S.E.K.C research: The quest for a spatial
environmental Kuznets curve.
Does economic growth improve the environment, considering overall environmental indicators and spatial interaction effects?
Abstract. Many studies investigating the environmental Kuznets curve (EKC) considered single
pollutants and have not accounted for spatial interaction effects. This research examines the relationship between economic development and ‗overall‘ measures of environmental degradation, adopting a spatial econometric approach. Results show that no EKC can be identified, but indicates an N-shaped relationship between per capita income and impact on the environment. Moreover, evidence of spatial interaction effects supports the estimation of the spatial Durbin model. It is found that an increase in either environmental degradation or per capita income in country i leads to respectively an increase and decrease in environmental degradation in neighbouring countries j ≠ i.
JEL classifications: C21, O13, Q01, Q56, R11
Key words: Environmental Kuznets curve, spatial interaction effects, GDP, environmental degradation
Master’s Thesis Economics – Final version, February 21, 2012 Bieuwe Pruiksma – 1421220
1. Introduction
Since the human species have been around on planet earth, they have interacted with their surroundings. Whereas the dependency of humanity on natural resources has been large for all times, the human impact on her environment has increased over time. The relationship between economic development and the state of the environment has been debated for many years. Although humans have been curious about the earth and its physical properties since at least the times of the ancient Greeks, the first known examples of this debate appeared in the eighteenth and nineteenth century. Malthus (1798) expressed concern about humanity‘s ability to feed itself given exponential population growth and linear increase of food production. Mill (1848) recognized wealth beyond the material and argued that the logical conclusion of unlimited growth is destruction of the environment and a reduced quality of life. Finally, Jevons (1865) argued that the limited amount of non-renewable resources, like coal as he considered the gradual exhaustion of UK‘s coal supplies, was the most important constraint on economic growth in the industrialized countries.
While the World experienced high population and economic growth rates since the industrial revolution, concerns about the finiteness of natural resources and potentially infinite resource demand remained. A growing economy is characterized by higher material and energy throughput. This leads to natural resources depletion because it requires more input, and causes more pollution like emissions and waste. One of the most famous publications addressing this issue is Limits to Growth by Meadows et al (1972), stating that unbounded growth in a finite world is physically impossible. The last decade experienced an increasing concern in broader and more complex environmental problems like global warming, climate change, ecosystem collapse and resource dependency. The relevance of these issues is by now recognized by more or less all experts, governments and international institutions throughout the world, although the public opinion remains divided
(IPCC, 2007)1. In their 2009 report Growing within Limits (PBL, 2009), the Dutch
Environmental Assessment Agency refers to Turner (2008). Turner (2008) finds that data2 of
the years 1970-2000 matches with the standard scenario of Limits to Growth, which in the
end results in a global collapse in the middle of the 21st century. Some scientists state that
the human element dominates her environment and has become the main cause of change, disruption and deterioration in the environment. In this respect, Crutzen (2002) argues that nowadays man is a geological force of importance and that we live in the ‗anthropocene‘.
1 See also 1972, 1992 and 2002 UN Conferences on Environment, and results of United Nations Framework
Convention on Climate Change (UNFCCC) like 1997 Kyoto Protocol and 2009 Copenhagen Accord.
2 Regarding changes in industrial production, food production, pollution, population growth and depletion of
On the other side, rather than being a threat, economic growth could be the solution to environmental improvement. This is also reflected by the concept of sustainable development as advanced by the World Commission report Our Common Future (Brundtland, 1987). Moreover, some argue that doom stories about finite natural resources like fossil fuels and certain materials are exaggerated as still new sources are found, and are skeptical about the severeness of environmental problems (Ljomborg, 2001). Besides, there was and is a strong belief in the potential of technological advancement and the market mechanism to help the environment cope with continued economic growth (Ruttan, 1971). Nordhaus (1992) stated that for the past two centuries technological change has been the clear victor in the race with resource depletion and contributed significantly to the growth of the world economy. Notable examples are the green revolution in the 1960s leading to higher agricultural yields, the digital revolution starting in the 1980s and nowadays nanotechnology and the potential of various renewable energy forms.
Apparently, the growth versus environmental quality debate is from all times. However, in the 1990s the discussion about the linkage of the state of the environment with
economic development intensified (Dinda, 2004)3. Exemplary is the construction of
alternative measures of economic development, accounting for environmental impact like
green GDP, Genuine Progress Indicator, etc4. The relationship between economic
development and the health of the environment is undoubtedly complex and also known as the pollution-income relationship (PIR). An explanation is the Environmental Kuznets Curve (EKC) hypothesis, which postulates that there exists an inverted U-relationship between economic activity and environmental burden, see figure 1 below:
Figure 1. The Environmental Kuznets Curve (EKC)5
3This debate can also be represented by the different view of ecological economics, emphasizing ‗strong‘
sustainability, vs. environmental economics, adopting ‗weak‘ sustainability (Bergh, 2001; Ayres, 2007).
4 Many more other indices have been developed, more related to the sustainability concept, incorporating not only
the economic but also the social and environmental dimension, like the Index for Sustainable Economic Welfare (ISEW), Human Development Inidicator (HDI), etc. See Bohringer and Jochem (2007); Kerk and Manuel (2008);
and also http://en.wikipedia.org/wiki/Gross_domestic_product#Other_Metrics.
This curve implies that in the early stages of economic growth environmental degradation and pollution increase, but that the trend reverses beyond some turning point. Empirical results that support this theory are mixed, which is due to the use of different methods and data.
The EKC is still investigated, since its presumed presence has implications for both economic and environmental policy. The message of the EKC is that economic development is both the cause and cure of environmental problems. Decoupling the use of natural resources from economic growth has been set as one of the policy goals in the sixth
Environmental Action Programme by the European Commission6. A number of contextual
factors and developments make this bell-shaped curve a relevant study subject. Well-known
and responsible for many environmental problems and threats7 is the concern about climatic
change (IPCC, 2007). Related to this, the capacity of the environment to accommodate (human induced) changes is not unlimited, and some important energy and material resources are finite (Kemp, 2004). Moreover, in 2050 there will be around 2.3 billion more people on this earth (UN, 2011), who all require basic needs like food, water, shelter, clothes and energy. More important, almost all of this population increase will be realized in developing countries. According to the EKC, most of these countries have not yet reached the turning point implying a deteriorating environment for these countries.
A number of interesting questions can be asked. Will economic development and technological advancement be the solution in order to lower the environmental burden? Can developing countries outgrow the situation of a deteriorating environment? This research aims to find answers by testing the simple reduced form of the EKC with per capita income as the explanatory variable, adopting a spatial econometric approach. There are various reasons to consider spatial interaction effects. First, checks for spatial interaction effects have become standard practice in economics, as incorrectly ignoring spatial correlation in the data leads to biased results (Anselin, 2006). Second, the literature review indicated that several EKC studies produced differing results and parameter instability, which might be due to spatial interaction effects (Maddison, 2006). Moreover, the EKC might lead to transboundary pollution via trade, and pollution itself is a transboundary problem. Therefore, it is reasonable to expect spatial spillovers at work, due to countries cross-border mimicking environmental standards and regulations and/or to technology diffusion (Maddison, 2006).
First, it is investigated whether a bell-shaped curve can be found using exploratory data analysis and panel regression techniques. Later, the presence of spatial interaction
6 See: http://ec.europa.eu/environment/newprg/intro.htm.
7 Like global warming, rise of the sea-level, increase in infectious diseases, and increased frequency of droughts and
effects is checked and incorporated if present by running spatial panel regressions. This research does not focus on one single pollutant to represent the environmental burden. Instead, various measures are used as dependent variable which can be qualified as indicators of overall environmental impact, knowing: the Ecological Footprint (EF); carbon dioxide emissions per capita; and per capita energy and material use. This is chosen to best proxy the qualitative state of the environment, as the environment has many different elements which can be influenced by economic activity (Rupasingha et al, 2004). Consequently, the main research question sounds: Is there any empirical evidence in
support of the EKC hypothesis for the various environmental indicators, considering spatial interaction effects? The outcome might be relevant in the light of natural resource scarcity,
climate change and related policy. As far as I know, this is the first paper to combine the use of overall indicators of environmental quality with a spatial econometric approach.
The structure of this paper is as follows. Section 2 gives an overview of existing literature on the subject of the EKC. In section 3 the research setup is explained, including a description of spatial interaction effects. Next, the various dependent and explanatory variables are discussed in section 4. Then in section 5 the results are reported and analysed. Finally, section 6 presents conclusions and offers directions for future research.
2. Literature Review
The following four subsections discuss the background and some explanations of the EKC hypothesis, provide an overview of theoretical studies and empirical evidence, and present some EKC studies that employed a spatial econometric approach.
2.1 Background
GDP growth in higher income levels, where the turning point marks a 'delinking' of environmental pressure from economic growth. Correspondingly, Beckerman (1992) stated that the surest way to improve your environment is to become rich, while Bhagwati (1993) argued that economic growth is a precondition for environmental improvement.
Subsequently, many studies on the subject of the EKC appeared after the above-mentioned pioneering studies (about 150 until 2011 as far as I know). But this increase in EKC studies was also due to the fact that from the beginning of the 1990s empirical data of concentrations and emissions of various pollutants became available (Dinda, 2004).
Various explanations have been advanced to rationalize the hump-shaped curve, although two main explanations can be identified (Galeotti and Lanza, 2005). The first is micro-based, utilizing the income elasticity of environmental quality demand by defining environmental quality as a luxury good. This implies that initially individuals do not invest in environmental protection and conservation. However, as a society becomes richer, preferences start to change and people demand a healthier and cleaner environment. This demand can be satisfied in various ways, like defensive expenditures, donations to environmental organizations, buying less environmentally damaging (‗green‘) products, but also by mobilizing pressure for environmental protection and regulation. The other common explanation is macro-based and reflects the changing structure of an economy as it evolves over time. Grossman and Krueger (1991) identify three different effects on the environment of economic development: a negative scale effect since more output means more extraction of natural resources and more wastes and pollution; an initial negative but later positive composition effect because of structural change of the economy from agriculture to energy intensive industry to service and knowledge based technology intensive industry; and a positive technology effect as old and dirty production capital is replaced by new and cleaner technologies. According to the EKC, in the first phases of economic development the negative scale effect dominates, but in later stages it will be outweighed by the positive composition and technology effect (Vukina et al, 1999). Stern (2004) defines these effects as the proximate causes of any change in pollution levels. All other explanations, like environmental regulation, awareness and education are underlying causes and can only have an effect via the proximate variables.
pollution-intensive industries to other, poorer countries with lower environmental standards. Consequently, while experiencing diminishing levels of environmental degradation, people living in developed countries do not have to change their consumption pattern as they can still import those goods (Arrow et al., 1995; Stern et al., 1996; Suri and Chapman, 1998; and Rothman, 1998). This explanation is criticized as imposing negative externalities internationally can create an illusion of sustainability (Rothman, 1998). Another factor that can justify the EKC is a ‗self-regulatory market mechanism‘, implying a shift from non-market to non-market natural resources as an economy grows which might prevent environmental degradation (World Bank, 1992; Unruh and Moomaw, 1998). Lastly, (in)formal regulation and property rights can cause an inverted-U PIR. This might occur when the institutional environment essential to design, implement and enforce environmental regulation improves with economic growth (Dasgupta et al, 2001).
Most studies adopt a simple reduced form of the EKC, which is a polynomial function of income and linear in the parameters. In the beginning, it was common to apply a cross-country analysis, but later on more panel data is used. The standard EKC model involves regressing environmental pressure on a constant, per capita income, per capita income squared, and an error term:
yit = αi + ηt + β1xit + β2x2it + εit (1)
where y is the environmental indicator, x is income per capita, α and η intercept parameters,
ε an error term and β1 and β2 unknown regression coefficients. The subscript i is an index for
the cross-sectional dimension (with spatial units i=1,…,N, like states or countries) and t stands for the time dimension (with time periods t=1,…,T, like years). Model (1) can be estimated in either levels or logarithms. Various alternatives for both the left-hand and right-hand side variables are available. Generally, environmental pressure is measured by emissions per capita of air pollutants. However, concentration figures of pollutants and other indicators of environmental quality are also used. Income is usually represented by per capita GDP (often based on PPP).
2.2 Criticism to early empirical evidence
Shortly after the publication of the first EKC studies, papers that criticized the EKC concept started to appear. One reason that led to this criticism is its relevance for economic and environmental policy in both developed and developing countries. Various points of criticism can be observed, of which one is differing results. After the three pioneering studies (Grossman and Krueger, 1991; Shafik and Bandyopadhyay, 1992; and Panatyou, 1993),
many other scientists estimated the EKC for various environmental indicators8. However, no
single conclusion with respect to the EKC hypothesis can be drawn from all these studies. In fact, contrasting results have been found when focusing for example solely on sulfur dioxide EKC empirical analyses (See table 1 of Stern, 2004, page 1425; and table 1 of He, 2007, page 37-40). The results indicate different shapes of the PIR (N-curve, inverted-U curve or EKC, monotonically increasing and decreasing, U-curve) and different turning point values ranging from about $2000 till about $100,000 (in 1990 US$).
Although these discrepancies can be due to the use of different data (data source, emissions or concentration, time period, country or city sample, GDP based on PPP yes or no) and different estimation methods (OLS, GLS, panel data random or fixed effects, etc.), this raises questions about the validity of the EKC (Harbaugh et al, 2002). In this respect, Stern (1998) claims that EKC relationships have been found for only a subset of countries and indicators. These findings might have tempted Dinda (2004) to state that the EKC is a statistical artifact. Similarly, Stern (2004) argues that the EKC is an essentially empirical phenomenon, focusing on exploring empirical regularities rather than theoretical foundation. So the criticism is that the EKC is not driven by any particular economic model, which ignores the underlying real causality between the two indicators. Related to this is omitted variable bias, as the simple reduced form of the EKC is criticized. Reduced-form relationships reflect correlation rather than the causal mechanism (Cole et al, 1997), and lack of insight into which processes cause delinking of environmental degradation from economic development makes designing specific policies difficult (Dinda, 2004). Besides, the specific functional forms assumed by reduced-form analyses is criticized, since the quadratic concave and the cubic function in the end imply zero (or even negative) and infinite environmental degradation respectively (Cole et al, 1997).
8 Like air pollution (nitrogen oxides, carbon monoxide and dioxide, sulfur dioxide, suspended particulate matter,
But there are more points of criticism to dispute the hypothesized empirical relationship between income and environmental quality, keeping the economics versus environment debate alive. First, the early empirical EKC studies adopted a static approach, while the EKC is an inherently dynamic process (Kidd, 2009). Related to this is the key criticism of Arrow et al (1995) about the unidirectional causality from income to environment. Income is assumed to be an exogenous variable in the EKC model, implying that there is no feedback from environmental damage to economic production. However, several studies contest the assumption that there is no threshold on environmental damage that can produce irreversible ecological consequences. This can be problematic when resource depletion or degradation is involved which has productivity-related effects on the economy (Perrings, 1987; Stern, 1998; Stern, 2004). Third, it is argued that not enough attention has been paid to country-specific EKCs. Many studies used cross-section data and examined the EKC for groups of countries, assuming that economic development trajectory would be the same for all (Dinda, 2004). De Bruyn et al. (1998), List and Gallet (1999), Munasinghe (1999), and Dijkgraaf and Vollebergh (2005) all criticize this EKC assumption that every country must pass through a similar development path. Social, economic, political and biophysical factors may vary among countries and developing countries might learn from developed countries via technology diffusion and ‗tunnel‘ through the EKC (Munasinghe, 1999; Anderson, 2001). Another point of criticism concerns econometric problems of EKC studies. Stern (2004) argues that most of the EKC literature is econometrically weak and that EKC results have a very flimsy statistical foundation. Perman and Stern (2003) find that the inverted-U EKC does not exist, when taking diagnostic statistics and specification tests into account and using appropriate techniques. Finally, Rothman (1998) criticizes the focus of many EKC studies on production-based indicators, which are in general local in nature and have relatively inexpensive abatement costs. He emphasizes that EKCs are attributable to changes in the production structure of economies, but argues that consumption is the principal driving force behind environmental impact. Consequently, to prevent handing off environmental problems by passing them to people in other places (see PHH) or times, the author advocates the use of consumption-based measures of environmental impact.
2.3 New developments
be characterized as the main criticism. This led to the development of ‗second generation empirical studies‘ which are more complex and inclusive, and to the construction of economic models providing a theoretical explanation for the EKC hypothesis.
2.3.1 Modern EKC empirics
Initially, none (Panatyou, 1993) or only a time trend and locational dummies (Shafik and Bandyopadhyay, 1992) were considered as explanatory variables besides income. Sometimes population levels were also accounted for, as higher population density might lead to greater environmental stress (Grossman and Krueger, 1995). However, in the last fifteen years many additional explanatory variables have been incorporated to address omitted variables bias. These studies often test the proximate and underlying causes (discussed in section 2.1) that are advanced to explain the EKC. Cole et al (1997) estimate technology level as an empirical determinant, as cleaner production capital is more efficient in the use of energy and materials. Many have investigated the role of trade in shaping the EKC, and found evidence for the existence of the PHH (Arrow et al, 1995; Stern et al, 1996; Suri and Chapman, 1998; Rothman, 1998; and Copeland and Taylor, 2004). However, Stern (2004) thinks this empirical evidence is not fully clear and convincing, which is supported by Cole (2004) and Kearsly and Riddel (2010) who find evidence of pollution havens for only some pollutants and of limited significance. The effect of economic structure is examined by Suri and Chapman (1998) and Hettige et al (2000), who looked at the share of industrial production in national output. Besides, according to Merlevede et al (2006) and Cole et al (2005), larger average firm size is more successful in improving the environment.
Instead of incorporating additional explanatory variables, some studies adopted a different approach and tested for other influences. Kijima et al (2011) state that there is no EKC between pollution and income but between pollution and time, while Stern (2004) argues that emissions rise monotonically with respect to income but can be reduced by time related technique effects. The emphasis on ‗income determinism‘ was already doubted by Moomaw and Unruh (1997) and Unruh and Moomaw (1998), who proposed nonlinear dynamic systems analysis and concluded that instead reductions in emissions have been triggered by external shocks. Specific events such as the 1973 oil crisis and the discovery of the ozone layer appeared to provide a sufficient incentive for new policy initiatives to lower pollution. Finally, some also used other measures like the human development index to reflect economic development, like Jha and Murthy (2003) and Costantini and Monni (2008). Generally, the additional variables turn out to be significant. But again it should be emphasized that the results are dependent on the used data and applied method. Furthermore, Stern (1998) and Rupasingha et al (2004) notice that no study has incorporated these ‗new‘ EKC determinants into a single conceptual framework. As a result, these studies are still subject to the problem of omitted variables bias. Finally, Agras and Chapman (1999) argued that income has the most explanatory power and thus the most significant effect on indicators of environmental quality.
2.3.2 Theoretical Foundation
In response to criticism of a missing theoretical foundation and weak econometrics, various economic models were constructed to explain the EKC and more attention was paid to econometric issues like heteroskedasticity, autocorrelation, multicollinearity, homogeneity, non-stationarity, cointegration and simultaneity bias (Lieb, 2003; Stern, 2004). The theoretical frameworks vary from static to dynamic models and introduce a trade-off between benefits and losses caused by pollution. They are either founded on macroeconomic production functions or utility functions only, and have been extended by including effects like technological progress, tax policy and environmental regeneration (Kijima et al, 2010).
One of the first to analyze the EKC hypothesis from a static theoretical perspective is Lopez (1994), who models the environment as a production factor. The author shows that
for environmental factors that do not have stock effects on production9, an inverted-U PIR
emerges if the welfare function is of non-homothetic preference such that economic growth increases the value of environment for consumers. A number of papers modeled the EKC
9 For resources with a productive feedback effect, it is shown that economic growth leads to lower environmental
considering only the utility function of the representative agent, starting with McConnell (1997). Different conditions for the EKC can be observed, like assuming that environmental quality is a normal good with respect to income instead of a luxury good (Lieb, 2002), increasing returns to scale of pollution abatement effort (Andreoni and Levinson, 2001), or just the marginal disutility of pollution alone (Di Vita, 2004). De Bruyn et al (1998) utilize a different approach based on decomposition analysis, and argue that the hump-shaped curve may arise due to changes in environmental policy rather than in economic structure.
Although static models may provide an understanding of certain causes and characteristics of the EKC, Kidd (2009) states that these are limited in their ability to describe the interconnected nature of the economic and environmental dimension. Consequently, dynamic models are better to describe the EKC and to gain insights into its origins. Most of these contributions consider an overlapping generations model in which each agent lives for two periods and faces the problem of income allocation between consumption and abatement effort, and where pollution is thus consumption-based. John and Pecchenino (1994), in discrete time setting and also considering the case of technological externalities, and Selden and Song (1995), in continuous-time setting and with infinitely lived agents, are one of the first who constructed a dynamic EKC model. Many adaptations and/or extensions to these models exist. Lieb (2004) distinguishes stock and flow pollutants, Ranjan and Shortle (2007) consider the environment in the presence of hysteresis, while both Chimeli and Braden (2008) and Prieur (2009) assume a limited assimilation capacity of pollution by the environment. Stokey (1998), Hartman and Kwon (2005), Tahvonen and Salo (2001) all endogenize technology, while Jones and Manuelli (2001) and Egli and Steger (2007) study the effects of environmental regulation. Finally, the real options approach is applied recently to include uncertainty in the economy (Di Vita, 2008b; Wirl, 2006; Kijima et al, 2011), as Pindyck (2007) points out that deterministic models are insufficient to capture uncertainty over the physical and ecological impact of pollution, over the economic costs and benefits of reducing it, and over the discount rates that should be used to compute present values.
2.3.3 Different Environmental Measures
More developments can be identified in modern EKC literature, like the use of different measures to indicate environmental quality. As mentioned before in footnote 7, many indicators have been used for the dependent variable, like various flow and stock pollutants, and measures of deforestation, waste, material use, energy use, land use, water quality and use, and biodiversity.
However, recently other indicators like the Ecological Footprint (EF) are employed to represent the state of the environment. So far, only three studies examined the effect of income on EF. As mentioned before, Rothman (1998) advocates the use of consumption-based instead of traditional production-consumption-based measures of environmental impact and is the first to test the EKC considering EF data. Plotting per capita EF of 52 countries for the year 1997 against GDP per capita, and estimating four alternative specifications does not produce strong evidence in favour of the EKC. Bagliani et al (2008) extend on Rothman‘s cross-country analysis in terms of spatial coverage by using 2001 EF data for 141 countries and by looking in more detail at the EF components. Instead of decoupling, the authors find evidence of unbounded growth of environmental pressure as income increases. However, they mention that the inverted-U shape might yield for the energy component of the EF. Caviglia-Harris et al (2009) work with an unbalanced panel covering EF data for 146 nations spanning eight, 5-year periods from 1961-2000. They estimate the traditional EKC model utilizing various methods. The authors also do not find empirical evidence of an EKC relationship. Interestingly, they conclude that energy is largely responsible for the lack of an EKC, and argue that this component should be discounted with 50% to realize an EKC.
Galeotti and Lanza, 2005; Galeotti et al, 2006; Musolesi et al, 2010), of which most have a
quantified emission limitation or reduction commitment in the Kyoto protocol10. However,
the common finding is that this PIR is monotonically rising (Lieb, 2003; Dijkgraaf and Vollebergh, 2005; Azomahou et al, 2006; Aslanidis and Iranzo, 2009; Narayan and Narayan, 2010).
2.3.4 Overall Empirical Results
The debate about the environment-economy relationship has not settled and still continues. Dinda (2004) states that the EKC subject is open-ended and that evidence for its existence
is inconclusive. Lieb (2003) and Dinda (2004) observe that most authors agree on the
importance of local and (inter)national policies in shaping the income-environment relationship. Secondly, they identify that the scale and shape of the EKC may vary per environmental indicator and per region, which also leads to different turning points. This raises questions about the EKC, like the issue whether the EKC is permanent and sustainable. Uncertainty about possible irreversible environmental effects also plays a role here, as most studies only consider environmental impact on a per capita or unit of economic activity basis, but do not examine total impact on environment (Rothman, 1998).
Generally, an EKC can be observed for the so-called short-lived flow pollutants (like sulfur dioxide, nitrogen oxides and carbon monoxide) causing mainly immediate and local problems, but not for pollution problems where the effects are far-displaced or long-delayed. In fact, the PIR seems to be monotonically rising for the long-lived global stock pollutants
like carbon dioxide (CO2) and natural resource use indicators like land, energy and material
use (Cavlovic et al, 2000; Lieb, 2003; Dinda, 2004). This picture fits environmental
economics theory (Stern, 2004) and can be explained by Hardin‘s theory about the tragedy of the commons (Hardin, 1968), because of the public good nature of clean air and of air pollution abatement. As a result, local impacts which are visible and cause direct damage can be overcome by local governments that provide the locally efficient level of abatement. However, this is much more difficult in the case of global, stock pollutants which involve transboundary and intergenerational externalities and provide countries incentives to free-ride on the efforts of other countries. This is because governments are not always that farsighted and the institutional capacity to make intergenerational transfers is constrained (Ansuategi and Escapa, 2002). Ansuategi and Perrings (2000) and Ansuategi (2003) argue
10 The so-called Annex I countries, who signed and ratified the Kyoto protocol, but also committed themselves to
that international cooperation is needed to internalize transnational external effects of pollution, but this is not easy due to high costs of policy implementation (Kijima et al, 2011).
2.4 EKC studies with spatial econometric approach
Florax and Van der Vlist (2003) argued that spatial econometrics continues to be an underutilized technique in environmental and resource economics. Most empirical papers pay attention to econometric issues like heteroskedasticity, non-stationarity, cointegration, and autocorrelation but ignore the spatial properties of the data. While it is reasonable to expect spatial interaction effects among units in cross-section analysis, tests for proximity-based relationships are rarely performed (Maddison, 2007). However, erroneously ignoring spatial correlation in the data has important implications like faulty inference testing procedures, possible bias and inconsistency (Anselin, 2006). Also regarding the EKC topic, most studies excluded the influence of neighbouring states on a countries‘ per capita emission. However, a few papers have investigated and incorporated spatial interaction effects, reflecting the increased interest in spatial modeling in economics in the last decade (Anselin, 2006).
Various explanations have been put forward to rationalize the presence of spatial relationships in EKCs, of which the most common are policy-mimicking by governments, the PHH and technology diffusion. Many of these cross-sectional relationships are based on Tobler‘s (1979) first law of geography, which states that ―everything is related to everything else, but nearby things are more related than distant things‖. Maddison (2006, 2007) argues that countries‘ strategic response to transboundary pollution flows is the most obvious explanation to expect spatial relationships. Specifically, governments might resort to cross-border copying of environmental standards and policies, as has been shown for states in the US (Fredriksson and Millimet, 2002). This mimicking behaviour can be motivated by economic reasons in order to attract capital or for trade purposes (Ulph, 1992) and by political reasons to reduce the costs of decision-making and to legitimize political actions (Frank et al, 2000). Furthermore, since geographical distance is an important determinant of technology diffusion (Keller, 2004) and volume of trade (Anderson and Van Wincoop, 2004), technological change and the PHH might be responsible for the existence of spatial patterns in measures of environmental quality. Maddison (2006) states that it is difficult to determine the exact reason of spatial spillovers, but that its presence provides a possible explanation of why many authors have observed instability in the parameters of the EKC.
groups and estimated separate EKCs for each group. While McPherson and Nieswiadomy (2005) claim to be the innovating empirical EKC research by incorporating spatial interaction effects, Rupasingha et al (2004) already performed this in their paper. The latter investigate the EKC for US counties, which is also pioneering since only state-level pollution data has
been used. The authors consider a spatial error model11, because they expect that the
residuals from different counties may display spatial covariance. Evidence is found for the existence of spatial interaction effects and estimation results indicate an N-shaped PIR. McPherson and Nieswiadomy (2005) examine the EKC for threatened bird and mammal species in a cross-country analysis. They detect spatial interaction effects, estimate a spatial lag model and conclude that an EKC may exist for both birds and mammals.
Whereas the previous two studies treated spatial interaction effects as an extension, Maddison (2006) is the first who adopts a comprehensive spatial econometric approach to test the EKC for some pollutants. The author uses a panel of various pollutants, performs a number of tests to determine whether and what form of spatial relationships are present, and considers four different spatial weights matrices. Estimating the spatial lag model does not support the EKC, but does indicate that national emissions per capita of sulphur dioxide and nitrogen oxides are heavily influenced by the per capita emissions of neighbouring countries. Maddison (2007) elaborates on the sulphur dioxide case, analysing emission data for 25 European countries over the period 1951-1990. Appendix A provides an extensive comparison of the four EKC articles that employed a spatial econometric approach.
3. Research setup
This section describes the methods utilized in this research. After motivating the application of exploratory analysis, I will present the model that is used to test the EKC hypothesis. Finally, I will discuss spatial interaction effects and incorporate those in the panel regression if test results indicate its existence.
3.1 Exploratory Analysis
In order to explore the possible presence of EKCs, I will construct scatterplots to explore possible EKCs. First, this is done for all observations plotting each environmental measure against per capita income. An order 3 polynomial trendline will be added to check whether an inverted-U PIR can be detected. Second, an individual approach is adopted by making scatterplots for each country separately. The time dimension of the data panels allows such
an analysis. This approach addresses the criticism that not enough attention is paid to country-specific EKCs. However, to investigate the possible presence and significance of an EKC in a more formal way, regression techniques need to be deployed.
3.2 Regression Model
Since this research is executed using data from several countries covering a number of years, a panel regression will be performed. The advantage of panel data over cross-sectional data is that it leads to higher efficiency of econometric estimates, since it simply provides more observations and in general is more informative. Moreover, it allows to control for unobserved heterogeneity, addressing the problem of omitted variables bias. The model to be estimated is based on the simple reduced form of the EKC, including only per capita income as explanatory variable. Accordingly, no additional variables are considered because literature has found that a variety of factors are significant. Incorporating all these factors will be difficult and problematic, while including a selection is subjective. This way, income is used as a catch-all variable representing the net effect on the environment (He, 2007). Consequently, it cannot be determined what causes the EKC if present.
Whether the model will be estimated in levels or logarithms is an empirical question, although the use of logarithms is preferable in the case of a quadratic formulation, since minus infinity pollution as income goes to infinity is not rational. A preview of the scatterplots reveals to estimate the model in levels. Finally, I decide to add a cubic term,
because the empirical evidence in favour of an N-shaped PIR cannot simply be ignored12.
The cubic formulation also allows for both the inverted-U and monotonically rising PIR. If the cubic term turns out not to be statistically significant at 5%, it can simply be eliminated (Lieb, 2003). Summarizing, I will estimate the following model:
ENVit = αi + ηt + β1GDPit + β2GDP2it + β3GDP3it + εit (2)
where ENVit represents the variable that indicates environmental burden for country i at year
t (i=1,…,N; t=1,…,T), GDPit is per capita income for country i at year t, βk is the coefficient
of the k explanatory variables, αi and ηt stand for intercept parameters which capture
country-specific and year-specific effects, and εit is an independently and identically
distributed error term for i and t with zero mean and variance σ2. Incorporating
spatial-specific effects accounts for space-spatial-specific time-invariant variables whose omission could
12 Kijima et al (2011) is the first who explains both the inverted-U and N-shaped PIR in an unified framework,
bias the estimates in a typical cross-sectional study. The same goes for time-period specific intercepts, which are intended to control for time-varying omitted variables and stochastic shocks that are common to all countries (Stern, 2004; Baltagi, 2005).
Model (2) is used to test which shape the PIR has for the various environmental
measures, since estimating (2) can produce different results: (i) β1=β2=β3=0. An insignificant, flat PIR.
(ii) β1>0 and β2=β3=0. A monotonically increasing PIR.
(iii) β1<0 and β2=β3=0. A monotonically decreasing PIR.
(iv) β1>0, β2<0 andβ3=0. An inverted U-shaped PIR, i.e. the EKC.
(v) β1<0, β2>0 andβ3=0. A U-shaped PIR.
(vi) β1>0, β2<0 andβ3>0. An N-shaped PIR.
(vii) β1<0, β2>0 andβ3<0. An inverted N-shaped PIR.
From above, it appears that the EKC is just one of the seven possible outcomes of the model. However, following Lieb (2003) it is assumed that the PIR must start at the origin, implying that man-made pollution must rise before it can fall. Accordingly, a monotonically decreasing PIR is interpreted as evidence for an EKC with a very low turning point. Similarly, a U-shaped PIR is interpreted as evidence of an N-shaped PIR.
Model (2) will be estimated by running a panel regression for four different panel
data specifications: without αi and ηt (a constant term will be included in this case), only αi,
only ηt, and both αi and ηt. Subsequently, a likelihood ratio (LR) test can be performed to
investigate which setting is preferred.
3.3 Spatial Econometric Analysis
different spatial models can be distinguished13: the spatial lag model, the spatial regression
model and the spatial error model. However, until recently only the first and last model have received considerable attention by spatial econometricians (LeSage and Pace, 2009).
First, it is investigated whether spatial correlation in the data is present, because this can have important implications for the estimation results. Incorrectly ignoring spatial correlation may lead to biased and inconsistent estimates of the model parameters for the spatial lag model, or inefficient estimates and biased t-test statistics in case of the spatial error model (Anselin, 2006). Like Maddison (2006) argued, this might explain why many authors have observed instability in the parameters of the EKC and possibly lead to a stronger acceptance or rejection of the EKC.
As identified in section 2.4, several explanations exist to justify the existence of spatial effects for the EKC. LeSage and Pace (2009) provide a number of theoretical motivations to include spatial autoregressive processes. First, time-dependence may play a role, as cross-sectional spatial relationships of the dependent variable may result from economic agents who consider past behavior of neighboring states. It can be argued that this motivation corresponds with the cross-country mimicking behavior explanation, where
countries look at environmental performance of their neighbours14. Accordingly, a spatial lag
model containing a spatially lagged dependent variable has to be estimated:
N
yit = ρ
∑
wijyjt + βxit + αi + ηt + εit (3)j=1
where yit is the dependent variable for cross-sectional unit i at time t (i=1,…,N; t=1,…,T).
The parameter ρ is the spatial autoregressive coefficient and measures the spatial
interaction. The variable ∑jwijyjt represents the interaction effect of the dependent variable
yit with the dependent variables yjt in neighbouring units, where wij is the (i,j)-th element of
a prespecified nonnegative NxN spatial weights matrix W describing the spatial arrangement
of the units in the sample. xit denotes a (1,K) row vector of exogenous variables, and β a
matching (K,1) vector of fixed but unknown parameters. The terms αi and ηt are optional
and capture spatial specific and time-period specific effects respectively, while εit is an error
term, independently and identically distributed for i and t with zero mean and variance σ2.
Second, spatial dependent omitted variables might cause spatial effects, like for instance technology diffusion. This can be estimated with the spatial error model, where the error terms are correlated across space:
13 Modeling spatial heterogeneity will not be considered, as Anselin et al (2008) discusses some forms but states
that none has seen application in panel data contexts.
14 This phenomena, based on the spillover or resource-flow model (see Brueckner, 2003), has been proved for
N
yit = βxit + αi + ηt + φit, φit = λ
∑
wijφit + εit (4)j=1
where φit reflects the spatially autocorrelated error term and λ denotes the spatial
autocorrelation coefficient.
Subsequently, it will be determined which type of spatial interaction effect should be accounted for, which can differ per environmental measure. However, LeSage and Pace (2009) argue that one should also consider the so-called spatial Durbin model when the non-spatial model is rejected in favor of the spatial lag or spatial error model. This model extends the spatial lag model with spatially lagged values of the independent variables:
N N
yit = ρ
∑
wijyjt + βxit + θ∑
wijxijt + αi + ηt + εit (5)j=1 j=1
where θ is the spatial correlation coefficient for the independent variables.
This research adopts the testing procedure as explained in Elhorst (2010b). First, after estimating the non-spatial model for the four different panel settings it will be examined whether there is evidence for spatial interaction effects (specific-to-general approach). This can be done by looking at the (robust) Lagrange multiplier (LM) test results, which indicate whether the spatial lag or spatial error model is more appropriate. The likelihood ratio (LR) test will be investigated to determine which setting best fits the data. Subsequently, a general-to-specific approach is employed by estimating the spatial Durbin model to test whether it can be simplified to the spatial lag or spatial error model. In order to decide on this, Wald and LR tests are performed. The criterion to choose the spatial lag or spatial error model is whether both approaches indicate the adoption of one of the two models, otherwise the more general spatial Durbin model will be utilized. Finally, the result of Hausman‘s specification test will be used to examine whether the random effects model should be accepted or rejected with respect to the fixed effects model.
To be able to model these spatial relationships, a spatial weight matrix W must be
specified which defines a neighbour structure for each observation. The spatial weights wij
are non-zero when the observations i (row) and j (column) are hypothesized to be
neighbours, and zero otherwise15. The value of the weights can be based on geographical
characteristics16 (contiguity, distance, general or graph-based weights) as well on
socio-economic factors like status scores and political resources (Anselin and Rey, 2006), and these elements of W can be standardized by row (common), column or matrix normalization. The specification of W is important and deserves attention, because another W affects the value and significance level of the interaction parameter and can lead to different empirical results (Leenders, 2002).
In this research I consider four different row-normalized spatial weights matrices. The first spatial weights matrix is contiguity-based, indicating which countries share a common border (applying queen contiguity order 1). The reason is that proximity might play a large role in stimulating cross-country mimicking behavior. This way, for instance Australia has no neighbours, which yields for all islands. Likewise, one can wonder what to do with countries connected by a road. These issues can be dealt with by changing the weights matrix manually, but this is subjective and a lot of work when considering all countries of the world.
Another option is to consider distance-based weights, which is done for the second weights matrix. Maddison (2006) suggests that this method is best to capture distance related phenomena (like trade and technology diffusion). The difficulty with distance-based weights is choosing the distance threshold criterion, which often leads to an unbalanced connectedness structure. This is the case when the spatial units have very different areas, like countries. As a result, smaller countries end up having many neighbors, while the larger
countries may have very few (or none, yielding ‗islands‘17). For example, the USA and
Canada might not be neighbours this way. Here the distance cut-off is set at 2,750 km18.
The minimum threshold distance to avoid islands may not be representative for the rest of the distribution. Anselin (2005) states that in such cases the use of k-nearest neighbor weights may be more appropriate. This method ensures that each observation has
15 The self-neighbour relation is excluded, so that the diagonal elements of W are zero.
16 Related to the fundamental concept in geography that nearby entities often share more similarities than entities
which are far apart, see Tobler‘s first law of geography.
17 Since for polygon shapefiles the polygon centroids are used as x-y coordinates.
18 Maddison (2006) uses 2,800 km, as the most remote country is 2,750 miles from its nearest neighbour. If I
exactly the same number (k) of neighbors. Which k to choose is a bit arbitrary, here I consider two more weights matrices with k=6 and k=10.
As noted, each W has its shortcoming. There is little guidance in the choice of the ‗correct‘ W, causing the choice for a certain W debatable. However, Seldadyo et al (2010) present a method to examine which spatial weights matrix best describes the data. The paper describes two approaches to compare different spatial weights matrices, one based on log-likelihood function values and the other based on Bayesian posterior model probabilities. Besides performing a sensitivity analysis to evaluate the influence of different Ws on the regression results, this research will adopt this method to determine which W is most appropriate.
Both the non-spatial and spatial panel regressions will be performed in Matlab, for
which I will use Matlab routines written by Elhorst19, Lacombe20 and LeSage21.
4. Data
This research estimates the PIR employing various indicators as the dependent variable to
reflect environmental degradation: the Ecological Footprint (EF); CO2-emissions per capita;
and per capita energy and material use. The independent variable economic development is measured by GDP per capita. Data from various countries over time is available which implies a panel data setting. Finding reliable and comparable pollution data is one of the main problems in estimating the PIR. This is due to the use of different methods and unrepresentative locations to measure data. Moreover, there is a lack of sufficiently long time series for many countries, as collecting environmental data started in the 1960s at the earliest. This holds especially for developing countries, whose data is also often considered unreliable. Accordingly, data quality of many EKC studies is poor and the use of certain pollution indicators is often determined by data availability (Lieb, 2003). Below I will describe each indicator and touch upon the reliability issue, after which I will elucidate why these indicators have been chosen and how I constructed the database to work with.
4.1 Data description
The EF22 is a metric that calculates human pressure on the planet, and is nowadays widely
used by scientists, businesses, governments and institutions working to monitor ecological
19 See http://regroningen.nl/elhorst/software.shtml.
20 See http://community.wvu.edu/~djl041/matlab.html.
21 See http://www.spatial-econometrics.com/.
22 Developed in 1990 by Mathis Wackernagel and William Rees at the University of British Columbia, see Global
resource use and advance sustainable development. The EF is the sum of the cropland, the grazing, the forest, the fishing ground, the carbon and the built-up land footprint, and is measured in global hectares per capita. It represents the amount of biologically productive land and sea area needed to regenerate the resources a human population consumes and to absorb and render harmless the corresponding waste. The EF data comes from The Ecological Footprint Atlas 2010, The Global Footprint Network. This is a nonprofit organization that works with scientific data to calculate the EF. The EF is available from 1961 for 116 countries to 2007 for 150 countries. The 2010 National Footprint Accounts use over 5,000 data points for each country, each year, derived from internationally recognized sources to determine the area required to produce the biological resources a country uses and to absorb its wastes, and to compare this with the area available. The calculations in the National Footprint Accounts are primarily based on international data sets published by the Food and Agriculture Organization of the United Nations, the International Energy Agency, the UN Statistics Division, and the Intergovernmental Panel on Climate Change. Other data sources include studies in peer-reviewed science journals and thematic collections.
Material use23 is measured by total material extraction in metric tons (1,000kg) per
capita on a national level. This is not the same as domestic material consumption since imports and exports of material flows are not incorporated. Material extraction data is
available for 229 countries over the years 1980-2007. Energy use24 is represented by
domestic energy consumption, as it refers to use of primary energy before transformation to other end-use fuels, which is equal to indigenous production plus imports and stock changes, minus exports and fuels supplied to ships and aircraft engaged in international transport. Energy use data is in metric tons of oil equivalent per capita, and available from 1960 (25 countries) to the year 2007 for 162 countries.
Carbon dioxide emissions per capita come from the World Development Indicators,
The World Bank, which is a known and widely used dataset25. The reported carbon dioxide
emissions are those stemming from the burning of fossil fuels and the manufacture of cement. They include carbon dioxide produced during consumption of solid, liquid, and gas fuels and gas flaring. This indicator is computed instead of measured and expressed in metric tons per capita and is available from 1960 (165 countries) till 2007 (196 countries).
Income best measures welfare and is as usually indicated by GDP per capita. GDP per capita is gross domestic product divided by midyear population. The gross domestic
23 Database provided and set up by Sustainable Europe Research Institute (SERI) in cooperation with the Wuppertal
Institute for Climate, Environment, and Energy. See http://www.materialflows.net/index.php.
24 Provided by the International Energy Agency (IEA), see http://www.iea.org/stats/index.asp.
25 World Development Indicators, The World Bank, see http://data.worldbank.org/. Data provided by the Carbon
product is the sum of gross value added by all resident producers in the economy plus any product taxes and minus any subsidies not included in the value of the products. It is calculated without making deductions for depreciation of fabricated assets or for depletion and degradation of natural resources. I use PPP-adjusted per capita GDP in 2005
international dollars (from 1960 for 112 countries till 2007 for 188 countries)26, since GDP
measured in market exchange rates underestimates the income of poor countries.
Considering the sources, the involved organizations and the methodology of constructing the indicators, the data can be interpreted to be reliable. The indicators are often used and familiar to most people. Naturally, it should be kept in mind that country data may be subject to different measurement methodologies and data manipulations. Besides, the EF is a composite and weighted index, which involves that choices have to be made about which criteria to include and what weights to attach. Accordingly, the EF is both praised and criticized because it condenses a large array of environmental data into a single measure.
4.2 Data Motivation
Besides its availability, I chose these indicators because of a number of reasons. First, I think these indicators best describe the overall impact on environment as a result of human
activity27. Second, most are consumption-based indicators that account for possible pollution
displacement effects. Remember, the PHH states that trade leads to global ‘NIMBY‘28 effects,
in which poor countries become the environmental dumping grounds for developed countries. Material extraction is a production-based measure that does not incorporate transboundary trade flows and thus does not capture potential delocalization effects, which might result in a bias toward acceptance of the EKC with PHH as a possible explanation.
Third, I decided to analyse the EKC for the pollutant CO229. Carbon dioxide is the most
prominent GHG and has effects distant in space (transnational) and time (i.e. it is a global
stock pollutant30). In addition, CO
2 is target of many (inter)national policies and agreements.
Furthermore, according to Lieb (2003) it is the best of any pollutant as relatively long time
series are available for both developed and developing countries. I chose to use CO2
emission instead of CO2 concentration data, because emissions are directly linked to
economic activity. Moreover, emissions are rather easy to calculate while concentrations are
26 From the Penn World Tables v 6.3, see http://pwt.econ.upenn.edu/php_site/pwt_index.php.
27 Of course there are other indicators like the environmental degradation index (EDI) constructed by Jha and
Murthy (2003) but these are not widely known and used.
28 ‗NIMBY‖ stands for ‗not in my backyard‘.
29 Unfortunately, there is no composite air (or water) pollution index available for many world countries over time.
30 Carbon dioxide has an average atmospheric lifetime of about 100 years before it is absorbed by terrestrial or
difficult to measure accurately. Besides, concentration figures just measure the local impact of emission activities without regard to the origin of these activities, while carbon dioxide can have global effects.
4.3 Database Construction
As identified above, data is not available for every year of every country. In general, more data is available for more recent years, but this differs per indicator. As a result, I will have to cope with missing data in building the database. I choose to exclude these missing data from the analysis, which implies that I have to decide which years and countries to include in the panel. During this process I aim to fulfill two conditions:
- The largest panel size to include as many observations as possible (as it is shown that the EKC may be sensitive to the choice of time period and country sample); - A representative composition of countries, as most missing data from early years
(1960s) is attributable to lower income, developing countries.
Sometimes countries show a period of missing data, while before and after these years data is present, which is often due to war/conflicts (like Somalia) and the separation of countries (like Czechoslovakia). These countries will be excluded if the data gap falls into the data range of the panel. Income data is guiding in this process since GDP per capita is the only explanatory variable, whereupon the availability of the environmental indicators will be examined. This process results in four different panels. Appendix B contains an overview of the data with descriptive statistics for all observations as well as for the four different data panels, and presents which countries are included in each panel. As can be seen, this research uses quite large panels (in both dimensions), which is important to find significant estimates of the indirect effects (Elhorst, 2010b).
Accordingly, also four different Ws need to be constructed, which is done from a shapefile using Geoda and Matlab. A number of shapefiles are inspected to find out which
renders the most complete and accurate Ws31, by looking at the connectivity histogram. It
turns out that the shapefile from Environmental Systems Research Institute, Inc. (ESRI) of
the year 2000 is the best to choose32.
31 As country boundaries are represented by polygons which may differ per shapefile, and one shapefile contains
more accurate boundaries and more countries than another.
5. Results
The coming three sections contain the results of this research and provide explanations regarding the scatterplots and the non-spatial and spatial panel regressions.
5.1 Scatterplot Results
The scatterplot graphs are depicted in Appendix C, where both overall and country-specific PIRs are provided. Regarding the overall scatterplot graphs including all observations, it turns out that for all four environmental measures no clear EKC can be observed. All overall scatterplots show a positive relationship between income and environmental degradation, but the variance is large. The added trendline shows an N-shaped relationship for all four
environmental measures CO2 emissions, Ecological Footprint (EF), material extraction and
energy use33.
Another approach is to investigate the PIR for each individual country. In table 1 below the number of countries are given classified by the type of relationship between per capita income and the different environmental measures:
Table 1: Classification of country-specific PIRs
Dependent Variable Positive Negative EKC Other (U-shape, constant)
Material extraction (154) 68 42 8 36
Energy use (100) 61 10 7 22
CO2 emission (145) 83 21 3 38
Ecological Footprint (108) 40 40 8 20
Total (507) 252 113 26 116
For the dependent variable material extraction, most PIRs (68) show a positive relationship with income per capita, and 42 countries show a negative PIR. Only Canada, Fiji, Japan, Luxembourg, Nicaragua, Papua New Guinea, St. Lucia and United States of America exhibit a clear EKC, all other countries demonstrate another or no clear relationship. The same holds
for the dependent variables energy use and CO2 emission, while the results for EF are in
balance regarding the upward and downward-sloping PIR. For energy use, an EKC can be observed for the countries Bulgaria, Colombia, Germany, New-Zealand, Poland, Romania
and Sweden. Considering CO2 emission, only Cuba, Gabon and Niger depict a clear EKC.
Lastly, regarding the ecological footprint, an EKC can be identified for Bulgaria, Canada, Denmark, Germany, Hungary, Poland, Romania and South Africa. No clear relation can be
33 This result does not come as a surprise because the trendlines are based on a polynomial function of order 3
without intercept, which is in fact the same as a linear regression (OLS). Basing the trendlines on a fractional
polynomial function led to N-shaped PIRs for CO2 emissions and the EF, while for material extraction and energy
detected among the various countries that show an EKC. Furthermore, it can be noted that Bulgaria, Canada, Germany, Poland and Romania show an EKC for two dependent variables.
Now both overall and country-specific PIRs are examined, one may ask whether there is enough evidence for the existence of an EKC. In case of the overall PIRs, it seems that there is no EKC but just a positive (N-shaped) relationship between income and environmental degradation. The country-specific PIRs show a more mixed picture, although the majority (252 of 507) also indicates a positive relation. Moreover, only 26 EKCs are observed (5%). One can argue that many countries show the first part of an EKC. However, either the delinking part is missing as there is no change in environmental measure for high incomes, or there is a (very) high turning point. Summarizing, the scatterplots do not give convincing evidence for the existence of an EKC, but rather indicate a positive PIR. The regression results in the next section may provide more information.
5.2 Panel regressions
Before the regression results will be discussed, first the correlation matrix is investigated to check the correlation between all variables separately (see Appendix D). The correlation between the dependent variable and GDP/capita is positive (0.41 – 0.75) for all environmental measures. The highest is for energy use, indicating that a growing economy consumes more energy in per capita figures, which is common. The lowest correlation with GDP/capita is for material extraction, which might be due to the fact that this measure does not account for material imports and exports. The correlation between the independent variables is quite high, ranging from 0.61 – 0.93, which is plain logic when also incorporating the squared and cubic values of a regressor. However, no perfect multicollinearity is observed.
Just running the regressions with the unadjusted values in 2005 I$ of per capita income, its squared and cubic value leads to a warning in Matlab, which holds that ―the
matrix might be badly scaled” with possibly inaccurate results as consequence. This can be
dealt with by adjusting the scale of the units of the explanatory variables: GDP/capita will be expressed in thousands of 2005 I$, GDP/capita squared in millions of 2005 I$ and GDP/capita cubic in billions of 2005 I$.
Below the OLS results are given in Table 2. These results are obtained including both
country-specific and year-specific effects34. Besides, the non-spatial regressions are run
34 A likelihood ratio (LR) test is performed to investigate the hypotheses that the spatial fixed effects and the
