Analysis on the nexus amid CO₂ emissions, energy intensity, economic growth, and foreign direct investment in Belt and Road economies: does the level of income matter?

Abstract

This study determines the relationship between economic growth, foreign direct investment, energy intensity, and carbon dioxide emissions along the Belt and Road initiative considering their income classification. The study employs data from 1995 to 2015, the panel unit root test, Westerlund cointegration test, augmented mean group estimation, and the Dumitrescu-Hurlin Granger causality test. The empirical results indicate that (1) the data from all income group had cross-sectional association; (2) the variables are integrated of order 1 after first difference; (3) The variables under discussion were cointegrated; (4) at 1% increase in energy consumption, carbon dioxide emissions increased by 0.8606%, 0.9082%, 0.91815%, and 0.8043% in high-, upper-middle-, lower-middle-, and low-income countries, respectively; (5) a bidirectional causal relationship was found between foreign direct investment and carbon dioxide across all income groups. Energy intensity has a bidirectional association with carbon dioxide in low-, upper-middle-, and high-income countries but one-way association in lower-middle-income countries. These recent methodologies take cross-sectional dependence into account in their estimation and findings show that the causal affiliations together with long-run estimated effects amid employed variables are influenced by the different income levels of Belt and Road countries in a tender to reduce carbon dioxide emissions. The empirical results point to some important policy implications.

Introduction

The Belt and Road initiative (BRI) is an ambitious move by China to promote economic cooperation. The “Belt” (Silk Road economic belt) links China with central Asia, South Asia, and Europe. The “Road” (new maritime Silk Road) connects China with Southeast Asian countries, gulf countries, North Africa, and Europe. As an economic cooperation, it provides interaction for organizations, enterprises, and governments along the BRI route. The interaction among these countries basically is to increase their economic growth. As stated by Mishkin (2009) in order to attain economic growth, countries must be open to the world through economic cooperation. Through economic cooperation, the host countries have several benefits not limited to wealth creation, increased in capita income, innovation products, and investments (Shahbaz et al. 2016).

In ripple effects, economic cooperation affects human lives in several ways, being it through overexploitation of natural resources, changes in consumption of energy, economic growth, foreign direct investment (FDI), jobs creation, and environmental changes (Benería et al. 2015). Among the various changes, CO₂ emissions resulting from economic growth and the concomitant changes in the environmental are of much concern. In order to achieve the sustainable developmental goals, especially goal number 13, reduction of CO₂ emissions which is known to be the main contributor to global warming cannot be underscored (Shahbaz et al. 2019). To this light Bulkeley and Newell (2015) stated that Climate change has become a key issue which has been debated worldwide and a global phenomenon which has become a threat to sustainable development.

The effects of economic cooperation on environmental pollution have had diverging opinions such that the endorsement of economic cooperation without paying attention to environmental pollution has been opined albeit the existence of opposing views. The principle that environmental pollutions are the trade-offs for economic benefits is arguable (Wang et al. 2016). Moreover, the economic cooperation benefits are felt only in developed economies as a result of social and political stability (Lipset 2018). Also, Ghosh (2010) stated that the effects of economic cooperation with regards to the environment pollutions are highly susceptible in developing countries owing to weak environmental standards and institutional quality.

The linkage between economic cooperation and pollution has been a subject of global discussion. Mishkin (2009) stated that economic cooperation leads to rise in gross domestic product (GDP), but some studies have hinted that the rise in GDP spikes CO₂ emissions since it lowers the economic credit limits. The quest to establish the long-term link between the environment quality and GDP was set off by early works of Grossman and Krueger (1995). The existing literature widely used Kuznets Curve Environment (EKC) concept to depict the association among economic growth and carbon emissions. The inverted U-shape of EKC suggests the direct relationship between economic growth and the deterioration of the quality of environment. However, as the economic growth reaches its turning limit, the further increase in economic growth leads to environmental improvements (Charfeddine and Mrabet 2017). Recently, the EKC hypothesis has utilized in several studies (Hanif et al. 2019; Kaika and Zervas 2013; Sarkodie and Strezov 2018). Considering an increase in gross domestic product in the BRI countries from US $23.3199 trillion to US$ 25.466 trillion (constant in 2010) between 2014 and 2016 with a growth rate averaged at 1.2% (author computation), the BRI could be marked for emissions of CO₂. Intriguingly, the findings by Sun et al. (2019) earlier envisaged the economic growth along the BRI upon using the PSM-DID to evaluate BRI. Therefore, this study included GDP into the discussion of carbon dioxide emissions in BRI countries.

The quality of the environment can as well be examined using foreign direct investment inflows (FDI). The relationship between FDI and environmental pollution has been controversial one. The pollution haven hypothesis has been the most famous hypothesis supporting the relationship between FDI and environmental pollution (López et al. 2018; Yang et al. 2018). The hypothesis stated that multi-state companies, mostly in developed countries, shift pollution-intensive industries to countries with lower environmental regulations to avoid costly compliance rules in their homeland. Hence, environmental pollution is heavily felt by developing countries making these developing countries a pollution haven. FDI can generate more pollution issues in the host country (Li et al. 2018; Zaidi et al. 2019). Hence, it was appropriate to include FDI in this study.

Energy intensity (EI) has been identified as another determinant of environmental quality. Energy intensity is the amount of energy required to produce a unit of an output (Greening et al. 2000). Since energy is equivalent to the value of converting energy into economic development, it functions as an indicator of energy usage and is debated as a necessary condition for economic cooperation (Overland 2016). Several studies have described a long-term link among EI and environmental pollution (Ben Jebli and Hadhri 2018; Solarin and Al-Mulali 2018). Therefore, this study included EI into the discussion of carbon dioxide emission in “One Belt And One Road” countries.

The relationship between CO₂ emissions, GDP, EI, and FDI has been investigated for both developing countries (Liu and Hao 2018; Salahuddin et al. 2018; Saud et al. 2019b) as well as developed countries (Cai et al. 2018; Sarkodie and Strezov 2019) for which the long-term cointegration between those variables was established. It is worth noting that the validity of the EKC was established for some studies (Abdouli et al. 2018; Phuong 2018). The summary of literature in relation to this study is presented in Table 1.

Table 1 Summary of some recent empirical studies on the relationship among GDP, EI, FDI, and CO₂ emissions

Although the relationship between CO₂ emissions, GDP, EI, and FDI has been well studied and reported in literature, the various income classification among the countries under studied remains vague. In response to this deficiency, we propose the significance of analyzing the different income samples from different countries. It is noteworthy that BRI can be a platform for countries to make greater contributions to achieving CO₂ emission reduction targets (Liu and Hao 2018). By exploring links between these variables, researchers may be able to help determine whether FDI or EI is the main driver of increased CO₂ emissions in various income groups. If FDI or EI helps mitigate emissions in income groups, then pursuing more FDI and the usage of EI will have beneficial effects on mitigating CO₂ emissions. On the other hand, if FDI or EI increases CO₂ emissions in income groups, then policies in reduction CO₂ emission should be considered.

Therefore, the present study seeks to establish the dynamic link between CO₂ emission and GDP, EI, and FDI of countries on the “One Belt And One Road” taking into consideration the income classifications. Also, the newly developed panel data augmented mean group (AMG) estimator employed in the study. More obviously, the advantage of the AMG is it robustness when dealing with cross-sectional dependency and the parameter of universal dynamic effect.

The remainder of this paper is organized as follows: section “Methodology” presents the methodology followed by empirical analysis; section “Long-run estimation and coefficient analysis” presents long run estimation and coefficient analysis, lastly with conclusion and policy recommendation.

Methodology

Theoretical model specification

Examining the causal relationship among GDP, EI, FDI, and CO₂ emissions, this study adopts Balsalobre-Lorente et al. (2018) and others’ model; hence, we write our carbon emission estimate function as:

$$ {\mathrm{CO}}_2=\left(\mathrm{GDP},{\mathrm{GDP}}^2,\mathrm{EI},\mathrm{FDI}\right) $$

(1)

Where CO₂ is carbon emission, GDP represents gross domestic product, GDP² represents gdp squared, to measure EKC, EI stands for energy intensity, and foreign direct investment is represented by FDI. To address the problem of heteroskedasticity, the variables converted into natural log. Therefore, our multivariate carbon emission function for our natural log model is given by

$$ {LnCO}_{2 it}={\beta}_0+{\beta}_1{LnGDP}_{it}+{\beta}_2{LnGDP^2}_{it}+{\beta}_3 Ln{EI}_{it}+{\beta}_4 Ln{FDI}_{it}+{\varepsilon}_{it} $$

(2)

where β₀ represents the slope coefficient, i denotes the countries selected in this study (1, 2… N), t indicates the time frame for the analysis, and ε_it designates the error term. β₁, β₂, β₃, and β₄ are the coefficients of GDP, GDP², EI, and FDI. The association between economic growth and CO₂ emissions is known as the environmental Kuznets curve hypothesis (EKC). According to the EKC hypothesis, economic growth is initially accompanied by high carbon dioxide emissions and then declines as the economy reaches a mature level and reaches the threshold of real income per capita (Stern 2004). We therefore expect β₁ > 0, β₂ < 0 if linkage between economic growth and CO₂ emissions is inverted U-shaped, i.e., EKC hypothesis otherwise β₁ < 0, β₂ > 0 if the relationship is U-shaped between economic growth and CO₂ emissions.

Econometric approach

This study employs newly developed panel data analysis to obtain the empirical results. Panel data analysis is generally superior to pure cross section analysis. The advantage of panel data is its ability to provide least collinearity between larger data sets, greater variability among variables, which is not a criterion in cross-sectional data analysis. Therefore, a more reliable estimate can be obtained in the empirical analysis. In addition, the advantage of using panel data is the power to check individual heterogeneity between groups. The framework of the methods is as follows:

Cross-sectional dependence

Cross-sectional dependence is usually found in panel data, since countries are related at regional, income, and global level. CSD in different groups may occur due to mutual shock, spillovers, or common factors that cannot be observed. If studies ignore the existence of CSD, then the efficiency of the estimated results is arguable (Urbain and Westerlund 2006). Hence, before the empirical analysis, the Pesaran (2004) CD test was used to access the cross-sectional dependency. The panel data model can be described as

$$ {y}_{\mathrm{i}\mathrm{t}}={\alpha}_{\mathrm{i}}+{\beta}_{\mathrm{i}\mathrm{t}}{x}_{\mathrm{i}\mathrm{t}}+{\mu}_{\mathrm{i}\mathrm{t}} $$

(3)

where i = 1, 2, ……N, t = 1, 2……T, β_it is the K × 1 parameter vector to be estimated, x_it is the K × 1 explanatory variable, α_i is the individual redundant parameter, and μ_it is the time invariant assuming independent and identical distributions. The Pesaran (2004) statistic test is given by

$$ CD=\sqrt{\frac{2T}{N\left(n-1\right)}}{\sum}_{i=1}^{N-1}{\sum}_{j=i+1}^N\hat{\rho_{ij}} $$

(4)

where N is the sample size, T represents the period, and ρ_ij is the product correlation errors of country i and j.

Panel unit root test

Pesaran (2007) CD cross-sectionally dependent augmented dickey fuller (CADF) test was adopted since it considers CSD. The regression for this test is given as

$$ \Delta {y}_{it}={\alpha}_i+{\beta}_i{y}_{it-1}+{\theta}_i{\overline{y}}_{t-1}+\sum \limits_{j=1}^{\rho }{\gamma}_{ij}\Delta {y}_{it-1}+\sum \limits_{j=0}^{\rho }{\delta}_{ij}\Delta {\overline{y}}_{t-j}+{d}_{it}+{\varepsilon}_{it} $$

(5)

where ${\overline{y}}_{\mathrm{t}}=\frac{1}{N}{\sum}_{i=1}^N{y}_{\mathrm{it}}$ and its inclusion in the equation can be used to replace the effect of the unseen common factor. α_i is the time invariant individual intercept parameters, β_i, θ_i, γ_ij and δ_ij represent individual specific effect, individual specific linear trend, and common time effect, respectively. ε_it is the error term. According to Pesaran (2007), stationarity test can be performed on the t-value of β_i, either seperately or jointly. Since the test is like to the IPS statistic of Im et al. (2003), it is given as;

$$ CIPS\left(N,T\right)=\frac{1}{N}\sum \limits_{i=1}^N{t}_i\left(N,T\right) $$

(6)

where t_i(N, T) is the measure of β_i in the equation above.

Panel cointegration

In observing the long run association amidst the variables in the model, the Westerlund-Edgerton bootstrap panel cointegration was adopted. They proposed four panel cointegration tests for the null hypothesis of no cointegration, taking into account structural dynamics. This test does not enforce any common constraints. This test not only provides good results, but also applies to all situations where CSD exists or does not exist. Westerlund and Edgerton (2007) suggested four residual test methods to evaluate the null hypothesis of no cointegration. Two of the tests are panel statistics (Gt and Ga), and the other two are group statistics (Pt and Pa), which are normally distributed. In effect, this test measures the existence of cointegration by judging whether there is an error correction in a single panel group and the whole panel. The model was built on

$$ {y}_{it}={\delta}_{0i}+{\delta}_{1 it}+{n}_i{D}_{it}+{x}_{it}^{\prime }{\beta}_i+{\left({D}_{it}{x}_{it}\right)}^{\prime }{\gamma}_i+{z}_{it} $$

(7)

where i = 1……N, t = 1……T, x_it = x_{i, t − 1} + v_it is the k-dimensional vector being I(1). D_it represents the break dummy variables. D_it = 1 if $t>{T}_{\mathrm{i}}^{\mathrm{b}}$ and zero otherwise.${T}_{\mathrm{i}}^{\mathrm{b}}$ represents the break for individual i.

Long run parameter estimation

AMG estimator formulated by Bond and Eberhardt (2013) was used in estimating the variables in this study. Ma (2015) stated that AMG method is robust in handling cross-sectional dependence. In addition, the AMG algorithm is not restricted by non-stationarities of variables during estimation (Balcilar et al. 2019). The main panel model Eq. (2) can be calculated as follows:

$$ \Delta {LnCO}_{2 it}={\beta}_0+{\beta}_1\Delta {LnGDP}_{it}+{\beta}_2\Delta {LnGDP^2}_{it}+{\beta}_3\Delta Ln{EI}_{it}+{\beta}_4\Delta Ln{FDI}_{it}+\sum \limits_{t=2}^T{q}_t\left({\Delta D}_t\right)+{\mu}_{it} $$

(8)

Equation 8 represents a standard ordinary least square (OLS) regression in the first difference (FD-OLS) with T-1 period dummies in the first difference(∆D_t), where q_t is the parameter of the periodic dummy. Equation 9 contains ω_t in the model. The inclusion of ω_t is to compensate any excluded idiosyncratic process which evolve over time. ω_t is subtracted from the dependent variable (equation 10), which means a common procedure is imposed on each set of unit coefficient. The AMG estimates are then derived as averages of the individual country estimates. The regression model for the group specific was first adjusted with β_t; then, the average group-specific parameters were computed.

$$ \Delta {LnCO}_{2 it}={\beta}_0+{\beta}_1\Delta {LnGDP}_{it}+{\beta}_2\Delta {LnGDP^2}_{it}+{\beta}_3\Delta Ln{EI}_{it}+{\beta}_4\Delta Ln{FDI}_{it}+{d}_1\left({\omega}_t\right)+{\mu}_{it} $$

(9)

$$ \Delta {LnCO}_{2 it}-{\omega}_t={\beta}_0+{\beta}_1\Delta {LnGDP}_{it}+{\beta}_2\Delta {LnGDP^2}_{it}+{\beta}_3\Delta Ln{EI}_{it}+{\beta}_4\Delta Ln{FDI}_{it}+{\mu}_{it} $$

(10)

Causality estimation method

AMG algorithm estimation only gives the long-run estimation for the variables and cannot give the direction of causality unlike pooled mean group (PMG). Hence, the panel causality test proposed by Dumitrescu and Hurlin (2012) was employed. The D-H causality test adapts heterogeneousness and cross-sectional dependence in panel data, which cannot in the case vector error correction model (VECM) granger causality. The data model can be specified as

$$ {y}_{i.t}={\alpha}_i+\sum \limits_{i=1}^p{\gamma}_i^{\left(\rho \right)}{y}_{i.t-n}+\sum \limits_{i=1}^p{\beta}_i^{\left(\rho \right)}{x}_{i.t-n}+{\mu}_{i.t} $$

(11)

where n refers to lag length, x and y are the basic variables for n cross section in t perios.${\gamma}_{\mathrm{i}}^{\left(\uprho \right)}$ and ${\beta}_{\mathrm{i}}^{\left(\uprho \right)}$ are the autoregressive parameters and regression coefficient for each panel or country, respectively.

Empirical results

Data source and description

About 70 countries make up the BRI; however, due to the availability of data and since this study used balanced data sets, only 44 countries were selected for this study with respect to the variables involved, with period 1995–2015. According to the 2014 world bank atlas method of gross national income per capita (GNI), the 44 countries were divided into four categories: high income (more than $12,736), upper-middle income ($4126–$12,736), lower-middle income ($1046–$4125), and low income (less than $1045). The HIC in this study consists of data from 15 countries, UMIC, LMIC, and LIC consist of data from 13, 10, 6 countries, respectively (Table 10). The data is converted into natural logarithms for the coefficient estimates to be explained as the elasticity of the dependent variable (carbon emissions). The countries involved were selected from the belt and road initiative book published in May 2016. Variables selected as a result of data are exemplified in Table 2 with their definition, symbol, and unit of measurement.

Table 2 Variable description and data source

Descriptive statistics for the selected variables with respect to the sampled countries and the various income groups are presented in Table 3. It reveals that for the selected sampled countries, GDP is on average of 8.491 with a standard deviation of 0.918 compared to CO₂ (M = 10.727, SD = 1.765). EI and FDI have M = 8.056 , SD = 1.411 and M = 15.662, SD = 1.701 , respectively. Comparing the descriptive for the various income groups, Table 3 again depicts that for GDP, HIC have M = 8.703 , SD = 0.720, UMIC have M = 8.881, SD = 0.304, LMIC have M = 8.407 , SD = 0.505, and LIC have M = 7.256, SD = 1.519, indicating that GDP is relatively high in UMIC followed by HIC then with LMIC and LIC having the lowest GDP among the income classification. In regard to CO₂ emissions, HIC (M = 10.851, SD = 1.429), UMIC (M = 11.118, SD = 1.811), LMIC (M = 10.911, SD = 1.998), and LIC (M = 9.261, SD = 1.208), depicting CO₂ emissions is averagely higher in UMIC and low in LIC among the groups. Considering energy intensity among the groups, HIC (M = 9.350 , SD = 0.767), UMIC (M = 7.990, SD = 0.914), LMIC (M = 7.026, SD = 0.746), and LIC (M = 6.680, SD = 1.598), indicating that energy consumption is averagely high in HIC and low in LIC among the classification. In regards to FDI, HIC (M = 14.756, SD = 1.304), UMIC (M = 15.761, SD = 1.731), LMIC (M = 16.615, SD = 1.841), and LIC (M = 16.122, SD = 1.037), revealing that FDI inflows is relatively high in LMIC while HIC has the lowest inflows among the income classification.

Table 3 Summary of descriptive statistics

For normality test, skewness and kurtosis were to specify the normality assumption. For normal distribution, skewness and kurtosis must be 0, and 3 respectively. The results considering the sampled BRI countries, Table 3 reveals skewness for GDP and the squared of GDP are to the left which means that these variables are more to the left than normal distribution. CO₂ emissions, FDI, and EI are skewed right, that is flattering to the right. This indicates that three out of five of the observations are heavily right tailed. With the peakness (kurtosis) of the distribution, CO₂ emissions and FDI were approximately 3, indicating a mesokurtic distribution. For the value of GDP, squared of GDP and EI were above 3 which also symbolizes that their distribution is leptokurtic. Therefore, kurtosis and skewness conditions for normality were not satisfied by any variables; hence, we affirm that none of the variables were normally distrusted. Table 3 again reveals the violation of the normality assumption among the income groups for all the variables. In HIC and UMIC, GDP, GDP², and EI are flattering to the left while CO₂ and FDI are skewed to the right. In LMIC, GDP, GDP², and CO₂ are skewed to the left while EI and FDI are flattering to the right. This results from the three groups that indicate that three of variables are heavily left tailed. However, in LIC, CO₂, EI, and FDI are skewed to the right while GDP and GDP² are skewed to the left indicating more variables tailed to the right. The rejection of the normality assumption was strongly supported by the Jarque-Bera test for normality with the null hypothesis that all the variables follow a normal distribution at a rejection of a probability less than 0.05.

Multicollinearity among the independent variables (GDP, the square of GDP, EI, and FDI) was done using correlation and variance inflating factor. Table 4 gives the results of the interconnection test. The VIF value must not be greater than 5 and the tolerance value should be greater than 0.2 (Craney and Surles 2002). Hence, it can be concluded that each independent variable is uniquely impacting the dependent variable. The results from the correlation test support that there is no strong correlation among the independent variable (threshold 0.7). Hence, each variable is independent of the other and a unique and significant impact on CO₂ emission.

Table 4 Multicolinearity test results

Cross-sectional dependence test

Cross-sectional dependency (CSD) has been the key focus in current energy-economic literature (Dogan and Inglesi-Lotz 2017; Ozcan and Ari 2017). Table 5 gives the value of the CSD test and their corresponding probability values. CSD test values of each series in each panel (sampled BRI, HIC, UMIC, LMIC, and LIC) are significant at 1% level. Hence, the countries involved are connected in some way. Having confirmed CSD among groups of countries, the study relies on the second-generation panel unit root test to check stationarities of the variables. Therefore, CIPS and CADF were used in this study.

Table 5 Cross-sectional independence test results

Panel unit root test

Table 6 indicates that variables in this study have unit roots at the level; however, in their first difference, they have no unit roots. Convinced that all variables are non-stationary at I(0), but become stationary at I(1), it is appropriate to study the existence of long-term association among variables.

Table 6 Results from CADF and CIPS panel unit root test

Panel cointegration test

Westerlund-Edgerton bootstrap panel cointegration test was utilized to analyze whether there was a long-run association among variables. Techniques such as Kao and Johansen’s panel cointegration ignore CSD when testing for the long-term association. Findings from the Westerlund-Edgerton bootstrap panel cointegration are displayed in Table 7. Using CO₂ as the dependent variable revealed that all the variables have long-term association in the various panels. Hence, the null hypothesis of no long-term association is declined at various significant level with respect to the statistics G_τ, G_α, P_τ, and P_α. Our argument was based on the robust p values which gives a robust evidence of long-term association within the series.

Table 7 Panel cointegration test (Westerlund and Edgerton (2007)

Long-run estimation and coefficient analysis

Long-run analysis (AMG estimation)

In the estimation analysis of our parameters, the paper employed AMG estimator to check the effects of GDP, GDP², EI, and FDI on CO₂ emissions. The results obtained are displayed in Table 8. The relationship between coefficients of GDP² and CO₂ emissions reflects the concept of EKC. When GDP increases by 1%, CO₂ discharge also goes up by 0.0189%, while 1% rise in GDP², shorten the quality of the environment by 0.2055% which is statistically significant for all countries in BRI. Likewise for HIC when GDP accelerates by 1%, the quality of the environment is shortened by 0.0208% while a percentage mount in GDP² diminishes CO₂ discharge by 0.3243% significantly. Also, for UMIC 1% increases in GDP vitiate the quality of environment by 0.01660% which is statistically significant, while 1% rise in GDP² quashes the quality of the environment by 0.1909% significantly. Similarly in LMIC countries, a 1%, rise in GDP also corrupts the environment by 0.0095%, while a 1% boost in GDP² increases CO₂ emissions by 0.3094% but not significant.

Table 8 Results of AMG panel data estimation method

Likewise for LIC when GDP increases by 1%, it weakens the environment quality by 0.0572%, and while 1% rise in GDP² diminishes the quality of the environment by 0.9552% but not statistically significant. Considering the insignificant GDP² in LMIC and LIC suggests that EKC does not exist in this income group. This may be due to the same concept as other developing economies, LMIC and LIC have not yet achieved a full-standard industrial economy. The inverted U-shaped in HIC suggests increased economic growth in HIC at early stage also stimulated carbon dioxide emissions to a certain level in the early stages, but after reaching this limit, carbon dioxide emissions began to decline as economic growth increased further.

At the initial stages of economic growth, these countries focused mainly on economic expansion, ignoring the environment and aiming at boosting trade and infrastructure with other countries through cooperation and financial development. Economic growth has increased personal income, thus increasing energy demand and worsening environmental conditions. Finally, rising income levels bring social and environmental awareness, which helps abridge environmental pollution (Zaidi et al. 2018). The evolution of EKC effect is due to the improvement of mass production technology and citizens’ demands on environmental quality. These results are consistent to those of Rafindadi and Ozturk, 2017for African economies, Haseeb et al. (2018) for BRICS countries, Bekhet and Othman, 2017 on Malaysia, and Sinha and Shahbaz (2018) for India. The U-shape in BRI panel and UMIC is an indication of no strong evidence for the existence of EKC for CO₂ emissions. Therefore, these results are consistent with Liu and Hao, 2018 research in investigating EKC hypothesis in the BRI, Balaguer and Cantavella (2016) for Spain, Pata (2018) for Turkey.

The results also display that EI has a significant positive relation with CO₂ emissions, as a 1% rise in EI corresponds to a 0.9441% rise in CO₂ emissions in BRI panel. In their income group levels, 1% increase in EI increases 0.8606%, 0.9082%, 0.91815%, 0.8043% rise in CO₂ emissions in HIC, UMIC, LMIC, and LIC, respectively, which are statistically significant. This result depicts that the establishment of more local and foreign companies in these countries enable local residents have more jobs and energy needs rises. The increment of firms enhances energy usage which then increases CO₂ emissions. However, at later stages, when the local and foreign firms have matured, they invest in energy-efficient infrastructure to lessen CO₂ emissions. Our results show that energy use is the main cause of carbon dioxide emissions. Therefore, BRI construction must shift energy consumption to renewable energy and achieve zero-emission growth. In addition as stated by Qasemi-Kordkheili and Nabavi-Pelesaraei (2014), energy efficiency demands innovative models to estimate the optimization of energy need and the potential decrease of greenhouse gases. We support their view that traditional energy can promote economic growth, but it is very harmful to the environment since it increases carbon dioxide emissions.

For the link between FDI with CO₂ emissions, a percentage increment in FDI pollutes the environmental quality by 0.6945% for all countries involved in the whole panel. For high-income countries, a 1% increase in FDI mitigates the quality of environment by 0.0581%. In low-income countries, a rise of 1% in FDI also corrupts the environment by 1.8287%. It can be concluded that FDI stimulates carbon emission in the long term in this panel group.

The RMSE value is a good measure of how accurately the model predict the response variable. Small value of RMSE indicates a better model fit; hence, it can be deduced that the values of RMSE for BRI countries, HIC, UMIC, LMIC, and LIC are 0.0553, 0.0353, 0.0383, 0.02558, and 0.0432, respectively. Therefore, each panel model is fit in predicting the CO₂ emissions.

Panel causality test

Dumitrescu and Hurlin (2012) causality test (D-H) was used due to the occurrence of CSD among the variables since AMG estimate does not propose a causal path. In the D-H test, the significance of causal relationship is tested by two kinds of statistics: w-bar statistics (average statistics are used for the test) and z-bar information (standard normal distribution is used for the test). The results of D-H causality test are reproduced in Table 9 for all countries in the Belt and Road initiative.

Table 9 Summary results from D-H granger causality test

For all county panel, the results proposed a bidirectional causality amidst CO₂ emissions and all the independent variables (GDP, EI, and FDI). A bidirectional causal effect was found between GDP and EI, GDP and FDI, and EI and FDI. This outcome implies that EI in countries on the BRI and their economic growth is correlated, such that increase in EI will spark economic growth of these countries. This outcome is in consonant with the study done (Li et al. 2015; Saud et al. 2019c). For HIC, there was an evidence of a unidirectional causality between GDP and CO₂ emissions. Notwithstanding, a bidirectional causality between CO₂ and EI, CO₂ and FDI, and EI and FDI was depicted. Last but not the least, there is a one-way causal effect from EI to GDP. In the case of LIC, there is a unidirectional causality from CO₂ emissions to GDP. Likewise, a bidirectional causal effect between CO₂ and FDI, EI and GDP, and EI and FDI was revealed. Finally, there is a one-way causal relationship from FDI and GDP.

For LMIC, there is a two-way causal relationship that was depicted between CO₂ and GDP, CO₂ and FDI, and GDP and EI. The results also suggest a unidirectional causality from EI to CO₂ emission, likewise from FDI to EI. Lastly for the UMIC, a bidirectional causal effect amidst CO₂ emissions and the each of the other variables (GDP, EI, and FDI), likewise a bidirectional association between GDP and EI and a unidirectional causality from EI to FDI was depicted. Interestingly, the relationship between EI and CO₂ emissions in all panel groups is in consonant with the findings of Liu and Hao (2018) in their study on BRI countries, Asafu-Adjaye et al. (2016) for of 53 countries globally. Not forgetting the causal relationship amid CO₂ emissions and GDP, the findings are consistent with that of Shahbaz et al. (2015) in studying some African countries, and Ali and Malik (2018) for the study on Pakistan. The results from the D-H granger causality links are summarized in Fig. 1.

Fig. 1

Graphical representation of two-sided and one-sided causal affiliations among variables in different country group

Conclusion and policy recommendation

This empirical study seeks to establish the dynamic nexus among CO₂ emissions, GDP, EI, and FDI along the BRI countries using data from 1995 to 2015. The main panel was divided into sub-panel group using the four income classification: HIC, UMIC, LMIC, and LIC. Analysis was done on the sampled panel and the sub-panels. In summary, Pesaran CD’s test was done to determine CSD among the variables. In addition, the CADF and CIPS panel stationarity tests were performed due to the presence of CSD. From the result, we infer that the variables are unstable at their level, but stable at their first difference. Westerlund-Edgerton panel bootstrap cointegration test was used to determine if the variables are cointegrated. From the results, the variables were cointegrated and we deduced that there is a structural long-run relationship. Hence, we employed AMG estimator which is more robust to CSD in estimation of the variables. Since AMG cannot indicate the direction of the long-run relationship, Dumitrescu and Hurlin (2012) panel causality test employed. The result from the causality test indicated that the results depicted that LIC, UMIC, and HIC followed the feedback hypothesis, which is indicated by a bidirectional causal effect between CO₂ and energy consumption. That income groups (HIC, UMIC, and LIC) exhibited a bidirectional association among EI and CO₂ emissions with exception of LMIC which has a one-way causal effect from CO₂ to EI. However, the main sampled BRI countries, a bidirectional causal effect amidst CO₂ and EI. In addition, all panel exhibits a bidirectional causal effect between CO₂ and FDI, indicating that environmental pollution is enhanced by FDI inflows. Again, UMIC and LMIC displayed a bidirectional causal relationship among GDP and CO₂ emissions. While a unidirectional causal effect was depicted from GDP to CO₂ emissions in HIC and LIC, EKC was found in HIC (inverted U-shape) while in UMIC, LMIC, and LIC, no presence of EKC (U-shape).

Policy implications in regards to this study will be as follows; first, the results indicate that for most countries, CO₂ emissions are energy-driven; the long-term two-way granger causality affirms this conclusion. This result shows that countries along the BRI must strive to establish effective energy policies to reduce CO₂ emissions in order to achieve the sustainable development goal (13). Hence, reducing the use of fossil fuels which is the main cause of CO₂ is a shared responsibility. Secondly in regards to FDI-CO₂ emission relationship, countries along the BRI need to develop policies that encourage public-private partnerships in producing safe energy from different sources. Thirdly from these findings is the role that a country’s income level plays, the most common factor in increasing CO₂ emissions among all income groups (HIC, UMIC, LMIC, and LIC) is FDI. However, EI also plays an important role in increasing CO₂ emissions in LIC, UMIC, and HIC while GDP increases CO₂ emissions in UMIC and LMIC. Given these differences, decision makers need to take into account the effects of various variables in their decision-making process and consider them in different ways depending on the income level of a given country. Lastly, for all income groups, GDP tends to mean increase in energy consumption and increased CO₂ emissions; therefore, promoting the transition to a renewable or nuclear energy usage is best way to reduce environmental pollution associated with economic cooperation.

The limitation of this study is that some BRI countries were excluded from the panel set due to missing data. Although this study conducted a preliminary quantitative study on CO₂ emission, GDP, FDI, and EI of countries along the BRI, still has some limitations, which may become the direction of future research. Future research may include all countries using unbalanced data when investigating the income groups in BRI countries. Again, since the relationships discussed in this paper may be quadratic, some appropriate methods, such as panel smooth transition regression models, can be used to establish relationships when the time frame is much longer. Due to non-linear reasons, the longer the sample time and the longer the period, the clearer the result.

Appendix

Table 10 Income classification of countries

Table 11 D-H causality for BRI countries

Table 12 D-H causality for high-income countries

Table 13 D-H causality for Low Income countries

Table 14 D-H causality for Lower middle Income countries

Table 15 D-H causality for upper-middle-income countries