Factors affecting per capita ecological footprint in OECD countries: Evidence from machine learning techniques a

Introduction

For several decades, scientists have discussed global warming, climate change, and their consequences on society and the environment. Finally, these issues have been recognized by world public opinion, and it has been taken into action several international initiatives. For instance, in 1992, the United Nations organized a conference called the Rio Summit, and they accepted that there is an issue related to environmental deterioration in the world. After, the Kyoto Protocol was adopted in 1997 in order to mitigate greenhouse gas emissions, especially in developed countries. Then, the Paris Agreement was signed in 2015 to control an increase in the global temperature for the purpose of decreasing global warming. ¹ Also, in 2015, the Sustainable Development Goals (SDGs) were adopted by United Nations to provide sustainable development including cleaner environment and more efficient use of energy. The 2021 United Nations Climate Change Conference (COP26) held in the United Kingdom to discuss how to combat with climate change. Recently, European Union countries introduced the European Green Deal which aims to make the union climate neutral in 2050.

As discussed above, for a few decades, environmental deterioration has been at the center of much interest Also, it is known that developed countries (including OECD countries) are primarily responsible for the high level of emission levels in the atmosphere. ² Therefore, this study aims to investigate the factors affecting environmental quality in 27 OECD countries for the period 1971–2016. Factors affecting environmental degradation can be determined and effective policies can be designed with this investigation thanks to novel machine learning methods. The study utilizes annual longitudinal data and uses per capita ecological footprint as an ecological indicator, while per capita income, types of energy consumption (oil products, coal, natural gas, electricity, and renewable energy), dimensions of globalization (trade, financial, social, and political), and urbanization rate are determined as input variables. The aforementioned series are the most used variables as determinants of environmental pollution in the empirical literature. In addition, the research data are drawn from four primary sources: Global Footprint Network, World Bank, International Energy Agency (IEA), and KOF Swiss Economic Institute.

We choose OECD countries as the current study's sample since we believe that the reliability of these data can be higher than in developing countries. In some developing countries, governments have control over the statistical institutions and force them to declare faulty statistics. Furthermore, it is accepted that this set of countries are primarily responsible for the high level of emission levels in the world currently. ² For this reason, these developed countries try to increase environmental quality through international initiatives such as the Kyoto Protocol, the Paris Agreement, and the European Green Deal. Therefore, their environmental actions and policies should be considered. Moreover, we try to work on a group of countries that share similar economic characteristics rather than geographical connections because the economic properties of countries have a more prominent role in environmental degradation. As known, most of the OECD countries have high-income levels except a few. For the above-mentioned reasons, we focus on OECD countries in this study.

The present study has shed a contemporary light on the contentious issue of the impact of economic variables on environmental deterioration through various machine learning techniques (tree regression, boosting, bagging, and random forest). These techniques have many advantages compared to conventional econometric methods. For instance, machine learning algorithms provide lots of instruments to capture nonlinear relationships in the data, ³ and they allow functional form flexibility. ⁴ Besides, they perform more efficiently handling a large dataset in comparison to other empirical methods. ⁵ Lastly, there is no need for a priori assumption on theoretical links and distribution of the variables.^6,7

To our knowledge, the present research explores, for the first time, the effects of income level, energy consumption, globalization, and urbanization on the ecological footprint for OECD countries through a bunch of machine learning techniques. There are only a few initiatives that examine the determinants of ecological footprint using similar methods in the empirical literature; Sun, ⁸ Janković et al., ⁹ Roumiani and Mofidi, ¹⁰ and Yu. ¹¹ In these studies, Sun ⁸ finds that the random forest methodology is the best algorithm for the model examined, among others, and energy structure, energy efficiency, energy intensity, and economic development affect energy carbon footprint in Liaoning province for the period 1998–2016. Besides, Janković et al.'s ⁹ results for random forest algorithm show that coal, oil, and natural gas consumption are the most significant factors affecting ecological footprint for 41 countries covering the period 1971–2014. Roumiani and Mofidi ¹⁰ conducted a few machine learning techniques to determine which variables can affect the ecological footprint significantly of G20 countries during 1999–2018. According to the LASSO (least absolute shrinkage and selection operator) method results, carbon dioxide emissions, gross fixed capital formation, and gross domestic product were determined as the most significant variables for ecological footprint. Recently, Yu ¹¹ found that GDP, energy consumption, and total retail of consumer goods are the most significant factors for projections of ecological footprint in China.

The remaining part of the paper proceeds as follows: Section 2 presents the empirical literature, while Section 3 introduces the dataset and discusses the specific methods by which the study and analyses were conducted. Section 4 highlights the empirical results and discusses the results, while the final section concludes the current research.

Literature review

This section summarizes the literature on the determinants of environmental deterioration, primarily focusing on the ecological footprint as a degradation index.

Data and methodology

In this study, the research material uses quantitative data related to factors affecting per capita ecological footprint for 27 OECD countries, 1971–2016. For this purpose, the study collected publicly available data related to one output variable and eleven input variables; in other words, our sample consists of 14,904 observations. To process the research material collected, a bunch of tree-based machine learning techniques is utilized: tree regression, bagging, boosting, and random forest

Empirical results and discussion

This study employs various machine learning techniques to predict the target variable using a set of input variables. For this purpose, the study follows the steps suggested by Basuchoudhary et al. ⁶ :

Selecting the algorithm that best fits the training data using error measurements,

OPEN IN VIEWER

Table 1.

The predictive quality of empirical methods.

Methods	MSE	MAPE	MAE	RMSE
Panel OLS Regression	0.0283	0.0758	0.1278	0.1682
Tree Regression	0.0466	0.0985	0.1654	0.2159
Bagging	0.0252	0.0723	0.1204	0.1587
Boosting	0.0280	0.0798	0.1348	0.1673
Random Forest	0.0248*	0.0686*	0.1112*	0.1575*

Note: * shows the method that provides minimum errors.

Table 2 presents the variable importance results for four main machine learning techniques. Since the random forest model fits our dataset, we take into consideration its variable importance result. The variables are sorted according to the highest value to the lowest value. The table exhibits the percentage increase in errors when a variable for each algorithm is omitted from the model. To be more precise, if oil product consumption is removed from the random forest model, MSE increases by 36.23%. So, we can say that including this variable is very important to estimate our machine learning model.

OPEN IN VIEWER

Table 2.

Variable importance according to models.

Variable	Random Forest↓	Tree Regression	Bagging	Boosting
Oil Product Consumption	36.2332	29.2608	89.0932	81.7033
Electricity Consumption	15.4077	16.4889	0.6778	3.4825
GDP	14.4786	18.9663	0.6240	1.6635
Natural Gas Consumption	7.4669	1.9868	0.4924	1.3161
Social Globalization	6.8280	13.8014	4.4296	4.7868
Coal Consumption	4.7403	7.9076	0.7386	1.6729
Renewable Energy Consumption	3.9336	1.2525	0.8620	1.5389
Trade Globalization	3.8820	2.2801	1.4239	1.1517
Financial Globalization	2.4714	5.8367	0.6091	0.6149
Political Globalization	2.3576	0.9115	0.3084	0.8754
Urbanization	2.2008	1.3075	0.7409	1.1940
Total	100.00	100.00	100.00	100.00

Note: ↓ denotes sorting from the largest value to the smallest one.

The Figure 1 shows the partial dependence plots on various factors for ecological footprint. So, these graphs exhibit the incremental impacts of several variables on the environmental deterioration for OECD countries. The figures are listed according to their variable importance levels.

OPEN IN VIEWER

Figure 1.

Partial dependence plots.

In Figure 1.a, the horizontal axis denotes oil product consumption, while the vertical axis shows the ecological footprint. It is inferred from the figure that the relationship between these two variables is not linear. It is seen that the effect of oil product consumption on environmental deterioration is positive throughout the sample. However, its impact is slight up to a certain point, while its effect on the environment becomes harsh with increases in the consumption level. Oil products are fossil fuels, and it is accepted that fossil fuel combustion damages the environment. The pollutive effect of oil product (total energy, nonrenewable energy, or fossil fuel) consumption was also supported by Destek and Sinha, ²⁵ Ozcan et al., ²³ and Ulucak et al. ²⁴ for OECD countries.

There are three main sources of electricity generation in the world: fossil fuels, renewable energy sources, and nuclear energy. Although many countries have constructed renewable- and nuclear energy facilities for a few decades, the use of fossil fuels (coal, natural gas, and oil) in electricity generation is still dominant. ⁶¹ Therefore, in Figure 1.b, it is revealed that an increase in electricity consumption causes environmental degradation in OECD countries. However, after a threshold point, its pollutive effect decreases a bit. This movement can be attributed to an increase in the share of renewable energy sources over total electricity generation. The empirical results of Ergün and Polat ²⁶ and Lau et al. ²⁷ supported the pollutive effect of electricity consumption on the environment for OECD countries, which was similar to our findings.

A nonlinear relationship between ecological footprint and GDP is depicted in Figure 1.c. In the initial stage of economic development, an increase in the income level pollutes the environment too much. However, after a specific level of income, the environmental damage caused by economic growth slows down. This slowdown can be attributed to a few reasons; the environmental consciousness of people rises, and they demand more ecological quality from the governments. Thus, governments design and implement more stringent environmental policies to mitigate damage to the environment. As a result of tight environmental regulations and advancements in technology level, firms start to use cleaner production technologies. Also, the structure of the economy affects the pollution level in a country. Environmental damage decreases when the economy moves from the pre-industrial and industrial stage to the service economy. However, even if the increase in pollution level slows down with the increase in income after a specific point, it is seen that the EKC hypothesis is not valid in OECD countries. Therefore, one can state that the technology effect cannot dominate scale and composition effects too much in OECD countries. Besides, Stern and Cleveland ⁶² asserted that shifting from the industrial economy to the service economy does not eliminate environmental deterioration totally because the service sector has a large demand for energy consumption for construction and maintenance activities for office towers, shopping centers, restaurants, real estate services, etc. In addition, Gill et al. ⁶³ argued that the EKC hypothesis is mainly focused on the production side of the economy; however, the consumption side also has a significant impact on it. The demand for luxury goods and electronics (they might be energy-dependents goods) may increase in the higher level of income level due to the conspicuous consumption behaviors of the citizens. For instance, according to D'Arpizio et al., ⁶⁴ the value of personal luxury goods (e.g. jewelry, watches, cars, accessories) market value in the world reached 281 billion euros in 2019.

Moreover, although people living in developed countries acknowledge that airplanes do more damage to the environment than trains or that private cars pollute nature more than public transportation, ⁶⁵ they still prefer these goods/services for their own comfort and pleasure. All these issues can be accepted as the main reasons why the EKC hypothesis does not hold in the OECD countries. Nevertheless, if the environmental consciousness of every economic agent continues and is assisted by various policy tools, the pollution level may start to decrease. Some studies backed up the pollutive effect of the income level throughout history, such as Ulucak and Koçak ²² and Ozcan et al. ²³ for OECD countries, and Alola et al. ¹⁹ for European Union countries.

A nonlinear relationship (can be regarded as U-shaped) between ecological footprint and natural gas consumption is presented in Figure 1.d. Natural gas can be used as a substitute for other fossil fuels like coal, diesel fuel, gasoline, and oil in some circumstances. Since natural gas emits less carbon emission compared to coal, ⁶⁶ it can also be used in heating. Thus, the initial impact of natural gas use is less pollutive. According to Dong et al., ⁶⁷ these findings could be explained by natural gas is a cleaner substitute for other fossil fuels. However, its negative impact on the environmental quality rises harshly in the huge amount of consumption level because the drilling, extraction, and transportation of natural gas release a significant amount of methane (that has a detrimental impact on the environment). ⁶⁸ Some of the empirical studies in the literature supported our findings (e.g.^69–71 after the threshold point. Besides, the U-shaped relationship between these two variables was backed up by Xu and Lin ⁷² for China and the Central Region. This kind of shape can be explained by optimizing the energy structure at the initial stage; however, the environmental degradation level increases with consuming more and more natural gas (which is one of the fossil fuels). ⁷²

Figure 1.e shows the incremental impacts of social globalization on environmental pollution. It is clearly seen that there is a positive association between these two variables. Since the social globalization index covers the number of airports, international tourism activities, migration, and movements of international exchange students, an increase in the social globalization index causes ecological problems in the majority of the sample; there is only a small decline after a threshold level of globalization. So, fuel and gasoline consumption rise, and the emission level is enhanced in nature because tourism activities and student movements trigger the transportation volume. Besides, migration may cause overpopulation and an increase in population density within the country, which affects environmental quality negatively. Empirical studies in the literature supported our findings, namely, Bilgili et al. ³⁸ and Langnel and Amegavi. ³⁹

Figure 1.f displays that a small amount of coal consumption leads to a rapid deterioration in environmental quality. Its impact on environmental degradation is enormous, as coal is considered by scientists to be the most carbon-emitting fossil fuel source. This empirical finding was backed up by many studies such as Saboori and Sulaiman, ⁶⁹ Shahbaz et al., ⁷³ Pata, ⁴³ Xu et al., ⁷¹ and Magazzino et al. ⁷⁴ In addition, there are negative consequences of the massive amount of coal consumption on human health, such as respiratory illness, cancer, cardiovascular disease, preterm delivery, fluorosis, arsenism, selenosis, and adverse child development. ⁷⁵ Therefore, it should be hesitated to consume a massive amount of coal use for public health.

Figure 1.g displays the incremental impacts of renewable energy consumption on the ecological footprint for OECD countries. At the initial stage of consumption, providing energy through renewable sources decreases environmental deterioration. However, there is a limited level of installed capacity regarding renewable energy sources. Countries require new facilities to produce a higher level of renewable energy. Therefore, there is a need to construct renewable energy facilities, including turbines (solar thermal, geothermal, biomass) and solar photovoltaics, then, installed capacity increases. All of these constructions cause environmental deterioration in several ways. According to Willmott Dixon, ⁷⁶ buildings led to a rise in air pollution, water pollution, landfill waste, and ozone depletion. Hence, the impact of an increase in total renewable energy consumption on environmental pollution turns positive. However, it is seen that its pollutive effect is lower than the initial level. Empirical studies in the literature—Destek et al., ⁷⁷ Adedoyin et al. ⁷⁸ and Altıntaş and Kassouri ⁷⁹ considering the European countries; Destek and Sinha ²⁵ and Ulucak et al. ²⁴ considering OECD countries—supported our findings even if up to the threshold point.

Figure 1.h illustrates the relationship between per capita ecological footprint and trade globalization. The impact of trade globalization on environmental quality is positive up to a threshold point. However, after this point, its effect turns negative on the environment. The export volume of the countries grows in a more globalized trade environment because an increase in trade agreements and trade partner diversity and a decrease in trade taxes and tariffs affect the trade volume. Therefore, the production level in the countries increases, including energy-intensive manufacturing. Thus, the environment can be damaged by an increase in the production level. In the same vein, Bilgili et al., ³⁸ in their work, stated that trade globalization has a positive impact on environmental quality in Turkey. This result is consistent with our results up to a certain point; however, its impact turns to negative in the highly globalized trade environment, according to our findings. Supporting this empirical result, Yilanci and Gorus ⁸⁰ found a positive correlation between ecological footprint and trade globalization in 14 MENA countries.

The partial dependence plot for financial globalization in Figure 1.i shows that a more globalized financial environment causes a decline in the pollution level in OECD countries for a certain point, then, it starts to increase sharply. As it is known, the financial globalization index includes many items related to investment activities such as FDI inflows, portfolio investment, and investment agreements. So, an increase in financial globalization means a higher level of inflow, including both FDI and portfolio investment. Since OECD countries have tighter environmental regulations compared to developing countries, foreign investors do not invest in dirty industries in the host country; they make investments in clean industries and the service sector. Thus, environmental quality increases in these countries thanks to environmentally-friendly technologies and management capabilities of the foreign companies. ⁸¹ In short, financial globalization creates a pollution halo effect for OECD countries for a specific level of financial globalization. These results are in accord with recent studies (see,^38,41), indicating that financial globalization provides a cleaner environment. However, after a threshold point, financial globalization damages the environment too much because a high level of development in financial markets causes a decrease in interest rates. Since low-interest rates trigger investment activities and increase the consumption level of people, demand for goods and services rises, and the use of energy sources multiplies. ⁸²

The partial dependence plot for political globalization over per capita ecological footprint in Figure 1.j exhibits that an increase in the index heals the environment in OECD countries for the period 1971–2016. According to the figure, at the lower level of political globalization, it has a small or zero impact on environmental quality. However, when governments engage in more international political cooperation and treaties (including environmental summits and meetings), the demand for natural assets decreases gradually; then, environmental quality increases. These empirical findings were consistent with those of Shahbaz et al., ³⁷ Bilgili et al., ³⁸ Destek, ⁸³ Langnel and Amegavi, ³⁹ and Suki et al. ⁴⁰

Figure 1.k shows the relationship between environmental deterioration and the percentage of people who settled in urban areas. The above figure outlines that there is a positive association between these two variables. This kind of relationship was supported by a bunch of studies, including Beck and Joshi, ⁴⁷ Ozturk et al., ¹⁷ Alola et al., ⁴⁸ and Ulucak et al. ⁴¹ In detail, the figure shows that an increase in urbanization rate increases environmental degradation until a specific point gradually. This environmental deterioration can be attributed to the rise in population density, deforestation, transport services, heating, production level, and construction facilities. However, a sharp decline is observed after a threshold point; therefore, an inverted U-shaped relationship can be confirmed up to a particular level of urbanization level; less than 4.40 in terms of the natural logarithm. A sharp decrease in pollution level with an increase in urbanization rate can be attributed to the municipalities’ intervention. This can be explained by the better services carried out by municipalities, such as sewers, infrastructure, and waste management. An inverted U-shaped link between pollution level and urbanization rate was also revealed in the studies carried out by Bekhet and Othman ⁸⁴ and Ahmed et al. ⁸⁵

Nevertheless, the pollution level starts to increase again in the case of high urbanization level. That is, the massive migration from rural to urban areas leads to more environmental problems. Overall, the nonlinear relationship between the aforementioned variables shows an N-shaped pattern for the OECD countries. One of the main reasons behind this issue is the management capabilities of municipalities; they cannot handle and manage overpopulation problems compared to the optimal level of urban population appropriately. In addition, one of the indirect impacts of urbanization is energy efficiency, thanks to technological innovations and developments in the energy sector. A decrease in energy prices following a rise in energy efficiency causes a rebound effect if economic agents demand more energy and energy-related goods and services. ⁸⁶ Also, a rise in urban population triggers an expansion in the service sector, and an increase in the service sector may cause development in subsectors that are closely related to CO₂ emissions such as transport services, hotels, restaurants, real estate, and public administration services. ⁸⁷ Moreover, Ji and Chen ⁸⁸ argued that the lifestyle and consumption patterns of the people may change when they move from rural to urban areas (or during the urbanization process). Therefore, their demand for goods and services rises since their consumption behavior shifts from survival to development and/or enjoyment modes. All of these issues can be considered as the main reasons why a high level of urbanization damages the environment. A similar pattern (N-shaped) was found by Sheng et al. ⁸⁹ for the Yangtze River Delta region of China and Yassin and Aralas ⁹⁰ for the Asian countries.

Conclusion

This study set out to explore the influence of income level, energy consumption, globalization level, and urbanization on environmental deterioration—per capita ecological footprint—through utilizing various machine learning techniques for 27 OECD countries. This investigation used the annual panel dataset for the period between 1971 and 2016. There are plenty of advantages of machine learning techniques over traditional econometric methods; a) functional form flexibility, b) deals with a large volume of datasets, c) no need for a priori assumption of the variables.

This study selected the algorithm that best fits the training data using three error measurements. According to the findings, it was suggested that the random forest model provides the minimum error level. Then, this study presented the variable importance and partial dependence plots based on the random forest algorithm.

The empirical results based on the variable importance demonstrated that oil product consumption, electricity consumption, and GDP are the most significant variables for our model. If they were removed from the random forest model, errors increase significantly. To be more precise, if oil product consumption was removed from the random forest model, MSE increases by 36.23%, while if electricity consumption and GDP were excluded from the model, errors increase by around 15%. On the other side, it was seen that urbanization, political, financial, and trade globalization levels had little importance in determining ecological footprint per capita. Moreover, the partial dependence plots indicated that economic growth and especially fossil fuel energy consumption damage the environment. Contrarily, renewable energy consumption polluted the environment less than oil products, coal, natural gas, and electricity consumption. Besides, it could be revealed that the demand for natural assets decreases gradually when governments engage in more international political cooperation and treaties, while trade and financial globalization levels heal the environment only up to a certain level. After these levels, these two indicators drastically increased the per capita ecological footprint in OECD countries. Lastly, the empirical findings showed that an excessive level of urbanization causes environmental deterioration for the period examined.

Empirical results of this study indicated that the EKC hypothesis is not validated for the OECD countries even though the marginal impact of income level on environmental pollution decreases gradually after a specific level of income. This finding might be attributed to the inability of the technology effect to dominate the scale and composition effects. In addition, the enlargement of the service sector might cause development in subsectors that are closely related to fossil fuel energy use. Besides, the conspicuous consumption behaviors and usual consumption patterns of the people which are highly dependent on their own comfort and pleasure can prevent a high decline in the pollution level in the countries examined. However, if the environmental consciousness of every economic agent is increased and sustainable development is assisted by miscellaneous policy tools, the pollution level may start to decrease in the OECD countries.

Taken together, these empirical findings have important implications for developing appropriate environmental policies. First, policymakers should ensure the transition from fossil fuel energy sources to renewable energy sources. For this purpose, they can make investments in renewable and clean energy projects. Besides, they can support and subsidy the private sector to increase the use of cleaner technologies. Second, governments should help to provide the optimum level of globalization regarding trade and finance since an excessive amount of economic globalization deteriorates the environment. Third, governments should engage in more international cooperation regarding the environmental problems in the world. Finally, they should control the internal migration from rural to urban areas because massive migration movements to urban settings decrease environmental quality.

This study is unable to encompass the entire possible factors affecting environmental degradation such as tourism activities, per capita military expenditure, technology level because the balanced panel dataset is not available for OECD countries during the period 1971–2016. This can be accepted as the main limitation of this study.

In further investigations, it might be possible to use components of the ecological footprint as the ecological indicators. The impacts of input variables on the output variable such as built-up, land, carbon, cropland, fishing, forest, and grazing land footprints can be slightly or entirely different. Also, the causal relationship between these variables could be examined through the recent machine learning techniques; therefore, the policymakers can conduct more effective economic and environmental policies to mitigate environmental pollution.

Supplemental Material