Decarbonizing the Building Sector: The Integrated Role of Environmental, Social, and Governance Indicators

Abstract

Climate change mitigation for the built environment has become a subject of greatest urgency, as buildings account for nearly 40% of total energy consumption and nearly one-third of total CO₂ emissions. While environmental, social, and governance (ESG) indicators are increasingly used to monitor sustainability performance, their collective role in impacting building-related emissions is yet largely under-investigated. The current research closes that gap through an examination of the ESG dimension–CO₂ emissions intersection of 180 nations from 2000 to 2022, in the hope of illuminating how environmental, social, and governance elements interact to facilitate decarbonization. The research is guided by a multi-method design, including econometric examination, cluster modeling, and machine learning techniques, which provide causal evidence and predictive analysis, respectively. The findings reveal that the deployment of renewable energy significantly reduces emissions, while per capita energy use and PM_2.5 air pollution exacerbate this effect. The social indicators show mixed results: learning, women’s parliamentary representation, and women’s workforce representation reduce emissions, while food production and growth among the lowest-income individuals demonstrate higher emissions. Governance demonstrates mixed results as well, with good regulation reducing emissions under specific conditions yet primarily supporting high-income countries with superior infrastructure. The examination of clusters reveals that ESG-balanced performance is retained by countries in the low-emission clusters, whereas decentralized ESG pillars are associated with higher emissions. Machine learning confirms the existence of non-linear effects and identifies PM_2.5 exposure and renewable energy deployment as the strongest predictors of the relationship. In summary, the findings suggest that successful policies for decarbonizing the built environment are constructed upon the consistency of environmental, social, and governance plans, rather than single steps.

1. Introduction

Climate change is a defining issue of the 21st century. The building sector produces nearly 40% of global energy demand and a third of CO₂ emissions [1]. ESG indicators are now recognized as standards of sustainability, yet their macro-scale policy application is limited [2]. Except for investment screens, ESG proxies institutional strength, innovation, and environmental performance. Healthcare and regulation reduce carbon intensity in industrialized sectors [3]. This study examines the connections between governance, education, healthcare, scientific productivity, and construction-sector emissions, utilizing IPCC data, econometrics, and machine learning [4]. Clustering and variable-importance mapping reveal cross-country differences and non-linear institution-emissions relationships. Research gaps are apparent: institutional and socio-ecological determinants are under-investigated, while technology diffusion and urban energy receive the most attention [5]. Findings indicate that effective governance attracts investors and that scientific productivity reduces emissions through innovation and policy. Paradoxically, greater R&D and governance accompany increased emissions, revealing a relationship between infrastructure growth and emissions [6]. Overall, ESG evidence reveals complex environmental performance determinants and policies that balance development and sustainability.

The article continues as follows: Section 2 reviews the current literature on how ESG features map onto emissions in buildings, identifying gaps in integrated studies. Section 3 presents a multi-method approach that combines econometrics, clustering, and machine learning across a global sample from 2000 to 2022. Section 4, Section 5 and Section 6 study environmental, social, and governance features in turn. They find unique but correlated effects on emissions. Section 7 synthesizes findings while highlighting ESG coherence. Section 8 and Section 9 present limitations and conclude on the relevance of ESG to future emission mitigation policy.

The article also contains several appendices dedicated to exploring metric dimensions and analyzing data characteristics: Appendix A provides a comprehensive description of the variables used in the ESG analysis, along with data sources and definitions. Appendix B presents the summary statistics for the ESG model’s Environmental (E) component. Appendix C releases the summary statistics, as well as the robustness checks, pertaining to the Social (S) dimension. Appendix D provides a summary of the results for the Governance (G) indicators, as well as their relationship with building-sector emissions. Appendix E presents diagnostic tests for autocorrelation and heteroscedasticity, ensuring the robustness of the econometric estimations. Appendix F describes the K-Nearest Neighbors (KNN) regression algorithm used, under the part dedicated to machine learning, the hyperparameter optimization process.

2. Literature Review

E-Environmental. Several works criticize the application of environmental certification tools at a symbolic level and a low level of linkage to actual environmental performance. Ref. [7], for instance, found that ESG tools applied within Florence’s residential policy did not cause noticeable decreases in gas or energy consumption. Ref. [1] are similarly critical of the BREEAM system’s application within the UK, as it places too much emphasis on environmental parameters while losing sight of equity and governance themes. In India, green certification is shown by [8] to generate higher rent levels, which are thus socioeconomically exclusionary. The separation between environmental performance and communication is further exacerbated by the minimal attention to facility-level practices in the post-construction phase. Whilst some works note positive practices, such as the use of prefabricated infrastructure for energy use and waste reduction [9], these benefits remain largely negated mainly due to mixed facility-level practices and maintenance [10,11]. Whilst ref. [10] successfully demonstrates that ESG-certified Class A office stock within the urban area of Madrid is a reality and firm energy savings occur, these are dependent upon repeated operating disciplines, such as metering and tenant engagement. This is reiterated in [12], who emphasize that successful facility practice is a necessity if sustainability practices are to be more than a symbolic application, especially within restrictive bureaucratic regimes. It has been posited that digital innovation has a solution. Ref. [13] speculates about a digital monitoring ESG system that applies to commercial stock in real estate, but its hypothetical nature, coupled with actual non-verification, limits its current application. Ref. [14] promote the applications of technology but without any linkages between technology applications and their resultant sustainability impacts. Ref. [15] concur with this disconnect and note that Malaysian ESG reporting still tends to be at a rather transformational level, rather than a reputation level, in its operational application. Other environmental concerns highlight researchers who study wider energy and pollution infrastructures. Ref. [16] cooperate on materials and lifecycle emission traceability but highlight the basis of materials sustainability in successful use and procurement systems. Likewise, ref. [17] outline ESG as a “change engine” of construction but highlight common operating processes between procurement and waste treatment. At the same time, these works suggest that environmental performance within ESG thinking depends not only on innovation or design but also on process practices designed in and repeated, energy source quality, and ongoing administrative resolve.

S-Social. The social ESG of the built environment is conceptually applied but lacks theoretical depth at operational levels. The social ESG literature focuses primarily on design intention, occupant wellness, and occupant health but underscores the long-term socially desirable aspects that buildings do. Ref. [18] cite air and light performance research and studies on occupant satisfaction but remain constrained within first-order design, missing tenant use and after-occupancy engagement—the sustainability drivers where they have the most impact in the long term. Equity concerns apply equally to criticism claims. Ref. [8] conclude that India’s green-awarded buildings rent at higher grade levels and are consequently exclusionary to poorer rent grades, thus exacerbating urban poverty. By analogy, ref. [19] cites a dilemma where ESG uptake still largely accrues within cash-based revenues, but environmental resilience is seldom accompanied by investment in the social or governance quadrants. Ref. [20] cite environmental and societal forces within real assets but provide no further explanation at the operational level beyond facilities. Retrofit and inclusion sustainability reports within comparisons beyond their own region. Ref. [21] suggest that ESG retrofit within South African facilities might be initiated more often due to energy blackouts (e.g., load shedding) rather than sustainability targets and will consequently not be maintained in the long term. Ref. [22] further suggest that larger practicum and service-provider architecture and engineering departments largely direct provision primarily within the SDGs during preparation intervals but overlook the use of ESG findings some years later. Ref. [23] sketch further ESG research mapping within commercial real property but concede that facility-level completion is poorly expounded. Social ESG implementation deterrent concerns often provide structural barriers. Ref. [24] concludes that tomorrow’s knowledge, ignorance, and operational incompetence remain the main deterrents to green uptake within Kenya and calls for firm management reinforcement and institutional reinforcement. These deterrents predict ESG policy change extending beyond disclosure and design but broader, co-participative engagement and governance. Even if research works such as ref. [25] assume user behavior modelers and facilities managers to be promoters of ESG outcomes, only frail empirical evidence can be provided, and experimental confirmation needs to be carried out. G-Governance. Governance is ESG’s third pillar and key to long-term sustainability within the built environment but has been revealed to have an inadequate operational governance mechanism. Regulatory fragmentation and inability to maintain follow-up on constructed buildings are common attributes of certain studies.

G-Governance. Systemic ESG issues, such as inefficient regulatory coordination and regulatory gaps, are highlighted in [26]. Ref. [27] clearly suggest that operational data requirements remain a consideration; however, minimal consideration continues to be given to the long-term ESG performance effects emanating from facility-level operations. Physical elements of design often drive ESG implementation, but ref. [28] finds that it is further influenced by administrative rituals, such as communication between relevant parties and energy use tracking, thereby encapsulating building management within governance effectiveness. Contract provisions that interface with ESG obligations at inception, between design and everyday use, are featured in [29]. This continues to be carried forward by [30], who suggest that Chinese agendas incorporate ESG into property agendas, but making such incorporation a reality in everyday operational use remains elusive. But policy and financial arrangements fail in governance. Investment programs, such as sustainable investments, offer positive financial returns but do not necessarily raise doubts about whether they lead to reform at a facility level [31]. Likewise, ref. [32] note that, while tenants account for ESG certificates within rent premiums, administrators of buildings seldom receive assessments regarding their compliance with ESG standards in the long term, resulting in a gap between market incentives and governance accountability. A green certificate often fails to institutionalize governance in most instances. Ref. [33] note that standardized certificate requirements mostly overlook the operational dimension—the dimension where attention is focused on documentation but not on use. Only in a similar manner do ref. [34] identify gaps between taxonomy at an EU level and schemes such as BREEAM-SE, inducing distress in building managers in making long-term compliance plans. Refs. [35,36] go further to identify how reform efforts within facilities’ sectors primarily call upon ESG but without designing operationally successful observance systems. Furthermore, managers’ roles within buildings and facilities’ governance are repeatedly mentioned in some studies. Ref. [37] identify an increase in property uptake but overlook the facilities’ use of smart technology, while ref. [12] suggest that post-construction monitoring often falls victim to administrative inertia. Ref. [11] again sarcastically identifies ESG compliance and repeatable operating commitments as supporting governance’s role at a building level. If there are no enforcing processes in place, then the ESG will be susceptible to becoming a label rather than a change-based governing process.

Governance, ESG’s third pillar, is central to making the constructed environment sustainable in the long term; however, operating governance mechanisms are surprisingly absent in the literature. Some of the literature refers to regulatory gaps and disjointed follow-up after construction. Ref. [26] present evidence of structural ESG challenges, such as non-coordination and regulatory gaps. Ref. [27] refer to operational data requirements that are increasingly accepted; however, facility-level, everyday processes are essentially left out of consideration for their contribution to long-term ESG performance. Ref. [28] reveals that ESG is not only administratively applied to the physical side of processes but also in processes such as tracking edifice management and energy, and this is at the core of governance performance. Ref. [29] suggest that contractual clauses bring ESG requirements into conception and link design and everyday implementation. This is also reported by [30], who note that ESG adoption is becoming integrated but does not yet seem to be translating into regular operational practice. However, governance also raises questions about policy and the limits of finances. Ref. [31] reveals evidence that sustainable investment funds yield better monetary returns, but does not raise questions about whether these investment increments have a tangible impact on governance at the executive level. Likewise, ref. [32] report that tenants pay premiums for ESG certification, and facade managers work mainly but are seldom measured, on compliance with ESG requirements; an increasingly significant gap is emerging between market incentives and accountability in governance performance. Green certification does not institutionalize governance. Ref. [33] conclude that certified requirements tend to exclude the operational dimension—the theory of documentation, instead. Likewise, ref. [34] identify inconsistencies between the EU taxonomy and local sets, such as BREEAM-SE, which makes it challenging for edifice managers to devise long-term compliance strategies. Refs. [35,36] further outline, based on construction reform, often relying on ESG but not converting it into operationally reflective mechanisms of oversight. Lastly, facilities governance and the positions of building managers receive attention in several remarks. Ref. [37] refer to innovative technological development at a property level but not facility managers’ application of such technology. Ref. [12] note that post-construction monitoring is often hindered by administrative routine inertia. Ref. [11] personally equates ESG compliance with reproducible operational necessities and reiterates the governance function at a building level. Without enforceable, evolutionary standards, ESG risks amount to a label but not a transformational governance role (

Table 1. ESG dimensions and their property-level implications in the building sector.

ESG Dimension	References	Main Results	Comparison with Our Study
E-Environmental	[1,7,8,9,10,11,12,13,14,15,16,17,20,28,30,31,32,35,36,37,38,39,40].	ESG certifications and technologies are often disconnected from actual emission outcomes; operational routines are key but under-applied.	Our study directly quantifies the impact of environmental ESG indicators (e.g., renewable energy, air quality) on building emissions, validating their role with econometric models.
S—Social	[8,18,19,20,21,22,23,24,25,29].	Social aspects are conceptually acknowledged but rarely implemented operationally; issues of equity, inclusion, and behavior are underdeveloped.	Our study empirically links social variables (e.g., gender equity, income equality, labor participation) to emission levels, revealing both mitigation and rebound effects.
G—Governance	[1,10,11,12,15,17,19,24,26,27,28,29,30,31,32,33,34,35,36,39,40,41,42,43,44].	Governance is often fragmented; regulation and institutional routines are disconnected from long-term ESG outcomes.	Our study finds that governance indicators (e.g., government effectiveness, political stability) may paradoxically correlate with higher emissions, especially in higher-income countries, highlighting the complexity of governance–ESG links.

).

3. Modeling Building-Related Emissions Through ESG Dimensions: A Multi-Method Analysis Using Econometrics, Clustering, and Machine Learning

The built environment and climate change cannot be separated. The building sector accounts for approximately one-third of global CO₂ emissions [1]. The current research examines building-sector CO₂ emissions against the ESG prism. Multi-method research combines econometrics, clusterization, and machine learning. The panel econometrics comprises 180 nations (2000–2021). Fixed- and random-effect regressions are linked between building emissions and environmental, social, and governance indicators [45,46]. National typologies are highlighted by cluster analysis. Normalized ESG–emissions data cluster states into groupings with similar structures or fragmented profiles [47,48]. Supervised machine learning also preserves non-linear relationships. Random Forest, SVR, k-NN, and Boosting predict and rank ESG attributes by importance [49,50]. Results converge. Balanced ESG pillars are associated with lower emissions. Broken governance or weak social equity are associated with higher emissions [51,52]. Results show interdependence. Environmental action becomes effective only if governance and social equity co-evolve. Disjointed ESG pillars are ineffective for lowering emissions. This triangulative approach clears up causal, predictive, and structural dynamics. The strong triangulation of econometrics, cluster analysis, and machine learning enhances validity [53,54,55]. Policy prescriptions are immediate: decarbonization of buildings is achievable only with governance quality, social inclusiveness, and environmental enforcement simultaneously. Fragmented ESG strategies are unsuccessful. See Figure 1. A complete description of the variables used to estimate the ESG model is provided in Appendix A.

4. Decoding Building Emissions: Environmental Drivers in an ESG Framework (2000–2022)

Estimating the environmental determinants of building-related carbon dioxide emissions (BCE) holds the key to synchronizing the policies of decarbonization with ESG (environmental, social, and governance) targets. This section introduces the first phase of the empirical work on analyzing the effect of environmental indicators on BCE during the years 2000–2022 in 180 countries. With the help of fixed and random effects panel regressions, the study evaluates the effect of prominent energy and pollution-based indicators—access to clean fuels (CFTC), electricity access (ELEC), per capita energy usage (ENUC), particulate matter pollution (PM_2.5), and renewable energy contribution (RENC)—on building-related emissions. The outcomes serve as a quantitative underpinning for the interpretation of the environment’s sustainability through the lens of ESG and present direct and indirect channels through which energy systems impact outcomes on emissions. This model forms a crucial diagnostic tool through which one can assess the environment’s performance as well as the structural dynamics shaping carbon intensity in the built environment on an evidence-based trajectory. The metric characteristics of the variables used to estimate the ECB component with respect to the E-Environmental variables within the ESG model are summarized in Appendix B. Diagnostic tests for autocorrelation and heteroscedasticity for the econometric estimates are provided in Appendix E.

4.1. Modeling the Environmental Determinants of Building Emissions: A Global ESG-Informed Econometric Analysis

To quantify the environmental determinants of building-related carbon emissions (BCE), this section presents a panel regression model estimating the effects of key energy-related indicators. Drawing on data from 180 countries over the period 2000–2022, the model incorporates variables aligned with the “E” pillar of the ESG framework, including clean fuel access, electricity availability, per capita energy use, PM_2.5 exposure, and renewable energy consumption. The analysis employs both fixed and random effects estimators, with results providing robust statistical insights into how national energy infrastructure and environmental conditions shape BCE outcomes globally.

Specifically we have estimated the following equation:

$$B{CE}_{it}={\alpha }_1+{\beta }_1{\left(CFTC\right)}_{it}+{\beta }_2{\left(ELEC\right)}_{it}+{\beta }_3{\left(ENUC\right)}_{it}+{\beta }_4{\left(PM25\right)}_{it}+{\beta }_5{\left(RENC\right)}_{it}$$

where i = 180 and t = 2000–2022. The results are shown in the following

Table 2. Results of the panel data.

	Random Effects (GLS), Using 990 Observations, Dependent Variable: BCE			Fixed Effects, Using 990 Observations Dependent Variable: BCE
	Coefficient	Std. Error	z	Coefficient	Std. Error	t-Ratio
const	9.3166	11.6274	0.8013	10.2987	10.6061	0.9710
CFTC	0.318025 ***	0.0842985	3.773	0.355114 ***	0.0893023	3.977
ELEC	−0.245591 ***	0.0923988	−2.658	−0.273064 ***	0.0964139	−2.832
ENUC	0.00269113 ***	0.000691466	3.892	0.00284832 ***	0.000735797	3.871
PM25	0.664678 ***	0.131472	5.056	0.691987 ***	0.144470	4.790
RENC	−0.489147 ***	0.104604	−4.676	−0.513324 ***	0.112617	−4.558
Statistics	Mean dependent var		2.480			2.480
	Sum squared resid		5,836,005			90,688
	Log-likelihood		−5702			−3640
	Schwarz criterion		11,445			8.413
	Rho		0.756174			0.756174
	S.D. dependent var		7.753
	S.E. of regression		7.697
	Akaike criterion		11,416
	Hannan–Quinn		11,427
	Durbin–Watson		0.324935
	LSDV R-squared					0.984747
	LSDV F(163, 826)					3.271
Test	‘Between’ variance = 4971.73, ‘Within’ variance = 91.6047, mean theta = 0.935247, Joint test on named regressors—Asymptotic test statistic: chi-square(5) = 109.56 with p-value = $5.0646 \times {10}^{-22}$			Joint test on named regressors— Test statistic: F(5, 826) = 20.981 with p-value = P(F(5, 826) > 20.981) = 8.85584 × 10−20
	Breusch–Pagan test—Null hypothesis: Variance of the unit-specific error = 0, Asymptotic test statistic: chi-square(1) = 3155.73 with p-value = 0			Test for differing group intercepts—Null hypothesis: The groups have a common intercept Test statistic: F(158, 826) = 319.607 with p-value = P(F(158, 826) > 319.607) = 0
	Hausman test—Null hypothesis: GLS estimates are consistent, Asymptotic test statistic: chi-square(5) = 5.09644, with p-value = 0.404224

Note: *** p < 0.01.

.

Building-sector CO₂ emissions (BCE) are a key environmental (E) metric within the ESG framework, making use of 180 countries’ (2000–2022) panel data. The approach relates BCE to accessibility, quality, and composition of energy, including cleaner fuel transition (CFTC), electricity accessibility (ELEC), per capita energy use (ENUC), exposure to PM_2.5, and use of renewable energy (RENC). The results establish that CFTC stimulates social performance while increasing BCE unless fueled by low-carbon energy [48,54]. On the contrary, a transition to electricity reduces the rate if electricity is generated from cleaner sources, thereby maintaining ESG co-benefits for universal electricity accessibility [56]. Increasing ENUC matches with rising emissions, making efficiency policies relevant for high-consumption nations. From a governance (G) perspective, proper planning, conservation incentives, and regulating energy-intensive tech are central [57]. PM_2.5 exhibits a high and desirable correlation with BCE, supporting the convergence between environmental damage and public health hazards and integrating the E–S pillars within ESG. On the contrary, the use of renewable energy (RENC) shares a high negative correlation with emissions, validating green energy as a prime avenue for a resilient future. The fixed-effect formulation yields a high R² value for the panel estimates, validating the use of BCE as a robust metric for ESG while accounting for country-wise dissimilarity. The results, in general, reveal that electrification and renewables lead to abatement of emissions, while energy efficiency is crucial for offsetting per capita demand growth [54,56]. The linkage between BCE and PM_2.5 provides a compelling rationale for integrating policies on both air quality and climate that need to be globally instituted, especially for emerging economies with lower regulations [48]. The results position BCE at the core of ESG performance, where the building sector’s performance is dependent on accessibility, quality, efficiency, and governance [47,57].

4.2. Evaluating Clustering Strategies for Building Emissions: A Multi-Metric Comparison

To disclose building-related carbon emission (BCE) global patterns and underlying country groupings that are structurally regular, we employ clustering techniques. Six methods are systematically compared across ten standardized performance measures, which vary from statistical fit and cohesion to quality of separation. The comparative benchmarking identifies the best possible segmentation technique for emission profiles, ensuring appropriate classification across countries by BCE structure. Moving beyond the selection of techniques, the application of clustering supports the interpretational strength of diagnostics that are ESG-based through correlation with environmental indicators. The approach possesses both technical sophistication and analytical coherence and constitutes a tool that scales for comparing emission dynamics across diverse country settings (Figure 2).

Six cluster procedure performances were compared against ten standardized indices (R², AIC, BIC, silhouette, Dunn, maximum diameter, minimum separation, entropy, and Calinski–Harabasz), each on a 0–1 scale. The density-based algorithm performed exceptionally well, achieving a mean score of 0.599, surpassing the optimal AIC and BIC (1.0) and achieving perfect separation indices (1.0), which is typical for well-defined and tightly clustered data [58]. Its silhouette value was high (0.8), beaten only by hierarchical (1.0), and satisfactory performance on maximum diameter (0.696) and Pearson’s γ (0.489). On R² and entropy, they performed poorly, although these indices are better suited for non-linear cluster shapes and are more suitable for DBSCAN [59]. Hierarchical was best in terms of R², silhouette, and γ while performing poorly in terms of separation, suggesting potential overfitting. The neighborhood- and model-based techniques showed one-dimensional strengths and mixed performance. The density-based approach showed the best-rounded performance and handles noisy or cluttered data the best.

Evaluating Cluster Quality and Structure in ESG-Driven Density-Based Analysis

This section reports on results from density-based environmental and energy indicator clustering relevant to ESG (environmental, social, governance). The leading indicators—clean fuel access (CFTC), electricity access (ELEC), per capita energy use (ENUC), PM_2.5 exposure, and renewable energy use (RENC)—produced several sharply differentiated clusters and one noise group. The clusters varied in size, shape, and separation, as indicated by the within-cluster sum of squares (WSS), explained heterogeneity, and silhouette statistics. The latter two statistics corroborate both statistical stability and interpretational soundness of the solution. The results link country energy infrastructures and environmental configurations with building-related CO₂ emissions (BCE) and reveal clustering as a gauge for benchmarking and ESG-centered sustainability reporting (

Table 3. Cluster characteristics from density-based clustering on ESG-linked energy and environmental indicators.

Cluster	Noise Points	1	2	3	4	5
Size	1	949	6	6	20	8
Explained proportion within-cluster heterogeneity	0.000	0.999	8.471 × 10−4	2.529 × 10−5	1.102 × 10−5	4.580 × 10−4
Within sum of squares	0.000	4.584	3.888	0.116	0.051	2.103
Silhouette score	0.000	0.347	0.632	0.968	0.985	0.820

).

The DB model identified one noise point and five distinct clusters. The cluster quality was determined based on size, explained variance, WSS, and silhouette measure. The noise point, a typical DBSCAN characteristic [60], exhibited a WSS and silhouette value of 0, indicating evident separation from the cluster. Cluster 1, with 949 data points, exhibited high explained variance (0.999) but low isolation (silhouette = 0.347) and was therefore observed to create an overlap boundary artifact [61]. Clusters 2–5 were much smaller, each having between 6 and 20 data points. Cluster 4 was an outlier (WSS = 0.051; silhouette = 0.985), and Clusters 3 (silhouette = 0.968), 5 (0.820), and 2 (0.632) also passed the 0.5 interpretability cutoff [62]. The entire cluster display showed primarily spherical, compact shapes with acceptable isolation, although resolution within the noise was low. Cluster 1 showed lower discrimination compared to the small clusters, which exhibited sharper and clearer groupings (Figure 3).

Examining the environmental (E) pillar within ESG, the model correlates building-related CO₂ emissions (BCE) with a set of five indicators: clean cooking fuels (CFTCs), electricity access (ELEC), per capita energy use (ENUC), PM_2.5 exposure, and renewable energy share (RENC). The strategy is consistent with integrated sustainability indices [63,64]. The nations fall within six groups. Cluster 0 represents energy poverty, characterized by low electricity, high PM_2.5 levels, low renewable energy sources, and a moderate BCE. Cluster 1 has balanced and low BCE regimes. Cluster 2 is more closely associated with high pollution and low renewable energy sources and thus receives a moderate BCE from fossil-heavily dominated urban–industrial systems. Cluster 3 achieves high electricity and renewables, as well as moderate BCE, due to efficiency gains. Cluster 4 has very low BCE and PM_2.5 levels, which is likely due to its nuclear or hydrocarbon dependence. Cluster 5 balances clean energy and low pollution yet maintains a moderate BCE, characteristic of efficient, high-income fossil fuel economies. Overall, each cluster plots distinct energy–environment portfolios, testifying to how integrating indicators strengthens the ESG diagnostics (Figure 4).

4.3. Explaining Carbon Emissions in the Built Environment: A Comparative Machine Learning Approach

This section presents a rigorous comparison of multiple machine learning regression models—ranging from ensemble methods to neural networks and proximity-based learners—evaluated through standardized performance metrics. By identifying the most reliable and accurate algorithm for predicting building-sector emissions, the analysis sheds light on model suitability and interpretability, with a particular focus on the role of local energy consumption, pollution exposure, and access variables in shaping emissions profiles (Figure 5).

Model performance was evaluated using MSE, RMSE, MAE, MAPE, and R², with the metrics normalized and inverted for ease of comparison. Composite scores showed that k-Nearest Neighbors (k-NNs) strongly dominated the rest, with error rates being low and an R² level of 1.0, consistent with a perfect explanation of variance. This is the fit between k-NN’s similarity principle and patterned data structures, as well as its adaptiveness with weak functional assumptions [65]. Despite the risk of overfitting driven by a perfect score, k-NN performed strongly across various error measures. Compared to that, Boosting yielded similar results with higher errors and lower R² [66]. Decision Trees overfitted (low MSE but R² = 0), while linear, regularization, and neural network models lacked complexity. Random Forest was stable without localized precision. Overall, k-NN’s high level of accuracy, along with its parsimonious and flexible character, highlights its strength for locally regularized data in empirically tested, simple models, consistent with prior results in software engineering estimation [67]. See Figure 6.

By applying k-Nearest Neighbors (k-NNs) with dropout loss (RMSE on 50 permutations), we determined significant environmental associates for building-related CO₂ emissions (BCE) in the ESG “Environment” pillar. The power consumption of fossil fuel use, combined with simultaneous air conditioning and heating, was the top contributor to air pollution (PM_2.5) (90.969), highlighting the significant impact of fossil fuel use on power, heating, and air conditioning consumption, as well as the climate–health hazard associated with decarbonization policy [62]. Consumer energy consumption of renewable energy (RENC) came a close second (76.284), indicating a reduction in carbon intensity and facilitating policy-mandated decarbonization [65]. Clean fuel for cooking (CFTC) came in a distant third (68.672), associated with biomass dependence and the infrastructural vulnerability of low-income countries [68]. Energy use per capita (ENUC) (63.747) delivered demand-side increases in emissions, while electricity penetration (ELEC) delivered the weakest effect (40.879) due to near-universal penetration. Overall, PM_2.5 and RENC were chief initiators, followed by CFTC and ENUC, and followed by marginal ELEC. The implications are for renewable installation, clean fuel consumption, and air quality management as central BCE mitigation activities within ESG policy. See Figure 7.

Additive decompositions for k-NN predictions of building-related CO₂ emissions (BCE) provide contributions for clean fuels (CFTC), electricity access (ELEC), energy use per capita (ENUC), PM_2.5, and renewables (RENC) compared to a base level of 25.434. For Case 1 (BCE = 12.257), maximum reduction (−16.664) was brought on by low PM_2.5, in line with evidence for air quality and emissions associations [62]; ENUC and RENC reduced, while CFTC and ELEC increased emissions by higher accessibility [65]. For Case 2 (BCE = 0.500), a reduction was observed for PM_2.5 (−17.110) and ENUC/RENC, offset by CFTC and ELEC, which is typical for low-emitting systems. For Case 3 (BCE = 18.570), a notable reduction by ENUC (−13.891) was offset by a favorable effect by RENC (+9.796) and ELEC. For Cases 4–5 (BCE = 4.024), the high negative contributions for ELEC (−20.170; −18.398) indicate a preference for low-carbon generating alternatives, such as hydro. Generally, PM_2.5 and RENC are repeated movers for BCE reduction, while accessibility-related indicators (CFTC, ELEC) increase or decrease emissions according to system efficiency and mix. The exploratory findings for these jobs in ML reveal clean energy and air quality as the key factors for ESG policy [68]. See Figure 8.

5. Equity, Participation, and Emissions: Social Determinants of Building-Sector CO₂

This paper examines the linkage between building-sector CO₂ emissions (BCE) and social development indicators within the context of ESG, with a focus on the oft-overlooked “Social” pillar. Using panel econometrics and machine learning, we quantify the impact of distributive income inequality, labor force participation, educational parity, food output, and female political empowerment on the level of emissions. The results demonstrate robust relationships across specifications: distributive income inequality and higher food production are linked with higher emissions, while schooling and female political representation are linked with reductions. The relationships reveal that social equity, economic inclusion, and institutional representation are stimulants to sustainable environmental performance and not just moral prescriptions. The integration of social considerations into ESG assessment enhances predictive strength and highlights how inclusive policy directly impacts long-term decarbonization trajectories. The metric characteristics of the variables used to estimate the ECB component with respect to the S-Social variables within the ESG model are summarised in Appendix C. Diagnostic tests for autocorrelation and heteroscedasticity for the econometric estimates are provided in Appendix E.

5.1. Social Dimensions of Carbon Emissions: A Panel Data Approach to the Building Sector

This subsection applies panel data econometrics to estimate the degree to which social considerations drive building-sector CO₂ emissions (BCE). From a universe of 180 nations and over 1200 observations, drawn from a global dataset, the research tests associations between emissions and a group of five indicators: food production (FOOD), female parity in primary schooling (GPIE), income share for the bottom 20% (INC20), labor force participation (LABF), and women in legislatures (WPAR). Fixed- and random-effect regressions forecast the size and significance of impacts. Results show that country-level emission profiles are shaped by social equity, accessibility of schooling, and political representation, demonstrating the importance of social dimensions as a driver of ESG-related sustainability, rather than a side concern.

Specifically, we have estimated the following equation:

$$\mathrm{B}\mathrm{C}{\mathrm{E}}_{\mathrm{i}\mathrm{t}}=\mathsf{\alpha }+{\mathsf{\beta }}_1{\left(\mathrm{F}\mathrm{O}\mathrm{O}\mathrm{D}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_2{\left(\mathrm{G}\mathrm{P}\mathrm{I}\mathrm{E}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_3{\left(\mathrm{I}\mathrm{N}\mathrm{C}20\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_4{\left(\mathrm{L}\mathrm{A}\mathrm{B}\mathrm{F}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_5{\left(\mathrm{W}\mathrm{P}\mathrm{A}\mathrm{R}\right)}_{\mathrm{i}\mathrm{t}}$$

where i = 180 and t = 2000–2020 (

Table 4. Impact of social indicators on building-related CO₂ emissions: fixed-effect and random-effect panel regression results (2000–2020).

	Fixed Effects, Using 1246 Observations Dependent Variable: BCE			Random Effects (GLS), Using 1246 Observations Dependent Variable: BCE
	Coefficient	Std. Error	t-Ratio	Coefficient	Std. Error	z
const	46.5798 ***	8.08670	5.760	36.0405	10.5431	3.418
FOOD	0.0863615 ***	0.0236010	3.659	0.0868547	0.0233877	3.714
GPIE	−0.00173333 **	0.000727140	−2.384	−0.00171224	0.000720050	−2.378
INC20	0.454688 **	0.215227	2.113	0.451706	0.213527	2.115
LABF	−0.323534 ***	0.114301	−2.831	−0.299359	0.111774	−2.678
WPAR	−0.169214 ***	0.0504386	−3.355	−0.169434	0.0499015	−3.395
Statistics	Mean dependent var		31.32247	Mean dependent var		31.32247
	Sum squared resid		73,866	Sum squared resid		9479
	LSDV R-squared		0.992108	Log-likelihood		−7335.684
	LSDV F(142, 1103)		976.5221	Schwarz criterion		14,714
	Log-likelihood		−4311.281	Rho		0.727702
	Schwarz criterion		9641.822	S.D. dependent var		86.70740
	Rho		0.727702	S.E. of regression		87.39678
	S.D. dependent var		86.70740	Akaike criterion		14,683
	S.E. of regression		8.183426	Hannan–Quinn		14,694
	Within R-squared		0.032513	Durbin–Watson		0.486805
	p-value (F)		0.000000
	Akaike criterion		8908.562
	Hannan–Quinn		9184.263
	Durbin–Watson		0.486805
Tests	Joint test on named regressors— Test statistic: F(5, 1103) = 7.41335 with p-value = P(F(5, 1103) > 7.41335) = 7.51925 × 10−7			‘Between’ variance = 6423.31, ‘Within’ variance = 59.2827, mean theta = 0.953035, Joint test on named regressors—Asymptotic test statistic: chi-square(5) = 36.391 with p-value = 7.93233 × 10−7
	Test for differing group intercepts— Null hypothesis: The groups have a common intercept Test statistic: F(137, 1103) = 1002.32 with p-value = P(F(137, 1103) > 1002.32) = 0			Breusch–Pagan test—Null hypothesis: Variance of the unit-specific error = 0 Asymptotic test statistic: chi-square(1) = 5098.24, with p-value = 0
				Hausman test—Null hypothesis: GLS estimates are consistent, Asymptotic test statistic: chi-square(5) = 2.65474 with p-value = 0.753031

Note: *** p < 0.01, ** p < 0.05.

).

This paper positions building-sector CO₂ emissions in the ESG matrix with a focus on the overlooked “Social” pillar. Using panel econometrics with fixed- and random-effect specifications, we test four social indicators—gender parity in primary schooling (GPIE), income share for the poorest 20% (INC20), labor force participation (LABF), and women in parliament (WPAR)—for their roles in signaling educational equity, income distribution, labor inclusiveness, and political representation. Results reveal robust associations. Female schooling (GPIE) is associated with lower pollution, confirming the value of schooling in sustainability [69]. Political representation for women (WPAR) also reduces pollution, corroborating evidence that female empowerment produces cleaner building codes, energy standards, and solar subsidies [70,71]. Increasing participation in the labor force (LABF) has a negative impact on residential energy use, suggesting that jobs lead to greater efficiency [72]. Increasing incomes for the poorest (INC20), however, drive up pollution, since new consumption creates more carbon-demanding use from illumination and domestic appliances [73]. Poverty reduction thus demands simultaneous investments in solar infrastructure and energy literacy. Generally, social equity, inclusiveness, and political representation are key drivers of the low-carbon transition, not adjuncts. Building on data for 180 countries over one generation, the study confirms the universality of these associations [74]. Low-carbon inclusivity demands the union of social justice and environmental defense.

5.2. Clustering Social Determinants of Emissions: An Evaluation of Algorithmic Performance

This research positions building-sector CO₂ emissions within the ESG framework, aiming to address the often-overlooked “Social” pillar. With fixed- and random-effect panel econometrics, we study the effect on CO₂ emissions from female parity in primary schooling (GPIE), income share for the poorest 20% (INC20), labor force participation (LABF), and women in parliament (WPAR) as proxies for educational equity, income distribution, labor inclusion, and political voice. Results show high associations. GPIE is significantly linked to negative CO₂ impacts, highlighting the importance of female education for sustainability [69]. Likewise, a negative effect emerges for WPAR, supporting the results that female political leadership is associated with greener building codes, energy codes, and green and renewable subsidies [70,71]. A correlation emerges between residential energy consumption and LABF, indicating a link to job efficiency [72]. INC20 appears to be significantly linked with CO₂ increases, as higher incomes for the poor result in a boost to fossil-intensive consumption [73]. The result supports the notion that the “Social” pillar is central, not peripheral, to ESG. Across over 180 countries spanning multiple generations, the research establishes the universality of these associations [74], thus supporting the notion that low-carbon transitions that are inclusive depend on aligning social equity and environmental protection. See Figure 9.

We experimented with six clusters—density-based, Fuzzy C-means, hierarchical, model-based, neighborhood-based, and Random Forest—with ten validity indices (R², AIC, BIC, entropy, silhouette, maximum diameter, minimum separation, Pearson’s γ, Dunn, and Calinski–Harabasz). They measure fit, error minimization, and quality structure in high-dimensional ESG data. Model-based and Random Forest excelled in R² (0.003), although their superiority over other techniques was slight, while density-based and neighborhood-based performed the worst (0.001). Random Forest performed best on AIC and BIC (0.016; 0.017), with model-based a close second (0.011; 0.012), indicating that explanatory strengths predominated over penalties for complexity [75]. Separation and silhouette measures were close to zero across techniques, with density-based being slightly higher (0.004), indicating a weak cluster definition. Dunn indices were unanimous zero, supporting reports that compaction in ESG clusterings was low [76]. Entropy measures varied on a small scale, with model-based (0.008) and Random Forest (0.007) best. Calinski–Harabasz was a flat 1.0, likely due to normalization. Overall, Random Forest was the best-balanced performer, excelling in AIC, BIC, and entropy, while tying with model-based on R². Fuzzy C-means, hierarchical, and neighborhood-based were inconsistent performers, with near-zero separation and validity [77].

Decoding Emissions and Equity: A Density-Based Clustering Approach to ESG Social Metrics

The research employs a density-based cluster model to analyze an ESG-related environmental dataset in the context of sectoral CO₂ emissions and certain societal development indicators. The cluster structure identifies four clusters and one outlier, covering both leading and niche configurations in the societal development and emissions process. The evaluation determines compactness and distinctiveness within the cluster, as well as within-cluster heterogeneity, using Silhouette indices and the respective sum of squares, along with their explanatory power. The model acknowledges an implicit consideration of the interrelation between environmental impact and societal dimension, thereby providing a more refined explanation for the diverse sustainability trajectories across nations. See

Table 5. Cluster characteristics from density-based clustering of ESG and building CO₂ emissions data.

Cluster	Noise Points	1	2	3	4
Size	1	1188	21	15	21
Explained proportion within-cluster heterogeneity	0.000	0.956	0.006	0.002	0.036
Within sum of squares	0.000	2.288	13.632	5.502	84.952
Silhouette score	0.000	0.618	0.851	0.878	0.768

Note: The between sum of squares of the 4-cluster model is 5054.78. Note: The total sum of squares of the 4-cluster model is 7447.41.

.

The density-based strategy identified four large and one noise point group, based on the number of firms, captured heterogeneity, as accounted for by WSS, and silhouette statistics. The largest group (Cluster 1), comprising 1188 firms, exhibited 95.6% within-cluster heterogeneity; it had a high size-related WSS (2,288,542) and a moderate silhouette (0.618), characteristics typical of large heterogeneous groups [78]. The small lumps (Clusters 2–4), comprising 21, 15, and 21 firms, were better defined and higher on silhouettes (0.851, 0.878, 0.768) and low on WSS (13,632; 5502; 84,952). The overall model explained 67.9% of the variance (BSS = 5054.78, TSS = 7447.41), characteristic of strong unsupervised performance on non-convex shapes [78]. One noise point signals aptness for ESG data, where such points typically flag data errors or outlier performers [79,80]. The output produces one large, heterogeneous cluster and three small, tight groups, suggesting the possibility of extracting both pervasive shape and weak ESG-related information. See Figure 10.

The density-cluster model associates building-sector CO₂ emissions (BCE) and social indicators—food production (FOOD), gender parity in primary enrolment (GPIE), income share of the poorest 20% (INC20), labor force participation (LABF), and women in parliament (WPAR)—for the “S” component of ESG. Cluster 0 has a low BCE (−0.354), high food production (FOOD = 4.285); however, it exhibits low equity within genders and income levels (≈ −0.1) and low labor force participation (LABF = −1.442). High rates of women in parliament (1.410) reflect middle-income countries where political representation has caught up with educational and labor advances [81]. Cluster 1 represents a moderately negative BCE (−0.117) with balanced rates across indicators, typical of transitional economies close to global means, served by a limited mix and efficient social infrastructure [82]. Cluster 2 includes high BCE (6.826) and weak social performance—negative GPIE (−0.129), INC20 (−0.236), and WPAR (−0.107), with just slightly favorable participation in the labor force (LABF = 0.548)—typifying industrial economies that are pollution-increasing and structurally inequitable and relying on fossil fuel [51]. See Figure 11.

Cluster 3 has a low BCE (−0.337) yet extremely heterogeneous social indicators, characterized by extremely low labor participation (−4.439; −2.845) and exceptionally high female parliamentary representation (2.214). This profile characterizes scenarios in which narrow labor market inclusion coexists with political equity advances and low emissions being a by-product more of poverty limits than on-regulation across emissions, consistent with evidence that poor economies are lower emitters yet development challenges [83,84]. Cluster 4 exhibits a near-neutral BCE (0.044) but high social performance variability, characterized by high GPIE (7.518), INC20 (7.328), and WPAR (4.165), which are accompanied by low LABF (−4.759) and a negative FOOD (−3.916). They are possible high-income scenarios with equitable allocations, yet structural economic limits. Past research has confirmed non-linear and, at times, negative associations between social indicators and emissions [51]. On their own, Cluster 2 complements high emissions and inequity, Clusters 0 and 3 exhibit low emissions and mixed social performance, and Cluster 4 illustrates selected equity–emissions decoupling. The finding corroborates multisided ESG, where environmental and social elements exhibit interlocking and context-aware relationships.

5.3. Finding the Best Fit: A Comparative Evaluation of Regression Models on ESG Data

We compared a broad range of regression methods on ESG data with conventional error measures (MSE, RMSE, MAE, MAPE, and R²). The assessment covered classical techniques (linear regression, Decision Trees) and current-state methods (ensemble methods, neural networks). The findings reveal clear trade-offs: ensemble methods consistently achieved optimal predictive quality and generalization, while linear methods achieved the best interpretability and computational efficiency. The quality of neural networks was patchy, being superior on some occasions and unstable. The research presents the optimal regression methods for ESG prediction tasks, striking a balance between quality, efficiency, and interpretability. See Figure 12.

A comparison between error measures and R² values from regression models reveals sharp contrasts in the prediction and generalization capacities. Compared models are linear, regularized linear, tree-based, k-Nearest Neighbors (k-NNs), support vector machines (SVMs), neural networks, and Boosting. Multi-model benchmarking similar to that provided is common in cryptocurrency predictions [85], vehicle price prediction [86], and e-commerce satisfaction modeling [87]. The best R² was 1.0, a perfect explanation of variance. Decision Tree (0.677) and k-NN (0.655) were close, while linear and regularized linear regressions (0.014) and SVM (0.0) were poor. From error measures (MSE, RMSE, MAE), Random Forest stood out above the others, with 0s reflecting an extremely close fit that nonetheless raises some concern about overfitting in the absence of cross-validation. The best error was logged by SVM, corroborating its poor prediction performance. k-NN demonstrated consistent performance (RMSE = 0.285, MAE = 0.17), while Decision Tree was decent (RMSE = 0.782, MAE = 0.741). The best MAPE was logged by Boosting (1.0), followed by Neural Networks (0.376), k-NN (0.453), and SVM scored 0.0 again. On MAPE, linear models underperformed across the board, with regularization yielding little improvement. Overall, Random Forest demonstrated the best prediction appropriateness and reliability and thus is best suited for the current regression task, while k-NN and Decision Tree are fair options where interpretability or computational efficiency is paramount.

Social Drivers of Emissions: Interpreting BCE Through Machine Learning and ESG Indicators

Random Forest feature-importance methods were employed to select the most important building-sector CO₂ emissions (BCE) determinants at the prediction level with additive accounts. Income distribution, labor force participation, and gender equity were found to be the most significant socio-economic indicators influencing emission outcomes. The indicators were overwhelmingly the most important across measures of importance, revealing a crucial underlying driving force for patterns in BCE. Case-level decompositions also revealed that changes in social equity and participation directly reposition amounts predicted for emissions. The findings provide evidence at a microlevel that environmental performance has social components in an ESG context, and they establish a linkage between equity, governance, and decarbonization strategies. See Figure 13.

Random Forest importance indicators (MDA, TINP, MDL) are significant indicators of socio-economic push factors influencing building-sector CO₂ emissions (BCE). The income share for the bottom 20% (INC20) has the greatest predictivity in each measure (MDA = 56.464, TINP = 973.192, MDL = 83.505), suggesting income distribution—the bottom quintile in particular—as crucial for BCE and likely a consequence of unequal accessibility to energy efficiency in buildings, renewable energy, and support infrastructure. The labor force participation (LABF) variable comes in a close second. Although with a moderate level of MDA (4.113), high TINP (755.169) and MDL (81.929) support high predictivity. High participation rates are associated with urban, energy-intensive building use, while low participation rates indicate structural fragility and inefficient energy demand. See Figure 14.

Results from the Random Forest model indicate that social indicators are significant predictors of building-sector CO₂ emissions. Female parliamentary representation (WPAR) becomes significant (MDA = 1.018, Node Purity = 599.087, Dropout Loss = 64.896), consistent with its correlation with progressive building standards and energy-efficacy policy. Gender parity in education (GPIE) also emerges strongly (MDA = 1.209, Node Purity = 517.854, Dropout Loss = 61.618), indicating that knowledge and competence are key triggers for low-emissions building. The Food Production Index (FOOD) has a smaller yet complementary effect (Node Purity = 434.772, Dropout Loss = 58.385), which connects rural land use and agro-industrial energy networks. The findings align with studies that identify socio-economic and structural variables as underlying factors in emitters’ trajectories [88,89,90]. Among the myriad variables tested, the income share of the poorest quintile (INC20) emerges as the best predictor, indicating that inclusiveness and equity are at the heart of ESG-driven sustainability. See Figure 15.

Additive model decompositions reveal the respective contributions of social indicators to projected building-related CO₂ emissions (BCE) relative to a base level of 35.711 Mt CO₂. Case 1 (35.045) experiences the greatest negative effect (−12.268) for labor force participation (LABF), balanced in part by women in parliament (WPAR, +10.702) and parity in children’s educational attainment (GPIE, +6.325); low-infrastructure agrarian settings are experienced for food production (FOOD, −5.862). Case 2 drops to 15.396 on the strength of a sharp loss for the income share for the poorest 20% (INC20, −15.411), and marginal losses for WPAR and LABF. Case 3 (21.101) again experiences a sharp negative effect for LABF (−15.392), while INC20 becomes positive (+2.898), reflecting energy growth due to income equalization. Case 4 (18.029) experiences losses for both LABF (−14.523) and parity in girls’ and boys’ educational attainment (−6.516) against a small gain for INC20 (+7.341). Case 5 jumps to 45.117, as WPAR (+20.099) and LABF (+12.822) propel high-income and high-energy systems, balanced in part by negative INC20 (−15.323) and negative food production (−10.631). Altogether, LABF exerts a constant negative effect on emissions, while WPAR and GPIE act contextually, and INC20 alternates between these two effects.

6. Governance and Carbon: Unpacking the Institutional Drivers of Building-Sector Emissions

This part assesses how governance elements impact building-sector CO₂ emissions (BCE), which is highlighted in the “G” pillar of ESG. The fixed- and random-effect regressions are conducted on panel data (n = 982) to ascertain if government effectiveness (GOVT), expenditure on education (EDUE), political stability (STAB), rule of law (LAWR), R&D expenditure (RNDG), hospital expenditure (HOSP), and scientific productivity (SCIE) have significant impacts on BCE. The results reveal complex interplays. Elevated governance capacity often coexists with higher emissions. The latter is a product reflecting infrastructure-push development. Auxiliary clustering (see Section 6.2) categorizes countries on governance and emission patterns. The analysis highlights the various ways in which governance impacts environmental performance in the built environment. A metric summary of the G-Governance component variables within the ESG model used to estimate the value of BCE is shown in Appendix D. Diagnostic tests for autocorrelation and heteroscedasticity for the econometric estimates are provided in Appendix E.

6.1. Governance and the Carbon Cost of Development: A Panel Analysis of Building Emissions

This section examines how considerations regarding governance influence building-sector CO₂ emissions (BCE) in relation to the “G” pillar of ESG. On a 982-observation panel, fixed- and random-effects regressions examine government effectiveness, political stability, the rule of law, educational and research and development (R&D) expenditure, hospital infrastructure, and science output. The results present multifaceted and sometimes contradictory associations: higher governance capacity is often accompanied by higher emissions, providing evidence for development being infrastructure-driven. The study provides empirical evidence for a linkage between governance quality and environmental performance in the built environment, as well as for the interdependence between institutional strength and emission pathways across countries.

We have estimated the following equation:

$$\mathrm{B}\mathrm{P}=\mathsf{\alpha }+{\mathsf{\beta }}_1{\left(\mathrm{G}\mathrm{O}\mathrm{V}\mathrm{T}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_2{\left(\mathrm{E}\mathrm{D}\mathrm{U}\mathrm{E}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_3{\left(\mathrm{S}\mathrm{T}\mathrm{A}\mathrm{B}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_4{\left(\mathrm{R}\mathrm{N}\mathrm{D}\mathrm{G}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_5{\left(\mathrm{L}\mathrm{A}\mathrm{W}\mathrm{R}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_6{\left(\mathrm{H}\mathrm{O}\mathrm{S}\mathrm{P}\right)}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_7{\left(\mathrm{S}\mathrm{C}\mathrm{I}\mathrm{E}\right)}_{\mathrm{i}\mathrm{t}}$$

where i = 180 and t = 2000–2020 (

Table 6. Panel regression results: governance indicators and building-sector CO₂ emissions (BCE).

	Fixed Effects, Using 982 Observations Dependent Variable: BCE			Random Effects (GLS), Using 982 Observations Dependent Variable: BCE
	Coefficient	Std. Error	t-Ratio	Coefficient	Std. Error	z
const	14.7556 **	6.12013	2.411	4.37823	9.08037	0.4822
GOVT	12.7921 ***	2.42217	5.281	11.2943 ***	2.15647	5.237
EDUE	0.450092 *	0.245066	1.837	0.404457 *	0.237995	1.699
STAB	−3.19087 ***	1.16516	−2.739	−2.29649 **	0.964108	−2.382
RNDG	−4.11812 **	1.71293	−2.404	−4.25068 **	1.65044	−2.575
LAWR	−4.29609 **	2.00928	−2.138	−1.36151 *	0.755280	−1.803
HOSP	3.80015 ***	0.612354	6.206	3.75312 ***	0.589134	6.371
SCIE	0.000433491 ***	2.56973 × 10−5	16.87	0.000462643 ***	2.50122 × 10−5	18.50
Statistics	Mean dependent var		44.83643	Mean dependent var		44.83643
	Sum squared resid		88,339.25	Sum squared resid		6,648,511
	LSDV R-squared		0.992464	Log-likelihood		−5724.171
	LSDV F(108, 873)		1064.614	Schwarz criterion		11,503.46
	Log-likelihood		−3602.578	Rho		0.766045
	Schwarz criterion		7956.121	S.D. dependent var		109.3165
	Rho		0.766045	S.E. of regression		82.57715
	S.D. dependent var		109.3165	Akaike criterion		11,464.34
	S.E. of regression		10.05935	Hannan–Quinn		11,479.22
	Within R-squared		0.299880	Durbin–Watson		0.464047
	p-value (F)		0.000000
	Akaike criterion		7423.156
	Hannan–Quinn		7625.898
	Durbin–Watson		0.464047
Tests	Joint test on named regressors- Test statistic: F(7, 873) = 53.4184 with p-value = P(F(7, 873) > 53.4184) = 1.58042 × 10−63			‘Between’ variance = 5929.08 ‘Within’ variance = 89.9585 Mean theta = 0.942699 Joint test on named regressors- Asymptotic test statistic: chi-square(7) = 437.883 with p-value = 1.77433 × 10−90
	Test for differing group intercepts- Null hypothesis: The groups have a common intercept Test statistic: F(101, 873) = 78.0485 with p-value = P(F(101, 873) > 78.0485) = 0			Breusch–Pagan test- Null hypothesis: Variance of the unit-specific error = 0 Asymptotic test statistic: chi-square(1) = 2207.29 with p-value = 0
				Hausman test- Null hypothesis: GLS estimates are consistent Asymptotic test statistic: chi-square(7) = 73.1884 with p-value = 3.34305 × 10−13

Note: *** p < 0.01, ** p < 0.05, * p < 0.10.

).

The present study distinguishes the governance dimension of ESG (“G”) as a predictor for building-sector CO₂ emissions (BCE), including residential, commercial, and other building-related emissions (IPCC, 2006; AR5). From a global panel for 982 observations, fixed-effect (LSDV) and random-effect (GLS) specifications are estimated, with BCE as the dependent variable and governance quality proxied by government effectiveness (GOVT), political stability (STAB), rule of law (LAWR), expenditure on education (EDUE), R&D expenditure (RNDG), hospital infrastructure (HOSP), and scientific output (SCIE). The results show that GOVT enters significantly and positively for emissions (12.79 FE; 11.29 RE, p < 0.01), reflecting convergence between effective governments and energy-prolific economies with high incomes [91]. EDUE enters significantly and positively (p < 0.10), reflecting a growing demand for infrastructure to support human capital development phases. On the contrary, political stability reduces emissions (−3.19 FE; −2.30 RE), enabling the effective enforcement of regulations [92]. R&D expenditure (−4.12 FE; −4.25 RE) and LAWR (−4.30 FE; −1.36 RE) lower emissions, corroborating the importance of investing in science and legal capacity for building efficiency and ESG conformity. HOSP enters significantly and positively (∼3.80, highly significant), reasserting that healthcare has a high carbon intensity and that mitigation hinges on efficiency upgrades [93,94]. SCIE enters significantly and positively but non-significantly, reflecting that research systems are both high-energy intensity providers and innovation facilitators. The fixed-effect specifications dominate (R² = 0.992; AIC/BIC superior; Hausman χ² = 73.19, p < 0.01), supporting unobserved heterogeneity. Overall, governance is a key predictor of BCE, but its impact hinges on context: while better institutions, stability, and R&D lower emissions, governance intensity in high-income economies comes alongside simultaneous growth driven by infrastructure and supporting higher carbon output [95]. Aligning governance quality with focus-tuned policy and conformity mechanisms is thus crucial to align institutional capability for sustainable decarbonization.

6.2. Governance and Emissions: Clustering Insights from Neighborhood-Based Algorithms

As a supplement to regression analysis, we employed clustering to identify latent groupings among nations based on governance attributes and building-sector CO₂ emissions (BCE). Performance measures included R², AIC/BIC, silhouette statistics, and entropy, as well as robust tests for statistical fit, compaction, and interpretability. Neighborhood-based clustering was optimal, achieving perfect explanatory power (R² = 1.0) with superior cohesion and separation compared to other methods. The outcome provides a rigorous template to unveil how diverse governance structures are plotted against profiles of emissions, with a sounder analytics basis for ESG-driven diagnostics. See Figure 16.

Neighborhood-based clustering was the best performer among the criteria. Performance was strong and consistent across quality indices. The algorithm achieved an R² value of 1.0, accounting for 100% of the variance, a very stringent test for model fit. Calinski–Harabasz index was 1.0, supporting tight and well-separated clusters. The AIC and BIC were not at a minimum, reflecting higher complexity, yet better practical validity indices did exist. The Silhouette measure (0.449) registered strong cohesion and sharp separation between groups. Entropy was high (0.962), reflecting structural order. Minimum separation and Pearson’s γ were strong, yet not the best. Balanced strength across both statistical and structural indices confirms Neighborhood-based clustering as the best and most comprehensible approach. Neighborhood-based clustering is the best solution when explanatory power and cluster quality are key.

Mapping Governance-Emission Profiles: Insights from Neighborhood-Based Clustering

To further elucidate the relationship between governance traits and structure-associated CO₂ emissions (BCE), the following is an exploration of the resulting neighborhood-based clustering algorithm, deemed optimal based on comparisons with preceding models. By clustering nations based on eight governance and infrastructural indicators (including BCE, government effectiveness, education, research and development (R&D), and institutional quality), such an exploration identifies distinct profiles of countries. By identifying divergent relationships between governance and emissions within these clusters, it yields a profound observation regarding the institutional qualities that converge towards environmental ends amidst diverse development contexts. See Figure 17.

Countries cluster based on eight indicators: building-related CO₂ emissions, government effectiveness, education, political stability, R&D, rule of law, hospital density, and scientific output. Cluster 3 has high cohesion (silhouette = 0.911). Governance, stability, and law are high, while education and science are low, indicating weak knowledge-investing regimes. Cluster 6 has the highest emissions and is closely linked to scientific production and R&D. Healthcare is relatively weak. They are industrial states that trade innovation for environmental cost. Cluster 2 (204 countries) clusters around the global means. Low silhouette means an undifferentiated group. Cluster 7 has an emphasis on education, low on R&D. Emerging nations are characteristic of this profile. Cluster 10 invests heavily in education but falls short in health and R&D, thereby limiting development spillovers. Clusters 3 and 6 are distinct: knowledge-constrained states that are stable and high-carbon trade-offs that are stuck in knowledge-intensive states. Such profiles reinforce evidence that structural forces are driving emissions [96,97]. Diversified profiles inform bespoke CO₂ mitigation, according to policy scenarios supplied by [98].

6.3. Predicting Emissions with Precision: Machine Learning Models for Governance and BCE

The econometric regressions are augmented with machine-learning regressions. The regressions are K-Nearest neighbors, Decision Trees, Random Forests, and Boosting. The performance is calculated as MSE, RMSE, MAE, MAPE, and R². The results establish the optimal performing algorithms for non-linear and complex relationships. The forecast precision validates governance variables as significant predictors for building-sector CO₂ emissions (BCE). The exercise establishes a data-driven benchmark for tracking environmental outcomes that are shaped by governance. See Figure 18.

Evaluation confirms KNN as the best predictor. Errors are eradicated (MSE, RMSE, MAE = 0; MAPE = 0.054). R² is 1.0 for ideal predictions for building-sector CO₂ emissions (BCE). Decision Trees are also robust (R² = 0.957). Boosting and Random Forest are similarly close (R² = 0.843, 0.857). Linear and regularized forms perform poorly. Neural networks are unsuccessful, with an R² value of 0, indicating a failure in proper generalization. KNN’s strength lies in its non-parametric flexibility, yet it is vulnerable to overfitting without validation. Past research has confirmed both the superiority and limitations in scale-extensive tasks [85,99,100]. Despite the computational cost, KNN is the most accurate and reliable method under the research’s configuration. The hyperparameter optimization settings for the KNN machine learning regression algorithm are given in Appendix F.

What Drives Emissions? Feature Importance of Governance and Knowledge Indicators

This section explores the predictive influence of governance-related variables on building-sector CO₂ emissions (BCE) using feature importance metrics and additive model explanations. Feature importance, measured through mean dropout loss across 50 permutations, highlights the central role of scientific output (SCIE), healthcare infrastructure (HOSP), and education and R&D investment in driving model accuracy. See Figure 19.

RMSE (50 permutations) drop-out loss ranks building-sector CO₂ emissions predictors across fields of science, socio-economics, and governance. Most strongly ranked is scientific output (SCIE, 170.266), citing roles for knowledge intensity and research infrastructures, in line with [101]. Conflict between hospital and system adequacy ranks two (33.765). Education expenditure (EDUE, 31.846) and R&D expenditure (RNDG, 29.399) again feature strongly for human and technological development, in line with [100]. The governance indicators are less contributory yet significant longitudinally: government efficiency (21.112) and political stability (19.785) are favorable for institutional persistence. Rule of law (3.374) has the lowest rank, demonstrating an indirect, intangible effect. The results validate knowledge production, healthcare, and R&D as lead drivers, and governance quality as complementary infrastructure. See Figure 20.

Figure 20 reports additive contributions for five scenarios relative to a base of 50.263. SCIE, the quantity of scientific publications, is the most negative driver. Offsets range between −33.359 (Cases 1–2) and −4.794 (Case 5). There is a systematic negative correlation between building-related CO₂ emissions (BCE) and scientific output. Expanding scientific production is consistent with lower BCE. The mechanism accounts for technological innovation, efficiency in building design, and research-driven policy decisions. See Figure 21.

Government effectiveness (GOVT) and R&D expenditure (RNDG) boost emissions. Case 1 exhibits a +24 increase, and similar effects are observed in Cases 2 and 4. Education expenditure (EDUE) and political stability (STAB) exhibit erratic behavior. They are positive for Cases 1–2 and negative for Cases 3–5, which support the notion of non-linear, context-dependent dynamics. Rule of law (LAWR) makes a moderate contribution, peaking in Case 2 at +16.392. Hospital beds (HOSP) add little explanatory power. Scientific production (SCIE) uniformly reduces emissions across scenarios, and it is a successful mitigation force. Overall, GOVT and RNDG push emissions higher, while SCIE balances them, exposing the structural dualism of development: institutional strength drives growth, and science reduces emissions.

7. Harnessing ESG Dimensions for Effective Building Sector Climate Action

Comparative research finds country clusters in building-sector CO₂ and ESG. Policies are diverse in context while requiring a technological foundation in building engineering [102]. Emission profiles are derived from envelope performance, material choice, and HVAC efficiency. Strategies must be context-aware while being technologically distinct. Decarbonization in Germany, Denmark, Finland, and Nordic Europe is a sophisticated endeavor, albeit limited by the possibilities of engineering. Institutionalized actors require air-tight envelopes in the retrofit programs, triple glazing, and digitalized energy monitoring. Carbon pricing on building elements has the potential to provide lifecycle neutrality. Governance and innovation flag-bearers must lead domestic decarbonization and invest in green transitions abroad [103,104,105]. Disseminations in modular architecture, passive architecture, and retrofitting have the potential to transform global practices. Emerging markets, such as India, Indonesia, and Brazil, are experiencing rapid urbanization. Policy priorities include green transportation, affordable housing, and procuring ESG-compliant products. Prefabrication leads to waste reduction. District solar or wind cooling deters HVAC emissions. Cooperation enables leapfrogging [106,107,108]. Codes with specifications for bamboo composites or recycled steel reinforce sustainability. Capacity-constrained states such as Nigeria, Kenya, and Pakistan require institution-building and ESG-conformant programs. Ventilation and compressed earth block pilot retrofits suppress emissions while building social acceptability. School and hospital investments build legitimacy. Global fora build cooperative retrofitting with local adaptability [109]. The Orient and East Asia, comprising China and Korea, lead in governance while experiencing equity gaps. Redistributive policy is required. Subsidized low-cost building retrofits are immediate. Green roof and high-performance envelopes are imminent. AI-generational HVACs, fresh materials, and adaptive urbanism must be propagated [110,111]. Indoor air quality requires continuous monitoring. High-income emitters such as the US and Arab states require strict regulation. Near-zero codes, carbon taxes, and thermal storage constitute key policy measures. Double-skin buildings and recoverable heat systems must be the foundation for high-rise portfolios under green leasing schemes. All contexts require ESG integrity, dependent on coherence. Environmental, governance, and social pillars must be brought into alignment. Clean energy uptake requires governance capacity and social equity. Fragmented ESG does not deliver. Indicators must inform interventions rather than merely serving symbolic alignment [112,113]. Engineering indicators, including insulation transmittance, rates of recovery, and embodied carbon, must be delivered alongside ESG reports [114,115]. Clustering evidence confirms bespoke strategies: OECD leaders take leapfrogging and transfer approaches; new economies leap ahead with modularity and construction; at-risk states fortify their institutions; East Asia balances equity and governance; high-carbon OECD states set codes and establish carbon markets. All else being equal, without engineering integration, ESG remains symbolic. Through interconnection, buildings can achieve net-zero targets and make significant progress on climate mitigation at scale.

8. Limitations

Research correlates CO₂ output from buildings with environmental, social, and governance (ESG) indicators across 180 countries. There is strong evidence, and some limitations. However, the analysis depends on national aggregates spanning 22 years. Urban–rural gaps and intra-city differences remain largely invisible, despite their significant importance for emission dynamics. Most sectoral and energy policies take place subnationally. The limited availability of local data restricts the transferability to municipal and urban administrations. Approaches are based on machine learning, econometrics, and clustering. The models are computationally tractable but constrained by crude ESG proxies, specifically end-use energy and air quality [116,117,118]. National-level coverage for governance and social ESG data is sparse [102,119], which restricts its use at smaller scales. Neural-network models did not perform well. The R² value was approximately zero, and errors were large, indicating shortcomings in the dataset and hyperparameters [120]. Noise, low sample boards, and lack of parameter tuning prevented generalizability. Generalizable AI is expected to reveal non-linear and threshold dynamics in follow-up research. Another point of consideration is that the data ends in 2021. Pandemic-driven transformations—hybrid work, office underuse, and plans for recovery are lacking [121]. Given these factors, future research demands near-real-time data, enriched collections, and novel reporting standards. Only thus are evolving trajectories of emissions detectable by ESG models.

9. Conclusions

CO₂ releases by the building sector were compared with ESG indicators across 180 countries. Econometrics, cluster analysis, and machine learning detected multifaceted, multidimensional drivers. Releases are shaped by technology, governance, science, and social investment. Institutions and societies are just as important as engineering. Scientific output was associated with CO₂ releases. Higher research intensity nations are better buffered, spread knowledge, innovative, and codified sound policy. Institutional quality and R&D investments reported mixed findings. Strong institutions were ubiquitous in low-emissions economies and densely infrastructured, high-demand economies of aging economies. Governance, innovation, and country-tailored sustainability policies need to be aligned for effective buffering. Social investments build viability for sustainable practice and environmental reform capacity. Equity-led policies are ethical and more. They facilitate direct low-carbon transitions. Education, healthcare, and inclusiveness build resilience and transformation. Integration across methodologies is this study’s strength. Regression, feature importance, and cluster detection distinguish country profiles, institutional trajectories, and predictive dynamics beyond monomethod research. Decarbonizing buildings cannot rely solely on markets or technology. Success depends on governance, science, social investments, and ESG frameworks that value the interdependence between the environment, society, and institutions.