Analysis of Chinese Industry Environmental Kuznets
Curve: A Spatial Panel Data Model
Wang Xin1
July 2013
Supervised by prof. dr.J.P. Elhorst
Abstract
This paper examines the relationship between income per capita and carbon dioxide (CO2) emission using Chinese provincial data over the period
1995-2010 for two CO2 emission indicators. The model incorporates spatial
effects, as well as control variables that have used in previous studies. The results provide evidence in favor of indirect spatial spillovers for carbon dioxide emissions among provinces. The results point to either inverted-U or inverted-N relations for different functional forms employed. A robustness test with cross-country data shows that cross-country data does not improve the estimation of EKC for China.
Keywords: Economic growth; Environmental Kuznets Curve; Spatial econometrics
JEL classification: C23; C53; Q53; Q56
1
1. Introduction
The disastrous weather events due to the climate change, in particular global warming, have prompted worldwide concern since the beginning of the 1990s. The Earth's atmosphere is overloaded with carbon dioxide (CO2),
which is the main cause of global warming. It is now recognized by many scientists and governments that the climate change is the biggest and most serious public health threat to the world2.
In the 2007 Fourth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC)3, it is stated that climate change might lead to some irreversible adverse impacts. If the global average temperature increases 1.5 to 2.5 relative to 1980-1999, approximately 20 to 30% of species are likely to be extinct. Moreover, if the global average temperature increases over 3.5 , approximately 40 to 70% species are in the risk of extinction. In 2011, IPCC released the “Special Report on Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation4”. This report pointed out that the frequency of record-breaking extreme events increased substantially from 1980 through 2011 and is expected to worsen as the climate change continues.
2
Natural Resources Defense Council (NRDC), (2013) “Climate Change Threatens Health”; (2011) “Climate and Your Health: Addressing the Most Serious Health Effects of Climate Change”
3
IPCC (2007) “Fourth Assessment Report (AR4): Climate Change 2007: Synthesis Report”, available at:http://www.ipcc.ch/publications_and_data/publications_ipcc_fourth_assessment_report_synthesis_ report.htm.
4
The 1992 Rio Earth Summit (along with subsequent conferences) and the 1997 Kyoto Protocol have reached the international agreement on finding alternative source of energy and reducing the greenhouse gas emissions. Therefore, the relationship between economic growth and environmental quality captures the attention of researchers around the globe, who seek answers to the question how the environmental quality will evolve along with economic growth. These answers are crucial for policy makers since their objective is to reduce CO2 emissions as well as maximize economic growth.
There has been a long standing debate on the relationship between economic growth and environmental quality based on extensive literatures.
Grossman and Krueger (1991) first found an inverted U-shaped relation between per capita income and environmental quality by using cross-sectional data of 42 countries. The environmental deterioration increases at low levels of income; but beyond a certain level, the environmental deterioration starts to decrease along with economic growth. This inverted U-shaped relationship is commonly referred to as the Environmental Kuznets Curve (EKC), named after the Kuznets Curve describing an inverted U-shaped relationship between inequality and income per capita.
(1997) and Peng and Bao (2006) found an N-shaped relation rather than inverted-U shaped one. Carson (1997) found an inverted J-shaped relationship between environmental pollution and income per capita. Finally Stern and Common (2001) and Stern (2002) found monotonic income-pollution relationship.
Not only the shape of EKC but also the cross-country data used by previous studies have been questioned by researchers. Because environmental pollution is generally increasing in developing countries and decreasing in developed countries, the EKC based on cross-country data might be misleading for simply reflecting two opposite trends in developing and developed countries rather than describing the evolution in a single country over time. Vincent (1997) estimated the income-pollution relation in Malaysia and found that cross-country analysis failed to predict the income-environment relationship in a single-country. De Bruyn et al. (1998) got similar conclusions for other individual countries. The doubt on the reliability of the cross-country data suggests that the EKC needs to be estimated for an individual country.
consequence of the lack of data is that single-country estimation raises problems when constructing out-of-sample predictions. Empirically, this problem can be fixed by cross-country panel data. In cross-country panel data, countries with different income level are corresponding to different emission levels. If we could assume all countries follow the same pattern, then we will observe that there are countries at different stages of EKC. Although a large number of empirical studies have been carried out to estimate EKC for a variety of indicators in different countries, there has been relatively little empirical analysis using spatial panel data to estimate EKC. This is because most studies assume that the pollution emission from one region is unaffected by the emission of neighboring regions. This assumption is not in line with the reality of pollution emission. All these empirical studies have ignored the underlying spatial relationship among geographical areas. Anselin(1988) has pointed out that ignoring spatial dependence might cause model misspecification. Therefore, to estimate the relationship between economic growth and environmental pollution, the inter-regional interaction should be accounted for.
This paper investigates the relationship between China’s CO2 emission and
economic growth for three main reasons. First, since China’s CO2 emissions
have already surpassed those of the United States, it has become the leading emitter of CO2 in the world in 2006. Furthermore, China was responsible for
an important factor in determining the global emission in the future. Second, in the Kyoto Protocol, China and many other developing countries only agreed to reduce their CO2 emissions without binding targets. Estimating the
environment-pollution relationship in a single-country helps policy makers to set a baseline emissions level in the future. Last but not the least, since the Kyoto Protocol has expired in 2012, China’s CO2 emissions will play an
important role in other country’s decision on participating any international agreement in the future.
In this paper, we use the spatial Durbin model (SDM) to investigate the relationship between income and CO2 emission in China. This spatial
Durbin model includes spatial lag effects of the dependent and independent variables to identify influences on local CO2 emission from neighboring
observations. Panel data of 29 Chinese provinces for the period of 1995-2010 are used here. There are two CO2 emission indicators used in this paper,
aggregate CO2 emission and per capita CO2 emissions respectively. It is
important to note that the aggregate CO2 emission and per capita
CO2 emission can yield very different estimation results. For
instance, the aggregate CO2 emission is expected to increase,
however, due to population increases the per capita CO2 emission
will decrease. Therefore we use both CO2 emission indicators in
robustness check using cross-country data is performed to see whether cross-country data helps investigating the EKC for China.
The paper is organized as follows. In the next section, we provide an overview of the literature related to this topic. In section 3 we explain the model and discuss some model specifications. In section 4 we describe the data of this study and section 5 we analyze the estimation results of the model and discuss the estimation result with cross-country data as a robustness check. Finally, section 6 concludes the paper and provides suggestions for future research.
2. Literature review
The relationship between economic growth and environmental quality has been a controversial issue since the end of the 1960s. At one extreme, the economists from “Rome Club” argued that limited environmental resources prevented economic growth from growing forever and would finally lead to a steady state to avoid an ecological catastrophe (Meadows et al. 1972). At the other extreme, Beckerman (1992) and Bhagawati (1993) argued that technological progress could control environmental pollution, providing everlasting economic growth. But all these arguments are purely theoretical without any empirical evidence.
of growth until income reaches a certain level, after that, the environmental degradation starts to decrease. Hettige et al. (1993) used a sample of 80 countries and also found an inverted U-shaped relation between toxic intensity and income per capita. There are other studies that found evidences supporting the inverted U-shaped EKC between different environmental pollution indicators and per capita income (Shafik and Bandyopadhyay, 1992; Panayotou, 1993; and Selden and Song, 1995).
Later Moomaw and Unruh (1997), Torrace and Boyce (1998) and List and Gallet (1999) incorporated a cubic form of income into their model. This cubic function allows them to test whether there are two turning points in the income-environmental pollution relationship. If there is a positive and statistically significant relationship between the cubic formed income and pollution emission, it means that the environmental pollution first increase as income increases and it starts to decrease after the first turning point; however, the decline in pollution is only temporary, after the second turning point, environmental pollution starts to increase again as income grows further. This kind of income-environmental pollution relationship is usually referred as the N-shaped relationship. Panayotou (1997) found the N-shaped relation between the ambient sulfur dioxide level and GDP per capita. He also found the two turning points were around $5000-$15000 and $20000.
of per capita GDP on pollution emission decreases as the income grows; after the inflection point the marginal (per-unit) effects increases as income grows. Selden and Song (1995) explained this inverted J-shaped curve by the Engel’s curve for environmental quality. At low income levels, people focus on food and shelter rather than the environmental quality; at higher income levels, people increase their demand for environmental quality. Carson (1997) found an inverted J-shaped relationship between environmental quality and income per capita.
Most data sets used to estimate EKC in previous studies are formed by the basic units of geographical areas such as countries and cities. But all of them have ignored the underlying spatial relationship among the units because they assume that the pollution emission of one region is unaffected by the emission of another neighboring region. Anselin (1988) pointed out that ignoring the spatial relationship in the data might cause faulty inference testing procedures and possible bias. And Anselin (2001) explained this “scale mismatch” in detail as “neither economic nor physical data collection necessarily matches the spatial scale of the phenomenon under study, such as the geographic extent of a ‘market’ (p.705).
Therefore, if the “pollution haven” hypothesis is true, then a developing country neighboring a developed country is more likely to have higher pollution emissions. Secondly, Keller (2004) found empirical evidences in favor of the hypothesis that geography is a major determinant in international technology diffusion. Since technology is the key factor in the
technological change process that reduced CO2 emissions, we might
also find the spatial patterns in CO2 emissions.
Recently, several EKC studies added spatial correlation into the estimation and found evidence for the existence of environmental pollution spillovers among regions. (Ansuategi, 2003; Rupasingha, 2004; McPherson and Nieswiadomy, 2005; Maddison, 2006 and Maddison, 2007)
Ansuategi (2003) first found the trans-boundary deposition of sulfur as an effective force in the estimation of EKCs. He divided the European countries into four groups according to the standards whether or not more than 65% of sulfur emissions are deposited outside a country’s own borders and whether or not more than 1.5 tons of sulfur per square kilometer is deposited on the country by its neighbors. EKCs are estimated separately for each group and showed significant differences between each other. However, this paper failed to estimate how trans-boundary pollutions influence the estimation of EKCs in the overall sample of European countries.
may contain spatially correlated residuals and thus incorporate a spatial weights matrix in the error terms. The paper found that spatial effects are crucial in understanding the pattern of toxic pollution in US counties.
McPherson and Nieswiadomy (2003) have found evidence of EKCs for threatened bird and mammal species in 113 countries. In their paper, they assumed that the threats to a species in one country might spill over to the species in neighboring countries. It is the first empirical EKC research to take spatial autocorrelation in a direct way. The results proved the existence of spatial autocorrelation and found the pattern of this spatial autocorrelation as “shocks spilling over to surrounding countries” (p.395). Maddison (2006, 2007) estimated EKCs for a variety of pollutants including sulfur dioxide, carbon dioxide, biological oxygen demand, nitrogen oxides, smoke and particulate matter for European countries. He employed both models with a spatial lag and models with a spatial error in his regressions. The results proved that the national pollution emissions are heavily influenced by the emissions of neighboring countries.
income-environment relationship for a single-country over time. However, this assumption cannot be supported by empirical evidences.
Roberts and Grimes (1997) found an inverted U-shaped relationship between national carbon dioxide emissions intensity and GDP per capita based on cross-country data. However, after they divided the sample of countries into three groups which had their income level classed by the World Bank as high, middle and low levels of GDP per capita, they found that the carbon dioxide emission intensity fell steadily among high income countries, but increased among low income countries. Therefore, the EKC based on cross-country data might be misleading for simply reflecting two opposite trends among developing and developed countries rather than describing the evolution in a single-country over time.
developing and developed countries are just statistical works of art; to overcome this problem, EKC should be estimated for a single-country. Similarly, Carson et al. (1997) and De Bruyn et al. (1998) investigated emissions of sulfur dioxide and carbon dioxide in OECD countries between 1960 and 1993. Bruyn et al. (1998) found that air pollutants were positively correlated with growth and Carson et al. (1997) also obtained similar conclusions that single-country data could overcome problems associated with international data.
In conclusion, all single-country studies argue that a single EKC that applies to all countries do not hold for individual countries over time. The doubts on the reliability of the cross-country data suggest that the EKC needs to be estimated for individual country.
3. Model
3.1 The reduced-form model
Most empirical studies are based on reduced-form models in which the relation between income and CO2 emission indicators is a quadratic or cubic
function. The model can be specified as follows:
where Eit is the CO2 emissions indicators, in this paper we use both
aggregate CO2 emissions and per capita CO2 emissions as in previous
studies; and indicate regional and time fixed effects; is per capita income; represent a set of control variables and denotes an
consideration, while i represents a province; subscript k represents the kth control variable. The parameters are to be estimated from data5. With different values of parameters , equation (1) can be made into different forms.
Firstly, if , this means that there is no relationship between income and CO2 emission.
Secondly, If , then there is a monotonically increasing relationship between income and CO2 emission, on the other
hand, if , then we expect a monotonically
decreasing relationship.
Thirdly, if , then we come up with an
inverted-U-shaped relationship, this is the classic EKC curve; if , then we observe a U-shaped relationship. Note that, in these cases, the turning point level of income, where emissions are at
minimum or maximum can be found using the following formula:6
[
]
5
As explained, is the parameter for regional fixed effect and is the parameter for time fixed effect.
6
In this case, equation (1) can be rewritten as:
If , both and ,
reaches its minimum at the point where , thus [ ].
If , then
whereas ,
reaches its maximum at the point where
Finally, the cubic function has two kinds of curves, namely the N-shaped curve and the J-shaped curve. The shape of the curve is determined by the
value of .7
If , we then have an N-shaped curve with two turning points, in which
indicates an N-shaped relationship while
indicates an inverse N-shaped relationship. In the N-shaped function, the formula to calculate these two turning points is:
[ √
]
If , we have a J-shaped curve with only one turning point, in which indicates an inverted J-shaped relation while indicates an J-shaped relationship.
In the J-shaped case, the inflection point can be calculated through the formula below:
[
]
The control variables are usually exogenous and the selection of these control variables is based on previous empirical researches and data availability (Grossman, 1995; Antweiler et al, 2001). In our model we use
7 The derivative of equation (1) is
,
through the quadratic formula, the roots of it are given by: √ .
If , then the cubic function has a local maximum and a local minimum; if , then the cubic function has one inflection point;
if , then there are no critical points.
population density8 (popdenit) and degree of openness of an economy
(openit) which is the exports plus imports as a share of total GDP.
3.2 Spatial panel data models
Tobler (1979) first pointed out the importance of spatial relationship as first law of geography: “everything is related to everything else, but near things are more related than distant things” (p.379). In our case, the CO2 emission
in one province may spill over to neighboring provinces. Three types of spatial panel data models have been used in previous studies: spatial error model (SEM), spatial lag model (also known as spatial autoregressive model, SAR) and spatial Durbin model.
The SEM model assumes there are omitted spatially correlated variables and the spatial relationship is represented as interaction between error terms. Thus the SEM model is written as:
∑
where is the spatial autocorrelation coefficient and w is an element of an N by N spatial weights matrix (to be discussed below). Because the income coefficients do not change when using the SEM model, the turning points remain the same as in the reduced-form model and we can find them through equation (2)-(4).
8
The SAR model assumes that the spatial relationship is captured by interaction effects between the dependent variable. Thus the SAR model used in this paper is written as:
∑
where is the coefficient of the spatially lagged dependent variable. Moving the spatially lagged dependent variable in equation (6) to the left side, we can rewrite the equation in matrix form as:
where covers the error terms, spatial and time fixed effects as well as other control variables in our model. Equation (7) shows that the partial derivatives of CO2 emission with respect to income as well as its quadratic
and cubic forms are no longer single coefficients such as . The matrix of partial derivatives of the CO2 emission in different provinces with
respect to the linear form of income in different provinces at a particular point in time is [ ]
income. The partial derivatives of CO2 emission with respect to the
quadratic form of income can be seen to be
[ ]
And the partial derivative matrix for cubic form of income can be seen to be
[ ]
Because the partial derivatives of CO2 emission with respect to income as
well as its quadratic and cubic forms are N by N matrices, we first need to calculate the direct effects of the normal, quadratic and cubic form of income, and then use the direct effects to find the turning points. LeSage and Pace (2009) defined the direct effect as the average of the diagonal elements of the partial derivative matrix. This direct effect is the average of the diagonal elements of the partial derivative matrices on the right-hand side of equation (8), (9) and (10).
⁄ (11)
In equation (11) N is the number of columns/rows of the spatial weight matrix W (29*29), therefore N in our model equals 29.
The functions for the turning points and inflection point then change into9:
9
[ ] [ √ ] [ ]
The spatial Durbin model adds spatially lagged independent variable to the spatial lagged model and the model is written as:
∑ ∑ ∑ ∑
here also covers the error terms, spatial and time fixed effects as well as other control variables in this model. In the spatial Durbin model, we have both non-spatial coefficients and spatial lagged coefficients of income . The matrix of partial derivatives of the CO2
emission in different provinces with respect to the linear form of income in different provinces at a particular point in time is
[ ]
Similarly, the partial derivative matrix for quadratic and cubic form of income can be seen to be
Therefore the functions for the turning points change into: [ ] [ √ ] [ ] ⁄
3.3 Spatial weights matrix
The spatial weights matrix used in this paper is based on the binary contiguity matrix of Moran (1948) and Geary (1954). By their definition, if two areas share a common border then an element of W will be valued as 1; if there is no common border between two areas a value 0 will be assigned. The diagonal elements of the contiguity matrix are set to zero by definition, because spatial units cannot be viewed as their own neighbor. For example, in our single-country case of 29 provinces, the contiguity matrix has 841 cells of zeros or ones and the diagonal elements are all zeros. The spatial weights matrix is usually row standardized such that each row sums to one. 4. Data
In this paper we use a panel database of 29 Chinese provinces over the years 1995-2010 to examine the relationship between CO2 emissions and income
economic variables for different provinces. Because Chongqing was administrated as the fourth direct-controlled municipalities (the other three are Beijing, Shanghai and Tianjin) in China in 1997, all CO2 and economic
variables of Chongqing were included in the data of Sichuan province. The model uses per capita GDP as the income indicator, and aggregate CO2
emissions and per capita CO2 emissions as environmental pollution
indicators. All provincial socioeconomic data are published in the China Statistical Yearbook and the China Environmental Yearbook.
Considering that per capita GDP data in the China Statistical Yearbook are in current prices, it needs to be converted into a fixed price. We need to adjust the provincial GDP data by individual province price index (Consumer Price Index, setting year 1995 = 100). We use the price index to calculate all the considered variables per capita, e.g. per capita GDP (in US dollar).
In reality, there are no direct measurements for CO2 emissions because there
are a large number of widely dispersed activities which generate CO2 and
directly measuring the emissions of CO2 will be costly. Therefore,
emissions are calculated by fossil fuel consumptions times multipliers based on the average carbon content of the fuel. The State Environmental Protection Administration (SEPA) uses specific emission multipliers for different energy sources and reports the CO2 emission data annually.
Table 1 Descriptive Statistics
Note: All the variables measured in monetary values are converted into 1995 constant price of US dollar. Data source: China Statistic Yearbook (1995-2010) and The State Environmental Protection Administration (SEPA) Annual Report. (1995-2010).
From the table 1 we see great disparities for each variable, not only in the income level but also in CO2 emissions and the degree of trade openness.
Furthermore, the standard deviation is large and the mean is smaller than the median for most variables, this means that the data is highly skewed to the left. This proves the unbalanced development of the economy and industry in China.
Figure 1. 1995 and 2010 per capita CO2 emissions
0 4 8 12 16 Sh an gh ai Ti an jin B e iji n g Sh an xi Li ao n in g Ji lin N in gx ia X in jia n g In n er M o n go lia He b ei He ilo n gj ia n g Ji an gs u Zh ej ia n g Si ch u an G u an gd o n g Sh an d o n g G an su Q in gh ai Sh aa n xi Hu b e i G u iz h o u Hu n an He n an An h u i Ji an gx i Fu jia n Yu n n an G u an gx i Ha in an P e r ca p it a C O2 e m issi on s 1995 2010
Var Explanation Units Obs Mean Std.Dev Min Max
Dependent variable
CO2 Aggregate CO2 emissions Million ton 464 151.89 114.12 6.24 685.91
CO2 PC Per capita CO2 emissions Ton 464 3.912 2.166 0.861 13.219
Independent variable
Y Per capita GDP US dollar 464 1942.5 1871.7 215.1 11012.4
popden Population density Persons/km2 464 391.63 530.9 6.67 3702.62
open
Ratio of export and import to total GDP
Data source: The State Environmental Protection Administration (SEPA) (1995-2010) Annual Report.
Figure 1 shows the ranking of all provinces sorted by per capita CO2
emissions in 1995. We can see there is a significant heterogeneity in the growth rates of per capita CO2 emissions in different provinces. Industrial
provinces, especially iron and steel industrial provinces tend to have higher CO2 emissions.
Figure 2. 1995 and 2010 per capita income (in 1995 US$)
Data source: China Statistic Yearbook (1998-2012)
Figure 2 shows per capita income of all provinces in 1995 and 2010 with 1995 as the base year. Form figure 2 we can see that there is a large growth in income over this period and the growth rates between different provinces also experience huge differences. Shanghai, Beijing and Tianjin are the top three wealthiest provinces in China in 2010, China’s per capita income is now heavily concentrated in the east-coastal provinces such as Shanghai, Guangdong and Zhejiang.
The variation of population density in China is also very large， the minimum population density is only 10% of the maximum population density. If we check the original data, we can observe that most populations are gathered in the east-coastal provinces such as Shandong, Jiangsu province.
Another interesting number in table 1 is the openness level of individual province (open) in one year, since the maximum number is over 1 (175.7%). This means that this province in this year exported and imported more than it actually produced by itself. The reason is very simple; some provinces are export-oriented and import rural materials and export industrial products, such as Ghuangdong and Tianjin province.
Table 2 Correlation Matrix
Aggregate CO2 Per Capita CO2 GDPPC Population Density Trade Openness Aggregate CO2 1
Per Capita CO2 0.3843 1
GDPPC 0.3972 0.6359 1
Population Density 0.1134 0.3183 0.5384 1
Trade Openness 0.1618 0.2718 0.5105 0.6464 1
Table 2 shows the correlation coefficients among all variables. The values of the coefficients represent relationships between the dependent and independent variables and are positive or negative, strong or weak. From the table we can see that per capita CO2 emissions are highly correlated with per
5. Estimation results 5.1 Unit root test results
If the panel data used in this paper exhibits a time trend, then the variable mean and variance in this panel data are no longer time-invariant. Thus the panel data is non-stationary and may lead to spurious regression. In this paper, we confirm the panel stationary by using two different unit root tests, the LLC (Levin-Lin-Chu) test and the IPS (Im-Pesaran-Shin) test. Both of these two tests assume that there is a common unit root across the panel data； more specifically, the LLC test assumes a common (identical) unit root in the null hypothesis while the IPS test allows individual unit root for each cross section.
Table 3 Panel unit root test results
Figures in parentheses are the p- value of t statistics for test statistics.
The results of these two unit root tests are reported in table 3. Column 2 reports the LLC test results while column 3 shows the IPS test results; the p-values are reported in the parenthesis below. From the unit root test results we can conclude that all seven dependent, independent as well as control variables used in our model are panel stationary. Therefore, the panel data used in this paper will not lead to spurious regression.
5.2 Spatial autocorrelation tests
Moran’s I as well as classic and robust Lagrange Multiplier (LM) tests for spatial autocorrelation for the two CO2 emission indicators in both quadratic
and cubic functional form are presented in table 4. Table 4 Spatial autocorrelation tests
Variables
CO2 CO2 PC
Quadratic Cubic Quadratic Cubic
Moran’s I
-0.169*** -0.162*** 0.406*** 0.406***
(-4.81) (-4.58) (12.16) (12.21)
LM spatial lag 56.33*** 53.28*** 43.06*** 40.15***
LM spatial error 73.54*** 69.73*** 73.54*** 69.73***
Robust LM spatial lag 0.195 0.165 3.00* 3.85**
Robust LM spatial error 17.40*** 16.61*** 33.47*** 33.42***
Figures in parentheses are the value of t statistics for regression coefficients.
*, **, *** means absolute value of t statistic is significant at 10%, 5% and 1%, respectively.
For aggregate CO2 emissions, the coefficients of Moran’s I test for quadratic
from low CO2 emission provinces to high CO2 emission provinces. This can
be seen as the traces industrialization in industrial provinces. Through industrialization, these industrial provinces increase their fuel consumption thus increase the aggregate CO2 emissions. However, for per capita CO2
emissions, the coefficients of Moran’s I test are both positive and significant at the one percentage significance level. This indicates that after adjusting for population density, the per capita CO2 emission is actually spilling over
from high CO2 emission provinces to low CO2 emission provinces.
However, the Moran’s I test only proves the existence of spatial autocorrelation in CO2 emissions; it cannot identify the source of this
positive or negative spatial autocorrelation. In order to determine the source and direction of this spatial autocorrelation, the spatial lag and spatial error models should be considered.
The results of classic LM test show that both the no spatial lagged dependent variable hypothesis and the no spatial autocorrelated error term hypothesis are rejected at the 1% significance level. The classic LM test results confirm both the spatial lag and spatial error autocorrelation formulation of our model for two CO2 emission indicators. Therefore, we
still cannot decide which adjustment (spatial lag or spatial error) is more appropriate.
presence of spatial lag correlation; while the robust LM test for spatial lag autocorrelation can test for the spatial lag dependence even with the presence of spatial error correlation.
For aggregate CO2 emissions, the robust LM test statistic for spatial error
correlation is still significant at the 1% significance, whereas the test for a spatial lag becomes insignificant. This indicates that the SEM model is more appropriate for aggregate CO2 emissions. However, for per capita CO2
emissions, both robust LM tests produce significant results for the spatial lag and the spatial error model. This indicates that the spatial Durbin model is more appropriate for per capita CO2 emissions. Since we are investigating
the relationship between economic growth and environmental quality under the assumption of spatial spillover effects, we decided to use the spatial Durbin model for both aggregate and per capita CO2 emissions.
Finally, we have to choose between the fixed effects model and random effects model for time effects (T) and regional effects (P). In order to choose between fixed and random effects models, we carry out the Hausman test which can be found in Lee and Yu (2010). The chi-square statistics are significant at the 1% significance level for three regressions; for the quadratic regression of per capita CO2 emission, the Hausman test is
significant at the 10% level. These results indicate that we should choose the fixed effects model for all estimations10.
5.3 Single-country estimation results
In table 5 we report the results of our regression for aggregate and per capita
10
Table 5 Single-country estimation results
Variables
CO2 CO2 PC
Quadratic Cubic Quadratic Cubic
W*Log(E) 0.077 0.081 0.077 0.081 (1.27) (1.33) (1.27) (1.33) Log(Y) 1.069*** -4.578*** 1.069*** -4.577*** (5.01) (3.30) (5.01)) (3.31) Log(Y)2 -0.042*** 0.749*** -0.042*** 0.748*** (3.06) (3.91) (3.07) (3.91) Log(Y)3 -0.036*** -0.036*** (4.14) (4.14) W*Log(Y) 0.645* 6.140** 0.646* 6.138** (1.71) (2.26) (1.71) (2.25) W*Log(Y)2 -0.057 ** -0.847** -0.057** -0.847** (2.52) (2.29) (2.52) (2.29) W*Log(Y)3 0.037** 0.037** (2.21) (2.20) popden 0.725** 1.226*** -1.578*** -1.076*** (2.00) (3.23) (4.35) (2.84) W*popden 3.423*** 2.947*** 3.600*** 3.132*** (5.39) (4.64) (5.64) (4.94) open 0.0005 -0.0002 0.0005 -0.0002 (0.86) (0.33) (0.86) (0.33) W*open 0.001 0.002** 0.001 0.002** (1.19) (2.14) (1.19) (2.15) Adj-R2 0.212 0.245 0.321 0.348 Hausman Test Chi^2 30.90 30.92 15.16 30.92 P-value 0.000*** 0.001*** 0.086* 0.001***
CO2 emissions using the spatial Durbin model. Further note that both the
quadratic and the cubic functional forms are estimated here.
The first two columns of table 5 provide the estimation results of the quadratic and cubic EKC relationship for aggregate CO2 emissions.
The spatial Durbin model points to significant results for both the quadratic and cubic regression function. This means that the data shows the characteristics of both the cubic and quadratic EKC relationships, in other words, both quadratic and cubic model cannot be rejected for our single-country study. Therefore, both models are discussed below, and then we further discuss which model is better for estimating the EKC.
Looking at the estimated coefficients for these two models, the lagged dependent variables are insignificant in both models, thus we cannot prove that the CO2 emissions in neighboring provinces affect local CO2 emission.
However, the non-lagged independent variables as well as lagged independent variables are significant in both models. This proves that the local income level has significant effects on the local CO2 emission, at the
same time, income level in neighboring provinces also have significant influence on the local CO2 emission.
For quadratic model, the result depicts an inverted U-shaped curve like the classic EKC curve. The aggregate CO2 emission increases with income and
As to the cubic model, the result shows an inverted N-shaped relationship. The CO2 emission first decreases as per capita GDP increases; then it starts
to increase after the first turning point and finally decreases again. The existence of an inverted N-shaped curve seems to imply that at low income levels, the negative impacts of composition and technology effects on industrial waste water emission are greater than the positive scale effects; therefore economic growth has a negative effect on the CO2 emission. After
the income exceeds a certain level, there are no more technology spillovers from developed countries to China, the scale effects become larger than the sum of composition and technology effects, CO2 emission starts to increase
as the income increases. Finally, the negative effects of new technology and industrial structure changes will overcome the positive scale effects; the CO2 emission starts to decrease again.
Furthermore, we can find the turning point of this curve by equation
(20): . This means that the CO2 emission
starts to increase at $97.96 of per capita GDP and decreases again after $17166.3 per capita GDP.
According to these two turning points, China has already passed the first turning point but is still far behind the second one. If we exclude the decreasing part of EKC before $97.96 per capita GDP, the cubic model also shows an inverted U-shaped relationship between CO2 emission and per
As to the control variables, the coefficients for population density and spatial lagged population density are positive and significant at the 1% significance level. The non-lagged coefficients indicate that as the local population increases the local CO2 emission increases as well. And the
spatial lagged population density indicates that the population density in neighboring provinces also have positive effects on local CO2 emissions.
These test results are consistent with the realistic situation that the increase in local population also increases the CO2 emission related to human
activity and provinces with higher population density spill over their human activity to neighboring provinces which cause the increase of CO2 emission
in neighboring provinces.
In terms of goodness of fit, the cubic model seems to have a slightly better fit relative to the quadratic model with the adjusted R square of 0.245 whereas the quadratic model has the adjusted R square of 0.212.
Finally, if we compare the shape of the EKC curve between cubic model and quadratic model, we obtain an inverted U-shaped relationship from the quadratic model and an inverted N-shaped relation from the cubic model. The cubic model outperforms the quadratic model with more significant coefficients and higher explanatory power.
The third and fourth columns of table 5 provide the estimation results of the quadratic and cubic EKC relationship for per capita CO2 emissions.
The per capita CO2 emission estimation results appear to be similar to the
estimation results for aggregate CO2 emissions, both in terms of coefficients
The results are also significant for both the quadratic and cubic regression function which means the data contains characteristics for both the cubic and quadratic EKC relationships.
The lagged dependent variables are still insignificant in both models, thus we cannot prove the influence of per capita CO2 emission across provinces.
Both non-lagged and lagged independent variables are significant which confirm that the local income level has significant effects on the local per capita CO2 emission and the income level in neighboring provinces also
have significant influence.
The estimation results from quadratic model depict an inverted U-shaped curve EKC curve and turning point is found at $20238.1 per capita GDP through equation (19).
As to the cubic model, there is an inverted N-shaped relationship between per capita CO2 emission and per capita GDP. The turning point of this curve
can be calculated by equation (20): . This means the CO2 emission starts to increase after per capital GDP exceeds
$93.04 and decreases again after $18714. Much like the EKC curve for aggregate CO2 emission, the cubic model for pre capita CO2 also shows an
inverted U-shaped relationship with one short-term decrease at the beginning.
One significant difference between aggregate CO2 emission and per capita
CO2 emission is the control variables. In the previous aggregate CO2
However in our model for per capita CO2 emission, the coefficients for
population density become negative whereas the spatial lagged population density remains positive and both coefficients are significant at 1% significance level. These coefficients are consistent with the realistic situation that the increase in local population will reduce the per capita CO2
emission and provinces with higher population density spill over their human activity to neighboring provinces which cause the increase of per capita CO2 emission in neighboring provinces.
In terms of goodness of fit, the cubic model also has a slightly better fit relative to the quadratic model with the adjusted R square of 0.348 while the quadratic model has the adjusted R square of 0.321.
Finally, for per capita CO2 emission, the EKC curve is an inverted U-shaped
relationship from quadratic model and inverted N-shaped curve from cubic model. The cubic model also outperforms the quadratic model with more significant coefficients and higher explanatory power.
On the whole, the results from single-country study appear to be similar between aggregate CO2 emission and per capita CO2 emission. Both
emission indicators describe an inverted N-shaped EKC, or more specifically, an inverted U-shaped EKC with a short decrease at the beginning.
5.4 Cross country estimation results: the robustness check
GDP with one almost out of sample turning point. This EKC curve shows how the CO2 emission will evolve along with economic growth. However,
this prediction has no explicit explanation about time. The EKC describes a long-term relationship between environmental quality and economic growth; it is a trajectory of development for a single economy that goes through different stages of the income level over time. Therefore, single-country study raises problems when constructing out-of-sample predictions.
Empirically, this problem can be fixed by considering cross-country data, in which countries with different income level corresponding to different emission levels are included too. If we could assume all countries included in the data follow the same EKC, then we can observe that there are countries at the beginning of EKC, developing countries at the increasing stage towards the turning point and industrialized countries at the decreasing stage of EKC.
Therefore, we performed one robustness test base on cross-country data to see whether cross-country estimation provides better estimation.
Based on the provincial data used in the single-country estimation, we extend it by including the neighboring countries of China into the cross-country data.
Indonesia. Due to data availability, North Korea, Afghanistan and Myanmar are not included in the cross-country data.
The cross-country data includes 29 Chinese provinces and 18 neighboring countries during the period 1995-2010. The data on CO2 emission is
obtained from the International Energy Agency and data on per capita GDP is obtained from the World Bank.
The spatial weights matrix used in cross-country estimation is still based on simple binary adjacency matrix in which countries that share a common border with provinces or other countries are assigned with a value 1, while no common border are assigned of value 0. For maritime neighbors, we assign the value 1 to the nearest province, and the rest of provinces are assigned of value 0.
The cross-country estimation results are present in table 6, same as single-country estimation, we estimate the EKC for aggregate and per capita CO2 emissions using spatial Durbin model in both quadratic and cubic
functional forms.
The first two columns of table 6 provide the cross-country estimation results of the quadratic and cubic EKC relationship for aggregate CO2 emissions.
Table 6 Cross-country estimation results
Variables
CO2 CO2 PC
Quadratic Cubic Quadratic Cubic
W*Log(E) 0.523*** 0.531*** 0.511*** 0.514*** (13.24) (13.58) (12.86) (12.98) Log(Y) 0.795*** -1.814*** 0.860*** -1.709*** (7.03) (2.93) (7.62)) (2.77) Log(Y)2 -0.023*** 0.317*** -0.027*** 0.308*** (3.16) (3.97) (3.69) (3.86) Log(Y)3 -0.015*** -0.015*** (4.28) (4.22) W*Log(Y) -0.502** 1.181 -0.553*** 1.046 (2.45) (0.78) (2.71) (0.69) W*Log(Y)2 0.011 -0.120 0.014 -0.185 (0.83) (1.00) (1.08) (0.93) W*Log(Y)3 0.009 0.008 (0.99) (0.94) popden 0.107 0.098 -0.863*** -0.871*** (0.80) (0.74) (6.45) (6.56) W*popden 0.335 0.527 0.845** 1.041*** (0.98) (1.50) (2.46) (2.97) open 0.002*** 0.002*** 0.002*** 0.002*** (5.73) (4.83) (5.51) (4.63) W*open -0.002** -0.001 -0.002* -0.001 (1.88) (1.08) (1.66) (0.88) Adj-R2 0.236 0.253 0.311 0.326 Hausman Test Chi^2 23.78 16.00 30.87 26.37 P-value 0.005*** 0.014** 0.000*** 0.006***
The lagged dependent variables become significant in both models, thus the cross-country data proves the influence of CO2 emission between
neighboring provinces and countries. The positive coefficients mean that a 1% higher amount of aggregate CO2 emission in neighboring regions leads to
approximately 0.5% higher amount of local aggregate CO2 emission. This
shows that countries with higher CO2 emission are more likely to transfer
their high CO2-emission industries to those low CO2-emission neighboring
countries. In comparison, the single-country data yields insignificant results on the lagged dependent variables. One possible explanation for this might be that the free capital movement and labor migration in a single country cause the CO2 emission in the high emission provinces continue to rise
rather than transfer to neighboring provinces.
The non-lagged independent variables remain significant in both models, indicating that local income has significant effects on local CO2 emissions.
However, some of the lagged independent variables become insignificant, as a result of which we cannot prove the influence of income level in neighboring country on the local CO2 emission. This also could be the
consequence of the limited capital movement between countries; the capital cannot move freely across different countries and therefore the income level in neighboring country has no significant effect on the local CO2 emission.
relationship and the turning points of this curve are .
Unlike the turning points found in single-country estimation, the turning points here seem to be quite different between quadratic and cubic models. The third and fourth columns of table 6 provide the cross-country estimation results of the quadratic and cubic model for per capita CO2 emissions.
The lagged dependent variable also becomes significant in both models with positive value of 0.5; while on the other hand, the lagged independent variables become insignificant.
The turning point of quadratic model is found at $74182.66 per capita GDP; as to the cubic model, the turning points are . The turning points also show great differences between quadratic and cubic model for per capita CO2 emissions.
We provide a brief review of empirical findings on the turning point in the specific case of CO2 emission as a reference to evaluate the turning points
1995 dollar whereas the turning points in cross-country data are found between $29000 and $100000 in per capita 1995 dollars. It is obvious that the turning points which found in single-country data are much closer to the previous findings.
Generally speaking, the results from cross-country study appear to be quite different between aggregate CO2 emission and per capita CO2 emission. The
lagged dependent variable becomes significant, while the lagged independent variables become insignificant, the control variables become insignificant as well. In terms of goodness of fit, the adjusted R square shows no significant improvement comparing to the previous single-country estimation.
Thus, the conclusion we may draw from our robustness check is that the cross-country data does not provide better estimation results than the single-country data.
6. Conclusion
The Environmental Kuznets Curve has been studied both deeply and widely in the past decades. Although the theoretical analysis gives the prediction of an inverted U-shaped curve for the income-environment relationship, researchers criticized the assumption of spatially independence in previous studies. In reality, the CO2 emission will not only affect the environmental
region. Therefore, to analyze the relationship between economic growth and environmental pollution, spatial effects cannot be ignored.
In this paper, we estimate the relationship between income and CO2
emissions in China based on data of 29 Chinese provinces for the period of 1995-2010. Two different shapes of EKC, the classic inverted U-shaped and inverted N-shaped curve, are considered in this paper, relating per capita GDP to two CO2 emission indicators including aggregate CO2 emission and
per capita CO2 emission.
For the analysis, we calculate Moran’s I test, as well as classic and robust LM tests for spatial lag correlation and spatial error correlation. The results of these tests suggest that both spatial lag and spatial error autocorrelation are statistically significant for per capita CO2 emission. The test results
show significant spatial error dependence in aggregate CO2 emissions. By
considering the spatial Durbin model can test for both spatial lag and error autocorrelation; therefore we employed the spatial Durbin model for all the pollution indicators.
We consider both quadratic and cubic models in the analysis of the income-CO2 emission relationship. The estimation results suggest that the
cubic spatial Durbin model is better for estimating the EKC for both CO2
indicators describe an inverted U-shaped EKC with a short decrease at the beginning.
Then we perform a robustness check using cross-country data to see whether cross-country data provides better estimation results for out-of-sample prediction. We find significant coefficients for the spatial lagged dependent variable, whereas the spatial lagged independent variables become insignificant. Differences are more pronounced in the cross-country models. Therefore, we conclude that cross-country estimations cannot provide better estimation than single-country estimation.
There are still some limitations of this study concerning the econometric techniques. As mentioned before, the panel data used in this paper only covers a short time period. This shortcoming of lacking long time-series data might lead to estimation results that only describe part of the income-environment relation. In order to understand the whole relationship over time, long time-series data should be collected in future studies.
The other limitation of this study is that the model only estimates the influence of income on CO2 emission and there is no direction of causality
in this model. In reality, CO2 may have a feedback effects on income growth.
Appendix A. Geographic information of Chinese provinces
Geographic Information of Chinese Provinces
No. Province Neighboring provinces No. Province Neighboring provinces
1 Beijing 2,3 17 Hubei 12,14,16,18,22,27
2 Tianjin 1,3 18 Hunan 14,17,19,20,22,24
3 Hebei 1,2,4,5,6,15,16 19 Guangdong 13,14,18,20,21
4 Shanxi 3,5,16,27 20 Guangxi 18,19,24,25
5 Inner Mongolia 3,4,6,7,8,27,28,30 21 Hainan 19
6 Liaoning 3,5,7 22 Chongqing 17,18,23,24,27 7 Jilin 5,6,8 23 Sichuan 22,24,25,26,27,28,29 8 Heilongjiang 5,7 24 Guizhou 18,20,22,23,25 9 Shanghai 10,11 25 Yunnan 20,23,24,26 10 Jiangsu 9,11,12,15 26 Tibet 23,25,29,31 11 Zhejiang 9,10,12,13,14 27 Shaanxi 4,5,16,17,22,23,28 12 Anhui 10,11,14,15,16,17 28 Gansu 5,23,27,29,30,31 13 Fujian 11,14,19 29 Qinghai 23,26,28,31 14 Jiangxi 11,12,13,17,18,19 30 Ningxia 5,27,28 15 Shandong 3,10,12,16 31 Xinjiang 26,28,29 16 Henan 3,4,12,15,17,27
References
Anderson, J., and Van Wincoop, E., (2004), “Trade costs”, The Journal of Economic Literature, Vol.42, pp.691-751.
Anselin, L., (1988), “Model validation in spatial econometrics: a review and evaluation of alternative procedures”, Intentional Regional Science Review, Vol.65, pp.11-34.
Anselin, L., (2001), “Spatial effects in econometric practice in environmental and resource economics”, American Journal of Agricultural Economics, Vol.83, pp.705–710.
Ansuategi, A., (2003), “Economic growth and trans-boundary pollution in Europe: an empirical analysis”, Environmental and Resource Economics, vol. 26, pp. 305–328.
Antweiler, W., Copeland, R., and Taylor, S., (2001). “Is free trade good for the environment?”, American Economic Review, Vol. 91, pp.877-908. Bhagawati J., (1993), “The case for free trade”, Scientific American, pp142-149.
Beckerman W., (1992), “Economic growth and the environment: whose growth? Whose environment?”, World Development, Vol.20, pp.481- 496. Carson. R., Jeon, Y., and McCubbin, D., (1997), “The relationship between airpollution emissions and income: US data”, Environment and Development Economics, Vol. 2, pp.433-450.
Cole, M., Rayner, A., and Bates, J., (1997), “The environmental Kuznets curve: an empirical analysis”, Environment and Development Economics, Vol.2, pp. 401-416.
De Bruyn, S., Van den Bergh, J., and Opschoor, J., (1998), “Economic growth and emissions: reconsidering the empirical basis of environmental Kuznets curve”, Ecological Economics, Vol.25, pp.161-175.
Elhorst, J., (2003), “Specification and estimation of spatial panel data models”, International Regional Science Review, Vol.26, pp.244–268. Geary, R., (1954), “The contiguity ratio and statistical mapping”, The Incorporated Statistician, Vol.5, pp.115– 145.
Grossman, G and Krueger, A., (1991), “Environmental impacts of the North American Free Trade Agreement”, NBER, working paper, no. 3914.
Grossman, G and Krueger, A., (1995), “Economic growth and the environment”, Quarterly Journal of Economics, Vol. 110(2), pp. 353 –377. Hettige, H., Lucas, R., and Wheeler, D., (1992), “The toxic intensity of industrial production: Global patterns, trends and trade policy”, American Economic Review, Vol.82, pp.478–481.
Hill, J., and Magnani, E., (2002), “An exploration of the conceptual and empirical basis of the environmental Kuznets curve”, Australian Economic Papers, Vol.41, pp. 239–254.
Keller, W., (2004), “International technology diffusion”, The Journal of Economic Literature, Vol.42, pp.752–782.
LeSage, P., and Pace, R., (2009) “Introduction to Spatial Econometrics”, Boca Raton, FL:CRC Press Taylor & Francis Group.
List, J., and Gallet, C., (1999), “The environmental Kuznets curve: Does one size fit all? Ecological Economics”, Vol.31, pp. 409–424.
Ma. S and Li. G., (2006), “The Kuznets Curve of Relationship between Economic Growth and Environmental Pollution”, Statistical Research, Vol. 23 (8), pp. 37-40. (Chinese).
Maddison, D., (2006), ‘Environmental Kuznets curves: a spatial econometric approach”, Journal of Environmental Economics and Management, vol. 51, pp. 218–230.
Meadows. D., Randers. J and Behrens. W., (1972), “The Limits to Growth”, New York: Universe Books. ISBN 0-87663-165-0.
Moomaw, W., and Unruh, G., (1997), “Are environmental Kuznets curves misleading us? The case of CO2 emissions”, Environment and Development Economics , Vol.2, pp. 451–463.
Moran, P., (1948), “The interpretation of statistical maps”, Journal of the Royal Statistical Society, Series B10, pp. 243– 251.
Panayotou, T., (1993), “Empirical Tests and Policy Analysis of Environmental Degradation at Different Stages of Economic Development”, Working Paper WP238, Technology and Employment Programme, Geneva: International Labor Office.
Panayotou, T. (1997). “Demystifying the Environmental Kuznets Curve: Turning a Black Box into a Policy Tool”, Environment and Development Economics, Vol.2, pp. 465-484.
Peng, S. and Bao, Q (2006). “Economic Growth and Environmental Pollution: An Empirical Test for the Environmental Kuznets Curve Hypothesis in China”, Research on Financial and Economic Issues, Vol. 273, pp.3-17. Roberts, J., and Grimes, P., (1997), “Carbon intensity and economic development 1962-91: a brief exploration of the environmental Kuznets curve”, World Development, Vol.25, n.2,pp.191-198.
Roy, N. and Van Kooten, G., (2004), “Another Look at the Income Elasticity of Non-Point Source Air Pollutants: a Semi-Parametric Approach”, Economics Letters, Vol. 85, pp.17–22.
Rupasingha, A., Goetz, J., Debertin, L., and Pagoulatos, A., (2004), “The environmental Kuznets curve for US counties: a spatial econometric analysis with extensions”, Papers in Regional Science, vol. 83, pp.407–424. Selden, T, and Song, D., (1994). “Environmental quality and Development: Is there a Kuznets Curve for Air Pollution Emissions”, Journal of Environmental Economics and Management, Vol.27, pp.147-162.
Shafik, N., Bandyopadhyay, S., (1992), “Economic growth and environmental quality: time series and cross-country evidence.” Background paper for the World Development Report 1992. World Bank, Washington, D.C.
Schmalensee, R., Stoker, M., and Judson, A., (1998), “World carbon dioxide emissions: 1950–2050”, Review of Economics and Statistics, Vol.80, pp. 15–27.
Stern, D., (2002) “Explaining changes in global sulfur emissions: An econometric decomposition approach”. Ecological Economics, Vol. 42, pp.201–220.
Stern, D., and Common, M., (2001), “Is there an environmental Kuznets curve for sulfur?” , Journal of Environmental Economics and Management., Vol.41, pp.162–178.
Tobler,W., (1979), “Cellular geography”, Philosophy in Geography. Reidel, Dordrecht, pp. 379–386.
Torras, M., and Boyce. J., (1998), “Income, inequality and pollution: A reassessment of the environmental Kuznets curve”. Ecological Economics Vol.25, pp. 147–160.
