![, here]](https://www.researchgate.net/profile/Zacharias-Bragoudakis/publication/390691444/figure/tbl1/AS:11431281382527167@1744873737285/here_Q320.jpg)
Content uploaded by Zacharias Bragoudakis
Author content
All content in this area was uploaded by Zacharias Bragoudakis on Apr 17, 2025
Content may be subject to copyright.
BANK OF GREECE
EU RO SY ST EM
Working Paper
Economic Research Department
S p e ci al S tu di es D iv is io n
2 1 , E . Ve ni ze l os Av en ue
G R - 1 0 2 5 0 , A t h e n s
Tel.: + 3 0 2 1 0 3 2 0 3 6 1 0
Fax: + 3 0 2 1 0 3 2 0 2 4 3 2
w w w . b a n k o f g r e e c e . g r
BANK OF GREECE
EU RO SY ST EM
WORKINGPAPERWORKINGPAPERWORKINGPAPERWORKINGPAPER
ISSN: 1109-6691
BANK OF GREECE
EU RO SY ST EM
Working Paper
Economic Research Department
S p e ci al S tu di es D iv is io n
2 1 , E . Ve ni ze l os Av en ue
G R - 1 0 2 5 0 , A t h e n s
Tel.: + 3 0 2 1 0 3 2 0 3 6 1 0
Fax: + 3 0 2 1 0 3 2 0 2 4 3 2
w w w . b a n k o f g r e e c e . g r
BANK OF GREECE
EU RO SY ST EM
WORKINGPAPERWORKINGPAPERWORKINGPAPERWORKINGPAPER
ISSN: 1109-6691
BANK OF GREECE
EU RO SY ST EM
Working Paper
Economic Research Department
S p e ci al S tu di es D iv is io n
2 1 , E . Ve ni ze l os Av en ue
G R - 1 0 2 5 0 , A t h e n s
Tel.: + 3 0 2 1 0 3 2 0 3 6 1 0
Fax: + 3 0 2 1 0 3 2 0 2 4 3 2
w w w . b a n k o f g r e e c e . g r
BANK OF GREECE
EU RO SY ST EM
WORKINGPAPERWORKINGPAPERWORKINGPAPERWORKINGPAPER
ISSN: 1109-6691
BANK OF GREECE
EU RO SY ST EM
Working Paper
Economic Research Department
S p e ci al S tu di es D iv is io n
2 1 , E . Ve ni ze l os Av en ue
G R - 1 0 2 5 0 , A t h e n s
Tel.: + 3 0 2 1 0 3 2 0 3 6 1 0
Fax: + 3 0 2 1 0 3 2 0 2 4 3 2
w w w . b a n k o f g r e e c e . g r
BANK OF GREECE
EU RO SY ST EM
WORKINGPAPERWORKINGPAPERWORKINGPAPERWORKINGPAPER
ISSN: 1109-6691
3
Zacharias Bragoudakis
Emmanouil Taxiarchis Gazilas
Evidence from the EKC framework
Does primary and secondary education contribute to
environmental degradation?
APRIL 2025
WORKINGPAPERWORKINGP
APERWORKINGP
APERWORKINGPAPERWORKINGP
APER
4
1
BANK OF GREECE
Economic Analysis and Research Department – Special Studies Division
21, Ε. Venizelos Avenue
GR-102 50 Athens
Τel: +30210-320 3610
Fax: +30210-320 2432
www.bankofgreece.gr
Published by the Bank of Greece, Athens, Greece
All rights reserved. Reproduction for educational and
non-commercial purposes is permitted provided that the source is acknowledged.
ISSN: 2654-1912 (online)
DOI: https://doi.org/10.52903/wp2025341
DOES PRIMARY AND SECONDARY EDUCATION
CONTRIBUTE TO ENVIRONMENTAL DEGRADATION?
EVIDENCE FROM THE EKC FRAMEWORK
Zacharias Bragoudakis
Bank of Greece
Emmanouil Taxiarchis Gazilas
University of Piraeus
ABSTRACT
This paper investigates the impact of education on the Environmental Kuznets Curve
(EKC) hypothesis using a balanced panel dataset of 167 countries over 21 years. By
employing three econometric models with CO₂ emissions, NOx emissions, and total
greenhouse gas emissions as dependent variables, we analyze the role of primary and
secondary education in shaping environmental outcomes. Our results confirm the
presence of an N-shaped EKC, suggesting that economic growth initially worsens
environmental degradation, followed by an improvement, and later a potential rebound
in emissions. More importantly, we find that education plays a significant role in
environmental dynamics: higher enrolment in both primary and secondary education is
associated with increased emissions, particularly in developing economies, possibly
due to the expansion of industrial activity and energy consumption linked to a more
skilled workforce. However, at higher levels of economic development, education may
contribute to environmental awareness, innovation, and policy implementation that
foster sustainable practices. These findings highlight the need for targeted educational
policies that integrate environmental sustainability to ensure long-term ecological
benefits.
Keywords: Environmental Kuznets Curve (EKC); Education; CO₂ Emissions; Nox
Emissions; Greenhouse Gases
JEL Codes: Q53, Q56, I25, O44, C33
Disclaimer : The views expressed in this paper are those of the authors and not
necessarily those of the Bank of Greece.
Correspondence:
Zacharias Bragoudakis
Economic Analysis and Research Department
Bank of Greece
El.Venizelos 21, 10250 Athens, Greece
Tel.: +30-2103203605
email: zbragoudakis@bankofgreece.gr
3
1. Introduction
In recent years, parametric and semiparametric panel data approaches have been
used extensively to study the Environmental Kuznets Curve (EKC) hypothesis. These
investigations have produced inconsistent and often contentious results (e.g., Apergis
et al., 2017; Halkos, 2003; Cole, 2004; Millimet et al., 2003; Zaim and Taskin, 2000).
The EKC claims that environmental deterioration first increases with economic
development due to the "scale effect" of industrial expansion. However, after a certain
income threshold, environmental degradation begins to decrease as cleaner technology
and more efficient manufacturing processes emerge—known as the "technique" and
"composition" impacts.
However, education plays a key role in identifying these processes. The influence
of educational enrollment on pollutant emissions can also be separated out using scale,
technique, and composition impacts. Higher primary and secondary school enrollment
may initially lead to higher emissions as economic activity increases ("scale effect").
However, higher levels of knowledge may incentivize companies to adopt more
environmentally friendly manufacturing practices and support societal shifts toward
sustainable practices ("technique" and "composition" impacts). Accordingly, the EKC
hypothesis suggests that pollution will decrease as a result of the composition and
method effects becoming more apparent at higher income levels while the scale impact
predominates at lower income levels (Jayanthakumaran and Liu, 2012).
Despite a variety of limitations, the EKC has been the focus of extensive research.
First, many recent studies assume that random disturbances occur across panel
dimensions or that variables are cross-sectionally independent. This assumption is
commonly broken in macroeconomic datasets due to unobserved common causes, such
as changes in environmental legislation worldwide, which results in biased and
unreliable conclusions. Second, most studies do not explore the interplay between
environmental outcomes, economic development, and education.
As mentioned earlier, our study of the relationship between economic growth and
education and environmental deterioration is framed by the Environmental Kuznets
Curve (EKC) hypothesis. Similar to financial markets, education has a distinct impact
on economic and environmental outcomes through processes such as "scale,"
"technique," and "composition" impacts. For instance, whereas secondary education
4
fosters creativity and abilities that support cleaner technologies and sustainable
practices, primary education may promote industrialization and economic expansion,
which could increase emissions. These intricate connections demonstrate how
important it is to incorporate education into the EKC framework in order to understand
its effects on the environment.
This study aims to bridge these gaps by investigating the ways in which economic
growth and educational enrollment impact the validity of the EKC hypothesis. Using a
balanced panel dataset of 167 countries from 2000 to 2020, the study accounts for cross-
sectional dependence using econometric techniques such the Pesaran (2004) CD test.
By including education factors into a static and dynamic EKC framework, this study
seeks to uncover the intricate links between primary and secondary education and CO₂,
NOₓ, and other greenhouse gas emissions. By providing a more comprehensive
understanding of how education influences environmental outcomes in the context of
economic development, the findings are meant to add to the broader discussion on
sustainable growth and environmental policy.
Research Questions
RQ1: Does the Environmental Kuznets Curve (EKC) hypothesis hold for CO₂,
NOₓ, and other greenhouse gas emissions across 167 countries from 2000 to 2020?
RQ2: How do primary and secondary education enrollment levels affect the
relationship between economic development and environmental degradation?
Research Hypotheses
H1: The EKC hypothesis is valid, with CO₂, NOₓ, and greenhouse gas emissions
initially rising with GDP per capita but declining after a critical income threshold.
H2: Higher primary and secondary education enrollment contribute to increased
emissions due to "scale effects."
The paper is organized as follows: Section 2 provides an extensive literature
review on the impact of economic growth and education level on the Environmental
Kuznets Curve Hypothesis. Section 3 presents the data and the econometric
methodology, while in Section 4 the empirical results and discussion are presented.
Finally, in Section 5 some concluding remarks are summarized.
5
2. Literature review
For their survival and development, humans depend on a wide range of
environmental resources, such as oxygen from the atmosphere, food from aquatic and
terrestrial ecosystems, and energy from coal, oil, and other natural resources. Even if
these resources increase economic growth and raise living standards, one of the main
environmental repercussions of their extraction is the emission of pollutants such as
carbon dioxide (CO₂), nitrogen oxides (NOₓ), and other greenhouse gases (GHGs).
Concern over climate change and global warming is increasing as a result of these
emissions (Solomon et al., 2009; Jones et al., 2016; Jackson et al., 2019; Fuss et al.,
2014; Stocker et al., 2013).
The intricate relationship between economic growth and environmental
degradation has been studied using frameworks such as the Environmental Kuznets
Curve (EKC), which suggests that environmental degradation initially rises as a
country's income rises but eventually falls once a certain income threshold is reached.
A nation may adopt cleaner technology and better environmental practices if it reaches
a certain level of prosperity, which could result in a decrease in emissions like CO₂ and
NOₓ, according to this inverted U-shaped relationship (Grossman & Krueger, 1995;
Stern, 2004; Panayotou, 1993; Cole, Rayner, & Bates, 1997). However, this relationship
is influenced by many factors outside of wealth. . Education is one of the most
significant elements that can affect how societies engage with their environment. It can
affect choices and actions related to pollution, resource use, and environmental
conservation. In particular, education for sustainable development (ESD) emphasizes
the need to integrate environmental considerations into educational curricula and
practices to help people understand the long-term effects of their actions on the
environment (Leicht, Heiss, & Byun, 2018; UNESCO, 2012, 2014, 2017; Sterling,
2004; Tilbury, 2011; Hopkins & McKeown, 2002; Wals, 2007; Jickling & Wals, 2008;
Orr, 1992).
However, as education levels increase, people become more aware of
environmental issues, leading to more sustainable consumption patterns and a greater
willingness to support policies that reduce pollution and protect ecosystems (Zsóka et
al., 2013; UNESCO, 2012; Stern & Dietz, 1994; Schultz & Zelezny, 2003; Poortinga et
al., 2004). Education also promotes the development of eco-friendly practices and green
6
technologies on a personal and social level. In the past, economic growth as measured
by GDP per capita has been associated with increased energy use and pollution.
Growing income levels are typically linked to rising energy use, which raises CO₂, NOₓ,
and other GHG emissions, especially in emerging countries with industrializing
economies. However, when countries' incomes increase, they may have the resources
and incentive to invest in more environmentally friendly technologies and enforce
stricter environmental regulations. Education is essential throughout this shift since
educated individuals are more likely to advocate for environmental sustainability and
support laws intended to reduce emissions (Hines et al., 1987; O'Neill & Nicholson-
Cole, 2009; Stevenson, 2007).
Educational attainment, particularly at the basic and secondary levels, can
influence these processes by providing people with the knowledge and skills to make
informed decisions for environmental preservation. Primary education increases
awareness of environmental issues at a young age, whereas secondary education helps
people get a deeper understanding of complex environmental challenges and solutions.
Thus, primary and secondary school enrollment can directly impact national attitudes
toward sustainability and indirectly contribute to lowering CO₂, NOₓ, and GHG
emissions (Barro, 2001; Gylfason, 2003; Cole & Neumayer, 2004; Sato & Vörösmarty,
2016; Adger & Kelly, 1999).
Due in large part to international initiatives like the United Nations' Decade for
Education for Sustainable Development (DESD) (2005-2014), the idea of education for
sustainable development (ESD) has gained popularity in recent years. ESD aims to
integrate sustainability into educational institutions worldwide so that future
generations have the knowledge, values, and skills necessary to address environmental
concerns. ESD encourages responsible behavior, fosters a deep understanding of
environmental issues, and motivates individuals to take action to lower emissions and
safeguard the environment (Tilbury, 1995; UNESCO, 2005; Hopkins & McKeown,
2002; Wals, 2011). ESD's primary focus has historically been environmental education,
but recent studies show that it also makes a substantial contribution to the creation of
sustainable, carbon-emission-free economic growth.
For instance, countries with greater levels of knowledge are more likely to switch
to sustainable energy sources and use resources faster and more efficiently. By
analyzing the relationship between education (primary and secondary enrollment rates)
7
and environmental indicators like CO₂, NOₓ, and GHGs (greenhouse gases), it is
possible to assess how education contributes to sustainable development outcomes in
different countries (Cole & Neumayer, 2004; Khan & Banu, 2017; Sato & Vörösmarty,
2016).
In conclusion, education for sustainable development must be integrated into
national policy and educational systems in order to stop environmental degradation.
Examining the impact of GDP per capita and basic and secondary education on
environmental indicators such as CO₂, NOₓ, and GHGs (greenhouse gases) may provide
additional insight into how education influences environmental outcomes in connection
to economic growth. Higher education enrollment, particularly at the secondary level,
is anticipated to be associated with reduced emissions, helping countries balance
economic growth with environmental sustainability.
3. Data and Methodology
The econometric estimation in this study utilizes an unbalanced panel dataset
encompassing 167 countries over a 21-year period (n = 167 and T = 21), spanning from
[specific years, e.g., 2000–2020. The dependent variables—CO₂ emissions, NOₓ
emissions, and GHG emissions (metric tons of CO₂ equivalent)—are sourced from the
World Bank’s World Development Indicators Database, providing reliable and
standardized environmental indicators (see Table 1).
[Insert Table 1, here]
The independent variables include educational metrics such as primary education
enrollment (prmpul) and secondary education enrollment (secpup), alongside economic
variables like GDP per capita (gdpc) and its higher-order terms (gdpc² and gdpc³) to
capture potential nonlinear relationships between economic development, education,
and environmental outcomes.
The missing values of the variables of interest for the recent years of the time
span under consideration were predicted using moving average, single, and double
exponential smoothing techniques, while interpolation was employed when necessary,
in the case of missing values. Accuracy metrics including Mean Absolute Percentage
Error (MAPE), Mean Absolute Deviation (MAD), and Mean Squared Deviation (MSD)
were used to help choose the best approach. Smaller values signify a better-fitting
8
model, and using these statistics allows us to compare various forecasting fits and
smoothing techniques.
While concerns may raise regarding the potential distortion of stationarity and
cointegration properties due to interpolation, the scale of imputation in this study is
minimal and unlikely to affect the results meaningfully. The panel consists of 21 years
across 167 countries, yielding a total of 28,056 observations across eight variables.
Before interpolation, we had 27,980 complete observations, meaning only 76 values
(0.27%) were imputed. This negligible proportion ensures that any risk of artificially
inducing trends, persistence, or biasing unit root and cointegration tests is virtually non-
existent.
The use of a balanced panel allows for robust econometric analysis, avoiding
issues of missing data that can complicate interpretations. This dataset, with its global
coverage, enables a comprehensive examination of the interplay between education and
emissions, offering new insights into how education might influence CO₂, NOₓ, and
greenhouse gas dynamics at a cross-country level.
In Table 2 we provide a summary of descriptive statistics for the dependent and
explanatory variables, while in Table 3 the corresponding diagonal correlation matrix
is presented.
[Insert Table 2, here]
With mean values of 204,946, 275,579, and 16,291 metric tons, respectively, CO₂
emissions, NOₓ emissions, and total greenhouse gas (GRHGAS) emissions show
significant dispersion. With CO₂ emissions as high as 11 million metric tons, the huge
standard deviations show that certain nations contribute disproportionately to global
pollution, while others have comparatively low emission levels. The variation in
greenhouse gas emissions points to significant differences in national economic
systems, patterns of energy usage, and environmental regulations. NOₓ emissions,
which are frequently linked to traffic and industrial activities, exhibit a similar trend,
with some nations maintaining very low emissions and others surpassing half a million
metric tons.
The sharp difference between high- and low-income countries is demonstrated by
the GDP per capita (GDPC), which can range from as low as $137 to over $204,000.
The high standard deviation and mean GDP per capita of $19,065 indicate that
9
economic inequality may be a significant factor in determining emissions patterns and
environmental policies. These disparities are further highlighted by education variables
like primary school enrollment (PRMPUL) and secondary school enrollment
(SECPUP). The wide variations in population size and educational access are reflected
in the enrollment in primary education, which ranges from 1,235 students to 140
million, and secondary education, which ranges from 508 students to 130 million.
With correlation values of 0.996 between CO₂ and GRHGAS and 0.914 between
CO₂ and NOₓ, it is predictably the case that CO₂ emissions, total greenhouse gas
(GRHGAS) emissions, and nitrogen oxide (NOₓ) emissions are closely connected (see
Table 3). The three pollutants appear to be closely associated, according to their strong
positive correlations, suggesting that nations with high CO₂ emissions also have high
NOₓ and other greenhouse gas emissions. The fact that CO₂ contributes significantly to
total greenhouse gas emissions is probably the reason for the almost perfect correlation
between CO₂ and GRHGAS. Similarly, the association between NOₓ and GRHGAS
(0.938) supports the notion that shared industrial and economic activities are the source
of several types of pollution. The robustness of these associations is confirmed by their
statistical significance (p-values = 0.000).
[Insert Table 3, here]
However, there is a far smaller correlation between GDP per capita (GDPC) and
emissions. GDPC and CO₂ have a correlation of just 0.0479, and NOₓ has an even lower
correlation of 0.006, which is statistically insignificant (p = 0.688). The Environmental
Kuznets Curve (EKC) hypothesis, which contends that emissions first increase with
economic expansion before eventually declining at higher income levels, is consistent
with this weak association. This non-linear relationship is not discernible using
straightforward correlation analysis. It's interesting to note that there are significant
positive relationships between emissions and education variables, specifically primary
school enrollment (PRMPUL) and secondary school enrollment (SECPUP). There is a
strong correlation between PRMPUL and SECPUP and GRHGAS (0.620 and 0.657),
CO₂ (0.586 and 0.619), and NOₓ (0.732 and 0.793, respectively).
These correlations imply that higher emissions are initially linked to higher
educational enrollment, most likely as a result of the growth of economic and industrial
activity that comes with greater literacy and labor force involvement. The notion that
10
education is a key factor in determining economic and environmental dynamics is
supported by the substantial connection (0.919) between primary and secondary
education, which shows that nations with high primary enrollment also have robust
secondary education systems. The necessity for policies that combine education with
sustainable environmental measures is further highlighted by the negative correlation
between GDPC and education variables (-0.097 for PRMPUL and -0.096 for SECPUP),
which indicates that higher education enrollment is more common in developing
economies.
Similarly to other empirical studies (see for example Millimet et al., 2003;
Apergis,2016), we first estimate separately the following (polynomial) panel data
models in a static form. The degree of the polynomial for each equation has been
determined by the maximum number of statistically significant powers.
Where ,
and are the in metric tons pollution in country
at time ; and are country and time fixed effects used in order to capture common
factors across the cross-sectional element;
is real GDP per capita (powers) for
country at time , and PRMPUL and SECPUL are the primary and secondary
education enrollment (total number of pupils). Finally, are zero mean i.i.d. errors.
The basic model of unobserved effects may be expressed as:
11
The first method used is the fixed effects (FE) estimator, allowing a different
intercept for every country and treating the constants as regression parameters.
To account for potential cross-sectional dependence (CD) in our panel dataset,
we perform four widely used tests: Breusch-Pagan LM (1980), Pesaran Scaled LM
(2004), Bias-Corrected Scaled LM (2008), and Pesaran CD (2004). Cross-sectional
dependence arises when shocks affecting one country spill over to others, which is
particularly relevant for global environmental and economic studies. Ignoring CD can
lead to biased standard errors and misleading statistical inferences, making these tests
crucial for ensuring the robustness of our econometric approach.
The Breusch-Pagan LM (5) test is a classical test for CD, particularly suitable for
panels with a large number of cross-sections (N) and a small-time dimension (T). It
tests whether residuals are correlated across countries, with a significant test statistic
indicating the presence of CD. However, this test has limitations in large panels, as it
tends to over-reject the null hypothesis of cross-sectional independence.
Where:
• is the estimated correlation coefficient of residuals between cross-sectional
units i and j.
• N is the number of cross-sectional units.
The test statistic follows a x2 distribution with
degrees of freedom under
the null hypothesis of no cross-sectional dependence.
To address the shortcomings of the Breusch-Pagan LM test, we use the Pesaran
Scaled LM (6) test , which adjusts for the number of cross-sections and ensures more
reliable results in large panels. A statistically significant result confirms the presence of
CD. Additionally, the Bias-Corrected Scaled LM (7) test further refines the Pesaran
Scaled LM test by adjusting for bias in small samples. This correction improves the
12
accuracy of the test in finite samples, making it a more reliable indicator of cross-
sectional dependence.
Where:
• is the squared correlation coefficient of residuals.
• T is the number of time periods.
Under the null hypothesis, is asymptotically standard normal.
Where:
• The bias correction term
improves small-sample performance.
Under the null hypothesis LMBC follows a standard normal distribution asymptotically.
Lastly, we apply the Pesaran CD (8) test, which is particularly effective for large
panels and remains valid even when the time dimension is relatively small. Unlike the
previous tests, which are based on sum-of-squared residual correlations, the Pesaran
CD test is based on pairwise correlation coefficients of residuals. A statistically
significant result suggests that CD is present across countries, indicating that
environmental and economic shocks in one country influence others.
Where:
• is the pairwise correlation of residuals.
13
Under the null hypothesis of no cross-sectional dependence, CD is asymptotically
standard normal.
To examine the stationarity properties of our variables, we apply three panel unit
root tests: Im, Pesaran, and Shin (IPS) W-stat, ADF - Fisher Chi-square, and PP - Fisher
Chi-square. These tests allow us to assess whether the variables exhibit unit roots,
ensuring the appropriateness of our econometric methods.
The Im, Pesaran, and Shin (IPS) W-stat (9) test extends the traditional Dickey-
Fuller test to a panel setting by averaging individual unit root test statistics across cross-
sections. Unlike methods that assume a common autoregressive coefficient for all units,
the IPS test allows for heterogeneity in the persistence of the series across countries. A
rejection of the null hypothesis (which states that all series contain a unit root) suggests
that at least some of the series are stationary. This flexibility makes IPS particularly
useful in our dataset, given the differences in economic and educational development
across the 167 countries in our sample.
Where:
• is the ADF t-statistic for each individual time series.
• is the number of cross-sectional units.
IPS shows that under the null hypothesis H0 all series have unit roots, the standardized
W-stat follows a standard normal distribution:
Where are mean and variance of the ADF statistic under H0.
• Null Hypothesis (H0): All series contain a unit root.
• Alternative Hypothesis (H1): Some (but not necessarily all) series are stationary.
The ADF - Fisher Chi-square (11) test, proposed by Maddala and Wu (1999),
aggregates p-values from individual Augmented Dickey-Fuller (ADF) tests conducted
for each country in the panel. This method does not require a balanced panel and is
useful in accounting for cross-sectional heterogeneity. By combining information from
14
multiple independent unit root tests, the ADF-Fisher test provides a robust measure of
stationarity. If the test rejects the null hypothesis, it indicates that at least one country
in the sample has a stationary series, supporting the presence of stationarity in the
dataset.
-
Where:
• is the p-value from the ADF unit root test for cross-section .
• Under the null hypothesis x2 follows a Chi-square distribution with 2N degrees
of freedom.
• Null Hypothesis (H0): All series have a unit root.
• Alternative Hypothesis (H1): At least one series is stationary.
Finally, the PP - Fisher Chi-square (12) test, based on the Phillips-Perron
methodology, is similar in approach to the ADF-Fisher test but accounts for serial
correlation and heteroskedasticity without requiring lag selection. It is particularly
useful for handling structural breaks and heterogeneity in the data. Like the ADF-Fisher
test, it combines the results of individual country-level Phillips-Perron tests to produce
an overall test statistic for the panel. A significant result suggests that at least one of the
panel series is stationary, reinforcing the conclusions drawn from the other unit root
tests.
Where:
• is the p-value from the ADF unit root test for cross-section .
• Under the null hypothesis x2 follows a Chi-square distribution with 2N degrees
of freedom.
• Null Hypothesis (H0): All series have a unit root.
• Alternative Hypothesis (H1): At least one series is stationary.
15
To investstigate the long-run relationship between economic development,
education, and emissions, we employ the Pedroni (1999, 2004) cointegration tests,
which extend the Engle-Granger framework to a panel data setting. These tests assess
whether a stable long-run equilibrium exists among the variables, allowing for cross-
country heterogeneity. Specifically, we use four test statistics:
The Panel PP-Statistic (13) and Panel ADF-Statistic (14) fall under the within-
dimension category, meaning they pool data across all countries. The Panel PP-Statistic,
based on the Phillips-Perron (PP) test, accounts for serial correlation and
heteroskedasticity in the residuals while testing for unit roots. A significantly negative
test statistic provides evidence of cointegration, suggesting that emissions, education,
and economic growth move together in the long run. Similarly, the Panel ADF-Statistic,
based on the Augmented Dickey-Fuller (ADF) test, tests whether the residuals are
stationary, offering an alternative measure of cointegration. A statistically significant
result indicates that the variables maintain a stable relationship over time.
Δ
Where:
• : Residuals from the cointegrating regression for unit at time
• First difference of the residuals
• : Adjustment term for serial correlation in the residuals for unit
• N: Number of cross-sectional units.
• T: Number of time periods.
Under the null hypothesis (H0) all series contain a unit root, and the statistic follows a
standard normal distribution asymptotically.
Δ
Where:
• : Residuals from the cointegrating regression for unit at time
• First difference of the residuals
• : Adjustment term for serial correlation in the residuals for unit
16
• N: Number of cross-sectional units.
• T: Number of time periods.
The null hypothesis is that all panels have a unit root, while the alternative
suggests stationarity.
The Group PP-Statistic (15) and Group ADF-Statistic (16) fall under the between-
dimension category, meaning they allow for greater heterogeneity across countries. The
Group PP-Statistic, like the Panel PP-Statistic, is derived from the Phillips-Perron
methodology but treats each country separately rather than pooling them. A significant
and negative value implies that at least one country exhibits cointegration. Similarly,
the Group ADF-Statistic, based on individual ADF regressions for each country,
provides further robustness by testing for unit roots in the residuals without assuming
a common autoregressive coefficient across countries. If this test rejects the null
hypothesis, it confirms the presence of cointegration in at least some cross-sections.
Δ
Where is the Phillips-Perron statistic for each individual cross-section.
• : Residuals from the cointegrating regression for unit at time
• First difference of the residuals
• : Adjustment term for serial correlation in the residuals for unit
• N: Number of cross-sectional units.
• T: Number of time periods.
Under the null hypothesis (H0), all series have a unit root, while the alternative
suggests at least some series are stationary.
Δ
Where is the ADF test statistic for each cross-section.
• : Residuals from the cointegrating regression for unit at time
• First difference of the residuals
• : Adjustment term for serial correlation in the residuals for unit
17
• N: Number of cross-sectional units.
• T: Number of time periods.
Again, under H0, all series contain a unit root, while under H1, at least some are
stationary.
4. Results And Discussion
Panel Cross-section Dependence Test
In panel data models, it often seems that disturbances are cross-sectionally
independent, particularly when the cross-section dimension is large. Nonetheless, there
is strong evidence that panel regression settings frequently exhibit cross-sectional
dependence. Ignoring cross-sectional dependency in estimate can have detrimental
effects; if residual reliance is not taken into consideration, estimator efficiency will be
lost, and test statistics will be deemed invalid.
The potential correlation between the variables or random disturbances across the
panel dimension is one of the extra issues that come up when working with panel data
as opposed to the pure time-series situation. The assumption that there was no CD was
made in the early literature on unit root and cointegration tests. This assumption is
frequently broken by macro-level data, though, which causes poor power and size
distortions in tests that rely on cross-sectional independence. For instance, widespread
unobserved effects of changes in national environmental laws may be the cause of CD
in our data. Thus, we check for CD before moving on to the unit root and cointegration
tests. We use the CD tests proposed by Breusch-Pagan (1980) LM, Pesaran (2004)
scaled LM, Baltagi, Feng, and Kao (2012) bias-corrected scaled LM and Pesaran. The
tests are based on the estimation of the linear panel model of the form
where T and N are the time and panel dimensions respectively, the country-
specific intercept, vector of regressors and the random disturbance
term.
The null hypothesis in both tests assume the existence of cross-sectional
correlation: for all t and for all . This is
18
tested against the alternative hypothesis that for at least one pair of
. The Pesaran (2004) tests are a type of Lagrange multiplier test that is based on
the errors obtained from estimating Equation (20) by the OLS method.
In consideration of the previously stated, we conduct the initial empirical analysis
by looking into the existence of CD. Considering the statistical significance of the CD
statistics, all tests provide evidence of CD in the data by strongly rejecting the null
hypothesis of cross-sectional independence (P-value = 0.000) for all models. Given this
data, we use tests that are resistant to CD (referred to as "second generation" tests) to
determine whether unit roots exist (see Table 4).
[Insert Table 4, here]
With P-values of 0.000 for all models, all tests significantly reject the null
hypothesis of cross-sectional independence, suggesting that the data contains cross-
sectional dependence (CD). This conclusion implies that our models' residuals are not
cross-sectionally independent, which is essential for guaranteeing the validity of our
findings. We use tests specifically designed to be robust to cross-sectional dependency,
called ‘second generation,' tests, to investigate the existence of unit roots given the
statistical significance of the CD statistics. By taking into consideration the detected
cross-sectional dependence, these tests enable more trustworthy conclusions about the
data's stationarity.
Panel Unit Root Tests
Panel unit root tests both under the assumption of cross-section independence and
allowing for cross-section dependence. We perform panel unit root tests under both the
assumption of cross-section independence and allowing for cross-section dependence.
Specifically, we apply three independent cross-section panel unit root tests: Pesaran
and Shin (2003), Fisher-type tests using ADF and PP tests (Maddala and Wu, 1999;
Choi, 2001), and Hadri (2000).
To assess the stationarity properties of the variables in our models, we utilize the
‘second-generation’ unit root tests for panel data. This approach is particularly suited
for handling non-linear functions of I (1) variables, as is the case in our study where
GDP is included both in its level and in quadratic and cubic forms (Apergis, 2016). For
19
this purpose, we employ the Fisher test, as proposed by Maddala and Wu (1999), which
accounts for cross-sectional dependence in an unbalanced panel dataset. This
methodology is based on the p-values of individual unit root tests and assumes that all
series are non-stationary under the null hypothesis, with the alternative hypothesis
positing that at least one series in the panel is stationary.
Unlike the Im–Pesaran–Shin (1997) test, the Fisher test does not require a
balanced panel, making it well-suited for our dataset. This flexibility ensures that the
unit root testing results are robust and reliable, even in the presence of an unbalanced
panel structure.
Panel unit root test: Summary
The presence of unit roots across all sample variables is confirmed by the panel
unit root tests that were performed, specifically the PP-Fisher Chi-square, the ADF-
Fisher Chi-square, the Im, and the Pesaran and Shin W-statistic (see Table 5). None of
the variables are integrated of an order greater than one I (1), according to these tests,
which offer strong evidence that the variables under investigation only show
stationarity after first differencing. The validity of further econometric estimations is
guaranteed, and the trustworthiness of this conclusion is strengthened by the
consistency of these results across various testing techniques. The findings allay
worries about false regression problems that could occur from non-stationary data by
verifying the lack of higher-order integration. Additionally, the validation of I (1)
integration is consistent with common assumptions in panel data econometrics,
allowing for the proper use of estimate methods that depend on stationarity following
differencing, including fixed effects or dynamic panel models. These results are
essential for guaranteeing the methodological soundness and empirical validity of the
connections examined between economic factors, educational indicators, and
environmental consequences. Overall, the panel unit roots tests confirm that all sample
variables have a unit root. Stated otherwise, the test findings indicate that none of the
variables are integrated to a level higher than one (I-1).
[Insert Table 5, here]
20
Estimation of regressions
Moment estimators for the unconditional variances are used in place of residuals
in the subsequent techniques, which are improved versions of the original White
statistics. These methods, which are based on the Panel Corrected Standard Error
(PCSE) technique first presented by Beck and Katz (1995), are intended to handle
unconditional variance matrices with no limits while placing further limitations on
conditional variance matrices. The conditional variances matching the unconditional
variances is a sufficient, but not a necessary, criterion for using PCSE methods.
Furthermore, the variance structures must be constant across cross-sections and time
periods, much like with the SUR estimators. Only the diagonal elements of the cross-
section and period covariance matrices are used by the diagonal versions of these
estimators, known as Cross-section weights (PCSE). These estimators are not made to
deal with general residual correlation, even though they are resilient against
heteroskedasticity across cross-sections or periods. Lastly, the non-degree-of-freedom-
corrected variants of these estimators further customize them to particular panel data
sets by streamlining the calculation by eliminating the leading term involving the
number of observations and coefficients. The regression results according to Cross-
section weights (PCSE) are presented in Tables 6,7,8.
CO₂ Emissions (CO₂) Regression Analysis
The GDP per capita (GDPC) coefficient is positive and statistically significant (β
= 7.048, p = 0.001), indicating that at lower levels of income, economic growth
contributes to rising CO₂ emissions (see Table 6). However, the squared (GDPC²) and
cubic (GDPC³) terms of GDP per capita are also significant, with GDPC² having a
negative coefficient (β = -0.0001, p = 0.000) and GDPC³ having a positive coefficient
(β = 3.69E-10, p = 0.0001). This confirms the presence of an N-shaped Environmental
Kuznets Curve (EKC), where emissions first rise with economic growth, then decline,
but eventually increase again at higher income levels. This suggests that economic
development alone does not guarantee long-term environmental sustainability, as
emissions may rise again after surpassing a certain income threshold.
[Insert Table 6, here]
21
Education has a noteworthy effect on CO2 emissions as well. Higher primary
school enrollment is linked to higher CO₂ emissions, most likely as a result of the scale
effects of economic expansion, according to the strong positive and highly significant
influence of primary education enrollment (PRMPUL) on emissions (β = 0.017, p =
0.000). Although there is a positive correlation between emissions and secondary
education enrollment (SECPUP), the magnitude and statistical significance of this
relationship are smaller (β = 0.002, p = 0.055). This could suggest that secondary
education has a more complicated or delayed effect on emissions, either as a result of
policy participation, technological developments, or heightened environmental
consciousness. With an R-squared of 0.943, the overall model fit is strong. However,
this high explanatory power is largely driven by the inclusion of country fixed effects,
which control for unobserved heterogeneity across countries. While the explanatory
factors contribute to the variation in CO₂ emissions, the fixed effects play a crucial role
in capturing structural differences across countries.
Nitrogen Oxide Emissions (NOX) Regression Analysis
The coefficient for GDP per capita (GDPC) is positive but statistically
insignificant (β = 0.041, p = 0.460), indicating that at lower income levels, economic
growth does not have a clear effect on NOₓ emissions (see Table 7). However, the
squared term (GDPC²) is negative and marginally significant (β = -1.19E-06, p =
0.094), suggesting that emissions may decline at higher income levels. The cubic term
(GDPC³) is positive and significant (β = 5.01E-12, p = 0.042), reinforcing the presence
of an N-shaped Environmental Kuznets Curve (EKC) for NOₓ emissions. This implies
that while emissions initially increase with economic growth, they eventually decrease
before rising again at higher levels of development, similar to the pattern observed for
CO₂. However, the weaker significance levels of the GDP-related variables suggest that
the EKC effect for NOₓ may be less pronounced than for CO₂.
[Insert Table 7, here]
NOₓ emissions are strongly and consistently impacted by education
characteristics. Higher education levels are linked to higher NOₓ emissions, as seen by
the positive and very significant enrollments in both primary (PRMPUL) and secondary
(SECPUP) schools (β = 0.0003, p = 0.000; β = 0.0004, p = 0.000). This implies that
economic and industrial activity grow as educational attainment increases, which adds
22
to pollution. With an R-squared of 0.989, the overall model fit is remarkably high.
However, this is largely attributed to the inclusion of country fixed effects, which
account for unobserved heterogeneity across countries. While the explanatory factors
contribute to explaining variations in NOₓ emissions, the fixed effects significantly
enhance the model's ability to capture structural differences across countries.
Nonetheless, the low Durbin-Watson statistic (0.439) suggests that the residuals may
be autocorrelated.
Greenhouse Gas Emissions (GRHGAS) Regression Analysis
The coefficient for GDP per capita (GDPC) is positive and statistically significant
(β = 7.696, p = 0.001), indicating that as economies grow, emissions tend to rise (see
Table 8). However, the squared term (GDPC²) is negative and highly significant (β = -
0.0001, p = 0.0002), suggesting that emissions begin to decline after reaching a certain
income threshold. The positive and significant cubic term (GDPC³) (β = 4.01E-10, p =
0.0001) further supports the presence of an N-shaped EKC, implying that after an initial
decline, emissions may rise again at higher levels of economic development. This
suggests that while economic progress can lead to reductions in emissions through
technological improvements and policy measures, sustained growth may eventually
reverse these gains, potentially due to increased consumption and energy-intensive
activities.
[Insert Table 8, here]
Variables related to education consistently and significantly affect greenhouse
gas emissions. Emissions and primary school enrollment (PRMPUL) are strongly
positively correlated (β = 0.018, p = 0.000), suggesting that as economic activity
intensifies due to increased educational access, emissions rise. Higher education levels
are linked to both industrial expansion and energy consumption, as seen by the positive
and substantial influence of secondary school enrollment (SECPUP) (β = 0.003, p =
0.019). While the low Durbin-Watson statistic (0.279) raises the possibility of
autocorrelation issues, the high R-squared value (0.954) should be interpreted with
caution, as it is largely influenced by the inclusion of country fixed effects. These fixed
effects capture unobserved heterogeneity across countries, contributing to the model’s
explanatory power beyond the included variables.
23
Cointegration Testing
The concept of non-stationary time series analysis was developed as a result of
the finding that a unit root may be present in many macroeconomic time series.
According to Engle and Granger (1987), two or more non-stationary series could be
linearly combined to create a stationary series. The non-stationary time series are
regarded as cointegrated when there is such a stationary linear combination. A long-
term equilibrium relationship between the variables is represented by the stationary
combination, often known as the cointegrating equation. Using the approach of Pedroni
(1999) and Pedroni (2004), we apply cointegration tests in a panel data framework in
this section. These tests are designed to assess the presence of cointegration among the
variables, allowing us to examine whether there is a long-run equilibrium relationship
between the economic indicators in our models.
We apply two residual cointegration tests following Pedroni (1999, 2004) and
Kao (1999), which take into consideration cross-sectional dependence (CD) and
assume weakly exogenous regressors, as stated by Demetriades and James (2011), to
investigate whether a long-run equilibrium relationship exists among the variables in
our three models. It should be noted that unless all explanatory variables are very
exogenous, estimating the cointegrating connections using simple OLS would result in
skewed coefficient estimates. Furthermore, because they assume cross-sectional
independence, alternative OLS estimators that seek to mitigate endogeneity bias—such
as the dynamic OLS or fully modified OLS are inappropriate for our data.
Pedroni (Engle-Granger based) Cointegration Tests
The basis for the Engle-Granger (1987) cointegration test is a review of the
residuals of an I(1) variable spurious regression. The residuals should be I(0) if the
variables are cointegrated. Conversely, the residuals will be I (1) if the variables are not
cointegrated. The Engle-Granger paradigm is extended to tests involving panel data by
Pedroni (1999, 2004) and Kao (1999). Pedroni suggests a number of cointegration tests
that take into account different trend coefficients and intercepts across cross-sections.
CO2 Model: The CO2 model's Pedroni Residual Cointegration Test yields
conflicting results about cointegration. With a probability of 0.054 and a value of -
1.602, the Panel PP-Statistic indicates poor evidence against the null hypothesis of no
24
cointegration. This value is around the 0.05 significance level. With a probability of
0.000 and a value of -9.669, the Panel ADF-Statistic is far more significant and shows
compelling evidence for cointegration. Cointegration is also suggested by the Group
PP-Statistic and Group ADF-Statistic, which have respective values of -3.683
(probability 0.0001) and -2.879 (probability 0.002). These findings suggest that the
CO2 model's series most likely show cointegration or long-term correlations, with the
group statistics offering more convincing support (see Table 9).
[Insert Table 9, here]
NOX Model: In the case of common AR coefficients, the NOX model's results
show a stronger argument against cointegration. The null hypothesis is strongly rejected
by the Panel PP-Statistic of -14.681 (probability 0.000), which indicates that the
residuals are probably stationary and that the series are cointegrated. There is no
substantial evidence for cointegration based on this test, nevertheless, as indicated by
the Panel ADF-Statistic of 4.136 with a probability of 1.000. With probability of 0.000
for both, the Group PP-Statistic and Group ADF-Statistic offer compelling evidence
against the absence of cointegration. It is more difficult to draw firm conclusions about
the existence of cointegration in the NOX model because, whereas the Panel PP statistic
points to cointegration, the ADF statistic offers contradictory data (see Table 10).
[Insert Table 10, here]
GRHGAS Model: The findings broadly support the existence of cointegration in
the GRHGAS model. The null hypothesis of no cointegration is strongly rejected by
the Panel PP-Statistic of -4.841 probability 0.000) and the Panel ADF-Statistic of -3.893
(probability 0.000), indicating that the series are cointegrated. The existence of a long-
term relationship between the variables is further supported by the Group ADF-Statistic
of -2.740 (probability 0.0031) and Group PP-Statistic of -2.796 (probability 0.002).
Overall, the evidence points to cointegration in the GRHGAS model, however the
weighted statistics reveal more conflicting findings, with the Panel PP-Statistic being
positive (see Table 11).
[Insert Table 11, here]
25
Environmental Kuznets Curves
Using the fixed-effects regression model with cross – section covariance error,
the Figure 1 shows an N-shaped association between GDP per capita and CO2
emissions. To account for the non-linear dynamics, the model includes GDP per capita,
CO2 (the dependent variable), its squared term (gdpc2), and its cubed term (gdpc3).
Plotting the projected values (co2_hat) against GDP per capita showed that CO2
emissions first climb as economies expand, then fall after a certain income threshold,
and finally rise at higher income levels. This N-shaped curve indicates that although
environmental laws and technology improvements may initially lower emissions,
higher economic growth at later stages may raise CO2 emissions, maybe as a result of
rising energy demand and consumption.
[Insert Figure 1, here]
In Figure 2, nitrogen oxide (NOx) emissions as a percentage of GDP per capita
are shown on an N-shaped Environmental Kuznets Curve. Plotting the predicted values
(nox_hat) against GDP per capita was done using the same regression, NOx (dependent
variable), with gdpc, gdpc2, and gdpc3. The curve indicates that NOx emissions
increase as economic development progresses, primarily due to urbanization and
industrialization. Emissions peak when income levels rise and subsequently fall as a
result of better technologies and more stringent environmental laws. However, NOx
emissions start to increase once more at very high-income levels, possibly as a result of
increased industrial and transportation activities in developed economies.
[Insert Figure 2, here]
In Figure 3, which also follows an N-shaped curve, looks at the connection
between GDP per capita and total greenhouse gas (grhgas) emissions. The projected
values (grhgas_hat) were plotted versus GDP per capita using the same methods, using
grhgas (dependent variable) and gdpc, gdpc2, and gdpc3. The graph shows that when
economies embrace cleaner technology and regulations, greenhouse gas emissions first
rise with economic expansion, peak at a particular income level, and then start to fall.
Emissions do, however, increase with affluence, most likely as a result of rising energy
use, agricultural production, and industrial operations in wealthier countries.
[Insert Figure 3, here]
26
5. Conclusions
The study's conclusions show that education and environmental outcomes have a
complicated and ever-changing relationship. This should not be construed as a criticism
of educational development, even if our study shows a statistically significant positive
association between emissions and primary and secondary school enrollment across all
three models. Instead, it reflects the fact that, in the early phases of economic growth,
more access to education stimulates economic expansion and industrial activity, both
of which can raise emissions. However, via heightened awareness, technological
innovation, and civic involvement, education also has enormous revolutionary potential
for long-term environmental sustainability. The unit root and cointegration tests
confirm the long-run relationship between emissions, growth, and education,
underscoring the necessity of long-term policy planning that aligns educational
development with environmental goals.
Integrating environmental education into primary and secondary school curricula
must be a top priority for policymakers in order to reduce the immediate environmental
costs linked to educational expansion. In order to truly integrate ideas like climate
change, biodiversity, sustainability, and environmental justice into fundamental topics,
this integration should go beyond cursory education. To develop a generation of
environmentally conscious citizens, it is important to encourage experiential learning,
critical thinking, and active involvement in environmental projects. In order to
guarantee that sustainability is a cornerstone of the educational system, governments
must simultaneously implement comprehensive national programs that link education
with development and climate policy. This covers curriculum change, teacher
preparation, and standardized tests that take environmental competencies into account.
Furthermore, funding for training and vocational education programs that are
adapted to the demands of a green economy is crucial. This covers classes on
environmental management, energy-efficient building, sustainable agriculture, and
renewable energy. Education may immediately aid in the shift to low-carbon industries
and lower emissions linked to traditional economic growth paths by giving young
people green skills. In order to make schools into role models for environmental
responsibility and climate resilience, governments should require green building
standards for schools that incorporate sustainable materials, solar energy, and efficient
waste management systems.
27
Additionally, stronger cross-sectoral cooperation is essential. To guarantee policy
coherence, especially when extending educational systems in emerging nations, the
ministries of labor, education, the environment, and the economy must cooperate. By
combining environmental protections with educational expansion, coordinated
measures can avoid the unexpected result of increased emissions. In order to dissociate
economic growth from environmental deterioration, governments need also enact more
comprehensive economic policies like carbon pricing, emissions caps, and incentives
for the adoption of clean energy. These policies can all be used in conjunction with
education. International collaboration and knowledge exchange are necessary to
support these initiatives, especially when it comes to helping developing nations adopt
sustainable education practices without sacrificing their development objectives.
Finally, future studies should investigate the causal pathways by which education
affects environmental outcomes, particularly when considering institutional
transformation, behavioral change, and technology innovation. Education must be
acknowledged by policymakers as a long-term lever for sustainability as well as a short-
term source of emissions during early development. The conflicting effects of
education on emissions may be balanced, and a fair, sustainable transition for all
economies can be ensured, with the backing of a deliberate, forward-looking
educational strategy that encourages green innovation, clean technology adoption, and
global environmental citizenship.
References
Adger, W. N., & Kelly, M. (1999). Social vulnerability to climate change and the
architecture of entitlements. Mitigation and Adaptation Strategies for Global
Change, 4(3), 253-266. https://doi.org/10.1023/A:1009601904210
Apergis N. 2016. Environmental Kuznets curves: new evidence on both panel and
country-level CO2 emissions. Energy Economics 54: 263–271.
Apergis, N., Christou, C., & Gupta, R. (2017). Are there environmental Kuznets curves
for US state-level CO2 emissions?. Renewable and Sustainable Energy Reviews,
69, 551-558. https://doi.org/10.1016/j.rser.2016.11.219
Barro, R. J. (2001). Human capital and growth. American Economic Review, 91(2), 12-
17. https://doi.org/10.1257/aer.91.2.12
28
Beck, N., & Katz, J. N. (1995). What to do (and not to do) with time-series cross-section
data. American political science review, 89(3), 634-647.
Baltagi, B. H., Feng, Q., & Kao, C. (2012). A Lagrange Multiplier test for cross-
sectional dependence in a fixed effects panel data model. Journal of Econometrics,
170(1), 164-177.
Choi, I. (2001). “Unit Root Tests for Panel Data,” Journal of International Money and
Finance, 20: 249–272.
Cole MA. 2004. Trade, the pollution haven hypothesis and the environmental Kuznets
curve: examining the linkages. Ecological Economics 48: 71–81.
https://doi.org/10.1016/j.ecolecon.2003.09.007
Cole, M. A., & Neumayer, E. (2004). Examining the impact of demographic factors on
air pollution. Population and Environment, 26(1), 5-21.
https://link.springer.com/article/10.1023/B:POEN.0000039950.85422.eb
Cole, M. A., Rayner, A. J., & Bates, J. M. (1997). The environmental Kuznets curve:
An empirical analysis. Environment and Development Economics, 2(4), 401–416.
https://doi.org/10.1017/S1355770X97000211
Engle, R. F., & Granger, C. W. (1987). Co-integration and error correction:
representation, estimation, and testing. Econometrica: journal of the Econometric
Society, 251-276.
Fisher, R. A. (1932). Statistical Methods for Research Workers, 4th Edition, Edinburgh:
Oliver & Boyd.
Fuss, S., Canadell, J. G., Peters, G. P., Tavoni, M., Andrew, R. M., Ciais, P., ... &
Yamagata, Y. (2014). Betting on negative emissions. Nature Climate Change, 4(10),
850-853. https://doi.org/10.1038/nclimate2392
Grossman, G. M., & Krueger, A. B. (1995). Economic growth and the environment.
The Quarterly Journal of Economics, 110(2), 353–377.
https://doi.org/10.2307/2118443
Gylfason, T., & Zoega, G. (2003). Education, social equality and economic growth: a
view of the landscape. CESifo Economic Studies, 49(4), 557-579.
https://doi.org/10.1093/cesifo/49.4.557
Hadri, Kaddour (2000). “Testing for Stationarity in Heterogeneous Panel Data,”
Econometric Journal, 3, 148–161.
Halkos G. 2003. Environmental Kuznets curve for sulphur: evidence using GMM
estimation and random coefficient panel data models. Environment and
Development Economics 8(4): 581–601.
29
Hines, J. M., Hungerford, H. R., & Tomera, A. N. (1987). Analysis and synthesis of
research on responsible environmental behavior: A meta-analysis. The Journal of
environmental education, 18(2), 1-8.
https://doi.org/10.1080/00958964.1987.9943482
Hopkins, C., & McKeown, R. (2002). Education for sustainable development: an
international perspective. Education and sustainability: Responding to the global
challenge, 13, 13-24.
Im, K. S., M. H. Pesaran, and Y. Shin (2003). “Testing for Unit Roots in Heterogeneous
Panels,” Journal of Econometrics, 115, 53–74.
Jackson, R. B., Friedlingstein, P., Andrew, R. M., Canadell, J. G., Le Quéré, C., &
Peters, G. P. (2019). Persistent fossil fuel growth threatens the Paris Agreement and
planetary health. Environmental Research Letters, 14(12), 121001.
https://doi.org/10.1088/1748-9326/ab57b3
Jayanthakumaran K, Liu Y. (2012). Openness and the environmental Kuznets curve:
evidence from China. Economic Modelling 29: 566–576.
https://doi.org/10.1016/j.econmod.2011.12.011
Jickling, B., & Wals, A. E. J. (2008). Globalization and Environmental Education:
Looking Beyond Sustainable Development. Journal of Curriculum Studies, 40(1),
1-21. https://doi.org/10.1080/00220270701684667
Jones, G. S., Stott, P. A., & Mitchell, J. F. B. (2016). Uncertainties in the attribution of
greenhouse gas warming and implications for climate prediction. arXiv preprint.
https://arxiv.org/abs/1606.05108
Kao, C. (1999). “Spurious Regression and Residual-Based Tests for Cointegration in
Panel Data,” Journal of Econometrics, 90, 1–44.
Khan, M. H., & Banu, R. (2017). Impact of education on the environment: A case study
on the relationship between education and environmental sustainability.
Environmental Science & Policy, 71, 1-9.
https://doi.org/10.1016/j.envsci.2017.01.001
Leicht, A., Heiss, J., & Byun, W. J. (Eds.). (2018). Issues and trends in Education for
Sustainable Development. UNESCO Publishing.
Maddala, G. S. and S. Wu (1999). “A Comparative Study of Unit Root Tests with Panel
Data and A New Simple Test,” Oxford Bulletin of Economics and Statistics, 61,
631–52.
Millimet D, List J, Stengos T. 2003. The environmental Kuznets curve: real progress or
misspecified models? The Review of Economics and Statistics 85(4): 1038–1047.
O'Neill, S., & Nicholson-Cole, S. (2009). “Fear won’t do it”: Promoting positive
engagement with climate change through visual and iconic representations. Science
Communication, 30(3), 355-379. https://doi.org/10.1177/1075547008329201
30
Orr, D. W. (1992). Ecological Literacy: Education and the Transition to a Postmodern
World. SUNY Press.
Panayotou, T. (1993). Empirical tests and policy analysis of environmental degradation
at different stages of economic development. International Labour Organization
Working Paper Series.
Pedroni, P. (1999). “Critical Values for Cointegration Tests in Heterogeneous Panels
with Multiple Regressors,” Oxford Bulletin of Economics and Statistics, 61, 653–
70.
Pedroni, P. (2004). “Panel Cointegration; Asymptotic and Finite Sample Properties of
Pooled Time Series Tests with an Application to the PPP Hypothesis,” Econometric
Theory, 20, 597–625.
Pesaran, M. H., & Smith, R. (1995). Estimating long-run relationships from dynamic
heterogeneous panels. Journal of econometrics, 68(1), 79-113.
Pesaran, M. H. (1997). The role of economic theory in modelling the long run. The
economic journal, 107(440), 178-191.
Pesaran MH. 2004. General diagnostic tests for cross section dependence in panels.
Cambridge Working Papers in Economics 0435, Faculty of Economics, University
of Cambridge.
Poortinga, W., Steg, L., & Vlek, C. (2004). Values, environmental concern, and
environmental behavior: A study into household energy use. Environment and
Behavior, 36(1), 70–93.
https://journals.sagepub.com/doi/pdf/10.1177/0013916503251466?casa_token=F8
YqRioK9CgAAAAA:u1lEToEE36WzbTVss5DkzGSX1SPZ_b3tU2ep4zc6cc8RP
qTJfvLUjOB1vAB0weInZmInyjh-TDgT
Schultz, P. W., & Zelezny, L. (2003). Reframing environmental messages to be
congruent with American values. Human Ecology Review, 10(2), 125–136.
Solomon, S., Plattner, G. K., Knutti, R., & Friedlingstein, P. (2009). Irreversible climate
change due to carbon dioxide emissions. Proceedings of the National Academy of
Sciences, 106(6), 1704-1709. https://doi.org/10.1073/pnas.0812721106
Sterling, S. (2004). Sustainable Education: Re-visioning Learning and Change. Green
Books.
Stern, D. I. (2004). The rise and fall of the environmental Kuznets curve. World
Development, 32(8), 1419–1439. https://doi.org/10.1016/j.worlddev.2004.03.004
Stern, P. C., & Dietz, T. (1994). The value basis of environmental concern. Journal of
Social Issues, 50(3), 65–84. https://doi.org/10.1111/j.1540-4560.1994.tb02420.x
31
Stevenson, R. B. (2007). Schooling and environmental education: Contradictions in
purpose and practice. Environmental Education Research, 13(2), 139-153.
https://doi.org/10.1080/13504620701295726
Stocker, T. F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S. K., Boschung, J., &
Midgley, P. M. (Eds.). (2013). Climate Change 2013: The Physical Science Basis.
IPCC Fifth Assessment Report. Cambridge University Press.
Tilbury, D. (1995). Environmental education for sustainability: Defining the new focus
of environmental education in the 1990s. Environmental Education Research, 1(2),
195-212. https://doi.org/10.1080/1350462950010206
Tilbury, D. (2011). Education for Sustainable Development: An Expert Review of
Processes and Learning. UNESCO.
UNESCO. (2005). UN Decade of Education for Sustainable Development (2005-2014)
Draft International Implementation Scheme. UNESCO.
UNESCO. (2012). Education for Sustainable Development Sourcebook.
UNESCO. (2014). Shaping the Future We Want: UN Decade of Education for
Sustainable Development (2005-2014) Final Report.
UNESCO. (2017). Education for Sustainable Development Goals: Learning
Objectives.
Wals, A. E. J. (2011). Learning our way to sustainability. Journal of Education for
Sustainable Development, 5(2), 127-138.
https://doi.org/10.1177/097340821100500208
Wals, A. E. J. (Ed.). (2007). Social Learning towards a Sustainable World: Principles,
Perspectives, and Praxis. Wageningen Academic Publishers.
https://doi.org/10.3920/978-90-8686-594-9
Westerlund, J. (2007). Testing for error correction in panel data. Oxford Bulletin of
Economics and statistics, 69(6), 709-748.
Wooldridge, Jeffrey M. (2002). Econometric Analysis of Cross Section and Panel Data,
Cambridge, MA: The MIT Press.
Zaim O, Taskin F. 2000. A Kuznets curve in environmental efficiency: an application
on OECD countries. Environmental and Resource Economics 17(1): 21–36.
https://link.springer.com/content/pdf/10.1023/A:1008318605976.pdf
Zsóka, Á., Szerényi, Z. M., & Kocsis, T. (2013). Greening education and changing
behavior: The impact of environmental education on students' pro-environmental
behavior and environmental knowledge. Sustainability, 5(11), 4189–4204.
https://doi.org/10.1016/j.jclepro.2012.11.030
32
TABLES AND FIGURES SECTION
Table 1. Variables’ Definitions
Variable Name
Variable Description
CO2
CO₂ emissions (total metric tons of CO₂ equivalent), sourced from the
World Bank.
NOX
NOₓ emissions (total metric tons of CO₂ equivalent), sourced from the
World Bank.
GRHGAS
Greenhouse gas (GHG) emissions (total metric tons of CO₂
equivalent), sourced from the World Bank.
GDPC
GDP per capita (current USD), representing economic development
levels in each country.
GDPC2
Square of GDP per capita, capturing nonlinear effects of economic
growth on emissions.
GDPC3
Cube of GDP per capita, capturing higher-order nonlinear
relationships between economic growth and emissions.
PRMPUL
Primary education enrollment (total number of pupils), reflecting
participation in basic education.
SECPUP
Secondary education enrollment (total number of pupils), reflecting
participation in secondary education.
Source: World Bank
33
Table 2. Descriptive Statistics
Variable
N
Mean
Standard deviation
Min.
Max.
Dependent Variables
CO2
3,507
204946.3
777439.9
6.6
1.10E+07
GRHGAS
3,507
275579.1
946736.7
19.690
1.30E+07
NOX
3,507
16291.22
47492.37
0.362
551683
Explanatory
variables
GDPC
3,507
19065.9
25224.01
137.182
204097
PRMPUL
3,507
3377524
1.19E+07
1235
1.40E+08
SECPUP
3,507
3529175
1.10E+07
508
1.30E+08
Source: Authors’ calculations
34
Table 3. Correlation Matrix
CO2
GRHGAS
NOX
GDPC
PRMPUL
SECPUP
CO2
1
GRHGAS
0.996
1
0.000
NOX
0.914
0.938
1
0.000
0.000
GDPC
0.047
0.038
0.006
1
0.004
0.022
0.688
PRMPUL
0.586
0.620
0.732
-0.097
1
0.000
0.000
0.000
0.000
SECPUP
0.619
0.657
0.793
-0.096
0.919
1
0.000
0.000
0.000
0.000
0.000
The values bellow the coefficients indicate the significance level
Source: Authors’ calculations
Table 4. Residual Cross-Section Dependence Test
Variable
Breusch-
Pagan
LM
Prob.
Pesaran
Scaled
LM
Prob.
Bias-
Corrected
Scaled
LM
Prob.
Pesaran
CD
Prob.
CO2
98,787.53
0.000
510.071
0.000
505.896
0.000
125.524
0.000
NOX
102,635.70
0.000
533.184
0.000
529.009
0.000
117.758
0.000
GRHGAS
91,221.82
0.000
464.631
0.000
460.456
0.000
86.5261
0.000
Null hypothesis: No cross-section dependence (correlation) in residuals
Source: Authors' Calculations
35
Table 5. Panel Unit Root Tests
Im, Pesaran and
Shin W-stat
ADF - Fisher Chi-
square
PP - Fisher Chi-
square
Variable
Statistic
Prob.**
Statistic
Prob.**
Statistic
Prob.**
Cross-
sections
Obs
CO2
3.080
0.999
339.637
0.374
365.195
0.101
167
3240
D(CO2)
-28.06
0.000
1478.02
0.000
1924.17
0.000
167
3101
NOX
1.462
0.928
364.706
0.119
366.132
0.109
167
3256
D(NOX)
-40.222
0.000
2020.40
0.000
4114.51
0.000
167
3069
GRHGAS
2.878
0.998
376.882
0.052
339.874
0.400
167
3221
D(GRHGAS)
-30.517
0.000
1548.82
0.000
1775.99
0.000
167
3123
GDPC
0.840
0.790
317.295
0.736
385.328
0.027
167
3284
D(GDPC)
-23.927
0.000
1194.17
0.000
1179.29
0.000
167
3138
PRMPUL
-34.37
0.997
1157.19
0.054
758.342
0.054
162
3105
D(PRMPUL)
-45.432
0.000
2497.48
0.000
7546.81
0.000
161
2924
SECPUP
-8.824
0.996
697.572
0.053
1063.92
0.053
167
3227
D(SECPUP)
-45.914
0.000
2359.23
0.000
8545.62
0.000
166
3025
Source: Authors’ estimations
** Probabilities for Fisher tests are computed using an asymptotic Chi-square
distribution. All other tests assume asymptotic normality.
36
Table 6. Panel fixed effects regression results with PCSE
(cross – section covariance error) (Dependent variable: CO2)
Effects specification: Cross – section fixed (dummy variables), Period fixed (dummy
variables)
Dependent Variable: CO2
Variable
Coefficient
Std. Error
t-Statistic
Prob.
GDPC
7.048443
2.156
3.268
0.001
GDPC^2
-0.000103
2.72E-05
-3.803
0.000
GDPC^3
3.69E-10
9.48E-11
3.892
0.000
PRMPUL
0.017284
0.00079
21.735
0.000
SECPUP
0.0026
0.00135
1.914
0.055
C
73583.17
24211.94
3.039
0.002
Root MSE
185198.9
R-squared
0.943
Mean dependent var
204946.3
Adjusted R-squared
0.939
S.D. dependent var
777439.9
S.E. of regression
190486.6
Akaike info criterion
27.205
Sum squared resid
1.2E+14
Schwarz criterion
27.543
Log likelihood
-47513.3
Hannan-Quinn criter.
27.326
F-statistic
288.405
Durbin-Watson stat
0.274
Prob(F-statistic)
0.0000
Source: Authors’ estimations
37
Table 7. Panel fixed effects regression results with PCSE
(cross – section covariance error) (Dependent variable: NOX)
Effects specification: Cross – section fixed (dummy variables), Period fixed (dummy
variables)
Dependent Variable: NOX
Variable
Coefficient
Std. Error
t-Statistic
Prob.
GDPC
0.041885
5.68E-02
0.737
0.460
GDPC^2
-1.19E-06
7.11E-07
-1.671
0.094
GDPC^3
5.01E-12
2.46E-12
2.031
0.042
PRMPUL
0.000363
2.10E-05
17.249
0.000
SECPUP
0.000461
3.54E-05
13.029
0.000
C
13384.6
637.5883
20.992
0.000
Root MSE
4981.549
R-squared
0.988
Mean dependent var
16291.22
Adjusted R-squared
0.988
S.D. dependent var
47492.37
S.E. of regression
5123.781
Akaike info criterion
19.974
Sum squared resid
8.70E+10
Schwarz criterion
20.311
Log likelihood
-34833.05
Hannan-Quinn criter.
20.094
F-statistic
1559.693
Durbin-Watson stat
0.439
Prob(F-statistic)
0.000
Source: Authors’ estimations
38
Table 8. Panel fixed effects regression results with PCSE
(cross – section covariance error) (Dependent variable: GRHGAS)
Effects specification: Cross – section fixed (dummy variables), Period fixed (dummy
variables)
Dependent Variable: GRHGAS
Variable
Coefficient
Std. Error
t-Statistic
Prob.
GDPC
7.696222
2.35848
3.263
0.001
GDPC^2
-0.000113
2.97E-5
-3.786
0.000
GDPC^3
4.01E-10
1.04E-10
3.867
0.000
PRMPUL
0.018837
0.000872
21.600
0.000
SECPUP
0.003489
0.001489
2.343
0.019
C
129780.3
26487.28
4.899
0.000
Root MSE
203457.6
R-squared
0.953
Mean dependent var
275579.1
Adjusted R-squared
0.951
S.D. dependent var
946736.7
S.E. of regression
209266.7
Akaike info criterion
27.393
Sum squared resid
1.45E+14
Schwarz criterion
27.731
Log likelihood
-47843.03
Hannan-Quinn criter.
27.514
F-statistic
358.340
Durbin-Watson stat
0.278
Prob(F-statistic)
0.000
Source: Authors’ estimations
39
Table 9. Pedroni Residual Cointegration Test Results (CO2 Model)
Alternative Hypothesis: Common AR Coefficients (within-dimension)
Statistic
Value
Probability
Weighted Statistic
Value
Panel PP-Statistic
-1.602
0.054
0.968
0.833
Panel ADF-Statistic
-9.669
0.000
2.431
0.992
Alternative Hypothesis: Individual AR Coefficients (between-dimension)
Statistic
Value
Probability
Group PP-Statistic
-2.879
0.002
Group ADF-Statistic
-3.683
0.0001
Source: Authors’ estimations
Table 10. Pedroni Residual Cointegration Test Results (NOX Model)
Alternative Hypothesis: Common AR Coefficients (within-dimension)
Statistic
Value
Probability
Weighted Statistic
Value
Panel PP-Statistic
-14.68128
0.000
-2.358
0.009
Panel ADF-Statistic
4.136051
1.000
-0.571
0.283
Alternative Hypothesis: Individual AR Coefficients (between-dimension)
Statistic
Value
Probability
Group PP-Statistic
-8.227
0.0000
Group ADF-Statistic
-7.063
0.0000
Source: Authors’ estimations
40
Table 11. Pedroni Residual Cointegration Test Results (GRHGAS Model)
Alternative Hypothesis: Common AR Coefficients (within-dimension)
Statistic
Value
Probability
Weighted Statistic
Value
Panel PP-Statistic
-4.841
0.000
1.745
0.959
Panel ADF-Statistic
-3.893
0.000
3.564
0.999
Alternative Hypothesis: Individual AR Coefficients (between-dimension)
Statistic
Value
Probability
Group PP-Statistic
-2.796
0.002
Group ADF-Statistic
-2.740
0.003
Source: Authors’ estimations
Figure 1. Environmental Kuznets Curve (CO2-GDP per capita)
Source: Authors’ estimations
100000 200000 300000 400000 500000
CO2 emissions (kt)
0 50000 100000 150000 200000
GDP per capita
Linear prediction Linear prediction
Environmental Kuznets Curve
41
Figure 2. Environmental Kuznets Curve (NOX-GDP per capita)
Source: Authors’ estimations
Figure 3. Environmental Kuznets Curve (Green House Gases-GDP per capita)
Source: Authors’ estimations
10000 15000 20000 25000
NOX emissions (kt)
0 50000 100000 150000 200000
GDP per capita
Linear prediction Linear prediction
Environmental Kuznets Curve
100000 200000 300000 400000 500000 600000
Green House Gases emissions (kt)
0 50000 100000 150000 200000
GDP per capita
Linear prediction Linear prediction
Environmental Kuznets Curve
42
BANK OF GREECE WORKING PAPERS
322. Degiannakis, S. and E. Kafousaki “Forecasting VIX: The illusion of forecast
evaluation criteria”, June 2023.
323. Andreou C. P., S. Anyfantaki, C. Cabolis and K. Dellis, “Exploring country
characteristics that encourage emissions reduction”, July 2023.
324. Dimakopoulou, V., Economides, G., Philippopoulos, A., and V. Vassilatos, “Can
central banks do the unpleasant job that governments should do?”, December 2023.
325. Chrysanthakopoulos, C. and A. Tagkalakis, “The medium-term effects of fiscal
policy rules”, January 2024.
326. Manou, K. and E. Papapetrou, “Does uncertainty matter for household
consumption? A mean and a two tails approach”, February 2024.
327. Kakridis, A., “War, mobilization, and fiscal capacity: testing the bellicist theory in
Greece, 1833-1939”, March 2024.
328. Mavrogiannis, C. and A. Tagkalakis, “From policy to capital: assessing the impact
of structural reforms on gross capital inflows”, April 2024
329. Delis, P., S. Degiannakis, G. Filis, T. Palaskas and C. Stoforos, “Determinants of
regional business cycle synchronization in Greece”, May 2024.
330. Sideris, D. and G. Pavlou, “Market power and profit margins in the Euro area
countries in the post-pandemic period”, June 2024.
331. Kasimati, E. and N. Veraros, “The dry-bulk shipping market: a small econometric
model”, September 2024.
332. Mermelas, G. and A. Tagkalakis, “Monetary policy transmission: the role of
banking sector characteristics in the euro area”, November 2024.
333. Anastasiou, D., Pasiouras, F., Rizos, A., and A. Stratopoulou, “Do macroprudential
policies make SMEs more-or-less discouraged to apply for a bank loan?”,
December 2024.
334. Malliaropulos, D., Passari, E., and F. Petroulakis, “Unpacking commodity price
fluctuations: reading the news to understand inflation”, December 2024
335. Degiannakis, S. and E. Kafousaki, “Disaggregating VIX”, January 2025
336. Degiannakis, S., Delis, P., Filis, G., and G. Giannopoulos, “Trading VIX on
volatility forecasts: another volatility puzzle?”, February 2025
337. Papadopoulos, G., Ojea-Ferreiro, J., and R. Panzica, “Climate stress test of the
global supply chain network: the case of river floods”, February 2025
338. Papaoikonomou, D., “Stochastic debt sustainability analysis: a methodological
note”, March 2025
339. Dellas, H. and G. Tavlas, “The great dollar shortage debate: a modern perspective”,
March 2025
340. Hall, S. and G. Tavlas, “Quantifying Federal Reserve credibility”, April 2025
