Threshold Effects on South Africa’s Renewable Energy–Economic Growth–Carbon Dioxide Emissions Nexus: A Nonlinear Analysis Using Threshold-Switching Dynamic Models

Abstract

The transition of South Africa from coal-dependent energy systems to renewable energy alternatives presents economic and environmental trade-off complexities that require empirical investigation. This study employed threshold-switching dynamic models, NARDL analysis, and threshold Granger causality tests to investigate nonlinear relationships between renewable energy generation, economic growth, and carbon dioxide emissions in South Africa from 1980 to 2023. The threshold-switching dynamic models revealed critical structural breakpoints: a 56.4% renewable energy threshold for carbon dioxide emissions reduction, a 397.9% trade openness threshold for economic growth optimisation, and a 385.32% trade openness threshold for coal consumption transitions. The NARDL bounds test confirmed asymmetric effects in the carbon dioxide emissions and renewable energy relationship. The threshold Granger causality test established significant unidirectional causality from renewable energy to carbon dioxide emissions, economic growth to carbon dioxide emissions, and bidirectional causality between coal consumption and trade openness. However, renewable energy demonstrated no significant causal relationship with economic growth, contradicting traditional growth-led energy hypotheses. This study concluded that South Africa’s energy transition demonstrates distinct regime-dependent characteristics, with renewable energy deployment requiring critical mass thresholds to generate meaningful environmental benefits. The study recommended that optimal trade integration and renewable energy thresholds could fundamentally transform the economy’s carbon intensity while maintaining sustainable growth patterns.

1. Introduction

The transition to renewable energy has become critical for lowering carbon dioxide emissions while maintaining economic growth, especially in emerging market economies that rely heavily on fossil fuels [1]. Mapuru et al. [2] argued that transitioning from fossil fuels to clean energy could be costly for the country. South Africa is a unique case study because it is one of Africa’s largest greenhouse gas emitters and has a coal-dependent economy, with coal-fired power plants accounting for more than 80% of total electricity generation as of 2024 [3]. The country faces the dual challenge of meeting rising energy demands while also meeting climate commitments under the Paris Agreement and achieving its Nationally Determined Contributions (NDCs).

Figure 1 illustrates South Africa’s fundamental energy transition over time, with renewable energy generation increasing from 1.0 TWh in 1980 to 20.1 TWh by 2023. This increase in renewable energy generation capacity reduced carbon dioxide emissions from a peak of 476.0 Mt in 2008 to 425.0 Mt by 2023. Renewable energy generation remained consistently low in the 1980s and 1990s, ranging from 0.1 TWh (1993) to 3.6 TWh (2011), before a transformative acceleration began in 2015. Prior to 2015, renewable energy generation in South Africa was limited due to the country’s historically coal-dependent energy system, which relied on coal for electricity generation. The transformative period began with the Renewable Energy Independent Power Producer Procurement Programme (REIPPPP), which was launched in 2011 but experienced significant growth after 2015 [4]. The program achieved significant cost savings after the initial procurement rounds, with solar PV tariffs dropping by 68% and wind tariffs falling by 42% between the first and fourth bid windows [5]. As a result, by 2023, the programme had received over USD 20 billion in investment across 123 projects, resulting in more than 6200 MW of operational renewable capacity [6].

Carbon dioxide emissions from energy increased steadily from 205.9 Mt in 1980 to their peak in 2008. This period reflected the country’s economic expansion and industrialisation phase. However, the subsequent decline to 425.0 Mt by 2023 was caused by several converging forces. These factors include, but are not limited to, the economic downturn caused by the 2008 global financial crisis, reliability issues with ageing coal fleets, and the replacement of coal-fired generation with renewable energy sources [7]. The relationship between renewable energy growth and carbon dioxide emissions reduction became more apparent after 2015, indicating a clear inverse correlation. Consequently, renewable energy generation tripled from 6.9 TWh in 2015 to 20.1 TWh in 2023. Carbon dioxide emissions fell from 455.3 Mt to 425.0 Mt over the same period. This resulted in a reduction of 30.3 Mt in carbon dioxide emissions.

In addition, the most significant acceleration occurred between 2020 and 2023. Renewable energy generation increased from 14.0 TWh to 20.1 TWh (a 44% increase), while carbon dioxide emissions fell from 448.9 Mt to 425.0 Mt. This acceleration occurred following significant policy changes, such as the elimination of licensing thresholds for embedded generation in 2020 and 2022 [8]. Regulatory barriers to renewable energy deployment were removed [9]. The transition took place during a period of significant energy security challenges; notably, the South African load shedding crisis (a local term for power outages), which cost the economy approximately ZAR 35 billion (South African Rand) between 2007 and 2019 [10]. This has increased the demand for energy solutions and complicated long-term growth and development planning [11]. However, Eskom’s Energy Availability Factor (EAF) increased significantly from approximately 57% in April 2024 to 67.02% by July 2024, indicating a structural shift in fleet performance following years of decline, with the utility aiming for 70% EAF by March 2025 as part of their Generation Recovery Plan (GRP) [12].

Following this background, the challenge of balancing economic growth with environmental sustainability through renewable energy generation poses a complex dilemma for South Africa. The country’s economy, according to Ateba et al. [13], primarily depends on electricity production and consumption, with energy-intensive industries such as mining and manufacturing contributing significantly to Gross Domestic Product (GDP). This dependency creates significant political implications, particularly concerning the coal industry. Moving away from fossil fuel-based energy systems could potentially cause political instability due to entrenched economic and political interests in the coal industry [14]. This includes, but is not limited to, employment concerns in coal-dependent regions and resistance from established energy stakeholders [15]. In addition, the lack of renewable energy infrastructure integrated into the wider electricity grid presents substantial technical and logistical barriers to transitioning toward more sustainable energy sources [1].

Mirzania et al. [14] pointed out that South Africa’s transition to renewable energy appears to be hampered by inadequacies in its national energy infrastructure. The technical challenges are further complicated by capacity constraints within distribution networks, which result in the loss of a significant portion of renewable energy-generated electricity during the transition and distribution processes [14]. The lack of adequate electricity infrastructure creates operational vulnerabilities, as companies suffer equipment damage and production interruptions due to electrical grid irregularities [13]. These infrastructure deficiencies not only impede the physical integration of renewable energy technologies but also create unsustainable economic barriers for the country [16]. Yasmeen and Shah [17] stated that the lack of dependable electrical grid connections reduces the viability and attractiveness of renewable energy investments.

This study addresses three key research questions: (i) Is there a long-term equilibrium relationship (cointegration) among South Africa’s renewable energy generation, economic growth, and carbon dioxide emissions? (ii) What are the threshold effects that determine regime changes in these relationships? (iii) Does renewable energy cause changes in economic growth and carbon dioxide emissions, and what are the directional causalities? Therefore, the study formulated several hypotheses based on the research questions. Firstly, this study hypothesises that South Africa’s renewable energy generation, economic growth, and carbon dioxide emissions converge over time. Secondly, as renewable energy capacity increases, carbon emissions decrease, which has become evident since South Africa’s major policy shifts after 2015. Thirdly, the research also expects renewable energy generation to increase economic growth by creating new investment opportunities, reducing coal infrastructure dependence, and improving energy security, despite transition costs. Fourthly, the study also hypothesises that South Africa’s energy transition has tipping points where these variables’ relationships change significantly around key policy interventions such as licensing barrier elimination between 2020 and 2022. Finally, the study expects to show that renewable energy investments lead to environmental and economic improvements, supporting the claim that South Africa can meet its Paris Agreement climate commitments and NDCs while maintaining economic growth and addressing energy security challenges.

The remainder of this paper is organised as follows: Section 2 reviews the theoretical framework and relevant empirical literature for the study. Section 3 outlines the methodology and data employed in the study. Section 4 provides the empirical findings of the study. Section 5 concludes the study by presenting recommendations for South Africa’s renewable energy transition and outlining future research directions.

2. Theoretical and Empirical Literature Review

This study builds upon three interconnected theoretical frameworks that provide the conceptual foundation for understanding nonlinear relationships between renewable energy, economic growth, and environmental outcomes in emerging market economies. These theoretical frameworks include the Environmental Kuznets Curve (EKC) hypothesis, Energy Transition Theory (ETT), and Green Growth Theory (GGT).

2.1. Environmental Kuznets Curve (EKC) Theory

The EKC hypothesis is the primary theoretical framework for studying the relationship between economic development and environmental degradation. Grossman and Krueger [18] proposed the EKC theory, which suggests an inverted U-shaped relationship in which environmental degradation increases with economic development before decreasing after reaching a certain income threshold. The EKC hypothesis is based on three mechanisms. These are the scale effects (increased economic activity causes pollution), composition effects (structural economic shifts towards cleaner industries), and technique effects (technological improvements reduce pollution intensity) [19]. As economies develop, the relative importance of these effects shifts, resulting in the distinctive inverted U-shaped pattern. The general equation for the EKC hypothesis is expressed as follows:

$$E=\alpha +{\beta }_{1}Y+{\beta }_2{Y}^2+\epsilon$$

(1)

where $E$ denotes environmental degradation (i.e., carbon emissions), and $Y$ represents economic growth. ${\beta }_1,{\beta }_2$ are coefficients, while $\alpha$ and $\epsilon$ represent a constant and error terms, respectively. For the inverted U-shaped pattern to hold, ${\beta }_1>0,$ indicates a positive relationship between pollution and income, while ${\beta }_2<0,$ indicates a negative coefficient on the squared term, resulting in a downward slope after the turning point. Then, the turning point $\left(tp\right)$ occurs at:

$$Y_{tp}=-{\displaystyle \frac{{\beta }_1}{2{\beta }_2}}$$

(2)

However, other versions of the EKC, such as Omay et al. [20], include additional variables in the general Equation (1) to be expressed as follows:

$$E=\alpha +{\beta }_{1}Y+{\beta }_2{Y}^2+{\beta }_{1}X+\epsilon$$

(3)

where $X$ represents additional factors, including, but not limited to, trade openness, coal consumption or quality of the institutions in the country.

Recent research finds mixed evidence for EKC validity in various economic contexts. Zhang et al. [21] validated the inverted U-shaped relationship in China’s high-speed rail effects on pollution. They discovered significantly positive GDP per capita coefficients and negative squared terms. Niu and Yang [22] confirmed EKC patterns in their study of carbon reduction effects from industrial transfer, emphasising declining environmental quality before improving at advanced development stages. However, empirical support varies significantly by region and methodology. He and Lin [23] discovered evidence supporting EKC for energy intensity in China, identifying a threshold value of 0.7670 where increasing income is associated with lower carbon dioxide emissions. Holtz-Eakin and Selden [24] and Bertinelli and Strobl [25] discovered no empirical evidence to support the traditional inverted U-shaped pattern.

The EKC relationship shows significant regional variation. He and Lin [23] discovered that only seven Chinese provinces (primarily developed coastal areas) exceeded the EKC threshold, while the majority of central and western regions remained below it. Udeagha and Breitenbach [26] used dynamic ARDL simulations in South Africa to investigate financial development’s moderating role in environmental quality from 1960 to 2020, confirming EKC validity while demonstrating that financial development promotes environmental sustainability. The EKC hypothesis is criticised for oversimplifying complex relationships and assuming universal patterns [27]. Critics argue that EKC ignores important factors such as globalisation, institutional quality, and energy consumption patterns, as well as pollution relocation from developed to developing countries [28]. These limitations drive our threshold analysis approach, which enables more nuanced regime-dependent relationships.

2.2. Energy Transition Theory

Energy Transition Theory (ETT) provides an analytical framework for understanding the transition from fossil fuel-based systems to sustainable energy alternatives. This theory focuses on the complex interactions of technology, policy, and societal preferences that influence renewable energy adoption patterns. Lin et al. [29] emphasised that ETT investigates the relationships between technological advancement, policy frameworks, and social acceptance that drive renewable energy transitions. Energy transitions necessitate significant structural changes, such as infrastructure development, institutional reforms, changes in consumer behaviour, and business model innovations.

The ETT has been expressed in one form or another by a number of researchers, including Fisch-Romito et al. [30], Akaev and Davydova [31], and Cherp et al. [32]:

$$E_t=f(T_t,P_t,S_t,I_t)$$

(4)

where Equation (4) represents the basic energy transition model. $E_t$ denotes the energy transition level at time $t$. $T_t,P_t,S_t,I_t$ denotes technology factors, policy frameworks, social acceptance or preferences and institutional factors. Equation (4) is transformed to express the dynamic transition Equation (5):

$${\displaystyle \frac{dE}{dt}}=\alpha T\left(t\right)+\beta P\left(t\right)+\gamma S\left(t\right)+\delta I\left(t\right)-\theta E(t)$$

(5)

where ${\displaystyle \frac{dE}{dt}}$ denotes energy transition rate, $\alpha$, $\beta$, $\gamma ,\delta$ are positive coefficients indicating the impact of each variable, while $\theta$ represents the resistance parameter. $\theta E\left(t\right)$ is the negative term that signifies institutional path dependency.

Izadpanahi et al. [33] emphasised that energy transitions require significant capital investment and careful policy planning, especially in manufacturing sectors where energy is a critical operational input. Aziz et al. [34] improved traditional ETT frameworks by incorporating the role of financial technology in facilitating transitions by means of better resource allocation and green finance mechanisms. Energy transitions demonstrate important geographic characteristics with significant governance implications. Davies et al. [35] argue that discussions about renewable energy should prioritise the geographical distribution of renewable infrastructure, which influences power dynamics and governance structures. This geographical perspective enables decentralised and multidisciplinary governance strategies, as demonstrated by South Africa’s REIPPPP implementation.

The multi-level nature of energy transitions is becoming more widely recognised, according to recent research. Tabrizian et al. [36] highlighted the shift from global governance frameworks to subnational, regional, and local energy policy perspectives. This multi-level approach acknowledges that energy transitions vary across geographical and administrative scales. Artificial intelligence (AI) and technological advancement are critical components in energy transitions. Tao et al. [37] showed that AI promotes low-carbon energy transformation by increasing efficiency and upgrading industries, while also causing energy rebound effects that partially offset AI-related benefits. This exemplifies the “Jevons Paradox”, in which efficiency improvements can increase rather than decrease energy consumption [38].

The implementation of energy transition varies greatly across different contexts. Guesmi et al. [39] demonstrated that government climate programs, summits, and taxation have a significant impact on renewable energy transitions in the United States. Sinha et al. [40] show that social inequality reduces the positive effects of energy transition drivers in OECD countries, whereas governance quality increases those effects.

2.3. Green Growth Theory

Green Growth Theory (GGT) is the third theoretical foundation for the study. This theory describes the simultaneous pursuit of economic growth and environmental sustainability. GGT calls into question long-held beliefs that environmental protection must hinder economic growth. GGT proposes that economic systems can be restructured to reduce environmental impact while increasing economic productivity [41]. Hence, the basic green growth decoupling equation can be expressed as follows:

$${\displaystyle \frac{dE}{dY}}<0,\mathrm{w}\mathrm{h}\mathrm{i}\mathrm{l}\mathrm{e}{\displaystyle \frac{dY}{dt}}>0$$

(6)

The basic green growth decoupling Equation (6) shows that environmental impact ($E$) decreases while economic output or GDP ($Y$) increases, indicating absolute decoupling, in which economic growth correlates with environmental improvement. The decoupling equation focuses on the rates of change. Lu and Zhao [42] emphasised the importance of GGT in achieving sustainable economic growth through innovation, renewable energy investment, and energy efficiency.

GGT effectiveness has been studied, but the empirical evidence is mixed. Monawwar et al. [43] proposed targeted demand-side management policies to reduce carbon footprints while maintaining economic growth. Zeng et al. [44] emphasised environmental governance’s critical role in mitigating the negative effects of resource exploitation on green growth. Fan and Wang [45] show that natural resources, renewable energy, financial development, and globalisation all have positive long-term effects on green growth, despite some negative short-term consequences. Guo et al. [46] demonstrated that technological innovation has a significant and positive impact on regional green growth performance, whereas environmental regulation relationships are more sophisticated.

However, there are a few studies that have challenged the GGT. Gazheli et al. [47] found that high carbon-intensity sectors in Denmark, Germany, and Spain grow faster than cleaner sectors, implying either unfeasible decarbonisation requirements or unavoidable economic contraction [48]. Gómez-Baggethun [49] contends that technological optimism about growth–environment decoupling lacks empirical support, concluding that increased growth inevitably leads to environmental destruction, despite efficiency advances.

2.4. Empirical Studies Specific to South Africa

Inglesi-Lotz [11] discovered a positive long-term relationship between renewable energy production and economic growth in South Africa. However, subsequent research by Ndlovu and Inglesi-Lotz [50] discovered that renewable energy has no significant GDP impact in South Africa, despite its strong effects in other BRICS countries. Saba [51] discovered that, while militarisation reduces short-term economic growth, renewable energy generation has a positive and significant long-term impact on South African economic growth. Shikwambana et al. [52] discovered strong positive linear correlations between economic growth and carbon dioxide emissions in South Africa, emphasising the ongoing environmental degradation that occurs concurrently with economic development. This contrasts with findings from developed economies, where studies such as Aslan et al. [53] discovered inverted U-shaped relationships between various carbon dioxide emission components in the United States. Tagwi [54] discovered that carbon dioxide emissions rise with agricultural economic growth in the short run, while renewable energy usage remains insignificant in both the short and long run in South Africa’s agricultural sector.

2.5. Empirical Regional and Comparative Studies

Renewable energy deployment poses unique challenges and opportunities for Africa as a whole. Nathaniel and Iheonu [55] found that renewable energy helps to reduce emissions in 19 African countries, albeit to a limited extent. Inglesi-Lotz and Dogan [56] discovered that renewable energy reduces emissions significantly in Sub-Saharan Africa, whereas Zoundi [57] found strong negative correlations across 25 African countries. Saidi and Omri [58] found that renewable energy boosts economic growth while lowering carbon dioxide emissions in 15 major renewable energy-consuming countries. Chen et al. [59] discovered that non-renewable energy generation increases carbon dioxide emissions, whereas renewable energy helps to reduce emissions. Danish et al. [60] validated the EKC hypothesis for Pakistan, demonstrating that renewable energy plays a dominant role in reducing carbon dioxide emissions, whereas non-renewable energy consumption is the main contributor to increased CO₂ emissions. This finding directly influences our threshold analysis approach for South Africa.

3. Methodology

This section of the study used advanced econometric methods to investigate threshold effects and regime-dependent linkages in South Africa’s renewable energy, economic growth, and carbon dioxide emissions nexus from 1980 to 2023. The study utilised five primary variables, including gross renewable energy generation (REC) measured in terawatts per hour (TWh), carbon dioxide emissions (CDE) from fossil fuels (oil, gas, and coal) measured in million tonnes (Mt), annual GDP growth rate (GDP_growth) as a proxy for economic growth, trade openness as a percentage of GDP, and coal consumption measured in TWh. The annual data was obtained from the World Bank development indicators and Energy Institute (Statistical review of world energy processed by World in Data).

This methodological approach extends previous studies that had used traditional linear econometric methods by incorporating nonlinear threshold dynamics, asymmetric cointegration relationships, and regime-switching mechanisms to better capture South Africa’s sophisticated economic growth and development, as well as energy transition trends. This methodology relies on threshold econometrics [61], asymmetric cointegration [62], and structural break analysis [63]. The methodological approach used in this study is significant for emerging market economies such as South Africa, which has undergone structural changes as a result of political transitions, economic liberalisation, and energy policy reforms over the years [64].

3.1. Theoretical Motivation for the Methodological Approach

Three economic principles serve as the theoretical foundation for the methodology applied in this study. These include threshold effects in economic relationships, asymmetric adjustment mechanisms, and regime-dependent causality patterns, as shown in

Table 1. Theoretical foundation for the methodology applied.

Key Economic Principles	Motivation
Threshold effects in economic relationships	The EKC literature suggests a threshold relationship between economic development and environmental quality [65]. According to [66], the relationship between environmental degradation and income growth is nonlinear and varies by regime. Azam et al. [67] found mixed evidence for Environmental Kuznets Curve validity across MENA countries, with threshold income levels (turning points) varying significantly, from USD 109 per capita in the UAE to USD 458 per capita in Turkey. This had indicated that environmental improvements occur at different economic development stages across the region. Jóźwik et al. [68] found threshold effects in the energy-growth relationship among EU countries, where high levels of non-renewable energy consumption (above the identified threshold) are associated with decreased economic activity, while renewable energy shows consistent positive effects on growth across all regimes during the green transition period.
Asymmetric adjustment mechanisms	Modern growth theory emphasises that positive and negative economic shocks can have distinct effects on long-term relationships [69]. Alper and Oguz [70] found mixed evidence for renewable energy’s impact on economic growth among new EU member countries using asymmetric causality analysis, with positive effects observed across all countries but statistical significance only in Bulgaria, Estonia, Poland, and Slovenia. This asymmetry is particularly pronounced in countries with high resource dependence and infrastructure constraints [71].
Regime-dependent causality patterns	Different economic regimes may have distinct causal relationships between energy consumption, economic growth, and environmental quality [72]. Tiba and Omri [73] found significant differences in the causality patterns between economic growth and renewable energy based on policy regime, institutional quality level, and stage of development. Resource curse dynamics and commodity price cycles amplify regime dependencies in resource-rich countries such as South Africa [74].

Source: Authors’ own assessment.

below.

3.2. Structural Unit Root Testing

The presence of a unit root in macroeconomic time series data has significant implications for econometric modelling and policy analysis. Nelson and Plosser [75] demonstrated that most macroeconomic series demonstrate unit root behaviour. Perron [76] showed that structural breaks can cause traditional unit root tests to incorrectly conclude non-stationarity when a series is actually stationary around a changing deterministic trend. Zhang [77] highlighted that the presence of a unit root indicates that the time series follows stochastic trends, whereby previous shocks have long-term effects on the series, resulting in memory-like characteristics within the data. Then, proper identification of integration orders is critical for selecting appropriate econometric methodologies and ensuring valid statistical inference.

In general, unit root tests are employed under two hypotheses. The first hypothesis, ${\mathrm{H}}_0:\mathsf{\theta }=0$, assumes the presence of a unit root and implies that the data is non-stationary and requires transformation to be used efficiently. The second alternative hypothesis, ${\mathrm{H}}_1:\mathsf{\theta }<0$, proposes the absence of a unit root, implying that the data is already stationary and may be examined without any further changes.

Therefore, the selected variables were tested for stationarity using the Zivot–Andrews (ZA), Narayan–Popp (NP), and Lee–Strazicich (LS) structural unit root tests. The economic rationale emerges from the fact that many macroeconomic series appear non-stationary, which can be attributed to structural changes rather than true unit root behaviour [75]. Failure to account for structural breaks can result in a false acceptance of the unit root hypothesis. This will lead to incorrect conclusions about the persistence of economic shocks [78].

3.2.1. Zivot–Andrews (ZA) Structural Unit Root Test

ZA test takes into account two primary model specifications (models A and C) to test for different types of structural changes. Model A only records the intercept break, whereas model C records both the intercept and trend breaks [79]. The ZA test (model A) consists of estimating the following regression:

$$∆y_t=\mu +\beta t+\theta {DU}_t\left(T_B\right)+\alpha y_{t-1}+{\sum }_{j=1}^k{c}_j∆y_{t-j}+{\epsilon }_t$$

(7)

where ${\mathsf{\epsilon }}_{\mathrm{t}}$ is a white noise error term and $∆y_t$ = ($y_t$ − $y_{t-1}$) is the first difference of the series. $\mu$ is the intercept term, while $\beta t$ is the deterministic trend coefficient. $\alpha$ represents the unit root coefficient and ${DU}_t\left(T_B\right)$ is a dummy variable for a break in intercept. ${\sum }_{j=1}^k{c}_j∆y_{t-j}$ are lagged differences that are used to correct autocorrelation. Model A accounts for a single shift in the series mean level, making it appropriate for modelling events such as policy regime changes that affect the long-term average of economic variables [80].

Additionally, the following regression is estimated using the ZA test (model C):

$$∆y_t=\mu +\beta t+\theta {DU}_t\left(T_B\right)+\gamma \theta {DT}_t\left(T_B\right)+\alpha y_{t-1}+{\sum }_{j=1}^k{c}_j∆y_{t-j}+{\epsilon }_t$$

(8)

where $\gamma$ denotes the magnitude of the trend break and ${DT}_t\left(T_B\right)$ represents the dummy variable for the trend break. Model C incorporates both level and slope variations in the deterministic component, making it appropriate for the analysis of time series that undergo shifts in both mean and growth rate [81]. The break date $T_B$ is not fixed but estimated by searching through possible break points and selecting the one that provides the most robust proof against the unit root hypothesis. This represents the lowest t-statistic for $\alpha$.

3.2.2. Narayan–Popp Structural Unit Root Test

The NP unit root test advances structural break econometrics by allowing for two endogenous structural breaks in both the level and slope of the trend function [82]. This methodology addresses the critical limitations of single-break tests, which are important when analysing economic time series from emerging market economies that have experienced multiple regime changes, economic transitions, and policy reforms [83]. The general specification of the NP unit root test (models AA and CC) allows for a variety of break types, providing flexibility in modelling different structural change scenarios. Equation (9) illustrates model AA, which tests intercept breaks (level shifts), while Equation (10) illustrates model CC, testing two intercept and trend breaks:

$$∆y_t=\alpha +\beta t+{\theta }_1{DU}_{1t}\left(T_{B1}\right)+{\theta }_2{DU}_{2t}\left(T_{B2}\right)+\rho y_{t-1}+{\sum }_{j=1}^k{\delta }_j∆y_{t-j}+{\epsilon }_t$$

(9)

$$∆y_t=\alpha +\beta t+{\theta }_1{DU}_{1t}+{\theta }_2{DU}_{2t}+{\gamma }_1{DT}_{1t}+{\gamma }_2{DT}_{2t}+\rho y_{t-1}+{\sum }_{j=1}^k{\delta }_j∆y_{t-j}+{\epsilon }_t$$

(10)

where ${DU}_{1t}$, ${DU}_{2t}$ denotes intercept break dummies, while ${DT}_{1t}$, ${DT}_{2t}$ denotes trend break dummies. $T_{B1}$, $T_{B2}$ are two break dates that are estimated endogenously by minimising the t-statistic for $\rho =0$. The rationale for allowing two breaks is relevant to South Africa, which has gone through numerous structural changes, including the sanctions era (1970s–1980s), the political transition (1990s), the implementation of inflation targeting (2000), and various external shocks such as the 2008–2009 global financial crisis and, more recently, COVID-19 pandemic.

3.2.3. Lee–Strazicich Structural Unit Root Test

The LS unit root test employs the Lagrange Multiplier (LM) principle to provide unbiased inference in the presence of structural breaks [83]. This methodology overcomes significant limitations in previous structural break unit root tests. One limitation is the spurious rejection problem, which occurs when there are breaks for both the null and alternative hypotheses [84]. The LS unit root test allows for breaks in both the null and alternative hypotheses. This ensures that the null hypothesis is rejected due to stationarity rather than a misunderstanding of the break procedure [83]. In basic terms, Equation (11) expresses the LS general framework:

$$\begin{array}{c}{y}_t={\delta }^{\prime }{Z}_t+x_t·\left(\text{deterministic}\right)\\ x_t={\beta x}_{t-1}+{\epsilon }_t·\left(\text{stochastic}\right)\end{array}$$

(11)

where $Z_t$ denotes the deterministic regressors (trend, break dummies) and ${\delta }^{\prime }$ are coefficients for deterministic terms. $\beta$ denotes an autoregressive root. The LS unit root test includes several model specifications (model A and C) to accommodate different types of structural breaks. For a single structural break at time $T_B$, the exogenous variables are defined as:

Model A, level break-intercept shift:

$$Z_t=[1,t,{DU}_t{]}^{\prime }$$

Model C, level and trend break:

$$Z_t=[1,t,{DU}_t,{{DT}_t]}^{\prime }$$

(12)

Then, Equation (13) specifies the LM test regression:

$$∆y_t={\delta }^{\prime }∆Z_t+\varnothing {\tilde{S}}_{t-1}+{\sum }_{j=1}^k{\gamma }_j∆y_{t-j}+{\mu }_t$$

(13)

where ${\tilde{S}}_t=y_t-{\tilde{\psi }}_x-Z_t^{\prime }\tilde{\delta }$ is the detrended series. $\tilde{\delta }$ denotes coefficients from regressing $y_t$ on $Z_t$ and $\varnothing$ is the key parameter for hypothesis testing. The economic rationale for LM-based unit root testing arises from the recognition that structural breaks in economic time series are frequently enduring characteristics that persist, regardless of whether the series demonstrates unit root behaviour [85]. For example, a permanent shift in government policy, institutional change, or technological advancement may cause permanent changes in economic variables, regardless of their stationarity [86].

3.3. Structural Break Tests

The Chow, Quandt–Andrews, and Bai–Perron tests were used to determine the presence and timing of structural breaks in the selected economic variables. The economic rationale arises from the observation that many macroeconomic time series demonstrate parameter instability over time, which is due to changes in policy regimes, external shocks, or institutional transformations rather than random fluctuations [87]. Disregarding structural breaks may lead to incorrect conclusions about economic relationships and mis-specified econometric models, resulting in biased parameter estimates [63].

The Chow test investigated structural stability at predetermined break points. The test assesses structural stability by comparing the sum of squared residuals from restricted and unrestricted regressions [88]. For a linear regression model with a potential structural break at time $T_B$, the test consists of estimating the following:

restricted model—no break:

$$y_t={\alpha }_0+{\alpha }_{1}t+{\sum }_{j=1}^k{\gamma }_j{X}_{jt}+{\epsilon }_t,t=1,\dots .,T$$

(14)

unrestricted model (with break at ${\mathrm{T}}_{\mathrm{B}}$):

$$y_t={\alpha }_{01}+{\alpha }_{11}t+{\sum }_{j=1}^k{\gamma }_{j1}{X}_{jt}+{\epsilon }_t,pre-break(t\le T_B)$$

(15)

$$y_t={\alpha }_{02}+{\alpha }_{12}t+{\sum }_{j=1}^k{\gamma }_{j2}{X}_{jt}+{\epsilon }_t,post-break(t>T_B)$$

(16)

test statistic:

$$F_{chow}={\displaystyle \frac{({RSS}_{\tau }-{RSS}_{ur})/k}{{RSS}_{ur}/(T-2k)}}$$

(17)

where ${RSS}_{\tau }={\sum }_{t=1}^T{ϵ}_t^{^2}$ denotes restricted residuals, while ${RSS}_{ur}={\sum }_{t=1}^{T_B}{ϵ}_{1t}^{^2}+{\sum }_{t=T_{B+1}}^T{ϵ}_{2t}^{^2}$ denotes unrestricted residuals. $k$ is the number of regressors including intercept and trend. Then, if $F_{chow}>F_{critical}\left(k,T-2k\right)$ we reject $H_0$ (no break) and vice versa.

The Quandt–Andrews test overcomes the limitation of requiring prior knowledge of break timing by identifying all possible break points within a given interval [89]. This methodology applies to emerging market economies where the timing of structural changes is uncertain or there are multiple potential break points [90]. The test recognises that economic systems can undergo gradual transitions rather than abrupt changes, making the precise timing of structural breaks difficult to predict before they occur [78].

The test entails calculating Chow statistics for all possible break points within a trimmed sample range [$T_1,T_2]$ and selecting the maximum value and the estimated break date in Equation (13):

estimated restricted model (no break):

$${RSS}_t(T_a)={\sum }_{t=1}^T{\left[y_t-\widehat{{\alpha }_0}-\widehat{{\alpha }_1}t-{\sum }_{j=1}^k\widehat{{\gamma }_j}{X}_{jt}\right]}^2$$

(18)

estimated unrestricted model (with break at ${\mathrm{T}}_{\mathrm{a}}$):

$${RSS}_1={\sum }_{t=1}^{T_a}{\left[y_t-\widehat{{\alpha }_{01}}-\widehat{{\alpha }_{11}}t-{\sum }_{j=1}^k\widehat{{\gamma }_{j1}}{X}_{jt}\right]}^2,pre\text{-}breakperiod(t\le T_a)$$

(19)

$${RSS}_2={\sum }_{t=T_a+1}^T{\left[y_t-\widehat{{\alpha }_{02}}-\widehat{{\alpha }_{12}}t-{\sum }_{j=1}^k\widehat{{\gamma }_{j2}}{X}_{jt}\right]}^{2}post\text{-}breakperiod(t>T_a)$$

(20)

total unrestricted RSS:

$${RSS}_{ur}\left(T_a\right)={RSS}_1+{RSS}_2$$

(21)

Then Equation (16) calculates the Chow statistic for $T_a$:

$$F_{chow}\left(T_a\right)={\displaystyle \frac{({RSS}_r\left(T_a\right)-{RSS}_{ur}\left(T_a\right))/k}{{RSS}_{ur}\left(T_a\right)/(T-2k)}}$$

(22)

where $k$ represents the number of regressors, including intercept and trend. Equations (23) and (24) determine the most likely break point by calculating the maximum Chow statistic using the estimated break date:

maximum Chow statistic:

$$F_{max}={}_{T_{B\in \left[T_1,T_2\right]}}{}^{max}F_{chow}(T_B),\mathrm{where}{T}_1=\left[0.15T\right],T_2=\left[0.85T\right]$$

(23)

estimated break date:

$$\widehat{T_B}=\arg \begin{array}{c}max\\ T_B\end{array}{F}_{chow}(T_B)$$

(24)

The hypothesis testing procedure entails the comparison of the calculated $F_{max}$ statistic with critical values obtained from the non-standard asymptotic distribution. Based on the null hypothesis of no structural break, the $F_{max}$ statistic adheres to a non-standard distribution, with critical values ($\alpha$, $k$, $T_1,T_2$) dependent on several parameters. These are the 5% significance level $\alpha$, the number of regressors $k$ in the model, and the trimming fractions $T_1$(0.15) and $T_2$ (0.85), which guarantee adequate observations in each sub-sample for reliable parameter estimation. Then, the null hypothesis is rejected when the calculated $F_{max}$ exceeds the critical value, indicating the presence of a structural break at the estimated break date within the specified confidence level.

The Bai–Perron test improves single-breakpoint analysis by allowing multiple structural breaks [63]. This test overcomes the significant limitations of single-break tests in the analysis of economic systems that undergo recurring structural transformations [91]. The test identifies multiple unknown break points in a time series, allowing for different regression parameters between regimes. The general equation of the Bai–Perron test is expressed as follows:

$$y_t={\alpha }_{0i}+{\alpha }_{1i}\mathrm{t}+{\sum }_{j=1}^k{\gamma }_{ji}{X}_{jt}+{\epsilon }_{it},t=T_{i-1}+1,\dots ,T_i$$

(25)

where $T_0$ = 0 and $T_{m+1}$ = T representing the conventional start and end points of the sample period, and $T_1$, $T_2$,…, $T_m$ denoting the m unknown break dates that divide the sample into $m+1$ regimes. Each regime $i$ has its own set of regime-specific parameters, such as ${\alpha }_{0i}$ (the intercept), ${\alpha }_{1i}$ (the trend coefficient), and ${\gamma }_{ji}$ (the slope parameter). ${\epsilon }_{it}$ denotes independently and identically distributed error terms that adhere to standard assumptions of zero mean, constant variance, and absence of serial correlation within each regime. Equation (25) is followed by estimating the break dates, which are chosen to minimise the sum of squared residuals (SSR) across all regimes, as shown in Equation (26):

$$\left(\widehat{T_1},\dots ,\widehat{T_m}\right)=arg\underset{T_1,\dots ,T_m}{\min }{S}_T\left(T_{1,\dots ,}{T}_m\right)$$

where:

$$S_T\left(T_{1,\dots ,}{T}_m\right)={\sum }_{i=1}^{m+1}{\sum }_{t=T_{i-1}+1}^{T_i}{\left[y_t=\widehat{{\alpha }_{0i}}-\widehat{{\alpha }_{1i}}t-{\sum }_{j=1}^k{\widehat{\gamma }}_{ji}{X}_{jt}+\right]}^2$$

(26)

The Bai–Perron test establishes critical constraints to ensure the statistical validity and reliability of the estimation process. It requires that each regime includes at least $h\ge k+2$ observations for meaningful parameter estimation, where $k$ denotes the number of explanatory variables. The additional two observations account for the intercept and trend parameters. The minimum regime length $h$ is typically set to $\left[0.15T\right]$, representing 15% of the total sample size $T$, to ensure sufficient degrees of freedom for robust least squares estimation within each regime while preserving adequate power for detecting structural breaks throughout the sample period. The parameters for each regime $i$ are estimated using Ordinary Least Squares (OLS) in accordance with Equation (27):

$${\widehat{\theta }}_i=(X_i^{\prime }{X}_i{)}^{-1}{X}_i^{\prime }{y}_i$$

(27)

where ${\widehat{\theta }}_i=[\widehat{{\alpha }_{0i}},\widehat{{\alpha }_{1i}},\dots ,{\widehat{\gamma }}_{ji}{]}^{\prime }$ denotes the vector of estimated parameters, which includes the regime-specific intercept, trend coefficient, and all slope parameters for the explanatory variables. The design matrix $X_i$ is defined as the ($T_i-T_i)\times (k+2)$ matrix, incorporating the constant term, time trend, and all explanatory variables for regime $i$. The vector $y_i$ represents the corresponding $\left(T_i-T_{i-1}\right)\times 1$ observations of the dependent variable within regime $i,$ thus providing the necessary specification for deriving regime-specific parameter estimates through OLS estimation. Next on the Bai–Perron test is to determine the number of breaks as highlighted in Equation (28):

$$F\left(\mathcal{l}+1|\mathcal{l}\right)={\displaystyle \frac{T-(\mathcal{l}+1)(k+2)}{k+2}}.{\displaystyle \frac{S_T\left({\widehat{T}}_1^{\left(\mathcal{l}\right)},\dots ,{\widehat{T}}_{\mathcal{l}}^{\left(\mathcal{l}\right)}\right)-S_T\left({\widehat{T}}_1^{\left(\mathcal{l}+1\right)},\dots ,{\widehat{T}}_{\mathcal{l}+1}^{\left(\mathcal{l}+1\right)}\right)}{S_T\left({\widehat{T}}_1^{\left(\mathcal{l}+1\right)},\dots ,{\widehat{T}}_{\mathcal{l}+1}^{\left(\mathcal{l}+1\right)}\right)}}$$

(28)

The anticipated breaking points for all break tests in the study include, but are not limited to, the 1994 democratic transition, which significantly altered economic institutions and policy frameworks [92]. Beginning in 1995, trade liberalisation brought the economy into line with international markets [93]. Inflation targeting was implemented in the early 2000s, changing the way monetary policy was applied [94]. The 2007/08 global financial crisis highlighted vulnerabilities in emerging market economies [95]. Renewable energy policies were implemented in 2011, resulting in the formation of new economic sectors and shifting energy market dynamics [5].

3.4. Applied Empirical Approach

The empirical methodology included the threshold-switching dynamic models, the NARDL model, and the threshold Granger causality test. To adhere to the accepted econometric methods for time series analysis, all selected variables were converted to logarithmic values. This was performed in accordance with standard econometric practices [96]. This log transformation is critical in econometric analysis for a number of reasons. For example, Franses [97] claims that the logarithmic transformation reduces the impact of outliers while stabilising variance in economic time series data, particularly for variables with exponential growth patterns. Jong [98] highlighted that using logged variables allows for easier interpretation of differences as percentage changes, which is more intuitive and economically relevant than absolute changes. The logarithmic transformation normalises the data distribution, resulting in a more uniform variance [99]. The log transformation reduces heteroskedasticity by representing changes in proportional terms, making the data suitable for time series modelling techniques [100].

3.4.1. Threshold-Switching Dynamic Model Applied

The economic rationale for threshold effects in this study includes critical mass effects in renewable energy deployment, which result in transformational changes once renewable capacity reaches a large enough scale to influence grid operations. Threshold effects are significant given the country’s unique energy transition challenges. These include, but are not limited to, the reliance on coal-fired electricity generation, the implementation of the REIPPPP, which may result in regime shifts, and the complex interactions between trade liberalisation and environmental outcomes in a resource-intensive economy. This study utilised the following threshold models for the low and high regimes:

CDE-REC threshold model:

$${CDE}_t={\alpha }_1+{\beta }_{11}{REC}_t+{\beta }_{12}{GDP\_growth}_t+{\beta }_{13}{CC}_t+{\beta }_{14}{TO}_t+{\epsilon }_{1t},if{REC}_t\le \gamma$$

(29)

$${CDE}_t={\alpha }_2+{\beta }_{21}{REC}_t+{\beta }_{22}{GDP\_growth}_t+{\beta }_{23}{CC}_t+{\beta }_{24}{TO}_t+{\epsilon }_{2t},if{REC}_t>\gamma$$

(30)

The CDE-REC denotes carbon dioxide emissions and the renewable energy generation model. This model identifies two distinct regimes based on renewable energy generation capacity levels, each with fundamentally different energy–environment dynamics. Low regime (${REC}_t\le \gamma )$ refers to the traditional energy system phase where REC capacity falls below the critical threshold $\gamma$. This is characterised by limited grid integration of renewable sources due to technological and institutional constraints [101], as well as the dominance of coal-fired electricity generation. Higher emissions elasticity in relation to economic growth is consistent with the early stages of the EKC, when economic growth is closely linked to environmental degradation [102]. In this regime, the energy system demonstrates characteristics typical of resource-intensive economies, where fossil fuel dependence creates significant connections between economic activity and CDE [103].

On the contrary, the high regime (${REC}_t>\gamma$) captures the energy transition phase where renewable energy capacity exceeds the critical threshold $\gamma$, triggering fundamental changes in energy system dynamics. This includes, but is not limited to, the displacement of fossil fuel generation as renewable sources achieve competitive cost parity and scale economies [104]. Accelerated grid flexibility investments to accommodate variable renewable energy sources through smart grid technologies and energy storage systems [105]. Lower emissions elasticity with respect to economic growth reflects the decoupling of economic expansion from environmental impacts [106]. This regime is characterised by positive feedback effects where increasing REC creates learning curve benefits that further reduce costs and accelerate adoption [107]. Enhanced institutional capacity for managing complex energy systems and coordinating multiple stakeholders and structural changes in the economy that reduce the carbon intensity of growth through technological innovation and sectoral transformation [108].

In addition, economic growth (GDP_growth) and trade openness (TO) low and high regime models are specified below:

GDP_growth-TO threshold model:

$${GDP\_growth}_t={\delta }_1+{\varnothing }_{11}{TO}_t+{\varnothing }_{12}{CDE}_t+{\varnothing }_{13}{REC}_t+{\varnothing }_{14}{CC}_t+{\mu }_{1t},if{TO}_t\le \tau$$

(31)

$${GDP\_growth}_t={\delta }_2+{\varnothing }_{21}{TO}_t+{\varnothing }_{22}{CDE}_t+{\varnothing }_{23}{REC}_t+{\varnothing }_{24}{CC}_t+{\mu }_{2t},if{TO}_t\le \tau$$

(32)

In the low regime (${TO}_t\le \tau$), trade liberalisation reflects the muted growth effects due to structural barriers including inadequate infrastructure, weak institutional frameworks, and limited absorptive capacity that constrain the economy’s ability to exploit international market opportunities. On the contrary, the high regime (${TO}_t\le \tau$) demonstrates substantial growth benefits through technology spillovers, enhanced competition, and learning-by-exporting effects once trade integration surpasses the critical threshold value $\tau$ [109].

CC-TO threshold model

$${CC}_t={\lambda }_1+{\phi }_{11}{TO}_t+{\phi }_{12}{GDP\_growth}_t+{\phi }_{13}{REC}_t+{\phi }_{14}{CDE}_t+{\upsilon }_{1t},if{TO}_t\le \psi$$

(33)

$${CC}_t={\lambda }_2+{\phi }_{21}{TO}_t+{\phi }_{22}{GDP\_growth}_t+{\phi }_{23}{REC}_t+{\phi }_{24}{CDE}_t+{\upsilon }_{2t},if{TO}_t>\psi$$

(34)

The CC-TO model denotes the coal consumption (CC) and trade openness (TO) threshold model. This model reveals distinct CC patterns across TO regimes with important environmental policy implications. In the low regime (${TO}_t\le \psi$), CC demonstrates weak responsiveness to climate policy enforcement due to limited international pressure and reduced exposure to global environmental standards. These tend to constrain domestic energy policy effectiveness. Insufficient technology transfer mechanisms limit access to cleaner production technologies and energy-efficient alternatives, as well as a lack of institutional capacity to implement environmental regulations in the absence of international monitoring and compliance mechanisms [104]. The high regime (${TO}_t>\psi$), however, demonstrates strengthened trade-linked consumption of coal. Increased trade integration exposes the economy to global environmental standards, green technology spillovers, and market-based incentives for cleaner production processes. This leads to more responsive coal consumption patterns to climate policy interventions [108].

Threshold-Switching Dynamic Model Estimation Procedure

Hansen [110] highlighted that threshold parameters $\gamma$, $\tau$, and $\psi$ are estimated using a two-step procedure that addresses the fundamental identification challenges in threshold models. The first step is to estimate the threshold value by minimising the sum of squared residuals (SSR) across all possible threshold values:

• where:

$$\widehat{\gamma }=arg\underset{\gamma }{\min }{S}_T\left(\gamma \right)=arg\underset{\gamma }{\min }\sum \widehat{u_t^2}(\gamma )$$

(35)

Equation (35) represents the optimal threshold for achieving the best model fit. This is accomplished by using a grid search algorithm that systematically tests all feasible threshold values within the range of the threshold variable. To prevent small sample bias and ensure accurate parameter estimation, each regime must have at least 15% of the total sample [111].

The second step constructs confidence intervals using the likelihood ratio (LR) statistic:

$${LR}_T\left(\gamma \right)=T\times {\displaystyle \frac{S_T\left(\gamma \right)-S_T\left(\widehat{\gamma }\right)}{\widehat{{\sigma }^2}}}$$

(36)

which follows a non-standard asymptotic distribution due to the nuisance parameter problem. This is where the threshold is not identified under the null hypothesis of linearity, requiring bootstrap-derived critical values for robust statistical inference [112]. The confidence interval for $\gamma$ includes all threshold values where ${LR}_T\left(\gamma \right)\le$ is the critical value. This provides a comprehensive assessment of threshold parameter uncertainty and accounts for the searching process inherent in threshold estimation [113].

3.4.2. NARDL Model Applied

NARDL builds on Granger and Yoon’s [114] work on hidden cointegration and extends Pesaran et al.’s [115] bounds testing approach to account for nonlinear dynamics. Shin et al. [62] formalised this framework and demonstrated that traditional linear cointegration tests may fail to detect long-run relationships when the underlying data generation process has asymmetric adjustment patterns. The NARDL model’s theoretical foundation is based on the premise that economic relationships can be asymmetric, with positive and negative shocks to explanatory variables having different magnitudes and persistence effects on the dependent variable [116]. This asymmetric specification is particularly relevant for analysing South Africa’s economic and environmental dynamics, as structural rigidities and institutional constraints may create differential adjustment mechanisms for positive versus negative shocks.

The general NARDL model can be expressed as follows:

$$∆y_t={\rho y}_{t-1}+{\theta }^{+}{x}_{t-1}^{+}+{\theta }^{-}{x}_{t-1}^{-}+{\sum }_{i=1}^{p-1}{\alpha }_i{\Delta y}_{t-i}+{\sum }_{i=0}^{q-1}\left({\pi }_i^{+}{\Delta x}_{t-i}^{+}+{\pi }_i^{-}{\Delta x}_{t-i}^{-}\right)+{\sum }_{j=1}^k{\beta }_j{Z}_{j,t}+{\epsilon }_t$$

(37)

where the key innovation lies in the decomposition of the explanatory variable into positive and negative partial sums:

$$x_t^{+}={\sum }_{j=1}^t\max \left(\Delta xj,0\right)and{x}_t^{-}={\sum }_{j=1}^t\min \left(\Delta xj,0\right)$$

(38)

The NARDL model categorises explanatory variables as positive ($x_t^{+})$ and negative ($x_t^{-})$ over time. This allows us to estimate asymmetric positive and negative long-run effects (${\theta }^{+}\&{\theta }^{-})$, as well as asymmetric short-run effects (${\pi }_i^{+}\&$ ${\pi }_i^{-})$. Consequently, similar to those studies conducted by Uprasen et al. [117], Lee et al. [118], and Borozan [119], our NARDL models are expressed as follows:

CDE-REC model:

$$\Delta {CDE}_t={\alpha }_0+{\rho }_{ec}{ECT}_{t-1}+{\sum }_{i=1}^{p-1}{\beta }_i\Delta {CDE}_{t-i}+{\sum }_{i=0}^{q-1}\left({\pi }_1^{+}\Delta {REC}_{t-i}^{+}+{\pi }_1^{-}\Delta {REC}_{t-i}^{-}\right)+{{\delta }_2\Delta GDP\_growth}_t+{{\delta }_3\Delta \mathrm{T}\mathrm{O}}_t+{{\delta }_4\Delta CC}_t+{\epsilon }_t=$$

(39)

$${ECT}_{t-1}={CDE}_{t-1}-\left[{\alpha }_0+{\theta }_1^{+}{REC}_{t-1}^{+}+{\theta }_1^{-}{REC}_{t-1}^{-}+{{\theta }_{2}GDP\_growth}_{t-1}+{{\theta }_{3}TO}_{t-1}+{{\theta }_{4}CC}_{t-1}\right]$$

(40)

GDP_growth-TO model

$$\Delta {GDP\_growth}_t={\alpha }_0+{\rho }_{ec}{ECT}_{t-1}+{\sum }_{i=1}^{p-1}{\beta }_i\Delta {GDP\_growth}_{t-i}+{\sum }_{i=0}^{q-1}\left({\pi }_1^{+}\Delta {TO}_{t-i}^{+}+{\pi }_1^{-}\Delta {TO}_{t-i}^{-}\right)+{{\delta }_2\Delta CDE}_t+{{\delta }_3\Delta \mathrm{R}\mathrm{E}\mathrm{C}}_t+{{\delta }_4\Delta CC}_t+{\epsilon }_t$$

(41)

$${ECT}_{t-1}={GDP\_growth}_{t-1}-\left[{\alpha }_0+{\theta }_1^{+}{TO}_{t-1}^{+}+{\theta }_1^{-}{TO}_{t-1}^{-}+{{\theta }_{2}CDE}_{t-1}+{{\theta }_{3}REC}_{t-1}+{{\theta }_{4}CC}_{t-1}\right]$$

(42)

CC-TO model

$$\Delta {CC}_t={\alpha }_0+{\rho }_{ec}{ECT}_{t-1}+{\sum }_{i=1}^{p-1}{\beta }_i\Delta {CC}_{t-i}+{\sum }_{i=0}^{q-1}\left({\pi }_1^{+}\Delta {TO}_{t-i}^{+}+{\pi }_1^{-}\Delta {TO}_{t-i}^{-}\right)+{{\delta }_2\Delta GDP\_growth}_t+{{\delta }_3\mathrm{C}\mathrm{D}\mathrm{E}}_t+{{\delta }_4\Delta REC}_t+{\epsilon }_t$$

(43)

$${ECT}_{t-1}={CC}_{t-1}-\left[{\alpha }_0+{\theta }_1^{+}{TO}_{t-1}^{+}+{\theta }_1^{-}{TO}_{t-1}^{-}+{{\theta }_{2}GDP\_growth}_{t-1}+{{\theta }_{3}CDE}_{t-1}+{{\theta }_{4}REC}_{t-1}\right]$$

(44)

where ${CDE}_t$, ${GDP\_growth}_{t}and{CC}_t$ are the dependant variables which captures the short and long-term fluctuations. ${ECT}_{t-1}$ is the error correction term that measures how quickly the system returns to its long-term equilibrium after being disturbed. ${\rho }_{ec}$ quantifies the proportion of disequilibrium corrected in each period, with values closer to −1 indicating rapid adjustment and values closer to zero suggesting slow convergence. ${\pi }_1^{+},{\pi }_1^{-}$ represent the impact of positive and negative changes in the short-run, while ${\theta }_1^{+},{\theta }_1^{-}$ represent the impact in the long-run. The models include several control variables in the short run (${\delta }_2,{\delta }_3\&{\delta }_4)$ and the long-run (${\theta }_2,{\theta }_3\&{\theta }_4)$ to account for confounding factors and ensure robust estimation. ${\epsilon }_t$ is the error term.

The following is a representation of the null hypothesis to be tested for the models, which suggests that there is no cointegration:

$$H_0:\rho ={\theta }^{+}={\theta }^{-}=0$$

(45)

In contrast, the following null hypothesis to be tested for both models suggests that there is a possibility of cointegration that may exist between the variables:

$$H_0:\rho \ne {\theta }^{+}\ne {\theta }^{-}\ne 0$$

(46)

The NARDL bounds test (F-statistic) was used for two different tests, as described in Pesaran et al. [115] and extended by Shin et al. [62]. These tests include cointegration and asymmetry effects. If the calculated F-statistic is greater than the upper bound critical value I(1), the null hypothesis of no cointegration is rejected, indicating the presence of a long-run relationship and asymmetry between the variables. The inverse is true if the calculated F-statistic is less than the lower bound critical value I(0). The null hypothesis is accepted, suggesting that the variables are symmetric and not cointegrated in the long run. When the calculated F-statistic is between the lower I(0) and upper I(1) bounds, the test results are considered inconclusive and we cannot confirm whether or not the variables are symmetric or asymmetric (also cointegrated or not cointegrated).

3.4.3. Threshold Granger Causality Test Applied

The threshold Granger causality test extends the traditional linear Granger causality framework by including regime-switching dynamics, which allow causal relationships to vary across different economic conditions [120]. This methodology corrects a fundamental flaw in traditional Granger causality tests, which assume that causal relationships remain constant over time and across economic conditions [121]. The theoretical foundation is based on Granger’s [122] original concept of causality, but it also incorporates Tong’s [123] threshold autoregressive (TAR) modelling approach, which Hansen [111] refined.

The economic justification for using threshold Granger causality in this study is based on the recognition that relationships between environmental and economic variables frequently demonstrate nonlinear dynamics due to structural breaks, policy regime changes, and threshold effects in environmental systems [124]. Traditional linear causality tests may overlook significant causal relationships that emerge only under specific economic or environmental conditions, resulting in inadequate policy recommendations and mis-specified models [125].

The threshold Granger causality test uses a two-regime threshold vector autoregression (TVAR) model, with the regime determined by whether a threshold variable exceeds a critical threshold value. Hansen and Seo [120] describe the general specification as follows:

regime 1:

$$y_t={\mu }_1+{\sum }_{i=1}^p{A}_{1i}{Y}_{t-1}+{\epsilon }_{1t},q_{t-d}\le \gamma$$

(47)

regime 2:

$$y_t={\mu }_2+{\sum }_{i=1}^p{A}_{2i}{Y}_{t-1}+{\epsilon }_{2t},q_{t-d}>\gamma$$

(48)

where $y_t=[{CDE}_t,{REC}_t,{GD{P}_{growth}}_t,{TO}_t,{CC}_t]$. $q_{t-d}$ determines regime switches (with delay $d$), while $\gamma$ represents a threshold value. $A_{1i}$ and $A_{2i}$ capture the autoregressive dynamics (lagged effects of $y_t$) specific to each regime. The intercept vectors ${\mu }_1$ and ${\mu }_2$ account for regime-specific baseline levels of endogenous variables. ${\epsilon }_{1t}$ and ${\epsilon }_{2t}$ denote error terms. The threshold Granger causality test employs a grid search procedure to determine the optimal threshold value that maximises the likelihood function while preserving adequate observations in each regime. In accordance with Hansen and Seo [120], the threshold parameter $\widehat{\gamma }$ is estimated as follows:

$$\widehat{\gamma }=arg\underset{\gamma }{\min }det\left(\widehat{\mathsf{\Omega }}\left(\gamma \right)\right)$$

(49)

where $\widehat{\mathsf{\Omega }}\left(\gamma \right)$ denotes the covariance matrix of residuals conditional on $\gamma$. In order to be consistent with Andrews’ [89] recommendations for structural break testing, the search is conducted over the range of the 15th and 85th percentiles of the threshold variable to ensure sufficient observations in each regime.

The threshold Granger causality test examined whether past values of variable $x$ can predict current values of variable $y$ better than $y$ own past values. This allows for different causal relationships across regimes. Therefore, the null hypothesis of no Granger causality in either regime is expressed as follows:

$$H_0:A_{1i,xy}=A_{2i,xy}=0,\forall i=\mathrm{1,2},\dots ,p$$

(50)

However, the alternative hypothesis that states that Granger causality exists in at least one regime is expressed as follows:

$$H_1:A_{1i,xy}\ne 0{\displaystyle \frac{and}{or}}{A}_{2i,xy}\ne 0forsomei$$

(51)

where $A_{ji,xy}$ denotes the coefficient of $x$ in $y$ equation for regime $j(j=\mathrm{1,2})$ and $p$ is the optimal lag length. The decision rule employed the standard econometric hypothesis testing procedures by utilising the calculated test statistic and corresponding p-value. If the computed test statistic has a p-value less than the 5% level of significance, then the null hypothesis of no Granger causality is rejected, and it is concluded that variable $x$ Granger causes variable $y$ in at least one of the two regimes. In contrast, if the p-value is greater than or equal to the significance level, then the null hypothesis of no Granger causality is accepted, and it is concluded that previous values of variable $x$ do not help predict current values of variable $y$ beyond what $y$’s own past values can predict in either regime.

3.5. Diagnostic Test Applied

The NARDL model in this study has been subjected to diagnostic tests. Diagnostic testing is crucial for establishing the credibility of empirical results and ensuring that statistical inferences drawn from models are sound and reliable [126]. As highlighted in

Table 2. Diagnostic tests applied.

Type of Diagnostic Test	Description	Source
Jarque–Bera Test	Tests whether model residuals follow a normal distribution by examining skewness and kurtosis. Essential for valid statistical inference in econometric models. Non-normal residuals may indicate that positive and negative shocks have fundamentally different distributional properties, affecting the interpretation of asymmetric effects.	Jarque & Bera [127]
Breusch–Godfrey LM Test	Detects serial correlation in model residuals using a Lagrange Multiplier approach. Tests the null hypothesis of no serial correlation against the alternative of autocorrelation up to the specified lag order. Absence of serial correlation confirms the model captures all systematic adjustment patterns.	Breusch [128]; Godfrey [129]

Source: Authors’ own assessment.

below, Jarque–Bera and Breusch–Godfrey LM tests are two diagnostics tests that had been applied.

Therefore, all diagnostic tests had been interpreted using standard econometric criteria. For these diagnostic tests, if the probability value exceeds the 5% (0.05) level of significance, the null hypothesis is accepted [130]. However, the null hypothesis of these tests has been rejected when the probability value is less than the 5% (0.05) level of significance [131].

4. Results and Discussion

4.1. Structural Unit Root Test Results

Table 3. Structural unit root testing results and decision.

Variable	ZA-A Stat	ZA-C Stat	NP-A Stat	NP-C Stat	LS-1 Stat	LS-2 Stat
Carbon Dioxide Emissions (CDE)	−1.63	−4.31	−2.75	−4.43	−229.56 ***	−226.33 ***
Renewable Energy Generation (REC)	−2.21	−6.78 ***	−2.48	−4.64	−13.30 ***	−9.61 ***
Economic Growth (GDP_growth)	−5.91 ***	−6.57 ***	−6.56 **	−6.50 **	−76.22 ***	−68.89 ***
Coal Consumption (CC)	−2.54	−5.78 ***	−3.09	−4.99	172.46	175.46
Trade Openness (TO)	−4.53	−4.37	−5.22	−5.55	−121.18 ***	−117.53 ***

Source: Authors’ own assessment. *** Rejected the null hypothesis at 1% level of significance; ** Rejected the null hypothesis at 5% level of significance.

displays the results from structural unit root tests. The GDP growth rate showed the most consistent evidence of stationarity among all structural unit root tests. ZA tests produced test statistics of −5.9078 (model A) and −6.5697 (model C), indicating rejection of the null hypothesis at the 1% level of significance. NP tests also confirmed stationarity, generating test statistics of −6.5593 and −6.5020 for models A and C, respectively. This allowed the null hypothesis to be rejected with a 5% level of significance. The LS tests provided evidence with negative test statistics (−76.2154 and −68.8923), rejecting the null hypothesis at the 1% level of significance for both one- and two-break specifications. As a result, the GDP growth variable is integrated with order zero, I(0), because it tends to be stationary at levels.

The structural unit root test results for CDE produced different results. The ZA and NP tests did not reject the unit root null hypothesis, with test results ranging from −1.6277 to −4.4303. This indicated non-stationarity at different levels of these testing frameworks. The LS tests yield contrasting results, decisively rejecting the null hypothesis with test statistics of −229.5617 (one break) and −226.3281 (two breaks). Both LS tests were significant at the 1% level. In light of structural breaks, CDE is most likely stationary, indicating an I(0) order of integration.

Additionally, depending on the model specification and unit root testing type, the stationarity properties of REC produced mixed results. With a test statistic of −2.2136, the ZA-A model failed to reject the unit root hypothesis, indicating non-stationarity. However, after accounting for both intercept and trend breaks, the trend-break model C demonstrated stationarity and rejected the null hypothesis (−6.7797) at a 1% level of significance. The NP tests failed to reject the unit root hypothesis in both models, yielding test statistics of −2.4840 and −4.6365. LS tests, on the other hand, rejected the null hypothesis at the 1% level of significance for one break (−13.3005) and two breaks (−9.6070), indicating stationarity. As a result, REC appears to be stationary near structural breaks, indicating the I(0) order of integration.

The CC variable provides the most consistent evidence of non-stationarity across multiple structural unit root tests. With a test statistic of −2.5412, the ZA-A model failed to reject the null hypothesis. However, when trend breaks were considered, the ZA-C model rejected the null hypothesis at the 1% level of significance (−5.7751). The NP tests in both models failed to reject the null hypothesis, with test statistics of −3.0926 and −4.9934. Similarly, the LS tests did not reject the null hypothesis, producing positive test statistics of 172.4646 and 175.4576. CC is thus considered integrated of order one, I(0), due to the ZA-C results.

Furthermore, in all TO model specifications, the ZA and NP tests failed to reject the null hypothesis, with test statistics ranging from −4.3702 to −5.5468. All of these test results fell short of their corresponding critical values. The null hypothesis, however, was significantly rejected by the LS tests, which produced significant test statistics at the 1% level of significance (−121.1843) for one break and −117.5330 for two breaks. As a result, TO remains stationary around structural breaks, indicating an I(0) order of integration.

4.2. Structural Break Assessment Outcome

Figure 2 highlights the results received from conducting a structural breaks assessment using three tests. These tests are the Chow test, Quandt–Andrews and Bai–Perron. The tests uncovered key break points between 1980 and 2023. These breaking points are related to the post-apartheid transition (1994–1995), the global financial crisis (2008), the renewable energy policy implementation (2011), and the political fragmentation (2014–2016).

The 1994–1995 structural breaks in GDP growth and trade openness coincide with South Africa’s democratic transition and massive economic liberalisation program. The 1994 democratic elections marked the beginning of economic restructuring under the Reconstruction and Development Programme (RDP), followed by the Growth, Employment and Redistribution (GEAR) strategy in 1996 [132].

Trade liberalisation fundamentally changed economic openness patterns. The 1995 elimination of the dual exchange rate system (financial rand and commercial rand) represented a critical structural intervention [133]. Additionally, South Africa’s 1995 World Trade Organisation (WTO) membership committed the country to neo-liberal trade policies, with average tariffs falling significantly from 22% to 7.9% by the year 2000 [134]. This rapid trade liberalisation explains the structural break in the trade openness variable around 1994–1995, as the economy shifted from the apartheid-era toward global integration [135]. Monyela and Saba [136] analysed trade openness–economic growth relationships using Vector Error Correction Models (VECM), finding substantial structural differences between pre-BRICS (1991–2010) and post-BRICS (2011–2021) periods. Manwa and Wijeweera [137] demonstrated that South Africa clearly benefited from trade liberalisation policies both in the short run and the long run, using the ARDL framework for 1980–2011.

The 2008 structural breaks across multiple variables reflect South Africa’s integration into global financial markets and vulnerability to external shocks. The economy entered its first recession in 17 years, contracting 1.5% in 2009 and losing nearly one million jobs [136]. This represented a fundamental shift from the growth path established in the post-apartheid period. Consequently, monetary and fiscal policy responses created structural changes in economic relationships. The South African Reserve Bank (SARB) implemented combative monetary easing, cutting interest rates by 450 basis points between December 2008 and May 2009 [138]. The government implemented large front-loaded fiscal easing equivalent to 4.5% of GDP in 2008/09. The budget deficit widened to 7.3% of GDP to support economic stimulus, representing a structural shift in the government’s economic role [139].

In addition, the mining sector disruption explains structural breaks in coal consumption and carbon dioxide emissions. Platinum group metals production fell by 48.5% year-on-year, while manufacturing contracted due to reduced global demand [140]. The electricity crisis (known locally as load-shedding) that began around 2007–2008 exposed the vulnerability of South Africa’s coal-dependent energy system. Load-shedding created the first recognition that coal-based electricity generation was unsustainable [141]. The crisis laid the foundation for subsequent energy policy reforms.

The 2011 structural break in renewable energy generation marks the beginning of South Africa’s most significant energy sector transformation since electrification. The 2011 launch of REIPPPP represented a fundamental policy intervention that created an entirely new economic sector [142]. REIPPPP’s quantitative impact explains the magnitude of structural change. The program mobilised over USD 20 billion in private sector investment and awarded 6200 MW of renewable energy capacity through 123 projects by 2023 [4]. Technology cost reductions were significant, with wind costs falling by 55% from ZAR1.51 to ZAR0.62 per kWh, while solar PV costs decreased 76% from ZAR3.65 to ZAR0.62 per kWh between 2011 and 2016 [143]. This reduction in costs created a structural shift in the economics of electricity.

Weideman et al. [144] applied Bai–Perron break tests to renewable energy data from 1990 to 2010, finding structural breaks linked to policy changes rather than market-driven adoption. This confirmed that government intervention, particularly REIPPPP, created genuine structural transformation rather than gradual evolution. Espoir et al. [145] found regime-switching behaviour in carbon dioxide emissions relative to business cycles, with negative emissions elasticity during recessions and positive elasticity during expansions. This time-varying relationship explains why structural breaks appear in carbon dioxide emissions during crisis periods (2008) and policy intervention periods (2011, 2016).

The 2014–2016 structural breaks correspond to increasing political fragmentation and the emergence of state capture as a systematic constraint on economic performance. State capture has been described as one form of systemic political corruption, in which private interests heavily influence a state’s decision-making procedures for their own benefit [146]. This phenomenon is thought to have undermined the performance of key institutions in South Africa, ultimately leading to government failure, which South Africa has yet to recover from [147].

4.3. Model Selection Criteria

Table 4. Model selection criteria results.

Threshold Variable	Threshold Value	Log-Likelihood	AIC	BIC	RMSE	R-Squared	Model Selected
Carbon Dioxide Emissions (CDE) model
REC	0.5644	129.935	−229.869	−203.804	0.010581	0.996	Yes
TO	3.979	129.222	−228.444	−202.379	0.010762	0.996	No
Economic growth (GDP_growth) model
REC	1.4587	−81.211	192.423	218.488	1.613815	0.549	No
TO	3.979	−76.926	183.852	209.917	1.457276	0.633	Yes
Coal Consumption (CC) model
REC	0.5644	121.788	−213.577	−187.512	0.012846	0.992	No
TO	3.8532	122.749	−215.498	−189.433	0.012555	0.992	Yes

Source: Authors’ own assessment.

highlights the outcomes of model selection criteria that compared two threshold variables.

REC emerged as a suitable threshold variable in the CDE model. REC performed significantly with a threshold value of 0.5644 and achieved a high log-likelihood of 129.935, which corresponds to higher AIC (−229.869) and BIC (−203.804) values compared to the TO threshold variable. REC accurately predicts CDE, with an RMSE of 0.010581 and an R-squared of 0.996. This outcome explains 99.6% of the variance and implies that REC is the critical structural variable that determines how other factors influence CDE.

The GDP_growth model presented an alternative narrative by identifying TO as the optimal threshold variable instead of REC. The threshold value of 3.979 signified a critical level of openness to trade at which the relationships between economic growth experience structural change. TO demonstrated better results with a log-likelihood of −76.926, a higher AIC of 183.852, BIC of 209.917, lower RMSE of 1.457276, and a higher R-squared of 0.633. This outcome highlighted that trade openness is the primary factor influencing regime changes in the relationships between economic growth and other variables, rather than the generation of renewable energy.

Similarly, the CC model identified TO as the threshold variable, with a threshold value of 3.8532, suggesting distinct structural breaks for varying outcomes. The TO threshold variable demonstrated significant performance compared to REC across all evaluated criteria. TO achieve a higher log-likelihood (122.749), improved AIC (−215.498) and BIC (−189.433), a lower RMSE (0.012555), and an identical R-squared of 0.992. Therefore, this indicates that trade openness influences the structural regime regulating the consumption of coal and potentially illustrates the impact of international trade on domestic policy energy decisions.

4.4. Empirical Assessment Outcomes

This section of the study discusses the results obtained from performing the threshold-switching dynamic models, the NARDL model and the threshold Granger causality test.

4.4.1. Threshold-Switching Dynamic Model Results

Figure 3 illustrates the results received from performing the threshold-switching dynamic models (CDE-REC, GDP-TO and CC-TO). These models demonstrated how South Africa’s environmental and economic transformation yields critical threshold effects that fundamentally change the relationships between renewable energy adoption, trade openness, and environmental outcomes. This study identified three critical threshold values in Table 4 that influence regime switching behaviour: 56.4% (0.5644) renewable energy share for reducing carbon dioxide emissions, 397.9% (3.979) trade openness ratio for GDP growth optimisation, and 385.32% (3.8532) trade openness for coal consumption transitions. These thresholds represent structural breakpoints where economic and environmental relationships shift from distinct behavioural regimes.

The CDE-REC model’s exceptional explanatory power (R-squared, 0.996), combined with a 45.2% low regime and 54.8% high regime distribution, suggests that South Africa’s CDEs are primarily characterised by high-carbon development patterns. High-carbon emission trends can be observed in the low regime, while low-carbon development patterns can be observed in the high regime. South Africa’s environmental transition demonstrates clear threshold-driven regime changes, which are consistent with EKC theory. The 56.4% renewable energy threshold marks a critical structural breakpoint at which carbon dioxide emissions relationships shift from positive to negative correlations with economic development. South Africa follows traditional developing economy patterns, with economic growth driving CDE increases. As of 2023, renewable energy accounts for approximately 12% of the electricity mix, placing the economy firmly in the high-emission zone [148].

Above the renewable energy threshold, econometric models predict a shift towards environmental improvements alongside economic growth [149]. This is consistent with the EKC turning points identified for South Africa, which are around USD 7573 GDP per capita [150]. As the economy transitions to clean energy systems, the theoretical regime predicts negative correlations between economic growth and carbon dioxide emissions. Ganda [151] and Shahbaz et al. [152] found that EKC is valid for South Africa across a wide range of environmental indicators. Trade openness consistently improves environmental quality by reducing the growth of energy pollutants, which supports the threshold model’s predictions. REIPPPP has demonstrated regulatory effectiveness in approaching these thresholds, procuring 6200 MW of capacity with 55% reductions in costs in wind and 76% in solar PV since 2011 [4].

The GDP-TO model’s moderate fit (R-squared, 0.633) indicates that economic growth dynamics in South Africa are influenced by factors other than trade openness, reflecting structural challenges documented in post-apartheid development literature. The 64.3% low regime and 35.7% high regime distributions indicate that South Africa’s economy has primarily operated in a low-openness regime from 1980 to 2023, with globalisation effects becoming dominant only when trade openness exceeds 3.9790 (397.9%) of GDP growth. This threshold outcome is consistent with Kovak’s [153] research on trade liberalisation and local labour market adjustment. Kovak [153] discovered that South Africa’s manufacturing sector was underdeveloped even when trade liberalisation started to rise. This exceptionally high threshold (397.9%) indicates that South Africa has not yet achieved optimal trade integration levels.

The CC-TO model’s exceptional fit (R-squared, 0.992) demonstrates that trade openness forces induce highly predictable regime shifts in CC trends. The threshold value of 3.8532 (385.32%) marks a critical point at which international competitive pressures and technology transfer begin to fundamentally change coal dependency. This threshold implies that greater global integration could accelerate the coal transition through technology transfer, competitive pressure, and access to clean energy financing. The 45.2% low regime and 54.8% high regime distributions correspond to the CDE model, implying that both CDE and CC undergo similar regime transitions driven by REC and globalisation forces. Monyela and Saba [136] confirmed that TO increased after BRICS membership (2011–2021), with a faster adjustment to equilibrium (28.97% per period) than before BRICS membership (2.18% per period). This suggests that threshold effects may vary over time and are influenced by patterns of international integration.

Low regime characteristics (below thresholds) reflect South Africa’s historical development patterns. This includes, but is not limited to, high carbon intensity, protected manufacturing, limited export diversification, and a traditional resource-based economic structure. In this regime, economic growth is positively correlated with environmental degradation, trade protectionism benefits inefficient industries, and coal consumption rises with economic activity [154]. High regime characteristics (above thresholds) represent the transformed economy. This includes, but is not limited to, high levels of renewable energy integration capacity to reduce carbon intensity, competitive manufacturing in global value chains, export diversification beyond primary commodities, and decoupling of economic growth from fossil fuel consumption [155]. Therefore, regime shift generates new growth dynamics in which environmental improvements and economic competitiveness mutually reinforce.

4.4.2. NARDL Model Analysis

NARDL Bounds Test and Asymmetric Analysis

Table 5. Cointegration results from performing the NARDL bounds test.

Model of Interest	F-Statistic	Asymptotic Critical Values *						Results
		10%		5%		1%
		I(0)	I(1)	I(0)	I(1)	I(0)	I(1)
CDE-REC model	3.4477	3.47	4.45	4.13	5.00	5.15	6.36	No cointegration
GDP_growth-TO model	19.1848							Cointegration exists
CC-TO model	1.4356							No cointegration

Source: Authors’ own assessment. Notes: * Pesaran et al. (2001) [115], p300, Table CI(iii), Case III.

presents the results received from conducting the NARDL bounds test on three separate models. The tests were conducted to determine whether or not the variables of interest are cointegrated.

As a result, the NARDL bounds test revealed mixed evidence for cointegration. The CDE-REC model yielded an F-statistic of 3.4477, which falls below the critical values established by Pesaran et al. [115] at all significance levels (10%, 5%, and 1%). This indicated that there is no long-run cointegration between the variables of interest. GDP_growth-TO model demonstrated clear evidence of cointegration. The calculated F-statistic of 19.1848 exceeded all critical value thresholds at all significance levels, confirming the existence of a stable long-run relationship between the variables of interest in the model. The CC-TO model yielded an F-statistic of 1.4356, which also fell below the critical values at all significance levels, providing no evidence for cointegration between the variables of interest in the model.

The cointegration between the GDP_growth-TO model aligns with economic theory, suggesting that more open economies tend to experience sustained growth patterns. For South Africa, this implies that trade liberalisation policies may have lasting effects on economic performance. The lack of cointegration in both environmental models (CDE-REC and CC-TO) suggests that South Africa’s environmental outcomes during 1980–2023 were not systematically linked to either renewable energy adoption or trade policies in the long run. This could indicate that environmental policies operated independently of these economic factors, or that other variables (not captured in these models) were more influential.

Table 6. NARDL asymmetry analysis summary.

Model	F-Statistic	p-Value	Decision
CDE-REC	3.9401	0.0471 **	Asymmetric effects detected
GDP_growth-TO	0.323	0.5698	No evidence of asymmetry
CC-TO	0.0535	0.817	No evidence of asymmetry

Source: Authors’ own assessment. Significant at ** 5% level.

summarises the asymmetry analysis results for NARDL models. Asymmetric effects were detected in the CDE-REC model, as the F-statistic (3.9401) produced a p-value (0.0471) that was less than the 5% level of significance. This had allowed the null hypothesis to be rejected. The asymmetric pattern represents the structural characteristics of South Africa’s energy system. Renewable energy additions have an immediate displacement effect due to merit order dispatch, in which low marginal cost renewable generation reduces coal plant utilisation [156]. Due to system reliability constraints and load shedding patterns, reductions in renewable energy do not proportionally increase coal generation. The country’s ageing coal fleet (with an Energy Availability Factor of 58.1% in 2022) creates capacity constraints that limit symmetric responses [69].

The asymmetric effects were not detected on GDP_growth-TO (F-statistic 0.323, p-value 0.5698) and CDE-TO (F-statistic 0.0535, p-value 0.817) models. This reflected South Africa’s structural economic characteristics, which generate symmetric rather than asymmetric responses. South Africa’s exports are still dominated by carbon-intensive commodities such as coal, metals, and minerals. In their mineral–energy complex analysis, Fine and Rustomjee’s [157] explained how this industrial structure generates long-term carbon-intensive growth patterns while avoiding asymmetric responses to trade liberalisation phases. Limited economic diversification into clean technology exports mitigates asymmetric trade-environmental responses [26].

As opposed to economies with differentiated export portfolios, South Africa’s comparative advantage in polluting industries results in linear relationships in which TO consistently raises CDE. Udeagha and Ngepah [158] demonstrated that South Africa has not reached the EKC turning point (estimated at USD 6502.88 per capita). This implies that economic growth continues to increase CDE without reaching decoupling phases. Infrastructure constraints, such as persistent load shedding (3773 h in 2022), led to consistent patterns that eliminate asymmetric responses. Energy bottlenecks have an equal impact on both economic expansion and contraction periods, reducing the environmental impacts that differ between business cycles. The ZAR 25 billion Eskom debt burden and ageing coal fleet impose systemic constraints that limit asymmetric policy responses [28].

NARDL Short- and Long-Run Effects

Table 7. CDE-REC NARDL model results.

Parameter	Coefficient	t-Statistic	p-Value	Significance
REC long-run effects
Error Correction Coefficient	0.2961	2.193	0.0354	**
Long-run Positive Effect (θ+)	−0.0506	−2.097	0.0437	**
Long-run Negative Effect (θ−)	−0.0025	−1.135	0.2647	-
Long-run Multiplier (Positive)	0.1707	-	-	-
Long-run Multiplier (Negative)	0.0085	-	-	-
REC short-run effects
REC Positive Change (π+)	−0.0049	−1.614	0.1161	-
REC Negative Change (π−)	−0.0051	−0.993	0.3281	-
Control variables
GDP Growth (δ2)	0.0002	0.041	0.9672	-
TO (δ3)	0.0027	3.529	0.0013	***
CC (δ4)	−0.0051	−0.271	0.7884	-
Constant	0.7354	21.304	0	***

Source: Authors’ own assessment. R-squared 0.9616; Adjusted R-squared 0.9526; Significance levels: *** 1%, ** 5.

displays the results of the CDE-REC model. The error correction coefficient of 0.2961 indicates that approximately 29.61% of any disequilibrium between carbon emissions and renewable energy consumption is corrected each period, implying a moderate adjustment speed of about 3.4 periods to achieve long-run equilibrium. This adjustment mechanism is consistent with Stern’s [102] environmental economics convergence hypothesis, which proposes that environmental improvements follow predictable convergence patterns, even if they are asymmetric.

The long-run positive shock coefficient of −0.0506 and the long-run negative shock coefficient (−0.0025) show that REC has a negative impact on CDE in both scenarios, although in small quantities. This finding provides support for Tapio’s [106] decoupling theory framework, which identifies varying degrees of separation between economic activity and environmental pressure. The negative coefficient indicates weak decoupling, in which renewable energy growth enables economic activity with a lower environmental impact. However, the small magnitude suggests that achieving significant decoupling, in which environmental pressures decrease while economic activity increases, will require a greater deployment of renewable energy than current levels suggest [106].

The asymmetric effects identified in this model provide direct support for the ratchet effect theory in environmental economics. Aghion et al. [159] confirmed that environmental degradation frequently results in irreversible changes that are more difficult to reverse than they were to initiate. The CDE-REC model’s asymmetric responses show that positive renewable energy shocks (expansion) and negative shocks (contraction) have varying degrees of impact on carbon dioxide emissions. Rahman et al. [160] discovered similar asymmetric effects in South and East Asian countries, with positive renewable energy shocks having a significantly greater impact on environmental outcomes than negative ones. This pattern reflects the path-dependent nature of energy transitions, in which investments in renewable energy infrastructure generate long-term benefits even when renewable energy growth slows [159].

The long-run positive multiplier of 0.1707 represents the overall long-term impact of a permanent positive shock to REC on CDE. This multiplier, calculated as the ratio of the long-run coefficient to the error correction term, represents the total adjustment that takes place over time as the system reaches its new equilibrium [62]. Similarly, the long-run negative multiplier of 0.0085 is much smaller, highlighting the asymmetric nature of the REC–CDE relationship.

The short-run coefficients for both positive (−0.0049) and negative (−0.0051) changes in REC are statistically insignificant, implying that the primary environmental benefits of renewable energy come from long-term structural changes rather than immediate adjustments. This finding is consistent with the energy transition literature, which emphasises that renewable energy deployment takes time to yield measurable environmental benefits due to grid integration challenges and technological learning curves [161].

The control variables in the CDE-REC model produced varying results. GDP_growth (0.0002) and CC (−0.0051) are not statistically significant. This suggests that the traditional connection between economic growth and CDE may be diminishing. This finding may indicate progress towards the EKC tipping point, at which higher income levels begin to correlate with improved rather than degraded environmental quality [18]. The CC effect on CDE appears counterintuitive, but it is most likely due to fuel switching dynamics, in which coal is replaced by other fossil fuels rather than eliminated. The negative coefficient may reflect periods when coal-fired power plants operate at reduced capacity factors to accommodate renewable energy generation, resulting in lower emissions per unit of coal consumed. The effect of TO (0.0027) on CDE is the only controlled variable with a significant positive relationship at the 1% level of significance. This provides support for the pollution haven hypothesis. This theory, developed by Copeland and Taylor [19], contends that increased international trade can result in carbon leakage, where production shifts to countries with less stringent environmental regulations, potentially raising global emissions.

The error correction coefficient of 0.6885 of the GDP_growth-TO model, as shown in

Table 8. GDP_growth-TO NARDL model results.

Parameter	Coefficient	t-Statistic	p-Value	Significance
TO long-run effects
Error Correction Coefficient (ρ)	0.6885	0.624	0.5371
Long-run Positive Effect (θ+)	−0.9446	−7.171	0.0001	***
Long-run Negative Effect (θ−)	−0.0911	−0.061	0.9518
Long-run Multiplier (Positive)	1.372	-	-	-
Long-run Multiplier (Negative)	0.1324	-	-	-
TO short-run effects
TO Positive Change (π+)	−0.4165	−0.199	0.8435
TO Negative Change (π−)	7.6633	1.683	0.1019
Control variables
CDE (δ2)	−0.4471	−0.107	0.9151
REC (δ3)	82.5317	4.821	0.0001	***
CC (δ4)	0.1451	0.52	0.6062
Constant	−60.9463	−4.521	0.0001	***

Source: Authors’ own assessment. R-squared: 0.7441; Adjusted R-squared: 0.6839; Significant at *** 1% level.

, is statistically insignificant, indicating weak adjustment mechanisms towards long-run equilibrium in the trade–growth relationship. According to Rodríguez and Rodrik [162], the complex and nonlinear relationship between trade policy changes and economic outcomes is mediated by institutional quality and complementary policies. The most significant finding is a long-run positive shock effect coefficient of −0.9446 at the 1% level of significance, indicating that increasing trade openness actually slows economic growth in the long run. This unexpected finding calls into question the traditional trade liberalisation hypothesis advocated by neoclassical economists such as Krueger [163], who claimed that trade openness boosts growth through efficiency gains and technology transfer.

The negative coefficient can be explained using several mechanisms identified in the literature. Firstly, premature trade liberalisation in countries lacking institutional capacity can result in resource misallocation and economic instability, rather than efficiency gains [162]. Secondly, trade liberalisation may eliminate domestic firms before they reach economies of scale, resulting in net welfare losses and reduced industrial capacity [164]. Lastly, the factor endowment theory suggests that countries may lose comparative advantages in higher-value activities when trade opens before domestic capabilities are fully developed [165]. The long-run negative shock effect coefficient of −0.0911 is statistically insignificant, implying that reductions in trade openness have no significant impact on growth. This suggests that trade restrictions may not always have a negative impact on economic growth, providing empirical support to Baldwin’s [164] infant industry protection arguments. The findings suggest that countries can maintain growth while implementing strategic trade policies to protect emerging industries or address market failures [164].

The long-run positive multiplier of 1.372 indicates that sustained increases in TO have cumulative negative effects on economic growth that surpass the immediate impact. According to Rodríguez and Rodrik [162], trade liberalisation may lead to additional negative effects such as industrial restructuring, factor reallocation, or technological displacement due to its high multiplier effect. The significantly long-run negative multiplier (0.1324) emphasises the asymmetric nature of trade policy effects, implying that the costs of trade liberalisation might outweigh the benefits of trade protection.

Both the positive (−0.4165) and negative (7.6633) short-run effects of trade openness changes are statistically insignificant. This implies that the primary trade-growth relationship is driven by long-term structural adjustments rather than immediate effects. This finding is consistent with the trade literature, which emphasises that trade policy changes take time to produce measurable economic impacts due to adjustment costs, learning effects, and institutional adaptation [166]. The large magnitude of the negative short-run coefficient, despite its statistical insignificance, indicates significant volatility in the trade–growth relationship. This volatility may reflect the diverse effects of trade policy across sectors and time periods, which is consistent with firm-level research indicating that trade liberalisation creates both winners and losers within economies [167].

The CDE coefficient (−0.4471) is not statistically significant, indicating that environmental factors may not directly limit economic growth in the short to medium term. The REC coefficient (82.5317) is significant at the 1% level and positive, supporting the Porter hypothesis. Porter and van der Linde [168] proposed that environmental regulations and the adoption of clean technology can stimulate economic growth by increasing innovation and efficiency. The large magnitude of this coefficient indicates that renewable energy investments have significant positive spillover effects throughout the economy, potentially through job creation, technological innovation, and improved energy security [159]. The CC coefficient (0.1451) is statistically insignificant, implying that traditional fossil fuel consumption may not be an important driver of economic growth in modern economies. This finding reflects the declining economic importance of coal-intensive industries in developed economies, where services and high-tech manufacturing have taken over heavy industry as primary growth drivers [169].

The error correction coefficient of −0.0241 for the CC-TO model is statistically insignificant as shown in

Table 9. CC-TO NARDL model results.

Parameter	Coefficient	t-Statistic	p-Value	Significance
TO long-run effects
Error Correction Coefficient (ρ)	−0.0241	−0.906	0.3713
Long-run Positive Effect (θ+)	−0.0011	−0.035	0.9719
Long-run Negative Effect (θ−)	−0.0091	−0.562	0.5776
Long-run Multiplier (Positive)	−0.0451	-	-	-
Long-run Multiplier (Negative)	−0.3779	-	-	-
TO short-run effects
TO Positive Change (π+)	−0.0248	−0.968	0.3399
TO Negative Change (π−)	0.0653	1.284	0.2081
Control variables
GDP Growth (δ2)	−0.0669	−1.479	0.1485
CDE (δ3)	−0.0037	−3.644	0.0009	***
REC (δ4)	1.251	21.366	0	***
Constant	0.0015	0.476	0.6373

Source: Authors’ own assessment. R-squared: 0.9566; Adjusted R-squared: 0.9463; Significant at *** 1% level.

, indicating that there are weak or absent adjustment mechanisms towards long-run equilibrium in coal consumption patterns. The ineffective adjustment mechanism may reflect the structural rigidity of coal-dependent infrastructure and the slow pace of energy system transitions.

The relationship between TO and CC is not significant in either direction, with both long-run positive (−0.0011) and negative (−0.0091) shock effects on the consumption of coal. This lack of relationship calls into question the pollution haven hypothesis, which holds that trade liberalisation increases dirty energy consumption in countries with comparative advantages in pollution-intensive industries [19]. The insignificant coefficients support factor endowment theory, implying that domestic resource availability, infrastructure constraints, and energy security considerations have a greater influence on coal consumption patterns than trade policies [165].

The long-run multipliers for both positive (−0.0451) and negative (−0.3779) changes highlight the weak connection that exists between TO and CC. These multipliers imply that trade policy changes have minimal cumulative effects on CC trends, which is consistent with the notion that energy systems have high inertia and path dependence [80]. However, the larger magnitude of the negative multiplier may reflect an asymmetry in trade effects, with trade restrictions having a slightly greater impact on coal consumption than trade liberalisation, but both effects are economically insignificant. In addition, both positive (−0.0248) and negative (0.0653) short-run effects of TO on CC are statistically insignificant. This demonstrates that trade policy changes do not result in immediate changes in coal consumption. This finding is consistent with the energy economics literature, which emphasises long adjustment lags in energy systems due to capital durability, technological constraints, and institutional inertia [161].

The control variables in the CC-TO model produce a mix of outcomes. The GDP growth coefficient (−0.0669) is statistically insignificant but negative, which could reflect structural economic shifts towards less energy-intensive activities. The CDE coefficient (−0.0037) is negatively significant at the 1% level, which appears counterintuitive but likely reflects fuel switching dynamics in which coal is displaced by other energy sources [170]. This relationship may include periods when coal-fired power plants operate at lower capacity factors to accommodate renewable energy generation, or when natural gas replaces coal in electricity generation and industrial applications. The significant negative relationship between CDE and CC may also be due to policy feedback effects, in which higher CDE trigger regulatory responses that specifically target CC through carbon pricing, renewable energy mandates, or direct regulations on coal-fired power plants [171]. The REC coefficient (1.251) is positively significant at the 1% level, indicating that renewable energy expansion coincides with increased coal consumption. This unexpected finding reflects the complementary rather than substitutive relationship between coal and renewable energy during energy transitions, as documented in recent energy economics literature [161].

4.4.3. Diagnostics Tests

Several diagnostic tests were performed to assess the reliability of the NARDL models (CDE-REC, GDP_growth-TO and CC-TO). As a result, the outcomes from performing the diagnostic tests are highlighted in

Table 10. Diagnostic test results.

Test Type	Test Statistic	p-Value	Decision
CDE-REC model
Ljung–Box Test (Serial Correlation)	1.8410	0.7650	Accepted the null hypothesis
Jarque–Bera Test (Normality)	1.0950	0.4565	Accepted the null hypothesis
GDP_growth-TO model
Ljung–Box Test (Serial Correlation)	2.1329	0.7113	Accepted the null hypothesis
Jarque–Bera Test (Normality)	229.9115	0.0010	Rejected the null hypothesis
CC-TO model
Ljung–Box Test (Serial Correlation)	1.0705	0.8989	Accepted the null hypothesis
Jarque–Bera Test (Normality)	0.1246	0.5000	Accepted the null hypothesis

Source: Authors’ own assessment.

below.

The three models had varying degrees of statistical reliability, with the CDE-REC and CC-TO models showing significant diagnostic performance and GDP_growth-TO producing different results. The CDE-REC and CC-TO models meet all diagnostic criteria, with Ljung–Box statistics of 1.8410 and 1.0705, respectively, and Jarque–Bera values of 1.0950 and 0.1246. These statistics are well above the probability value of 5%, indicating that the models are properly specified, have reliable parameter estimates, and provide valid statistical inference.

The GDP_growth-TO model provided mixed diagnostic test results. The GDP_growth-TO model passes the serial correlation test (Ljung–Box statistic, 2.1329, p-value, 0.7113), but the Jarque–Bera statistic of 229.9115 and p-value (0.0010) violate the normality assumption. This departure from normality can be attributed to the inherent volatility of GDP growth data, structural breaks over four decades of economic development, and the presence of crisis periods, which resulted in fat-tailed distributions with excess kurtosis [172]. Economic growth data typically reflects such properties as a result of business cycle asymmetries, which cause sharp negative deviations during recessions and more gradual expansions. However, this violation has little effect on NARDL validity because the methodology’s asymptotic properties remain robust to non-normal residuals, as opposed to classical linear models, which require strict normality assumptions for valid inference.

4.4.4. Threshold Granger Causality Test Results

Table 11. Threshold Granger causality test results.

Causality Direction	Null Hypothesis	Threshold Value	T-Statistic	p-Value	Low Regime %	High Regime %	Decision
CDE to REC	Past CDE levels do not help predict current REC beyond what REC own past values predict	0.60	−8.90	1.00	51.20%	48.80%	Accepted the null hypothesis
REC to CDE	Past REC levels do not help predict current CDE beyond what CDE own past values predict	5.93	7.80	0.00	51.20%	48.80%	Rejected the null hypothesis
GDP_growth to TO	Past GDP growth rates do not help predict current TO beyond what TO own past values predict	3.87	6.29	0.00	51.20%	48.80%	Rejected the null hypothesis
TO to GDP_growth	Past TO levels do not help predict current GDP growth rates beyond what GDP_growth own past values predict	0.87	−8.60	1.00	51.20%	48.80%	Accepted the null hypothesis
CC to TO	Past CC levels do not help predict current TO beyond what TO own past values predict	3.87	4.29	0.01	51.20%	48.80%	Rejected the null hypothesis
TO to CC	Past TO levels do not help predict current CC beyond what CC own past values predict	0.01	5.94	0.00	51.20%	48.80%	Rejected the null hypothesis
CDE to GDP_growth	Past CDE levels do not help predict current GDP_growth beyond what GDP_growth own past values predict	0.87	−16.40	1.00	51.20%	48.80%	Accepted the null hypothesis
GDP_growth to CDE	Past GDP_growth rates do not help predict current CDE beyond what CDE own past values predict	5.93	9.37	0.00	51.20%	48.80%	Rejected the null hypothesis
REC to GDP_growth	Past REC levels do not help predict current GDP_growth beyond what GDP_growth own past values predict	0.87	−6.53	1.00	51.20%	48.80%	Accepted the null hypothesis
GDP_growth to REC	Past GDP growth rates do not help predict current REC beyond what REC own past values predict	0.60	1.96	0.12	51.20%	48.80%	Accepted the null hypothesis
CDE to CC	Past CDE levels do not help predict current CC beyond what CC own past values predict	0.01	6.94	0.00	51.20%	48.80%	Rejected the null hypothesis
CC to CDE	Past CC levels do not help predict current CDE beyond what CDE own past values predict	5.93	13.34	0.00	51.20%	48.80%	Rejected the null hypothesis

Source: Authors’ own assessment.

summarises the results of the threshold Granger causality test using Hansen and Seo’s [120] threshold approach with a median split for the period 1980–2023. The interpretation of these findings follows a standard econometric approach. For instance, if the test returned probability values (p-values) equal to or greater than 0.05 (5%), we accepted the null hypothesis that variable X does not predict or influence future changes in variable Y. When the p-value is less than 0.05 (5%), the null hypothesis is rejected, implying that variable X predicts or influences future changes in variable Y [173]. The threshold structure enabled the identification of regime-dependent relationships in which causality trends differ significantly between high and low threshold variables.

The threshold Granger causality test revealed a significant causal relationship from REC to CDE. This finding implies that previous levels of REC have strong predictive power for current CDE, extending beyond what CDE’s own historical patterns indicate. The relationship supports Phiri’s [74] recent findings using continuous complex wavelets analysis, which showed that renewable energy in South Africa requires a critical mass to generate significant environmental benefits. This relationship is regime-dependent (threshold value: 5.93), implying that the effectiveness of REC in reducing CDE varies significantly across economic development stages. This nonlinearity is consistent with the literature on technology adoption and reflects the fact that South Africa’s renewable energy sector experienced rapid expansion only after 2011 with the implementation of the REIPPPP. The REIPPPP had attracted over USD 20 billion in investments and aims to integrate 17.8 GW of renewable energy capacity by 2030 [64]. In contrast, the test found no significant causal relationship between CDE and REC (threshold value: 0.60), indicating that previous emission levels do not always result in increased renewable energy adoption. This finding is consistent with Inglesi-Lotz and Dogan’s [56] study of South Africa, which discovered that environmental quality does not always drive renewable energy adoption without deliberate policy intervention.

The significant causal relationship from GDP_growth to CDE provides support to Grossman and Krueger’s [18] EKC hypothesis. This significant relationship suggests that economic growth patterns are highly predictive of CDE course of events. The threshold structure demonstrates that this relationship varies across development regimes. This finding is consistent with previous research conducted in South Africa. Ganda [151] used the ARDL methodology to investigate the EKC hypothesis in South Africa from 1980 to 2014 and discovered evidence supporting the inverted U-shaped relationship between economic growth and carbon emissions. Shahbaz et al. [152] used a variety of cointegration techniques to confirm the presence of EKC patterns in South Africa.

The threshold structure identified in this analysis corresponds to specific periods in South Africa’s economic development with varying emission patterns. During the 1980s and early 1990s, economic growth was significantly correlated with mining and energy-intensive manufacturing, resulting in proportionally higher emissions. However, the post-apartheid period (after 1994) saw structural economic changes towards services and less energy-intensive activities, which may explain the regime-dependent nature of the GDP_growth–CDE relationship [11]. The lack of significant causality from CDE to GDP_growth (threshold value: 0.87) indicates that environmental quality did not directly constrain economic growth in observable ways over the study period. This finding is consistent with Menyah and Wolde-Rufael’s [174] conclusion that carbon dioxide emissions do not drive economic growth in South Africa, implying that environmental degradation has not yet imposed binding constraints on economic development.

The significant causal relationship between CC and TO demonstrates how South Africa’s coal-dependent energy system has a significant impact on its international trade integration strategies. This relationship is especially relevant for South Africa, which has traditionally used its abundant coal resources to develop energy-intensive export industries such as mining, metals processing, and chemical manufacturing [157]. The threshold nature of this relationship is probably due to periods when coal consumption intensity reached levels that fundamentally changed South Africa’s competitive advantage in international markets. During high coal consumption regimes, South Africa developed stronger trade ties with countries importing energy-intensive products, whereas lower consumption periods resulted in different trade pattern evolution. The CC to TO causality identified here provides empirical support to the carbon trade relationship and demonstrates how domestic energy consumption patterns influence international economic integration in resource-rich developing economies.

The reverse causality from TO to CC (p-value: 0.000) suggests that trade liberalisation has a significant impact on domestic coal consumption patterns in South Africa. The low threshold value (0.01) indicates that even minor increases in TO can cause significant changes in CC patterns, reflecting the high sensitivity of South Africa’s energy-intensive industrial structure to international competitive pressures. The threshold effect at such a low level suggests that South Africa’s coal consumption is highly responsive to trade policy changes, which may reflect the country’s reliance on energy-intensive export industries that are sensitive to international competitive conditions. Malefane and Odhiambo [175] discovered that trade openness in South Africa increased significantly following 1994, coinciding with economic liberalisation and integration into global markets.

The significant causal relationship between GDP_growth and TO demonstrates the nonlinear nature of economic openness in South Africa’s development experience. This finding is in line with Malefane and Odhiambo [175]. The threshold value (3.87) is most probably associated with critical periods in South Africa’s economic development, when growth patterns fundamentally altered the country’s approach to international trade integration. Economic growth during apartheid (prior to 1994) was associated with import substitution and reduced trade openness because of international sanctions. However, in the post-apartheid period, economic growth became increasingly associated with trade liberalisation and integration into global value chains [176].

The absence of causality from REC to GDP_growth supports Phiri’s [74] conclusion that renewable energy in South Africa has more short-term than long-term growth effects, at least under current policy frameworks. The bidirectional causal relationship between CC and CDE reflects coal’s central role in South Africa’s energy system and emissions profile. Ganda [151] discovered similar bidirectional relationships between various energy sources and emissions in South Africa, emphasising the country’s complex feedback mechanisms within its fossil fuel-dependent energy system. Therefore, the causal relationship findings from this study extend support to arguments for regime-specific policy design that acknowledges the conditional nature of South Africa’s environmental and economic relationships.

5. Conclusions and Policy Recommendations

This study used threshold-switching dynamic models, NARDL analysis, and threshold Granger causality tests to investigate the nonlinear relationships between renewable energy generation, economic growth, and carbon dioxide emissions in South Africa from 1980 to 2023.

5.1. Key Research Findings

The threshold-switching dynamic models identified three critical structural breakpoints that fundamentally change the relationships between South Africa’s environmental and economic variables, demonstrating that the country’s long-term development progression is marked by distinct regime-dependent dynamics rather than smooth transitions. The study identified a 56.4% renewable energy threshold for significant carbon dioxide emissions reductions, indicating a critical mass effect in which renewable energy deployment must reach a significant scale before yielding substantial environmental gains.

The threshold analysis of this study can provide guidance on South Africa’s Paris Agreement climate strategy. The country must meet the renewable energy thresholds to peak carbon dioxide emissions between 2020 and 2025 and decline thereafter. Our threshold model predicts that the regime shift from high-carbon to low-carbon development patterns will occur at 56.4%.

The analysis revealed a 397.9% trade openness threshold for GDP growth rate optimisation, implying that South Africa has yet to reach optimal trade integration levels to maximise economic benefits from globalisation. This unusually high threshold indicates that the country’s economy has primarily operated in a low-openness regime throughout the study period, with little exposure to the growth-promoting effects of international trade integration. The 64.3% low regime and 35.7% high regime distributions show that South Africa’s globalisation effects become dominant only when trade openness exceeds this critical threshold.

A trade openness threshold of 385.32% for coal consumption transitions demonstrates that international competitive pressures and technology transfer only begin to fundamentally change coal dependency patterns once trade integration exceeds this critical level. This finding suggests that greater global integration might accelerate South Africa’s coal transition by improving access to clean energy technologies and financing mechanisms. The 45.2% low regime and 54.8% high regime distributions correspond to similar regime transitions driven by renewable energy and globalisation forces.

Our analysis of trade openness thresholds (397.9% and 385.32%) suggests that Article 6 of the Paris Agreement’s international cooperation and technology transfer mechanisms could accelerate South Africa’s transition to these beneficial regimes. This supports increased international climate finance and technology transfer to help emerging economies reach a renewable energy critical mass.

The NARDL analysis found mixed evidence for asymmetric cointegration in the three models examined. GDP_growth-TO model showed clear evidence of cointegration, with an F-statistic of 19.1848 that exceeded all critical value thresholds. This had confirmed the existence of a stable long-run relationship between GDP growth rate and trade openness in South Africa. However, there was no evidence of asymmetric effects, with an F-statistic of 0.323 and a p-value of 0.5698 indicating symmetric rather than asymmetric responses.

The CDE-REC model produced an F-statistic of 3.4477, which fell below critical values and indicated that there is no long-term cointegrating relationship between carbon dioxide emissions and renewable energy generation. However, this model showed significant asymmetric effects (F-statistic: 3.9401, p-value: 0.0471), implying that positive and negative renewable energy shocks have different effects on carbon dioxide emissions despite the lack of long-run cointegration. According to the study, positive renewable energy shocks have significantly greater environmental impacts than negative shocks, demonstrating the path-dependent nature of energy transitions in which renewable infrastructure investments provide long-term benefits even when growth slows.

The CC-TO model revealed no cointegration relationship, with the F-statistic (1.4356) falling below all critical values and thus finding no evidence of cointegration between the variables of interest in the model. The NARDL asymmetry analysis for the CC-TO model revealed no evidence of an asymmetric relationship. This reflects South Africa’s structural economic characteristics, which result in symmetric rather than asymmetric responses. Resource dependence and trade composition show consistent patterns across economic cycles. South Africa’s exports remain dominated by carbon-intensive commodities such as coal, metals, and minerals.

The threshold Granger causality analysis revealed significant unidirectional causality from renewable energy to carbon dioxide emissions, as well as from GDP growth to carbon dioxide emissions. Additionally, the test revealed a bidirectional relationship between coal consumption and trade openness. Trade policy can be strategically used to support NDC implementation by reducing coal dependency and encouraging the adoption of clean energy technologies. The regime-dependent causality (threshold value: 5.93) indicates that the effectiveness of renewable energy in reducing carbon dioxide emissions varies significantly across economic development stages, which is consistent with technology adoption literature that emphasises critical mass requirements. However, renewable energy showed no significant causal relationship with GDP growth rate, contradicting traditional growth-led energy hypotheses and implying that South Africa’s renewable energy deployment has yet to reach a sufficient scale to drive economic growth directly. This finding provides support to other studies that have concluded that renewable energy in South Africa has more environmental benefits than economic growth under current policy frameworks.

5.2. Policy Recommendations

Based on these empirical findings, this study makes a few policy recommendations. The South African government, in collaboration with the energy industry, must prioritise and accelerate the REIPPPP’s expansion beyond current commitments in order to reach the critical threshold (56.4%) identified in this study through systematic capacity expansion.

Establish renewable energy industrial clusters to capitalise on asymmetric effects. Given the asymmetric effects identified in the CDE-REC model, policy should prioritise maximising positive renewable energy shocks by establishing concentrated renewable energy manufacturing hubs in provinces with high renewable energy potential.

Develop consistent trade liberalisation policies that recognise symmetric trade-growth relationships. South Africa can implement consistent and predictable trade liberalisation policies without regard for differential impacts across business cycles, focusing on selective liberalisation in sectors where the country has comparative advantages while developing institutional capacity for long-term trade integration.

Promote systematic trade in renewable energy technology. The bidirectional causality between coal consumption and trade openness, combined with symmetric trade-growth relationships, suggests that sustained international integration may accelerate renewable energy transitions.

Strengthen energy transition governance frameworks for threshold management. To achieve the identified threshold effects efficiently, policy implementation must be coordinated across multiple government departments. South Africa should establish a dedicated Energy Transition Authority with the mandate and resources to coordinate renewable energy deployment, grid integration, and industrial policy, with a focus on monitoring progress towards the 56.4% renewable energy target and implementing adaptive policies as the economy approaches regime transitions.

5.3. Limitations and Future Research Directions

Building on the study’s findings, future research could take this analysis in a variety of directions. Future research should extend the analysis to multi-country panel data to increase sample size while controlling for country-specific heterogeneity. This would improve threshold estimate precision and allow for cross-country regime comparison. Also, future research should utilise monthly or quarterly data. This can reveal short-term adjustment mechanisms that annual data may not capture, especially for rapid environmental changes and policy responses.