Causality analysis of CO 2 emissions, foreign direct investment, gross domestic product, and energy consumption: Empirical evidence from SAARC countries

Over the period 1980-2016, this study looks into the causal relations among CO 2 emissions, energy consumption (EC), foreign direct investment (FDI), and gross domestic product (GDP) in SAARC countries. To achieve the research objectives, panel unit root tests, panel cointegration autoregressive lag model, and Granger causality tests are used. In the long-run, income has a positive impact on CO 2 emissions, while squared income has a negative impact, conrming the occurrence of the Environmental Kuznets Curve (EKC) theory in SAARC countries. Among all SAARC countries, Bangladesh and Nepal support the pollution haven hypothesis, but India, Pakistan, and Sri Lanka support the pollution halo hypothesis. EC has a large positive impact on CO 2 emissions across the country. In the long-run, the Granger causality test conrms one-way causation from EC to CO 2 emissions and bidirectional causality of FDI and CO 2 . As a result, these countries might encourage clean energy technology through FDI without jeopardising GDP and environmental quality. This will allow these countries to reduce CO 2 emissions to achieve a long-term green GDP and combat global warming. (1980) and the CSD test suggested by Pesaran (2021) to regulate whether the panel data are cross-sectionally dependent or Both tests are calculated with cross-sectional independence as the null hypothesis and CSD as the alternative hypothesis. By assuming potential cross-sectional independence within the panel data, this work analyzed stationary qualities utilizing the second generation unit root tests such as cross-sectionally augmented dickey (CADF) and cross-sectionally Im-Pesaran-Shin (CIPS). Pearson (2007) suggested that the CADF and CIPS are superior to Levin, Lin and Chu (LLC) and Im Pesaran and Shin (IPS) panel unit root tests. The study used the Pedroni and Kao cointegration tests, as suggested by Pedroni (2004) and Kao (1999), to ensure that the investigated variables were cointegrated. Later, the study used ARDL to measure the long-run and short-run effects of our proposed variables on CO 2 . The study also used the pooled mean group (PMG) estimator for further assessment. The Akaike information criterion (AIC) has been applied to select the best lag structure because it reduces the loss of degrees of freedom. It is also treated as the “parsimonious lag structure.” In this case, the PMG estimator is utilized because it maintains long-run coecient control across all cross-sections while allowing short-run coecient diversity across all cross-sections. The ARDL model has some practical advantages, such as the ability to estimate short-and long-term consistent estimates simultaneously, regardless of whether the series is I(0) or I(1). With such a model and a modest sample size, reliable ndings can be produced. Due to the presence of rich

The subsequent segment includes an ephemeral literature review followed by econometric models and data sources. Next, the technical speci cs of several econometric approaches are discussed. Following that, the test and model results are shown. The nal section contains a summary of the entire study as well as a conclusion with policy recommendations.

Related Literature
For the past few years, the causal link between GHG emissions, EC, GDP, and FDI has been a topic of debate (Achour and Belloumi 2016). Social and environmental scientists have spent a lot of effort trying to gure out what causes the variables to interact in the way they do. However, no agreement has been reached on the causal link between the factors because of model misspeci cation, omitted variables bias, and the inclusion of other irrelevant factors (Niemand and Mai 2018). The technique utilized in the studies on the connection of CO 2 emissions, GDP, EC, and FDI varies, as do the variables considered and the area covered (single country or multi-country). However, the current study, based on earlier research, divides the extant literature into three categories:

GDP and GHG emissions
The rst group of studies examines the association between GDP and CO 2 emissions. Nonetheless, empirical evidence from the existing literature shows

FDI and GHG emissions
The association of FDI and GHG emissions is examined in the second area of research. This connection has sparked a discussion among the researchers concerning the positive and negative effects of both variables. Ra ndadi et al. (2018a) used panel data from 1990 to 2004 to observe the in uence of FDI on GHG emissions in a few member states of the Gulf Cooperation Council (CGC). Using the ARDL method, they discovered that FDI encourages GHG emissions. Nguyen (2018) looked into the relationship between FDI and GHG emissions in Vietnam. They discovered a two-way (bidirectional) connection between FDI and CO 2 emissions using the ARDL-Granger causality test. Ssali et al. (2019) examined the causal association of FDI and GHG emissions in six Sub-Saharan African nations. They employed the pooled mean estimator with the ARDL model (ARDL-PMG). In the long-run, their ndings revealed a unidirectional causal relationship between CO 2 and FDI, while no causal associations were discovered in the short-run. Likewise, many studies have determined that FDI drives GDP and has a substantial impact on CO 2 emissions (Behera and Dash 2017a; Riti et al. 2017). However, Mert and Bölük (2016) and Sung et al. (2018) found that FDI improves energy e ciency and reduces CO 2 emissions in host nations.

EC and CO 2 emissions
The nal section emphasizes the link between EC and CO 2 emissions. A large body of empirical research found a strong relationship between EC and CO2 emissions. Through direct or indirect causal relationships, EC is one of the main catalysts of CO 2 emissions (Leito and Balogh 2020; Nguyen 2018). Boutabba (2014) showed unidirectional causality from EC to India's carbon emissions using the ARDL-VECM causality test. Lu (2017), who employed Granger causality in sixteen Asian nations, found similar ndings. On the other hand, Dogan and Aslan (2017) found that EC and CO 2 emissions are bidirectional in a few countries of the European Union. Based on the panel cointegration test, Sterpu et al. (2018) showed that an increase in EC leads to an increase in CO 2 emissions in the atmosphere. They also pointed out that, in contrast to fossil fuels, renewable energy sources minimize CO 2 emissions in the atmosphere. All of the research presented in Tables 1, 2, and 3 used single or multiple nation analyses, different timeframes, and several econometric estimations to explain the association between energy consumption and CO 2 emissions. In terms of the number of studies, the results are inconclusive, citing mixed evidence in terms of direction. As a result, it's critical to remember that while studying the link between variables, the in uence or direction of the association can change.

Econometric Model, Data Sources And Econometric Approach
Using a quadratic model as stated in Eq. (1), the study observed the relationship between GHG emissions, FDI, GDP, and EC in a few SAARC member states.
lnCO 2t = ϕ 1 + ϕ 2 lnGDP t + ϕ 3 lnGDP 2 t + ϕ 4 FDI t + ϕ 5 TEC t + ϵ t 1 where metric tones per capita is used as a unit to measure CO 2 , current US dollars is used to measure GDP, FDI is calculated as a net in ow percentage of GDP, and British Thermal Units (BTU) per capita is used to measure total EC. The subscript t indicates the time period and ϕ i is the intercept where i = 1…5. The error term is represented by ϵ. Except for total energy consumption, the time series data were captured from the database of the World Bank development indicator, 2021. The data on energy consumption comes from the World Energy Statistics 2021. All of the variables used in the regression model are converted to logarithms, which show the elasticity (or percentage change) of the dependent variable. The sign of the parameters ϕ 2 and ϕ 3 is predicted to be positive or negative based on the theory of the environmental Kuznets curve (EKC). At the primary stages of growth of any country, the positive sign of the coe cient ϕ 2 implies a positive association between lnCO 2 and lnGDP. Similarly, the nonlinear inverted U-shaped relationship between CO 2 emissions and GDP per capita is con rmed by the negative sign of the squared parameter ϕ 3 of the EKC. As a result, the presence of EKC for the panel is con rmed by the statistical signi cance of both the positive and negative signs of the concerned parameter. Furthermore, the sign ϕ 4 associated with FDI is used to test the legitimacy of the pollution halo hypothesis or the pollution haven hypothesis. The rst hypothesis contends that FDI reduces GHG emissions, whereas the last hypothesis contends that FDI increases GHG emissions. As a result, the sign of the parameter ϕ 4 contributes to the validation of these two hypotheses. Generally, an intensi cation in energy consumption is projected to result in an escalation of CO 2 emissions as a result of increased economic activity; consequently, the expected sign of ϕ 5 is positive.
Cross-sectional dependence (CSD) is fairly prevalent and regularly encountered in practice when dealing with repeated cross-sectional data. It is self-evident to test for CSD before examining the stationary features of a time series or panel data. The use of unit root and cointegration methods could be adjusted by the CSD; otherwise, the estimate from the unit root and cointegration properties could be inconsistent (Silva et al. 2018). We used the lagrange multiplier (LM) test suggested by Breusch and Pagan (1980) and the CSD test suggested by Pesaran (2021) to regulate whether the panel data are cross-sectionally dependent or not. Both tests are calculated with cross-sectional independence as the null hypothesis and CSD as the alternative hypothesis. By assuming potential crosssectional independence within the panel data, this work analyzed stationary qualities utilizing the second generation unit root tests such as cross-sectionally augmented dickey (CADF) and cross-sectionally Im-Pesaran-Shin (CIPS). Pearson (2007) suggested that the CADF and CIPS are superior to Levin, Lin and Chu (LLC) and Im Pesaran and Shin (IPS) panel unit root tests. The study used the Pedroni and Kao cointegration tests, as suggested by Pedroni (2004) and Kao (1999), to ensure that the investigated variables were cointegrated. Later, the study used ARDL to measure the long-run and short-run effects of our proposed variables on CO 2 . The study also used the pooled mean group (PMG) estimator for further assessment. The Akaike information criterion (AIC) has been applied to select the best lag structure because it reduces the loss of degrees of freedom. It is also treated as the "parsimonious lag structure." In this case, the PMG estimator is utilized because it maintains long-run coe cient control across all cross-sections while allowing short-run coe cient diversity across all cross-sections. The ARDL model has some practical advantages, such as the ability to estimate short-and long-term consistent estimates simultaneously, regardless of whether the series is I(0) or I(1). With such a model and a modest sample size, reliable ndings can be produced. Due to the presence of rich dynamics, the problem of multicollinearity can be readily solved (Dougherty 2016). Finally, because all variables are expected to be endogenous, the ARDL model removes the endogeneity issues that plague the Engel-Granger technique (Seker et al. 2015). The ARDL (pq) model contains lag p used as the outcome and lag q used as the explanatory variable. This model was developed by Ssali et al. (2019) and it takes the following form: where y it denotes the outcome variable. The number of countries and time period used for analysis are symbolized by the subscript i = 12…N and t = 12…T. The other subscript j is the number of cross-sections. The quantity of lags for the outcome and explanatory variables is expounded by p and q.
The explanatory variables x it − j represent a m(row) × n(column) vector. The scalar vector is also expressed by λ ij . The symbol δ ′ ij shows the m × 1 coe cient vector and ϵ ij indicates the error term.
If there is cointegration between the variables under investigation, the ARDL model should be reformulated. In the ARDL model, cointegration could be a concern in short-run dynamics. To obtain a consistent estimate, the error correction term (ECT) should be added. The ECT is used to measure how quickly a dependent variable approaches long-run equilibrium while the independent variable changes. The ECT can be written like this: ij and show the long-run association between outcome and explanatory variables. The symbol φ i is a part of ECT and is likely to be non-positive or less than one. The presence of cointegration is indicated by a negative sign, whereas the pace of adjustment is indicated by a fraction less than one. If φ i = 0 there is no con rmation of a long-run relationship (no cointegration). Therefore, by combining the ECT, our model follows Eq. (4), where CO 2 emission is the outcome variable, and other variables (GDP, GDP 2 , and FDI energy use) are the explanatory variables.
While estimating Eq. (4), we chose the PMG estimator given by Pesaran et al. (1999) since it has various advantages over the others. For example, it accounts for differences in short-run coe cient intercepts and error variation among countries. Long-run estimates are similarly restricted by this estimator to being constant. According to Pesaran et al. (1999), the advantage of PMG is that it is robust to outliers as well as lag orders. The PMG has grown popular among researchers due to its application capacity ( In order to justify the suitability of the pooling coe cient in the ARDL situation, we used Hausman poolability to diagnose the above-mentioned model. This diagnostic test checks if the null hypothesis of pooled long-run coe cients being identical for all cross-sections is true or not (Ssali et al. 2019). Finally, we applied the panel Granger causality test to measure the causality-based vector error correction's direction. The Granger causality test was done in two steps. The long-run relationship is estimated in the rst phase in order to create an ECT for the second step. In the long run, an ECT is de ned as one-period lagged residuals. The ECT sign speci es whether one or two variables are used to correct divergence from the long-run relationship. The VECM is also calculated using biassed-corrected least square dummy variables (Bruno 2005). Boutabba (2014) de ned the VECM as follows: where i = 12…N shows the country, t = 12…Tpresents the time, p indicates the lag length, (1 − B) is the rst difference operator, rϵ it is treated as a serially uncorrelated error term and a nite covariate matrix, and ECT it − 1 shows for lagged ECT. The VECM is used to capture both long-run and short-run Granger causality. The F-test of lagged explanatory factors may be used to determine short-run dynamics, while the t-statistics on the coe cient of lagged error term can be used to determine the signi cance of long-run contributing effects. Our VECM-based modi ed model is applied to test the way of the causality, and that can be written as follows:  Note. The asterisk *** **and * specify statistical signi cance levels at 1% 5% and 10% .

Panel unit root test
To obtain the stationary property and analyze the sequence of integration amongst examining panel time series, this study uses CADF and CIPS unit root tests rather than the traditional unit root test. This is because the CADF and CIPS unit root tests produce valid results in the presence of CSD. As a result, Table 5 shows the outcomes of these two tests. The CADF and CIPS tests reveal that all variables except lnFDI have unit root (non-stationary) at levels, as indicated in Table 5. However, following the initial difference, they become stationary. Given the substantial evidence for a common sequence of integration among the variables under examination, I (1). After the rst difference, the CADF and CIPS tests reject the null hypothesis that the variables are non-stationary, and the involved variables demonstrate the stationary property and I (1). Because the variables are stable at the rst difference, the cointegration test is used to see if there is a long-term relationship between them. Note. The Akaike Information Criterion is used to determine the optimal lags.

Panel cointegration test
The cointegration tests of Pedroni (2004) and Kao (1999) were utilized to con rm the existence of cointegration. Pedroni's panel cointegration test results for several models are shown in Table 6, along with seven different test statistics. At the 1%, 5%, and 10% levels of signi cance, the majority of test ndings, both panel-based and group-based, indicated evidence of panel cointegration among the variables. As a result, the null hypothesis of no cointegration is rejected at a different level of signi cance for the majority of the test statistics. By using the Kao (1999) cointegration test, the homogenous slope coe cient over the cross-section is also shown in Table 6. This test revealed that the investigating variables are cointegrated with a heterogeneous slope by rejecting the null hypothesis of no cointegration with a homogenous slope coe cient at a 10% level of signi cance. Both tests are statistically signi cant, indicating that the panel variables in the models have a long-run cointegrating relationship.

ARDL long-run and short-run elasticities using PMG
To estimate long-run and short-run estimates of CO 2 emissions, GDP, FDI, and ECe, this study used the PMG estimator using ARDL. Table 7 shows the ndings of the PMG estimator. The Chi-squared test statistics of the Hausman test have a value of 72.425 with a connecting p-value of 0.377, according to the ARDL model using the PMG estimator. This p-value is greater than 0.05, indicating that the null hypothesis cannot be rejected and demonstrating that long-run coe cients for all cross-sections are equal. The null hypothesis is rejected, indicating that the homogeneity restriction is valid and that the PMG estimator is more e cient. With an appropriate sign, the coe cient of lag ECT is statistically signi cant (the negative sign and less than unity). Any divergence from the long-run equilibrium of GDP per capita in one year is recti ed by around 72 percent the following year, according to the coe cient of -0.715. Both the short-run and long-run coe cients of lnGDP are positive and signi cant at a 5% level. The coe cient of lnGDP 2 on the other hand, as expected, reveals a negative association and is signi cant in both the short-and long-run. The presence of an inverted U-shaped relationship with CO 2 emissions is revealed by the statistical signi cance of both linear and nonlinear factors of GDP per capita. The conclusion shows that a 1% increase in GDP per capita will result in a 0.68 percent increase in CO 2 emissions per capita, whereas the negative nonlinear term appears to con rm the delinking of CO 2 emissions by 0.024 percent with a higher level of GDP at 5% in the long-run. This positive and negative relationship holds true for estimates of short-run coe cients. These ndings are consistent with the EKC hypothesis, which claims that economic growth is both a cause and a clari cation for the rising GHG emissions. As a result, SAARC member states would not be afraid to pursue their economic growth strategies because, in the long run, economic growth would address the environmental  Note. The asterisk *** **and * specify statistical signi cance levels at 1% 5% and 10% .

Country-speci c ARDL model with PMG estimator
For each cross section, we also run the ARDL model using the PMG estimator. In order to be certain about the presence or lack of autocorrelation, we used the B-G LM test and the Autoregressive Conditional Heteroskedasticity (ARCH) test to evaluate heteroscedasticity. In addition to testing the model's stability, we ran cumulative CUSUM and CUSUMSQ tests on each cross-section. The serial correlation (B-G LM test) and heteroscedasticity (ARCH test) results in Table 8 show that these ARDL models are free from the autocorrelation and heteroscedastic problems. The estimated results of CUSUM and CUSUMSQ corroborate the stability of these models. The EKC can be veri ed by examining the long-run elasticity characteristics of two variables: ΔlnGDP and ΔlnGDP 2 . The longrun coe cients of ΔlnGDP and ΔlnGDP 2 are statistically signi cant, with a positive and negative sign for Pakistan and Sri Lanka. Both the results for Pakistan and Sri Lanka support the EKC hypothesis and economic expansion for CO 2 emissions in Pakistan and Sri Lanka. On the contrary, the anticipated parameters for ΔlnFDI, has a favourable in uence on CO 2 emissions in Bangladesh and Nepal but has a negative impact in India, Pakistan, and Sri Lanka.
The ndings upkeep the presence of the pollution haven hypothesis in Bangladesh and Nepal as well. Table 8 shows that the ΔlnTEC has a large favourable impact on CO 2 emissions in all nations.   Note. The asterisk *** **and * specify statistical signi cance levels at 1% 5% and 10% .

Panel Granger causality test
After applying the ARDL approach to measure the direction of the effects, the VECM-based Granger causality test was used to evaluate the short-and long-run directions of causal effects.  Note. The asterisk *** **and * specify statistical signi cance levels at 1% 5% and 10% .

Conclusion And Policy Implication
Climate change is attracting traction, as is the link between pollution and GDP, FDI, and energy consumption. The EKC hypothesis can be applied to assess the link between ecological impurity and economic growth. Furthermore, rising energy demand and FDI in ows may wreak havoc on developing countries' environmental quality. However, there is continuing discussion over the impact of FDI and energy consumption on environmental norms. Over the last few decades, the trade-off between energy expansion and FDI emissions has also been a hotly disputed topic. Therefore, these variables are signi cant because of the assorted roles they play in designing policies for reducing CO 2 emissions. It is also rational for policymakers to set suitable plans to curb CO 2 emissions. Pakistan, and Sri Lanka. As a result, the pollution haven theory has been discovered for Bangladesh and Nepal, while the pollution halo hypothesis has been found to be true for India, Pakistan, and Sri Lanka. Energy use has a substantial positive impact on CO2 emissions for all countries. Table 9 depicts the direction of the causal relationship between our proposed variables. The VECM-based Granger causality test shows that unidirectional Granger causation exists between GDP and GHG emissions, FDI and GHG emissions, and energy use and GHG emissions. Similarly, there are four bidirectional causalities between GHG emissions and energy use, GHG emissions and FDI, energy usage and GDP, as well as FDI and energy use, which have been detected.
However, there is no causal association between GDP and FDI. The statistical signi cance of ECT suggests that there are two long-run panel causal links between energy use FDI and GDP to GHG emissions, as well as between energy use GDP and GHG emissions to FDI.
The present study has gone some way towards enhancing our understanding of environmental pollution and four vibrant economic variables. It's important to remember that any changes in energy and FDI policies that reduce CO 2 emissions while maintaining economic growth must take into account variables other than the underlying variables in this study. This research can be expanded to include deforestation, institutional quality, rural development concerns, and other environmental variables in the context of SAARC countries as part of a prospective research opportunity. Furthermore, the impact of renewable and nonrenewable energy use on CO 2 emissions by country could be a fascinating issue for additional research.
The generalisability of the present study is subjected to certain limitations. For instance, the study's scope was limited to looking into the causal relationship between CO 2 emissions, GDP, FDI, and energy use in SAARC countries. Despite the fact that SAARC is made up of eight countries, this study only looks at ve of them, based on data from two different sources. Furthermore, due to a lack of data, this study only covers a small number of observations and a short time period . Another potential aw in this study is that it does not account for the structural breaks in individual data series when performing cointegration analysis. Finally, there is a limitation that is linked to the omitted variable bias. In order to avoid omitted variable bias, similar studies include variables that signi cantly contribute to CO 2 emissions and GDP. However, it was not done for this study, and it may be improved by adding more factors.
Our study's outcomes have a number of important insights for future practice. SAARC countries must change their investment strategies to include renewable energy sources in order to achieve economic progress. However, because they are energy-dependent economies, there must be a balance between economic growth and energy use. Alternative energy sources, such as wind energy, hydropower, nuclear energy, and solar energy, should be prioritized. This can be accomplished through expanding and strengthening renewable energy supplies in the targeted countries, as well as increasing energy e ciency around the globe. To counteract GHG emissions, green investment should be encouraged, and clean industries using green technology should be prioritized in these selected countries. Any plan that encourages current polluting industries to move to cleaner technology will have a particularly positive impact on the implementation of policies aimed at lowering GHG emissions. Furthermore, increased public awareness of the bene ts of green technology and the environment would hasten the reduction of GHG emissions.

Declarations
Funding