Bridging Policy, Infrastructure, and Innovation: A Causal and Predictive Analysis of Electric Vehicle Integration Across Africa, China, and the EU

Abstract

Electric vehicles (EVs) are central to the decarbonisation of transport systems and achievement of the Sustainable Development Goals (such as SDGs 7 and 13, affordable and clean energy and climate action, respectively). This study adopts a hybrid methodological framework, merging panel econometric models with machine learning (ML), to examine the drivers of EV adoption across Africa, China, and the European Union between 2015 and 2023. We analyse the influence of charging station density (CSD), GDP per capita, renewable energy share (RES), urbanisation, and electricity access using both first-difference and fixed-effects models for causal insight and Random Forest, XGBoost, and neural network algorithms for predictive analytics. While CSD emerges as the most significant driver across models, results reveal a paradox—GDP per capita demonstrates a negative relationship with EV adoption in econometric models yet ranks among the top predictive features in ML models. This divergence highlights the limitations of assuming linear causality in high-income settings and underscores the value of combining causal and predictive approaches. SHAP and PCA analyses further illustrate regional disparities, with Africa showing low feasibility scores due to infrastructure and grid limitations. Sub-regional case studies (Kenya, South Africa, Morocco, Nigeria) emphasise the need for tailored, integrated policies that address both energy infrastructure and transport equity. Findings highlight the value of combining interpretable models with predictive algorithms to inform inclusive and region-specific EV transition strategies.

1. Introduction

Over the past decade, there has been growth in electric vehicle adoption stemming from their environmental benefits, propelled by global commitments like the Paris Agreement [1] and the Sustainable Development Agenda by the United Nations [2]. The integration of electric vehicles (EVs) into transportation systems has accelerated, offering a pathway to reduce greenhouse gas emissions and enhance urban sustainability [3]. Governments worldwide have bolstered this transition through policies such as subsidies, carbon pricing, and investments in charging infrastructure. However, the effectiveness of these interventions depends on techno-economic factors, such as battery costs and grid reliability, which vary by region. For instance, EVs achieve significant emission reductions only when powered by reliable renewable energy sources, yet intermittency challenges often compromise grid stability, necessitating costly battery storage solutions, a persistent issue for both grid reliability and EV feasibility [4]. Moreover, the financial burden of developing charging infrastructure and offering incentives poses challenges, particularly for resource-constrained countries.

Despite growing EV market shares in certain regions, progress remains geographically uneven [5], underscoring the need for a just and inclusive energy [6]. Developed regions like the European Union (EU) and China have made significant strides in EV adoption, leveraging robust infrastructure and policy support. In contrast, emerging economies, particularly in Africa, face persistent barriers: over 600 million, i.e., 53% of Africans, lack electricity access [6], corruption hampers regulatory frameworks [7], and funding shortages limit infrastructure investment [8]. These challenges deter large-scale EV deployment and threaten the 1.5 °C CO₂ emissions target, necessitating region-specific strategies.

Understanding EV adoption requires a theoretical lens to identify key variables and their regional variations. This study draws on the diffusion of innovation theory [9], which posits that technology adoption hinges on economic, social, and policy drivers. The existing literature on EV integration often focuses on individual case studies or global trends, overlooking the nuanced regional differences shaped by economic structures and policy environments [10]. Traditional econometric methods based on innovation theory [9], widely used in these studies (Inglesi-Lotz & Dogan [11]; Oosthuizen et al. [12]; Zhang et al. [13]; Hashmi et al. [14]) excel in causal inference but falter with high-dimensional non-linear data. Machine learning (ML), however, adeptly handles such complexities, offering superior pattern recognition and predictive power [15].

This study addresses a critical gap by integrating panel econometric models with machine learning (ML) to examine causal drivers and predict EV adoption feasibility across Africa (Morocco, South Africa, Kenya, Nigeria), China, and the EU from 2015 to 2023. Unlike prior econometric studies that focus on causality but miss non-linear patterns [15] or ML studies that prioritise prediction without policy depth [16], this work combines fixed-effects models: Random Forest, XGBoost, and neural networks, enhanced by SHAP for interpretability. This dual-method approach, later detailed, is novel, particularly for African sub-regions, where EV adoption is underexplored due to data scarcity and methodological limitations.

This research seeks to answer three key questions:

• How do policy interventions impact EV adoption? Fixed-effects models quantify causal effects, controlling for regional heterogeneity.
• Which techno-economic variables (e.g., charger density, energy prices, GDP) dominate EV adoption? ML feature importance and SHAP identify key drivers and non-linear patterns.
• Does ML outperform econometrics in predictive accuracy? Comparative metrics (R², MSE, MAE) evaluate forecasting power.

This study integrates econometric and machine learning models to assess EV adoption feasibility, providing data-driven insights for policymakers, investors, and researchers to advance the global clean energy transition. The paper is structured as follows: Section 2 reviews the literature on EV adoption factors; Section 3 details the methods, data sources, and descriptive statistics; Section 4 presents the results from the linear regression and machine learning models; Section 5 discusses the findings and summarises the contributions, limitations, and future directions.

2. Literature Review

Transportation, responsible for about 25% of global CO₂ emissions, is a climate action priority [17] that has triggered the adoption of EVs. EVs are powered by traction motors rather than fossil fuels, thus reducing emissions while mitigating noise and air pollution [18]. To accelerate EV adoption, governments have implemented policies, including subsidies, tax exemptions, and investments in charging [19]. However, barriers to global EV adoption persist, particularly in developing regions, where oil consumption is projected to rise due to growing automobile markets [20]. OPEC [21] forecast that oil and gas will constitute 53% of the energy mix (29.3% oil, 24% gas), despite a 24% increase in energy demand driven by non-OECD countries. This contrasts with the projected 30 million EV sales by 2027 and 73 million by 2040, led by Europe (86%), China (81%), and North America (78%) in automobile sales, highlighting the very lopsided growth in global EV adoption [22,23]. Numerous researchers have extensively studied EV adoption. From a technical perspective, the prevailing issues with battery storage capacity, including a shorter driving range, are compounded by the long charging time and are an established constraint [24,25]. Given the possible digital footprint and interconnectedness of EVs, issues of personal data safety and system cybersecurity risks exist [26], ranging from sensor malfunctioning, electromagnetic interference, and temperature sensing manipulation to even fault injection [27]. Xue et al. [28] highlighted financial incentives, including subsidies and tax reductions, and non-financial policy incentives, like free parking, access to restricted traffic zones, charging infrastructure, and awareness creation, as background factors affecting EV adoption given the positive response of customers to EVs [29]. Additionally, individual characteristics involving socio-economic variables, such as costs, including upfront costs and costs per mile [30,31], age and gender [32,33,34], familiarity with EV technology [32,35], and educational awareness [36,37,38], have all been found to affect the EV adoption rate. Research further shows that several of the highlighted incentives have been adopted in countries like Norway, China, and Sweden, leading to increased adoption of EVs.

The literature reveals mixed findings on the relative importance of these factors. Although Chonsalasin et al. [39], Pamidimukkala et al. [40], Maybury et al. [41], and Xue et al. [28] consistently identify charging infrastructure and financial incentives as pivotal, with Xue et al.’s random-effects model across 20 countries (2015–2019) highlighting subsidies and charger density as key predictors of market share, Mekky and Collins [42] used a fixed-effects model across U.S. states (2012–2020) to demonstrate that a higher share of renewable or nuclear energy in the grid exerts a stronger influence than incentives or infrastructure, likely due to controlling for state-specific heterogeneity, which was absent in Xue et al.’s broader, less granular approach. Wang et al. [43], analysing cross-sectional data from 15 countries, further complicate the narrative by downplaying the role of fiscal incentives in explaining adoption disparities, instead emphasising charger density, fuel prices, and road priority; however, their static model may overlook temporal policy dynamics. These discrepancies stem from methodological trade-offs; random-effects models risk overestimating policy effects by assuming no unobserved heterogeneity, while fixed-effects models better capture regional nuances but may miss global trends and data differences, such as Xue et al.’s international scope versus Mekky and Collins’ U.S.-focused sample. This variability underscores the need for ensemble methods, such as machine learning, to integrate diverse datasets and model non-linear interactions, enhancing decision-making confidence.

2.1. Global North vs. South Disparities

Electric vehicle (EV) adoption trajectories differ markedly between the Global North and Global South, driven by disparities in infrastructure, institutional capacity, economic readiness, and policy design [44]. Countries like Norway and China have emerged as global leaders, albeit via contrasting models. Norway’s consumer-centric strategy, backed by a high renewable energy share and fiscal incentives, has resulted in over 82% of new car sales being electric [45]. Generous tax exemptions, high environmental awareness, and widespread access to home charging enable this success [46,47]. Norway’s model reflects the intersection of affordability, infrastructure maturity, and cultural environmentalism.

China, on the other hand, has relied on a producer-centric, state-led industrial policy, leveraging subsidies, automaker mandates, and urban licensing systems to drive uptake [48]. With 6.8 million EVs sold in 2022, 59% of total car sales, China’s success is mainly urban and top-down, powered by companies like BYD and NIO [22,45]. However, gaps remain in rural infrastructure, revealing internal disparities even within advanced contexts.

In contrast, the Global South, especially sub-Saharan Africa, faces deeply embedded constraints: low access to electricity, limited charger infrastructure, fiscal limitations, and weaker institutional frameworks [49,50]. Over 600 million Africans lack access to electricity, while high battery costs and fuel import dependence exacerbate adoption challenges [8,45]. Although EV policies in countries like South Africa and Morocco exist, implementation is uneven, and strategies borrowed from the Global North often fail due to contextual mismatches [51].

Moreover, while Norway and China offer replicable insights like tax incentives and centralised coordination, Africa requires tailored interventions. For instance, import duty waivers for EVs, the expansion of two-wheeler EVs, and decentralised microgrid charging systems may be more viable alternatives than direct subsidy replication [52,53]. The Global South’s EV transition thus hinges on not just financial capital or technology transfer but adaptive policy innovation rooted in local realities.

2.2. Equity and Justice in African EV Transitions

The discourse on electric vehicle (EV) adoption in Africa is incomplete without a thorough interrogation of equity and justice dimensions. While EVs are globally promoted as tools for decarbonisation, they risk exacerbating socio-spatial and economic inequalities if transition frameworks are not designed inclusively. Beyond infrastructure and affordability, distributional, procedural, and recognitional justice are critical for inclusive transitions, as framed by Just Transition Theory [54] and the Multi-Level Perspective (MLP) [55]. Distributional inequities stem from Africa’s urban–rural divide, with charging stations concentrated in cities like Nairobi, Cape Town, and Johannesburg, excluding rural areas where 53% lack electricity access [56]. Two-wheeler EVs, affordable and prevalent in East Africa (e.g., Kenya’s 72,000 units by 2023) [57], offer inclusive mobility but are sidelined by policies favouring urban four-wheelers, reflecting elite biases and donor-driven agendas [58]. Procedural justice is limited, as marginalised groups, informal transport operators, rural communities, and women are rarely involved in EV policy design [59]. In Nigeria, EV policies prioritise urban infrastructure, sidelining informal okada (motorcycle taxi) drivers, risking economic displacement without participatory planning [60]. Inclusive governance is essential to address such exclusions [54]. Recognitional justice, which values diverse needs and identities, is under addressed in African contexts. Policies often ignore women’s reliance on non-motorised transport and their caregiving roles, exacerbating gender inequities [61]. In South Africa, only about 12% of women access private vehicles, necessitating gender-responsive EV subsidies [62]. Similarly, informal economies reliant on ICE vehicles require transition support to avoid economic displacement [50]. The MLP frames EV adoption as a socio-technical transition, with two-wheeler EVs as niche innovations challenging fossil fuel regimes amid landscape pressures (e.g., SDGs) [55,63]. Just Transition Theory emphasises inclusive governance to align with Africa’s heterogeneous transport landscape [54]. Policies must prioritise rural charging (e.g., solar-powered stations), subsidies for low-income households, and mechanical training for two-wheeler EVs [58]. Community engagement and awareness campaigns can counter fossil fuel resistance, ensuring alignment with SDG 11 (sustainable cities) [63]. This holistic approach ensures Africa’s EV transition is green and just, foregrounding marginalised voices.

2.3. Econometric vs. Machine Learning Approaches in EV Policy Analysis

Understanding how policy interventions influence electric vehicle (EV) adoption requires not only robust data but also methodological approaches that reflect the complexity of energy transitions across different contexts. The existing literature draws primarily on two methodological paradigms: econometric models, valued for their causal identification, and machine learning (ML) models, prized for their predictive flexibility. Each offers distinct insights yet also reveals important trade-offs that motivate hybrid modelling strategies.

Econometric methods, such as fixed-effects panel regressions and OLS, have been the dominant approach for evaluating EV policy impacts. For instance, Peng et al. [64] used cross-sectional OLS on EU and U.S. regions (2018–2022), confirming that income, tax incentives, and charger availability shape adoption patterns—though their findings diverged by region, with income negatively associated with EV uptake in parts of the U.S. These differences reveal the sensitivity of econometric models to contextual heterogeneity and model assumptions, such as linearity and exogeneity. While fixed-effects models offer credible within-region estimates, they often struggle to capture complex, non-linear interactions, such as policy threshold effects or regional policy synergies.

Diffusion models, such as the Bass model used by Abas and Tan [65] in Brunei, complement econometric approaches by emphasising adoption timing and peer effects yet remain sensitive to assumed innovation–imitation dynamics. These approaches, grounded in the Diffusion of Innovations theory [9], help explain temporal shifts in consumer behaviour but offer limited guidance on institutional readiness or infrastructure constraints.

In response to these limitations, machine learning models have gained traction for their ability to model high-dimensional, non-linear, and interactive relationships in EV adoption. Shil et al. [66], employing Random Forest and XGBoost on U.S. data (2015–2020), identified rebates and subsidies as strong predictors, achieving an R² of 0.85 with Random Forest. Kamis and Abraham [67], using gradient tree boosting and random-effects regression across U.S. states, found that solar energy share, education, and air quality were important predictors of EV penetration. These models outperform traditional regressions in capturing non-additive effects and interaction terms, though they typically lack causal interpretability unless paired with model explanation techniques like SHAP or counterfactual inference.

More recent studies extend ML beyond prediction to infrastructure optimisation. Liu and Meidani [68] introduced an end-to-end heterogeneous graph neural network (GNN) model for traffic assignment, enabling real-time flow optimisation in complex road networks. This method captures spatial dependencies and user equilibrium behaviour, offering a promising tool for modelling charger network expansion and EV-related traffic dynamics. In parallel, surrogate modelling techniques such as neural network approximators are applied to aerodynamic truck platooning [69], demonstrating how ML can simulate physical systems with high accuracy, enabling scenario modelling for EV energy efficiency, routing, and infrastructure planning. These innovations underscore ML’s potential to support infrastructure design, deployment planning, and real-time system optimisation, which are rarely addressed in traditional econometric models.

Despite these advances, several limitations persist. Many ML models rely on historical patterns and assume stationarity, which may not hold under disruptive policy interventions. Others, particularly those using multi-criteria or optimisation-based approaches (e.g., Tripathi et al., [70]; Naseri et al., [16]) are often region-specific and difficult to generalise. Moreover, while ML excels at forecasting, it often lacks the causal depth needed to inform policy effectiveness, especially in emerging markets, where data is sparse or unevenly reported.

This study addresses these gaps by employing a hybrid approach—panel regression (fixed-effects and first-difference models) to estimate the structural and policy-level drivers of EV adoption and supervised ML models (Random Forest, XGBoost, neural networks), enhanced by SHAP to uncover non-linear relationships, policy interactions, and regional feasibility profiles across Africa, China, and the EU. This transdisciplinary design recognises that econometrics explains why adoption patterns occur, while ML forecasts how and where they are likely to evolve.

By integrating causal inference with dynamic predictive modelling, this study advances an evidence base for policy design that is both theoretically grounded and practically actionable, particularly in rapidly changing and data-scarce contexts such as sub-Saharan Africa. It reframes the methodological choice not as an either–or decision but as a strategic complementarity, essential for driving equitable and efficient EV transitions under the SDG 7 and 13 mandates.

This review frames the methodological challenge not as a binary choice between econometrics and ML but as a strategic complementarity—econometrics for explaining why policies work and ML for forecasting how and where adoption is likely. A comparative summary of both approaches is presented in

Table 1. Comparative overview of econometric vs. machine learning approaches in EV adoption studies.

Dimension	Econometric Models	Machine Learning Models
Primary Purpose	Causal inference and policy evaluation	Pattern recognition, prediction, and scenario forecasting
Key Assumptions	Linearity, exogeneity, and stationarity	Minimal assumptions; captures non-linear, high-dimensional relationships
Data Requirements	Structured panel data with sufficient temporal and spatial variation	Flexible—handles structured and unstructured data; performs well with large datasets
Interpretability	High—model coefficients are directly interpretable	Often opaque; interpretability improved via SHAP, feature importance, or surrogate models
Handling Endogeneity	Addressed using instrumental variables, fixed effects, or lagged variables	Not inherently addressed; must be modelled externally or combined with econometric techniques
Complexity Capture	Limited in capturing interaction effects and non-linearity	Strong capacity to capture non-additive, complex interactions
Policy Usefulness	Strong for evaluating “what works” and understanding policy mechanisms	Strong for forecasting “where” and “how much”; weaker in explaining underlying causality
Limitations	May overlook threshold effects or policy synergies; less effective in high-dimensional data	Risk of overfitting; lacks causal interpretability without additional methods
Use Case Examples	Assessing the effects of subsidies, income, infrastructure, and urbanisation on EV adoption	Predicting adoption feasibility, identifying drivers across regions, and scenario simulation
Best Used When	Testing policy effectiveness and structural relationships	Forecasting EV trends, uncovering latent patterns, or prioritising interventions

.

As shown, econometric models offer strong causal insights under well-specified assumptions, while ML methods excel in identifying complex, non-linear relationships and variable interactions. This paper combines both to address gaps in existing EV policy analysis.

2.4. Research Gaps

Despite the growing body of literature on electric vehicle (EV) adoption, several critical gaps remain, particularly concerning methodological integration and regional equity considerations.

First, most existing studies rely exclusively on either econometric models or machine learning (ML) techniques. Econometric models offer causal insights but often struggle with non-linearities and high-dimensional data. Conversely, ML approaches provide strong predictive performance but lack causal interpretability and policy relevance. Few studies have successfully integrated these complementary methodologies to evaluate both the drivers and the feasibility of EV adoption across diverse regions.

Second, prior research tends to concentrate on high-income countries, with limited attention to emerging markets, particularly in Africa. These regions face unique infrastructural, economic, and policy challenges, such as limited access to electricity, low charger density, and inconsistent subsidy regimes, which are inadequately captured by models built for the Global North. Moreover, the equity and justice implications of EV deployment in these regions are under-theorised in the existing literature.

Third, while variables such as GDP per capita and charger density are frequently included in EV adoption models, their interactions and contextual relevance differ across regions. Notably, few studies have explored the paradoxical role of GDP, where it may simultaneously support and hinder adoption depending on regional saturation, vehicle type preferences, or affordability perceptions.

Finally, existing studies rarely quantify EV feasibility in a multidimensional, data-driven manner. There is limited use of composite indices or interpretable ML frameworks like SHAP to uncover policy-relevant insights at regional or sub-national levels.

This study addresses these gaps by

• Merging panel econometrics and supervised ML to offer both causal explanations and predictive insights;
• Developing a composite EV feasibility index using PCA;
• Applying SHAP analysis to bridge black-box ML with policy interpretation;
• Conducting a disaggregated regional analysis across Africa, China, and the EU to capture spatial heterogeneity and equity dimensions.

3. Methodology

This study employs a novel dual-method design, combining causal inference with predictive analytics. First, panel data analyses quantified the causal impact of policy interventions, economic variables, and techno-economic factors on EV adoption rates. Second, a supervised machine learning (ML) approach employed Random Forest (RF), XGBoost, and neural network models enhanced by SHAP to predict EV adoption feasibility scores and identify non-linear patterns and sensitive variables. This dual-method framework combined the strengths of causal inference with predictive accuracy, offering a comprehensive understanding of EV integration dynamics across diverse regions. The methodology is designed to address the research questions outlined in the Introduction while accounting for regional variations and data limitations.

3.1. Choice of Machine Learning and Panel Data Regression Models

For panel data analysis, regression model selection employed the Hausman test to compare random versus fixed-effects models under the null hypothesis that individual effects are uncorrelated with explanatory variables [71,72,73]. A rejected null (p < 0.05) indicated endogeneity, justifying the use of fixed-effects models to control for unobserved heterogeneity across regions.

Machine learning (ML) models were selected to complement econometric causality with predictive accuracy, capturing non-linear patterns in EV adoption feasibility. Despite ML’s strengths, challenges like data representativeness and interpretability persist [74]. To address these, we adopted Random Forest (RF), XGBoost, and neural networks (NNs). RF was chosen for its robustness to overfitting, ideal for small datasets, and its feature importance metrics, which clarify policy drivers [75]. However, RF can produce biased importance scores, especially when dealing with categorical variables or correlated features, and may struggle with high-dimensional data, where many features are irrelevant [76,77]. XGBoost, a gradient-boosting framework, optimises accuracy through regularisation, efficiently handling structured, heterogeneous data with non-linear interactions [78]. Nevertheless, XGBoost requires careful hyperparameter tuning, which can be complex and time-consuming. Improper tuning may lead to suboptimal performance or overfitting, particularly in small or noisy datasets [79]. Finally, neural networks (NNs) were included to model deeper, high-order relationships, which is particularly beneficial when working with the PCA-derived composite feasibility index. NNs can capture intricate patterns in data. Yet, they are inherently less interpretable and more data-hungry than tree-based models. Their “black box” nature complicates causal inference, and performance can degrade without extensive tuning, regularisation, and sufficient training data. To offset this, we applied SHAP (SHapley Additive exPlanations) to enhance transparency and extract interpretable policy insights [80]. In sum, while each ML model provides unique advantages for forecasting EV adoption feasibility, their effectiveness depends on data quality, dimensionality, and the trade-off between prediction and explanation—factors we address through methodological triangulation with econometric analysis [81]. Alternatives like SVM were avoided due to their poor performance on small, high-dimensional datasets, and LightGBM risked overfitting without XGBoost’s regularisation [82,83].

3.2. Data Collection and Variable Specification

The study compiled a panel dataset covering the period from 2015 to 2023, reflecting the rapid growth of EV markets and renewable energy transitions. The dataset includes annual observations for the regions of Africa (aggregated) and is further narrowed into sample countries in sub-regions (North: Morocco, South: South Africa, East: Kenya, and West: Nigeria), China, and the EU (aggregated). This results in a balanced panel of 63 observations (seven regions and sub-regions × 9 years). The literature review informed the selection of variables, aiming to capture the multifaceted dimensions of electric vehicle (EV) integration across Africa, including Morocco, South Africa, Kenya, and Nigeria, as well as China and the European Union (EU). These variables encompass policy interventions, economic factors, techno-economic indicators, and relevant control factors, ensuring a comprehensive analysis of EV adoption feasibility. Each variable was precisely defined, measured, and sourced from reputable international databases, like the IEA, IRENA, World Bank, EU, AFDB, National Policy documents, etc., to ensure reliability and consistency. Table 2 shows the variables and their sources.

3.2.1. Data Limitations

While this study utilises reputable data sources, such as the IEA, World Bank, and IRENA, certain limitations warrant acknowledgement. First, some variables had missing values for specific years or countries, particularly in African sub-regions. Where feasible, median imputation and linear interpolation were applied to preserve panel balance, though this may introduce smoothing biases. Second, sub-national data were unavailable, requiring all variables to be aggregated at the national level, potentially masking regional disparities within countries. Additionally, battery cost and subsidy data were not consistently available across all countries and years, leading to their exclusion from certain model specifications to avoid introducing noise. These limitations were mitigated by robustness checks and sensitivity analyses but remain important for interpreting model accuracy and generalisability.

3.2.2. Dependent Variable

This study employs two primary outcome variables to assess electric vehicle (EV) integration. The EV adoption rate (EVAR) is defined as the percentage of new vehicle sales, comprising battery electric vehicles (BEVs) and plug-in hybrid electric vehicles (PHEVs), sourced from IEA and regional databases, and serves as the dependent variable for panel data analysis to estimate causal effects. The EV adoption feasibility score, used in machine learning models, is a composite index (scaled 0–100) derived via Principal Component Analysis (PCA) of three indicators: EVAR, charging station density (CSD), and renewable energy share (RES). In PCA, these indicators are standardised, and the principal component is rescaled to form the score. This score captures the multidimensional feasibility of EV adoption, integrating infrastructure, policy effectiveness, and energy sustainability, and is used to predict regional adoption potential.

3.2.3. Independent Variables

The independent variables are categorised into four main groups: policy interventions, economic variables, techno-economic, and control variables based on their theoretical relevance to EV adoption, as identified in prior studies (e.g., Xue et al. [28] and Sierzchula et al. [53]). These variables are selected to examine their influence on EV adoption rates and feasibility scores.

3.2.4. Policy Interventions

This category includes variables representing governmental and institutional measures designed to promote EV adoption.

• Subsidies and Tax Exemptions: A binary variable indicating the presence (1) or absence (0) of subsidies and tax exemptions for EV purchases, such as exemptions from value-added tax (VAT) or registration fees or rebates.
• Charger Density: Defined as the number of public charging stations per 100,000 population, it reflects the availability of charging infrastructure.

3.2.5. Economic Variables

Economic factors influence consumer purchasing power and the cost-effectiveness of EVs relative to internal combustion engine vehicles (ICEVs).

• GDP per Capita: Measured in USD, it proxies for income levels and economic development, which affect consumers’ ability to afford EVs.
• Access to Electricity: Measured in the fraction of the population that utilises electricity in their daily lives.

3.2.6. Techno-Economic Variables

Techno-economic variables reflect technological and energy-related factors that influence the feasibility and scalability of EV integration.

• Renewable Energy Share in the Grid: Measured as the percentage of electricity generated from renewable sources (e.g., solar, wind), indicating the sustainability of the energy mix powering EVs, which affects their environmental benefits and public acceptance [42].
• Battery Cost: Measured in USD per kWh, it represents the cost of EV batteries, a major determinant of EV purchase prices and affordability.

3.2.7. Control Variables

Control variables are included to account for demographic and geographic factors that may influence EV adoption, ensuring that the estimated effects of the independent variables are not confounded by omitted variable bias.

• Urbanisation Rate: Measured as the percentage of the population living in urban areas, this variable accounts for the higher likelihood of EV adoption in urban settings due to better infrastructure and shorter commuting distances.
• Population Density: Measured as the number of people per square kilometre, this variable controls for spatial constraints that may affect infrastructure deployment (e.g., charger installation) and EV suitability.

3.2.8. Rationale for Variable Selection

The selection of these variables was grounded in both theoretical frameworks and empirical evidence from the literature. Policy interventions (subsidies, tax exemptions, charger density) are widely recognised as key drivers of EV adoption, as they address financial and infrastructural barriers [28]. Economic variables (GDP per capita, access to electricity) reflect affordability and reach, which are critical in regions with diverse income levels like Africa and the EU [53]. Techno-economic variables (renewable energy share, battery cost) capture the technological and sustainability dimensions of EV integration, which are particularly relevant in the context of global energy transitions [42]. Control variables (urbanisation rate, population density) ensure that the analysis accounts for structural differences across regions, enhancing the robustness of the findings.

3.3. Data Sources and Reliability

All data were sourced from established international databases (IEA, IRENA, World Bank, United Nations) and regional reports (e.g., AfDB, Eurostat for the EU), ensuring consistency and reliability. Where regional data gaps exist, particularly for Africa and sampled countries, aggregated estimates or imputation methods (e.g., median imputation) were applied, with sensitivity analyses conducted to assess the impact of such adjustments on the results. Missing variables greater than 15% were dropped to prevent issues of data imbalance, leading to attrition or selection bias [84]. The use of standardised units and constant price adjustments ensured comparability across regions and years.

Table 2. Response variables and data sources.

Variable	Sub Variable	Unit	Abbreviation	Variable Sources
Dependent	EV adoption rate	(%)	EVAR	IEA [56]
Independent	GDP per capita	($)	GDP	World Bank Group [85]
	Access to electricity	(%)	AE	IEA [45]
	Subsidies & tax exemption	binary	ST	IEA [56]
	Charging station density	per 105 population	CSD	Statista [86]
	Battery cost	($)	BC	IEA [45]
	Renewable energy share of generation	(%)	RES	IRENA [87]
	Urbanisation	(%)	URB	World Bank Group [85]
	Population density	(people/km)	PD	World Bank Group [85]

Table 2. Response variables and data sources.

Variable	Sub Variable	Unit	Abbreviation	Variable Sources
Dependent	EV adoption rate	(%)	EVAR	IEA [56]
Independent	GDP per capita	($)	GDP	World Bank Group [85]
	Access to electricity	(%)	AE	IEA [45]
	Subsidies & tax exemption	binary	ST	IEA [56]
	Charging station density	per 105 population	CSD	Statista [86]
	Battery cost	($)	BC	IEA [45]
	Renewable energy share of generation	(%)	RES	IRENA [87]
	Urbanisation	(%)	URB	World Bank Group [85]
	Population density	(people/km)	PD	World Bank Group [85]

3.4. Theoretical Framework

3.4.1. Panel Data Analysis

The model was designed using the GRETL package to estimate the causal impact of policy interventions, economic variables, and techno-economic factors on EV adoption rates while controlling for unobserved heterogeneity and temporal trends. The analysis determined both the short-term and long-term effects of variables by utilising first-difference and fixed-effects models for a robust understanding of EV adoption rates [88]. The data used exists in three dimensions—individual, index, and time—such that the dependent variable Y and the explanatory vector set $X_{it}$ are in k-dimension. A panel roots test was performed to ensure stationarity of the data [89]. In cases of non-stationarity, a first difference is applied to ensure consistent distributions, thus ensuring stationarity. The first-difference model is given as follows:

$$∆X=X_{i+1}-X_i$$

(1)

such that $X$ and i represent the variable and time step, respectively.

The Levin–Lin–Chu (LLC) and Kwiatkowski–Phillips–Schmidt–Shin (KPSS) tests were employed to assess the stationarity of panel data variables, with the LLC test evaluating the null hypothesis (H₀) that all panels exhibit a unit root, implying non-stationarity [90]. A p-value below 0.05 rejects this null, indicating stationarity (I (0)), whereas high p-values (e.g., 1.00) suggest non-stationarity (I (1)). The LLC tests were configured with a constant, two lags of first differences, and a Bartlett kernel with 2–3 lags. Conversely, the KPSS test assumes the null hypothesis that all units are stationary, with p-values below 0.05 indicating non-stationarity and p-values above 0.10 supporting stationarity [91]. Additionally, Choi’s meta-tests, combining KPSS p-values via inverse chi-square, normal, and logit methods to yield p-values, were adopted [92]. Given that the LLC and KPSS tests assess opposing nulls, divergence in their outcomes was expected and necessitated robustness checks using the Augmented Dickey–Fuller Generalised Least Squares (ADF-GLS) test. This complemented prior LLC and KPSS analyses for a more informed conclusion [93]. The panel root test was infeasible for data exhibiting exact or near collinearity.

Fixed-effects models employed for this analysis were due to the control of unobserved heterogeneity since the regions under study exhibit significant heterogeneity in time-invariant factors like cultural attitudes, historical infrastructures, and governance structures. This is in line with Shrimali and Kniefel [94], who proposed that a country’s fixed effects are crucial to addressing unobserved heterogeneities. Furthermore, this unobserved heterogeneity was assumed to be correlated with the independent variables, hence justifying the application of the fixed-effects model. This was further confirmed using the Hausman test in preliminary analyses [73].

The relationship is developed as follows:

$$\begin{gathered} {EVAR}_{it}={\beta }_0+{\beta }_1{ST}_{it}+ {\beta }_2{CSD}_{it}+{\beta }_3{GDP}_{it}+{\beta }_4{AE}_{it}+{\beta }_5{RES}_{it}+{\beta }_6{BC}_{it} \\ +{\beta }_7{URB}_{it}+{\beta }_8{PD}_{it}+++u_{it}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em} \end{gathered}$$

(2)

where ${\mathrm{E}\mathrm{V}\mathrm{A}\mathrm{R}}_{it}$ is the EV adoption rate—the dependent variable, ${\alpha }_i$ is the region-specific term, ${\beta }_1$ to ${\beta }_9$ represent the coefficients of the independent variables, ${\gamma }_t$ represents the time-specific effects that control for unobserved temporal shocks or trends that affect all regions simultaneously each year, and $u_{i\mathrm{t}}$ is the idiosyncratic error. The fixed-effects model removes the time-invariant unobserved heterogeneity by demeaning the data within each region. The estimator minimises the sum of squared residuals:

$$Minimize {\sum }_{i=1}^N{\sum }_{T=1}^T(Y_{it-}{\widehat{Y}}_i-\beta \left(X_{it}-{\widehat{X}}_i\right){)}^2$$

(3)

where ${\widehat{Y}}_i$ and ${\widehat{X}}_i$ are the means of the dependent and independent variables for the region (i), and T and N refer to the period and number of regions, respectively. Nevertheless, addressing the dynamic features of the variables induced the logarithm transformation given below:

$$\begin{gathered} {ln(EVAR}_{it})={\beta }_0+{\beta }_{1}ln({ST}_{it})+ {\beta }_2{ln(CSD}_{it})+{\beta }_3{ln(GDP}_{it})+{\beta }_4{ln(AE}_{it})\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em} \\ +{\beta }_5{ln(RES}_{it})+{\beta }_6{ln(BC}_{it})+{\beta }_{7}ln\left({URB}_{it}\right)+{\beta }_8{ln(PD}_{it})+{\alpha }_i \\ +{\gamma }_t+u_{it} \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em} \end{gathered}$$

(4)

3.4.2. Machine Learning: PCA, Random Forest, XGBoost, and Neural Network Models

This study employed three supervised machine learning (ML) models—Random Forest (RF), XGBoost, and neural network (NN)—to predict EV adoption feasibility scores, uncover non-linear relationships, and identify variable importance. This complements the causal insights derived from the panel data analysis. The target variable was the EV adoption feasibility score (scaled 0–100), constructed via Principal Component Analysis (PCA) of EV adoption rates, charger density, and the renewable energy share, capturing the multidimensional aspects of feasibility.

PCA is used to reduce the dimensionality of high-dimensional techno-economic variables, mitigate multicollinearity, and improve ML model performance. It transforms the original dataset into a new set of orthogonal axes (principal components) that maximise variance. This procedure entails matrix standardisation, covariance matrix determination, eigen decomposition, and principal components estimation. Below is the step-by-step mathematical relationship.

Let X be the standardised matrix of size n × p, where n is the number of observations, p the number of variables, and x_ij the standardised value of variable j for observation i, with mean 0 and standard deviation 1.

$$\mathit{X}=\left[\begin{array}{c}\begin{array}{c}{\mathit{x}}_{11}\\ {\mathit{x}}_{21}\\ ⋮\end{array}\\ {\mathit{x}}_{\mathit{n}1}\end{array} \begin{array}{c}\begin{array}{ccc}{\mathit{x}}_{12}& \dots & {\mathit{x}}_{1\mathit{p}}\\ {\mathit{x}}_{22}& \dots & {\mathit{x}}_{2\mathit{p}}\\ ⋮& \ddots & ⋮\end{array}\\ \begin{array}{ccc}{\mathit{x}}_{\mathit{n}2}& \dots & {\mathit{x}}_{\mathit{n}\mathit{p}}\end{array}\end{array}\right]$$

(5)

Since variables have different units, standardise each variable:

$${x^{\prime }}_{ij}={\displaystyle \frac{x_{ij}-{\mu }_j}{{\sigma }_j}}$$

(6)

where ${\mu }_j$ is the mean of variable j, ${\sigma }_j$ is the standard deviation of variable j, and ${x^{\prime }}_{ij}$ is the standardised value in $\mathit{X}$.

The covariance matrix is given by

$$\mathit{C}={\displaystyle \frac{1}{\mathit{n}-1}}{\mathit{X}}^{\mathit{T}}\mathit{X}$$

(7)

where ${\mathit{X}}^{\mathit{T}}$ is the transpose of $\mathit{X}$, and c_jk is the covariance between variables j and k. Since $\mathit{X}$ is standardised, C is the correlation matrix, with diagonal elements of 1 and off diagonal elements representing correlations.

Performing an eigen decomposition on C gives the following:

$$\mathit{C}{\mathit{v}}_{\mathit{k}}={\mathit{\lambda }}_{\mathit{k}}{\mathit{v}}_{\mathit{k}}$$

(8)

where ${\lambda }_k$ is the eigenvalue k, representing variance explained by the k-th principal component (PC), ${\mathit{v}}_{\mathit{k}}$ is the eigenvector k, the direction of PC k, of length p. k = 1, 2, …, p, with ${\lambda }_1$ ≥ ${\lambda }_2$ ≥ … ${\lambda }_p$.

Finally, the principal components are found by projecting X onto the eigenvector such that

$$\mathit{Z}=\mathit{X}\mathit{V}$$

(9)

where V is the matrix of eigen vectors [v₁, v₂, …, v_p], size p × p, and Z is the matrix of PC scores, size n × p, where column k is the score for PC k.

To ensure optimal performance and convergence in the machine learning models, all feature variables were standardised using z-score normalisation, as shown in Equation (6). This scaling approach centres each variable at zero mean with unit variance, preventing model bias towards features with larger numerical ranges. For the charging station density (CSD) variable in the African region, missing values were imputed using the median of observed entries, implemented via the SimpleImputer(strategy = ‘median’) class from the scikit-learn library. Median imputation was chosen for its robustness to outliers and skewed distributions, preserving the integrity of sparsely populated regional data while maintaining the variable’s scale consistency across the dataset. The features used were identical to the independent variables in the panel data analysis from Table 2. The dataset is split into training (70%) and testing (30%), stratified by region to maintain representativeness. The models employed were as follows:

• Random Forest (RF): An ensemble method that constructs multiple decision trees and averages their predictions for regression tasks. RF was trained with 500 trees, and the hyperparameters (e.g., number of trees, max depth) were tuned using five-fold cross-validation to optimise performance while mitigating overfitting.
• XGBoost: A gradient-boosting framework that builds trees sequentially to minimise the loss function. XGBoost was trained with 300 estimators, a learning rate of 0.1, and a max depth of 5, with the hyperparameters tuned via a grid search (e.g., depth: 3–7, estimators: 100–500) to enhance predictive accuracy.
• Neural Network (NN): A deep learning model that learns complex patterns through interconnected layers. The NN was configured with an architecture of two hidden layers (8 and 4 nodes) and ReLU activation and trained using the Adam optimiser over 500 epochs, with early stopping applied to prevent overfitting.

Model evaluation assessed performance using mean squared error (MSE), mean absolute error (MAE), and R-squared (R²) on the test set for all three models. Feature importance was computed differently across models: RF used tree-based feature importance scores, XGBoost relied on gain metrics from boosting, and NN employed SHAP values to quantify each feature’s contribution, addressing NN’s black-box nature. RF, XGBoost, and NN predictions were compared with actual EV adoption rates and panel data coefficients to evaluate consistency and complementarity. A comparative analysis was conducted to assess the strengths and trade-offs of each model (that is, RF’s robustness, XGBoost’s precision, and NN’s capacity for complex patterns) against their interpretability and computational demands.

The implementation was coded in Python version 3.13 using libraries such as scikit-learn (for RF), xgboost (for XGBoost), TensorFlow (for NN), and shap (for SHAP analysis). Steps included data loading, preprocessing, model training, evaluation, and the visualisation of feature importance (e.g., SHAP summary plots for NN, gain plots for XGBoost), ensuring a robust predictive framework that enhances the interpretative depth of the study. This multi-model approach, as shown in Figure 1, provides a comprehensive lens to address the research questions, balancing predictive power with policy-relevant insights.

3.4.3. Case Country Selection Rationale

Multiple strategic considerations, extending beyond geographic diversity, informed the selection of the four African case countries: South Africa, Morocco, Kenya, and Nigeria. These countries were chosen based on their levels of EV policy engagement, data availability, energy infrastructure characteristics, and regional economic and political significance.

First, all four countries demonstrate varying degrees of engagement with electric mobility policy and pilot initiatives. South Africa, for instance, is home to the continent’s most developed automotive industry and has outlined its ambitions in the Draft Green Transport Strategy [95]. Morocco has emerged as a hub for automotive exports, supported by its expanding renewable energy base and industrial collaboration with global EV manufacturers such as Renault and Stellantis [96]. Kenya’s EV sector has been driven by grassroots innovation and policy support for electric two-wheelers and electric buses, placing it at the forefront of East African e-mobility [97]. As Africa’s largest and most populous economy, Nigeria represents a critical site for understanding affordability constraints, grid reliability challenges, and policy inertia in large developing nations.

Second, these countries offer relatively consistent and publicly available data across the key variables used in the econometric and machine learning models. This includes GDP per capita, urbanisation, electricity access, charging infrastructure density, and renewable energy penetration drawn from sources like the World Bank, IEA, and national energy agencies.

Third, these countries collectively represent diverse infrastructural and socio-economic contexts. Kenya has one of the highest shares of renewables in sub-Saharan Africa, while Nigeria has one of the lowest electricity access rates globally. South Africa’s grid is carbon-intensive and heavily centralised, whereas Morocco invests in solar-based distributed energy systems. These differences enable comparative insights into how structural and policy contexts shape EV feasibility.

Finally, each country plays a leadership role in its respective regional bloc: South Africa in the Southern African Development Community (SADC), Kenya in the East African Community (EAC), Morocco in the Arab Maghreb Union (AMU) and wider MENA region, and Nigeria in the Economic Community of West African States (ECOWAS). As such, their adoption trajectories may have broader implications for regional diffusion, financing strategies, and policy harmonisation.

Together, these selection criteria ensure that the analysed countries capture heterogeneity in Africa’s mobility transition landscape and offer meaningful insights into the institutional, infrastructural, and policy pathways necessary for a just and feasible EV transition across the continent.

4. Results

To contextualise the subsequent statistical analyses,

Table 3. Feature variables across regions.

Countries	EVAR (%)	RES (%)	CSD (per 100k Population)	Access to Electricity (%)
China	75.2	26.3	60.6	~99%
Europe	15.9	35.6	59.5	~100%
South Africa	0.1	3.6	0.3	~86%
Nigeria	0.1	24.6	0.0	~57%
Morocco	0.4	17.2	0.6	~99%
Kenya	0.5	85.5	0.1	47–79%

presents a concise comparative summary of electric vehicle (EV) adoption conditions across the study regions. The European Union and China exhibit high EV adoption rates, robust charging infrastructure, and near-universal electricity access. In contrast, African countries show significant variation. South Africa leads in infrastructure and adoption, while Nigeria lags due to limited grid access and policy support. Kenya excels in renewable energy penetration but suffers from infrastructure gaps, whereas Morocco demonstrates strong renewable capacity yet slow EV uptake due to bureaucratic inertia and limited charger density. This heterogeneity underscores the need for tailored, region-specific strategies that reflect each country’s institutional and infrastructural context.

This section presents the findings from descriptive analysis, the panel data model, and the machine learning algorithms.

Table 4. Descriptive statistics of variables.

Variables	Mean	Std	Min	Max	Q1	Q3
EVAR (%)	12.16	26.20	0.00	109.10	0.09	2.8
CSD (per 100,000 population per km)	879.5	2158.3	0.024	12189.2	0.83689	5235.01
GDP per capita ($)	8826.75	11,115.63	1489.12	39,180.38	2233.71	8275.05
RES (%)	30.46	24.57	1.65	91.46	18.18	31.7
Access to electricity (%)	81.57	19.49	47.00	99.00	61.75	99.00
Battery cost ($)	70.25	29.28	44.54	131.00	47.16	82.20
Urbanisation (%)	54.74	15.18	25.66	75.46	42.82	65.47
Sub & Tax	0.43	0.50	0.00	1.00	0	1.00

shows the descriptive statistics, and Figure 2a shows the correlation matrix between response variables. A wide range of correlations, from −0.66 to 0.84, between response variables is observed, with the highest correlations between EV adoption rate (EVAR), charging station density (CSD), access to electricity, and subsidies/taxes. Figure 2b further displays the pairwise correlation broken down by the EV adoption rate, revealing that charging station density is significantly correlated with urbanisation, subsidies and tax exemptions (Sub & Tax), access to electricity, and GDP per capita.

Notably, from Figure 2a, the renewable energy share of the grid (RES) shows significant inverse correlations with access to electricity, subsidies and tax exemptions (Sub & Tax), and urbanisation in our dataset.

4.1. Panel Data Analysis Results

The stationarity test performed is displayed in

Table 5. Panel unit root results (LLC and KPSS).

Variable	LLC z-Score	LLC p-Value	KPSS p-Value Range (Units)	Choi Meta-Tests p-Values	Conclusion
EVAR	11.1827	1	0.049–>0.10	0.0000–0.0006	Non-stationary (I (1))
RES	−4.34677	0	<0.01–>0.10	Not calculated	Stationary (I (0))
GDP	1.40254	0.9196	<0.01–>0.10	Not calculated	Non-stationary (I (1))
BC	10.7824	1	0.070 (all units)	0.0000–0.0007	Non-stationary (I (1))
URB	−1.76847	0.0385	0.066–0.084	0.0001–0.0007	Likely stationary (I (0))
CSD	14.5838	1	0.050–0.094	0.0000–0.0005	Non-stationary (I (1))

. The KPSS results showed a range of p-values across units (T = 8 or 9), reflecting heterogeneity. Choi meta-tests yielded p-values below 0.001, reinforcing non-stationarity where calculated. They were not computed for RES and GDP per capita, possibly due to insufficient variation or software limitations [92]. Non-stationary variables, such as EVAR, GDP per capita, battery cost, and CSD, require first differencing to achieve stationarity. This is consistent with methodologies for handling non-stationary panel data. RES, being stationary, was suitable for direct modelling without differencing. The urbanisation variable presented ambiguity, as the LLC test (p = 0.0385) suggested stationarity, yet the KPSS meta-tests rejected stationarity, necessitating further testing. The ADF-GLS test, applied to seven units (T = 6) with a constant trend and two lags of first differences, tests the null hypothesis (H0H_0H₀) that all units have a unit root, implying non-stationarity. The test statistics, ranging from 2.37 to 2.69 with p-values of 0.15 to 1.0, fail to reject H0H_0H₀ across all units, as none fall below the 0.05 threshold. This implies that urbanisation is non-stationary. The first difference applied to the data allowed the determination of short-run impacts of variables, and the fixed-effects approach informed long-run impacts.

To decide between the fixed-effects and random-effects models, the Hausman test was performed. The F-statistics from the chi-squared test of the model coefficients yielded a p-value < 0.05. This rejects the null hypothesis, establishing endogeneity in explanatory variables. Hence, a fixed-effects model was adopted.

Table 6. Panel data analysis results (first-difference and fixed-effects models).

Dependent Variable	EVAR (Electric Vehicle Adoption Rate)
Explanatory Variables	First Difference				Fixed Effect
	Coefficient	Standard Error	p-Value	t-Ratio	Coefficient	Standard Error	p-Value	t-Ratio
CSD	2.10	0.53	0.0003 ***	3.94	2.01	0.36	1.04 × 10−6	5.67
URB	−126.67	107.03	0.2442	−1.18	−12.97	7.74	0.1008	−1.68
GDP	−2.18	1.05	0.0432 **	−2.09	−5.75	1.30	6.23 × 10−5 ***	−4.43
RES	−0.03	0.04	0.4003	−0.85	−0.06	0.05	0.222	−1.24
AE	−35.20	28.29	0.2212	−1.24	4.88	1.94	0.0157 **	2.51
Constant	2.93	2.16	0.1837	1.36	52.30	27.24	0.0613 *	1.92
R2 (LSDV-Within)	0.47–0.41				0.97–0.74
F-statistics	5.17 (p = 0.0011)				24.66 (p = 9.4904 × 10−12)

Significance Codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

shows the output from the regression models; the first-difference and fixed-effects models capture the short-term and long-term impacts, respectively, offering a robust casualty analysis for overarching policy considerations [88,98,99]. Five of the predictor variables are statistically significant. The variance inflation factor (VIF) scores, all below 5, confirm the absence of multicollinearity among the included variables. The models explain 41% and 74% of the within-unit variance, as indicated by their respective R² Within values. Coefficient estimates were evaluated for both their magnitude and direction to assess their impact.

From Table 2, the battery cost variable was found to be invariant across regions and over time in the dataset used. This lack of both cross-sectional and temporal variation significantly undermines its utility in a panel regression framework. Variables with near-zero variance tend to be absorbed into time fixed effects, leading to multicollinearity and inflated standard errors [100]. Moreover, battery cost reductions have largely followed a global trend, driven by supply-chain innovation and economies of scale in East Asia, which means national-level battery cost differences are difficult to observe in aggregated datasets. Consequently, battery cost acts more as a global time trend proxy than a country-specific determinant and was therefore excluded to preserve estimation reliability.

Similarly, the subsidies and tax variable, coded as a binary dummy indicating whether a country has any fiscal incentives for EVs, was excluded due to both distributional imbalance and the risk of incidental parameter bias. With only about 40% of observations coded as ‘1’ and the remainder as ‘0’, the limited within-country variation over time made it unsuitable for inclusion in fixed-effects estimation, where time-invariant or rarely changing dummies are often dropped or poorly estimated [73]. The binary nature of the variable also raises concerns about functional form misspecification, particularly when such policies vary widely in scope and magnitude, i.e., from full import waivers to minor tax credits, which cannot be captured in a single dichotomous variable.

Furthermore, initial variance inflation factor (VIF) diagnostics indicated potential multicollinearity when subsidies and tax were included alongside the urbanisation and GDP per capita variables, which are often endogenously linked to fiscal capacity and the likelihood of EV policy deployment. Its inclusion also impaired the overall model fit and reduced the precision of key coefficients such as charging station density (CSD) and access to electricity (AE). Given these statistical and conceptual challenges, the variable was excluded from the panel regression but retained in the machine learning models, where its non-parametric treatment allowed for interaction effects and non-linear responses without violating classical assumptions.

The CSD is the most significant variable, with the maximum t-ratio, with urbanisation and economic level, proxied by GDP and RES, also proving relatively significant across both models. Access to electricity maintains a negative coefficient across both models—on average, −0.036. Nevertheless, this is negligible due to saturation in high-access-to-electricity regions. Comparing both models, as indicated by the F-statistics together with the p-value, identifies the first-difference and fixed-effects models as significant and highly significant respectively.

4.2. Machine Learning Results

The results for machine learning are described based on the accuracy metrics highlighted and compared to panel data analysis. Figure 3 displays a line chart of feasibility scores for EV adoption across seven regions from 2015 to 2023, derived from PCA on standardised data (EVAR, RES, CSD), rescaled to 0–100. China consistently achieves the highest scores (18–99), indicating a 500% increase across the years, with a widening gap over time, driven by robust charging infrastructure, aligned with econometric findings. Europe follows (3–35), reflecting moderate infrastructure and wealth. Africa and its sub-regions remain at the bottom (<2), largely due to low EV adoption and sparse charging infrastructure despite relatively high renewable energy shares. These trends highlight the central role of combined EVAR, RES, and CSD performance in shaping regional EV adoption feasibility over time. These feasibility scores represent the target variable across all machine learning models employed in this study. See Supplementary Materials for time series trend of PCA component variables.

Each machine learning model was tuned using empirically derived hyperparameters to optimise predictive performance and avoid overfitting. The Random Forest (RF) model was configured with a maximum depth of 3, a minimum sample split of 2, and 100 estimators. For reproducibility, the XGBoost model was tuned using 300 estimators, a learning rate of 0.1, and a fixed random state of 42. For the neural network (NN), a single hidden layer architecture with four nodes was employed, using an L2 kernel regularisation factor of 0.01 and a learning rate of 0.001. These configurations were selected through iterative experimentation to balance model complexity, accuracy, and generalisability.

Across all three models, Random Forest, XGBoost, and neural network, as represented in

Table 7. Feature importance across machine learning models.

Feature	Random Forest	XGBoost	Neural Network
CSD	0.911	0.758	0.501
GDP	0.039	0.219	0.071
RES	0.020	0.023	0.032
AE	0.000	0.000	0.122
BC	0.017	0.000	0.033
URB	0.013	0.000	0.069
ST	0.000	0.000	0.173

, CSD consistently shows its most significant importance, highlighting its critical impact on electric vehicle adoption. This supports previous findings on infrastructural challenges as adoption barriers [29,53,101,102]. GDP per capita emerges as the second most crucial variable in the XGBoost model. It retains substantial influence across all machine learning models, suggesting that higher economic standing is positively associated with EV adoption (Figure 4).

With machine learning models widely regarded as black-box models, SHAP plots provide a more detailed understanding [103,104]; Figure 5 displays these. This bolsters earlier findings on the CSD’s dominance of the EV adoption rate while amplifying neural networks’ ability to capture complex relationships, offering a more nuanced causality analysis and robust policy considerations.

4.3. Model Comparison

The accuracy metrics presented in

Table 8. Accuracy metric comparison of econometric models with machine learning models.

Model	Method	Adj. R2	RMSE	MAE
PDA	First-Difference Model	0.41	3.86	3.05
	Fixed-Effects Model	0.74	4.21	3.33
ML	Random Forest	0.81	11.02	4.76
	XGBoost	0.88	8.85	3.35
	Neural Network	0.99	2.20	1.49

highlight the comparative strengths and limitations of econometric and ML models in modelling electric vehicle adoption rates (EVAR). The panel data analysis (PDA) models provide valuable insights into the structural and temporal dynamics influencing EVAR across regions. Their strength lies in theoretical interpretability, policy relevance, and the ability to control for unobserved heterogeneity [73]. The fixed-effects model achieves an adjusted R² of 0.74, indicating strong explanatory power for within-entity variance over time. However, its root mean squared error (RMSE) of 4.21 and mean absolute error (MAE) of 3.33 suggest limitations in predictive precision, particularly for out-of-sample forecasting.

In contrast, the ML models, particularly XGBoost and neural networks, demonstrate higher predictive accuracy by better capturing the non-linearities and complex feature interactions often overlooked in traditional econometrics [105]. The neural network model stands out with an almost perfect fit, adjusted R² of 0.99, along with the lowest RMSE (2.20) and MAE (1.49). This signals exceptional performance in both variance explanation and error minimisation. However, the near-perfect R² raises concerns about overfitting, especially given the relatively small sample size. Overfitting in neural networks is a well-documented issue when the model captures noise in the data rather than underlying trends, compromising generalisability [106].

XGBoost, a gradient-boosted tree-based model, also performs well, with an adjusted R² of 0.88, RMSE of 8.85, and MAE of 3.35, providing a balance between accuracy and feature interpretability through tools like SHAP values. However, its higher RMSE relative to PDA models may indicate a sensitivity to outliers or variance inflation due to recursive partitioning methods [107]. Similarly, while the Random Forest model achieves a respectable R² of 0.81, its error metrics, RMSE of 11.02 and MAE of 4.76, reflect a weaker predictive reliability than both the econometric and advanced ML models [75]. In all, the predictive superiority of machine learning is limited by the stationarity assumption of historical data [66].

These findings underscore a critical distinction—PDA models are well-suited for policy analysis and hypothesis testing due to their theoretical grounding and causal inference capabilities. On the other hand, ML models excel in predictive tasks and pattern recognition, especially in the presence of high-dimensional or non-linear data structures [108]. Therefore, combining both approaches offers a complementary toolkit—PDA for structural insights and ML for forecasting and scenario analysis, answering RQ3: “Does ML outperform econometrics in predictive accuracy?”

5. Discussion and Conclusions

Correlation analysis on control variables across all regions, as shown in Figure 2, aligns with previous studies, consistently highlighting the interplay between control variables for EV adoption. For instance, Javadnejad et al. [109] found that government subsidies, tax credits, and charging infrastructure significantly drive EV adoption in the U.S., particularly in urban areas, supporting our CSD–subsidies and tax and CSD–urbanisation correlations. Similarly, the IEA [56] reported that countries like the UK and Korea have increased charging infrastructure through policy incentives while noting that limited electricity access in rural areas hinders CSD growth, corroborating our CSD–access to electricity correlation. Additionally, Boushey [110] highlighted that wealthier regions in the U.S. benefit from federal investments in charging networks, reinforcing the CSD–GDP per capita correlation observed in our study.

Notably, from Figure 2a, the renewable energy share of the grid (RES) shows significant inverse correlations with access to electricity, subsidies and tax exemptions (Sub & Tax), and urbanisation in our dataset. Regions with high renewable energy shares like Kenya (83–91% RES) often face limited electricity access (47–79%), few EV incentives (e.g., no subsidies in Kenya and Nigeria), and low urbanisation (25–29%). In contrast, Europe, with a lower RES (31–40%), enjoys near-universal electricity access (99%), widespread subsidies, and higher urbanisation (73–75%). Urbanisation demands robust energy infrastructure, often supported by a diversified mix, including fossil fuels, in developed regions like Europe and China, contributing to their lower RES, as shown in Figure 2b. Europe’s cities rely on natural gas and nuclear, unlike Kenya’s rural grid, which depends primarily on hydropower. This finding aligns with recent studies highlighting the high-RES regions’ structural challenges. The IEA [56] notes that countries with high renewable energy shares, such as those in Africa, often face electricity access challenges due to underdeveloped grids, supporting our study’s inverse RES–access to electricity correlation. Similarly, Javadnejad et al. [109] found that EV adoption in the U.S. is driven by financial incentives and high urbanisation, which are more prevalent in regions with diversified energy mixes (lower RES), corroborating the inverse RES–subsidies and tax and RES–urbanisation correlations. These insights suggest that while renewable energy is critical for sustainable EV adoption, regions with high RES may face barriers related to electricity access and policy support, which could limit their ability to deploy charging infrastructure and promote EV adoption.

From Table 6, the analysis reveals that CSD consistently and significantly promotes EV adoption in both the short- and long-term, aligning with previous research like that of Chonsalasin et al. [39], Pamidimukkala et al. [40], Maybury et al. [41], Xue et al. [28], and Sierzchula et al. [53]. GDP shows a significant negative relationship, suggesting that higher economic output may not directly translate to higher EV adoption, potentially due to existing infrastructure or market preferences. Hardman et al. [111] found that wealthier individuals preferred ICEVs, perhaps mainly because the most ostentatious vehicles are predominantly ICEVs. Buhmann and Criado [112] also found that the desirability of EVs as sustainable products only increases if prices are higher. Access to electricity (AE) becomes a significant positive factor over time, highlighting the importance of infrastructure development for sustained EV adoption. A study conducted by Lukuyu et al. [113], focusing on Nairobi, posits that increased EV adoption could moderately raise electricity demand, utilising existing excess generation capacity, hence the need for increased access to electricity, leading to a reduction in energy prices and the seamless adoption of EVs [114]. URB and RES do not show statistically significant effects in this analysis, indicating that their impact on EV adoption may be influenced by other factors not captured in this model. This observation aligns with findings from recent studies that highlight the complexity of these relationships. For instance, Zaino et al. [115] conducted a comprehensive systematic review of technological, environmental, organisational, and policy impacts on EV adoption. Their study emphasises that while factors like urbanisation and renewable energy integration are important, their effects on EV adoption are often indirect and influenced by a combination of other variables, such as infrastructure development, policy incentives, and consumer behaviour. Similarly, Clegg [115] examined the impact of EV adoption in urban and suburban contexts, revealing that urbanisation alone does not guarantee higher EV adoption rates. Instead, factors like the availability of charging infrastructure, local policies, and socio-economic characteristics play more significant roles in influencing adoption patterns. These studies suggest that the relationship between URB, RES, and EV adoption is multifaceted, and their impacts may be contingent upon a broader set of contextual factors. Furthermore, Li et al. [116] applied panel data analysis to data from across 14 countries and also found that urbanisation had no impact on EV demand. Therefore, the non-significant findings in the model could reflect the need to incorporate additional variables or consider interaction effects to capture the dynamics at play fully. These findings underscore the importance of infrastructure, such as charging stations and electricity access, in promoting EV adoption, while also suggesting that economic factors and urban planning play complex roles that warrant further investigation.

Findings from Table 6 present GDP as one of the critical factors affecting EV adoption, indicating a direct relationship with EV adoption. This aligns with prior studies by Patil [117] and Xue et al. [28], which emphasise income levels as a key enabler of EV uptake. However, econometric results reveal an inverse relationship between GDP per capita and EV adoption, highlighting a potential mismatch between observed patterns and predictive mechanisms, hence a GDP–EV adoption relationship paradox. This paradox underscores a fundamental divergence between predictive correlations captured by non-parametric algorithms and the causal inferences drawn from linear econometric models, calling for a more nuanced interpretation of GDP’s role in EV transitions. A plausible explanation for this inverse relationship lies in the saturation dynamics of wealthier regions, where early adopters may have already transitioned to EVs and subsequent growth slows among mainstream consumers who require additional incentives. This phenomenon has been observed in several high-income contexts, where policy incentives plateau and marginal gains in adoption diminish over time [111,118]. Moreover, in affluent regions, consumers may retain a cultural and aesthetic preference for internal combustion engine vehicles (ICEVs), especially premium SUVs and luxury sedans, which are perceived as status symbols. Buhmann and Criado [112] argue that EVs must be sufficiently premium-priced and branded as aspirational products to appeal to high-income buyers, suggesting that the relationship between wealth and EV adoption may not be linear. Additionally, high-GDP countries often feature alternative transportation systems, including efficient public transit, bike sharing, and hybrid vehicle usage, which reduce the perceived need for individual EV ownership. For example, Zhao et al. [119] report that in urban European and East Asian cities, modal substitution effects, where individuals switch from private EVs to transit or micromobility options, undermine the presumed correlation between income and adoption. Furthermore, in regions with high living costs, the marginal cost of EV ownership, including home charging installations or parking space scarcity, may deter uptake despite sufficient income levels. In contrast, low- and middle-income countries may experience policy-induced EV growth through international donor financing, concessional loans, or tax waivers, where GDP does not necessarily dictate infrastructure rollouts or consumer behaviour [45]. In conclusion, while GDP per capita is a critical factor in EV adoption, its influence is mediated by market saturation, consumer preferences, policy environments, and social norms. Understanding this complex interplay is essential for policymakers and stakeholders aiming to promote EV adoption across diverse economic contexts. Figure 5 visualises this inversion using SHAP force plots; hence, for researchers, it affirms the value of hybrid modelling frameworks that bridge interpretability and complexity, offering a fuller picture of how wealth shapes the pathways and pace of EV adoption.

The neural network model captured variables undervalued by tree-based models. Subsidies and taxes posed a 17.3% importance, indicating the pivotal role of incentives for EV adoption rates, as found by Halse et al. [120] and the case studies reviewed earlier. Access to electricity (AE) plays a vital 12.2% contribution to EV adoption, in line with Powell et al. [121], who found that while locally optimised controls and high home charging adoption can support EV integration, they may also strain the grid if not adequately managed. This underscores the importance of strategic infrastructure planning and the implementation of smart charging solutions to accommodate the growing electricity demands associated with widespread EV adoption. The more nuanced understanding of the feature variables could inform neural networks’ improved ability to detect complex interdependencies and high-order relationships amongst variables, as further shown in Figure 4.

6. Policy Recommendations

6.1. Disaggregated Analysis of Electric Vehicle Adoption in African Sub-Regions: Infrastructure, Renewable Energy, and Policy Implications

The transition to electric mobility in Africa presents a heterogeneous landscape, with distinct regional trajectories influenced by varying levels of infrastructure development, renewable energy integration, and policy frameworks. Using the Multi-Level Perspective (MLP) on socio-technical transitions [55], the continent’s current EV landscape can be seen as a fragile convergence of landscape pressures (e.g., climate goals, fuel import dependence), regime challenges (e.g., weak institutions, fossil fuel lock-ins), and niche innovations (e.g., solar microgrids, local EV startups). The empirical results from this study, particularly the dominance of charging station density (CSD) and the role of access to electricity (AE), underscore how technological transitions require enabling socio-political frameworks, not just capital and hardware.

6.1.1. Morocco: Harnessing Renewable Potential Amid Infrastructure Constraints

Morocco exemplifies a landscape where niche innovations in renewable energy (RES of 42%) have not translated into scalable EV adoption due to low CSD (0.61 per 100,000). Our regression analysis showed a strong positive link between CSD and EV adoption, while ML results emphasised AE as a key predictor. Despite high RES, only 0.44% of vehicles are electric, affirming that energy source quality alone is insufficient without infrastructural access. Morocco’s transition is stalled at the regime level—bureaucratic inertia, fragmented permitting, and limited fiscal space hinder scaling. To accelerate progress, Morocco could issue a series of climate-aligned green bonds, potentially under its existing sovereign green bond framework, to raise at least USD 150–200 million for the development of 1000 fast and semi-fast public charging stations by 2030. These stations should be strategically distributed along key intercity corridors (e.g., El Jadida–Rabat–Tangier) and urban centres, with a minimum of 30% reserved for underserved peripheral districts to ensure inclusive access. Implementation should be overseen by the Ministry of Energy Transition and Sustainable Development, in partnership with Masen and international financiers such as the African Development Bank or the Green Climate Fund. To de-risk private investment and ensure rapid deployment, Morocco should fast-track permitting through a centralised digital platform and offer land lease incentives on publicly owned properties. Integration with existing renewable assets, such as the Noor solar complex, will not only reduce lifecycle emissions but also enhance the green credentials of the bond issuance, improving its international rating and attractiveness to ESG investors.

6.1.2. South Africa: Advancing EV Adoption Amid Energy Mix Challenges

South Africa leads the continent in EV adoption, with an EVAR of approximately 0.5% and a relatively high CSD of 0.3 per 100,000 population. Yet, this progress is undermined by a coal-dominant energy regime and only 20% renewable generation. This contradicts the environmental rationale of EVs, a concern noted in both IEA reports and in our analysis, where RES and AE showed mixed predictive outcomes. The MLP framing suggests South Africa’s regime-level inertia, particularly Eskom’s dependence on legacy coal assets, is impeding a full transition. The policy path forwards involves electrifying urban charging infrastructure using decentralised renewables, particularly in provinces like Gauteng and Western Cape. The country’s success in fiscal incentives, highlighted by their influence in the ML models, should be extended to support EV leasing schemes and peri-urban two-wheeler fleets for logistics and last-mile delivery [52].

6.1.3. Kenya: Leveraging Renewable Energy Amid Access and Infrastructure Gaps

Kenya boasts a high renewable energy share, with renewables accounting for 86% of its energy mix, predominantly from hydropower sources [45]. Nevertheless, the socio-technical transition is hindered by limited electricity access, ranging from 47% to 79%, and a low CSD of approximately 0.14 per 100,000 population, a mismatch clearly demonstrated by our econometric findings linking AE and EVAR. Despite grid constraints, Kenya is rich in niche-level innovations, such as M-KOPA’s pay-as-you-go solar and microgrid deployment, which could form the backbone of decentralised EV charging. Policymakers should scale up off-grid solar charging networks tailored for two- and three-wheeler vehicles, particularly in informal transport sectors. These approaches would not only improve AE scores but also reflect the SHAP-identified significance of electricity access in ML models. Kenya also requires policy innovation to legalise and standardise EV components, improving investor confidence and consumer safety [122].

6.1.4. Nigeria: Addressing Systemic Barriers to EV Adoption

Nigeria faces significant obstacles in transitioning to electric mobility, with an EVAR of less than 0.1% and electricity access at approximately 50% [85]. The country grapples with chronic power outages and a negligible CSD of around 0.01 per 100,000 population. Econometric models show strong dependence on CSD and AE, while ML algorithms emphasise institutional and policy signals, including subsidies and tax waivers. Nigeria suffers from a broken socio-technical regime, where governance failures, electricity unreliability, and a lack of formal EV policy intersect. Using MLP framing, Nigeria remains trapped in a pre-transition state. Reform must begin with urban EV pilot corridors in Lagos and Abuja, using carbon credit schemes and duty exemptions to finance infrastructure and stimulate demand. Regulatory transparency, especially in procurement and community engagement via public awareness campaigns (e.g., radio, churches, transit unions), are vital to overcoming trust and knowledge deficits. This reflects our ML models’ 17.3% feature importance for subsidy and tax incentives, particularly in the neural network [8].

6.1.5. Africa’s Policy Diversity and the Case for Contextualised Transitions

Africa’s EV transition is marked by sharp national contrasts, as seen, shaped by differing levels of institutional maturity, grid reliability, and policy commitment. Morocco has strong renewable generation but suffers from regulatory delays and low charging infrastructure, reflecting a stable yet slow-moving regime. South Africa shows moderate EV uptake, aided by fiscal incentives and urban infrastructure, but its coal-heavy grid illustrates deep regime lock-in. Kenya exemplifies a more adaptive, bottom-up transition, supported by startup-led innovations and microgrid expansion yet constrained by uneven electrification. Nigeria, despite its economic size, remains policy-fragmented and pre-transition, with weak electricity access and limited institutional coordination.

These differences show that Africa’s EV future cannot rely on uniform solutions. Applying the Multi-Level Perspective (MLP), transitions must reflect each country’s position, whether niche-driven (Kenya), regime-stagnant (South Africa, Morocco), or systemically blocked (Nigeria). A successful pan-African EV strategy must therefore be context-specific, aligning national transitions with local capabilities, political realities, and decentralised infrastructure innovations.

6.1.6. Progress and Learnings from China and the EU

China’s electric vehicle (EV) success is primarily attributed to centralised policies, such as urban licence plate lotteries that restrict internal combustion engine vehicles and incentivise EV adoption, alongside strategic industrial policies that have positioned Chinese companies as a global battery leader. However, this urban-centric growth has exposed rural charging infrastructure gaps, which offer cautionary insight for Africa—replicating China’s urban-focused strategy could yield early wins but must be balanced with rural access plans to ensure inclusive EV uptake. Meanwhile, the European Union’s progress is supported by a diversified energy mix (31–40%) of renewables and high urbanisation rates (73–75%), creating a favourable environment for EV integration. Nonetheless, Europe’s residual dependence on fossil fuel-based electricity in certain regions raises concerns about the net sustainability of EVs. This duality suggests that African nations should simultaneously pursue clean energy expansion and EV adoption, avoid the pitfalls of urban exclusivity or carbon-intensive grids while learning from both China’s central planning and the EU’s systemic integration.

Table 9. Comparative synthesis of regional EV adoption: empirical drivers, barriers, and transition pathways.

Region	EVAR	Key Drivers	Major Barriers	Policy Type	MLP/Transition Stage
China	High	ST, GDP, AE	Rural infrastructure lag	State-led	Consolidated transition
EU	High	CSD, ST, URB	Fossil dependency in certain nations	Consumer-driven	Stabilised regime
Morocco	Medium	RES, CSD, URB	Urban–rural divide, bureaucratic delays	Centralised hybrid	Regime-aligned
South Africa	Medium	AE, URB, RES	Fossil-heavy grid, moderate incentives	Partial regime support	Fragmented transition
Nigeria	Very Low	CSD, AE, GDP	Grid unreliability, policy gaps	Nascent/Niche	Transition-inhibited
Kenya	Low	RES, CSD, AE	Low access to electricity	Microgrid-led (Niche)	Experimentation

compares EV adoption levels across case regions, identifying dominant variables from machine learning and econometric models. It also synthesises policy characteristics and classifies each country’s position within the Multi-Level Perspective (MLP) transition framework, offering empirical and theoretical insight into regional disparities.

6.2. Machine Learning: Relevance, Limitations, and Future Use in Policymaking

This study employed Random Forest, XGBoost, and neural networks to predict EV adoption using a variety of technical and socio-economic variables. XGBoost and SHAP were chosen for their robustness in variable interpretability, while neural networks demonstrated the capacity to model high-order, non-linear interactions. However, the limited sample size (n = 63) raises overfitting concerns, particularly in the neural network. Additionally, data sparsity, poor digitisation, and inconsistencies across African countries limit model portability. These challenges, common in Global South contexts, hinder the integration of ML into national policy planning. There is a pressing need for institutional investment in data systems, capacity building in ML techniques within ministries, and creating open-access datasets to enable robust, Africa-owned modelling. Until these gaps are addressed, ML’s policy relevance will remain under-realised.

6.3. Conclusions

This study bridges causal and predictive frontiers to offer a multidimensional analysis of electric vehicle integration across Africa, China, and the European Union. Combining fixed-effects panel regression with supervised machine learning models (Random Forest, XGBoost, and neural networks) and enhancing interpretability through SHAP analysis, it addresses the why and how behind EV adoption feasibility across regions with distinct infrastructural and policy landscapes.

Key findings reveal that charging station density is the most consistently influential determinant across econometric and ML models, followed by access to electricity and renewable energy share. Notably, the regression analysis uncovered a counterintuitive negative relationship between GDP per capita and EV adoption, particularly in high-income regions, suggesting the presence of behavioural, cultural, or saturation effects. Meanwhile, the SHAP values affirmed GDP’s predictive relevance, highlighting non-linear patterns missed by traditional models. These insights underscore the importance of integrating econometric causality with ML-based prediction to capture complex, real-world dynamics.

Regionally, the study finds that Southern Africa and North Africa exhibit the highest EV adoption feasibility among African sub-regions. At the same time, Nigeria and Kenya lag due to infrastructural deficits and policy inertia. The predictive models, achieving high accuracy (R² of 0.99 in neural networks), offer targeted foresight to policymakers and investors alike.

From a policy standpoint, the findings advocate for context-specific strategies: high-income regions should address behavioural barriers and infrastructure optimisation, while emerging markets require targeted subsidies, off-grid charging innovations, and localised industrial policies. The methodological integration used here can be replicated for other sustainability transitions (such as renewable energy deployment or clean cooking adoption), underscoring its broader utility.

This study contributes to the growing body of transdisciplinary sustainability research by demonstrating that combining econometric rigour with machine learning’s predictive power enhances interpretability and informs more effective and inclusive policy design.

6.4. Future Work

This study establishes charging station density (CSD) as the most influential predictor of EV adoption across econometric and machine learning models. However, future work should pursue a more multidimensional and causally robust approach. First, addressing the potential endogeneity of CSD, future research should apply instrumental variable (IV) techniques or lagged variable specifications to mitigate reverse causality bias in panel regressions. While fixed effects and time controls partially account for unobserved heterogeneity, the simultaneity between infrastructure deployment and EV uptake warrants more rigorous causal strategies.

Second, future work should construct continuous or ordinal indicators of policy intensity, such as EV-related spending per capita or composite fiscal incentive indices, to replace binary policy dummies, which lack granularity and risk bias. Incorporating these indicators across econometric and ML frameworks will enable more precise distinctions between structural and policy-driven effects.

Third, to explain observed model discrepancies (e.g., the GDP paradox), researchers should integrate consumer behavioural data, including survey evidence, sentiment analysis, and willingness-to-pay studies. This supports mixed-method triangulation, enriching quantitative models with insights into user preferences, trust in technology, and socio-cultural adoption barriers—especially in diverse contexts like Nigeria and China.

Methodologically, future studies should adopt panel Error Correction Models (ECMs) to capture long-run dynamics and non-linear regression tools (e.g., splines or threshold models) to identify tipping points in variables like CSD and electricity access. Expanding the dataset to include countries such as Ghana, Ethiopia, and Rwanda will improve the external validity of findings. Enhancing ML robustness through regularisation techniques (e.g., L2, dropout layers) and using scenario modelling for smart charging deployment in weak-grid contexts will also be valuable for infrastructure planning.

Finally, future analyses should disaggregate EV types, especially two- and three-wheelers, which dominate peri-urban African markets. This will improve the precision and inclusivity of EV policy recommendations, aligning more closely with local transport realities and SDG-aligned equity goals.