy intensity. increase this form of pollution. At higher Structural change largely explains the increase income levels they reduce pollution, but toge- in energy intensity that occurred over 1870– ther with the technique effect, which acts 1913 and contributed to the reduction in against the scale effect at every level of energy intensity during 1913–70. In the latter income, they only just offset the effects of ris- period, rising energy quality may have also ing scale. contributed. But technical change dominated Bruvoll, Fæhn, and Strøm (2003) combine in that period and was the only important index number decomposition with a CGE factor in reducing energy intensity before 1870 model which allows decomposition into the and after 1970. Once energy intensity is proximate factors of scale, composition, and explained, carbon per unit energy is explained technique, as well as an underlying policy fac- by shifts in the fuel mix. In the 19th century tor which acts through the first three factors. coal use replaced wood and muscle power, Their study is, however, a policy simulation increasing the carbon intensity, coal was then rather than an empirical analysis. The endoge- replaced by oil, and then by electricity (nuclear nous policy simulation they run over 2000– and hydro) to some extent. In the last two 2030 uses a carbon tax to target a level of decades however, coal and wood have staged a carbon emissions predicted by a cubic loga- comeback again raising the carbon emitted per rithmic EKC. The model endogenously com- unit of energy used. putes the changes in trade over the forecast Hilton and Levinson (1998) use a hybrid horizon and therefore the changes in emissions approach. Data are available on both the total associated with imports and exports. They consumption of gasoline and the lead content estimate that emissions leakage would increase of gasoline for a large number of countries for throughout the forecast horizon. specific years. Hence, decomposition into scale Koop (1998) and Zaim and Taskin (2000) and technical change effects is easy in this estimate global econometric production fron- special case and the two structural effects are tiers for carbon and Stern (2002) for sulfur. Of absent. They carry out an EKC type analysis these three studies, however only Stern (2002) of these variables. Per capita gasoline use rises actually derives the decomposition into proxi- strongly with income. In 1992, lead content mate factors. These studies are part of an per gallon of gasoline was a declining function emerging literature on environmental efficiency. of income. There is however a wide scatter in A frontier representing the best-practice tech- developing countries with many low and nology is constructed using either data envel- middle-income countries having low lead con- opment analysis (e.g., Lansink & Silva, 2003) or tents. Before 1983, there is no evidence of an estimated econometrically (e.g., Fernandez, EKC type relation in the data.The inference is Koop, & Steel, 2002; Reinhard, Lovell, & that there was a technological innovation that Thijssen, 1999) using data on inputs, outputs, was preferentially adopted in high-income and pollution emissions for a group of coun- countries. As Gallagher (2003) suggests, these tries or firms. 14 This approach allows some innovations may be adopted with a relatively countries to be on the production frontier that short lag in developing countries. Hettige et al. represents best practice technology and other (2000) use a similar approach to model indus- countries to be behind the frontier using a trial BOD emissions in a range of devel- technology that is less efficient than the best oping and developed countries. The share of practice. Relative movement of the frontier in manufacturing industry in national income different directions measures emissions specific and the share of polluting industries within technical change and changes in general total total manufacturing represent composition factor productivity in the best practice tech- effects and actual plant-level end-of-pipe nology. Movement of countries relative to the BOD emissions per unit output represent frontier reflects the degree to which they adopt technique effects. Each component is modeled the best practice technology. The impacts of as a function of income and other variables structural shifts in inputs and outputs and the and then the components are reassembled to scale effect are defined by the estimated pro- predict emissions at different levels of devel- duction technology. 1434 WORLD DEVELOPMENT

Stern (2002) uses the following econometric representation of the data. A nested test of this model to decompose sulfur emissions in 64 model and the EKC showed that the income countries during 1973–90: squared term in the EKC added no explanatory power to that provided by the decomposition S Y E YJ y aj XK e it ¼ c it A it jit kit e ; ð5Þ model. P i P t Y Y E it it it it j¼1 it k¼1 it In an attempt to determine the effects of trade on scale, composition, and technique where S is sulfur emissions and P population effects, Antweiler et al. (2001) estimate a dif- and the RHS decomposes per capita emissions ferent type of econometric decomposition into the following five effects: model that derives a reduced form equation Y from a theoretical structural model of the it scale––GDP per capita. P demand and supply of pollution. The technique it effect is however assumed to be induced by the At a common global time effect represent- ing the effects of emissions specific increase in income due to trade. This model, technical progress over years t. therefore, takes the EKC hypothesis as a given. Eit The composition effect is expected to differ in energy intensity––the effect of general Y capital intensive and labor intensive economies. it productivity on emissions. y y Trade is likely to increase pollution in the 1it ; ...; Jit output mix––shares of the output of former and reduce it in the latter. Therefore, YJit Yit different industries y in total GDP Y . the capital/labor ratio is controlled for. The e e model is estimated using the GEMS sulfur it ; ...; Kit input mix––shares of different energy Eit Eit dioxide concentration data. Rather than a sources e in total energy use E. decomposition of changes in emissions they compute elasticities with respect to the different

Additionally, ci is the relative efficiency of factors. The ‘‘scale elasticity’’ is estimated to country i compared to best practice, and eit is a average 0.266. The sample mean of the tech- random error term. The contributions of the nique elasticity (elasticity of concentrations five effects to changes in global emissions at the w.r.t. GNP per capita) is )1.15. The composi- global level are given in Table 3. Input and tion elasticity (elasticity of concentrations w.r.t. output effects contributed little globally, capital/labor ratio) is 1.01 and trade intensity though in individual countries they can have has an elasticity of )0.864. Combining the important effects. At the global level the two effects, trade has a negative impact on emis- forms of technological change reduced the sions. Cole and Elliott (2003) extend the anal- increase in emissions to half of what it would ysis to other pollutants and to emissions. The have been in their absence with emissions spe- results are more mixed. cific technological change lowering aggregate The conclusion from all these studies is that emissions by around 20%. The residuals from the main means by which emissions of pollu- the model show it to be a statistically adequate tants can be reduced is by time related tech- nique effects and in particular those directed specifically at emissions reduction, though Table 3. Contributions to total change in global sulfur emissions productivity growth or declining energy inten- sity has a role to play. Though structural Weighted logarithmic change and shifts in fuel composition may be percent change (%) important in some countries at some times, Total change their average contribution seems less important Actual emissions 28.77 quantitatively. Those studies that include Predicted emissions 27.37 developing countries (e.g., Antweiler et al., Unexplained fraction 1.40 2001; Judson et al., 1999; Stern, 2002) find that these technological changes are occurring in Decomposition both developing and developed countries. Scale effect 53.78 Innovations may first be adopted preferentially Emissions related tech- )19.86 in higher income countries (Hilton & Levinson, nical change 1998) but seem to be adopted in developing ) Energy intensity 10.20 countries with relatively short lags (Gallagher, Output mix 3.77 2003). This result is in line with the evidence of ) Input mix 0.13 Dasgupta et al. (2002) and the EKC-based THE ENVIRONMENTAL KUZNETS CURVE 1435 estimates of time effects in Stern and Common first with a short lag before it is adopted in the (2001) and Stern (2002). majority of poorer countries. However, emis- sions may be declining simultaneously in low- and high-income countries over time, ceteris 9. CONCLUSIONS paribus, though the particular innovations typically adopted at any one time could be The evidence presented in this paper shows different in different countries. that the statistical analysis on which the envi- It seems that structural factors on both the ronmental Kuznets curve is based is not robust. input and output side do play a role in modi- There is little evidence for a common inverted fying the gross scale effect though they are U-shaped pathway that countries follow as mostly less influential than time-related effects. their income rises. There may be an inverted U- The income elasticity of emissions is likely to be shaped relation between urban ambient con- less than one––but not negative in wealthy centrations of some pollutants and income countries as proposed by the EKC hypothesis. though this should be tested with more rigorous In slower growing economies, emissions- time-series or panel data methods. It seems reducing technological change can overcome unlikely that the EKC is an adequate model of the scale effect of rising income per capita on emissions or concentrations. I concur with emissions. As a result, substantial reductions in Copeland and Taylor (2004), who state that: sulfur emissions per capita have been observed ‘‘Our review of both the theoretical and in many OECD countries in the last few empirical work on the EKC leads us to be decades. In faster growing middle income skeptical about the existence of a simple and economies, the effects of rising income over- predictable relationship between pollution and whelmed the contribution of technological per capita income.’’ change in reducing emissions. The true form of the emissions–income rela- The research challenge now is to revisit some tionship is likely a mix of two of the scenarios of the issues addressed earlier in the EKC lit- proposed by Dasgupta et al. (2002) illustrated erature using the new decomposition and in Figure 3. The overall shape is that of their frontier models and rigorous panel data and ‘‘new toxics’’ EKC––a monotonic increase of time-series statistics. For example, how can the emissions in income. But over time this curve effects of trade on emissions be modeled in the shifts down, which is analogous to their context of the decomposition and econometric ‘‘revised EKC’’ scenario. Some evidence shows frontier models? 15 Rigorous answers to such that a particular innovation is likely to be questions are central to the debate on global- adopted preferentially in high-income countries ization and the environment.

NOTES

1. The increasing number of exceptions include Cole relative to the prices of services and therefore manufac- (2003b), Coondoo and Dinda (2002), Day and Grafton turing’s share of GDP declines when measured at (2003), de Bruyn (2000), Stern and Common (2001), current prices but not when measured at constant prices. Perman and Stern (2003), Friedl and Getzner (2003), Due to this productivity growth in manufacturing, its and Heil and Selden (1999, 2001). pollution intensity falls over time relative to the pollu- tion intensity of services. 2. Exceptions include Stern and Common (2001) and Magnani (2001). 5. Grossman (1995) calls the combination of 3 and 4 the ‘‘technique effect.’’ 3. For example: Dasgupta et al. (2002), Harbaugh, Levinson, and Wilson (2002), Perman and Stern (2003), 6. Later published as Grossman and Krueger (1994). and Koop and Tole (1999).

4. Kander (2002) argues that structural shift in the 7. Later published as Shafik (1994). economy may largely be an illusion. Due to rising productivity in manufacturing, manufacturing prices fall 8. Also see Ezzati, Singer, and Kammen (2001). 1436 WORLD DEVELOPMENT

9. Solid waste does not necessarily result in ‘‘pollu- are flat to declining in industry and construction. tion,’’ especially as techniques to reduce seepage and Technical innovations tend to introduce more energy emissions from landfills and increase recycling have using appliances to households and energy saving progressed. But landfills still disturb the landscape and techniques to industry. geology and recycling and waste processing require additional energy and resource use. 13. Judson et al. (1999) decompose emissions by sector and then fit EKC models––see the previous endnote. 10. See the discussion of energy EKCs above. 14. Lansink and Silva (2003), Reinhard et al. (1999), 11. In an alternative test for omitted variables, Mag- and Fernandez et al. (2002) all look at agricultural firms nani (2001) uses the Ramsey test on cross-sectional rather than countries. EKCs and finds that the null of no omitted variables is rejected in almost every case. 15. Bruvoll et al. (2003) provide a possible method for a single country model. But they only investigate 12. Judson, Schmalensee, and Stoker (1999) estimate emissions directly associated with imports and exports separate EKC relations for energy consumption in each and not the effects of trade on the overall pollution of a number of energy-consuming sectors for a large intensity of the economy. Antweiler et al. (2001) and panel data set using spline regression, which allows them Cole and Elliott (2003) provide another approach to estimate different time effects in each sector. These with different limitations––for example, the technique vary substantially energy consumption rises over time, effect is assumed to be induced by an EKC like income ceteris paribus, in the household and other sectors but effect.

REFERENCES

Andreoni, J., & Levinson, A. (2001). The simple and the (reverse) environmental Kuznets Curve. analytics of the environmental Kuznets curve. Jour- Journal of Development Economics, 65, 225–235. nal of Public Economics, 80, 269–286. Cole, M. A. (2003a). Environmental optimists, environ- Ansuategi, A., & Perrings, C. A. (2000). Transboundary mental pessimists and the real state of the world––an externalities in the environmental transition hypoth- article examining The Skeptical Environmentalist: esis. Environmental and Resource Economics, 17, 353– Measuring the Real State of the World by Bjorn 373. Lomborg. Economic Journal, 113, 362–380. Ansuategi, A., Barbier, E. B., & Perrings, C. A. (1998). Cole, M. A. (2003b). Development, trade, and the The environmental Kuznets curve. In J. C. J. M. van environment: How robust is the environmental den Bergh & M. W. Hofkes (Eds.), Theory and Kuznets curve? Environment and Development Eco- implementation of economic models for sustainable nomics, 8, 557–580. development. Dordrecht: Kluwer. Cole, M. A., Elliott, R. J., & R (2003). Determining the Antweiler, W., Copeland, B. R., & Taylor, M. S. (2001). trade–environment composition effect: The role of Is free trade good for the environment. American capital, labor and environmental regulations. Journal Economic Review, 91, 877–908. of Environmental Economics and Management, 46, Arrow, K., Bolin, B., Costanza, R., Dasgupta, P., Folke, 363–383. C., Holling, C. S., Jansson, B.-O., Levin, S., Maler,€ Cole, M. A., Rayner, A. J., & Bates, J. M. (1997). The K.-G., Perrings, C. A., & Pimentel, D. (1995). environmental Kuznets curve: An empirical analysis. Economic growth, carrying capacity, and the envi- Environment and Development Economics, 2(4), 401– ronment. Science, 268, 520–521. 416. Beckerman, W. (1992). Economic growth and the Coondoo, D., & Dinda, S. (2002). Causality between environment: Whose growth? Whose environment? income and emission: A country group-specific World Development, 20, 481–496. econometric analysis. Ecological Economics, 40, Bruvoll, A., & Medin, H. (2003). Factors behind the 351–367. environmental Kuznets curve: A decomposition of Copeland, B. R., & Taylor, M. S. (2004). Trade, growth the changes in air pollution. Environmental and and the environment. Journal of Economic Litera- Resource Economics, 24, 27–48. ture, 42, 7–71. Bruvoll, A., Fæhn, T., & Strøm, B. (2003). Quantifying Dasgupta, S., Laplante, B., Wang, H., & Wheeler, D. central hypotheses on environmental kuznets curves (2002). Confronting the environmental Kuznets for a rich economy: A computable general equilib- curve. Journal of Economic Perspectives, 16, 147–168. rium study. Scottish Journal of Political Economy, Day, K. M., & Grafton, R. Q. (2003). Growth and the 50(2), 149–173. environment in Canada: An empirical analysis. Bulte, E. H., & van Soest, D. P. (2001). Environmental Canadian Journal of Agricultural Economics, 51, degradation in developing countries: Households 197–216. THE ENVIRONMENTAL KUZNETS CURVE 1437 de Bruyn, S. M. (1997). Explaining the environmental Heil, M. T., & Selden, T. M. (1999). Panel stationarity Kuznets curve: Structural change and international with structural breaks: Carbon emissions and GDP. agreements in reducing sulphur emissions. Environ- Applied Economics Letters, 6, 223–225. ment and Development Economics, 2, 485–503. Heil, M. T., & Selden, T. M. (2001). Carbon emissions de Bruyn, S. M. (2000). Economic growth and the and economic development: Future trajectories environment: An empirical analysis. Dordrecht: Klu- based on historical experience. Environment and wer Academic Press. Development Economics, 6, 63–83. Dijkgraaf, E., & Vollebergh, H. R. J. (1998). Growth Hettige, H., Lucas, R. E. B., & Wheeler, D. (1992). The and/or (?) environment: Is there a Kuznets curve for toxic intensity of industrial production: Global carbon emissions? Paper presented at the 2nd bien- patterns, trends, and trade policy. American Eco- nial meeting of the European Society for Ecological nomic Review, 82(2), 478–481. Economics, Geneva, March 4–7th. Hettige, H., Mani, M., & Wheeler, D. (2000). Industrial Ekins, P. (1997). The Kuznets curve for the environment pollution in economic development: The environ- and economic growth: Examining the evidence. mental Kuznets curve revisited. Journal of Develop- Environment and Planning A, 29, 805–830. ment Economics, 62, 445–476. Ezzati, M., Singer, B. H., & Kammen, D. M. (2001). Hilton, F. G. H., & Levinson, A. M. (1998). Factoring Towards an integrated framework for development the environmental Kuznets curve: Evidence from and environmental policy: The dynamics of environ- automotive lead emissions. Journal of Environmental mental Kuznets curves. World Development, 29, Economics and Management, 35, 126–141. 1421–1434. Holtz-Eakin, D., & Selden, T. M. (1995). Stoking the Fernandez, C., Koop, G., & Steel, M. F. J. (2002). fires. CO2 emissions and economic growth. Journal Multiple-output production with undesirable out- of Public Economics, 57, 85–101. puts: An application to nitrogen surplus in agricul- Hsiao, C. (1986). Analysis of Panel Data. Cambridge: ture. Journal of the American Statistical Association, Cambridge University Press. 97(458), 432–442. IBRD (1992). World Development Report 1992. Devel- Friedl, B., & Getzner, M. (2003). Determinants of CO2 opment and the Environment. New York: Oxford emissions in a small open economy. Ecological University Press. Economics, 45, 133–148. Im, K. S., Pesaran, M. H., & Shin, Y. (2003). Testing for Gallagher, K. S. (2003). Development of cleaner vehicle unit roots in heterogeneous panels. Journal of technology? Foreign direct investment and technol- Econometrics, 115, 53–74. ogy transfer from the United States to China, Paper Jaffe, A. B., Peterson, S. R., Portney, P. R., & Stavins, presented at United States Society for Ecological R. N. (1995). Environmental regulation and the Economics 2nd Biennial Meeting, Saratoga Springs competitiveness of US manufacturing: What does the NY, May. evidence tell us? Journal of Economic Literature, 33, Galli, R. (1998). The relationship between energy 132–163. intensity and income levels: Forecasting log-term Jiang, T., & McKibbin, W. J. (2002). Assessment of energy demand in Asian emerging countries. Energy China’s pollution levy system: An equilibrium pol- Journal, 19(4), 85–105. lution approach. Environment and Development Eco- Grossman, G. M. (1995). Pollution and growth: What nomics, 7, 75–105. do we know?. In I. Goldin & L. A. Winters (Eds.), John, A., & Pecchenino, R. (1994). An overlapping The Economics of Sustainable Development (pp. 19– generations model of growth and the environment. 47). Cambridge: Cambridge University Press. Economic Journal, 104, 1393–1410. Grossman, G. M., & Krueger, A. B. (1991). Environ- John, A., Pecchenino, R., Schimmelpfennig, D., & mental impacts of a North American Free Trade Schreft, S. (1995). Short-lived agents and the long- Agreement. National Bureau of Economic Research lived environment. Journal of Public Economics, 58, Working Paper 3914, NBER, Cambridge MA. 127–141. Grossman, G. M., & Krueger, A. B. (1994). Environ- Judson, R. A., Schmalensee, R., & Stoker, T. M. (1999). mental impacts of a North American Free Trade Economic development and the structure of demand Agreement. In P. Garber (Ed.), The US–Mexico Free for commercial energy. The Energy Journal, 20(2), Trade Agreement. Cambridge MA: MIT Press. 29–57. Grossman, G. M., & Krueger, A. B. (1995). Economic Kander, A. (2002). Economic growth, energy consump- growth and the environment. Quarterly Journal of tion and CO2 emissions in Sweden 1800–2000, Lund Economics, 110, 353–377. Studies in Economic History, 19, University of Hamilton, C., & Turton, H. (2002). Determinants of Lund, Lund. emissions growth in OECD countries. Energy Policy, Kaufmann, R. K., Davidsdottir, B., Garnham, S., & 30, 63–71. Pauly, P. (1998). The determinants of atmospheric Harbaugh, W., Levinson, A., & Wilson, D. M. (2002). SO2 concentrations: Reconsidering the environmen- Reexamining the empirical evidence for an environ- tal Kuznets curve. Ecological Economics, 25, 209– mental Kuznets curve. Review of Economics and 220. Statistics, 84, 541–551. Koop, G. (1998). Carbon dioxide emissions and eco- Hausman, J. A. (1978). Specification tests in economet- nomic growth: A structural approach. Journal of rics. Econometrica, 46, 1251–1271. Applied Statistics, 25, 489–515. 1438 WORLD DEVELOPMENT

Koop, G., & Tole, L. (1999). Is there an environmental Pearson, P. J. G. (1994). Energy, externalities, and Kuznets curve for deforestation? Journal of Devel- environmental quality: Will development cure the opment Economics, 58, 231–244. ills it creates. Energy Studies Review, 6, 199–216. Kuznets, S. (1955). Economic growth and income Perman, R., & Stern, D. I. (2003). Evidence from panel inequality. American Economic Review, 49, 1–28. unit root and cointegration tests that the environ- Lansink, A. O., & Silva, E. (2003). CO2 and energy mental Kuznets curve does not exist. Australian efficiency of different heating technologies in the Journal of Agricultural and Resource Economics, 47, Dutch glasshouse industry. Environmental and Re- 325–347. source Economics, 24(4), 395–407. Rees, J. A. (1990). Natural Resources: Allocation, Lefohn, A. S., Husar, J. D., & Husar, R. B. (1999). Economics and Policy (2nd edn.). London: Routledge Estimating historical anthropogenic global sulfur and Kegan Paul. emission patterns for the period 1850–1990. Atmo- Reinhard, S., Lovell, C. A. K., & Thijssen, G. (1999). spheric Environment, 33, 3435–3444. Econometric estimation of technical and environ- Lieb, C. M. (2001). The environmental Kuznets curve mental efficiency: An application to Dutch dairy and satiation: A simple static model. Environment farms. American Journal of Agricultural Economics, and Development Economics, 7, 429–448. 81(1), 44–60. Lindmark, M. (2002). An EKC-pattern in historical Schmalensee, R., Stoker, T. M., & Judson, R. A. (1998). perspective: Carbon dioxide emissions, technology, World carbon dioxide emissions: 1950–2050. Review fuel prices and growth in Sweden 1870–1997. Eco- of Economics and Statistics, 80, 15–27. logical Economics, 42(1–2), 333–347. Selden, T. M., & Song, D. (1994). Environmental quality List, J. A., & Gallet, C. A. (1999). The environmental and development: Is there a Kuznets curve for air Kuznets curve: Does one size fit all? Ecological pollution? Journal of Environmental Economics and Economics, 31, 409–424. Management, 27, 147–162. Lomborg, B. (2001). The skeptical environmentalist: Selden, T. M., & Song, D. (1995). Neoclassical growth, Measuring the real state of the world. Cambridge: the J curve for abatement and the inverted U curve Cambridge University Press. for pollution. Journal of Environmental Economics Lopez, R. (1994). The environment as a factor of and Management, 29, 162–168. production: The effects of economic growth and Selden, T. M., Forrest, A. S., & Lockhart, J. E. (1999). trade liberalization. Journal of Environmental Eco- Analyzing reductions in US air pollution emissions: nomics and Management, 27, 163–184. 1970 to 1990. Land Economics, 75, 1–21. Lucas, R. E. B., Wheeler, D., & Hettige, H. (1992). Shafik, N. (1994). Economic development and environ- Economic development, environmental regulation mental quality: An econometric analysis. Oxford and the international migration of toxic industrial Economic Papers, 46, 757–773. pollution: 1960–1988. In P. Low (Ed.), International Shafik, N., & Bandyopadhyay, S. (1992). Economic Trade and the Environment. World Bank Discussion Growth and Environmental Quality: Time Series and Paper No. 159, World Bank, Washington, DC. Crosscountry Evidence. Background Paper for the Maddala, G. S. (1977). Econometrics. Singapore: World Development Report 1992, The World Bank, McGraw Hill. Washington, DC. Magnani, E. (2001). The environmental Kuznets curve: Stern, D. I. (1998). Progress on the environmental Development path or policy result? Environmental Kuznets curve? Environment and Development Eco- Modelling and Software, 16, 157–166. nomics, 3, 173–196. Martinez-Alier, J. (1995). The environment as a luxury Stern, D. I. (2002). Explaining changes in global sulfur good or ‘too poor to be green’. Ecological Economics, emissions: An econometric decomposition approach. 13, 1–10. Ecological Economics, 42, 201–220. McConnell, K. E. (1997). Income and the demand for Stern, D. I. (2004). Comment on: Cole M. A. (2003) environmental quality. Environment and Development Development, trade, and the environment: how Economics, 2, 383–399. robust is the environmental Kuznets curve? Environ- Meadows, D. H., Meadows, D. L., Randers, J., & ment and Development Economics, 8, 557–580. Behrens, W. (1972). The Limits to Growth. New Stern, D. I., & Common, M. S. (2001). Is there an envi- York: Universe Books. ronmental Kuznets curve for sulfur? Journal of Envi- Millimet, D. L., List, J. A., & Stengos, T. (2003). The ronmental Economics and Management, 41, 162–178. environmental Kuznets curve: Real progress or Stern, D. I., Common, M. S., & Barbier, E. B. (1996). misspecified models? Review of Economics and Sta- Economic growth and environmental degradation: tistics, 85(4), 1038–1047. The environmental Kuznets curve and sustainable Panayotou, T. (1993). Empirical tests and policy analysis development. World Development, 24, 1151–1160. of environmental degradation at different stages of Stokey, N. L. (1998). Are there limits to growth? economic development. Working Paper WP238, Tech- International Economic Review, 39(1), 1–31. nology and Employment Programme, International Streets, D. G., Jiang, K., Hu, X., Sinton, J. E., Zhang, Labour Office, Geneva. X.-Q., Xu, S., Jacobson, M. Z., & Hansen, J. E. Panayotou, T. (1997). Demystifying the environmental (2001). Recent reductions in China’s greenhouse gas Kuznets curve: Turning a black box into a policy emissions. Science, 294, 1835–1837. tool. Environment and Development Economics, 2, Suri, V., & Chapman, D. (1998). Economic growth, 465–484. trade and the energy: Implications for the environ- THE ENVIRONMENTAL KUZNETS CURVE 1439

mental Kuznets curve. Ecological Economics, 25, World Commission on Environment and Development 195–208. (1987). Our common future. Oxford: Oxford Univer- Torras, M., & Boyce, J. K. (1998). Income, inequality, sity Press. and pollution: A reassessment of the environmental Zaim, O., & Taskin, F. (2000). Environmental efficiency Kuznets curve. Ecological Economics, 25, 147–160. in carbon dioxide emissions in the OECD: A non- Viguier, L. (1999). Emissions of SO2,NOx, and CO2 in parametric approach. Journal of Environmental Man- transition economies: Emission inventories and Div- agement, 58, 95–107. isia index analysis. Energy Journal, 20(2), 59–87. Zhang, Z. (2000). Decoupling China’s carbon emissions Wang, H., & Wheeler, D. (2003). Equilibrium pollution increase from economic growth: An economic anal- and economic development in China. Environment ysis and policy implications. World Development, 28, and Development Economics, 8, 451–466. 739–752.