Aspects of Governance and $\hbox {CO}_2$Emissions: A Non-linear Panel Data Analysis

Abstract

The reduction of $\hbox {CO}_2$emissions has been at the centre of worldwide debates on environmental issues, though its inclusion as one of the millennium development goals (MDGs) by the United Nations has changed the focus of relative literature. Among many, one of the World Bank’s “recipes” to achieve a higher position toward MDGs has been to undertake reforms for a better governance. While, the majority of researches’ focus has been on one single aspect and its relationship with the environment, some studies, have simultaneously considered two governance dimensions. In this paper, we focus on the role of several aspects of governance on $\hbox {CO}_2$emissions. This provides us with a chance to explore the possible impacts of all aspects of governance on a more direct measure of emissions, that is different to previous researches which have focused on the indirect transmission and considered the maximum of two. Using the IV method within panel data analysis, we show that only one single aspect of governance, Control of Corruption, has a negative significant effect on $\hbox {CO}_2$emissions and its effect has a non-linear relationship. The non-linearity exists in both parametric and nonparametric analysis after controlling for endogeneity.

Appendices

Appendix 1

Appendix 1.1: Variables Definition and Data Source

See Table 4.

Table 4 Definition of variables and their data source

1.1 Appendix 1.2: Countries in the Sample

See Table 5.

Table 5 Countries in the sample

1.2 Appendix 1.3: Two Way Fixed Effects Regression

See Table 6.

Table 6 Two way effects regression

1.3 Appendix 1.4: Details of the Solution for Endogeneity in Semi-Parametric Regression

Following proofs are from Vincenzo Verardi (author of relevant Stata package) presentation in 2013 UK Stata Users Group meeting.

In Standard IV, the variable X is called endogenous when in $y=X \beta + \epsilon$ it is possible to show that $E(\epsilon | X) \ne 0$. As a solution an exogenous variable($\backslash$Instrument) W is found and therefore the estimated coefficient would be $\hat{\beta _{IV}}=\left( X'P_{w} X \right) ^{-1} X'P_{w}y$ in which $P_{w}=W(W'W)^{-1}W'$ is the part of endogenous variable that can be explained by the instrument (i.e. W). In other words, $P_{w}$ is the fitted values of $X=W+\nu$. And by using the IV the equation would be $y=X \beta + + \gamma {\hat{\nu }} + \epsilon$ (CFA).

However, in Semi-parametric, where the equation can be presented as $y=X \beta + m(z)+ \epsilon$, if $E(\epsilon | X) = 0$ and $E(\epsilon | z) \ne 0$ finding the a solution can be complicated and difficult. To explain the source of difficulty, take the expected value from the equation, condition on z : $E(y|z)=E(X|z) \beta + m(z)+\underbrace{E(\epsilon | z)}_{\ne 0 }$. Because z is endogenous it is impossible to efficiently estimate the E(y | z) and E( X |z).

However, if we find a variable W that is correlated to z and not $\epsilon$ and have $z= W \pi + \nu$ and $E(\nu | W)= 0$. Then if we assume that there is a linear relationship between $\nu$ and $\epsilon$ such as $\epsilon = \rho \nu + \eta$.¹⁷

With this new notations, the main equation would be $y=X \beta + + m(z) + \rho \nu + \eta$ ¹⁸ and can be estimated using the semi-parametric method.

1.4 Appendix 1.5: Nonparametric Estimates of $\theta (\cdot )$ with the Actual Values

See Fig. 6.

Fig. 6

Nonparametric estimates of $\theta (\cdot )$

1.5 Appendix 1.6: A New Nonparametric Estimator in Fixed Effects Panel Data

The earlier estimator by Baltagi and Li (2002), Verardi and Libois (2012) which used series estimation of B-Splines of a single nonparametric variable was used in several studies such as Desbordes and Verardi (2012). However, the recent estimator by Pütz and Kneib (2016) accommodates a multidimensional nonparametric estimator, by penalizing the splines for each nonparametric variable. In order to show the robustness of the results with regard to recent development and the choice of model and estimator, in this section the model is estimated assuming the only variables that are considered as nonparametric are CC and LGDPPC (following EKC literature). In this settings, the rest of variables assumed to have a linear function. As illustrated by Fig. 7, despite the relatively larger dispersion the general results are consistent. Another interesting aspect of the results, shown in Table 7, is the close estimates of variables to their parametric counterparts in Table 3.

Fig. 7

Nonparametric estimates of $\theta (\cdot )$—penalized splines

Table 7 Results of nonparametric estimation following Pütz and Kneib (2016)