Carbon Emissions and Economic Growth: Causality Testing in

Heterogenous Panels

Dr David Maddison, Katrin Rehdanz

Department of Economics,
University of Birmingham, UK

Abstract
Numerous papers have examined data on energy and GDP for evidence of Granger causality.
Using time series techniques these analyses not infrequently reach differing conclusions
concerning the existence and direction of Granger causality. This paper presents a heterogenous
panel approach to Granger causality testing. This technique is used to examine a panel of data for
evidence of a causal relationship between GDP and carbon emissions per capita allowing for
heterogeneity in short run dynamics and even the long run cointegrating vector. This technique is
compared to the standard fixed dynamic effects approach to pooling individual error correction
models. In one important case the heterogenous panel test for Granger causality reaches
conclusions quite different to those from conventional tests of Granger causality. Except for Asia
there is strong evidence for the existence of a bidirectional causal relationship between GDP per
capita and CO2 emissions per capita.
Keywords: Energy; Carbon Emissions; Granger Causality; and Heterogenous Panel

1. Introduction

Since the seminal work of Kraft and Kraft (1978) researchers have attempted to determine
the existence and direction of causal linkages between energy and income. The original
impetus to this literature was concern over energy price rises, the finite nature of key
energy resources and the presumed importance of providing energy services for facilitating
the development process. Initially these causality tests involved bivariate relations between
energy and income alone. Latterly however causality tests have grown to include other
variables such as capital stock and labour. Sometimes analyses of causality focus on
particular sectors of the economy rather than the economy as a whole.

A causal link running from energy use to income is usually interpreted as meaning that
attempts to reduce energy consumption would hamper economic growth whereas causality
running in the other direction implies that reducing economic growth will effect a
reduction in energy use. Unidirectional causality from GDP to energy also indicates that
energy forecasts can be prepared assuming that income is exogenous. No causal
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

relationship indicates that development takes place independently of energy use and vice-
versa whereas bidirectional causality demonstrates the mutual interdependence of energy
and income.

The idea that energy causes income emerges from the neoclassical paradigm exemplified
by the work of Jorgenson and Wilcoxen (1993). This approach emphasises the importance
of energy in the production function alongside capital, labour and raw materials.
Alternative paradigms such as the ecological approach also suggest that energy causes
income. From an ecological perspective energy is required to extract low entropy materials
and turn them into goods and high entropy waste (e.g. Cleveland et al, 1984).

For a recent review of the literature on causal relationships between energy and income the
interested reader can consult Lee and Chang (2007), Yoo (2006) or Lee (2005). However,
virtually all recent contributors to the literature have commented on the unclear nature of
the evidence generated by the research. More specifically, it appears that evidence on
causality and the direction of causality depends on the countries considered, the time
period under scrutiny and the empirical techniques employed.1

In a meritorious although isolated attempt to trap potential sources of disagreement

Zachariadis (2006) compares some of the methods used by researchers to test for bivariate
causality between energy and GDP including the two-step error correction technique
(Engle and Granger, 1987), the autoregressive distributed lag method (Pesaran and Shin,
1999) and the vector autoregressive method (Toda and Yamamoto, 1995). Using identical
datasets he confirms that different techniques can lead researchers to different conclusions
when the sample size is limited. For this reason recent research papers have been keen to
combine cross-sectional and time-series data on energy and GDP.

Concern about the environmental consequences of energy use has provided further
stimulus to the literature testing for causality between energy use and income although
given the concern about climate change, the relationship between carbon and income is
arguably of more immediate relevance. But so far only four papers have examined causal
relationship between carbon emissions and income. Instead, a parallel literature has now

1
In his review of the literature Lee (2005) calls the evidence on causality “mixed and conflicting” and cites a
number of papers providing conflicting results of the nature and direction of causality for the identical-same
countries. Mahadevan and Asafu-Adjaye (2007) refer to the evidence as “mixed”. Referring to tests of the
relationship in transition countries Al-Iriani (2006) refers to the evidence as “mixed and sometimes
conflicting” and in developing countries “not more conclusive than [that] obtained from developed and
transition economies”. Soytas and Sari (2007) describe the research as having “failed to achieve unanimous
results”.

2
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

grown up surrounding the idea of an environmental Kuznets curve for pollutants associated
with economic activity in general and energy use in particular. The key difference between
the two strands of literature is that even if it exists a Kuznets curve relationship does not
reveal the direction of causality.2 And the question of whether reductions in carbon
emissions involve reductions in GDP has been investigated using computable general
equilibrium models e.g. Jorgenson and Wilcoxen (op cit).

Turning to the literature on causal relationships between carbon and GDP, using VAR
methods on differenced data Coondo and Dinda (2002) examine the existence of causal
relationships between income and carbon emissions using a cross country panel of data.
The results do not provide much evidence for the existence of a universal causal
relationship between income and carbon emissions. Instead the authors find that in some
regions there is a causal relationship running from carbon to income and in other regions
the causal relationship runs from income to carbon. In Africa and Asia bidirectional
causality is observed.

Dinda and Coondoo (2006) re-examine the evidence using more modern time series
econometric techniques. Surprisingly, they reject the null hypothesis of a long run
cointegrating relationship for the North America, South America, Asia and Oceania
country groupings. For the remaining country groupings as well as for the panel in its
entirety the evidence strongly points to the existence of bivariate causality.

Using the two-step error correction approach, Soytas et al (2007) study long run causality
between carbon emissions, energy use and income in the US. Once again, they find no
evidence of causality between either income and carbon emissions, or income and energy
use. By means of the method of Toda and Yamamoto (1995) Soytas and Sari (2007)
discover that carbon emissions Granger cause energy use in Turkey. As the authors
themselves remark this result appears to be counter intuitive. The same study suggests that
neither carbon emissions nor energy use Granger cause income.

The presumption that carbon emissions cause economic growth is thus hardly borne out by
the limited evidence available. The evidence however is extremely limited when compared

2
Unfortunately, just like the literature on causal relations between energy and income, the environmental
Kuznets curve literature has similarly failed to reach firm conclusions regarding whether pollution bears an
inverse U-shaped relationship with income (e.g. Stern, 2004). It is notable however that the literature tends to
find that carbon emissions tend to bear a monotonic relation with income e.g. Holtz-Eakin and Selden (1985).
This is possibly because carbon dioxide is a perfectly mixing pollutant and there is currently no economically
feasible method of removing carbon dioxide from waste gases.

3
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

to the mass of studies dealing with the existence of causal relationships between energy
and income; and the sheer number of studies estimating Kuznets curve-type relationships
between carbon and income. The present study is an attempt to partially redress the
imbalance in the literature and also to employ more recent methods to test for unit roots
and cointegration in panel data.

There is one further feature that characterises both of the multi-country research papers on
causal relations between GDP and carbon emissions per capita, as well as rather many
multi-country studies linking GDP and energy use per capita. This is the reliance on fixed
dynamic effects techniques to estimate the error correction model used in tests of Granger
causality involving nonstationary panel data. These studies pool individual panels allowing
only for fixed effects whilst simultaneously constraining the short run dynamics and the
long run cointegrating vectors to be identical. Failure to account for such heterogeneity
may result in a heterogeneity bias (Pesaran and Smith, 1995).3 Alternative estimators are
available allowing for slope-heterogeneity in the short run and even the long run
cointegrating parameter.4

Admitting the existence of panel heterogeneity also has consequences for the conduct of
Granger causality tests. Hurlin (2008) distinguishes between homogenous causality and
heterogenous causality in panels. In homogenous causality the hypothesis is that there is a
causal relationship between two variables X and Y and that the data generating process is
identical for all panels. The alternative hypothesis is that of homogenous non causality in
which there is no causal relationship between X and Y. Under the heterogenous causality
hypothesis there exists a causal relationship between variables X and Y but the data
generating process is potentially differs across panels. The heterogenous non causality
hypothesis describes a situation in which there is no causal relationship present in any of
the heterogenous panels.
The approach taken in this paper is to test whether a subset of the data exhibits a Granger
causal relationship in a heterogenous panel data context against the alternative that none of
the panel members exhibit a causal relationship. By contrast the existing literature tests the
hypothesis of homogenous non causality against the alternative hypothesis of a
homogenous causal relationship. The possibility exists that adopting a more cautious to

3
The nature of this heterogeneity bias is later described in more detail.
4
Interestingly the heterogenous panel approach has been used by Martinez-Zarzoso and Bengochea-
Morancho (2004) to estimate an environmental Kuznets curve for carbon but not for the purposes of causality
testing.

4
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

pooling data from potentially heterogenous panels advocated by this paper may diminish
the number of instances in which seemingly contradictory findings are drawn regarding the
possible existence and nature of any causal relationships.

The remainder of the paper is organised as follows. The next section describes panel data
on GDP per capita and carbon emissions per capita for 134 countries from 1990 to 2005
and examines its time series properties. Section three describes in detail the approach to
testing for causality in heterogenous panels. Section compares the results of conventional
tests of Granger causality with a test for heterogenous panels Granger causality. To
anticipate the main findings, overall it appears that the null hypothesis of no causal
relationship between per capita carbon emissions and GDP per capita can be rejected.
Nevertheless, there are instances in which the results of causality tests hinge on
assumptions about panel heterogeneity. The final section concludes.

2. Description of the data

The data is taken from the IEA (2007). It includes both GDP per capita (GDPPC) and CO2
per capita (CO2PC) measured in metric tonnes. The data covers 134 countries and runs
from 1990 to 2005. The countries are furthermore grouped into income quartiles as well as
by geographic area. These include: Western Europe; Eastern Europe; Africa; Asia; Middle
East; Latin America; North America; and Oceania. The data is described in Table 1 and the
membership of the country groupings is rendered explicit in Annex 1.

The data are subject to a variety of tests for stationarity. The test of Levin et al (2002), Im
et al (2003) and the test of Hadri (2000) are all employed. Surprisingly, the Levin et al test
suggests that the hypothesis of a common unit root can be rejected when a time trend is
included. The Im et al test indicates that when individual effects and deterministic trends
are included the hypothesis of a common unit root is occasionally strongly rejected. The
test of Hadri by contrast indicates that the hypothesis of stationarity is always strongly
rejected. Henceforth it is assumed that the CO2PC and GDPPC data is characterised by a
unit root with a deterministic trend but it is clear that the tests for stationarity yield mixed
messages.

The Kao (1999) and Pedroni (1999) tests are used to establish the existence of a long run
cointegrating relationship between GDPPC and CO2PC. Although Pedroni provides a suite
of seven different tests for cointgration in panel data, Pedroni (2004) establishes the greater

5
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

power of the Panel ADF and Group ADF tests in small samples. The results of these tests
are displayed in Table 3. With the exception of Western Europe the tests uniformly
indicate that the null hypothesis of no cointegration can be rejected. To some extent this
reinforces the earlier decision to treat the variables as nonstationary despite the somewhat
diverse findings of the panel unit root tests. The results for Western Europe are statistically
significant at the five percent level of confidence but not at ten percent. In what follows a
cointegrating relationship is assumed.

These findings are very different to those obtained by Dinda and Coondoo (2006) who
failed to find cointegration for several large country groupings. This may be because the
authors use country groupings that differ in terms of membership to those employed here
or because the data itself is different. Also, unlike this paper the authors do not use the now
standard Pedroni test for panel cointegration. Instead they obtain the residuals from an
estimate of the panel specific long run relationships linking GDP and carbon emissions per
capita and then subject these to the Im et al (op cit) unit root test. Critically this procedure
provides an invalid test of cointegration.

The failure to find a cointegrating relationship for major country groupings means that
Dinda and Coondoo are unable to proceed to the next stage of estimating the error
correction model used to test for causality for several important regions.

Table 1. Description of the data

Number of countries = 134
Number of time periods = 16
Variable Definition Mean Std Dev Min Max
GDPPC GDP per capita (000s USD) 8.95 8.26 0.45 46.50
CO2PC CO2 per capita (tonnes) 5.52 6.68 0.02 57.92
YEAR Calendar year 1997.50 4.61 1990 2005
Source: IEA (2007)

6
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Table 2. Results of Panel Stationarity Tests

Group Variable Levin et al Test IPS Test Hadri Test
All log GDPPC 0.000 0.000 0.000
log CO2PC 0.000 0.004 0.000
Q1 Income log GDPPC 0.000 0.095 0.000
log CO2PC 0.000 0.161 0.000
Q2 Income log GDPPC 0.000 0.033 0.000
log CO2PC 0.000 0.544 0.000
Q3 Income log GDPPC 0.000 0.000 0.000
log CO2PC 0.000 0.011 0.000
Q4 Income log GDPPC 0.000 0.000 0.000
log CO2PC 0.000 0.018 0.000
W Europe log GDPPC 0.002 0.002 0.000
log CO2PC 0.000 0.006 0.000
E Europe log GDPPC 0.000 0.000 0.000
log CO2PC 0.000 0.429 0.000
Africa log GDPPC 0.000 0.028 0.000
log CO2PC 0.000 0.045 0.000
Asia log GDPPC 0.112 0.458 0.000
log CO2PC 0.000 0.434 0.000
Middle East log GDPPC 0.003 0.062 0.000
log CO2PC 0.001 0.116 0.000
Latin America log GDPPC 0.042 0.240 0.000
log CO2PC 0.000 0.470 0.000
N America log GDPPC 0.541 0.352 0.049
log CO2PC 0.520 0.041 0.000
Oceania log GDPPC 0.045 0.102 0.001
log CO2PC 0.043 0.006 0.040
Note: The Levin et al test includes individual effects and time trends. The null hypothesis is of a
common unit root. The Im, Pesaran and Shin (IPS) test includes individual effects and individual
trends. The optimal lag length is chosen on the basis of the Schwartz Information Criterion. The
null hypothesis is the existence of individual unit root processes. The Hadri test includes individual
effects and individual trends. The null hypothesis is stationarity. The entries refer to the probability
of incorrectly rejecting the null hypothesis.

7
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Table 3. Results of Panel Cointegration Tests

Kao Test Pedroni Test
Group Panel ADF Group ADF
All 0.000 0.000 0.000
Q1 Income 0.000 0.002 0.000
Q2 Income 0.003 0.000 0.000
Q3 Income 0.000 0.000 0.000
Q4 Income 0.000 0.000 0.000
W Europe 0.000 0.078 0.060
E Europe 0.000 0.000 0.000
Africa 0.000 0.000 0.000
Asia 0.000 0.005 0.005
Middle East 0.000 0.000 0.000
Latin America 0.000 0.000 0.000
N America 0.000 0.000 0.010
Oceania 0.000 0.000 0.003
Note: The null hypothesis is that of no cointegration. The entries refer to the probabilities of
incorrectly rejecting the null hypothesis.

3. Testing for Granger causality in heterogenous panels

The basis for assessing causal relationships in this paper is the pooled two-step Engle
Granger procedure. The majority of analyses examining causal relationships between
carbon or energy and GDP also employ this approach allowing for fixed dynamic effects
(FDE) in a limited attempt to control for panel heterogenity.

Testing for causality assuming FDE and a logarithmic relationship between GDPPC and
CO2PC involves running the following regression:

i=n i =n
∆ ln GDPPC jt = α j + βYEARt + γ i ∑ ∆ ln GDPPC jt -i + δ i ∑ ∆ ln CO 2 PC jt −i + ζECT jt −1 + ε jt
i =1 i =1

8
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Where j signifies the country, t the time period and the maximum number of lagged
variables is i = n.5 The Greek letters αj, β, γi, δi and ζ are parameters to be estimated. The
error correction term ECT corresponds to the residual u from the long run relationship:

ln GDPPC jt = φ j + ϕ ln CO 2 PC jt + u jt

Alternatively, we may allow the long run cointegrating parameter φ to vary across panels:

ln GDPPC jt = φ j + ϕ j ln CO 2 PC jt + u jt

Testing for long run causality involves testing whether ζ = δi = 0.

In order to test reverse causality i.e. the hypothesis that in the long run GDPPC causes
CO2PC the following regression is required:

i=n i=n
∆ ln CO 2 PC jt = α j + βYEARt + γ i ∑ ∆ ln GDPPC jt -i + δ i ∑ ∆ ln CO 2 PC jt − i + ζ ( ECT jt −1 ) + ε jt The
i =1 i =1

Testing for Granger causality involves determining whether ζ = γi = 0.

Unfortunately Pesaran and Smith (op cit) have shown that if the parameter of interest is the
averaged effect of some exogenous variable on the dependent variable then the pooled
estimator is not consistent in dynamic models even for large N and T. The reason is that
when the regressors are serially correlated incorrectly ignoring parameter heterogeneity
generates serial correlation in the disturbances which in turn generates inconsistent
estimates in models with lagged dependent variables. This source of inconsistency is
distinct from that suffered by the fixed effects estimator in panels with small T.

In contrast to the FDE approach the mean group (MG) estimator provides consistent
although inefficient estimates of the average effect of exogenous variables on the
dependent variable. Each panel is estimated separately and the mean values of the
parameter estimates calculated along with their variances. This estimator allows for

5
In view of the relatively short time period covered by the data we shall assume that i = 1 in the analysis that
follows.

9
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

heterogeneity in both the short run dynamics and the cointegrating parameters. Somewhere
between the FDE and the MG estimator, the pooled mean group (PMG) estimator (Pesaran
et al, 1999) allows only for the possibility of heterogenous short run dynamics.

Regrettably, none of these estimators is suitable for testing heterogenous panel Granger
causality. The MG estimator provides consistent estimates of the average effect of the
independent variables. But a test of Granger causality cannot be conducted on the basis of
the average effect since such averaging may conceal causal relationships in individual
panels. And the FDE estimator does not even unbiased estimates of the average effect in
the face of parameter heterogeneity. The PMG estimator is also biased unless the long run
cointegrating parameters are homogenous.

The method employed to test for heterogenous panel Granger causality in this paper
involves estimating separately the long run relationship for each unit and obtaining the
residuals in the conventional manner. For each individual time series the error correction
model is then estimated and subjected to a test of Granger causality by means of a
conventional F-test. This then results in j separate P-values denoted Pj. These P values may
be combined all together or in groups to test the null hypothesis that Granger causality
does not exist in any panel against the alternative hypothesis that Granger causality exists
in at least one of the units. The technique used to combine the P values is that proposed by
Fisher (1948):

χ 22 j = −2∑ log Pj
j

Note the possibility that none of the individual Pj values achieve a given level of
significance e.g. 0.05 but that the Fisher test is significant at the 0.05 level. Note also the
fact that if any one probability Pk is equal or close to 0.000 then the Fisher test will be
statistically significant at equal or close to 0.000. In the next section we compare two
homogenous panel tests of Granger causality using FDE estimators with the Fisher test for
heterogenous panel Granger causality.

4. Results

The results of the FDE tests of Granger causality are presented in Tables 4 and 5. The
results shown in Table 4 assume slope homogeneity in the short run dynamics and in the
long run cointegrating parameters. Those shown in Table 5 assume slope homogeneity in

10
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

the short run dynamics but slope heterogeneity in the long run cointegrating parameters.
The results shown in these tables are qualitatively identical and are discussed together.

Combining all 134 countries the data suggests that there is bidirectional causal relationship
between CO2PC and GDPPC. Examining the results for countries divided into income
quartiles also reveals very strong evidence in favour of a bidirectional causal relationship
between CO2PC and GDPPC. Turning to the geographical groupings, it appears that there
is once more a strong bidirectional relationship between CO2PC and GDPPC in all regions
apart from North America. We note in passing that such findings already differ greatly
from those reported in Dinda and Coondoo (op cit).

Table 6 presents results for the heterogenous-panels test of Granger causality. These results
indicate that there is Granger causality for the World as a whole, and also for the four
income quartiles (although one of these tests is significant at only the 10 percent level).
Granger causality also exists for the majority of country groupings except for North
America and Asia. The results for North America mirror those already obtained. The result
for Asia however is strikingly different. More specifically, whilst the hypothesis that
GDPPC Granger does not cause CO2PC in any Asian country is strongly rejected the
reverse hypothesis that CO2PC does not Granger-cause GDPPC in any Asian country
cannot be rejected at any conventional level of significance.

Although this is the only result that changes, this country grouping includes China and
India; the World’s most populous countries. Much of the anticipated global increase in
CO2 emissions is expected to come from these two countries. The surprising implication is
that curtailing carbon emissions there will not reduce economic growth in this region, or at
least would not have done so over the period 1990-2006.

11
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Table 4. The Fixed Dynamic Effects Test of Granger Causality

GDPPC → CO2PC CO2PC → GDPPC
All F(2, 1737) = 264.15*** F(2, 1737) = 70.93***
Q1 Income F(2, 424) = 57.98*** F(2, 424) = 15.31***
Q2 Income F(2, 438) = 99.21*** F(2, 438) = 16.82***
Q3 Income F(2, 425) = 48.13*** F(2, 425) = 13.46***
Q4 Income F(2, 438) = 65.61*** F(2, 438) = 40.37***
W Europe F(2, 282) = 38.92*** F(2, 282) = 13.13***
E Europe F(2, 347) = 98.39*** F(2, 347) = 27.18***
Africa F(2, 333) = 27.92*** F(2, 333) = 16.42***
Asia F(2, 243) = 23.58*** F(2, 243) = 19.73***
Middle East F(2, 165) = 16.86*** F(2, 165) = 19.18***
Latin America F(2, 295) = 19.87*** F(2, 295) = 14.65***
North America F(2, 22) = 1.43 F(2, 22) = 0.86
Oceania F(2, 22) = 7.26*** F(2, 22) = 7.87***
Note: These results assume slope homogeneity in the short run coefficients and the long run
cointegrating parameters. *** denotes significance at the one percent level of confidence; **
denotes significance at the five percent level; and * denotes significance at the one percent level.

Table 5. The Fixed Dynamic Effects Test of Granger Causality Allowing for Slope
Heterogeneity in the Cointegrating Parameters
GDPPC → CO2PC CO2PC → GDPPC
All F(2, 1737) = 259.61*** F(2, 1737) = 70.97***
Q1 Income F(2, 424) = 55.14*** F(2, 424) = 16.97***
Q2 Income F(2, 438) = 99.59*** F(2, 438) = 14.44***
Q3 Income F(2, 425) = 65.40*** F(2, 425) = 19.45***
Q4 Income F(2, 438) = 55.10*** F(2, 438) = 48.76***
W Europe F(2, 282) = 56.13*** F(2, 282) = 12.20***
E Europe F(2, 347) = 81.38*** F(2, 347) = 15.72***
Africa F(2, 333) = 25.31*** F(2, 333) = 20.37***
Asia F(2, 243) = 21.80*** F(2, 243) = 24.79***
Middle East F(2, 165) = 18.00*** F(2, 165) = 21.67***
Latin America F(2, 295) = 35.53*** F(2, 295) = 10.92***
North America F(2, 22) = 3.54** F(2, 22) = 0.58
Oceania F(2, 22) = 5.26** F(2, 22) = 11.44***

12
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Note: These results assume slope homogeneity in the short run coefficients and slope heterogeneity
in the long run cointegrating parameters. *** denotes significance at the one percent level of
confidence; ** denotes significance at the five percent level; and * denotes significance at the one
percent level.

Table 6. The Fisher Test of Heterogenous Panel Granger Causality

GDPPC → CO2PC CO2PC → GDPPC
All χ2(268) = 684.15*** χ2(268) = 482.32***
Q1 Income χ2(66) = 174.64*** χ2(66) = 128.43***
Q2 Income χ2(68) = 129.60*** χ2(68) = 84.96*
Q3 Income χ2(66) = 169.40*** χ2(66) = 145.10***
Q4 Income χ2(68) = 174.50*** χ2(68) = 123.82***
W Europe χ2(44) = 116.00*** χ2(44) = 69.23***
E Europe χ2(54) = 112.83*** χ2(54) = 155.80***
Africa χ2(52) = 152.81*** χ2(52) = 95.31***
Asia χ2(38) = 75.35*** χ2(38) = 39.39
Middle East χ2(26) = 67.88*** χ2(26) = 45.41**
Latin America χ2(46) = 106.09*** χ2(46) = 63.57**
North America χ2(4) = 5.11 χ2(4) = 5.33
Oceania χ2(4) = 12.03** χ2(4) = 8.24*
Note: These results allow for slope heterogeneity in both the long run cointegrating parameters
and the short run coefficients. *** denotes significance at the one percent level of confidence; **
denotes significance at the five percent level; and * denotes significance at the one percent level.

5. Conclusions

Previous research into causal relationships between energy or carbon and GDP has reached
differing conclusions regarding the existence and direction of causality. Some of these
discrepancies might be explained by differences in the time period under scrutiny as well
as by the application of different empirical techniques to small samples. But some of the
discrepancies might be caused by inappropriately pooling data from different countries
thereby implicitly assuming that countries share the same long run cointegrating
relationship and short run dynamics.

This paper introduces the concept of heterogenous panels Granger causality in which the
null hypothesis is no Granger causality in any panel and the alternative Granger causality
in at least one time series. Using data on GDPPC and CO2PC the paper compares various
tests of Granger causality based on FDE estimators with a test for heterogenous panels
Granger causality. The comparison demonstrates that allowing for heterogeneity can in

13
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

some instances produce results concerning causality that differ markedly from those that
do not.

In particular, this paper indicates that a bidirectional causal relationship exists between
carbon emissions per capita and GDP per capita at different levels of per capita income and
over wide geographical areas. But critically in Asia, where the majority of the anticipated
growth in future emissions is expected to occur, there is no longer any evidence that
CO2PC caused GDPPC once heterogenous panel tests of Granger causality are applied.

References

Al-Iriani, A.M. (2006), Energy-GDP relationship revisited: An Example from the GCC
Countries using Panel Causality, Energy Policy, Vol. 34, pp. 3342-3350.
Cleveland, C.J., Costanza, R. Hall, C.A.S., and Kaufman, R.F. (1984), Energy and the US
Economy: A Biophysical Perspective, Science, Vol. 225, pp. 890-897.
Coondoo, D. and Dinda, S. (2002), Causality between Income and Emission: A Country
Group Specific Econometric Analysis, Ecological Economics, Vol. 40, pp. 351-367.
Dinda, S. and Coondoo, D. (2006), Income and Emissions: A Panel Based Cointegration
Analysis, Ecological Economics, Vol. 57, pp. 167-181.
Engle, R.F. and Granger, C.W.J. (1987), Cointegration and Error Correction:
Representation, Estimation and Testing, Econometrica, Vol. 55, pp. 251-276.
Fisher, R.A. (1948), Combining Independent Tests of Significance, American Statistician,
Vol. 2, pp. 30.
Hadri, K. (2000), Testing for Stationarity in Heterogeneous Panel Data, Econometrics
Journal, Vol. 3, pp. 148-161.
Holtz-Eakin, D. and Selden, D.M. (1995), Stoking the Fires? CO2 Emissions and
Economic Growth, Journal of Public Economics, Vol. 57, pp. 85-101.
Hurlin, C. (2008), Testing for Granger Non-Causality in Heterogenous Panels, Mimeo.
Department of Economics: University of Orleans.
Im, K.S., Pesaran, M.H., Shin, Y. (2003), Testing for Unit Roots in Heterogenous Panels,
Journal of Econometrics, Vol. 115, pp. 53-74.
Jorgenson, D.W. and Wilcoxen, P.J. (1993), Reducing US Carbon Emissions: An
Econometric General Equilibrium Assessment, Resource and Energy Economics, Vol. 15,
pp. 7-25.
Kao, C. (1999), Spurious Regression and Residual-Based Tests for Panel Cointegration,
Journal of Econometrics, Vol. 90, pp. 1-44.
Kraft, J. and Kraft, A. (1978), Relationships between Energy and GNP, Journal of Energy
and Development, Vol. 3, pp. 401-403.
Lee, C-C. (2005), Energy Consumption and GDP in Developing Countries: A Cointegrated
Panel Analysis, Energy Economics, Vol 27, pp. 415-427.

14
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Lee, C-C. (2006), The Causality Relationship between Energy Consumption and GDP in
G-11 Countries Revisited, Energy Policy, Vol. 34, pp. 1086-1093.
Lee, C-C. and Chang, C-P. (2007), Energy Consumption and Economic Growth in Asian
Economies: A More Comprehensive Analysis using Panel Data, Resource and Energy
Economics (forthcoming).
Levin, A., Lin, C. F. and Chu, C.S. (2002), Unit Root Tests in Panel Data: Asymptotic and
Finite Sample Properties, Journal of Econometrics, Vol. 108, pp. 1-24.
Mahadevan, R. and Asafu-Adjaye, J. (2007), Energy Consumption, Economic Growth and
Prices: A Reassessment using Panel VECM for Developed and Developing Countries,
Energy Policy, Vol. 35, pp. 2481-2490.
Martinez-Zarzoso, I. and Bengochea-Morancho, A. (2004), Pooled Mean Group
Estimation of an Environmental Kuznets Curve for CO2, Economics Letters, Vol. 81, pp.
121-126.
Pedroni, P. (1999), Critical Values for Cointegration Tests in Heterogenous Panels with
Multiple Regressors, Oxford Bulletin of Economic Statistics, (Special Issue) pp. 653-669.
Pesaran, M.H. and Smith, R. (1995), Estimating Long Run Relationships from Dynamic
Heterogenous Panels, Journal of Econometrics, Vol. 68, pp. 79-113.
Pesaran, M.H., Shin, Y. and Smith, R. (1999), Pooled Mean Group Estimators of Dynamic
Heterogenous Panels, Journal of the American Statistical Association, Vol. 94, pp. 621-
634.
Pesaran, M.H., Shin, Y. and Smith, R.J. (2001), Bounds Testing Approaches to the
Analysis of Level Relationships, Journal of Applied Econometrics, Vol. 16, pp. 289-326.
Soytas, U. and Sari, R. (2007), Energy Consumption, Economic Growth and Carbon
Emissions: Challenges Faced by an EU Candidate Member, Ecological Economics
(forthcoming).
Soytas, U., Sari, R. and Ewing. B.T. (2007), Energy Consumption, Income and Carbon
Emissions in the United States, Ecological Economics, Vol. 60, pp. 482-489.
Stern, D. (2004), Economic Growth and Energy. In Cleveland, C.J. et al (Eds.)
Encyclopedia of Energy Volume 2, Elsevier, Amsterdam.
Toda, H.Y. and Yamamoto, T. (1995), Statistical Inference in Vector Autoregression with
Possibly Integrated Processes, Journal of Econometrics, Vol. 66, pp. 225-250.
Yoo, S-H. (2006), The Causal Relationship between Electricity Consumption and
Economic Growth in the ASEAN Countries, Energy Policy, Vol. 34, pp. 3573-3582.
Zachariadis, T. (2006), On the Exploration of Causal Relationships between Energy and
the Economy, Discussion Paper 2006-05, Department of Economics, University of Cyprus.

15
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Annex 1. Definitions of Country Groupings

Q1 low income countries: Angola; Benin; Cameroon; Congo; Democratic Republic of
Congo; Cote d’Ivoire; Eritrea; Ethiopia; Ghana; Kenya; Mozambique; Nigeria; Senegal;
Sudan; United Republic of Tanzania; Togo; Zambia; Zimbabwe; Iraq; Yemen; Armenia;
Kyrgyzstan; Republic of Moldova; Tajikistan; Uzbekistan; Bolivia; Haiti; Bangladesh;
India; Democratic People’s Republic of Korea; Nepal; Pakistan; and Vietnam.

Q2 medium low income countries: Algeria; Egypt; Morocco; Tunisia; Islamic Republic of
Iran; Jordan; Lebanon; Syria; Albania; Romania; Bosnia and Herzegovina; Former
Yugoslavian Republic of Macedonia; Serbia and Montenegro; Azerbaijan; Georgia;
Kazakhstan; Turkmenistan; Ukraine; Dominican Republic; Ecuador; El Salvador;
Guatemala; Honduras; Jamaica; Nicaragua; Panama; Paraguay; Peru; Venezuela;
Indonesia; Myanmar; Philippines; Sri Lanka; and People’s Republic of China.

Q3 medium high income countries: Mexico; Czech Republic; Greece; Hungary; Poland;
Slovak Republic; Turkey; Korea; Gabon; Libya; Namibia; South Africa; Oman; Saudi
Arabia; Bulgaria; Malta; Croatia; Belarus; Estonia; Latvia; Lithuania; Russia; Argentina;
Brazil; Chile; Colombia; Costa Rica; Cuba; Netherlands Antilles; Trinidad and Tobago;
Uruguay; Malaysia; and Thailand.

Q4 high income countries: Canada; United States; Austria; Belgium; Denmark; Finland;
France; Germany; Iceland; Ireland; Italy; Luxembourg; Netherlands; Norway; Portugal;
Spain; Sweden; Switzerland; United Kingdom; Australia; Japan; New Zealand; Bahrain;
Israel; Kuwait; Qatar; United Arab Emirates; Cyprus; Gibraltar; Slovenia; Brunei; Chinese
Taipei; Singapore; and Hong Kong.

Western European countries: Austria; Belgium; Denmark; Finland; France; Germany;

Greece; Iceland; Ireland; Italy; Luxembourg; Netherlands; Norway; Portugal; Spain;
Sweden; Switzerland; Turkey; United Kingdom; Cyprus; Gibraltar; and Malta.

Eastern European countries: Czech Republic; Hungary; Poland; Slovak Republic; Albania;
Bulgaria; Romania; Bosnia and Herzegovina; Croatia; FYR of Macedonia; Slovenia;
Serbia and Montenegro; Azerbaijan; Belarus; Estonia; Georgia; Kazakhstan; Kyrgyzstan;
Latvia; Lithuania; Republic of Moldova; Russia; Tajikistan; Turkmenistan; Ukraine; and
Uzbekistan.

African countries: Algeria; Angola; Benin; Cameroon; Congo; Democratic Republic of

Congo; Cote d’Ivoire; Egypt; Eritrea; Ethiopia; Gabon; Ghana; Kenya; Libya; Morocco;

16
 Globalization, Energy and Environment
Warsaw School of Economics, 29-30 May, 2008

Mozambique; Namibia; Nigeria; Senegal; South Africa; Sudan; United Republic of

Tanzania; Togo; Tunisia; Zambia; and Zimbabwe.

Asian countries: Bangladesh; Brunei; Chinese Taipei; India; Indonesia; Democratic

People’s Republic of Korea; Malaysia; Myanmar; Nepal; Pakistan; Philippines; Singapore;
Sri Lanka; Thailand; Vietnam; People’s Republic of China; and Hong Kong.

Middle Eastern countries: Bahrain; Islamic Republic of Iran; Iraq; Israel; Jordan; Kuwait;
Lebanon; Oman; Qatar; Saudi Arabia; Syria; United Arab Emirates; and Yemen.

Latin American countries: Argentina; Bolivia; Brazil; Chile; Colombia; Costa Rica; Cuba;
Dominican Republic; Ecuador; El Salvador; Guatemala; Haiti; Honduras; Jamaica;
Mexico; Netherlands Antilles; Nicaragua; Panama; Paraguay; Peru; Trinidad and Tobago;
Uruguay; and Venezuela.

North American countries: Canada and United States of America.

Oceania countries: Australia and New Zealand.

17