Content uploaded by Perry Sadorsky
Author content
All content in this area was uploaded by Perry Sadorsky on Jul 17, 2021
Content may be subject to copyright.
The effect of urbanization on CO
2
emissions in emerging economies
Perry Sadorsky ⁎
Schulich School of Business, York University, 4700 Keele Street, Toronto, Ontario M3J 1P3, Canada
abstractarticle info
Article history:
Received 14 December 2012
Received in revised form 29 October 2013
Accepted 10 November 2013
Available online 21 November 2013
JEL classification:
Q43
R11
O14
Keywords:
CO
2
emissions
Emerging economies
Energy intensity
Urbanization
The theories of ecological modernization and urban environmental transition both recognize that urbanization
can have positive and negative impacts on the natural environment with the net effect being hard to determine
a priori. This study uses recently developed panel regression techniques that allow for heterogeneous slope
coefficients and cross-section dependence to model the impact that urbanization has on CO
2
emissions for a
panel of emerging economies. Theestimated contemporaneouscoefficients on the energyintensity and affluence
variablesare positive, statistically significant and fairly similar across different estimationtechniques. By compar-
ison, the estimated contemporaneous coefficient on the urbanization variable is sensitive to the estimation
technique. In most specifications, the estimated coefficient on the urbanization variable is positive but statistical-
ly insignificant. The implications of these results for sustainable development policy are discussed.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
The year 2010 marked a milestone in urbanization as this was the
year that world urbanization passed 50%.
1
While urbanization in devel-
oped countries continues to increase, developing countries are expected
to experience the greatest increase in urbanization. For example, the
United Nations Population Division (2007) predicts that in the year
2020, urbanization in the less developed regions of the world will pass
50%. Furthermore, it is expected that urbanization in the less developed
regions of the world will more than triple, from 18% in 1950 to 67% in
2050. Changes in urbanization can affect economic growth, energy use
and carbon dioxide emissions (CO
2
). If urbanization has a significant
impact on carbon dioxide emissions then this will have implications
for sustainable development and climate change policies.
If urbanization is found to have a positive and statistically significant
impact on CO
2
emissions then this can affect forecasting models and
climate change policy. Forecasting models of CO
2
emissions that fail to
take into account the impact of urbanization on CO
2
emissions will
under forecast carbon dioxide emissions. Energy and environmental
policies that omit the impact of urbanization on CO
2
emissions will
likely lead to inaccurate outcomes making sustainable development
objectives more difficult to achieve. If urbanization is found to have a
negative and statistically significant impact on CO
2
emissions then this
will make sustainable development objectives easier to achieve.
The theories of ecological modernization and urban environmental
transition both recognize that urbanization can have positive and nega-
tive impacts on the natural environment with the net effect being hard
to determine a priori. If urbanization is found to have a statistically
insignificant impact on CO
2
emissions then urbanization will have no
meaningful impact onCO
2
emissions. This is consistent withthe positive
and negative effects of urbanization on CO
2
emissions canceling each
other out.
This paper makes several important contributions to the literature.
First, the relationship between urbanization and carbon dioxide emis-
sions has been studied by a number of authors (eg. Cole and Neumayer,
2004; Hossain, 2011; Liddle and Lung, 2010; Martinez-Zarzoso and
Maruotti, 2011; Parikh and Shukla, 1995; Poumanyvong and Kaneko,
2010; Sharma, 2011; York et al., 2003) but most of this research uses a
static model applied to a panel data set. A panel data set offers advantages
over a cross-section data set by including a time dimension. This increases
the number of observations and allows for variation in both the cross-
section and time dimension. Static models cannot, however, capture dy-
namic relationships. Dynamic models are advantageous because both
long-run and short-run impacts are modeled. This paper uses a static
and dynamic framework to model the impact of urbanization on carbon
dioxide emissions in order to compare the results obtained by the two dif-
ferent models.
Second, previous studies have mostly assumed that the impact of
urbanization on carbon dioxide emissions is homogeneous across
countries. This is a very strong assumption to make and one that is
unlikely to hold across a large grouping of countries. In this paper
panel regression models are estimated using recently developed
Energy Economics 41 (2014) 147–153
⁎Tel.: +1 416 736 5067; fax: +1 416 736 5687.
E-mail address: psadorsk@schulich.yorku.ca.
1
Data sourced from http://esa.un.org/unup/.
0140-9883/$ –see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.eneco.2013.11.007
Contents lists available at ScienceDirect
Energy Economics
journal homepage: www.elsevier.com/locate/eneco
techniques that allow for heterogeneity in the estimation of the slope
coefficients and cross-section dependence. If panel data exhibits cross-
section dependence, estimating models with homogeneous slope coeffi-
cients (as in the case of pooled OLS, fixed effects, or GMM) may yield
estimated coefficients that are biased (Andrews, 2005). In order to
account for cross-section dependence, models are estimated using the
mean group (MG) estimator of Pesaran and Smith (1995),Pesaran's
(2006) Common Correlated Effects Mean Group (CCEMG) estimator,
and the Augmented Mean Group (AMG) estimator of Eberhardt and
Teal (2010) and Bond and Eberhardt (2009).
The purpose of this paper is to investigate theimpact of urbanization
on CO
2
emissions for a panel of 16 emerging economies. Empirical
models areestimated using heterogeneous panel regression techniques.
The following sections of the paper set out the contextual material, the
empirical model, data, empirical results, implications, and conclusions.
2. A brief review of the literature
“Although urbanization is often discussed in the context of economic
modernization, it is a demographic indicator that increases urban
density and transforms the organization of human behavior, thereby
influencing household energy use patterns (Barnes et al., 2005;
Poumanyvong and Kaneko, 2010).”
According to Poumanyvong and Kaneko (2010), the existing
literature points to three theories (ecological modernization, urban
environmental transition and compact city theories) that are useful
for explaining how urbanization can impact the natural environ-
ment. The theory of ecological modernization details how urbaniza-
tion is a process of social transformation that is an important
indicator of modernization. As societies move from low to middle
stages of development, environmental problems may increase be-
cause in these stages of development, economic growth takes prior-
ity over environmental sustainability. As societies continue to evolve
to higher stages of development, environmental damages become
more important and societies seek ways to make their societies
more environmentally sustainable. The damaging impact of econ om-
ic growth on the environment may be reduced by technological
innovation, urbanization, and a shift from a manufacturing based
economy to a service based economy (Crenshaw and Jenkins, 1996;
Gouldson and Murphy, 1997; Mol and Spaargaren, 2000).
The theory of urban environmental transition links environmental
issues with urban evolution at the city level (McGranahan et al., 2001).
In modern history, cities often become wealthier by increasing their
manufacturing base and this can lead to industrial pollution problems
that impact the land, air and water. As cities continue to become wealth-
ier, industrial pollution problems may be lessened via environmental
regulations, technological innovation, or changes in economic sector
composition. Wealthier cities create wealthier residents and wealthier
residents demand more energy intensive products and this puts further
stress on the environment. The net impact of the wealth effect is difficult
to determine a priori.
The compact city theory mostly focuses on the benefits of
increased urbanization. Higher urban density helps to facilitate
economies of scale for public infrastructure (eg. public transporta-
tion, water supply, electricity production, schools, hospitals) and
these economies of scale lead to lower environmental damages
(Burton, 2000; Capello and Camagni, 2000; Jenks et al., 1996;
Newman and Kenworthy, 1989).
The empirical relationship between urbanization and CO
2
emis-
sions has been studied by a number of authors. In one of the earliest
studies, Parikh and Shukla (1995) use a data set of 83 developed
and developing countries for the year 1986 to investigate the
impact of urbanization on energy use and toxic emissions. They
find that urbanization has a positive and significant impact on CO
2
emissions, CH
4
emissions, and CFC emissions. In particular, they
find a CO
2
emissions elasticity of urbanization of 0.036. York et al.
(2003) use a cross section of 137 countries to test for a relationship
between urbanization and CO
2
emissions. They find evidence that
increases in urbanization lead to increases in CO
2
emissions.
Cole and Neumayer (2004) study 86 countries over the period
1975to1998andfind a positive relation between urbanization
and CO
2
emissions. More specifically, they find a 10% increase in ur-
banization increases CO
2
emissions by 7%. For developing coun-
tries, Fan et al. (2006) find a negative relationship between
urbanization and carbon dioxide emissions. Liddle and Lung
(2010), using a panel data set of 17 developed countries followed
over 10, 5 year periods, find a positive but insignificant impact of
urbanizationoncarbondioxideemissionswhenaggregatecarbon
dioxide emissions are used as the dependent variable. Urbanization
has a positive and statistically significant impact on carbon dioxide
emissions when carbon dioxide from transport is used as the de-
pendent variable.
Poumanyvong and Kaneko (2010) use a Stochastic Impacts by Re-
gression on Population, Affluence and Technology (STIRPAT) model
to investigate the impact of urbanization on CO
2
emissions in a
panel of 99 countries over the period 1975 to 2005. A variety of
panel regression techniques are used but the empirical approaches
are all static in nature. They find that urbanization has a positive
and significant impact on CO
2
emissions for each income group
but its impact is greatest for the middle income group of countries.
For all income groups, the estimated coefficient on urbanization
varies between 0.350 and 0.506. For low income groups, the esti-
mated coefficient on urbanization varies between 0.430 and
0.615. For middle income groups, the estimated coefficient on ur-
banization varies between 0.210 and 0.512. For high income
groups, the estimated coefficientonurbanizationvariesbetween
0.041 and 0.358.
Martinez-Zarzoso and Maruotti (2011) use a STIRPAT model in a
panel of 88 developing countries over the period 1975 to 2003. Empiri-
cal results estimated using the complete panel of countries supports an
inverted U shaped relationship between urbanization and CO2 emis-
sions. From their results reported in Table 2, the estimated coefficient
on the urbanization variable is statistically significant at the 10% level
and ranges in value between 0.506 and 1.329. The estimated coefficient
on the squared urban term varies between −0.089 and −0.209. A novel
feature of their paper is the use of a semi-parametric mixture model to
classify countries into groupings based on CO
2
emissions. Three groups
of countries are found and in two of the country grouping, the estimated
coefficient on urbanization is statistically significant at 10% and varies
between 0.445 and 0.801. This approach is useful in showing that the
impact of urbanization on CO2 emissions differs across the country
groups providing support to the idea that panel regression models esti-
mated assuming homogenous slope coefficients may give rise to miss-
leading results. A one period lag of CO
2
emissions is included in the re-
gression models to account for dynamics.
Sharma (2011) studies a large panel of 69 countries (including high
income, middle income and low income countries) and finds that ur-
banization does have a negative and statistically significant impact on
carbon emissions for the global panel. For the global panel, a 1% increase
in urbanization decreases CO
2
emissions by 0.7%. Urbanization has a
negative but insignificant impact on carbon emissions in the low
income, middle income and high income panels. Hossain (2011) stud-
ied 9 newly industrialized countries (Brazil, China, India, Malaysia,
Mexico, Philippines, South Africa, Thailand, and Turkey) over the period
1971 to 2007 and found the existence of a long-run cointegrating vector
between CO
2
emissions, output, energy consumption, trade openness
and urbanization. In the long-run, a 1% increase in: energy use increases
CO
2
emissions by 1.2%, income increases CO
2
emissions by 0.2%, and
urbanization lowers CO
2
emissions by 0.6%.
In summary, the existing empirical literature is inconclusive on the
impact that urbanization has on CO
2
emissions. The magnitude and
sign of this effect depends upon the data set and estimation technique.
148 P. Sadorsky / Energy Economics 41 (2014) 147–153
3. Empirical model
Following other authors, a Stochastic Impacts by Regression on Popu-
lation, Affluence and Technology (STIRPAT) model is used to investigate
the relationship between urbanization and CO2 emissions (eg. Liddle
and Lung, 2010; Martinez-Zarzoso and Maruotti, 2011; Poumanyvong
and Kaneko, 2010). The STIRPAT model is based on the Influence, Popu-
lation, Affluence, and Technology (IPAT) model developed by Ehrlich
and Holdren (1971). The IPAT model relates environmental impact to
population, affluence (consumption per capita), and technology.
I¼PxAxT ð1Þ
The IPAT model has been criticized as 1) being primarily a mathemat-
ical equation or an accounting identity which is not suitable for hypothe-
sis testing, and 2) assuming a rigid proportionality between the variables.
In response, Dietz and Rosa (1997) proposed a stochastic version of IPAT.
Iit ¼aiPb
it Ac
it Td
it eit ð2Þ
In Eq. (2), countries are denoted by the subscript i (i = 1,…,N) and
the subscript t (t = 1,…,T) denotes the time period. Country specific
effects are included through a
i
and ε
it
represents the random error
term.TakingnaturallogarithmsofEq.(2) provides a convenient linear
specification for panel estimation. When all variables are expressed in
natural logarithms the estimated coefficients can be interpreted as
elasticities.
ln Iit
ðÞ
¼bln Pit
ðÞ
þcln Ait
ðÞ
þdln Tit
ðÞ
þνiþεit ð3Þ
Country specific effects are included through ν
i
and ε
it
represents the
random error term. The STIRPAT model can easily accommodate addi-
tional explanatory variables and in this paper, model (3) is augmented
with urbanization (U). The augmented model is
ln Iit
ðÞ¼bln Pit
ðÞþcln Ait
ðÞþdln Tit
ðÞþfln Uit
ðÞþνiþεit ð4Þ
When it comes to estimating Eq. (4), a distinction can be made be-
tween models with homogeneous slope coefficients and models with
heterogeneous slope coefficients. If the assumption of homogeneous
slope coefficients is made then these models can be estimated using
standard panel regression techniques like pooled OLS (POLS) and vari-
ous fixed effects (FE), random effects (RE), or GMM specifications.
Models with heterogeneous slope coefficients can be estimated using
mean group (MG) estimators (eg. Pesaran and Smith, 1995; Pesaran,
1997) or variants on mean group estimators. Estimating panel models
with heterogeneous slope coefficients is currently an active area
of econometrics (eg. Coakley et al. 2006; Eberhardt and Teal, 2011;
Eberhardt et. al., 2013).
Eq. (4) can be transformed into a dynamic model by adding lags of
the dependent and each independent variable to the right-hand side
of Eq. (4). This formulation is known as an autoregressive distributed
lag (ARDL) model. Model selection criteria like SIC can beused to deter-
mine the appropriate lag lengths. Dynamic models are advantageous
over static models because dynamic models facilitate the calculation
of both short-run and long-run elasticities.
Models are estimated using the mean group (MG) estimator of
Pesaran and Smith (1995),Pesaran's (2006) Common Correlated Effects
Mean Group (CCEMG) estimator, and the Augmented Mean Group
(AMG) estimator (Bond and Eberhardt, 2009; Eberhardt and Teal,
2010). The mean group (MG) approach incorporates heterogeneity
across countries by allowing all slope coefficients and error variances to
vary across panels (or countries as is the case in this paper) (Pesaran
and Smith, 1995). The MG approach applies OLS to each panel/country
to obtain panel specific slope coefficients and then averages the panel
specific coefficients. For large T and N the MG estimator is consistent.
For inference on the long-run parameters, Pesaran (1997) and Pesaran
and Shin (1999) show that including the appropriate number of lags in
the order of the ARDL model can simultaneously correct for the problem
of residual serial correlation and endogenous regressors. The MG estima-
tor does not, however, incorporate any information on common factors
that may be present in the panel data set. Pesaran's (2006) Common Cor-
related Effects Mean Group (CCEMG) estimator includes cross-sectional
dependence and heterogeneous slope coefficients. The cross-sectional
dependence is modeled using cross-sectional averages of the dependent
and independent variables. These cross-sectional averages account for
the unobserved common factors. The unobservable common factors
may be nonlinear or non-stationary. The slope coefficients are averaged
across panel members. The CCEMG estimators are very robust to struc-
tural breaks, lack of cointegration and certain serial correlation
(Kapetanios et al., 2011). The Augmented Mean Group (AMG) estimator
is an alternative to the Pesaran (2006) CCEMG estimator (Bond and
Eberhardt, 2009; Eberhardt and Teal, 2010). In the CCEMG approach
the set of unobservable common factors is treated as a nuisance. In the
AMG approach the set of unobservable common factors are treated as a
common dynamic processes that, depending upon the context, may
have useful interpretations.
4. Data
The data set is an unbalanced panel of 16 emerging countries follow-
ed over the years 1971–2009. The countries are selected from the MSCI
classification of emerging markets.
2
MSCI classifies countries as emerg-
ing if their financial markets meet certain size, liquidity, and market ac-
cessibility criteria.
3
The list of countries includes: Brazil, Chile, China,
Colombia, Egypt, Hungary, India, Indonesia, Korea (South), Malaysia,
Mexico, Morocco, Philippines, South Africa, Thailand, and Turkey.
4
In the empirical analysis, CO2 is the natural logarithm of CO
2
emis-
sions (metric tons of carbon dioxide emissions), Affluence is the natural
log of real per capita GDP (GDP per capita, in constant 2005 US dollars),
Population is the natural log of total population,and Urban is the natural
log of urbanization (measured by the fraction of the population living in
urban areas). Following Liddle and Lung (2010),Poumanyvong and
Kaneko (2010),andMartinez-Zarzoso and Maruotti (2011) the technol-
ogy variable in Eq. (4) is measured using energy intensity. Intensity is
the natural log of total energy use per dollar of GDP (energy use in kg
of oil equivalent relative to GDP, constant 2005 US dollars). The data
Table 1
Average annual growth rates in percent (1971–2009).
Country CO2 Affluence Population Urban intensity
Brazil 3.500 2.047 1.796 1.045 −0.407
Chile 2.552 2.615 1.470 0.422 −0.755
China 5.894 7.420 1.248 2.596 −4.110
Colombia 2.357 1.918 1.958 0.795 −1.721
Egypt, Arab Rep. 5.895 3.247 1.918 0.066 0.622
Hungary −0.932 2.142 −0.079 0.333 −1.219
India 5.941 3.189 1.955 1.121 −1.389
Indonesia 6.499 3.860 1.880 2.710 −1.150
Korea, Rep. 5.765 5.230 1.108 1.815 0.554
Malaysia 6.689 3.756 2.398 1.934 0.251
Mexico 3.497 1.293 2.018 0.699 0.388
Morocco 4.874 2.249 1.732 1.259 0.864
Philippines 2.609 1.214 2.417 0.990 −1.188
South Africa 3.086 0.449 2.060 0.628 0.576
Thailand 7.364 4.147 1.503 1.206 −0.254
Turkey 4.806 2.154 1.839 1.542 0.313
Total 4.400 2.933 1.701 1.197 −0.539
2
http://www.msci.com/products/indices/tools/index_country_membership/emerging_
markets.html
3
http://www.msci.com/products/indices/market_classification.html
4
MSCI lists 21 countries as emerging. Due to data limitations, Czech Republic, Peru,
Poland, Russia, and Taiwan were omitted from the analysis.
149P. Sadorsky / Energy Economics 41 (2014) 147–153
are obtained from the World Bank (2013) World Development Indica-
tors online data base.
5
Average annual growth rates for the variables are shown in Table 1.
The average annual growth rate in CO
2
emissions ranges from a high of
7.36% in Thailand to a low of −0.93% in Hungary. Thailand, Malaysia,
and Indonesia each have average annual growth rates in CO
2
emissions
greater than 6%. The average annual growth rate in affluence (per capita
GDP) ranges from a high of 7.42% (China) to low of 0.45%(South Africa).
China, Korea, and Thailand each experienced average annual growth
rates in affluence greater that 4%. Onaverage, the fastest growth in pop-
ulation occurred in the Philippines, Malaysia, South Africa, and Mexico.
Each of these countries has average annual growth rates in population
greater than 2%. Hungary actuallyexperienced a negative average annu-
al growth rate in population. Indonesia and China each recorded aver-
age annual increases in urbanization greater than 2%. Egypt had the
lowest average annual growth rate in urbanization (0.67%). Average
annual growth rates in energy intensity were the highest in Morocco,
Egypt, South Africa and Korea (each larger than 0.50%). Negative aver-
age annual growth rates in energy intensity were recorded for about
half of the countries with China having the lowest average annual
growth rate in energy intensity (−4.11%).
Correlations are presented in Table 2. CO2 emissions correlate
negatively with affluence and urbanization. CO2 emissions correlate
positively with population and energy intensity.
Prior to formal econometric modeling, it is important to have an
understanding of the time series properties of the data. Unit root tests
that assume cross-sectional independence can have low power if esti-
mated on data that have cross-sectional dependence. Pesaran's (2004)
cross-section dependence (CD) test was used to check for cross-
sectional independence. The CD tests indicate that, except for energy
intensity, each series exhibits cross-sectional dependence (Table 3).
As a first investigation into unit root s, the Im et al. (2003) (IPS) panel
unit root test was run for each series (Table 4). These tests were run
with a constant and trend term and an automatic lag selection process
using the AIC with a maximum of five lags. These tests clearly indicate
that four of the series are first difference stationary and urban is station-
ary in levels. However, as indicated from Table 3, four out of five vari-
ables exhibit cross-sectional variation. As a result, Pesaran's (2007)
CIPS (Z(t-bar)) test for unit roots was calculated. This is a unit root
test that allows for cross-sectional dependence. All tests were estimated
with a constant term and trend. The CIPS tests indicate that, except for
urbanization, each series contains a unit root. These results are similar
to Liddle (2013) who also found urbanization is stationary in levels.
The mixture of I(0) and I(1) variables indicates that standard panel re-
gression techniques won't be applicable in this case. Consequently,
modeling is carried out using recently developed techniques for hetero-
geneous panels that are robust to cointegration and cross-sectional de-
pendence (Chudik et al., 2011; Kapetanios et al., 2011; Pesaran and
Tosetti, 2011).
5. Empirical results
The empirical analysis is conducted by estimating a series of regres-
sion models under different assumptions about slope coefficients and
dynamics. The first suite of results is for static models with homoge-
neous slope coefficients. Empirical results are presented for specifica-
tions estimated using pooled OLS with panel corrected standard errors
(PCSE), fixed effects (FE), and random effects (RE) (Table 5). The esti-
mated coefficient on the affluence variable is between 1.091 and 1.143
and is statistically significant. The estimated coefficient on the popula-
tion variable ranges between 0.956 and 1.884 and is statistically signifi-
cant. The estimated coefficient on the urbanization variable under fixed
effects or random effects is within the range of values reported by
other authors (eg. Cole and Neumayer, 2004; Martinez-Zarzoso and
Maruotti, 2011; Parikh and Shukla, 1995; Poumanyvong and Kaneko,
2010). The estimated coefficient on the energy intensity variable is posi-
tive and statistically significant. The results in Table 5 indicate that in-
creases in affluence, population or energy intensity increase CO
2
emissions. An increase in urbanization can have either positive or nega-
tive impacts on CO
2
emissions, depending upon the estimation tech-
nique. The residuals are tested for cross-sectional dependence using
Pesaran's (2004) CD test and stationarity is tested using Pesaran's
(2007) CIPS. It is important to test for stationarity in the residuals because
residual stationarity is an important part of a good fitting econometric
model. Applying the CD test to the regression residuals provides strong
evidence of cross-section dependence in each specification. More trou-
bling is the CIPS test indicates that all regressions have non-stationary re-
siduals which indicate a poorly fitting model.
Table 6 presents empirical results for static models with heteroge-
neous slope coefficients. The estimated coefficient on the affluence var-
iable is between 1.032 and 1.158 and statistically significant at the 1%
level. The estimated coefficient on the population variable is between
1.789 and 3.529 and significant at the 5% level. The estimated coefficient
on the energy intensity variable is between 0.830 and 0.978 and statis-
tically significant at the 1% level. The estimated coefficient on the urban
variable is negative but statistically insignificant in two specifications
and positive and significant in one specification. The CD test indicates
some evidence of cross section dependence in the CCEMG and AMG
specifications but no statistically significant evidence of cross section
dependence in the MG specification. All three specifications have
stationary residuals which may be the result of controlling for heteroge-
neous parameters and cross section dependence. Notice that the RMSE
values reported in Table 6 are one order of magnitude smaller than
those reported in Table 5 indicating a preference for models estimated
with heterogeneous slope coefficients.
Table 4
Tests for unit roots.
Variable IPS IPS 1ST DIFF CIPS CIPS 1ST DIFF
CO2 −0.367 −17.315
a
0.598 −2.999
a
Affluence −0.099 −8.896
a
1.474 −3.372
a
Population 4.015 −4.413
a
−5.341 −1.561
c
Urban −2.389
a
2.631 −1.675
b
4.474
Intensity −0.657 −17.337
a
2.111 −4.724
a
The superscripts a, b and c denote significance at the 1%, 5% and 10% levels respectively.
5
Urbanization refersto The World Bank's definitionof the percentage of thepopulation
living in urban areas as defined by national statistical offices (http://data.worldbank.org/
indicator/SP.URB.TOTL.IN.ZS). However, urban areascan be defined differently by different
national statistical offices and in general there is no universally accepted definition of ur-
banization (eg. Vlahovand Galea, 2002). Moreover, a country'sdefinition of an urban area
can change across time.
Table 2
Correlations.
CO2 Affluence Population Urban lntensity
CO2 1.000
Affluence −0.066 1.000
Population 0.791 −0.575 1.000
Urban −0.194 0.863 −0.557 1.000
lntensity 0.485 −0.737 0.626 −0.774 1.000
(obs = 624)
Table 3
Tests for cross-section dependence.
Variable CD-test p-value corr abs(corr)
CO2 49.49 b0.001 0.723 0.912
Affluence 54.78 b0.001 0.801 0.804
Population 53 b0.001 0.775 0.969
Urban 54.63 b0.001 0.799 0.873
lntensity 0.12 0.904 0.002 0.484
150 P. Sadorsky / Energy Economics 41 (2014) 147–153
Heterogeneous parameter estimates from the dynamic panel model
are reported in Table 7.
6
Looking first at the contemporaneous variables,
the estimated coefficient on the affluence variable is positive and statis-
tically significant at the 1% level. The values range between 1.041 and
1.128 indicating little variation between the three estimation methods.
The estimated coefficient on the population variable ranges between
0.742 and 2.165 and is statistically significant at the 1% level in two of
the specifications. The estimated coefficient on the urban variable is
positive but not statistically significant. Notice that the estimated coeffi-
cients on the urban variable from the heterogeneous dynamic specifica-
tions are larger than those found from thestatic specifications (Tables 5
and 6) and similar to the values reported by other authors (eg.
Martinez-Zarzoso and Maruotti, 2011; Poumanyvong and Kaneko,
2010). The estimated coefficient on the energy intensity variable is pos-
itive and statistically significant at the 1% level. The estimated coefficient
on lagged CO
2
emissions is positive and statistically significant at the
10% level and ranges between 0.101 and 0.471 indicating a low to mod-
erate degree of persistence. The estimated coefficient on the lagged af-
fluence variable is negative and statistically significant at the 1% level
in two of the specifications. The estimated coefficient on the lagged in-
tensity variable is negative and statistically significant at the 5% level
in two of the specifications. The CD test indicates no evidence of cross
section dependence at the 1% level. There is no evidence of nonstation-
ary residuals in any of the specifications at conventional levels of signif-
icance. The dynamic specification with heterogeneous slope coefficients
is preferred over static specifications with homogeneous slope coeffi-
cients because of the lower root mean square error (RMSE) values and
stationary residuals.
Since the estimated coefficient on the urbanization variable is statisti-
cally insignificant, the dynamic models were re-estimated omitting the
urbanization variable. The estimated coefficients indicate that lagged
CO2 emissions has a positive and statistically significant impact on cur-
rent period CO2 emissions while, current period affluence and intensity
each have positive and statistically significant impacts on CO2 emissions
(Table 8). The estimated coefficient on the population variable is positive
in each specification and statistically significant in two specifications. The
residual diagnostic tests indicate stationary residuals and no evidence of
cross section dependence at the 1% level of significance. Notice however,
that for each specification, the estimation with the urbanization variable
(Table 7) produces slightly lower root mean square error (RMSE) values
(4% smaller in the case of MG, 11% smaller in the case of CCEMG, and 5%
smaller in the case of AMG) than estimation without the urbanization
variable (Table 8). The differences in RMSE values are not too great and
in the interest of parsimony, the parameter estimates from Table 8 are
preferred.
6. Implications
The empirical results reported in Table 8 can be used to calculate
short-run and long-run CO
2
elasticities (Table 9). The short-run afflu-
ence elasticities range from 1.125 to 1.193 while the long-run affluence
elasticities range from 0.904 to 0.996. For each specification, the short-
run affluence elasticity is slightly larger than the corresponding long-
run elasticity.
Table 7
Heterogeneous estimates (dynamic).
MG CCEMG AMG
CO2(−1) 0.471
a
0.101
c
0.376
a
(6.47) (1.83) (5.52)
Affluence 1.128
a
1.041
a
1.091
a
(13.89) (12.83) (15.33)
Affluence(−1) −0.575
a
−0.134 −0.490
a
(−3.79) (−0.99) (−3.52)
Population 0.742
c
2.165 1.318
b
(1.87) (0.98) (2.55)
Urban 0.495 1.518 0.536
(1.26) (1.42) (1.18)
Intensity 0.794
a
0.781
a
0.772
a
(8.22) (7.74) (7.46)
Intensity(−1) −0.347
a
−0.105 −0.290
b
(−2.91) (−1.01) (−2.49)
Constant −12.27
b
−33.71 −22.04
a
(−2.24) (−1.53) (−2.95)
RMSE 0.0351 0.0257 0.0329
Observations 608 608 608
Groups 16 16 16
CD 1.27 −2.40
b
−0.65
CIPS −7.91
a
−12.47
a
−8.48
a
tstatistics in parentheses.
Estimation is from an unbalanced panel of 16 emerging economies covering the period
1971–2009. Estimated coefficients are un-weighted averages across countries.
a
pb0.01.
b
pb0.05.
c
pb0.10.
6
This model is selected using the SIC from a maximumof one lag on each right-hand-
side variable.
Table 5
Pooled estimates (static).
PCSE FE R E
Affluence 1.111
a
1.143
a
1.091
a
(109.56) (7.45) (6.67)
Population 0.956
a
1.884
a
1.322
a
(120.78) (8.27) (10.57)
Urban −0.0924
b
0.0555 0.219
(−2.13) (0. 21) (0.95)
Intensity 1.094
a
0.814
a
0.865
a
(73.11) (4.85) (4.90)
Constant −12.55
a
−30.61
a
−20.56
a
(−108.18) (−8.54) (−8.08)
RMSE 0.247 0.112 0.122
Observations 624 624 624
Groups 16 16 16
CD −2.41
b
−3.89
a
−3.99
a
CIPS 1.28 0.01 0.15
tstatistics in parentheses.
Estimation is from an unbalanced panel of 16 emerging economies covering the period
1971–2009.
Time dummy variables included in FE and RE specifications.
Robust t statistics reported.
a
pb0.01.
b
pb0.05.
Table 6
Heterogeneous estimates (static).
MG CCEMG AMG
Affluence 1.133
a
1.158
a
1.032
a
(12.95) (15.29) (13.29)
Population 1.789
a
3.529
b
2.410
a
(2.80) (2.03) (3.97)
Urban −0.0855 1.286
c
−0.00328
(−0.15) (1.76) (−0.01)
Intensity 0.978
a
0.830
a
0.907
a
(5.83) (8.22) (6.10)
Constant −27.41
a
−20.63 −38.35
a
(−3.10) (−1.10) (−4.21)
RMSE 0.0488 0.0303 0.0418
Observations 624 624 624
Groups 16 16 16
CD 0.60 −2.09
b
−1.88
c
CIPS −4.36
a
−10.59
a
−5.29
a
tstatistics in parentheses.
Estimation is from an unbalanced panel of 16 emerging economies covering the period
1971–2009. Estimated coefficients are un-weighted averages across countries.
a
pb0.01.
b
pb0.05.
c
pb0.10.
151P. Sadorsky / Energy Economics 41 (2014) 147–153
The short-run population elasticities range between 0.776 and
1.963. The long-run elasticities range between 1.748 and 2.862. For
each specification, the long-run population elasticities are larger than
the short-run population elasticities. The short-run energy intensity
elasticities are in the range 0.819 to 0.856and these values are very sim-
ilar to the long-run elasticities. All of these short-run and long-run elas-
ticities are positive indicating that increases in affluence, population, or
energy intensity increase CO
2
emissions in both the short-run and the
long-run.
These results have serious implications for the buildup (stock) of
CO
2
in the atmosphere. CO
2
emitted into the atmosphere lasts for a
long time and today's emissions (a flow) contribute to the total stock
of CO
2
in the atmosphere.
7
Since affluence and population are likely to
continue increasing in emerging economies, this leaves reductions in
energy intensity as the only practical way to reduce CO
2
emissions.
Since urbanization is found to have no statistically significant impact
on CO2 emissions at conventional levels, omitting urbanization from the
types of econometric models estimated in this paper will not have much
impact on forecasts of carbon dioxide emissions. Energy and environ-
mental polices formulated without considering the impacts of urbaniza-
tion on carbon dioxide emissions will probably meet their stated
objectives.
7. Conclusions
This paper uses a STIRPAT model to explore the impact that urbani-
zation has on carbon dioxide emissions in emerging economies. It isex-
pected that urbanization will continue rising in emerging economies
and understanding how urbanization affects CO
2
emissions is an impor-
tant and timely topic to study. A better understanding ofhow urbaniza-
tion affects CO
2
emissions is necessary from a sustainable development
perspective.
This study uses recently developed heterogeneous panel regres-
sion techniques like mean group estimators and common correlated
effects estimators to model the impact that energy use, income, and
urbanization has on CO
2
emissions for a panel of emerging econo-
mies. In particular, models are estimated using the mean group
(MG) estimator of Pesaran and Smith (1995),Pesaran's (2006) Com-
mon Correlated Effects Mean Group (CCEMG) estimator, and the
Augmented Mean Group (AMG) estimator of Eberhardt and Teal
(2010) and Bond and Eberhardt (2009). In addition, results are pre-
sented for static and dynamic specifications.
The estimated contemporaneous coefficient on the affluence vari-
able is fairly similar across different estimation techniques. This is also
the case for the energy intensity variable. These results are important
in establishing how remarkably robust the estimated coefficients on af-
fluence andenergy intensity are to different estimation techniques (ho-
mogenous slope coefficients or heterogeneous slope coefficients) and
assumptions about the dynamics (static or dynamic) even in cases
where formal diagnostic tests reveal evidence of miss-specification.
The estimated contemporaneous coefficient on the urbanization
variable is, however, sensitive to the estimation technique. For static
specifications estimated with homogeneous slope coefficients, the esti-
mated coefficient on the urbanization variable is negative and statisti-
cally significant in one of the specifications and positive and
statistically insignificant in two of the specifications. For the fixed effects
and random effects specifications, the estimated coefficient on the ur-
banization variable is within the range of values foundby other authors.
Residual diagnostic tests indicate that static specifications with ho-
mogeneous slopes are, however, miss-specified.
For static specifications estimated with heterogeneous slope coeffi-
cients, the estimated coefficient on the urbanization variable is positive
and statistically significant in one of the specifications and negative but
statistically insignificant in two of the specifications. For dynamic spec-
ifications estimated with heterogeneous slope coefficients, the estimat-
ed coefficient on the urbanization variable is statistically insignificant at
the 10% level indicating that the urbanization variable can be dropped
from the model. Based on residual diagnostic tests, the dynamic model
with heterogeneous slope coefficients is preferred.
One of the implications of these results is that omitting the urbaniza-
tion variable will have little impact on emissions reduction strategies or
sustainable development policies. The theories of ecological moderniza-
tion and urban environmental transition recognize that urbanization
can have both positive and negative impacts on the natural environ-
ment with the net effect being hard to determine a priori. Higher urban-
ization is associated with higher economic activity. Higher economic
activity generates higher wealth and wealthier residents often demand
more energy intensive products (eg. automobiles,air conditioning, etc.)
which can increase carbon dioxide emissions. Wealthier residents are
also likely to care more about the environment. Increased urbanization
also helps to facilitate economies of scale for public infrastructure and
these economies of scale lead to lower environmental damages. The re-
sults of this paper indicate that the two opposing effects of urbanization
on CO
2
emissions tend to cancel each other out leaving the net impact of
urbanization on CO
2
emissions statistically insignificant from zero.
Table 8
Heterogeneous estimates (dynamic) without urbanization variable.
MG CCEMG AMG
CO2(−1) 0.556
a
0.314
a
0.470
a
(9.88) (6.1 3) (8.46)
Affluence 1.193
a
1.125
a
1.145
a
(15.33) (12.72) (16.07)
Affluence(−1) −0.766
a
−0.442
a
−0.666
a
(−5.98) (−3.90) (−5.78)
Population 0.776
b
1.963 1.333
a
(2.38) (1.1 9) (3.66)
Intensity 0.856
a
0.843
a
0.819
a
(9.59) (8.2 2) (8.33)
Intensity(−1) −0.485
a
−0.207
c
−0.399
a
(−5.46) (−1.76) (−4.79)
Constant −11.16
b
−11.56 −20.40
a
(−2.26) (−1.50) (−3.56)
RMSE 0.0365 0.0290 0.0345
Observations 608 608 608
Groups 16 16 16
CD 1.47 −2.00
b
−0.14
CIPS −6.40
a
−10.24
a
−6.87
a
tstatistics in parentheses.
Estimation is from an unbalanced panel of 16 emerging economies covering the period
1971–2009. Estimated coefficients are un-weighted averages across countries.
a
pb0.01.
b
pb0.05.
c
pb0.10.
7
It is estimated that between 65% and 80% of CO
2
released into the atmosphere dis-
solves into the oceans over a period of between 20 to 200 years. The remaining CO
2
dis-
solves through various weathering processes and this can tak e several thousands of
years (http://www.guardian.co.uk/environment/2012/jan/16/greenhouse-gases-remain-
air).
Table 9
CO
2
emissions elasticities.
Elasticities MG CCEMG AMG
short-run
Affluence 1.193 1.125 1.145
Population 0.776 1.963 1.333
Intensity 0.856 0.843 0.819
long-run
Affluence 0.962 0.996 0.904
Population 1.748 2.862 2.515
Intensity 0.836 0.927 0.792
Elasticities derived from parameter estimates in Table 8.
152 P. Sadorsky / Energy Economics 41 (2014) 147–153
Estimates from the dynamic model with heterogeneous slope coeffi-
cients indicate that long-run population elasticities (between 1.748 and
2.862) are greater than long-run affluence elasticities (between 0.904
and 0.996) and long-run affluence elasticities are greater than long-
run energy intensity elasticities (between 0.792 and 0.927). All of
these long-run elasticities are positive indicating that increases in
affluence, population, or energy intensity increase CO
2
emissions in
the long-run.
The results of this paper show that the most direct way for emerging
economies to reduce CO
2
emissions is to reduce affluence, population,
and energy intensity. Emerging economies are, however, currently on
a trajectory of increasing affluence and population. Consequently, a re-
duction in CO
2
emissions is going to have to come from an increase in
energy efficiency and a greater effort at fuel switching from fossil fuels
to renewables. Pricing of carbon dioxide emissions either through
taxes or a cap and trade system would also be beneficial in helping to
reduce the consumption of fossil fuels.
Acknowledgements
I thank an anonymous reviewer for helpful comments.
References
Andrews, D.W.K., 2005. Cross-section regression with common shocks. Econometrica 73,
1551–1585.
Barnes, D.F., Krutilla, K., Hyde, W.F., 2005. The Urban Household Energy Transition: Social
and Environmen tal Impacts in the Developing World. Resour ces for the Future,
Washington, DC.
Bond, S., Eberhardt, M., 2009. Cross-section dependence in nonstationary panel models: a
novel estimator. Paper presented at the Nordic Econometrics Conference in Lund
Sweden.
Burton, E., 2000. The compact city: just or just compact? A preliminary analysis. Urban
Stud. 37, 1969–2001.
Capello,R., Camagni, R., 2000.Beyond optimal city size:an evaluation of alternative urban
growth patterns. Urban Stud. 37, 1479–1496.
Chudik, A., Pesaran,M.H., Tosetti, E., 2011. Weak and strongcross section dependence and
estimation of large panels. Econ. J. 14, C45–C90.
Coakley, J., Fuertes, A.-M., Smith, R.P., 2006. Unobserved hete rogeneity in panel time
series models. Comput. Stat. Data Anal. 50, 2361–2380.
Cole, M.A., Neumayer, E., 2004. Examining the impact o f demographic factors on air
pollution. Popul. Dev. Rev. 2, 5–21.
Crenshaw,E.M., Jenkins, J.C., 1996.Social structure andglobal climate change: sociological
propositions concerning the greenhouse effect. Sociol. Focus 29, 341–358.
Dietz, T., Rosa, E., 1997. Effects of population and affluence on CO2 emissions. Proc. Natl.
Acad. Sci. U.S.A. 94, 175–179.
Ehrlich, P., Holdren, J., 1971. The impact of population growth. Science 171, 1212–1217.
Eberhardt, M., Teal, F., 2010. Productivity analysis in global manufacturing production.
Economics Series Working Papers 515, University of Oxford, Department of
Economics.
Eberhardt, M., Teal, F., 2011. Econometrics for grumblers: a new look at the literature on
cross-country growth empirics. J. Econ. Surv. 25, 109–155.
Eberhardt, M., Helmers, C., Strauss, H., 2013.Do spillovers matter when estimatingprivate
returns to R&D? Rev. Econ. Stat. 95 (2), 436–448.
Fan, Y., Lui, L.-C., Wu, G., Wie, Y.-M., 2006. Analyzing impact factors of CO2 emissions
using the STIRPAT model. Environ. Impact Assess. Rev. 26, 377–395.
Gouldson, A.P., Murphy, J., 1997. Ecological modernization: economic restructuring and
the environment. Polit. Q . 68, 74–86.
Hossain, M.S., 2011. Panel estimation for CO
2
emissions, energy consumption, economic
growth, trade openness and urbanization of newly industrialized countries. Energy
Policy 39, 6991–6999.
Im, K.S., Pesaran, M.H., Shin, Y., 2003. Testing for unit roots in heterogeneous pane ls.
J. Econ. 115, 53–74.
Jenks, M., Burton, E., Williams, K. (Eds.), 1996. The Compact City: A Sustainable Urban
Form? E & FN Spon, New York.
Liddle, B., Lung, S., 2010. Age-structure, urbanization, and climate change in developing
countries: revisiting STIRPAT for disaggregated population and consumption related
environmental impacts. Popul. Environ. 31, 317–343.
Liddle, B., 2013. The energy, economic growth, urbanization nexus across development:
Evidence from heterogeneous panel estimates robust to cross-sectional dependence.
Energy J. 34, 223–244.
Kapetanios, G., Pesaran, M.H., Yamagata, T., 2011. Panels with nonstationary multifactor
error structures. J. Econ. 160, 326–348.
McGranahan, G., Jacobi, P., Songsore, J., Surjadi, C., Kjellen, M., 2001. The Citizen at Risk:
From Urban Sanitation to Sustainable Cities. Earthscan, London.
Martinez-Zarzoso, I., Maruotti, A., 2011. The impact of urbanization on CO2 emissions:
evidence from developing countries. Ecol. Econ. 70, 1344–1353.
Mol, A.P.J., Spaargaren, G., 2000. Ecological modernization theory in debate: a review.
Environ. Polit. 9, 17–49.
Newman, P.W.G., Kenworthy, J.R., 1989. Cities and Automobile Dependence: An Interna-
tional Sourcebook. Gower Technical, Aldershot.
Parikh, J., Shukla, V., 1995. Urbanization, energy use and greenhouse effects in economic
development —results from a crossnational study of developing countries. Glob. En-
viron. Chang. 5, 87–103.
Pesaran, M.H., 1997. The role of economic theory in modelling the long run. Econ. J. 107,
178–191.
Pesaran,M.H., 2004. General diagnostictests for cross section dependence in panels. Cam-
bridge Working Papersin Economics No. 0435. University of Cambridge (June 2004).
Pesaran, M.H., 2006. Estimation and inference in large heterogeneous panels with a mul-
tifactor error structure. Econometrica 74, 967–1012.
Pesaran, M.H., 2007. A simple panel unit root test in the presence of crosssection depen-
dence. J. Appl. Econ. 22, 265–312.
Pesaran, M.H., Shin, Y., 1999. An Autoregressive Distributed Lag Modelling Approach to
Cointegration Analysis. In: Strom, S. (Ed.), Econometrics and Economic Theory in
the 20th Century: the Ragnar–Frisch Centennial Symposium. Cambridge University
Press.
Pesaran, M.H., Smith, R.P., 1995. Estimating long-run relationships from dynamic hetero-
geneous panels. J. Econ. 68, 79–113.
Pesaran, M.H., Tosetti, E., 2011. Large panels with common factors and spatial correla-
tions. J. Econ. 161, 182–202.
Poumanyvong, P., Kaneko, S., 2010. Does urbanization lead to less energy use and lower
CO
2
emissions? A cross-country analysis. Ecol. Econ. 70 (2), 434–444.
Sharma, S.S., 2011. Determinants of carbon dioxide emissions: empirical evidence from
69 countries. Appl. Energy 88, 376–382.
United Nations Population Division,2007. World urbanization prospects —the 2007 revi-
sion population database. (Retrieved March 2012, from http://esa.un.org/unup/.).
Vlahov, D., Galea, S., 2002. Urbanizat ion, urbanicity, and health . J.Urban Health 79,
S1–S12.
World Bank, 2013. World Development Indicators. Accessed at: http://www.worldbank.
org/data/onlinedatabases/onlinedatabases.html.
York, R., Rosa,E.A., Dietz, T., 2003.STIRPAT, IPAT and ImPACT:analytic tools for unpacking
the driving forces of environmental impacts. Ecol. Econ. 46, 351–365.
153P. Sadorsky / Energy Economics 41 (2014) 147–153
