Trade Openness and Environmental Emissions: Evidence from a Meta-Analysis

Abstract

How trade affects environmental emissions has generated heterogeneous results over the years. This is due to empirical ambiguities that are endemic in the literature. In order to evaluate and explain the discrepancy in the literature, this paper conducts a meta-analysis of 88 empirical studies published until 2018. Our results show that trade contributes to environmental emissions after controlling for publication bias and heterogeneity. In explaining the heterogeneous results across the primary studies, our findings largely suggest the estimated elasticities depend systematically on the estimation characteristics, the choice of pollutants and the publication characteristics of the primary studies. Accounting for heterogeneity, the result remains robust only for CO₂ emissions compared to SO₂. Overall, the trade elasticity of emissions effect remains robust when we decompose the analyses for different groups of countries, however, the emissions-content of trade is more pronounced for developed compared to developing countries.

1 Introduction

The trade-environmental literature has generated intense political and economic debates over the years mainly because of the contradictory results, especially in the empirical literature (van Bergeijk 1991). The debate is about whether the degree of exposure of a country in global trade or trade openness affects the level of environmental emissions. Not surprising, the debate has generated several empirical studies (see, e.g., Copeland and Taylor 2003; Rose and Stanley 2005; Managi et al. 2009; Shahbaz et al. 2017a). However, there is still ambiguity in the results because of the endemic heterogeneity in the empirical studies. The relationship between trade openness (hereafter trade) and environmental emissions is a sensitive global concern according to van Bergeijk (1991). This is due to the anecdotal evidence that shows a simultaneous surge in trade and environmental emissions globally.¹ These heightened global concerns and the anecdotal evidence have generated a keen research interest both from theoretical and empirical perspectives.

The traditional trade theories do not consider the welfare losses if international trade results in environmental emissions (hereafter emissions) (van Bergeijk 1991). However, Antweiler et al. (2001) provide a plausible mechanism through which trade openness can partially improve environmental quality. Through the positive impact of trade on income, in line with the popular evidence provided by Frankel and Romer (1999), Antweiler et al. (2001) argue a positive consequence of trade on the environment. This argument is premised on the basis that there is a higher tendency for the populace to demand more environmentally friendly (eco) products as their income rises. This is mainly because the environment is considered a normal good (Copeland and Taylor 2003). In contrast, environmentalists and economists alike have also argued that increased global trade generates increased production through economies of scale.² The growth in production directly contributes to over exploitation of the environment through increased production and consumption of pollution-intensive goods (Aklin 2016). The main argument is premised on the fact that the free market economy lacks the necessary impetus to efficiently allocate these environmental resources in a way that ensures the negative externalities are internalized.

Increasingly, many studies have also provided empirical evidence supporting either side of the theoretical positions. However, available empirical literature has deepened this uncertainty about the effect of trade on the environment. The uncertainty about the trade-environment nexus was acknowledged by van Bergeijk (1991), while a number of recent studies such as Rose and Stanley (2005), Managi et al. (2009) and Li et al. (2015) also reiterate it. These studies acknowledge the rapidly growing literature, as well as the contradictory results endemic in the empirical literature. Reviewing the past and recent literature confirms the uncertainty of these results. This emphasizes the fact that the trade-emissions debate has been lingering on for a long time without a consensus. Figure 1 highlights the discrepancy that is endemic in the 88 empirical studies used in our meta-analysis. The figure shows a 50-50 split between the results that the trade elasticity of emissions can either be positive or negative. The discrepancy also extends to the statistical significance of the effect as we see 55% of the studies find significant trade-emissions elasticity compared to 45% that find insignificant results.

Fig. 1

The trade-emissions elasticity reported in the studies published in 1991-2018 (N=789). Data Source: Based on studies for the meta-analysis

The opposing evidence in the empirical literature makes it necessary to re-visit the subject to critically assess the state of knowledge. From the political economy perspective, this may have wider implications for many governments because the trade-environment nexus is a sensitive and policy-relevant concern. In response to this uncertainty, the main objective of this paper is to use the tool of meta-analysis to systematically determine whether the increasing level of trade openness is good or bad for the environment, especially after accounting for publication bias and heterogeneity.

Specifically, this paper has two main objectives. First, to estimate the average effect size or magnitude of the trade-emissions elasticity which would provide the underlying impact of trade on emissions. This is important because policy-makers are particularly interested in knowing how changes in trade openness would affect emissions. Therefore, it is imperative to identify the economic and statistical significance of these elasticities (Stanley and Doucouliagos 2012). For instance, Copeland and Taylor (2003) estimate that if trade liberalization increases economic activity by 1%, pollution concentrations would increase within the range of 0.25–0.5%, while an increase in income per capita decreases pollution concentration by 1.25–1.5%. Using the tool of meta-analysis, we can derive the underlying trade-environment elasticities.

Second, the paper seeks to explain the heterogeneity of the results in previously published empirical studies. Through a multivariate meta-regression analysis (MRA), we provide a quantitative assessment of the factors that can explain the heterogeneity of these studies. Thus, we can understand how the different characteristics of the studies such as study design, environmental indicators, sample of countries, econometric methods, and time period influence the results both qualitatively (sign of estimated coefficients) and quantitatively (size and significance of coefficients).

The next section of the paper provides a theoretical note and a brief empirical literature review on the relationship between trade and the environment. Section 3 discusses the data and the empirical strategy adopted in accordance with the Meta-Analysis of Economics Research Network (MAER-Net) reporting guidelines. In Sect. 4, we provide and discuss the results. Section 5 concludes the paper and offers some policy implications and limitation of the study.

2 Relationship Between Trade and the Environment

2.1 A Theoretical Perspective

The basic theory of comparative advantage in international economics proposes countries specialize in the production of different products and then trade those products among themselves by transporting the goods from the point of production to consumption. Through the channels of production and transportation, these introduce a simple complementarity relationship between trade and the environment as increases in trade contribute to increases in emissions. However, Grossman and Krueger (1993) who were the first to postulate the trade-emissions nexus discard this simple linkage between trade and emissions as they analyze the environmental implications of the North America Free Trade Agreement (NAFTA). Following on the works of Grossman and Krueger (1993), Copeland and Taylor (2003) use a general equilibrium model that decomposes the net effect into three components; scale, technique and composition effects. In doing so, they take into consideration the possible exogenous and endogenous variations that are likely to exist between trade and the environment.

According to Grossman and Krueger (1993), the scale effect measures the increase in emissions as a result of an increase in the value of production in the economy, holding constant the mix of goods and production techniques. This scale effect is an adverse effect of trade on the environment as the scaling up of the endowments of the economy also expands the scale of production and emissions proportionately. The opening of trade could lead to a greater use of unemployed factors of production such as land and capital resources that may constitute environmental damage. The scale effect is aligned with the popular perception that more opened economies would adopt looser environmental standards for fear of losing competitiveness in international markets. This mechanism was confirmed by Copeland and Taylor (2003) in their pollution haven model of international trade in which they explain that trade may encourage pollution-intensive production processes to move from regions with high environmental standards to regions with low environmental standards.

A number of studies support this deleterious view of trade on the environment (see e.g., Managi et al. 2009; Shahbaz et al. 2017a; Li et al. 2015). They base their view on the theoretical premise that globalization leads to an increase in global demand, resulting in increased production that exploits the natural resources and pollutes the environment (van Bergeijk 1991). For instance, references are made to the Maquiladora zone established in Mexico with high concentration of polluting industries because of NAFTA, increased deforestation as result of timber trade, threats to endangered species due to international demand and low environmental quality in China as a result of export-led growth (Copeland and Taylor 2003; Aklin 2016) .

Conversely, through the technique effect, the increasing level of global trade may be benefical to the environment by curbing the the level of emissions. Grossman and Krueger (1993) identify two main channels through which this could occur. First, improvements in the production methods through increased access and the adoption of climate-friendly technologies. Firms that want to remain competitive in the global market would have incentives to develop and produce environmentally friendly goods. In addition, opening up for international trade could lead to a more efficient allocation of resources. This reduces the duplication of production processes in different locations, hence would reduce emissions. If every country has to produce to meet its domestic demand, this could result in duplication in the production process and therefore higher emissions. In line with the gains-from-trade hypothesis, trade can spur managerial and technical innovations and this can trigger the adoption of environmentally friendly technologies. Second, trade liberalization would also increase the availability of eco-friendly goods. Without international trade, consumers would have limited choices, and could be forced to purchase only domestic goods that may have been produced under lax environmental standards. In addition, Copeland and Taylor (2003) point to the fact the increasing levels of income that are as a result of the gain from trade may lead the general public to demand less pollution-intensive products thereby lowering emissions .

Also, the WTO through its green provisions has spurred enthusiasm among its members for environmental sustainable trade policies. Free trade agreements (FTAs), which are a major appendage of global trade, are also increasingly making an effort to make trade and environmental policies compatible. As countries seek to join FTAs, they are also made to simultaneously embrace environmental cooperation agreements. According to the joint report by the United Nations Environment Program (UNEP) and the WTO (2009), the European Communities (EC), Canada and the US are all ensuring that their trade and environmental policies are compatible as they require environmental impact assessments to be carried out before any FTA is signed. These countries have promulgated national policies that ensure trade is environmentally sustainable. For example, in 1999 the Canadian government issued a directive that requires a strategic environmental assessment of any policy, plan, or program before it receives ministerial or cabinet approval. Demonstrating how relevant the trade-environment nexus is to Canada, the government developed a framework in 2001 that stipulates environmental impact assessments are carried out before it signs any trade agreement.

Similarly, in the 1999, the EC also stipulated that environmental considerations should be integrated into the EC’s trade negotiations strategy. The US also through an executive order in 1999, committed to conduct environmental assessments of certain trade agreements. Most proponents of trade liberalization emphasize that restricting international trade may not be a panacea to reduce emissions. For instance, Paarlberg (2013) argues that one of the unintended consequences of the US restriction of sugar imports using quotas contributed to a high domestic price of sugar for industrial use in the US and this led to substituting away from sugarcane ethanol to corn ethanol. However, the processing of ethanol from corn is less energy efficient and the agriculture of corn also depletes the land though the applications of fertilizer (Paarlberg 2013).

The composition effect measures the share of polluting goods in the total or national output. Depending on whether the pollution-intensive sector expands (contracts), this may be harmful (helpful) for the environment. The composition effect is ambiguous as this depends on changes in the structure of the economy. The net effect of trade on the environment depends on these three components and this effect can either be positive or negative depending on which of the effect dominates the others. Thus, we construct our main hypothesis taking into consideration that the effect of trade openness on emissions could be either positive or negative.

Hypothesis 1

An increase in trade openness leads to a significant change (i.e, increase or decrease) in environmental emissions.

The level of development of a country also dictates the depth of integration of that country in global trade and emissions. Frankel and Romer (1999) provide robust evidence of this by stating that countries that trade more experience higher economic growth as trade raises income by spurring the accumulation of physical and human capital and by increasing output for given levels of capital. Similarly, the level of development of a country is an important determinant of pollution. Copeland and Taylor (2003) argue that changes in per capita income will lead to an increase in the demand for lower environmental emissions. Peters et al. (2011) clarify this further by differentiating the effect of international trade on emissions from production and consumption perspectives for both developing and developed countries. From the consumption perspective, they argue that although emissions from developed countries have stabilized over the years, they are still largely responsible for emissions generated in developing countries since developed countries are large importers of emission-intensive products produced in developing countries. In a sense that through emission transfer a country that is not directly involved in the production of pollution-intensive goods can also contribute to emissions increase through its consumption (import) of pollution or energy-intensive products.

From the production perspective, developed countries are also specializing in knowledge intensive service, research and development and high-value segments of the global value chains. At the same time there is reallocation of dirty or pollution-intensive production from developed to developing countries. Thus, it is intuitive to expect a lesser production-induced emission in developed countries compared to developing countries. This is important as Peters et al. (2011) made the point that significant share of global emissions are from production of internationally traded goods and services. Differentiating the effect of FDI on emissions for developed and developing countries, Demena and Afesorgbor (2020) find a pronounced reducing effect of FDI on emissions for developed countries compared to developing countries, and they explain that this could be due to stricter environmental regulations on emissions in developed countries than in developing countries. That notwithstanding, for manufacturing products which tend to be more capital and pollution-intensive, this is still more concentrated in developed countries than developing countries. For example, WTO (2019) indicates that among the top 10 largest exporters of manufacturing goods in 2018, the EU accounts for 39% , the US accounts for 9% compared to China which accounts for 18%. China and Mexico are the only developing countries among the 10 top exporters of manufacturing goods.

This therefore leads to our second hypothesis in which we hypothesize that the trade-emission elasticity differs qualitatively and quantitatively for developed and developing countries. We define qualitative in terms of sign - whether the trade-emission elasticity is positive or negative while quantitative is define in terms of the size or magnitude of the trade-emission elasticity.

Hypothesis 2

The trade-emission elasticity differs qualitatively or quantitatively for developed and developing countries.

2.2 Empirical Overview of the Literature

Since, the overall effect depends on the three effects (scale, technique and composition effects), this introduces an empirical uncertainty about the effect of trade on the environment. Empirically, a number of studies have tested this theoretical exposition by decomposing the trade effect on the environment into the three different components. For instance, Antweiler et al. (2001) confirm that the impact of trade on the environment can be positive or negative. Using the sulfur dioxide (SO₂) concentration, they show that trade-induced technique effect on pollutant concentration outweighs the scale effect, while the composition effect is close to zero. Cole and Elliott (2003) also indicate that the net effect of trade in terms of sign and size on the environment depends to a large extent on the specific type of pollutant. In particular, they find that trade liberalization increases nitrogen oxide (NO_x) and carbon dioxide (CO₂) while it decreases the emissions of SO₂. The distinction of the trade effect on a specific pollutant is important as Frankel and Rose (2005) argue that the effect would depend on whether a pollutant is a local or global in nature.

The negative effect of trade on the environment is similarly supported in a recent study by Shahbaz et al. (2017a). They indicate that trade openness increases emissions through significant expansions in industrial output. As a result of rapid globalization and the quest to be competitive in global markets, both developing and developed countries have greater energy utilization leading to potential emissions of CO₂. Other empirical studies that find similar results that trade impedes air quality for both developing and developed countries (Li et al. 2015) and intensifies pollution in non-OECD countries compared to OECD countries (Managi et al. 2009). Contrary to this popular view, there are also studies such as Eiras and Schaefer (2001), Frankel and Rose (2005) and Kohler (2013) who also claim that the trade appears to have a beneficial effect for certain environment pollutants.

The pollution haven hypothesis (PHH) has direct implication for the effect of trade on the environment as the effect would depend on the specific country or level of development of the country. Antweiler et al. (2001) emphasize this point by stating that it is implausible to only focus on how trade affects the environment, but rather suggest that trade should be conditioned on a country’s characteristics. In evaluating the effect of NAFTA on the environment, Grossman and Krueger (1993) identify a differential effect of NAFTA for the US and Mexico. They explain that through the mechanism of comparative advantage, increased trade among the NAFTA member countries would lead to an expansion of the labor-intensive sector in Mexico while leading to an expansion of the capital-intensive sector in the US. With labor-intensive production being a less pollution-intensive sector, emissions may decline in Mexico. Conversely, the US may see an increase in emissions because capital-intensive production is more pollution-intensive.

This example provides the nuance of the effect of trade on the environment, explaining that the effect may be conditional on a country’s level of development. Different studies focus on a specific country as a case study or use a group of countries to assess the impact of trade on emissions. Using a large sample of developing and developed countries, Kim et al. (2019) confirm the heterogeneous results of trade on CO₂ emissions for different permutations of trade flows in terms of North-North, North-South, South-North and South-South trade. They find that for developed countries’ trade with the North or the South, increases in trade flows lead to a reduction in CO₂ emissions. Whereas for developing countries, their trade with the North leads to an increase in emissions; however trade among themselves mitigates emissions. These differential results point to the fact that the effect of trade on emissions can be influenced by various institutional, economic and political factors. These factors differ for many countries and from one empirical study to the other as they evaluate the effect of trade on emissions using different set of countries, different proxies to control for institutional, economic and political factors. Thus, to effectively account for this variation in the different studies, it is very important to employ a multivariate meta-regression technique . By so doing, we can explain how country characteristics influence the effect of trade on the environment, as well as how different sets of control variables influence the reported results.

These conflicting positions in the empirical literature point to the need to synthesize existing knowledge in order to have a consensus about the effect of trade on the environment. The closest attempt made in the literature to derive a general consensus on the trade-environmental linkage was done almost a decade ago by Kirkpatrick and Scrieciu (2008). Although Kirkpatrick and Scrieciu conclude that trade liberalization is bad for the environment, they employ a conventional narrative review that relies on simple summaries instead of a systematic and objective review that is possible with meta-analysis. A meta-analytical review is an important tool to objectively review literature and also explain the heterogeneity in the previous studies. By using meta-analysis, we can filter out the personal biases and influences of the authors by controlling for biasing factors in econometric models (Stanley and Doucouliagos 2012). Apart from the methodological challenges of Kirkpatrick and Scrieciu (2008), the literature has also grown rapidly since their study was published in 2008. More specifically, 94% (i.e., 83 of the 88) of the empirical studies included in our analysis were published after 2008.

3 Data and Empirical Strategy

3.1 Data

Following the MAER-Net reporting guidelines of Stanley et al. (2013), we identify and collect data from empirical studies that estimate the effect of trade on the environment. To do this, we employ a search method from the economics, energy and environmental literature using a combination of keywords. Specifically, we use the Boolean connectors for the two keywords: Trade (OR trade openness, globalization, import, export, international trade), AND Emissions (OR environment, pollution, CO₂, SO₂, NO_x, climate change, environmental quality). The search was conducted using different bibliographic databases including Google Scholar, Web of Science and Scopus. Using, for instance, the keywords (trade and emissions) produce about 2,300 individual studies. In addition, we complement our search using a backward search approach by looking at the bibliographies of previous studies. Apart from the backward search, we also use forward search by checking for papers that have cited previous papers. The search period was between May to August 2018, and coding of the papers was done by two independent researchers (one author and a research assistant) and the other author double-checked the data entry to ensure the highest standard of quality.

Out of the large number of studies produced from the search, we restrict the selection to journal articles and unpublished working papers that employed an econometric approach to determine the effect of trade on the environment. Through the inspection of the titles, abstracts, introductions and conclusions, we finally identify 88 studies that were relevant for our study objectives. Our final data set contains 789 observations which are estimated coefficients from the 88 primary studies published between 1991 and 2018. The large number of studies confirm that the trade-environment nexus is a topical issue and hence conducting a meta-analysis is long overdue.

In the identification of the relevant papers, we realize that while some studies adopt an econometric approach, they restrict their estimations to only Granger causality tests rather than elasticities. Thus, we eliminate such studies from the identified sample of studies. In terms of publication status, about 90% of the primary studies are peer-reviewed journal articles (78 out of the 88 primary studies included) and the other 10 are working papers. The top five outlets in which the studies are published include Energy Policy (11 studies), Renewable and Sustainable Energy Reviews (9 studies) , Environmental Science and Pollution Research (6 studies), Economic Modeling (5 studies) and Natural Hazards (4 studies). The newest primary study was published in 2018 in Natural Hazards. Although the oldest study was published in 1991 by the Journal of World Trade, half of our primary studies were published in the last three years. As a result, the median study was published in 2015, indicating the environmental consequences of globalization have become a more sensitive issue recently as concerns related to climate change grow.

$$\begin{aligned} ln({{emissions}}_{it})\,= \,& {} \beta _{0} +\beta _{1} ln(GDPpc)_{it}+ \beta _{2}[ln(GDPpc_{it}]^2 + \beta _{3} ln(\mathbf{C} )_{it}\nonumber \\&+ \mathbf {\delta } ln\left( \frac{X+M}{GDP}\right) _{it}+ \alpha _{i}+ \alpha _{t} +\epsilon _{it} \end{aligned}$$

(1)

Data on elasticities (coefficients), standard errors (t-statistics), sample size, econometric methods and empirical settings (city or country) from the identified empirical papers were then collected. A typical econometric model estimated by the studies from which we collect the necessary statistics uses the log-log regression as specified in Eqn. (1). The model is estimated at country level, where i and t capture the country and time effects respectively. The dependent variable is captured by $emissions_{it}$ which measures the emissions of different environmental pollutants. Different types of environmental pollutants were used as the outcome variable by different studies: 66% of the studies used CO₂, 19% using SO₂, while the remaining used other pollutants. The main variable of interest is the trade openness index, which produces $\delta$ that captures the trade-emissions elasticity. Table 9 in the appendix provides information on the studies used in our meta-analysis.

The conventional approach of measuring trade openness, $\frac{export (X) + import(M)}{GDP}$ is dominant in the empirical studies. The studies selected mostly measure trade using the conventional definition of trade openness, hence making them comparable. However, some studies compute trade openness as a ratio or percentage.³ In terms of the functional form, about one-third of the studies employ a log-linear form and thus instead estimated semi-elasticities rather than elasticities. With such studies, we employ the Delta method in Gujarati (2009) to transform these effect sizes and their standard errors from semi-elasticities into full elasticities using the means, thus making the estimates comparable. If semi-elasticity is reported in the case of a log-linear functional form where the outcome variable is specified in a log form whereas the explanatory variable is in level, we use the sample mean for the trade variable to convert the semi-elasticity coefficients into full elasticity coefficients (Iršová and Havránek 2013; Demena and Afesorgbor 2020).

In addition, different studies used different sets of control variables ($C_{it}$). Apart from including GDP per capita (pc) and its square term to test for Environmental Kuznets Curve (EKC) hypothesis, most studies also control for the political, economic and institutional characteristics of the countries. These variables can also affect environmental quality in line with Torras and Boyce (1998) who show that literacy, political rights, and civil liberties tend to create a better environmental quality. In the case of panel regressions, the studies also control for individual country fixed effects ($\alpha _{i}$) and time effects ($\alpha _{t}$). Furthermore, there are also studies that employ a time series model, especially if they focused on a single country over longer time period. In case of the time series estimation, the estimated model is similar to (1), but without the i dimension. Because of the similarity of the primary studies in the specification of their econometric models, our coefficients are comparable across different studies. Although some studies employ different control variables, this could be accounted for by using the multivariate MRA.

3.2 Empirical Strategy

A meta-analysis involves collecting empirical estimates from previously published studies with the purpose of summarizing, integrating and synthesizing the combined effect sizes of the contrasting studies for a similar hypothesis, research question, and/or empirical effect (Stanley 2005). The combination of different studies helps to derive more precision and investigates the discrepancies. Stanley (2001) indicates that combining the results taken from individual studies would give more insight and greater explanatory power. Historically, meta-analysis has long been a standard for evidence-based research in medicines (Stanley and Doucouliagos 2012). Recent application of the meta-analysis approach in the trade literature include studies such as Afesorgbor (2017) and Cipollina and Salvatici (2010), which examine the effect of regional trade agreements on trade; Rose and Stanley (2005) which examine the effect of common currency on trade. For the environmental literature, Bel and Gradus (2016) examine the effects of unit-pricing on household waste collection while Ma et al. (2015) also examine consumers’ willingness to pay for renewable energy. More recently, Demena and Afesorgbor (2020) also look at the effect of FDI on environmental emissions using 65 empirical papers.

Table 1 Simple and weighted means of the trade-emissions elasticity

For our empirical approach, we adopt a two-step method in line with the meta-analysis literature. In the first-step, we calculate the meta-average or the combined effect size using both unweighted and weighted averages. According to Card (2015), to avoid the erroneous method of giving equal weight to estimates from different studies, it is important to also construct weighted averages using the variance as weights. Thus, in Table 1, we compute the simple and weighted average effect of trade openness on environmental emissions. The simple average shows a negative value contradicting that of the weighted average and both effects are insignificant. However, in the presence of potential publication bias and heterogeneity, these mean values would be spurious.

$$\begin{aligned} \delta _{j}=\beta _{0} +\beta _{1}SE_{j} \end{aligned}$$

(2)

Publication bias occurs largely when research papers with statistical significant results or consistent with the theory are preferred by editors, referees, or authors themselves, and thus, more likely to be accepted, and over-represented within research record (Stanley and Doucouliagos 2012; Havranek and Irsova 2011). To test and account for publication bias, we require the funnel asymmetry test (FAT) and the precision effect test (PET) analyses which are specified in Eqn. (2), where j captures the individual studies. The FAT is used to test the absence of publication bias. In the FAT test, a simple regression of the effect sizes $(\delta _{j})$ on standard errors ($SE_j$) is conducted. A significant $\beta _{1}$ parameter indicates the presence of publication bias because without publication bias, the effect sizes will be independent of the standard errors. Apart from an econometric approach for testing publication bias, it can also be examined graphically using the funnel plot (see Fig. 2). From Eqn. (2), we can also derive the PET which indicates whether there is any genuine effect beyond publication bias. The coefficient $\beta _{0}$ captures the PET and the size of the coefficient indicates the average effect of trade on environmental emissions.

$$\begin{aligned} \delta _{j}=\beta _{0} +\beta _{1}SE_{j} + \beta _{k} \sum _{j=1}^{n} X_{kj} \end{aligned}$$

(3)

In the second-step, a meta-regression or moderator analysis is conducted to determine how the differences in the individual studies affect the heterogeneity of the estimates. The individual studies differ in various dimensions, including sampled countries, sample size, environmental pollutants ($CO_2, SO_2, NO_x$), econometric models (log-log or log-linear, short and long-run, OLS or fixed effects), and types of data (panel or time series). These differences can drive the heterogeneous results in the literature. Taking stock of the literature shows substantial heterogeneity in terms of different pollutants, countries, econometric methods, time period and empirical results. To do this, we augment Eqn. (2) with the individual study characteristics ($X_{k}$). $X_{k}$ represents the vector of variables that captures the individual heterogeneity in the studies. The multivariate MRA is relevant in order to filter out how the individual differences in the design of each empirical study affect the effect sizes reported. The relevant moderator variables are categorized into data, estimation and publication characteristics as well as macroeconomic, emissions and trade variables. Table 2 provides information on the detailed variables that were used for the moderator analysis.

Table 2 Definition and descriptive statistics of collected variables

In the estimation of these equations, there are some econometric concerns. First, the presence of outliers that deviate from the majority of the effect sizes that are clustered around zero. To deal with this and identify the possible outliers, we use the Hadi (1994) method to filter out both the effect sizes and their standard errors. This method has been identified as suitable for outliers in multivariate data and commonly used in various applications of meta-analysis (for recent literature, see Havranek and Irsova (2011); Demena and van Bergeijk (2017)). The application of this method resulted in the availability of 688 reported estimates from 84 studies for the analysis after excluding 101 observations (13% of the reported elasticities).⁴ Second, the presence of heteroskedasticity makes ordinary least squares (OLS) econometrically infeasible to estimate the meta-regressions. Stanley and Doucouliagos (2012) indicate that since the effect sizes are coming from different studies, the variances of the effect sizes and the error terms would vary across studies and this violates the classical linear regression model. Homoskedasticity is a relevant assumption in estimation of classical linear regression model. Thus, they strongly suggest using the weighted least squares (WLS) approach by dividing the equations by the standard errors of the effect sizes. Using the WLS approach solves the problem of heteroskedasticity as this makes the variance approximately constant. Mathematically, the WLS procedure means that  Eqn. (2) is equivalently estimated using Eqn. (4). $t_j$ is the t-statistics, while $\frac{1}{SE_j}$ is the precision of the effect size.

$$\begin{aligned} t_{j}= \beta _{1} +\beta _{0}\frac{1}{SE_j} \end{aligned}$$

(4)

Another important empirical concern is the issue of within-study dependence which emanates from the practice of collecting multiple effect sizes from each study. To guard against this, we cluster the standard errors at the level of the individual study to address the issue of within-study correlation for all the different estimators (Oczkowski and Doucouliagos 2015). In addition, we also employ fixed effect (FE) estimation clustered at the level of studies to address the issue of individual within-variation. Apart from these, between-study dependence is also a concern as we have collected reported estimates across studies published by the same authors. This is an important concern as reported estimates from different studies are likely to be correlated. Empirically, we have tested for the existence of between-study correlation, applying the Breusch-Pagan Langrange multiplier (BP-LM) method. The BP-LM revealed between-study level effect of 558.32 with p< 0.001, statistically significant at 1% level for the presence of between-study dependence. To account for this concern, we employ the mixed-effects multilevel (MEM) model as recommended by Stanley and Doucouliagos (2012). MEM is estimated using the restricted maximum likelihood when there is data dependence as a result of multiple effect sizes from the same study and authors.⁵ Following this representation, therefore, estimates of trade-emissions elasticities as nested within each study and the estimates are modeled to differ between studies. Importantly, the relevance of controlling for between-study correlation using the MEM has been emphasized in recent studies such as Demena and van Bergeijk (2017), Floridi et al. (2020) and Havranek and Irsova (2011).

4 Results and Discussion

4.1 Bivariate FAT-PET Analysis

We present the bivariate results for the FAT-PET in Table 3. The results are estimated using three estimation techniques, OLS, FE and MEM, however, the MEM is our preferred method because it controls for both between and within heterogeneity. In all these estimations, we adjust for the data by using the WLS and cluster at the individual study level. The results show that there is a genuine effect of trade on the environment. The positive effect size for the PET is consistent across the three different estimation methods. The positive coefficient indicates that trade openness leads to a significant increase in emissions. Put differently, this means that more open-economies contribute more in terms of global emissions. On the basis of the MEM estimator, the size of the coefficient indicates that a 1% increase in trade openness contributes to about 0.02% increase in environmental emissions. The magnitude of meta-average is also plausible as Cole and Elliott (2003) find that 1% increase in trade intensity increases emissions for a median country by 0.04%. This confirms that increases in trade stimulate the scale of economic activity and through the scale effect, trade leads to an increase in emissions. Plausibly, increased global trade is mostly accompanied with increases in capital-intensive production techniques that have high emissions intensity (Copeland and Taylor 2003). The results therefore signals that the emission-increasing impacts of the scale and composition effects dominate the emission-reducing technique effect, thereby resulting in a positive net effect on emissions.

Table 3 Bivariate MRA for FAT-PET: publication bias and underlying effect

Turning to the FAT analysis, our results indicate the absence of publication bias as $\beta _1$ is statistical insignificant. This indicates that the effect sizes are independent of the standard errors. Thus, we can objectively conclude that there is no selection bias for significant effect sizes in the literature. The FAT result is also confirmed by the funnel plot in Fig. 2. The figure is a scatter plot of the relationship between the effect sizes and the their precision (inverse of the standard errors). The figure is also symmetric from a pictorial point of view and that is in line with the assertion of Stanley and Doucouliagos (2012) that an asymmetric funnel is an antecedent for publication bias.

Fig. 2

The funnel plot: relationship between precision and effect sizes. Data Source: Based on studies for the meta-analysis

In line with the arguments of Copeland and Taylor (2003), we decompose the results for studies that use countries at different levels of development. We classify the countries into two main groups; developing and developed on the basis of the United Nations (UN) classification. Developed countries represent single country (e.g., Canada, the US and Japan) and multi-country (e.g., EU, OECD and groups of hing-income countries) case studies.

Similarly, developing countries also involve single country as well as multi-country case studies. In terms of single country studies, the countries include: Algeria, Bahrain, Bangladesh, Benin, Botswana, Brazil, Burkina Faso, Cameroon, China , Côte d’Ivoire, Egypt, Gabon, Gambia, Ghana, India , Indonesia, Iran, Ivory Coast, Jordan, Kenya, Korea, Kuwait, Laos, Lebanon, Malaysia, Mexico, Morocco, Myanmar, Niger, Nigeria, Oman, Pakistan, Philippines, Qatar, Saudi Arabia, Senegal, Sierra Leone, South Africa, Syria, Thailand, Togo, Tunisia, Turkey, UAE, Vietnam, and Yemen. Whereas, a group of developing countries represented as, Africa, LDCs, Middle East and North African (MENA) countries, low-income economies (LIE), emerging and developing economies. To account for studies that used a mix of these groups, we add another category for studies that include both developing and developed countries.⁶

Table 4 presents the FAT-PET results for different groups of countries. Consistently we find that trade increases emissions under the three different categories of countries. Although the results are similar qualitatively in terms of sign of the PET coefficient, quantitatively, the size or magnitude differs. We find a pronounced effect for developed countries compared to developing countries. The greater trade-emission elasticity for developed countries compared to developing countries supports the emission transfer argument by Peters et al. (2011) in which developed countries may not be directly involved in the production of pollution or energy-intensive products, but can also contribute to emission increase through their consumption (import) of those dirty goods. In addition, looking at it from consumption rather than production perspective, developed countries are more-opened economies, and they still account for relatively larger share of import of goods and services produced in developing countries. Thus, emission transfer via international trade as argued by Peters et al. (2011) may be responsible for why developed countries still have a pronounced effect compared to developing countries. This may not be surprising especially when Peters et al. (2011) indicate that the level of emissions reduction in developed countries is far less than emission transfer via international trade.

Table 4 Bivariate MRA for FAT-PET: publication bias and underlying effect for different group of countries

4.2 Multivariate Analysis

In addition to the bivariate analysis, we also conduct a multivariate or moderator analysis in order to account for the heterogeneity in the studies. This is important as the effect sizes reported in the empirical studies may have been influenced by several factors such as different countries, pollutants, time period, methodology and explanatory variables included in the econometric models. Stanley and Doucouliagos (2012) highlight that if the heterogeneity is non-random, then the results from bivariate analysis may be biased. Econometrically, we account for the heterogeneity in line with the MAER-Net guidelines by coding explicitly any differences in research dimensions such as data, estimation and publication of the empirical studies (Stanley et al. 2013).

To capture these dimensions, we use dummy variables as defined in Table 2 that could potentially explain the effect sizes. However, because of the high dimensionality and possible multicollinearity of the moderator variables, we employ the general-to-specific (G-to-S) approach. The G-to-S approach begins with the inclusion of all potential moderator variables in the Eqn. (3) and then removing the insignificant variable one at time until only statistically significant variables remain (Stanley and Doucouliagos 2012). This is in line with the MAER-Net reporting guidelines. The guidelines recommend the use of the G-to-S modeling approach to simplify the multivariate MRA. The G-to-S approach has an added advantage of improving the degrees- of-freedom and avoiding the situation of multicollinearity. Table 5 reports the results for the multivariate analysis. The moderator variables included in Table 2, but not in Table 5, are insignificant variables that are excluded due to the G-to-S approach. Thus, the following 13 moderator variables; trade openness-ratio, long run, year FE, endogeneity, log duration of the data, data source, institution, panel, time series, GDP, log of observation, journal impact, and energy consumption were excluded. Empirically, the joint test [F(13, 657)=0.99] of the excluded moderator variables did not reject the null hypothesis of a zero-joint effect (p-value: 0.459), supporting that the excluded variables are not only individually but also jointly insignificant.

Table 5 Explaining heterogeneity in the trade-emissions elasticity for all countries

The genuine effect of trade on the environment remains robust even after accounting for various dimensions of heterogeneity. The PET finds a positive and significant coefficient for the trade-emissions elasticity. Thus, confirming the robustness of the result that trade spurs environmental emissions. The result for FAT also confirms the absence of publication bias. Our results from the multivariate MRA also show that the different moderator variables affect the effect sizes differently. First of all, the estimation characteristics tend to have a significant effect on trade-emission elasticities. Based on column (3) of Table 5, studies that employ OLS estimation tend to report lower effect sizes by 0.0279. Conversely, including country fixed effects in the regression models also produces a greater effect size. For instance, for studies that control for country fixed effects on average have emissions that are higher by 0.0407 compared to studies without country fixed effects. In addition, the double logarithmic functional form leads to less pronounced effect of trade on environmental emissions by 0.0857.

Similarly, studies that define trade openness in terms of exports plus imports as a percentage of GDP as compared to the studies that measure trade openness with values of export plus imports (i.e., not weighted by GDP) find a more pronounced effect of trade on emissions, reporting on average 0.0287 higher emissions. Importantly, we find that the choice of the pollutant matters and has an underlying effect on the trade-emissions elasticity. Studies that use CO₂ as a pollutant find larger effects compared to studies that used other pollutants. For CO₂ as a pollutant, reported effects increase by 0.0159, providing more positive trade-emissions effects as compared to other pollutants such as nitrogen dioxide, other volatile organic compounds or other pollutant measures. For SO₂ the effect goes in the opposite direction, reducing emissions by 0.0629.

Studies that include urbanization as an additional control variable in explaining the variation in emissions also have a smaller effect of trade on emissions. This simply confirms trade and urbanization are correlated such that if urbanization is omitted this is likely to bias the trade-emission elasticity. Theoretically, this is expected as urbanization and trade can be related (Charfeddine and Khediri 2016). For instance, Ades and Glaeser (1995) and Krugman and Elizondo (1996) relate trade and urbanization by arguing that trade restriction and low levels of international trade increase the concentration of urban population while increasing trade liberalization would likely weaken urbanization. Empirically, Kasman and Duman (2015) find that there is a unidirectional causality running from energy consumption, trade and urbanization to emissions. A sub-sample analysis for different groups of countries (developing vs. developed countries in Table 7) indicates that controlling for urbanization has a negative and significant effect for developing countries but insignificant for developed countries. Martínez-Zarzoso and Maruotti (2011) confirm the differential effect of urbanization on emissions for both developing and developed countries. Most importantly, the pace of urbanization has been more rapid for developing countries while that of developed countries has remained largely the same (Yuan and Guanghua 2015).

For the publication characteristics, our results show that the year of publication has a negative and significant effect. This means that more recent studies tend to produce lower trade-emissions elasticities (on average 0.0486 lower). This finding is in line with the ‘economics research cycle’ hypothesis (Goldfarb 1995) which suggests that seminal research often produces large and significant estimates. While skeptical follow-up studies become common and may dominate the literature and commonly exhibit a downward trend of the empirical effect in question. Similarly, the quality of the papers considered in our meta-analysis as captured by the moderator variable, number of citations in Google Scholar, has a significant effect on the reported effect sizes. If the number of citations increases by one this significantly reduces the effect sizes by a magnitude of 0.0297. Apart from these, studies that have been reviewed and published in academic journals tend to report a greater impact (peer-reviewed studies likely to report about 0.0863 higher emissions).

Fig. 3

The plot of the estimated coefficients and their 95% confidence intervals. Note: This is based on the the MEM results for Table 5 (All) and Table 6 (developing and developed countries)

4.3 Further Analyses and Best Practice Approach

We conduct two main further investigations by running the multivariate analysis for different groups of countries and also for different pollutants. Table 6 presents the results for different groups of countries. Consistently, we find that there is a genuine underlying effect of trade on the environment irrespective of the country of study. From the qualitative point of view, the results indicate that the trade-emissions elasticity is positive and significant for both developing and developed economies. Quantitatively, we again find the trade-emissions elasticity is more pronounced for developed countries than for developing countries. For studies that mix developed and developing countries (both countries), although the effect is positive and this is not significant. The results as to how the heterogeneity or differences in the design of the studies explain the effect sizes are summarized in Fig. 3.

Table 6 Explaining heterogeneity in the trade-emissions elasticity for different group of countries

A further robustness check is the multivariate regression for different pollutants as indicated in Table 7. First, we find a significant FAT which is a clear indication of publication bias when we disaggregate the sample by the type of pollutants. Specifically, using the partitions of Doucouliagos and Stanley (2013), the size of the bias for CO₂ will be classified as little to modest, but substantial for SO₂. This means that publication selectivity was only significant for studies that focused on SO₂. In addition, this could be an indication that there was a possible offset between CO₂ and SO₂ when these studies are combined. Second, we find that the trade-emissions elasticity is positive and strongly significant when environmental pollutant of interest is CO₂. Although the trade-emissions elasticity is positive for SO₂, the effect is not statistically significant. This differential effect on CO₂ and SO₂ is in line with Frankel and Rose (2005). They argue CO₂ has a free-rider problem as its adverse effects transcend national borders. Thus, individual countries are less committed to reduce their CO₂ emissions, especially for the reason of remaining competitive in the global market. This differential effect for CO₂ and SO₂ is also supported empirically by Demena and Afesorgbor (2020) in which they found that the effect of FDI in reducing emissions was more pronounced for SO₂ compared to CO₂. They explain that since SO₂ is a local pollutant, governments are more eager to clamp down on its emissions in order to prevent the associated health hazards for the local populace.

Table 7 Robustness: Effect of trade on different pollutants

The inclusion of observable sources of heterogeneity provides a better estimation of the genuine effect (PET) of trade on the environment. However, the PET coefficient is no longer the measure of the underlying effect like in the case of bivariate MRA. The underlying effect is obtained from the combination of the PET variable and other variables that are included in the multivariate MRA. With the multivariate MRA, neither publication bias nor the authentic effect is represented by one single moderator variable (Stanley and Doucouliagos 2012). They argue that there is no single variable that can be regarded as the authentic one, but rather a combination of selected moderator variables used to capture the heterogeneity. Largely, the underlying effect would depend on the professional judgment of the meta-analyst conditional on the selected sources of heterogeneity.⁷

In what follows, we explore the findings presented in Table 5 to derive the “best practice” genuine effect. We apply the ‘best practice’ approach by limiting the explanatory variables to a subset consisting of indicator variables for CO₂, country fixed effect, double-log and urbanization. We select these variables because they are widely used by most of the primary studies included in our analysis. We also consider a set of publication characteristics of studies included in our analysis. Thus, a subset of the following variables; publication year of the study, peer-reviewed, as well as the citation of the primary studies were included. We select these variables for three reasons. First, recent studies report lower elasticities (as shown in Table 5) and this is consistent with the ‘economics research cycle’ hypothesis. Second, 80% of our samples is constructed from peer-reviewed publications. Most researchers would argue that peer-reviewed studies are of higher quality than those that are not (Stanley and Doucouliagos 2012). Third, we also use the number of citations of the studies as this could be an indicator of the quality of study and the reliability of the effect sizes.

For our ‘best practice’ results, we present the results based on Table 5 (the full sample) and Table 6 (for developing and developed economies data separately) to obtain the genuine effect conditional on observable sources of heterogeneity, using our preferred estimation - MEM. The results are provided in Table 8. The best practice effect produces an elasticity of 0.113 which is strongly statistically significant at 1% level of significance for the full sample. When we conduct the analysis for the sub-sample groups of developing and developed countries separately, the approach yields similar positive results, but different in terms of the magnitude of the effect. Specifically, we find genuine and significant effects of 0.097 and 0.947 for developing and developed countries, respectively. In this regard, the genuine effect results suggest that trade openness contributes to emissions more significantly for developed countries compared to developing countries. Therefore, after correcting for publication bias, controlling for the potential moderator variables and the best-practice approach, all these findings indicate that the meta effect is positive and significant, suggesting trade openness contributes to environmental emissions.

Table 8 Meta-effect based on best practices

5 Conclusion

Without doubt, the effect of trade on the environment is a controversial topic. Although there is a rapidly growing literature on the linkage between trade and the environment, there is no certainty about the underlying effect as there is a wide range of conclusions from similar studies. Thus, the simple question of whether trade is good or bad for the environment has received a great deal of attention both theoretically and empirically, however, we are far from reaching a definite conclusion or consensus.

The conflicting positions in the literature may have implications for environmental policies in many countries as countries may be selective of research to support their (in)actions for the environment. Individually, many countries have also recognized the strong linkage between globalization and the environment, and are thus making polices to ensure that their quest to remain competitive globally are not achieved at the expense of the environment. In contrast, there are also countries that have not taken the needed actions or designed environmental regulations to formalize their environmental commitments of reducing emissions. Shahbaz et al. (2017a), for instance, argue that the lack of consensus on the trade-environmental linkage as one of the key reasons for inaction on appropriate environmental regulation policies to curb emissions. This paper makes the first attempt to synthesize the trade-emissions effect size by employing the tool of meta-analysis which is the most objective and quantitatively rigorous approach to do a systematic review of empirical literature. Our main findings are as follows:

First, our results discount the presence of publication bias in the literature. This means that, the personal bias of researchers, reviewers and editors does not follow any systematic trend in the literature. After accounting for publication bias, we find that trade increases emissions level. Specifically, employing the bivariate FAT-PET analysis in accordance with the MAER-Net guidelines, we find that a 1% increase in the trade openness of a country leads to an increase in the level of emissions ranging from 0.02 to 0.06%. However, estimating the corrected effect from the multivariate MRA, conditional on observable sources of heterogeneity, improves the size of the effect. We find that the underlying meta effect is 0.113 and this is strongly statistically significant. This suggests trade openness has a positive and significant effect on emissions notably after controlling for publication bias and heterogeneity. Overall, increases in trade intensity increase environmental emissions due to a large scale effect that dominates the technique effect. This finding supports the view that globalization can result in the adoption of loose-environmental production techniques. The increased production coupled with increased global demand indirectly contribute to the exploitation of the environment and the depletion of natural resources.

In addition, the magnitude of the effect size depends on whether a country is developed or developing as we find a more pronounced effect for developed economies. This could largely be explained from the concept of emission transfer where developed countries although are not dominantly involved in the manufacturing of energy-intensive products, but still contribute to emissions through their consumption of those environmental dirty products produced in developing countries.

With significant heterogeneity across the empirical studies, it is imperative that we account for differences in the study designs. By accounting for the heterogeneity, we can also explain how specific differences such as data, publication, estimation and other variations affect the validity of the results. Using moderator regression analysis and the G-to-S modeling approach, we find that the positive effect of trade on environmental emissions is strongly robust. For the various dimensions of the study that affect the effect sizes, we find that estimation techniques (OLS or FE) significantly reduce the trade-emissions elasticity. In addition, the functional form of the regression model, the inclusion of country fixed effects, the definition of trade openness as exports and imports as percentage of GDP, the type of pollutant, the inclusion of urbanization as an additional control variables, the year of publication, the number of citations and whether a study has been published or not all have significant effects on the effect sizes.

The findings of this study may have interesting policy implications, as trade could be used to minimize or avoid the potential adverse environmental effects and maximize the positive impacts of globalization on environment. Our results that trade contributes to increased global emissions highlight the importance of making trade policies more compatible with sustainable environment policies by incorporating environmental decision-making into trade policy formulation. For instance, this emphasizes that environmental taxes should be aligned with a firm’s participation in global trade. This also reinforces the relevance of including the distribution or transportation component in the calculations of carbon taxes. The results also put the WTO in the limelight as it may be required that they play a directly active role in clamping down on environmental emissions. The WTO could use the gains associated with global trade as effective bargaining strategies to demand environmental accountability from countries hoping to benefit from global trading systems.

In addition, regional trade agreements (RTAs) that are major appendage of the global trading systems could be used to effectively promote sustainable trade and environmental policies simultaneously. As countries seek to join RTAs, they also embrace environmental cooperation agreements. For instance, many countries including Canada and those in the EU, have developed national policies that stipulate that prior to signing any trade agreement, environmental impact assessments must be carried out. That means that any country that signs trade agreements with those countries would also automatically sign environmental cooperation deals.

In the light of our meta-analysis, future research should focus particularly on increasing the number of studies as this topic has generated keen debate within policy circles and the literature is continuously growing. Our data collection ended in August 2018 and the literature on this topic has increased already from that time period. In addition, we have excluded some studies because of the lack of information on key variables that are required for conducting meta-analysis (e.g., some studies that do not report t-values and standard errors). Although we did make frantic effort to contact the authors of these excluded studies to obtain those missing data in their papers, the success rates are very low.

Appendix

Table 9.

Table 9 Appendix A: List of studies