A panel data heterogeneous Bayesian estimation of environmental Kuznets curves for CO₂ emissions

Abstract

This article investigates the Environmental Kuznets Curves (EKC) for CO₂ emissions in a panel of 109 countries during the period 1959 to 2001. The length of the series makes the application of a heterogeneous estimator suitable from an econometric point of view. The results, based on the hierarchical Bayes estimator, show that different EKC dynamics are associated with the different sub-samples of countries considered. On average, more industrialized countries show evidence of EKC in quadratic specifications, which nevertheless are probably evolving into an N-shape based on their cubic specification. Nevertheless, it is worth noting that the EU, and not the Umbrella Group led by US, has been driving currently observed EKC-like shapes. The latter is associated to monotonic income–CO₂ dynamics. The EU shows a clear EKC shape. Evidence for less-developed countries consistently shows that CO₂ emissions rise positively with income, though there are some signs of an EKC. Analyses of future performance, nevertheless, favour quadratic specifications, thus supporting EKC evidence for wealthier countries and non-EKC shapes for industrializing regions.

Notes

¹ Waste, which is a very different externality with respect to impacts and local dimension, is the only pollutant other than CO₂ where there is a lack of robust evidence in favour of absolute delinking (Wang et al., 1998; Mazzanti and Zoboli, 2005; Mazzanti, 2008). There is some recent evidence of EKC trends in waste generation (Mazzanti et al., 2007, 2008).

² A recent seminal paper by Copeland and Taylor (2004) surveys the literature and presents a model in which sources of growth, increasing returns to abatement, income and threshold effects are the main drivers of EKC.

³ They also pointed out that for some pollutants, such as CO₂, the lack of homogeneity is not a surprising outcome given the trends in international specialization, differences in local features and absence of strongly coordinated policies at least at the international level.

⁴ We argue that this sub-division is useful for deriving policy conclusion in the CO₂ policy arena, where the reasoning mainly revolves around the role played by different areas according to their environmental and developmental path.

⁵ Although within the slightly different framework of policy evaluation, we note the work by Brock et al. (2003), who proposed model averaging methods as a statistical procedure to tackle model uncertainty. Within this reasoning, based on averaging models using a formalized statistical procedure rather than informal methods, the Bayesian paradigm and statistics (versus Waldean and frequentist) plays a crucial role. Since EKC issues are within the broader realm of empirical macroeconomics and are to some extent linked to policy evaluation studies, this type of reasoning has relevancy even within this framework. First, model uncertainty, which is discussed at length, is a key pillar of this literature. With or without a formalized theoretical model, the spectrum of empirical models is wide. Within the EKC framework, following the authors’ taxonomy of uncertainty, theory uncertainty (e.g. which empirical model, which nonlinear assumption), specification uncertainty (e.g. nonlinearity, number and content of covariates) and heterogeneity uncertainty (data sources, time span and statistical units) as inter-related and overlapping concepts, are relevant. Second, more and more studies are analysing the extent to which policies modify the endogenous EKC dynamic (reducing the turning point income level and/or the environmental indicator peak).

⁶ The inverted U-shape is shown only for the range of values not constrained to the current observed GDP values. It is just a possibility for a future time.

⁷ Among others, see Bohringer and Loschel (2003), Lindholt (2005) and de Brauw (2006) for policy-oriented empirical analyses on Kyoto protocol issues recently published in this journal.

⁸ See Fig. 1 for fitted and real values of the cubic specifications.

⁹ Even preliminary analysis from scatter plots clearly shows that heterogeneity concerning EKC trends is a key issue. We also argue that the scatter plots show that a high value added may derive from analyses based on country level data, possibly exploiting geographic/economic within-country heterogeneity. See, as rare examples, List and Gallet (1999) and Managi (2006) for EKC frameworks, and Kim (2004) in the original Kuznets curves literature. The simple but useful scatter plot investigation highlights that, at least in panels with long time series, the cross country heterogeneity is an even more crucial issue. Dynamic trends could differ sharply from country-to-country, leading to the (here) often stressed necessity of using either heterogeneous estimators, or country-specific time series/panel datasets.

¹⁰ EKC trends in the non-OECD countries have been, and will be more and more in the future, driven by fast growing and high energy consuming countries, such as India and China. Meuniè (2004) exploited data for the 30 Chinese regions for 1990–1999, and for CO₂ found some initial evidence in favour of an EKC. The peaks are quite sensitive to the specification used, ranging from Yuan 2900 to Yuan 8500 per capita in 1995.

¹¹ We decided to present both the quadratic and cubic specifications (although the robustness of the latter may make any comments on the former irrelevant) to show the consequential estimation procedures related to testing the usual EKC hypothesis against the relatively new N-shape modified hypothesis.

¹² For a recent example of (mixed) EKC evidence on CO₂ for a sample of 84 countries and 40 years, see Kahutu (2006).

¹³ Although these less than clear cut results have many possible interpretations, we argue that they may be evidence that the sub-sample analyses are more relevant. For these, the quadratic specification is always preferred.

¹⁴ Even this outcome is open to interpretation; given that the literature does not indicate an ‘optional’ rule for defining a proper forecast length, this result may mean that the loss of 5 years in the estimation process is relevant for achieving a good forecast. The 1-year forecast is the proper framework of assessment in this case. Scatter plots (Fig. 2 depicts plots for the EU) nevertheless exclude the fact that the underlying reason for high values in the 5-year case may be due to a preference for a linear specification rather than a quadratic or cubic specification.