The environmental Kuznets curve at different levels of economic development: a counterfactual quantile regression analysis for CO₂ emissions

Abstract

This paper applies the quantile fixed effects technique in exploring the CO₂ environmental Kuznets curve within two groups of economic development (OECD and non-OECD countries) and six geographical regions – West, East Europe, Latin America, East Asia, West Asia and Africa. A comparison of the findings resulting from the use of this technique with those of conventional fixed effects method reveals that the latter may depict a flawed summary of the prevailing income–emissions nexus depending on the conditional quantile examined. The paper also extends the Machado and Mata decomposition method to the Kuznets curve framework to explore the most important explanations for CO₂ emissions gap between OECD and non-OECD countries. We find a statistically significant OECD--non-OECD emissions gap and the decomposition reveals that there are non-income related factors working against the non-OECD group's greening. We tentatively conclude that deliberate and systematic mitigation of current CO₂ emissions in the non-OECD group is required.

Keywords:

• environmental Kuznets curve
• economic growth
• sustainable development
• quantile fixed effects
• decomposition analysis
• income–emissions nexus

JEL codes:

• Q56
• Q58

Notes

1. This paper uses the terms economic development, growth and progress interchangeably. The paper does not distinguish between the three concepts.

2. Moreover, transferring the stylised reasoning from the epidemiological relationship between outdoor concentrations of local pollutants and adverse human health to the Kuznets Curve subject – being that higher concentrations of pollutants, such as particulate matter and sulphur oxide, are more harmful to humans relative to lower concentration levels (see, for instance, Pope, 2007; Yaduma et al. 2013) – an analogous reasoning applies to the relationship between concentrations of CO₂ emissions and the greenhouse effect.

3. The two techniques – quantile regression and random coefficients models – are related as they can estimate flexible slope coefficients across quantiles or groups.

4. However, it is worth noting that Huang et al. (2007) employed the same method to investigate the original Kuznets relationship between income inequality and economic development. Also, recently Flores et al. (2014) used the technique in exploring the EKC for nitrogen and sulphur emissions at the US state level.

5. In addition to the movement of these industries, there are technical transfers, particularly of advanced and cleaner production technologies, from the developed to developing countries. Consequently, the overall effect of pollution havens on the environmental quality of the developing countries is not a simple one-way relationship. This also depends on whether the depletion effect of pollution havens outweighs the enhancement effect of technical transfers.

6. It is worth noting that the traditional framework does not control for other possible determinants of emissions; for instance population density, energy use, income distribution within the country and trade openness amongst other factors. This is not to imply that this framework belittles the role of these factors. The choice of income (in its level, quadratic and cubic polynomials) is based on three main reasons. Firstly, the EKC hypothesis is mostly concerned with the shape of the relationship between income and emissions without obtaining best predictions for emissions in subsequent years. Secondly, data limitations restrict the analysis to income and emissions. In this respect, the use of panel techniques that sweep cross-country effects allows to control implicitly for any invariant determinants. Thirdly, the framework allows comparability with similar studies (see Azomahou et al. 2006 for more details).

7. As in the left hand side variable, emissions, the income variables are measured in logs as well.

8 Conversely, the non-OECD covariates could be used to generate ems*_OECD(τ)=X_non-OECDβ_OECD(τ_j). Estimation results can be expected to differ. Given our research perspective, we prefer using the OECD sample as the ‘counterfactual’.

9. Again, a similar counterfactual density could be generated for the OECD group if necessary; f*(lnems_OECD; X_non-OECD). That is, the OECD distribution of emissions if its covariates were distributed as in the non-OECD group.

10 Similarly, Employing the original OB method, the emissions gap corresponding to Equation (8(8) ) is: , where the first and second terms on the left-hand side of the equation are the mean outcomes of the OECD and non-OECD per-capita emissions respectively; the right-hand side of the equation denotes the emissions gap, and are vectors of explanatory variables evaluated at their means for the OECD and non-OECD groups respectively; and are the conforming vectors of estimated coefficients for OECD and non-OECD groups. Thus, the first and second terms of the right-hand side of the equation are the explained and unexplained components of the emissions gap respectively [see Oaxaca (1973) and Blinder (1973) for more details].

11. We use Blaise Melly's publicly provided Stata code for this decomposition. For more details on this code, see http://www.econ.brown.edu/fac/Blaise_Melly/code_counter.html and http://www.econ.brown.edu/fac/Blaise_Melly/code_rqdeco.html

12. It is worth noting that the table omits the numerous year and country fixed effects accompanying the income coefficients.

13. As the income variable (in level) for the 0.75 quantile is marginally insignificant at the 10% level, this therefore makes the estimated relationship insignificant.

14. Precisely, the income and incomesq variables are marginally insignificant at the 10% level for the 75th percentile. Also, theoretically, quantile curves are not expected to cross each other. However, it is not unusual to have crossing curves in an empirical application of the quantile regression method but the crossings should not be too many (Koenker, 2005). In the Latin-American sample, it is quite inconceivable to avoid these crossings especially in the presence of both the monotonic and inverted U-shaped relationships found in this sample.

15. The paper bootstraps the procedure a 100 times for computing bootstrapped standard errors.