environmental kuznets revisited · 2015. 7. 2. · called environmental kuznets curve (ekc)....
TRANSCRIPT
Environmental Kuznets Curves for CO2:
Heterogeneity versus Homogeneity
Elbert Dijkgraaf
Erasmus University Rotterdam, Rotterdam
Betrand Meelenberg Department of Econometrics and CentER, Tilburg University
Herman R.J. Vollebergh
Erasmus University Rotterdam, Rotterdam
January 31th, 2004
Paper Submitted for the SURED-conference 2004 Ascona – Switzerland
June 7-10, 2004
Abstract We explore the emission-income relationship for CO2 in OECD countries using various modelling strategies. Even for this relatively homogeneous sample we find that the inverted U-shaped curve is quite sensitive to the degree of heterogeneity included in the panel estimations. This finding is robust, not only across different model specifications but also across estimation techniques, including the more flexible non-parametric approach. Differences in restrictions applied in panel estimations are therefore responsible for the widely divergent findings for an inverted U-shape for CO2. Our findings suggest that enough heterogeneity is essential to prevent spurious correlation from reduced form panel estimations. Moreover, this inverted U is likely to exist for some, but not for all countries. Keywords: Environmental Kuznets Curves, Semiparametric Estimation, Heterogeneity. JEL Code: C 33; O 50; Q 40
Correspondence: Herman R.J. Vollebergh, Department of Economics, Erasmus University Rotterdam PO Box 1738, 3000DR, Rotterdam, The Netherlands. Phone +31)104081498; Email: [email protected]
1. Introduction
Recently, attention has grown for the robustness of reduced form estimations of the typical
inverted U-shaped relationship between economic growth and the environment, the so-
called Environmental Kuznets Curve (EKC). Earlier findings reported in the literature have
been shown to be highly sensitive to data sets, model specifications, and differences in
employed estimation techniques. For instance, Harbaugh, Levinson and Wilson (2002)
show how the specification of functional forms in the parametric approach as well as the
inclusion of additional covariates and the use of different data sets causes much less
empirical support for the existence of an inverted U-shaped relationship for ambient air
pollution emissions. Millimet, List and Stengos (2003) go one step further and report that
modeling strategies are also important and they provide statistical evidence in favor of the
more flexible semi-parametric models. However, they still find support for so called EKC
Turning Points, that is points after which per capita emissions tend to decline with per
capita income levels, in the case of US NOx and SO2 emissions.
Estimations of inverted U-shaped patterns for CO2 emissions show an even more
scattered picture leaving an outsider confused behind. This literature started with the much
cited papers by Shafik (1994) and Holtz-Eakin and Selden (1995) who both employ the
traditional parametric approach. Both papers report a Turning Point though far out of
sample ($ 7 million per capita). In an interesting contribution to this literature Schmalensee
et al. (1998) (SSJ), using the much more flexible spline-based estimation technique, even
report clear evidence of a within sample Turning Point with negative income elasticities for
the highest income segment. However, with the even more flexible non-parametric
approach Azomahou and Van Phu (2000) have shown that the inverted U-shape would no
longer hold in the CO2 case. In their sample the richest countries do not reach a phase of
declining per capita emissions, which challenges the idea that countries become
1
automatically cleaner if they reach higher income levels. They also claim their estimation
technique would be preferred to the traditional parametric approach, and more recent
papers even seem to question the panel approach as such (Egli, 2002). So whether an
“inverted U” exist for CO2 still remains a puzzle, which is rather discomforting in view of
the important international environmental challenge of climate change.
In this paper we take up this challenge and ask whether it is likely that an EKC
pattern exist for CO2. To that end we compare the different estimation techniques
examined in the literature so far using a panel for CO2-emissions in OECD countries
between 1950-2000. Moreover, we extend the semiparametric partially linear regression
(PLR) model followed by Millimet et.al. (2003) to capture potential criticisms that their
approach does not allow for heterogeneity in the time dummy. Accordingly we also
include estimations that allow for as much heterogeneity as possible but still benefit from
panel estimations techniques.
Our empirical results provide interesting new insights. First, we confirm earlier
findings in the literature showing that EKC findings for countries depend on the type of
estimation technique employed. The more flexible trend corrected non-parametric
estimation further adds to the widely divergent picture of EKC patterns for countries.
Second, our results tend to support the non-existence of an inverted U pattern. We find
support for the existence of the EKC pattern only in 1 country (Sweden) regardless of
estimation technique. In contrast 11 countries do not show this pattern (including rich
countries like Australia, Canada, France, Italy), whereas different estimation techniques
yield different windows for the other 12 (including a rejection of the EKC-hypothesis for
Japan and the US based on both non-parametric estimations). Our findings are robust for
different data sets. Moreover, we explore sensitivity of this outcome for covariates
reflecting (homogeneous) exogenous factors not yet explored in this context, and for
2
different levels of aggregation. We conclude that an EKC for CO2 is rather unlikely for the
whole sample, but might still be likely for a subsample of countries. Our finding also
matches basic economic intuition, which suggest that “automatic” internalization of
climate change, with its large cross-country spillovers, is not very likely if countries grow
richer.
The paper is organized as follows. Section 2 describes our data set and shortly
discusses the econometric model specifications. Section 3 shows that applying commonly
used estimation techniques to our data set reproduces the mixed picture for the existence of
an EKC found in the literature. Section 4 explores in detail the role of (des)aggregation,
heterogeneity and exogeneity using both parametric and more flexible estimation
techniques. Section 5 concludes.
2. Empirical Strategy
2.1 Econometric approach
To maintain as much consistency as possible with previous studies on an inverted U for
CO2, we not only focus on polynomial specifications of country-level emissions as a
function of country per capita income allowing for both country and time (fixed) effects,
but we also include the spline function approach applied by SSJ, as well as non-parametric
models. In its most general form, we consider the specification
0],,[,),,( =+= tryEtryhc εε (1)
with c = C/N, y = Y/N, and where r stands for country/region r, and t stands for year t. To
start with, the function h is left unspecified. However, without further restrictions this
function is not identified, since for each (r,t)-combination only one (c,y)-observation is
available. In the traditional, homogeneous approach one imposes a structure like the
following:
3
tt
trr
r dtdrygtryh ∑∑ ++= λαβ ),(),,( (2)
with β = (β0, β1, β2, β3)′ such that
( ) 33
2210, yyyyg βββββ +++= (3)
and with drr a dummy for country/region r, and dtt a dummy for year t. This model can be
estimated using standard panel data techniques (after imposing appropriate distributional
assumptions). This traditional approach is quite restrictive because it is typically assumed
that every cross-sectional unit reacts similar to shifts in the income parameters, even if the
units are allowed to differ in their intercepts. This homogeneity assumption also
characterizes more flexible estimation techniques applied in the literature, like the spline
method used by Schmalensee et.al. (1989) and the semiparametric methods used by
Galeotti and Lanza (1999) and Millimet et.al. (2003). These semiparametric alternatives
estimate a version of
tt
trr
r dtdrygtryh ∑∑ ++= λα)(),,( (4)
with g(.) left unspecified. This model can be estimated using Robinson (1988), as, for
example, in the analysis of Millimet et al. (2003) who apply the semiparametric partially
linear regression (PLR) model.
The common procedure behind both the traditional and non-parametric approach is
based on the general premise is that a single cross-sectional unit r undergoes the inverted-
U relationship over time. Usually, only country specific heterogeneity intercepts are
allowed and not heterogeneous slope parameters, i.e., δh(y,r,t)/δy is postulated not to
depend on r. The typical exception is List and Gallett (1999) and also Millimet et.al.
(2003) point at the potential relevance of heterogeneity. Extending List and Gallett’s
(1999) approach to more flexible estimation techniques, we show that estimations for the
4
CO2-case are strongly dependent on the amount of heterogeneity allowed in either the
income and time parameters or both.
For both the parametric and spline approach we explicitly tested for the country
specific homogeneity assumption that all countries follow an isomorphic pattern for CO2-
emissions in relation to GDP. More precisely, for example, in case of equation (2) we
considered as a generalization
tt
trr
rrr dtdrygtryh ∑∑ ++= λαβα ),(),,( (5)
and we tested the null hypothesis βr = β. We also allowed for different degrees of
heterogeneity in the control variables (including country fixed effects as well as country
specific trends) and for both the polynomial and spline specifications.1
Note, however, that rejection of the null hypothesis might also indicate model
misspecification, and, thus, rejection of the hypothesis βr = β does not necessarily imply
non-homogeneity.2 So, to proceed it makes sense to consider
tt
t dtryftryh ∑+= λ),(),,( (6)
with f(.,.) left unspecified, and to test f(y,r) = g(y). Specification (6) fits in the Robinson
(1988)-framework, but there is one difficulty. In order to apply Robinson (1988) one has to
estimate in a first round
0],[ =rydtE t (7)
But for a given country/region r the dummy variable dtt is always zero, except for one
observation, namely year t, implying that there is not enough variation in the data to
estimate (7), making an application of Robinson (1988) to (6) impossible. As a
consequence, nonparametric testing for homogeneity does not seem to be possible if the
starting point is specification (4) as employed by, for instance, Millimet et al. (2003). To
5
proceed, one has to impose more structure on the time component, i.e., instead of (6) one
could consider
)(),(),,( tGryftryh += (8)
with G(t) less flexible than G(t) = tt
t dt∑λ . For instance, by taking G(t) = ( )Htdrdrt rtt
t∑λ
[NOTE: To be included: further explanation of the PLR estimate with time heterogeneity] 3
2.2 Data
Our results are based on national-level data for 24 OECD countries (excluding ‘new’
members like the Czech Republic, Hungary, Korea, Mexico and Poland) between 1950 and
2000. We thus concentrate exclusively on the subsample of traditional OECD countries,
which alone is responsible for 50% of overall world carbon dioxide emissions in 1996. The
data included are the following:
C = CO2 emissions from energy consumption, millions of metric tons of C [??]
Y = GDP, millions of 1990 U.S. dollars
N = population
E = energy consumption, million Tons of Oil Equivalent (TOE)
Our overall data set contains 984 observations on these variables, for each country we have
41 observations.
Data on C are calculated from E using OECD (2002) and IEA (1991). To calculate
CO2 emissions, we use data for Total Primary Energy Supply (TPES) per fuel, corrected
for non-energy use of fuels such as chemical feedstocks. Fuels incorporated in the
calculations are coal, other solid fuels (wood for example), crude oil, petroleum products
and natural gas. Total energy use per country as well as emissions are corrected for exports
and imports of fuels, as well as for stock changes and international marine bunkers.4
6
Data on Y and N were taken from the OECD Energy Balances. All figures are
expressed in 1990 dollars, using purchasing power parities. Time coverage of these data is
considerably more recent compared to the widely used Penn World Table, which has
figures only until 1992. The data on Germany require some additional attention due to the
country’s unification in 1991. The OECD reconstructed data on Y for Germany as a whole
(including the former GDR) for the years between 1970 and 1989. We further extrapolated
GDP figures backward until 1960 using adjusted GDP levels for Western Germany with
the number of inhabitants of Eastern Germany. See Table 1 for descriptive statistics for all
variables and Figure 1 for a scatter plot of our data. [INSERT TABLE 1] [INSERT
FIGURE 1]
Note that our findings are based on a subset of the countries that are usually
analyzed in the case of CO2-emissions. Our panel, however, is particularly useful for a
study of the homogeneity assumption at the country level because there is a wide overlap
of observations of different countries at similar income levels. Moreover, the range of
observations is long enough to test for each country whether their slope coefficients are
sufficiently close to allow for panel based estimations of an EKC for CO2. If a problem
arises for this (high-income) subsample of OECD countries, one might expect the
homogeneity assumption to be even more problematic for samples including both OECD
and non-OECD countries. Note, in addition, that this panel also includes data on CO2-
emissions covering the ‘90s and is by far the most reliable source of information on these
emissions compared to non-OECD sources as well as non-energy related CO2-emissions.
3. Empirical results based on the homogeneity assumption
This section reproduces earlier estimations in the literature of the inverted-U pattern based
on the homogeneity assumption for the case of CO2-emissions as our benchmark. Figure 1
7
summarizes our main findings for (pooled) parametric cubic specification, the (linear)
spline method and the standard non-parametric estimation based on Robinson (1988).
[INSERT FIGURE 2]
The response coefficients for income in the cubic specification with both time and
country fixed effects are significantly different from zero at the p <0.01 levels
Interestingly, our results present a much gloomier picture of an EKC pattern for CO2-
emissions compared to earlier results based on polynomial specifications reported by
Shafik (1994) and Holtz-Eakin and Selden (1995). We find a within sample Turning Point
(TP) at $ 14.365 which is at 43% of the maximum panel observation. Vertical lines added
at the predicted peak of this parametric EKC and its upper and lower limit of the 95%
confidence interval for the estimated parametric peak indicate the robustness of this result.5
Further evidence for an EKC pattern is produced with the much more flexible
piecewise (linear) spline framework first applied in this context by SSJ (1998). Like SSJ,
we first started with a model featuring 20- and 24-segment splines and time-fixed effects,
where each segment contains the same number of data points. In our case, we reject
simplifications to 12 and 10 splines that preserve this symmetry, but the differences are
small. The same holds for simplifications from 16 to 8 splines. Our findings indicate a TP
at a much higher income level than the standard parametric estimation, though still within
sample, i.e. at 64 % maximum value and significant at p <0.01%.6
These results seem to provide overwhelming evidence for the existence of an
inverted U for CO2. Applying the semiparametric estimation technique, however, yields a
completely different picture. Although the fitted line more or less closely follows the EKC
pattern produced by the (parametric) cubic specification for income levels up to $20.000 or
59% of the maximum income level, we observe an emission-income relationship that casts
doubt on the existence of an inverted U for the observations at the upper end of the
8
sample.7 Where both the traditional parametric as well as the spline method generate a
negative income-elasticity, the semiparametric results are much less decisive and even
suggest an overall positive elasticity (see figure 1). This finding seems to confirm
Azomahou and Van Phu (2000) who – using the same specification – conclude that the
overall pattern more or less follows a monotonic increasing pattern of CO2 emissions per
capita with rising (per capita) income levels. Therefore they claim that the existence of an
EKC pattern for CO2-emissions should be rejected. However, the number of observations
at the upper end is small as is reflected in the much wider confidence bound.
Thus we more or less reproduce the existing findings in the literature. Accordingly,
the different findings for the three basic models applied in the literature on the EKC for
CO2-emissions confirm the main finding reported by Millimet et.al. (2003) that modeling
strategies matter. For our data, however, not only the location of the TPs is different, but
also the answer to the question whether an inverted U exist or not. Moreover, applying the
same specification test using the semi-parametric PLR method as the alternative (see
Zheng, 1996 and Li and Wang, 1998), we also reject the parametric but not the spline
based specification. This is a bit surprising and may be due to the fact that the PLR method
is highly inconclusive at the upper tail of the income distribution. The last spline, however,
closely follows the parametric estimation that is rejected. Therefore we belief there is little
reason to belief the spline method is more accurate given the upward shift at the end of the
distribution.[NOTE: results of the specification tests]
Closer inspection of the data shows that only one or two countries dominate the
data in the upper tail of the income distribution, in particular Luxemburg and the USA (see
also Figure 1). Re-estimating the models specified before without the data for Luxemburg
does not alter the parametric or spline-based result, but has a considerable effect on the
standard PRL model. An inverted emission-income pattern without the data for Luxemburg
9
is highly unlikely in this case. Therefore we conclude that not only parametric but also
semiparametric results may be dependent on relatively few observations in the upper tail of
the (income) distribution. Indeed, the weight of the data for Luxemburg – with only
400.000 inhabitants –is entirely similar to those of countries like the USA, Japan or
Germany. Moreover, one major event – the closing down of a large steel firm in the ‘80s in
Luxemburg8 – may affect our ultimate judgment on whether or not an EKC for CO2 exist.
It goes without saying that this is undesirable.
This issue points at an issue hardly explored in the literature so far, which is the
role of international specialization and divergent patterns in technological change across
countries (over time), which may not be adequately captured by country and time fixed
effects in a modeling environment based on the homogeneity assumption. Even without an
effective regulatory policy on CO2-emissions, OECD countries show considerable
improvements in their energy-efficiency although with remarkable differences across
countries at the same time. Because of the gradual change in the dependence on fossil fuels
of several, but not all OECD countries one might expect significant differences across
countries in terms of their CO2-emission-income relationship. Moreover, international
specialization of industries over time also affects differences in country specific energy-
intensity and its closely related CO2-emission intensity.
4. The role of heterogeneity
4.1 Homogeneity and model specification
Essentially reduced form estimations of the EKC hypothesis focus on the role of the
income parameters β, while preserving as much homogeneity between different cross-
sections r as possible, i.e. βr = β. This typically has the advantage to yield predictions: one
expects a country at the lower end of the income observations to follow the same emission-
10
income pattern as the other cross-sections even if its emission level is different (controlled
for by country fixed effects). Therefore, maintaining homogeneity is a desirable
characteristic of panel based estimations.
Unfortunately, explicit testing of the null hypothesis of homogeneous country-
specific slopes (i.e., whether βr = β in (5)) of the models presented in the previous section
yields a clear rejection of this core assumption at the p <0.01 levels [INSERT TABLE 2]
The magnitude of the Wald-test for the model with country-fixed effects only (not
included), and the model with both time- and country-fixed effects, is Wald(69 888) =
5,694 and Wald(69 851) = 4,830 at the p <0.01 level, respectively. This result does not
change if one allows for more flexibility in the time parameter by including country-
specific trends (see second column of Table 2). Even though this more general model
performs considerably better than the commonly estimated parametric models, the
homogeneity assumption on the GDP coefficients is still rejected (the Wald-statistic is
Wald(69 864) = 1,328.9
Further testing of homogeneity in the case of the spline (piecewise linear) function
yields similar results. As far as the homogeneity assumption is concerned, we also find
clear indications that even spline models with country-specific trends do not allow for
enough heterogeneity. With the same income levels for the different segments applied to
the country level, we find a rejection of this crucial assumption for the preferred models in
all cases.10 As we explained in section 2 these results may indicate that reduced form
parametric or spline based estimations assuming homogeneity in either the income or the
time parameter or both might be misleading.11
Direct testing of the homogeneity assumption in the semi-parametric Robinson
framework is more complicated. More heterogeneity in the income parameter is only
possible with enough (parametric) structure on the time component. We find that a third
11
order time polynomial is the preferred (semi-parametric) model, which, in turn, could be
tested against its homogeneous specification. Again this yields a clear indication that the
homogeneity restriction on the income parameter is too strong.
The importance of heterogeneity is further illustrated by including country-specific
GDP variables for one country at a time in the panel model. Using a LR test, we have to
reject the hypothesis of homogeneity for 14 of the 24 countries at a 99% level of
significance (using the preferred model with country-specific trends).12 Furthermore, by
systematically testing the homogeneity of all possible sub-panels (in total, nearly 380,000
combinations are checked), we also found that sub-panels for which homogeneity is not
rejected are rare, and never exceed a group of five countries. Moreover, even for very
small sub-panels, homogeneity is rejected in nearly all cases. Thus, even for an apparently
homogeneous group of OECD countries, panel-based estimates for commonly used
polynomial estimators do not seem to allow for enough heterogeneity, and might yield
biased and inconsistent parameter estimates for CO2 emissions.
4.2 An EKC for CO2?
To test the implications of our result for the EKC hypothesis, we finally compare country-
specific income parameter estimates for the heterogeneous polynomial model including a
(country specific) trend with a PLR estimate for each country separately. We allow for
more heterogeneity in the time component of the PLR estimates by applying (9) which is
based on pairwise combinations of countries that have developed more or less closely over
time, like Belgium and the Netherlands (see Appendix for further details).
The result is remarkable (see Figure 3). [INSERT FIGURE 3] Now only 14 of the
24 countries have a within-sample TP, and of those 14 even 5 are (far) out of their own
country income range. Furthermore, of the 13 OECD countries mentioned explicitly by
12
SSJ as having a within-sample turning point (all of them dated in the ‘70s), only four
countries confirm this picture based on our estimates (Germany, Luxembourg,
Switzerland, and the US). Also, the data for three of the seven highest income countries do
not indicate a turning point according to our estimates.
Again the semi-parametric estimates present a strikingly different pattern. For
several countries the polynomial based parametric estimates suggest a clear TP where the
PLR estimate points at a different development over time. [TO BE TESTED]
4.3 Robustness checks
One obvious objection to our findings is that our results are sensitive to the data sets used.
To check this sensitivity, we also tested whether the homogeneity hypothesis is rejected for
the data sets used by Holtz-Eakins and Selden (1995) and SSJ. We first tested for a sample
period excluding data between 1990-1997. We also used income data taken from the Penn
World Table until 1992 for the same (OECD) sample (this also accounts for potential
problems with data on Y for Germany, as these data are restricted to West Germany only).
Finally, we used emission figures for the same panel taken from the Carbon Dioxide
Information Analysis Center of the Oak Ridge National Laboratory. In all of these cases
our basic findings are similar.1
5. Conclusion
Our findings suggest that panel-based estimations of the inverted-U hypothesis for CO2
should be treated with care. Although non-parametric estimations of a rather restrictive
specification for the entire panel suggest that no such pattern exist, and specification tests
suggest this technique to be preferable, allowing for country specific estimations shows
that such inverted U-shaped patterns do exist for several countries. Thus the existence of
13
an overall inverted-U for CO2 emissions ultimately depends on the balance between the
high-income countries with an (expected) inverted-U, and those high-income countries
with a still-growing amount of (per capita) emissions. An overall inverted-U seems
doubtful if so many counterexamples exist at the country level. Lack of homogeneity with
respect to CO2, however, should not come as a surprise, given the trends in international
specialisation, and other differences in local circumstances, as well as the absence of (co-
ordinated) policies against CO2 emissions in the past.
14
References
Azomahou, Theophile, and Nguyen van Phu (2000), “Economic Growth and CO2
Emissions: a Nonparametric Approach”, Strasbourg, France.
Bruyn, Sander M. de (2000), Economic Growth and the Environment, Dordrecht: Kluwer
Academic Publishers.
Grossman, Gene M. and Allen B. Krueger (1995), “Economic Growth and the
Environment,” Quarterly Journal of Economics 110(2), 353-377.
Harbaugh, William, Arik Levinson and David Wilson (2002), “Re-examining the
Empirical Evidence for an Environmental Kuznets Curve”, Review of Economics and
Statistics, 84(3) 541-551.
Holtz-Eakin, Douglas and Thomas M. Selden, “Stoking the Fires? CO2 Emissions and
Economic Growth,” Journal of Public Economics 57 (1995), 85-101.
IEA/OECD, Greenhouse Gas Emissions: The Energy Dimension (Paris: OECD, 1991).
Lanne, M. and M. Liski (2003), “Trends and Breaks in per-capita Carbon Dioxide
Emissions, 1870-2028, MIT-CEEPR Working Papers, 2003-002, Boston.
List, John A. and Craig A. Gallet, “The Environmental Kuznets Curve: Does One Size Fit
All?,” Ecological Economics 31 (1999), 409-423.
Millimet, Daniel L., John A. List and Thanasis Stengos (2003) “The Environmental
Kuznets Curve: Real Progress or Misspecified Models?”, The Review of Economics
and Statistics, 85, 1038-1047.
OECD (2000), Energy Balances, Statistical Compendium, Paris.
Schmalensee, Richard, Thomas M. Stoker, and Ruth A. Judson, “World Carbon Dioxide
Emissions: 1950-2050,” The Review of Economics and Statistics, 80(1) (1998), 15-27.
Shafik, Nemat, “Economic Development and Environmental Quality: An Econometric
Analysis,” Oxford Economic Papers, 46(0) (1994), 757-773.
15
Figure 1 Data plot of emission-income relationship
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
11000
12000
13000
2500 7500 12500 17500 22500 27500 32500
Per capita income
Per c
apita
CO
2
Other countries United States Luxembourg
16
Figure 2 Estimation results based on the homogeneity assumption
OECD: homogenous GDP coefficients
0
2000
4000
6000
8000
2500 7500 12500 17500 22500 27500 32500
Per capita income
Per c
apita
CO
2
Cubic Spline TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Figure 3 CO2 –emissions: OECD countries
17
Australia
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Austria
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Belgium
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Canada
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Denmark
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Finland
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
France
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Germany
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Greece
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Iceland
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Ireland
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Italy
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
NB: Check grote verschilNB: SD TP niet beschikbaar: afhankelijk van trend
Japan
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Luxembourg
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
11000
12000
13000
14000
15000
2500 7500 12500 17500 22500 27500 32500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Netherlands
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Norway
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
New Zealand
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Portugal
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Spain
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Sweden
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Switserland
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Turkey
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
United Kingdom
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
United States
0
1000
2000
3000
4000
5000
6000
2500 7500 12500 17500 22500 27500 32500
Per capita income
Per c
apita
CO
2
Cubic TP-2SD TP TP+2SD NP-LB NP-AV NP-UB
Table 1: Descriptive statisticsa,b Variable Mean (SD) Minimum Maximum
Per Capita Carbon 2,601 (1,801) 167 12,333
Per Capita Income (1990$) 13,172 (4,992) 2,771 33,635
Population (mln) 33 (50) 0.2 275
a) Descriptive statistics are for the 24 OECD countries for the period 1960-2000 (n = 984). b) Emission levels are measured in tons.
Table 2: Main test resultsa)
[NOTE: to be completed]
Parametric Parametric Spline?? Semiparam??.
Independent variables GDP
-31.12*** (7.58)
-30.88*** (6.12)
xxx***
(7.48) xx*** (6.12)
GDP2
4.22*** (0.83)
3.80*** (0.67)
Xxx***
(0.82) xx*** (0.67)
GDP3
-0.18*** (0.03)
-0.15*** (0.02)
Xxxx*** (0.03)
xx*** (0.02)
Fixed-effects countries
Yesd Yesd Yesd Yesd
Fixed-effects years
Yesd
General trend
Yesd
Country-specific trend
Yesd Yesd
Specification tests
xx**
yy***e z*** P***e
Homogeneity tests Wald (GDP variables)
4,830*** g
1,328*** g
xx***g
xx*** g
Wald (country-specific trends)
1,867***h xx***h
Wald (all variables)
11,635***i xx***i
a) Dependent variable is CO2 emissions per capita; standard errors in parentheses. b) Wald test with H0: a1i=a1i+1 and a2i=a2i+1 and a3i=a3i+1 . c) Wald test with H0: βi= βi+1 . d) Wald test with H0: a1i=a1i+1 and a2i=a2i+1 and a3i=a3i+1 and βi= βi+1 . (*** Significant at 99% confidence interval).
18
Notes
1 This approach is slightly different from List and Gallett (1999) who showed the importance of
slope heterogeneity by including both a cross-section specific time trend and income estimators
based on a SUR estimation for individual states in the US. However, List and Gallett (1999) do not
test for the spline specification nor for the semi-parametric framework.
2 This can easily be illustrated by considering the case of two countries whose y-values do not
overlap (like Luxembourg and Turkey). In case of rejection of βLuxembourg = βTurkey homogeneity
might still be present.
3 Panel estimations are preferable due to more efficient estimators. Indications of potential of a unit
root in the series (see e.g. Egli, 2002) are of some concern. However, a recent paper on structural
breaks in carbon emissions per capita by Lanne and Liski (2003) based on much longer time series
(1878-1994) presents evidence that (endogenously determined) structural breaks are mainly located
at the beginning of the 20th century. This suggests that the oil price shock cannot be seen as such
an event. Moreover, using a deterministic trend is one way to overcome a potential unit root in the
data. Furthermore, our concern is restricted to within sample behaviour and has no aim to predict.
4 Our procedure to calculate CO2-emission from OECD energy consumption data is similar to the
approach followed by the Oak Ridge National Laboratory (ORNL) whose data are usually included
in empirical research on CO2-emisisons (see in particular Holtz-Eakin and Selden, 1995 and
Schmalensee et.al. 1998).
5 We present results only for the cubic model because the quadratic models were all clearly
rejected vis-à-vis the cubic specifications. Furthermore, both the quadratic and cubic models
without any fixed effects were also rejected. Response coefficients for the quadratic model, as well
as for models with country-fixed effects and time-fixed effects, are available upon request.
6 For the 24-spline estimation only the first two and the last splines are significant. This finding is
robust for the 20-, 16- and 12-spline specification.
7 [INCLUDE NOTE ON CONFIDENCE BOUND DUE TO FEWER OBSERVATIONS AT THE
UPPER END OF OUR OBSERVATIONS]
19
8 Steel production has been responsible for over 50% of industrial production in 1980 but has been
reduced to 3% in 2000.
9 We generate our Wald-statistics by comparing the sum of squared residuals of the general model
with and without heterogeneous coefficients for either only the GDP variables (‘traditional
models’) and/or the time-specific trend variable (general model). Because in the last case all
coefficients are country-specific, we estimated this model with time-series analysis. Although using
the Seemingly Unrelated Regressions (SUR) model potentially increases the efficiency of
estimation, the sum of squared residuals for our data is larger under SUR (3.43 versus 2.44),
indicating that time-series estimates are preferable. Also, 51% of the individual residuals do not
improve with SUR. These results are consistent with our finding that testing of the general model is
not possible, due to a near singular matrix.
10 For instance, the Wald-test on heterogeneous coefficients of the income variables for the 8-spline
model is F(140 716) = 11.67. We found similar results for 12-, 10- and the (non-preferred) 6-
spline models. Results are available upon request.
11 We also tested whether common exogenous covariates, like differences in temperature,
geological structure (mountainous landscape) or availability of (fossil fuel) resources may affect
our findings for the income variables. Interestingly, we succeeded in producing similar explanatory
power as the (parametric) model including fixed country effects without having much effect on the
income parameters. This suggest that fixed effects capture these exogenous factors rather well.
Results available on request.
12 Repeating this procedure by excluding countries with the largest LR statistics does not result in a
panel for which homogeneity cannot be rejected. Not surprisingly, we also find the same results for
models with a general trend or time-fixed effects. These results are available upon request.
20