3.1 The model

This paper adopts an extension of the basic model of Grossman and Krueger (Reference Grossman and Krueger1995), which examines the relationship between the level of development and pollution indicators. In its canonical form, the model establishes an empirical and non-linear logarithmic relationship between a pollution indicator and GDP per capita (Stern, Reference Stern2004). Beyond the squared term that captures the nonlinear relation between pollution and GDP per capita (EKC hypothesis), Grossman and Krueger (Reference Grossman and Krueger1995) include a cubic term to test N-shape EKC (Stern, Reference Stern2004). This ‘recoupling’ effect captures the resumption (overlapping effect) of pollution with the increase in per capita income after a certain threshold. Depending on the objective of the research, the basic specification of the model can be improved through the inclusion of several variables. However the focus of this paper is on the role of institutions, captured by two groups of variables, namely the political regime $\textrm{(PR)}$ and the six governance indicators $\textrm{(Gov)}$ of the World Bank.

Adopting a panel data specification and including additional control variables, the estimated model is given as follows:

\begin{align*} \textrm{Log}({\textrm{Pol}{\textrm{l}_{it}}} )& = {\gamma _0} + {\gamma _1}\textrm{Log}({{y_{it}}} )+ {\gamma _2}{(\textrm{Log}{y_{it}})^2} + {\gamma _3}\textrm{P}{\textrm{R}_{it}}\nonumber\\ & \quad + {\gamma _4}\textrm{Go}{\textrm{v}_{it}} + {\gamma _5}{X_{it}} + {\mu _i} + {\varphi _t} + {\varepsilon _{it}}.\end{align*}

$\textrm{Poll}$ is the indicator of environmental degradation (CO₂ metric tons per capita); y is real per capita income; $\textrm{PR}$ is the political regime; $\textrm{Gov}$ captures governance indicators; X is a set of control variables such as the level of industrialization, natural resource rents, energy intensity, digitalization, and trade openness; ${\mu _i}$ is the country's fixed effects; ${\varphi _t}$ is the time fixed effects and $\varepsilon$ is the error term.

3.2 Variables and data

This paper uses three main types of variables (environmental, economic and institutional). The dependent variable is the log of CO₂ emissions. The emissions related to CO₂ are those stemming from the burning of fossil fuels and the manufacturing of cement. They include CO₂ produced during the consumption of solid, liquid, and gas fuels, and gas flaring, respectively. CO₂ emissions are considered in recent studies as a relevant measure of environmental quality (Marques and Caetano, Reference Marques and Caetano2022; Nchofoung and Asongu, Reference Nchofoung and Asongu2022). Other pollution variables such as methane, nitrous oxide, fine particulate matter (PM_2.5), GHGs, and environmental footprint are reserved for robustness.

The GDP per capita is one of the key predictor variables of our model. According to the World Bank (WDI, 2023), GDP per capita is calculated as gross domestic product divided by midyear population. GDP is the sum of gross value added by all resident producers in the economy, plus any product taxes, and minus any subsidies not included in the value of the products. The calculation supposes no deductions for the depreciation of fabricated assets or the depletion and degradation of natural resources. Data are in constant 2010 US$.

Institutional indicators are considered as variables of interest in this paper. They are relative to the political regime (Polity2) and the World Bank's six indicators of governance, namely: voice and accountability, political stability and absence of violence/terrorism, government effectiveness, regulatory quality, rule of law, and control of corruption. The variable Polity2 ranges from −10 (autocratic regime) to +10 (democratic regime), and the governance indicators are between −2.5 (very bad quality) and +2.5 (very good quality). We expect a negative sign for the coefficients associated with institutions. Also, we generate a governance index by simple averaging, given the similarity of the dimensionality of the six indicators. This homogeneity in their basic characteristics guarantees the robustness of the index obtained. For robustness purposes, we will use some political risk indicators of the International Country Risk Guide (ICRG).

Additional control variables were included to ensure model parsimony and to further control for variables' omission bias. For this purpose, we consider industrialization, measured as industrial value added as a proportion of GDP. Following Cherniwchan (Reference Cherniwchan2012), this variable captures the pace of structural transformation of the economy, whether this transformation is polluting or not. Also, the consideration of natural resources in the explanation of the dynamics of polluting emissions makes it possible to determine whether the rent which results from it can set up anti-pollution measures, but in particular to see if the exploitation of natural capital with regard to the equipment used is harmful to the environment.

The third variable is energy intensity, which is the ratio of a country's energy consumption to its GDP. In other words, concerning the environment, the amount of energy needed to produce a unit of product can be harmful to the environment (Li et al., Reference Li, Qiao, Li and Wang2022) if the mix does not promote renewable energy. In this context, a significant weight of non-renewable energies in the energy mix would be a determining factor in polluting emissions. The fourth additional control variable is digitalization, measured as the rate of internet penetration per 100 people. Following Avom et al. (Reference Avom, Nkengfack, Fotio and Totouom2020) and Evans and Mesagan (Reference Evans and Mesagan2022) among others, this variable allows us to understand how digitization influences pollution by reducing transactions and factor movements. One would expect a negative sign.

Finally, trade openness (ratio of exports and imports to GDP) is considered to test the pollution haven hypothesis. It affects the level of pollution through three channels materialized by three effects (Grossman and Krueger, Reference Grossman and Krueger1995; Copeland and Taylor, Reference Copeland and Taylor2005). The first is the scale effect, due to a positive association between income and pollution levels. The second effect is described as a composition effect, arising from the contribution of the respective sectors of the economy to pollution. The last effect, the technical effect, explains the level of pollution by the polluting quality or not of the production equipment. These three effects make the relationship between trade openness and pollution ambiguous.

The data comes from three main sources: the World Bank (WDI, 2023) for environmental and macroeconomic variables, Polity IV of the Centre for Systemic Peace for the democratic regime, and the Worldwide Governance Indicators for governance indicators. Other data sources such as the South Pacific Applied Geoscience Commission, the United Nations Environment Program, the Global Footprint Network, Varieties of Democracy (V-Dem), and the ICRG were used for sensitivity and robustness. The sample covers 50 African countriesFootnote ¹ over the period 1996–2020. The choice of this period is constrained by the availability of data on the World Bank's governance indicators.

3.3 Identification strategy and estimation technique

Four main empirical critiques related to heteroscedasticity, simultaneity, omitted variables bias, and cointegration are addressed in the modeling of pollution and its main determinants (Stern, Reference Stern2004). However, most studies failed to prove the existence of heteroscedasticity. Moreover, the question of cointegration is not a matter of urgency since we use panel data. Following Stern (Reference Stern2004) and Cole (Reference Cole2004), there remains the issue of endogeneity, which requires the use of instrumental variable techniques. In this paper, we use the generalized method of moments (GMM) estimator through a dynamic panel specification to capture the lagged effect of the emission on its current level and to efficiently correct the endogeneity bias. According to Baum et al. (Reference Baum, Schaffer and Stillman2003), this technique is more appropriate in the presence of an unknown form of heteroscedasticity.

One can attest to a bi-directional link between the level of emissions and some of its determinants (GDP per capita, industrialization, etc.), just as some variables can be measured with errors (institutions, etc.). In this context, we could suspect an endogeneity bias. Indeed, in a dynamic panel model, the countries' unobservable specific effects are correlated with the lagged dependent variable, which provides inconsistent estimators under ordinary least squares (OLS). However, as in Nchofoung and Asongu (Reference Nchofoung and Asongu2022), the GMM estimator becomes consistent under this condition. Using the lagged values of the first difference of the endogenous variable as instruments, Arellano and Bond (Reference Arellano and Bond1991) developed a consistent estimator, called the difference GMM estimator. However, Blundell and Bond (Reference Blundell and Bond1998) demonstrated that when the dependent variable is persistent over time, lagged values are very poor instruments. Using additional conditions of moment, they proposed a more robust alternative estimator called the system GMM estimator from a system of two equations, one in level and the other in first difference.

Several other arguments can be used to justify the GMM estimator. First, our specification obeys Roodman's (Reference Roodman2009) condition that the number of countries (50) must be much larger than the number of periods (25). Second, the inertial effect of CO₂ is high, as the correlation between the dependent variable and its lagged value is strongly significant; this leads us to retain the lagged CO₂ as an important determinant of current emissions. Third, the GMM estimator is biased when the estimation strategy imposes too many instruments. To overcome this problem, we follow the approach of Arellano and Bover (Reference Arellano and Bover1995) extended by Roodman (Reference Roodman2009), which consists of limiting the number of instruments and maximizing the sample size by using the forward orthogonal deviation technique. Also, since the one-step GMM estimator retains a heteroskedastic error structure, we apply the two-step estimator to correct this bias. Finally, our identification strategy poses problems related to the simultaneity of the variables and to the set of restrictions to be observed. Faced with these problems, the recent literature (see Nchofoung and Asongu, Reference Nchofoung and Asongu2022) advises considering all the explanatory variables as a potential source of endogeneity, with time-fixed effects having made it possible to instrument the equations in level and in difference.

For this purpose, the set of equations that generates our system GMM estimator is:

\begin{align*} \textrm{Log}(\textrm{Pol}{\textrm{l}_{it}}) & ={\gamma _0} + \alpha \textrm{Log}(\textrm{Pol}{\textrm{l}_{it - 1}}) + {\gamma _1}\textrm{Log}({y_{it}}) + {\gamma _2}{(\textrm{Log}{y_{it}})^2} + {\gamma _3}\textrm{P}{\textrm{R}_{it}}\\ & \quad + {\gamma _4}\textrm{Go}{\textrm{v}_{it}} + {\gamma _5}{X_{it}} + {\mu _i} + {\varphi _t} + {\varepsilon _{it}}\\ \textrm{Log}(\textrm{Pol}{\textrm{l}_{it}}) & - \textrm{Log}(\textrm{Pol}{\textrm{l}_{it - 1}}) + \alpha [\textrm{Log}(\textrm{Pol}{\textrm{l}_{it - 1}}) - \textrm{Log}(\textrm{Pol}{\textrm{l}_{it - 2}})]\\ & \quad + {\gamma _1}[\textrm{Log}({y_{it}}) - \textrm{Log}({y_{it}} - 1)] + {\gamma _2}[{(\textrm{Log}{y_{it}})^2} - {(\textrm{Log}{y_{it - \textrm{1}}})^2}]\\ & \quad + {\gamma _3}[\textrm{P}{\textrm{R}_{it}} - \textrm{P}{\textrm{R}_{it - 1}}] + {\gamma _4}[\textrm{Go}{\textrm{v}_{it}} - \textrm{Go}{\textrm{v}_{it - 1}}] + {\gamma _5}[{X_{it}} - {X_{it - 1}}]\\ & \quad + ({\varphi _t} - {\varphi _{t - 1}}) + ({\varepsilon _{it}} - {\varepsilon _{it - 1}}). \end{align*}

GMM is a generalization of the method of moments (MM) that allows for more moment conditions than there are parameters to estimate. This makes GMM more efficient than MM, as it can use more information to estimate the parameters. In addition to being more efficient than MM, GMM has several other advantages, including: (i) robustness to misspecification: GMM is robust to misspecification of the model's distribution. This means that the estimates will still be consistent even if the model is not exactly specified; (ii) ability to handle overidentifying restrictions: GMM can handle overidentifying restrictions, which are additional restrictions on the parameters of the model. These restrictions can be used to test the validity of the model and to improve the efficiency of the estimates; and (iii) ability to handle non-stationary data: GMM can handle non-stationary variables in mean and variance over time. This makes it a useful tool for analyzing time series data.

3.4 Data main characteristics

The validation of estimates must be preceded by a battery of tests applicable to panel data. These statistics/tests give us a glimpse of the main characteristics of the variables, their level of correlation and collinearity, and above all their cross-sectional dependence and stationarity.

As for the main statistical characteristics of the variables selected for our study, two main trends emerge, based on the proportionality of the first two moments of their distributions (table A1, online appendix). Thus, we observe variables clustered around their means (GDP, manufacturing value added for industrialization, energy intensity and trade openness), characterized by their under-dispersion. The other variables (natural resource rents and institutional indicators) are over-dispersed, with their standard deviations being greater than their means.

Then, before proceeding with the econometric analysis, it is important to consider the level of bilateral correlation between the variables in order to anticipate the expected signs of the different relationships. The correlation matrix (table A2, online appendix) shows that pollutant emissions are positively correlated with real GDP per capita, industrialization, natural resource exploitation, internet connectivity, and trade openness. In general, institutional indicators are negatively correlated with pollution levels. This bivariate analysis tends to reinforce the intuition that better institutions constrain polluting emissions. Indeed, such institutions would be governed by, among other things, a democratic political regime, low levels of corruption, political stability combined with an absence of violence, the rule of law, good regulatory quality, good representativeness of the people (voice and citizenship), and efficient government.

However, the high level of some correlation coefficients could induce a suspicion of multicollinearity. To verify this, we conduct the variance inflation factor (VIF) test. This test reveals that our variables are not strongly correlated with all VIF statistics being below 10 (table A3, online appendix), the threshold at which multicollinearity becomes critical. For this test, we retained several specifications depending on the type of governance indicator and according to the average level of governance using an aggregate index.

Finally, we check for the cross-sectional dependence of our series and we implement the related unit root tests. In general, the robustness of the results of a GMM estimation is enhanced in the presence of stationary variables. However, stationarity tests in panel data are conditioned by the presence or absence of cross-sectional dependence between countries (Pesaran, Reference Pesaran2015, Reference Pesaran2021), which first requires carrying out this test and implementing the unit root test(s) controlling for this dependence. Indeed, by ignoring the cross-sectional dependence between the variables used, most of the unit root tests conclude that the variables are stationary (Banerjee et al., Reference Banerjee, Marcellino and Osbat2004), which is not always the case in reality.

Apart from governance indicators, our results stipulate that the economic structures of the countries in our sample are strongly linked or integrated. In other words, real shocks (GDP), pollution shocks, natural resource price shocks, energy shocks, digital transformation, trade shocks and (non-) democratic waves, could easily be transmitted from one African economy to another, through more or less complex mechanisms. In other words, the dependence is strong for CO₂ emissions, real GDP per capita, natural resources, energy intensity, the level of internet connectivity, trade openness and the political regime, but weak for most governance indicators (table A4, online appendix).

Under these conditions, the second-generation panel data unit root tests effectively controlling for cross-sectional dependence are indicated. As a reminder, the first-generation tests come up against two identical problems, namely the heterogeneous nature of the unit root and the failure to consider inter-individual independence. We use two tests in this paper, namely the test of Breitung and Das (Reference Breitung and Das2005) and the test of Pesaran (Reference Pesaran2007). The results validate the stationarity of the majority of the pooled series (table A4, online appendix), which does not weaken our empirical strategy (the non-stationary variables are significantly stationary in first difference).

5.1 Sensitivity to institutional characteristics

The institutional variables modeled in this paper are designed on a scale that denotes their good and bad quality. For both the political regime and political governance indicators, it is possible to identify negative values (describing bad quality) and positive values (describing good quality). Thus, distinguishing them according to the identified signs, we control sequentially for the political regime and the overall governance index.

First, the definition of the political regime is made in a continuum integrating the two most extreme forms, which are autocracy and democracy. To do this, we distinguish, according to the typology proposed by Marshall and Elzinga-Marshall (Reference Marshall and Elzinga-Marshall2017), three categories of political regimes, namely autocracies (−10 ≤ political score ≤ −6), anocracies (−5 ≤ political score ≤ +5), and democracies (+6 ≤ political score ≤ +10). The results obtained from this categorization are reported in table 2 (see also online appendix table A5). Based on the three samples, it appears that countries with democratizing regimes are the ones that best deal with environmental pollution. In anocratic countries, the effect is certainly relevant, but it is weaker and less significant. Finally, autocratic regimes seem to be pollution havens (pollution attractors), although the effect remains insignificant. Moreover, in all the specifications, we detect the existence of a globally significant EKC.

Table 2. The political regime characteristics on CO₂ emissions

Note: Standard errors in parentheses.

Second, in terms of governance quality, we distinguish two modalities, namely poor governance (negative aggregated index) and good governance (positive aggregated index). The results obtained in table 3, with and without taking into account the additional control variables, broadly validate those previously established (see table A6, online appendix).

Table 3. The governance index characteristics on CO₂ emissions

Note: Standard errors in parentheses.

First, good quality governance strengthens the quality of institutions, which contributes to reducing ecological dumping. Thus, better quality institutions help to strengthen the pollution halo. Second, the pollution haven hypothesis is globally verified in a weak institutional context, which favors ecological dumping. The results also broadly validate the existence of an EKC.

5.2 Sensitivity to sub-regional heterogeneity

The second sensitivity test focuses on sub-regional samples, that is, the base model is specified for each sub-region. This specification violates the Roodman (Reference Roodman2009) condition of proportionality between the number of countries and the number of annual periods. To address this, we use the pooled OLS. This estimator (table 4) globally validates the existence of an EKC in Eastern, Central and Northern Africa (see table A7, online appendix).

Table 4. Subregional heterogeneity (pooled OLS estimates)

Note: Robust standard errors in parentheses.

The analysis also shows that the improvement of democratic conditions is favorable to the reduction of polluting emissions in all sub-regions, with the effect being more pronounced in Central, Southern and Western Africa. Also, improving the governance of political institutions appears to be a brake on pollution, although the effect remains insignificant in East Africa (bivariate specification) and West Africa (augmented specification).

5.3 Sensitivity to environmental vulnerability index

We aim to test whether the exposure of countries to certain environmental constraints influences their tolerance for pollution. For this purpose, we use the Environmental Vulnerability Index (EVI), which attempts to characterize the relative severity of various types of environmental problems. This index is used to focus on expected solutions to negative pressures on the environment. It also measures environmental sustainability. The index is calculated based on several factors, including exposure, sensitivity, and adaptive capacity.

The EVI is a composite index obtained from 50 indicators and can be subdivided into three sub-indices relating to hazards, resistance, and damage. The 50 indicators can also be broken down into climate change, biodiversity, water, agriculture and fisheries, human health aspects, desertification, and exposure to natural disasters. The EVI is a useful tool for policymakers and researchers to assess the environmental vulnerability of countries and to identify areas where interventions are needed to reduce vulnerability. Measured on a scale of 1 (high resilience / low vulnerability) to 7 (low resilience / high vulnerability), five categories of countries are encountered: resilient, at-risk, vulnerable, highly vulnerable and extremely vulnerable countries. The last category of countries is absent in our sample.

Applying pooled OLS to avoid violation of Roodman's (Reference Roodman2009) ‘large N and small T’ condition, we obtain results according to the bivariate and multivariate specifications in table 5 (see also table A8 in the online appendix). These results show that in vulnerable and highly vulnerable countries, environmental degradation is not sensitive to the quality of institutions. On the other hand, in resilient and at-risk countries, good institutional quality contributes significantly to ensuring a healthy environment. The effect is much more dependent on the political regime than on good governance indicators. Everything being equal, democratizing countries gives more freedom of speech, which in turn gives more space to environmental lobbies, whose daily actions help to protect the environment.

Table 5. Controlling for Environmental Vulnerability Index (pooled OLS estimates)

Note: Robust standard errors in parentheses.

6.1 The inertial effect of CO₂ on results

The effects of CO₂ can be persistent, i.e., CO₂ has a hysteresis effect, marked by a long or short memory. Thus, estimating the length of this memory effect or auto-regression effect yields important information on the stability of results and the forecasting of economic policies. Technically, we estimate a set of AR(p)-X panel models, i.e., auto-regressive panel models with exogenous variables of interest and control. This intuition is taken from Acemoglu et al. (Reference Acemoglu, Naidu, Restrepo and Robinson2019). In each of the specifications concerning lags, we retain two types of models, namely the basic model (which considers only GDP, its square and institutions) and the augmented model (which incorporates the control variables).

The results in table 6 show that the memory effect fades overall after the 4^th-order lag. In other words, the CO₂ dynamics have an inertia of four years at most in the basic model. In the model augmented with control variables, this effect is reduced to two years, making explicit the significance of dynamic interactions between past CO₂ values and its other predictors. Furthermore, the EKC hypothesis is preserved, and the good quality of institutions continues to be a powerful predictor of pollution reduction (see table A9 in the online appendix).

Table 6. CO₂ persistence effect

Note: Standard errors in parentheses.

6.2 Alternative indicators of pollution

This test is conducted in two sequences, namely the mobilization of certain components of GHGs and fine particles for the first, and the ecological footprint for the second. For the first sequence, the variables specifically used are methane (metric tons per capita in log), nitrous oxide (in log), fine particulate matter (PM_2.5), and total GHGs, whose use in this paper allows us to understand the impact of global pollution on global warming (Mignamissi and Djeufack, Reference Mignamissi and Djeufack2022). Applying the system GMM estimator, the results (table 7) validate the existence of an EKC for nitrous oxide and PM_2.5. In all these specifications, CO₂ pollution is significantly sensitive to the good quality of institutions (high political score and good level of governance on average) in the sample (see table A10, online appendix).

Table 7. Robustness checks with pollution indicators

Note: Standard errors in parentheses.

In the second sequence, it should be noted that the relationship between institutional quality and environmental footprint has rarely been analyzed. Following Uzar (Reference Uzar2021), we use the total ecological footprintFootnote ² and its different dimensions (built-up land, carbon, cropland, fishing grounds, forest products, and grazing land). Both are measured in global hectares and are extracted from the Global Footprint Network database. Using the average governance indicator, we obtain the results reported in table 8. They show that the negative association between good institutional quality and environmental footprint contributes to ecological sustainability in Africa (Uzar, Reference Uzar2021). Furthermore, the EKC hypothesis is globally confirmed (see table A11, online appendix).

Table 8. Institutions and ecological footprint

Note: Standard errors in parentheses.

6.3 Alternative institutional measures

The measurement of institutions is diverse and varied in the literature. In this robustness test, we use the V-Dem database to measure corruption according to its various sources (public sector corruption, executive corruption, and political corruption). We also use indicators such as government stability, military in politics, law and order, democratic accountability and bureaucracy from the ICRG. The effect of the indicators selected remains consistent with the general intuition (table 9).

Table 9. Alternative institutional measures and pollution in Africa

Note: Standard errors in parentheses.

A stable government ensures the sustainability of its institutions and therefore takes socio-economic measures that improve the living conditions of its people. Such a government is associated with the implementation of policies that promote sustainable development, which makes it possible to design effective strategies to combat environmental pollution. Moreover, more corruption weakens the functioning of the State through the development of rent-seeking behaviors, which prevents the design of inclusive social policies to the benefit of self-interest. In this context, large foreign polluting firms easily find havens to export their pollution. This result is significant for all measures of corruption, depending on its origin (see online appendix table A12).

Also, the analysis shows that the implementation of a hardline and interventionist regime (with more military) at all levels tends to reduce economic activity and, in general, discourage the action of large polluting firms that do not have a secure guarantee of their property rights. In this context, more military in politics could in the short term slow down polluting activities even if, in the long term, the search for credibility makes military governments relatively conciliatory, which in turn avoids the uprising of populations that have suffered restrictions since they came to power. The effect would therefore not be monotonous over time.

The variable ‘rule of law and order' tends to constrain polluting activities. In a regime that has already adopted environmental legislation, the coercive nature of the legislation and the compliance of the people with it reinforces its effect on environmental protection, which tends to reduce the level of pollution. However, this effect was not significant in our sample.

This tendency to control pollution is further enhanced when the government feels democratically committed, responsible, and accountable. Indeed, most governments are established based on their political, social, economic, and environmental commitments. Thus, their reappointment to power requires that they send signals about their accountability. In this context, they tend to promote activities in favor of the environment and impose pacts on firms concerning their social responsibility. However, the results show that democratic accountability in the majority of African states is not sufficient to significantly bend the pollution curve. Finally, a good quality bureaucracy survives successive governments, which allows the continuance of contracts related to environmental balance by avoiding renegotiations according to the system in place.