The Impact of Trade Openness on Carbon Emissions: Empirical Evidence from Emerging Countries

Abstract

Emerging countries are the main source of new CO₂ emissions and the major net carbon importers, and they have also become an important part of the global trade pattern. In this study, the impact of trade openness on CO₂ emissions was investigated by approaches such as fully modified least squares (FMOLS), dynamic ordinary least squares (DOLS), and pooled mean group-autoregressive distributive lag (PMG-ARDL) methods. Further estimations were conducted by employing methods such as DCCEMG (dynamic common-correlated effect mean group) and Driscoll–Kray to strengthen the robustness of the results. Moreover, the Granger causality between trade openness and CO₂ emissions was tested by using the Dumitrescu–Hurlin method. Conclusions can be drawn as follows: First, economic growth, energy consumption, trade openness, and CO₂ emissions are all interconnected in the long term. Specifically, higher levels of economic growth and trade openness are associated with lower CO₂ emissions, whereas energy consumption contributes to higher emissions. However, in the short term, economic growth and energy consumption lead to an increase in CO₂ emissions, while trade openness does not have a significant impact. Moreover, there is a two-way Granger causality between trade openness and CO₂ emissions. Additionally, economic growth and energy consumption have an indirect effect on CO₂ emissions by influencing trade openness. Given these findings, emerging market countries should focus on enhancing their service sectors, promoting technological advancements, and fostering international collaboration in green technologies. By actively engaging in efforts to combat climate change, these countries reach a point where trade expansion and carbon reduction are achieved.

1. Introduction

Global warming has become a challenge facing the world nowadays due to its severe negative impacts on ecosystems, economies, and societies. Although causes of global warming remain controversial [1], human activity causing global warming has been recognized by increasingly more research. For example, the increase in greenhouse gases from industrial production and residential consumption is one of the major causes. To control global warming, reducing greenhouse gas emissions has become a global consensus. Greenhouse gas emissions were restricted by international regulations for the first time in the Kyoto Protocol, requiring major industrial countries to assume responsibility for reducing greenhouse gas emissions. The Paris Agreement not only sets a target for temperature control less than 2 °C from the pre-industrial revolution level and strives to achieve a 1.5 °C rise but also mobilizes countries around the world to reduce emissions independently.

As shown by the World Development Indicators (WDI) data, emerging countries have become the main source of new CO₂ emissions, while CO₂ emissions in major industrial countries have declined significantly. Among the top ten countries and regions for CO₂ emissions, China, India, Indonesia, Russia, Brazil, and Iran are all emerging countries, and their CO₂ emissions have increased except Brazil. In Figure 1, China’s total CO₂ emissions increased from 4.424 in 2003 to 10.945 billion metric tons in 2020, and CO₂ emissions per capita increased from 2.65 in 2000 to 7.76 metric tons in 2020. India’s CO₂ emissions increased from 1.012 in 2003 to 2.201 billion metric tons in 2020, and CO₂ emissions per capita increased from 0.87 in 2000 to 1.58 metric tons in 2020. In contrast, America’s total CO₂ emissions declined from 5.659 in 2003 to 4.321 billion metric tons in 2020, and CO₂ emissions per capita declined from 20.10 in 2000 to 13.03 metric tons in 2020. Japan’s total CO₂ emissions decreased from 1.214 in 2003 to 1.014 billion metric tons in 2020, and CO₂ emissions per capita decreased from 9.94 in 2013 to 8.03 metric tons in 2020. Moreover, CO₂ conversion technology, particularly the conversion of CO₂ into fuels [2], is rapidly advancing. However, emerging market countries have lagged significantly behind developed countries in technology, making it difficult for them to mitigate CO₂ emissions through CO₂ conversion technology.

What is driving the rapid increase in carbon emissions among emerging market countries? Studies suggest that the surge in economic growth and energy consumption plays a significant role in driving up carbon emissions in these nations [3,4]. However, there are two notable facts in the global carbon emission landscape: the share of CO₂ emissions embodied in international trade accounts for 1/4 to 1/3, and rich countries are increasingly outsourcing CO₂ emissions. The data from “Our World in Data” on trade-embodied CO₂ emissions shows that China and India have consistently been net carbon exporters from 1992 to 2020, but the United States and Japan have been net carbon importers (Figure 2). This trend underscores the significant role that international trade plays in affecting environmental quality, particularly CO₂ emissions [5,6,7].

Nevertheless, the study on the nexus between trade openness and carbon emissions essentially belongs to the study on trade liberalization and environmental pollution. The trade impact on the environment has been classified into scale effects, technique effects, and composition effects [8,9]. That is, its impact depends on the net effect resulting from the interaction of the above three effects. The expansion of the economy in terms of scale or a focus on upstream industries in the industrial structure leads to increased CO₂ emissions with trade openness. On the other hand, when technological contributions drive the economy or when the service trade becomes more significant, trade openness can help decrease CO₂ emissions. This suggests that trade openness can contribute positively, negatively, or not at all to the growth of CO₂ emissions, implying that greater trade openness does not automatically result in higher CO₂ emissions.

To verify this view, the impact of trade openness on CO₂ emissions was estimated by FMOLS, DOLS, PMG-ARDL methods, and the causality between them was further explored in this study, with emerging countries as the objects of study. This paper contains seven parts: Introduction; Literature Review; Research Methods and Data; Unit Root Test and Cointegration Test; Estimation Results; Causality Test; and Conclusions, Policy Implications, and Limitations.

2. Literature Review

With the increasing concern about global climate change, exploring the impact of trade openness on CO₂ emissions has become a research topic of great concern. Herein, the existing literature was divided into two categories (Table 1).

2.1. Studies Based on Time Series Data

Some scholars analyzed the impact in each country using time series data, including emerging, developing, and developed countries. In terms of methods, econometric methods such as the ARDL bounds test and the Vector Error Correction Model (VECM) Granger Causality were adopted by most of the literature, and a little literature used other methods. For example, Mutascu [6] used the Wavelet Toolbox, Hdom and Fuinhas [10] applied FMOLS and DOLS, and Udeagha and Ngepah [11] utilized Novel dynamic ARDL simulation. The following conclusions were made: first, the higher the degree of trade openness, the higher the CO₂ emissions [12]. Second, trade openness has no impact on CO₂ emissions [13]. Third, a two-way nexus is present between them [14]. Fourth, the nexus between them is uncertain [15]. Despite these different conclusions, most studies argued that trade openness increases CO₂ emissions.

2.2. Studies Based on Panel Data

In recent years, the impact has been explored by some scholars using panel data models. Two major conclusions were drawn: First, enhanced trade openness increases CO₂ emissions. For instance, Ahmed et al. [16], and Zhang et al. [17] studied the impact of trade openness on CO₂ emissions using FMOLS and VECM Granger Causality and believed that enhanced trade openness increases CO₂ emissions. AI-Mulali and Ozturk [18], Jebli et al. [19], and Destek et al. [20] also employed FMOLS and VECM Granger causality to demonstrate the interaction between them. Chen et al. [21], by virtue of panel quantile models, concluded that enhanced trade openness in the 64 Belt and Road countries increased CO₂ emissions. Moreover, Zheng et al. [22] studied the nexus between trade openness and CO₂ emissions in ten Asian countries using CS-ARDL, AMG, CCEMG, and confirmed that enhanced trade openness increases CO₂ emissions.

Second, whether enhanced trade openness increases CO₂ emissions remains inconclusive. For example, Sun et al. [23] found through methods such as FMOLS, panel VECM (Vector Error Correction Model), and Granger causality that enhanced trade openness increases or decreases CO₂ emissions under different income levels. Iqbal et al. [24] used FMOLS, DOLS, and the system GMM to study this impact in countries with different income levels. They concluded that trade openness increases CO₂ emissions in global, upper-middle, and low-income economies but decreases CO₂ emissions in high-income economies. After conducting a classification study on 182 countries, Wang and Zhang [25] found that the increase in trade openness has reduced CO₂ emissions in high-income and upper-middle-income countries but has no significant impact on CO₂ emissions in lower- middle-income countries and has increased CO₂ emissions in low-income countries. Lv and Xu [26] used PMG to study the impact in 55 middle-income countries and confirmed that enhanced trade openness decreases CO₂ emissions in the short run but increases them in the long run. Suleman et al. [27] divided 85 countries into high trade openness economies and low trade openness economies and found that trade openness had a positive and significant impact on high trade openness economies, while it has a negative impact on low trade openness economies.

Table 1. Summary of empirical studies.

Authors	Country/Regions	Period	Methodology	Results
Study based on time-series data
Jayanthakumaran et al. [13]	India and China	1971–2007	ARDL bounds test	None
Kohler [14]	South Africa	1960–2009	ARDL bounds test, VECM Granger Causality	Two-way
Tiwari et al. [12]	India	1966–2011	ARDL bounds test, VECM Granger Causality	Positive
Ertugrul et al. [28]	Top ten emitters among developing countries	1971–2011	ARDL bounds test, VECM Granger Causality	Positive
Mutascu [6]	France	1960–2013	Wavelet tool	Positive
Hdom and Fuinhas [10]	1975–2016	Brazil	FMOLS, DOLS	Positive
Ansari et al. [15]	top CO2 emitters	1971–2013	ARDL, VECM Granger causality	Uncertain
Suhrab et al. [29]	Pakistan	1985–2018	Cointegration analysis, Granger Causality	Positive
Udeagha and Ngepah [11]	South African	1960–2020	Novel dynamic ARDL simulation	Uncertain
Study based on panel data
Al-Mulali and Ozturk [18]	14 MENA countries	1996–2012	FMOLS, VECM Granger causality	Two-way
Jebli et al. [19]	25 OECD countries	1980–2010	DOLS, FMOLS, VECM Granger causality	Two-way
Ahmed et al. [16]	BRICS	1970–2013	FMOLS, VECM Granger causality	Positive
Destek et al. [20]	Ten CEEC countries	1991–2011	DOLS, FMOLS, VECM Granger causality	Two-way
Zhang et al. [17]	Ten newly industrialized countries	1971–2013	OLS, FMOLS, DOLS, Panel VECM Granger causality	Positive
Afridi et al. [30]	SAARC	1980–2016	OLS, GLS, panel Causality tests	Negative
Lv and Xu [25]	55 middle-income countries	1992–2012	PMG	Uncertain
Sun et al. [23]	49 high-emission countries in BRI	1991–2014	FMOLS, panel VECM Granger causality	Uncertain
Iqbal et al. [24]	Heterogeneous income groups	1971–2020	FMOLS, DOLS, system GMM	Uncertain
Chen et al. [21]	64 BRI countries	2001–2019	Panel quantile regression	Positive
Salam and Xu [31]	BRI countries	2001–2018	Two-step GMM	Uncertain
Chhabra et al. [32]	23 middle-income countries	1994–2018	GMM, Dumitrescu-Hurlin causality test	Positive
Azam et al. [33]	Six countries from the OPEC	1975–2018	OLS	Positive
Zheng et al. [22]	10 Asian countries	1995–2018	CS-ARDL, AMG, CCEMG	Positive
Wang and Zhang [25]	182 countries	1990–2015	FMOLS	Uncertain
Ashraf et al. [34]	75 BRI countries	1990–2019	Spatial panel data models	Positive
Pata et al. [35]	6 ASEAN countries	1995–2018	Panel ARDL, Dumitrescu-Hurlin causality test	Negative
Suleman et al. [27]	85 countries	1995–2020	Stepwise regression, FMOLS et al.	Uncertain
Pham and Nguyen [36]	64 developing countries	2003–2017	BMA	None

Table 1. Summary of empirical studies.

Authors	Country/Regions	Period	Methodology	Results
Study based on time-series data
Jayanthakumaran et al. [13]	India and China	1971–2007	ARDL bounds test	None
Kohler [14]	South Africa	1960–2009	ARDL bounds test, VECM Granger Causality	Two-way
Tiwari et al. [12]	India	1966–2011	ARDL bounds test, VECM Granger Causality	Positive
Ertugrul et al. [28]	Top ten emitters among developing countries	1971–2011	ARDL bounds test, VECM Granger Causality	Positive
Mutascu [6]	France	1960–2013	Wavelet tool	Positive
Hdom and Fuinhas [10]	1975–2016	Brazil	FMOLS, DOLS	Positive
Ansari et al. [15]	top CO2 emitters	1971–2013	ARDL, VECM Granger causality	Uncertain
Suhrab et al. [29]	Pakistan	1985–2018	Cointegration analysis, Granger Causality	Positive
Udeagha and Ngepah [11]	South African	1960–2020	Novel dynamic ARDL simulation	Uncertain
Study based on panel data
Al-Mulali and Ozturk [18]	14 MENA countries	1996–2012	FMOLS, VECM Granger causality	Two-way
Jebli et al. [19]	25 OECD countries	1980–2010	DOLS, FMOLS, VECM Granger causality	Two-way
Ahmed et al. [16]	BRICS	1970–2013	FMOLS, VECM Granger causality	Positive
Destek et al. [20]	Ten CEEC countries	1991–2011	DOLS, FMOLS, VECM Granger causality	Two-way
Zhang et al. [17]	Ten newly industrialized countries	1971–2013	OLS, FMOLS, DOLS, Panel VECM Granger causality	Positive
Afridi et al. [30]	SAARC	1980–2016	OLS, GLS, panel Causality tests	Negative
Lv and Xu [25]	55 middle-income countries	1992–2012	PMG	Uncertain
Sun et al. [23]	49 high-emission countries in BRI	1991–2014	FMOLS, panel VECM Granger causality	Uncertain
Iqbal et al. [24]	Heterogeneous income groups	1971–2020	FMOLS, DOLS, system GMM	Uncertain
Chen et al. [21]	64 BRI countries	2001–2019	Panel quantile regression	Positive
Salam and Xu [31]	BRI countries	2001–2018	Two-step GMM	Uncertain
Chhabra et al. [32]	23 middle-income countries	1994–2018	GMM, Dumitrescu-Hurlin causality test	Positive
Azam et al. [33]	Six countries from the OPEC	1975–2018	OLS	Positive
Zheng et al. [22]	10 Asian countries	1995–2018	CS-ARDL, AMG, CCEMG	Positive
Wang and Zhang [25]	182 countries	1990–2015	FMOLS	Uncertain
Ashraf et al. [34]	75 BRI countries	1990–2019	Spatial panel data models	Positive
Pata et al. [35]	6 ASEAN countries	1995–2018	Panel ARDL, Dumitrescu-Hurlin causality test	Negative
Suleman et al. [27]	85 countries	1995–2020	Stepwise regression, FMOLS et al.	Uncertain
Pham and Nguyen [36]	64 developing countries	2003–2017	BMA	None

Most studies, regardless of data types or estimation methods, hold that enhanced trade openness leads to increased CO₂ emissions, while some conclude that they have no nexus, or their nexus varies with conditions. With the development of econometric methods, the study of the nexus between them is further deepened, which has become an important part of climate change research. Compared with the existing studies, this study makes these possible marginal contributions: Firstly, the emerging countries with certain economic scale, constantly increasing market openness and obviously increasing CO₂ emissions, rather than one or a few emerging countries, were included as the research objects; Secondly, the impact in emerging countries was estimated by using FMOLS, DOLS, and PMG-ARDL methods, and methods such as DCCEMG and Driscoll–Kray were employed to enhance the robustness of the estimations, the Dumitrescu–Hurlin causality test [37] was carried out in emerging countries, revealing its causal pathways. Thirdly, the current outbreak of the Russia Ukraine and Middle East wars has exacerbated geopolitical risks and seriously affected the stability of global industrial and supply chains. Compared with developed countries, the economies of emerging market countries are more vulnerable to the shock of geopolitical risks. In this situation, exploring the impact of trade openness on carbon emissions is not limited to addressing climate change, it can also achieve the dual goals of maintaining economic growth and reducing fossil energy demand by adjusting trade structure. This study can provide empirical support for how to achieve green and stable economic growth in emerging market countries.

To achieve these possible marginal contributions, Figure 3 presents the framework for empirical research.

3. Research Methods and Data

When the panel data are nonstationary and have a long-time dimension or slope heterogeneity is present, methods such as fixed-effects, mixed regression, and GMM are not suitable. Therefore, if the panel data in this paper have the above characteristics, FMOLS, DOLS, and PMG-ARDL methods could be used to study the impact in emerging countries.

3.1. Research Method

3.1.1. Basic Model Setting

To detect the impact of trade openness on CO₂ emissions in emerging countries, a basic panel data model was established as follows:

$$ln{CE}_{it}=\alpha +{\beta }_{1}ln{EG}_{it}+{\beta }_{2}ln{EC}_{it}{+\beta }_{3}ln{TP}_{it}+{\epsilon }_{it}$$

(1)

where $lnCE$, $lnEG$, $lnEC$, and $lnTP$ are logarithms of CO₂ emissions, economic growth, energy consumption, and trade openness, respectively; $i$ represents country $\left(i=1, 2,\cdots ,N\right)$; $t$ is time $\left(t=1, 2,\cdots ,N\right)$; $\alpha$ is an intercept term; and $\epsilon$ is a residual term.

3.1.2. FMOLS, DOLS, and PMG-ARDL Methods

(1)

FMOLS method

Phillips and Hansen [38] developed the FMOLS method for estimating optimal cointegrating regression. Incorporating lagged independent variables and lagged error terms, the FMOLS estimation approach is designed to handle endogeneity issues. The FMOLS estimation method can also effectively handle non-stationary data, and its estimator exhibits asymptotic unbiasedness and consistency under the large sample. Nevertheless, FMOLS estimation results may be biased and lack asymptotic properties in a small sample.

For convenience in describing the FMOLS method, Model (1) was rewritten as:

$$Y_{it}={\alpha }_{it}+\beta X_{it}+{\epsilon }_{it}$$

(2)

$$X_{it}=X_{it-1}+{\mu }_{it}$$

(3)

where $Y_{it}$ is the explained variable $ln{CE}_{it}$, $X_{it}$ is the explanatory variable $X_{it}=\left(ln{EG}_{it},ln{EC}_{it},ln{TP}_{it}\right)$, and $\beta$ is the parameter vectors to be estimated. $X_{it}$ was assumed to be equal to its first-stage lag term $X_{it-1}$ plus the residual term ${\mu }_{it}$.

Let $Z_{it}={\left(Y_{it},X_{it}\right)}^{\prime }~I\left(0\right)$, $\varpi =\left({\epsilon }_{it},{\mu }_{it}\right)~I\left(0\right)$ with a long-run covariance matrix ${\mathsf{\Omega }}_i=L_i{L}_i^{\prime }$; $L_i$ is the lower triangular decomposition of ${\mathsf{\Omega }}_i$. ${\mathsf{\Omega }}_i$ can also be decomposed into the sum of ${\mathsf{\Omega }}_i^0$, ${\mathsf{\Gamma }}_i$, and ${\mathsf{\Gamma }}_i^{\prime }$; ${\mathsf{\Omega }}_i^0$ represents the contemporaneous covariance; and ${\mathsf{\Gamma }}_i$ is the weighted sum of autocovariances. The panel FMOLS estimator coefficient $\beta$ is as follows:

$${\beta }_{NT}^{*}=N^{-1}\sum _{i=1}^N{\left(\sum _{i=1}^T{\left(X_{it}-{\overline{X}}_i\right)}^2\right)}^{-1}\left(\sum _{i=1}^T\left(X_{it}-{\overline{X}}_i\right)Y_{it}^{*}-T{\overline{\tau }}_i\right)$$

where $Y_{it}^{\mathrm{*}}=\left(Y_{it}-{\overline{Y}}_i\right)-{\displaystyle \frac{{\widehat{L}}_{21i}}{{\widehat{L}}_{22i}}}\Delta X_{it}$, ${\widehat{\tau }}_i={\widehat{\mathsf{\Gamma }}}_{21i}+{\widehat{\mathsf{\Omega }}}_{21i}^0-{\displaystyle \frac{{\widehat{L}}_{21i}}{{\widehat{L}}_{22i}}}\left({\widehat{\mathsf{\Gamma }}}_{22i}+{\widehat{\mathsf{\Omega }}}_{22i}^0\right)$, and ${\overline{X}}_i$ represents the average value of the $i$-th variable.

(2)

Panel DOLS method

Like the FMOLS method, the DOLS estimation approach can tackle endogeneity in panel data models. Yet, the DOLS method proves to be more efficient with small sample sizes compared to FMOLS. However, the DOLS estimation results are greatly affected by the lag order and outliers. The DOLS estimation method assumes that a dynamic linear relationship is present between the residual term in Model (2) and the first-order difference of the explained variable. Then, the residual term in Model (2) was written as:

$${\epsilon }_{it}=\sum _{j=-q}^{j=q}{\theta }_{ij}\Delta X_{it-j}+{\epsilon }_{it}^{*}$$

(4)

where ${\theta }_{ij}$ is the parameter to be estimated, $j$ indicates the time period change $\left(j=-q,-q+1,\cdots ,q-1,q\right)$, and $q$ is the upper limit of $j$. Equation (4) was substituted into Model (2) to obtain the panel DOLS estimation Model (5).

$$Y_{it}={\alpha }_{it}+\beta X_{it}+{\sum }_{j=-q}^{j=q}{\theta }_{ij}\Delta X_{it-j} +{\epsilon }_{it}^{*}$$

(5)

The estimator coefficient $\beta$ in Model (5) is:

$${\beta }_{dols}^{*}=N^{-1}\sum _{i=1}^N{\left(\sum _{t=1}^T{Z}_{it}{Z}_{it}^{\prime }\right)}^{-1}\left(\sum _{t=1}^T{Z}_{it}{Y}_{it}^{*}\right)$$

where $Z_{it}=\left(X_{it}-{\overline{X}}_i,\Delta X_{it-j},\cdots ,\Delta X_{it+j}\right)$.

(3)

Panel PMG-ARDL method

The general form of the panel ARDL model is usually as follows:

$$Y_{it}=\sum _{j=1}^p{\delta }_{ij}{Y}_{i,t-j}+\sum _{j=0}^q{\gamma }_{ij}^{\prime }{X}_{i,t-j}+{\phi }_i+{\epsilon }_{it}$$

(6)

In Model (6), ${\delta }_{ij}$ is the coefficient for the lag term of the explained variable, ${\gamma }_{ij}$ is the parameter to be estimated for the explained variable and its lag term, and ${\phi }_i$ is the individual effect. If the variable is first-order integrated $I\left(1\right)$ and cointegrated, its perturbation term is an $I\left(0\right)$ process. $p$ and $q$ represent the number of lag phases.

$$\Delta Y_{it}={\theta }_i\left[Y_{i,t-1}-{\lambda }_i^{\prime }{X}_{i,t}\right]+\sum _{j=1}^{p-1}{\varphi }_{ij}\Delta Y_{i,t-j}+\sum _{j=0}^{q-1}{\psi }_{ij}^{\prime }\Delta X_{i,t-j}+{\phi }_i+{\epsilon }_{it}$$

(7)

where ${\theta }_i=-\left(1-{\delta }_i\right)$ is the coefficient of velocity adjustment, ${\lambda }_i$ is the long-run relation vector, $ECT=Y_{i,t-1}-{\lambda }_i^{\prime }{X}_{i,t}$ is the error correction term, and ${\varphi }_{ij}$, ${\psi }_{ij}$ are the short-run dynamic coefficients. ${\theta }_i$ is the coefficient of the error correction term, which estimates the adjustment rate of the explained variable towards long-run equilibrium. ${\theta }_i<0$ ensures the existence of a long-run relation. The higher the absolute value of ${\theta }_i$ is, the higher the convergence rate towards long-run equilibrium will be. However, ${\theta }_i\ge 0$ indicates no stable correlation between variables in the long run. Therefore, adjusting the velocity parameter ${\theta }_i$ and long-run coefficient ${\lambda }_i$ is critical for model estimation.

Since bias may occur between the white noise term and the mean differenced independent variables in ARDL models, the traditional ARDL method cannot control for this bias when individual effects are present in panel dataset models. For this reason, this study applies the PMG-ARDL technique [39] to resolve the problem. The method can not only estimate both long-term and short-term parameters at the same time, but it also can effectively handle heterogeneity and mitigate cross-sectional dependence. Nevertheless, the model relies on the assumption of long-term homogeneity, which may not align with reality in some cases, and the setting of the lag order depends on the researcher’s experience or trial.

3.2. Data and Basic Statistics

3.2.1. Data Description

Emerging countries have different classifications across studies, based on the IMF classification of emerging countries (

Table 2. List of emerging countries by IMF.

America (6)	Europe (10)	Africa (1)	Asia (12)
Argentina Brazil Chile Colombia Mexico Peru	Bulgaria Czech Republic Greece Hungary Poland Romania Russian Federation Slovenia Turkey Ukraine	Egypt	Bangladesh China India Indonesia Israel Iran South Korea Malaysia Pakistan Philippines Thailand Vietnam

). In this study, the panel data from 29 emerging countries in 2003–2020 were used to explore the impact. CO₂ emissions (CE), economic growth (EG), and energy consumption (EC) were described by CO₂ emissions per capita, GDP per capita, and energy consumption per capita, respectively. Trade openness (TP) was measured by the ratio of total imports and exports to GDP. Specific descriptions and data sources are displayed in

Table 3. Explanations and sources of variables.

Variables	Explanations	Unit	Sources
$lnCE$	Logarithm of CO2 emission per capita	Metric tons	WDI
$lnEG$	Logarithm of GDP per capita	Constant 2015 US dollar	WIT
$lnEC$	Logarithm of energy consumption per capita	MBTU	EIA
$lnTP$	Logarithm of trade openness	Percent	WDI

.

3.2.2. Basic Statistics

To detect the variability and distribution characteristics of variables, the basic statistical descriptions of CO₂ emissions per capita, GDP per capita, energy consumption per capita, and trade openness are displayed in

Table 4. Descriptive statistics.

	CE	EG	EC	TP
Mean	4.834	9350.804	82.840	75.296
Median	4.190	7627.160	77.018	60.778
Min	0.210	698.732	4.742	22.106
Max	12.217	39591.40	240.364	210.374
Sd	3.213	7915.109	55.811	40.995
Skewness	0.531	1.518	0.744	0.933
Kurtosis	2.331	5.224	3.134	2.904
Jarque-Bera	34.303 ***	308.134 ***	48.511 ***	75.939 ***
Prob	0.000	0.000	0.000	0.000
Observations	522	522	522	522

Notes: *** represents 1% level of significance.

. On average, CO₂ emissions per capita, GDP per capita, energy consumption per capita, and trade openness of emerging countries were 4.834 metric tons, 9350.804 US dollars, 82.840 MBTU, and 75.296%, respectively. From Table 4, they varied widely among emerging countries. From the perspectives of skewness and kurtosis, they all presented positive skewness and high kurtosis. In terms of the normal distribution, the P-values for the Jarque-Bera tests were all zero, suggesting that the null hypothesis with a normally distributed series was rejected at the 5% significance level, i.e., the variable did not satisfy the normal distribution. This result may be attributed to the cross-sectional heterogeneity in the data used in this study, so cross-sectional heterogeneity should be corrected in the panel data model estimation.

According to

Table 5. Correlation matrix.

	$\mathit{l}\mathit{n}\mathit{C}\mathit{E}$	$\mathit{l}\mathit{n}\mathit{E}\mathit{G}$	$\mathit{l}\mathit{n}\mathit{E}\mathit{C}$	$\mathit{l}\mathit{n}\mathit{T}\mathit{P}$
$lnCE$	1 — —
$lnEG$	0.791 *** (0.000)	1 — —
$lnEC$	0.977 *** (0.000)	0.825 *** (0.000)	1 — —
$lnTP$	0.460 *** (0.000)	0.294 *** (0.000)	0.415 *** (0.000)	1 — —

Notes: *** represents 1% level of significance.

, which provides the correlation matrix for CO₂ emissions, economic growth, energy consumption, and trade openness, the correlation between energy consumption and economic growth is 0.825—much higher than 0.8. This indicates a possible multicollinearity problem between these variables.

To further explore whether multicollinearity exists among the explanatory variables,

Table 6. Variance inflation factors (VIF) of explanatory variables.

	lnEG	lnEC	lnTP
VIF	3.16	3.49	1.22
1/VIF	0.316	0.287	0.821

provides the Variance Inflation Factors (VIF). The highest VIF value observed is 3.49, which is significantly lower than 10 (a threshold commonly used to indicate serious multicollinearity). Consequently, considering VIF and the correlation coefficients between the variables, multicollinearity is not a major concern in our linear model estimations.

4. Unit Root Test and Cointegration Test

4.1. Slope Heterogeneity Test

Panel data models have been used in many empirical studies. Standard panel data regression models, all assume homogeneity of major parameters. Ignoring slope heterogeneity can bias estimates, slope homogeneity needs testing prior to model estimation. Pesaran and Yamagata [40] proposed a slope heterogeneity test method for panel data models under large N (cross-sections) and large T (time periods) cases based on Swamy’s test: assuming that the perturbation term is independently distributed and allowed to have heteroscedasticity, and its statistic is written as $\widehat{\Delta }$. If the error satisfies normal distribution, the mean-variance bias-adjusted statistic is written as ${\widehat{\Delta }}_{adj}.$ The null hypothesis for the slope heterogeneity test was that the slope coefficients were all the same. As in

Table 7. Slope heterogeneity test.

Homogenous/Heterogenous Slope Coefficient Testing
Test	Statistic	p-Value
$\widehat{\Delta }$	12.507 ***	0.000
${\widehat{\Delta }}_{adj}$	14.716 ***	0.000

Note: *** represents 1% level of significance. Same as below.

, the p-value for the slope heterogeneity test was less than 1%, so the null hypothesis was rejected, suggesting the presence of slope heterogeneity.

4.2. Cross-Section Dependence Test

No cross-section dependence (CD) is usually assumed by the standard panel data models, but the links of different countries are generally strengthened because of the impact of globalization, so the assumption of independence between different cross-sections in panel data may become invalid. Neglecting CD will lead to erroneous inference [41]. According to the differences in the cross-section dimension and the time dimension of panel data, the CD test [42] and bias-adjusted Lagrange multiplier (LM) test [43] are adopted. When the cross-section dimension is larger than the time dimension in the sample, i.e., N > T, the CD test is used. The null hypothesis for the CD test is that no dependence is present between cross-section units. If the p-value for the CD test is less than 5%, the null hypothesis is rejected. As in

Table 8. Results of Pesaran CD test.

Variable	CD Test	p-Value	Corr	Abs (Corr)
$lnCE$	10.40 ***	0.000	0.122	0.625
$lnEG$	63.00 ***	0.000	0.737	0.827
$lnEC$	18.03 ***	0.000	0.211	0.622
$lnTP$	7.65 ***	0.000	0.089	0.467

Note: *** represents 1% level of significance.

, the null hypothesis was rejected, i.e., CD was present among CO₂ emissions, economic growth, energy consumption, and trade openness.

4.3. Panel Unit Root Test

First- and second-generation panel unit root tests have been developed. The first-generation panel unit root test assumes no CD and slope heterogeneity, but the classical methods for the first-generation panel unit root test, such as the Breitung test, ADF-Fisher test, and PP-Fisher test, can cause significant bias if CD and slope heterogeneity are present [44,45]. Phillips and Sul [46], Moon and Perron [47], and Pesaran [48] proposed the second-generation panel unit root test for panel data with CD and slope heterogeneity. The results of the slope heterogeneity test and CD test revealed that the samples involved had slope heterogeneity and CD, so Pesaran’s [48] cross-sectional augmented IPS (CIPS) and covariate-augmented Dickey–Fuller (CADF) tests were applied to test the stationarity. The results of the CADF and CIPS tests are shown in

Table 9. Results of CADF Test.

Variables	Level Form			1st Difference
	T-Bar	Z [t-Bar]	p-Value	T-Bar	Z [t-Bar]	p-Value
$lnCE$	−1.982	−1.279	0.100	−2.269 ***	−2.792	0.003
$lnEG$	−1.421	1.686	0.954	−2.234 ***	−2.608	0.005
$lnEC$	−1.892	−0.801	0.212	−2.270 ***	−2.798	0.003
$lnTP$	−1.169	3.013	0.999	−2.374 ***	−3.347	0.000

Note: *** represents 1% level of significance.

and

Table 10. Results of CIPS Test.

Variables	Level Form		1st Difference
	Statistic	Critical Value (1%)	Statistic	Critical Value (1%)
$lnCE$	−1.737	−2.32	−3.276	−2.32
$lnEG$	−0.956	−1.76	−2.336	−1.76
$lnEC$	−1.868	−2.32	−3.465	−2.32
$lnTP$	−1.249	−2.32	−3.243	−2.32

, respectively. It could be found that when the variables were in a level form, the null hypothesis could not be rejected, indicating that these variables are not stationary in level form. However, the null hypothesis was significantly rejected when the first-order difference was taken for these variables, indicating that these variables are stationary following the first-order difference.

4.4. Results of Cointegration Test

Considering the presence of first-order unit roots, the long-run equilibrium relationship among them could be described by a cointegration test using the methods of Pedroni [49], Kao [50], and Westerlund [51] (see results in

Table 11. Pedroni’s cointegration test.

Within-Dimension
	Statistic	p-Value	W. Statistic	p-Value
Panel v-Statistic	1.968 **	0.025	2.021 **	0.022
Panel rho-Statistic	−0.718	0.237	−0.983	0.163
Panel PP-Statistic	−3.351 ***	0.000	−3.659 ***	0.000
Panel ADF-Statistic	−2.247 **	0.012	−2.003 **	0.023
Between-Dimension
	Statistic	p-Value
Group rho-Statistic	0.883	0.811
Group PP-Statistic	−4.380 ***	0.000
Group ADF-Statistic	−2.614 ***	0.005

Note: ***, ** represent 1% and 5% levels of significance, respectively.

,

Table 12. Kao’s cointegration test.

	t-Statistic	p-Value
ADF	−4.111 ***	0.000

Note: *** represents 1% level of significance.

and

Table 13. Westerlund’s cointegration test.

	Statistic	p-Value
Variance ratio	−1.843 ***	0.033

Note: *** represents 1% level of significance.

). It was found that emerging countries had cointegration, indicating a consistent long-run tendency.

5. Results of Estimation

5.1. Panel FMOLS and DOLS Estimates

Since CO₂ emissions, economic growth, energy consumption, and trade openness had cointegration, the impacts of economic growth, energy consumption, and trade openness on CO₂ emissions were estimated using mixed OLS, FMOLS, and DOLS methods, respectively (

Table 14. Panel long-run estimates.

Variables	FMOLS		DOLS
	Coefficient	p-Value	Coefficient	p-Value
$lnEG$	−0.153 ***	0.000	−0.197 **	0.010
$lnEC$	1.232 ***	0.000	1.251 ***	0.000
$lnTP$	−0.060 **	0.037	−0.112 **	0.036
Adjusted R-squared	0.996		0.997

Note: ***, ** represent 1% and 5% levels of significance, respectively.

). Both the FMOLS and DOLS estimation results reveal that economic growth and trade openness have a significant negative correlation with CO₂ emissions, whereas energy consumption shows a significant positive correlation with CO₂ emissions. In addition, the coefficients estimated by the DOLS method have a greater impact on CO₂ emissions than those estimated by the FMOLS method.

5.2. PMG-ARDL Estimate

The impacts on CO₂ emissions were estimated using the MG, PMG, and dynamic fixed-effects (DFE) methods (

Table 15. Panel ARDL model from the MG, PMG and DFE estimates.

	(1) MG	(2) PMG	(3) DFE
Long-term
$lnEG$	0.632 (0.223)	−0.060 ** (0.022)	−0.168 (0.107)
$lnEC$	0.699 * (0.083)	1.026 *** (0.000)	1.250 *** (0.000)
$lnTP$	0.182 (0.164)	−0.149 *** (0.000)	−0.056 (0.534)
Short-term
ECT	−0.583 *** (0.000)	−0.295 *** (0.000)	−0.192 *** (0.000)
$\Delta lnEG$	0.284 ** (0.019)	0.405 *** (0.001)	0.311 *** (0.000)
$\Delta lnEC$	0.165 (0.136)	0.432 *** (0.000)	0.482 *** (0.000)
$\Delta lnTP$	−0.014 (0.704)	0.036 (0.194)	−0.002 (0.943)
Constant	−0.193 (0.643)	0.038* (0.055)	−0.129 (0.173)

Note: ***, **, * represent 1%, 5%, 10% levels of significance, respectively.

). To identify the optimal estimation method among MG, PMG, and DFE, the Hausman test (

Table 16. Hausman test.

	Chi-Squared Statistic	p-Value
MG	3.12	0.373
PMG
MG	0.00	1.000
DFE

) was carried out. The results showed that the optimal results were obtained by PMG, followed by MG and DFE, so PMG could be used as the most suitable estimation method. In the long run, economic growth, and trade openness significantly reduced CO₂ emissions, while energy consumption significantly increased them. Specifically, a 1% increase in economic growth led to a 0.06% decrease in CO₂ emissions, a 1% increase in energy consumption resulted in a 1.026% increase in CO₂ emissions, and a 1% increase in trade openness caused a 0.149% decrease in CO₂ emissions. In the short term, economic growth and energy consumption significantly contribute to higher CO₂ emissions, and a 1% boost in economic growth sparked a 0.405% surge in CO₂ emissions, while a 1% jump in energy consumption ignites a 0.432% rise in CO₂ emissions, whereas trade openness has no significant effect. ECM was negative in the 1% interval, suggesting the ability to adjust from disequilibrium to equilibrium.

5.3. Further Estimation: Enhancing Robustness

Despite the PMG-ARDL model’s consideration of cross-sectional dependence, we utilized additional methods to bolster the robustness of our findings. The DCCEMG and Driscoll–Kray approaches were applied to assess the effects (see the

Table 17. Panel estimates from DCCEMG and Driscoll–Kray.

	DCCEMG	Driscoll–Kray
$lnEG$	−0.190 ** (0.042)	−0.129 ** (0.01)
$lnEC$	0.861 *** (0.000)	1.177 *** (0.000)
$lnTP$	−0.052 ** (0.022)	−0.058 *** (0.003)
$L.lnCE$	−0.714 *** (0.000)	_______
Constant	_______	−0.485 *** (0.000)

Note: ***, ** represent 1% and 5% levels of significance, respectively.

). Notably, the outcomes align closely with the results derived from the FMOLS, DOLS, and PMG-ARDL estimations.

5.4. Analysis and Discussion of Estimation Results

For the convenience of observing the results estimated by methods such as FMOLS, DOLS, PMG-ARDL, DCCEMG, and Driscoll–Kray, Figure 4 presents the long-term and short-term impacts of economic growth, energy consumption, and trade openness on CO₂ emissions, respectively.

In the long run, the increase in economic growth and trade growth has reduced the growth of CO₂ emissions, mainly for three possible reasons as follows: firstly, the expansion of the service sector and service trade. The economic growth of emerging market countries is not solely driven by industrialization. Increasingly, the development of the service sector is becoming a significant contributor to their economic expansion. A similar pattern is observed in foreign trade growth, where service trade is also emerging as a key driver for the trade development of these countries. Given that the service sector typically generates lower carbon emissions compared to the industrial sector, the development of the service industry can not only promote economic growth and enhance trade openness but also improve total-factor carbon emission efficiency [52]. Consequently, this relatively reduces the growth of CO₂ emissions per capita. For example, India is vigorously developing service trade represented by software service outsourcing, and the total export value of service trade has exceeded the total trade value of goods, while promoted economic growth and controlling the growth of CO₂ emissions.

Secondly, a possible reason is the technological upgrading in industrial transfer and the global diffusion of green technologies. In the early stages, developed nations transferred numerous high-polluting and high-emission industries to emerging market countries. However, as technology has advanced, the technological content of industries being transferred from developed to emerging market countries has also increased significantly. This transfer has demonstrated clear spillover effects, leading to improvements in labor productivity within emerging markets. Over the past four decades of industrialization, China initially accepted many industries transferred from developed countries, which not only led to serious environmental issues but also caused a rapid increase in CO₂ emissions. However, with China’s technological progress and industrial structure upgrading, this problem has been effectively controlled. Furthermore, environmental regulations in developed countries encourage green innovation among businesses. These green technologies are then disseminated globally through investment and trade, thereby accelerating the green and low-carbon transformation of emerging market economies. Consequently, technological factors play a dual role: they not only stimulate economic growth and expand the scale of trade but also contribute to CO₂ emissions reduction.

Thirdly, a possible reason is the actions of emerging market countries in addressing climate change. Even though the Kyoto Protocol did not impose emission reduction obligations on emerging market countries, they have not prioritized economic growth over addressing greenhouse gas emissions. Instead, emerging markets have proactively implemented measures to curb the rise in carbon emissions. To illustrate, China has established clear objectives to achieve “carbon peaking” by 2030 and “carbon neutrality” by 2060, while actively promoting renewable energy as a means of energy transition. Similarly, India has launched a renewable energy expansion initiative aimed at reaching 450 gigawatts of renewable energy capacity by 2030. The efforts of these emerging market nations to combat climate change are helping to lower carbon emissions in production processes, thus making it feasible for economic growth, trade openness, and CO₂ emissions to exhibit an inverse relationship.

However, given that traditional energy sources play an indispensable role in both industrial activities and everyday life, the growth of industrial scale and rising personal income levels will naturally result in greater consumption of fossil fuels, leading to increased CO₂ emissions. In the short term, since adjustments to economic structure, energy structure, and technological advancements all take time, these factors remain constant in the short run. As a result, economic growth and higher energy usage in the short term will cause CO₂ emissions to rise. Moreover, within a short timeframe, aspects such as the scale of trade, trade structure, and production standards in international trade are also fixed. Therefore, a temporary increase in trade openness may not necessarily bring about changes in CO₂ emissions.

6. Causality Test

The Granger causality test is usually applied to determine the presence or absence of causality between the current value of the dependent variable and the lagged value of the explanatory variable. Hoaltz–Eakin et al. [53] investigated the causality in panel data under the homogenous assumption but ignored individual heterogeneity. Dumitrescu and Hurlin [37] introduced heterogeneity into the causality test of panel data, which has relatively low requirements for sample size, and provides robust test results. However, it is sensitive to the specification of lag order.

The testing procedure is as follows: Firstly, estimate the individual-specific coefficients in Equation (6) using time series data to obtain the estimated values ${\widehat{\delta }}_{ij}$ and ${\widehat{\gamma }}_{ij}$. Secondly, based on the estimated individual coefficients, construct the Wald test statistic $W_{N,T}$ to test for causality between variables. Finally, compare the calculated test statistic $W_{N,T}$ with the critical value. If $W_{N,T}$ is greater than the critical value, reject the null hypothesis of no causality and conclude that there is a causal relationship from $X$ to $Y$.

To detect causality, the Dumitrescu–Hurlin causality test was carried out (Table 14). Firstly, two-way Granger causality existed between them (see

Table 18. Pairwise Dumitrescu–Hurlin panel causality tests.

Null Hypothesis	W-Stat.	Zbar-Stat	p-Value
$lnEG\to lnCE$	4.035 ***	2.795	0.005
$lnCE\to lnEG$	4.124 ***	2.952	0.003
$lnEC\to lnCE$	3.679 **	2.170	0.030
$lnCE\to lnEC$	4.553 ***	3.706	0.000
$lnTP\to lnCE$	5.508 ***	5.384	0.000
$lnCE\to lnTP$	4.461 ***	3.544	0.000
$lnEC\to lnEG$	2.615	0.299	0.765
$lnEG\to lnEC$	3.123	1.192	0.233
$lnTP\to lnEG$	2.813	0.648	0.517
$lnEG\to lnTP$	7.768 ***	9.355	0.000
$lnTP\to lnEC$	4.796 ***	4.133	0.000
$lnEC\to lnTP$	7.752 ***	9.327	0.000

Notes: ***, ** represent 1% and 5% levels of significance, respectively. “A→B” means A does not homogeneously cause B.

). Secondly, no Granger causality was found between economic growth and energy consumption. Finally, economic growth and trade openness had one-way causality, and economic growth was the Granger causality of trade openness, but trade openness was not the Granger causality of economic growth. To illustrate the causal pathways more clearly through which trade openness influences CO₂ emissions, Figure 5 shows that trade openness also indirectly influences CO₂ emissions through its direct effects on economic growth and energy consumption in addition to the direct impact on CO₂ emissions.

7. Conclusions, Policy Implications, and Limitations

7.1. Conclusions and Policy Implications

To explore the impact of trade openness on CO₂ emissions in emerging countries, the FMOLS, DOLS, and PMG-ARDL methods were used, based on which the Dumitrescu–Hurlin causality test was carried out. Hence, the following conclusions have emerged: Firstly, over the long term, economic growth and enhanced trade openness are associated with reduced CO₂ emissions, whereas increased energy consumption results in higher CO₂ emissions. Secondly, in the short term, trade openness does not significantly impact CO₂ emissions, while economic growth and energy consumption are key factors driving the increase in CO₂ emissions. Thirdly, a two-way causal relationship exists between trade openness and CO₂ emissions, with economic growth and energy consumption being capable of influencing CO₂ emissions by modifying trade openness. Considering the relationship between them, along with the potential reasons for their interplay, the following policy implications can be considered.

Firstly, focus on the expansion of the service trade. The growth of the service trade is heavily reliant on robust institutional frameworks. Emerging market countries should make it a priority to create systems that promote the liberalization and facilitation of service trade. This involves aligning service trade rules and enhancing regulatory coordination. By referencing the World Trade Organization’s “Reference Paper on Domestic Regulation in Services Trade”, these countries can standardize licensing, qualifications, and technical standards within the service sector. Simplifying the procedures for obtaining licenses and approvals, as well as increasing the transparency of regulatory policies, are also crucial steps. Furthermore, the implementation of a negative list system for cross-border service trade is recommended, with the need to update any existing laws and regulations that conflict with this system. Additionally, emerging countries should strengthen the information infrastructure construction, promote the application of new technologies such as cloud computing and big data in the service trade, improve transportation and logistics networks to enhance logistics efficiency, encourage innovation in the service trade, and promote the integration of the service trade with manufacturing [54,55].

Secondly, encourage the modernization of traditional industries and increase collaboration on green technology internationally. In response to the latest technological advancements, emerging market countries should leverage digital and smart technologies to upgrade traditional industries. This includes optimizing production processes and enhancing the efficiency of resource use. Emerging countries utilize tax incentives, tax subsidies, and preferential loan rates to support emerging industries and technological innovation. They increase investment in scientific and technological research and development, establish platforms for the transformation of scientific and technological achievements, and promote technology transactions and cooperation. They expand government-to-government cooperation in green technologies with developed countries, learning from and drawing on the policies, regulations, and standards of developed countries in the field of green technologies. They also promote green technology cooperation among enterprises, introducing green technologies from developed country enterprises through technology transfer, technology licensing, and other means, and carrying out joint research and development projects with enterprises from developed countries to jointly tackle challenges in the field of green technologies.

Thirdly, improve environmental management to curb the rapid increase in CO₂ emissions. Compared to the European Union Emissions Trading System (EU ETS), emerging market countries generally face challenges such as insufficient technical capacity, lack of funding, inadequate market mechanisms, and low participation of market entities. Moreover, the carbon emissions trading system is widely considered an effective policy tool. Emerging countries should take inspiration from the EU-ETS and adopt a phased and gradual approach to establish ETS that suits their specific national contexts. Quotas should shift from being mainly free to being mainly auctioned, the coverage of industries should gradually expand, and emission reduction targets should increase step by step. These systems can harness market mechanisms to drive carbon reduction efforts. It is important to align fiscal and financial policies to support these initiatives, offering financial incentives for renewable energy generation and promoting the adoption of new energy vehicles. Financial institutions should be encouraged to provide preferential loans to eligible green projects.

7.2. Limitations

The impact of trade openness on CO₂ emissions in emerging countries was studied through various methods and enlightening conclusions were drawn in this paper, but deficiencies were still exposed: First, emerging countries were used as a whole sample, and they were not classified according to some characteristics, such as BRICS countries, East Asian countries, and Eastern European countries, and due to the availability of data and the significant structural changes caused by COVID-19, this study did not include trends beyond 2020. Further, when dealing with structural breaks in panel data, two general approaches are adopted. One is to introduce panel data model estimation methods that can handle structural breaks [56,57], and the other is to utilize the event study method [58]. These two approaches facilitate further exploration of the impact of trade openness on CO₂ emissions under the influence of sudden events such as COVID-19. Second, CO₂ emissions were described by CO₂ emissions per capita, which objectively reflected the status of CO₂ emissions in emerging countries. However, the growth rate of total CO₂ emissions may be higher than that of carbon emissions per capita due to the rapid growth of populations in emerging countries, so the impact of trade openness on CO₂ emissions may be underestimated. Third, the nexus between trade openness and CO₂ emissions involves not only the division of labor in global industrial chains but also the specific resource endowments and economic development models of each country, so the mechanism between the two remains to be deeply explored. Therefore, further attention and research are required in the future to understand the nexus more deeply between trade and CO₂ emissions.