*2.3. Comment and Discussion*

From the above literature review, we can discover that the influence of financial development on carbon emissions is still under debate in both the theoretical and empirical research, which reflect the complexity of their relationship which cannot be readily detected or described.

Specifically, the theoretical research reveals that the financial development has both positive and negative effects on carbon emissions, the aggregate effect might be determined by the relative size of these positive and negative effects. The empirical research reflects that the influence of financial development on carbon emissions varies across countries and regions. Actually, it also demonstrates the viewpoint of the theoretical research to some extent, as it is reasonable to consider both the positive and negative effects are divergent in different countries and regions. Although the influence of financial development on carbon emissions remains in dispute, the relevant research on this topic have provided important theoretical values for environmental policy making.

However, two limitations exist in the literature on this topic: firstly, most of the researchers selected regional or individual country samples as the research objects, but few focused on this issue from a global perspective; secondly, the application of different methods, samples, and data, has created challenges for comparing research completed by different scholars.

Considering the above limitations, we collected a comprehensive country sample that contained the data of 155 countries to analyze the influence of financial development on carbon emissions from the global perspective, in order to provide more empirical evidence on this topic.

#### **3. Empirical Strategy and Data**

### *3.1. Empirical Model and Methodology*

In this paper, we focused on three research objectives: first of all, we analyzed the influence of financial development on carbon emissions from the global perspective based on panel data of 155 countries, which could enable us to detect their relationship on a macro angle; in addition, we researched this issue by dividing the sample countries into two sub-groups—developed countries, and emerging market and developing countries, to detect national differences under a unified analytical framework. The discrepancy of the empirical results reflects the heterogeneous effect of financial development on carbon emissions in different countries and regions. However, due to the inconsistent samples, proxy

variables and methodologies adopted by scholars, it was quite difficult to compare their empirical results in a unified framework. Therefore, after the full sample analysis, we further divided our sample into different sub-groups to examine the national effect of financial development on carbon emissions across different types of countries, in the same empirical framework. According to general practice in the empirical works [29–31], we divided our sample countries into two sub-groups—developed countries, and emerging market and developing countries, as it is widely believed by the scholars that a significant discrepancy in the aspects of economic structure, technical level, and resource endowment exists between these two groups, which may result remarkable implications on macro factors (please refer to Appendix A for more information about the country classification); lastly, we investigated the influence of different aspects of financial development on carbon emissions by adopting a series proxy variables of financial development. Much research deems financial development as a unique concept and takes one or two indexes to be its proxy variables. Actually, scholars commonly believe that financial development has rich connotation that can be divided into different aspects, therefore, besides the use of a comprehensive index, we further adopted five concrete indexes of financial development to analyze the effect of different aspects of financial development on carbon emissions, apart from its aggregate effect, to provide more specific evidence on this topic.

Considering the above research objectives and following the general practice on this topic [15,21,28], we established the dynamic panel model below:

$$CE\_{it} = \alpha + \beta\_0 CE\_{it-1} + \beta\_1 FD\_{it} + \gammaControl\_{it} + \mu\_i + \varepsilon\_{it} \tag{1}$$

where *CEit* represents carbon emission; *FDit* signifies financial development; *Controlit* denotes a series of control variables; β0, β1, and γ are the corresponding coefficients; μ*<sup>i</sup>* represents the unobserved country specific effect; ε*it* is the residual term; and *i* and *t* indicate the country and time, respectively.

We introduced the lag-term of carbon emissions into the regression equation to reflect the dynamic process of carbon emissions, which was consistent with reality. Adding a lag-term can eliminate the influence of uncontrollable factors, increasing the credibility of the regression results.

As a result of the existence of a lag-term, the model could not be estimated by ordinary least squares (OLS) or traditional panel model estimation methods (such as fixed-effect or random-effect), as they would have caused an endogenous problem and therefore could not provide effective estimators. To solve this problem, we adopted the generalized method of moments (GMM) [32–34] to estimate the above model. GMM can be divided into difference GMM and system GMM; each of them can further be divided into one-step and two-step estimation methods according to the selection of different weight matrixes.

Compared with the difference GMM, the system GMM can help mitigate the problems of weak tools and limited sample errors and can improve the efficiency of estimation. The two-step estimation performed better in handling the autocorrelation and heteroscedasticity problems than one-step estimation. Therefore, we adopted a two-step system GMM method to estimate our model. We adopted Stata (StataCorp, College Station, Texas, USA) and the command "xtabond2" to complete the estimations. Please refer to Roodman [35] for more details about this command. We conducted the test of serial correlation and the effectiveness of instrument variables to examine the consistency of the estimators, based on relevant statistics. In addition, the Hansen test was used to judge the effectiveness of the instrument variables rather than the Sargan test, as Roodman [35] shows that the Sargan test is not robust to heteroscedasticity or autocorrelation.

#### *3.2. Data*

So far, no indictors are widely accepted as proxy variables of financial development due to their rich and complex connotation. Scholars adopt many different indicators according to their research objectives and the data availability [36–40]. In our research, we used the comprehensive index of financial development proposed by Svirydzenka [41] in the main regression, which allowed us to investigate the "aggregate" effect of financial development on carbon emissions. Please refer to Appendix B for more information of this index.

To analyze the influence of different aspects of financial development on carbon emissions and to guarantee the reliability and accuracy of the empirical results, we also adopted another five variables as the proxies of financial development besides the comprehensive index proposed by Svirydzenka [41].

Following the common research on carbon emissions [27,42,43], we used carbon dioxide emissions (metric tons per capita) as its proxy, and selected four control variables: trade openness, urbanization, population growth, and industrial structure. The details of the variables used in our model are presented in Table 1.


**Table 1.** Variable descriptions.

Note: GDP denotes gross domestic product.

According to the data availability, our sample contained 155 countries, including 35 developed countries and 120 emerging market and developing countries. The main regression data cover the period from 1990 to 2014, and in the robustness checks, the period was extended to 1960–2014. All the variables were extracted from the World Development Indicators database of the World Bank, except for the financial development (FD)1 variable, which was sourced from the International Monetary Fund (IMF) database. All the variables were transformed into the natural logarithms to reduce nonnormality and heteroscedasticity [44], except for FD1 and population growth (POP) as they were already dimensionless or ratio indexes. Table 2 displays the descriptive statistics of the variables for the main regression. We adopted several regressions in the section of empirical analysis, and to save space, the descriptive statistics of the other regressions are not presented but are available from the authors upon request.

**Table 2.** Descriptive statistics of main regression.


Note: Obs. denotes number of observations. SD denotes standard deviation.

Table 3 presents the correlation matrix of all the variables in the main regression. Generally, when all the correlation coefficients of each of the variables are less than 0.85, the model is considered to not have a multicollinearity problem [45,46]. Table 3 shows that all the correlation coefficients were less than 0.85; therefore, we thought that our model was not affected by the multicollinearity problem. We also calculated the correlation matrixes for other regressions besides the main regression; all of the correlation coefficients were less than 0.85.


**Table 3.** Correlation matrix of the variables.

#### **4. Results and Comments**
