G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 201
emission estimates as well as an approximation of environmental degradation like deforestation.
The EKC concept is dependent on the state of the economic activity. The main explanatory variable is the per capita income. As population grows and economic
growth takes off, forests are being cut to provide materials for construction, land Ž
for cultivation, etc. Usually, the larger the size of economic activity approximated .
by GDP per capita the larger is the depletion of natural resources. By the laws of thermodynamics the use of natural resources implies the production of waste. The
EKC literature assumes that there are no limits to growth. Income per capita represents consumer’s purchasing power and approximates
the consumption patterns. We accompany income per capita with population density as explanatory variables to capture both scale and composition of consump-
tion activities. We also consider the importance of some variables like the distribu- tion of GDP in manufacturing in order to examine the importance of specific
sources of pollution in the analysis. We also take into consideration some other socio-economic variables to see their influence in this kind of analysis. Finally,
dealing with the long-run character of CO emissions we include in the regressors
2
the growth rate of the economies of the countries under examination. Thus, we have used the following variables: GDP per capita in purchasing power
Ž .
Ž . parity million , 1991 ; the average annual growth rate of GDP in 1980᎐1991;
Ž .
Ž population density 1993, per 1000 hectares ; infant mortality rate per 1000 live
. Ž
. births in 1990᎐1995 ; urban population of total, 1995 ; distribution of GDP in
Ž .
Ž .
manufacturing , 1991 ; deforestation average change, 1980᎐1990 ; and CO
2
Ž .
per capita emissions in tons . The source of the last variable is Rodenburg et al. Ž
. Ž
1995 while the rest of the variables were obtained from World Resources WRI, .
1991, 1994 . Frequency distributions for the main variables of the study are reported in Fig. 1.
3
6. Discussion of results
6.1. Separating hyperplane model 6.1.1. Magnitudes of  and  y 
1 2
1
Ž .
In the deforestation model estimates of  are y0.429 and y0.469 see Table 1
1
for the normal and Student’s t, respectively, i.e. higher GNPrc leads to lower deforestation. The estimate of  y  is not statistically significant thus the two
2 1
3
The countries considered in our study are: Ethiopia, Mozambique, Bangladesh, Madagascar, Zambia, Uganda, Niger, Benin, India, Kenya, Pakistan, Nigeria, Ghana, Sri Lanka, Indonesia, Senegal,
Zimbabwe, Philippines, Morocco, Dominican R., Honduras, Thailand, El Salvador, Nicaragua, Guate- mala, Cameroon, Jamaica, Paraguay, Botswana, Ecuador, Tunisia, Colombia, Chile, Peru, Mexico,
Brazil, Panama, Hungary, Algeria, Portugal, Venezuela, Greece, Spain, Ireland, Singapore, former Czechoslovakia, UK, Italy, Australia, Belgium, the Netherlands, Austria, France, Germany, Finland,
Denmark, Canada, Japan, Norway, USA and Switzerland.
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 202
Fig. 1. Frequency distributions for main variables of the study.
regimes are not different in terms of coefficients. In the CO model the estimates
2
of  are 1.078 and 0.992, respectively. Again  y  is not significant. This can
1 2
1
be seen also from 95 Bayes probability intervals for ␥ . They are, generally, very
5
dispersed and do include the origin. We conclude that the two regimes do not differ in terms of  and  .
1 2
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 203
Table 1
a
Posterior results for the SHSR Parameters
Dependent variables Deforestation
CO
2
Normal Student’s t
Normal Student’s t
 y
0.43 y
0.47 1.07
0.99
1
Ž .
Ž .
Ž .
Ž .
0.23 0.25
0.14 0.17
 y  y
0.20 y
0.08 y
0.25 y
0.097
2 1
Ž .
Ž .
Ž .
Ž .
0.46 0.47
0.22 0.3
␥ y
0.12 y
0.39 3.19
y 0.58
1
Ž .
Ž .
Ž .
Ž .
3.44 3.56
6.84 7.6
␥ 0.4
0.31 y
8.66 y
3.04
2
Ž .
Ž .
Ž .
Ž .
4.36 4.21
8.73 11.7
␥ 1.98
0.42 y
0.47 0.053
3
Ž .
Ž .
Ž .
Ž .
5.42 5.27
6.61 6.58
␥ 0.16
y 0.32
y 4.38
y 1.47
4
Ž .
Ž .
Ž .
Ž .
3.42 3.49
6.67 7.49
␥ 1.67
0.93 y
3.73 y
1.18
5
Ž .
Ž .
Ž .
Ž .
4.66 4.52
5.76 7.47
y
0.88 y
0.87 0.48
0.43
1
Ž .
Ž .
Ž .
Ž .
0.20 0.23
0.09 0.11
y 1.83
0.59 y
0.81 y
0.82
2 1
Ž .
Ž .
Ž .
Ž .
0.18 1.02
0.083 0.09
a
Ž .
Notes . Student’s t has 1 d.f. i.e. s 1 . Posterior standard deviations in parentheses.
6.1.2. Magnitudes of and y
1 2
1
In the deforestation case the difference y is 1.833 and 0.592 for the
2 1
normal and Student’s t models with S.E.s 0.175 and 1.016. Thus according to the normal specification but not according to the Student’s t specification. In
1 2
the CO model the variance differences are y0.807 and y0.815 with S.E.s 0.083
2
and 0.086, respectively, i.e. highly significant. Thus we may conclude that the two regimes differ in terms of error
¨
ariances of the en
¨
ironmental degradation equation. This means that the variances of deforestation or CO equations are different
2
and therefore policy changes, which affect GNP will not have the same result in high-income and low-income countries because of the different variances. When
we use CO as the dependent variable we find that
so low-income
2 1
2
countries have higher dispersion. Therefore policy measures which increase GNP will decrease CO emissions with higher certainty in high-income countries. The
2
opposite is true if we use deforestation as a dependent variable. 6.1.3. Magnitudes of ␥ . . . ␥
separating hyperplane coefficients
1 5
As can be seen from their posterior means and posterior standard errors none of Ž
. the regressors urbanization, industrialization, mortality, growth or GNP appears
Ž to be a separator. The sign of ␥
the coefficient of GNP in the separating
5
. hyperplane is positive when we consider the deforestation equation and negative
when we consider the CO equation although statistically insignificant in both
2
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 204
cases. In this sense there is no e
¨
idence to support separation according to GNP or other economic and demographic
¨
ariables, which implies no support for an EKC. 6.1.4. Discussion of country membership to regimes
We found that the two regimes are different in terms of
2
although this cannot
1
be explained by economic and demographic variables. Due to the fact that estimates of ␥ appear to be insignificant, allocating countries to the two regimes
i
according to the separating hyperplane should be viewed with conservatism. However, it is useful to compare separation probabilities across models and across
dependent variables.
6.1.5. Discussion of differences between normal and Student’s t specifications Fig. 2a,b presents graphs of the probability of regime 2 for the Student’s t vs. the
normal specification, using deforestation and CO as dependent variables, respec-
2
tively. Fig. 2c,d plots the same probability for CO vs. deforestation for the
2
Student’s t and normal distributions, respectively. From Fig. 2a,b, we see that Student’s t and normal distributions give comparable results only for CO . For
2
deforestation, there is significant disagreement in the separation probabilities ᎏ although they tend to be positively correlated, and strongly so under the Student’s t
specification. Fig. 2c,d reveals significant divergence in the separation properties of different
Ž .
dependent variables. Under a normal specification Fig. 2c probabilities of regime 2 for CO and deforestation, show significant scattering around their regression
2
line. Under a Student’s t specification not only that happens but they also tend to be negatively correlated. Therefore it seems that not only the choice of dependent
variable, but also the stochastic specification of switching regression models, affects separation significantly according to environmental degradation and economic᎐
demographic factors. Thus applied researchers who venture into testing for the existence of an EKC, should not only examine alternative dependent variables and
alternative sets of regressors, but they should also examine different stochastic specifications for their switching regime or linear regression models. This adds
another dimension of model uncertainty, yet it is necessary practice because EKCs are sensitive to all these factors. To put things in a different manner, given the
choice of dependent and explanatory variables, EKCs will be sensiti
¨
e to alternative
distributional assumptions about the error term.
4
In cross-sectional studies, this is something that applied researchers should expect, and therefore they should
properly account for it in order to avoid excess sensitivity of results to the nature of the disturbance term.
6.2. Exogenous break model The most striking result that emerges from the exogenous break model is our
4
Distributional assumptions are important. This is the case in that logs vs. levels or fixed effects vs. random effects can have an impact on the results in panel regression studies.
G.E. Halkos,
E.G. Tsionas
r Energy
Economics 23
2001 191
᎐ 210
205
Ž . Ž .
Ž . Ž .
Fig. 2. Probabilities of regime 2: a deforestation; b CO ; c normal distribution; and d Student’s t.
2
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 206
Table 2
a
Ž .
Posterior results for the EBSR dependent variable: deforestation Explanatory
Regime 1 Regime 2
Difference variables
coefficients coefficients
Constant 8.16
0.220 7.94
Ž .
Ž .
Ž .
2.40 9.81
9.98 Urbanization
y 0.00038
y 0.0084
0.008 Ž
. Ž
. Ž
. 0.014
0.016 0.02
Industrialization 0.00086
0.015 y
0.14 Ž
. Ž
. Ž
. 0.016
0.047 0.051
Mortality y
0.018 0.162
y 0.180
Ž .
Ž .
Ž .
0.0086 0.246
0.246 Growth
y 0.080
0.180 y
0.260 Ž
. Ž
. Ž
. 0.088
0.389 0.403
log GNP y
0.843 y
0.181 y
0.662 Ž
. Ž
. Ž
. 0.345
1.007 1.049
2
1.24
0.202 1.037
Ž .
Ž .
Ž .
0.27 0.48
0.572
a
Posterior standard deviations in parentheses.
failure to find a posterior mean for the break z
U
significantly lower than the Ž
.
U
maximum value of z i.e. log GNPrc in our specification . Estimates of z are
9.418 and 9.453 with posterior S.E.s 0.572 and 0.430 for deforestation and CO
2
equations, respectively. These are approximately 94 of the maximum value of log GNPrc. Moreover, the differences of  y 
appear to be insignificant. The
1 i 2 i
2 2
Ž same is true for the difference y
estimates are 1.037 and 0.217 with S.E.s
2 1
. 0.572 and 0.43, respectively ᎏ see Tables 2 and 3 . Thus, we find no e
¨
idence for a structural break from log GNP r c.
For the deforestation model, mortality rates and log GNPrc are significant in the first regime but none of the variables is statistically significant in the second
regime. For the CO model only industrialization and log GNPrc are significant in
2
the first regime. Notice that higher GNP implies lower deforestation in the first regime but higher CO
emissions. We conclude that measuring en
¨
ironmental
2
degradation using deforestation or CO makes a difference.
2
Only for the deforestation model we find that
2
y
2
is close to significance
2 1
Ž .
1.037 and with S.E. 0.572 . So it turns out that we find very little evidence to support the existence of two regimes with differences in coefficients andror error
variances. 6.2.1. Marginal posterior distribution of z
U
Marginal posterior distributions of z
U
for the deforestation and CO model are
2
shown in Fig. 3. From inspection it follows that the mass is concentrated between 90 and 95 of the maximum value of z s log GNP. It is remarkable that for
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 207
Table 3
a
Ž .
Posterior results for the EBSR dependent variable: CO
2
Explanatory Regime 1
Regime 2 Difference
variables coefficients
coefficients Constant
y 6.23
6.02 y
12.25 Ž
. Ž
. Ž
. 1.34
14.33 14.31
Urbanization 0.0063
0.0076 y
0.0013 Ž
. Ž
. Ž
. 0.007
0.0203 0.0217
Industrialization 0.025
y 0.022
0.048 Ž
. Ž
. Ž
. 0.009
0.045 0.045
Mortality y
0.0057 0.362
y 0.368
Ž .
Ž .
Ž .
0.0049 0.278
0.28 Growth
0.0448 0.434
y 0.389
Ž .
Ž .
Ž .
0.050 0.365
0.369 log GNP
0.771 y
0.731 1.502
Ž .
Ž .
Ž .
0.185 1.468
1.467
2
0.406
0.188 0.217
Ž .
Ž .
Ž .
0.09 0.418
0.430
a
Posterior standard deviations in parentheses.
both the log CO and deforestation models these distributions are approximately
2
the same.
Ž .
Fig. 3. Posteriors of break point relative to maximum log GNP .
G.E. Halkos, E.G. Tsionas r Energy Economics 23 2001 191᎐210 208
7. Concluding remarks