Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study attempts to classify climate vulnerability across countries by combining the ESG framework with an extreme heat indicator (HI35) using clustering methods, thereby constructing a corresponding taxonomy. Overall, the topic exhibits a certain degree of novelty, particularly in its attempt to integrate climate risk with the ESG framework. However, in its current form, the manuscript still presents several notable shortcomings in terms of overall structure, logical coherence, and academic rigor, and would benefit from systematic revision and refinement. I encourage the authors to carefully consider the following comments:

1. First, I believe there is a fundamental issue regarding the conceptual treatment of the HI35 indicator. The manuscript attempts to reinterpret HI35—from a purely climatic exposure measure—into a systemic vulnerability indicator “shaped” by ESG structures. However, this reinterpretation lacks sufficient theoretical support and involves a degree of conceptual ambiguity. By construction, HI35 (the annual number of days with apparent temperature above 35°C) is primarily determined by exogenous geographical and climatic conditions, such as latitude, solar radiation, atmospheric circulation, and regional climate systems. These factors are largely independent of a country’s ESG characteristics. Therefore, treating HI35 as being directly “determined” or “endogenously generated” by ESG factors appears conceptually tenuous. More importantly, the manuscript tends to blur the distinction between climatic exposure (exogenous) and institutional or socio-economic capacity (endogenous). While ESG factors may indeed influence how a country responds to or adapts to heat shocks, this is fundamentally different from the occurrence of extreme heat itself. Without clearly distinguishing these two layers, the interpretation of variables may become confusing, thereby weakening the overall theoretical rigor of the paper. I therefore recommend that the authors provide a clearer conceptual delineation and explicitly clarify the relationship between these two dimensions.
2. The manuscript constructs different “vulnerability types” based on clustering results; however, the economic interpretation of these classifications remains insufficient. The current discussion mainly focuses on descriptive statistics or feature summaries, lacking deeper comparative analysis. For instance, systematic differences across clusters in terms of development stage, institutional characteristics, or climatic conditions are not fully explored, which weakens the explanatory power of the results. I suggest that the authors enrich the interpretation by incorporating more in-depth analysis, such as cross-group comparisons or illustrative country cases, to enhance the economic meaning of the findings.
3. Although the literature review is extensive, it largely takes the form of parallel listing and lacks a clear problem-oriented structure. In particular, the limitations of existing studies are not systematically synthesized, which makes it difficult to highlight the marginal contribution of this paper. I recommend that the authors reorganize the literature review around the specific research question, clearly identifying gaps in the existing literature and explicitly explaining how the present study addresses them.
4. The policy implications proposed in the manuscript remain relatively general and are not closely linked to the empirical findings. In particular, the correspondence between different clusters and specific policy pathways is not sufficiently elaborated. I suggest that the authors strengthen the mapping between empirical results and policy recommendations, so that the implications become more targeted and practically relevant.
5. The study relies on country-level data to examine the relationship between HI35 and ESG structures. However, countries are highly heterogeneous units, especially those with large territories, diverse climate zones, or significant regional disparities. Aggregating such heterogeneity into a single national observation may obscure important within-country variation—for example, differences in heat exposure, infrastructure, and institutional capacity. I recommend that the authors justify the choice of the national level as the unit of analysis and discuss the potential information loss arising from such aggregation.
6. The manuscript identifies several “vulnerability types,” yet it does not clearly specify the boundary conditions of these findings—namely, under what conditions the results hold, when they may fail, and to which contexts they should not be generalized. For a study centered on classification, clarifying the scope and limitations of applicability is particularly important. I suggest that the authors explicitly address these issues in the conclusion to avoid overgeneralization.
7. ESG is originally developed as a framework to evaluate corporate sustainability performance. While extending it to the country level is not necessarily inappropriate, it requires stronger theoretical justification. At present, the manuscript combines a set of environmental, social, and governance indicators without sufficiently explaining why these variables can still be coherently interpreted as an ESG framework at the national level, rather than as a general collection of macro-level development indicators (particularly given that key factors such as economic development level are not explicitly considered). I recommend that the authors provide a more rigorous discussion on the validity and boundaries of applying ESG at the country level, thereby strengthening the conceptual foundation of the study.

Author Response

Point to Point Answers to Reviewer 1

Q1. This study attempts to classify climate vulnerability across countries by combining the ESG framework with an extreme heat indicator (HI35) using clustering methods, thereby constructing a corresponding taxonomy. Overall, the topic exhibits a certain degree of novelty, particularly in its attempt to integrate climate risk with the ESG framework. However, in its current form, the manuscript still presents several notable shortcomings in terms of overall structure, logical coherence, and academic rigor, and would benefit from systematic revision and refinement.

A1. We thank the reviewer for the overall assessment and for recognizing the novelty of integrating climate-risk analysis with the ESG framework through clustering methods. We also appreciate the reviewer’s constructive observation that the manuscript required systematic revision in terms of structure, logical coherence, and academic rigor. In response, we have substantially revised the manuscript across its main sections. First, we clarified the conceptual framework by distinguishing more explicitly between HI35 as an exogenous climatic exposure indicator and ESG variables as national conditions related to vulnerability, sensitivity, resilience, and adaptive capacity. This revision addresses potential ambiguity in the relationship between climatic exposure and socio-institutional capacity. Second, we reorganized the literature review around a problem-oriented structure, identifying the main gaps in existing studies and explaining how the present paper contributes to the literature by constructing a cross-country ESG–HI35 exposure–vulnerability taxonomy. Third, we strengthened the economic interpretation of the clustering results by moving beyond descriptive cluster summaries and linking vulnerability types to development stage, institutional capacity, demographic characteristics, climatic exposure, and representative country profiles. Fourth, we revised the policy implications section to create a clearer correspondence between empirical clusters and differentiated policy pathways, including environmental, social, governance, resource-dependent, fragile, advanced-economy, small-island, and large-system vulnerability profiles. Fifth, we added a justification for the use of the country as the unit of analysis while explicitly acknowledging the information loss arising from national aggregation. Sixth, we clarified the boundary conditions of the taxonomy, emphasizing that the results are macro-comparative, exploratory, data-dependent, and non-causal, and should not be generalized to subnational heat-risk mapping or local operational planning. Finally, we added a more rigorous theoretical discussion of the validity and limits of applying ESG at the country level, explaining why the selected indicators can be coherently interpreted as a bounded country-level ESG framework rather than as a general collection of macro-development indicators. Overall, these revisions improve the manuscript’s conceptual clarity, internal coherence, methodological transparency, and policy relevance, and we believe they directly address the reviewer’s general concerns.

 

I encourage the authors to carefully consider the following comments:

Q2. First, I believe there is a fundamental issue regarding the conceptual treatment of the HI35 indicator. The manuscript attempts to reinterpret HI35—from a purely climatic exposure measure—into a systemic vulnerability indicator “shaped” by ESG structures. However, this reinterpretation lacks sufficient theoretical support and involves a degree of conceptual ambiguity. By construction, HI35 (the annual number of days with apparent temperature above 35°C) is primarily determined by exogenous geographical and climatic conditions, such as latitude, solar radiation, atmospheric circulation, and regional climate systems. These factors are largely independent of a country’s ESG characteristics. Therefore, treating HI35 as being directly “determined” or “endogenously generated” by ESG factors appears conceptually tenuous. More importantly, the manuscript tends to blur the distinction between climatic exposure (exogenous) and institutional or socio-economic capacity (endogenous). While ESG factors may indeed influence how a country responds to or adapts to heat shocks, this is fundamentally different from the occurrence of extreme heat itself. Without clearly distinguishing these two layers, the interpretation of variables may become confusing, thereby weakening the overall theoretical rigor of the paper. I therefore recommend that the authors provide a clearer conceptual delineation and explicitly clarify the relationship between these two dimensions.

 

A2. We thank the reviewer for this important and constructive comment. We fully agree that HI35 should not be interpreted as a vulnerability indicator generated or determined by ESG structures. By construction, HI35 measures climatic exposure, namely the annual number of days with apparent temperature above 35°C, and is primarily driven by exogenous geographical and atmospheric factors such as latitude, solar radiation, humidity, atmospheric circulation, and regional climate systems. To address this concern, we revised the manuscript to clarify the distinction between two analytical layers. The first is climatic exposure, measured by HI35. The second is systemic vulnerability, captured by ESG-related variables, which describe environmental, social, and institutional conditions that may influence how countries experience, absorb, and respond to heat exposure. Accordingly, ESG variables are no longer presented as determinants of HI35, and HI35 is not treated as an endogenous outcome of ESG conditions. The clustering analysis is explicitly framed as non-causal and classificatory. Its purpose is to identify exposure–vulnerability configurations, where HI35 captures exogenous heat exposure and ESG indicators capture sensitivity, resilience, and adaptive capacity. We revised the Abstract, Introduction, Literature Review, Methodology, Results, and Conclusion accordingly. These changes preserve the climatic nature of HI35 while clarifying how ESG conditions shape vulnerability to heat exposure.

 

 

Q3. The manuscript constructs different “vulnerability types” based on clustering results; however, the economic interpretation of these classifications remains insufficient. The current discussion mainly focuses on descriptive statistics or feature summaries, lacking deeper comparative analysis. For instance, systematic differences across clusters in terms of development stage, institutional characteristics, or climatic conditions are not fully explored, which weakens the explanatory power of the results. I suggest that the authors enrich the interpretation by incorporating more in-depth analysis, such as cross-group comparisons or illustrative country cases, to enhance the economic meaning of the findings.

 

A3. We thank the reviewer for this valuable suggestion. We agree that the previous version of the manuscript relied too heavily on descriptive cluster statistics and did not sufficiently develop the economic interpretation of the vulnerability types. In response, we have revised the manuscript by adding a new subsection, In this new section, we complement the cluster membership table with a comparative interpretation based on development stage, institutional capacity, demographic structure, climatic exposure, and policy-relevant constraints. The clusters are now interpreted not only as statistical groupings, but also as economically meaningful exposure–vulnerability profiles. We also added a new table, which links selected clusters to representative countries, economic profiles, institutional/development characteristics, and policy implications. For example, Cluster 5 is interpreted as an advanced-economy profile with stronger institutional and technological capacity; Cluster 14 as a structurally fragile profile largely composed of Sub-Saharan African countries; Cluster 17, including China and the United States, as a large-system economy profile; Cluster 18 as a small-island vulnerability profile; and Clusters 2 and 20 as resource-dependent or hydrocarbon-intensive heat-exposed economies. These additions allow the taxonomy to move beyond descriptive statistics and provide a clearer economic interpretation of the results. We believe that the revised discussion strengthens the explanatory power of the clustering analysis and makes the findings more useful for policy design, since each vulnerability type is now linked to specific adaptation needs and structural constraints.

 

 

Q4. Although the literature review is extensive, it largely takes the form of parallel listing and lacks a clear problem-oriented structure. In particular, the limitations of existing studies are not systematically synthesized, which makes it difficult to highlight the marginal contribution of this paper. I recommend that the authors reorganize the literature review around the specific research question, clearly identifying gaps in the existing literature and explicitly explaining how the present study addresses them.

A4. We thank the reviewer for this constructive suggestion. We agree that the previous version of the literature review was too extensive and was organized mainly as a parallel presentation of existing studies, without sufficiently emphasizing the research problem, the limitations of previous work, and the marginal contribution of the paper. In response, we have substantially reorganized the literature review around the central research question of the study: how HI35, as an exogenous climatic exposure indicator, can be integrated with ESG variables to identify cross-country vulnerability profiles without conflating exposure with socio-institutional capacity. The revised literature review is now structured around problem-oriented research streams rather than a simple thematic listing. Specifically, we distinguish among firm-level ESG resilience to heat stress, integration of climatic exposure indicators into ESG frameworks, social vulnerability under heat exposure, governance and institutional adaptive capacity, sectoral innovation and adaptive investment, digital ESG and climate monitoring, and heat indicators for policy planning. For each stream, we now explicitly identify what the existing literature shows, what limitations remain, and how the present study addresses these gaps. To make this contribution clearer, we also revised Table 1 into a problem-oriented synthesis of the literature, research gaps, and study contribution. The revised review highlights four main gaps: the predominance of firm- and sector-level studies, the incomplete integration of heat exposure indicators into ESG frameworks, the limited operationalization of social and governance dimensions through measurable national indicators, and the lack of structured cross-country exposure–vulnerability taxonomies. We believe these revisions make the literature review more coherent, problem-oriented, and directly connected to the contribution of the paper. The manuscript now more clearly shows how our study advances the literature by combining HI35 and ESG indicators through unsupervised learning to identify heterogeneous cross-country exposure–vulnerability configurations.

 

Q5. The policy implications proposed in the manuscript remain relatively general and are not closely linked to the empirical findings. In particular, the correspondence between different clusters and specific policy pathways is not sufficiently elaborated. I suggest that the authors strengthen the mapping between empirical results and policy recommendations, so that the implications become more targeted and practically relevant.

 

A5.  We thank the reviewer for this valuable suggestion. We agree that the previous version of the manuscript presented policy implications in relatively general terms and did not sufficiently connect the empirical cluster results to specific policy pathways. In response, we have substantially revised the policy implications section to create a clearer mapping between the clustering results and targeted policy recommendations. The revised section now explicitly links different exposure–vulnerability profiles to differentiated adaptation strategies. Specifically, we now distinguish between several policy pathways. Environmental clusters characterized by high HI35, cooling degree days, natural resource depletion, or energy intensity are linked to policies such as ecological restoration, urban cooling, water-saving technologies, green infrastructure, and energy-efficient buildings. Clusters with high fossil-fuel dependence or emissions-related pressures are linked to energy-transition policies, renewable energy expansion, and low-carbon cooling systems. For the social dimension, clusters with high HI35 and indicators such as under-five mortality, fertility, undernourishment, or uneven access to safely managed drinking water are linked to public-health preparedness, early-warning systems, safe-water access, nutrition programs, maternal and child health, and social protection. For governance-related clusters, we distinguish between institutionally resilient profiles and partial-capacity profiles. The former are linked to innovation, ESG disclosure, and climate-risk dashboards, while the latter require stronger rule of law, accountability, legal rights, transparency, and implementation capacity. We also strengthened the country-level policy interpretation by linking specific cluster types to practical interventions. Advanced-economy clusters are associated with infrastructure renewal and energy-efficient adaptation; structurally fragile clusters with basic services, health systems, water access, and food security; resource-dependent and arid clusters with economic diversification and cooling-energy planning; densely populated heat-exposed clusters with urban adaptation and labor protection; small-island clusters with international climate finance and resilient infrastructure; and large systemic economies with multi-level and regionally differentiated adaptation strategies. We believe these revisions make the policy implications more targeted, empirically grounded, and practically relevant by directly connecting the clustering taxonomy to concrete adaptation pathways.

 

 

Q6. The study relies on country-level data to examine the relationship between HI35 and ESG structures. However, countries are highly heterogeneous units, especially those with large territories, diverse climate zones, or significant regional disparities. Aggregating such heterogeneity into a single national observation may obscure important within-country variation—for example, differences in heat exposure, infrastructure, and institutional capacity. I recommend that the authors justify the choice of the national level as the unit of analysis and discuss the potential information loss arising from such aggregation.

 

A6. We thank the reviewer for this important comment. We agree that countries are internally heterogeneous units and that the use of country-level data may obscure relevant within-country variation in heat exposure, infrastructure, institutional capacity, and social vulnerability. In response, we have added a new subsection to justify the choice of the national level as the unit of analysis. We clarify that the country level was adopted because the objective of the study is to construct a comparable cross-national taxonomy of heat exposure and ESG-based vulnerability conditions. This choice is also consistent with the structure of the World Bank ESG database, where HI35 and most ESG indicators are harmonized and available at the national level. We also explain that many ESG variables used in the study, such as governance quality, regulatory capacity, legal rights, political stability, research expenditure, public-health provision, social-service systems, and environmental management, are defined and implemented primarily at the national scale. For this reason, the country level represents an analytically meaningful unit for comparing systemic exposure–vulnerability configurations across a global sample. At the same time, we explicitly acknowledge the limitation raised by the reviewer. The revised manuscript now states that national aggregation may conceal important regional variation, especially in large countries with multiple climate zones, uneven infrastructure, and strong territorial disparities. We therefore clarify that the results should not be interpreted as capturing subnational vulnerability patterns, but rather as a macro-comparative taxonomy of broad national exposure–vulnerability profiles. Finally, we added this point to the limitations and future research discussion, noting that future work should extend the framework to subnational regions, metropolitan areas, or gridded datasets in order to capture local variation in heat exposure, adaptive capacity, infrastructure, and social vulnerability.

 

Q7. The manuscript identifies several “vulnerability types,” yet it does not clearly specify the boundary conditions of these findings—namely, under what conditions the results hold, when they may fail, and to which contexts they should not be generalized. For a study centered on classification, clarifying the scope and limitations of applicability is particularly important. I suggest that the authors explicitly address these issues in the conclusion to avoid overgeneralization.

 

A7. We thank the reviewer for this important suggestion. We agree that, because the study is based on classification and taxonomy-building, it is necessary to clarify the boundary conditions under which the identified vulnerability types should be interpreted. In response, we added a dedicated subsection to explicitly define the scope, applicability, and limitations of the proposed taxonomy. We clarify that the vulnerability types should be interpreted primarily as macro-comparative country-level classifications based on national HI35 and ESG indicators. Therefore, they are most appropriate for cross-country comparison, ESG benchmarking, international climate-risk screening, and broad national adaptation planning. We also specify the conditions under which the results may be less informative. In particular, the taxonomy may be less applicable in contexts where within-country heterogeneity is stronger than cross-country variation, such as large countries with multiple climate zones, strong territorial inequalities, uneven infrastructure, or decentralized institutional systems. In these cases, national-level clusters may conceal substantial regional differences in exposure, sensitivity, and adaptive capacity. Furthermore, we clarify that the results are data-dependent and exploratory. The vulnerability types may change if alternative ESG indicators, different data providers, longer time horizons, more spatially disaggregated datasets, or alternative clustering algorithms are used. We also emphasize that the clustering results should not be interpreted causally: the taxonomy identifies patterns of co-occurrence between exogenous climatic exposure and ESG-based vulnerability conditions, but does not establish that ESG variables determine HI35 or that cluster membership mechanically produces specific outcomes. Finally, we added a short reminder in the conclusion to avoid overgeneralization and to direct readers to the scope conditions. We believe these revisions make the applicability of the classification clearer and prevent the vulnerability types from being interpreted as universal, subnational, or causal risk categories.

 

 

Q8. ESG is originally developed as a framework to evaluate corporate sustainability performance. While extending it to the country level is not necessarily inappropriate, it requires stronger theoretical justification. At present, the manuscript combines a set of environmental, social, and governance indicators without sufficiently explaining why these variables can still be coherently interpreted as an ESG framework at the national level, rather than as a general collection of macro-level development indicators (particularly given that key factors such as economic development level are not explicitly considered). I recommend that the authors provide a more rigorous discussion on the validity and boundaries of applying ESG at the country level, thereby strengthening the conceptual foundation of the study.

 

A8. This is certainly an important and constructive comment, which we appreciate. We agree that while ESG was initially devised as a methodology for measuring corporate sustainability performance and associated risks, its expansion into the country-level context requires a clearer theoretical rationale. Accordingly, our manuscript has been updated to state clearly that neither country-level ESG nor related indicators should be viewed as an extension of firm-level corporate ESG performance and associated rating practices, but rather, that ESG is used as a bounded macro-sustainability framework for describing the country-level factors contributing to vulnerability and adaptive capacity in the face of exogenous climatic exposure. Concretely, our manuscript provides a description of how the three ESG pillars can be coherently reinterpreted at the national level. Environmental is used as an instrument capturing ecological pressure, resource depletion, energy dependence, emission intensity, and climate stress. Social pillar captures population sensitivity through population health and nutrition, demographic pressure, and availability of basic services. Governance is viewed as reflecting institutions and adaptability through indicators of regulatory quality, rule of law, political stability, and accountability, legal rights, innovation capability, and macroeconomic governance-related variables. We emphasize that while selected ESG indicators provide a convenient instrument for interpreting macro-exposure-vulnerability relationships, these are not meant to represent a comprehensive set of development indicators. In other words, their validity lies precisely in the bounded nature of interpretation. Accordingly, our manuscript has been updated to emphasize that our ESG-based approach is neither meant to reproduce a corporate-type rating system, nor to construct a development index, or cover all aspects of national vulnerability. HI35 serves as an exogenous indicator of climatic exposure, whereas ESG describes the national conditions that may exacerbate or attenuate consequences of such an exposure. While economic development is undoubtedly an important determinant of cross-country vulnerability analysis, our use of World Bank ESG indicator architecture implies that we view GDP growth as an indicator of macroeconomic governance conditions. However, we interpret this variable very cautiously, because it does not necessarily imply high institutional quality, as defined in the dataset. Future work should consider the inclusion of GDP per capita, income group classifications, fiscal capacity, and urbanization indicators to test the robustness of our taxonomic framework across development contexts.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

PEER REVIEW REPORT

The manuscript titled “Using Machine Learning Clustering to Build an ESG-Based Taxonomy of Heat Vulnerability” presents an unsupervised machine-learning framework to construct a cross-national taxonomy of heat vulnerability by integrating the novel indicator HI35 — the average annual days with an apparent temperature exceeding 35°C — with Environmental, Social, and Governance (ESG) variables drawn from the World Bank database. The study covers 193 countries and applies six clustering algorithms (Density-Based, Fuzzy C-Means, Hierarchical, Model-Based, K-Means, and Random Forest) across three ESG pillars (E, S, G) as well as a combined Principal Component Analysis (PCA) and hierarchical Ward linkage strategy, ultimately producing 20 national vulnerability clusters. A central conceptual contribution is the authors' explicit rejection of the standard assumption of local exogeneity for climatic variables, arguing instead that HI35 should be reinterpreted as an endogenous, diagnostically rich indicator of systemic ESG fragility.

However, there are several questions that require resolution. Below, the authors find a list of suggestions and comments per section.

1. The introduction is generally well-structured and provides a broad intellectual context for the research. The authors make a credible case for repositioning HI35 beyond its meteorological origin and for challenging local exogeneity assumptions. The literature review is extensive, touching on corporate ESG performance under heat stress, social vulnerability, governance accountability, digital ESG platforms, and sectoral dynamics. However, several concerns should be addressed:

• The dual labelling of sub-sections — for instance, 'Relaxing the common assumption of local exogeneity' appears twice (as a sub-heading in the Introduction and again as a standalone sub-section) — creates redundancy. The authors should consolidate or clearly differentiate these passages.
• Self-citation density is notably high. Works by Costantiello & Leogrande (2023a, 2023b), Drago et al. (2025a, 2025b), and Leogrande & Costantiello (2023) appear repeatedly across both the introduction and analytical sections. While not disqualifying, the editors may wish to flag potential over-reliance on authors' own prior output.
• Several references in Table 1 (Literature Synthesis) cite works in the future tense (e.g., 'Drago et al., 2025a'; 'Yu et al., 2025c') without confirmed publication details. These should be verified and updated with full bibliographic information.
• The discussion of 'Emerging Technologies and Digital ESG' (Section 2) reads as tangential to the core empirical contribution. Unless computational or digital ESG infrastructure is meaningfully connected to the clustering results, this sub-section should be condensed or removed.

2. The conceptual innovation of reframing HI35 as an endogenous, systemic vulnerability indicator is intellectually stimulating and well-articulated. However, the empirical design creates a significant tension with the conceptual claims:

• The authors state explicitly that their framework is 'deliberately descriptive and associative' and that 'no assertion of pinpointing causality' is made. This is methodologically appropriate for unsupervised clustering, but it sits in tension with the repeated assertion in the introduction and conclusions that HI35 'results from' ESG conditions or that countries' ESG setups 'contribute to its formation'. The language of causality recurs throughout the manuscript while the methods are expressly non-causal. This inconsistency must be resolved — either the causal language should be removed throughout, or a qualifying caveat should be added wherever directional relationships are implied.
• The decision to treat HI35 as the 'dependent variable' in some sub-sections (Section 4) contradicts the unsupervised clustering framework in which no variable is designated as dependent or independent. The authors should clarify whether they are running any parametric regressions or whether all analyses are purely cluster-based.
• The justification for using country-level ESG indicators to characterize vulnerability (rather than, for example, sub-national or population-weighted data) is not explicitly defended. Large federal states (China, USA, India, Brazil) are treated as single observations, which aggregates over enormous intra-national heterogeneity.

3. The data section (Section 3 and Table 2) provides a clear inventory of all variables. The sourcing from the World Bank ESG Portal is appropriate and publicly reproducible. Nonetheless, the following issues need attention:

• Temporal scope is underspecified in the main body. The authors aggregate observations by country across the entire available panel, which spans up to 64 years for some indicators. How aggregation is performed (mean, median, last available year?) is not stated in Section 3 or 4; it only appears obliquely in Section 6. This must be clarified explicitly in the methods section.
• The Under-5 Mortality Rate (MR) variable in the social cluster analysis records a negative minimum value of -55.61 per 1,000 live births, which is physically impossible. The authors acknowledge this anomaly in Appendix B but do not address it analytically. A data cleaning protocol should be described, and observations with impossible values should be excluded or imputed before analysis.
• Missing data are handled using the missForest algorithm (Section 6), which is a reasonable choice. However, the extent of missingness — notably for EIN (2,388 missing of 4,623 total observations, >50%), RDE (2,580 missing), and PUSM (1,839 missing) — is substantial enough to warrant sensitivity analysis. Results derived from imputed majority-missing variables may be unstable.
• The rationale for including Cooling Degree Days (CDD) as an 'environmental' ESG variable rather than as a climate exposure variable is not discussed. CDD is closely correlated with HI35 by construction, introducing potential collinearity in the environmental clustering step.

4. The methodological framework is ambitious and multi-layered. The use of six competing clustering algorithms with a multi-criteria selection procedure (Section 5) is a genuine strength and represents a more rigorous approach than single-algorithm studies. However, there are important weaknesses:

• Algorithm selection across the three ESG pillars is inconsistent: Model-Based clustering is selected for both the Environmental and Governance dimensions, whereas Random Forest proximity clustering is selected for the Social dimension. The decision rule for this asymmetry is not made explicit. The authors state that structural balance (cluster size distribution) was used as a tiebreaker, but they do not provide a pre-registered or reproducible decision protocol. This ex post selection risks circularity.
• The PCA step in Section 6 is incompletely described. The number of principal components retained, the proportion of variance explained, loadings for each component, and the criterion used for component retention (e.g., eigenvalue > 1, scree plot, parallel analysis) are not reported. This information is essential for reproducibility and for interpreting the final cluster assignments.
• The choice of 20 clusters in the hierarchical taxonomy (Section 6) is determined via the gap statistic, but the gap statistic results themselves are not shown. A figure displaying the gap statistic as a function of the number of clusters should be included, along with the rationale for selecting k = 20 over other local optima.
• The dendrogram from the hierarchical clustering (Ward's method) is not presented. Its inclusion would substantially aid interpretation of the cluster structure and provide evidence of separation between groups.
• The description 'pure, unintelligent machine learning methodology' in the abstract is imprecise and potentially misleading. The authors presumably mean 'unsupervised' (as opposed to 'supervised'). The term 'unintelligent' has no standard methodological meaning and should be replaced.
• For each clustering analysis, standard validation measures (silhouette plots, within-cluster sum of squares curves) should be reported visually, not only as tabular normalized scores.

5. The results are presented in considerable detail across Sections 5–7. The cluster mean tables (Tables 6, 10, 14) are informative. However, the narrative interpretation of results requires improvement:

• Cluster 17 (Section 6) groups China and the United States together, described as reflecting 'large economies with comparable exposure and adaptive capacity'. This is an unusual finding that warrants deeper scrutiny. The two countries differ dramatically across almost every ESG dimension. The authors should investigate whether this co-clustering is robust across different numbers of clusters or is an artefact of the Ward/PCA pipeline.
• The discussion of Cluster 6 — which records the highest HI35 with strong regulatory quality but weak accountability and legal rights — would benefit from naming which countries are included in this cluster. Azerbaijan alone is listed for Cluster 6 in Table 15, which raises questions about whether any substantive cross-national generalizations can be drawn from a single-country cluster.
• Several single-country clusters (Azerbaijan, Bahrain, Barbados, Belgium, Belize, Benin, Cambodia, China + USA) indicate that the 20-cluster solution may over-partition the data. The authors should discuss whether a more parsimonious solution (e.g., k = 10–12) would yield substantively similar taxonomy patterns with greater statistical stability.
• Results are interpreted largely in terms of the cluster means of standardized variables. Unstandardized or back-transformed values (e.g., actual average HI35 days, actual mortality rates) would help readers assess real-world significance.
• The taxonomy table (Table 16) is a useful synthesis, but the 'Effect on HI35' column uses signs ('+', '-', 'mixed') without confidence intervals or test statistics. In an associative, non-causal framework, this notation risks being interpreted as directional causal claims.

6. Regarding figures and tables, the authors must address the following questions:

• Figures A1–A5 and B1–B5 and C1–C5 are all placed in appendices and are described only in text within the appendix sections. Key distributional plots (e.g., Figure A1) and the scatterplot matrices should be referenced directly in the main analysis sections where they are most relevant.
• Table 15 (country-cluster assignments) contains a formatting error: Cuba appears in both Cluster 4 and Cluster 13. This duplication must be corrected.
• Several figures (A2, B2, C2, C5) are dense scatterplot matrices that are difficult to read at journal resolution. The authors should either replace these with selective pair-plot panels or increase figure resolution and label clarity.
• Appendix D (national HI35 time series) is valuable but its link to the main analytical narrative is weak. The authors mention that it 'complements the main analytical outline' but do not integrate the temporal trends into any clustering or inference.

 

The English requires substantial revision. While the core ideas are discernible, the prose is often convoluted, repetitive, and at times ungrammatical. Specific issues include:

• Extremely long and syntactically complex sentences that impede comprehension (e.g., the paragraph beginning 'Such analysis explicitly diverges...' in Section 4).
• Repetitive phrasing across sections — for example, the argument about relaxing local exogeneity is restated at least four times across Sections 1, 2, 4, and 9.
• Idiosyncratic word choices and apparent machine-translation artifacts: e.g., 'hedonic stress' (line 1003), 'extremurry' (line 136), 'favorite son metrics' (line 1082), 'decrement of GNI' used ambiguously.
• Several section titles are grammatically non-parallel or redundant (compare 'Relaxing the common assumption of local exogeneity' in Section 2 with 'Relaxig the assumption of local exogeneity' [sic] in Section 4).

The authors should engage a professional academic English editing service before resubmission.

 

The references are generally appropriate and relevant to the research topic. However, several forward-dated references (2025) lack confirmed publication details and should be verified. Moreover, four of the five listed authors contribute co-authored works that are cited ten or more times in the manuscript. While the cited works may be relevant, the density of self-citation — particularly given that several cited papers appear to be preprints or working papers rather than peer-reviewed publications — warrants editorial scrutiny.

In conclusion, with respect to the necessary adjustment to be made, as suggested above, the overall recommendation is to reconsider after a major revision.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The English requires substantial revision. While the core ideas are discernible, the prose is often convoluted, repetitive, and at times ungrammatical. Specific issues include:

• Extremely long and syntactically complex sentences that impede comprehension (e.g., the paragraph beginning 'Such analysis explicitly diverges...' in Section 4).
• Repetitive phrasing across sections — for example, the argument about relaxing local exogeneity is restated at least four times across Sections 1, 2, 4, and 9.
• Idiosyncratic word choices and apparent machine-translation artifacts: e.g., 'hedonic stress' (line 1003), 'extremurry' (line 136), 'favorite son metrics' (line 1082), 'decrement of GNI' used ambiguously.
• Several section titles are grammatically non-parallel or redundant (compare 'Relaxing the common assumption of local exogeneity' in Section 2 with 'Relaxig the assumption of local exogeneity' [sic] in Section 4).

The authors should engage a professional academic English editing service before resubmission.

Author Response

Point to Point Answers to Reviewer 2

 

Q1. The manuscript titled “Using Machine Learning Clustering to Build an ESG-Based Taxonomy of Heat Vulnerability” presents an unsupervised machine-learning framework to construct a cross-national taxonomy of heat vulnerability by integrating the novel indicator HI35 — the average annual days with an apparent temperature exceeding 35°C — with Environmental, Social, and Governance (ESG) variables drawn from the World Bank database. The study covers 193 countries and applies six clustering algorithms (Density-Based, Fuzzy C-Means, Hierarchical, Model-Based, K-Means, and Random Forest) across three ESG pillars (E, S, G) as well as a combined Principal Component Analysis (PCA) and hierarchical Ward linkage strategy, ultimately producing 20 national vulnerability clusters. A central conceptual contribution is the authors' explicit rejection of the standard assumption of local exogeneity for climatic variables, arguing instead that HI35 should be reinterpreted as an endogenous, diagnostically rich indicator of systemic ESG fragility.

A1. Thank you for the comprehensive summary and overall assessment of the manuscript. We appreciate the reviewer’s recognition of the methodological ambition of the study, including the use of multiple unsupervised clustering algorithms, the ESG–HI35 exposure–vulnerability framework, and the effort to construct a cross-national taxonomy of heat vulnerability across 193 countries. Following the reviewer’s observations, the manuscript has been substantially revised in order to improve conceptual consistency, methodological clarity, empirical transparency, and overall readability. In particular, we revised the theoretical framing of the ESG–HI35 relationship and clarified throughout the manuscript that HI35 is treated as an exogenous climatic exposure indicator rather than as an endogenous outcome generated by ESG conditions. The revised version now consistently distinguishes climatic exposure from ESG-based vulnerability and adaptive-capacity conditions, avoiding causal interpretations that are not supported by the unsupervised clustering methodology. We also clarified that the empirical strategy is entirely unsupervised and classificatory. No parametric regression models are estimated, and no variables are treated as dependent or independent in an econometric sense. The methodological framework now explicitly states that HI35 and ESG indicators jointly enter the clustering space as classification variables used to identify long-run exposure–vulnerability profiles. In addition, the revised manuscript includes:

• a clearer explanation of the temporal aggregation procedure and the construction of the country-level dataset;
• a dedicated subsection on data cleaning and plausibility checks;
• the removal of problematic or conceptually redundant variables;
• revisions to reduce repetitive phrasing and improve linguistic clarity;
• improved figure quality and appendix integration;
• corrections to tables, cluster assignments, and formatting inconsistencies;
• a substantial reduction in self-citations and verification of recent references;
• stronger differentiation between descriptive classification and causal interpretation.

Overall, the revised manuscript now presents a more coherent and methodologically consistent ESG–HI35 exposure–vulnerability taxonomy while preserving the original objective of identifying heterogeneous national profiles of climatic exposure, sensitivity, resilience, and adaptive capacity.

 

Q2. However, there are several questions that require resolution. Below, the authors find a list of suggestions and comments per section.

A2. We thank the reviewer for the careful reading of the manuscript and for the detailed and constructive comments provided throughout the review process. We carefully addressed all the observations and substantially revised the manuscript in order to improve its conceptual clarity, methodological consistency, empirical transparency, structure, and overall readability. The point-by-point responses below describe the revisions implemented in the new version of the paper.

 

Q3. The introduction is generally well-structured and provides a broad intellectual context for the research. The authors make a credible case for repositioning HI35 beyond its meteorological origin and for challenging local exogeneity assumptions. The literature review is extensive, touching on corporate ESG performance under heat stress, social vulnerability, governance accountability, digital ESG platforms, and sectoral dynamics. However, several concerns should be addressed.

 

A3. Thank you for this positive assessment of the Introduction and literature review. In response to the reviewer’s concerns, we substantially revised the Introduction and Section 2 in order to improve conceptual clarity, reduce redundancy, and strengthen the alignment between the literature review and the empirical contribution of the paper. Specifically, we consolidated overlapping discussions regarding local exogeneity and the role of HI35, removing duplicated subsection labels and relocating the full methodological clarification to a dedicated subsection in the empirical framework. We also reduced the density of self-citations, replacing several references with broader external literature and retaining only those strictly necessary for the conceptual framing. In addition, all recent references were verified and updated with complete bibliographic information. Furthermore, the discussion of emerging technologies and digital ESG was substantially condensed and integrated into the broader literature on ESG data quality and climate-risk comparability, in order to keep the review more closely connected to the core empirical objective of the paper. Finally, the literature review was reorganized around problem-oriented research streams and explicitly linked to the contribution of the study, namely the construction of cross-country ESG–HI35 exposure–vulnerability profiles through unsupervised clustering techniques.

 

 

 

Q4. The dual labelling of sub-sections — for instance, 'Relaxing the common assumption of local exogeneity' appears twice (as a sub-heading in the Introduction and again as a standalone sub-section) — creates redundancy. The authors should consolidate or clearly differentiate these passages.

A4. Thank you for this helpful observation. We agree that the previous version created redundancy by using similar subsection labels and by repeating the same conceptual distinction in different parts of the manuscript. We revised the manuscript accordingly. First, we removed the duplicated subsection heading from the Introduction and replaced the previous extended discussion with a shorter introductory paragraph. The Introduction now presents only the general motivation of the exposure–vulnerability framework, clarifying that HI35 is treated as a climatic exposure indicator while ESG variables capture national conditions related to vulnerability and adaptive capacity. Second, we consolidated the full conceptual and methodological explanation in the methodological framework. To clearly differentiate this discussion from the Introduction, we added a short introductory paragraph and created a dedicated subsection. This subsection now develops the distinction between climatic exposure and ESG-based vulnerability in the appropriate methodological context. Third, we shortened the opening of the clustering results discussion and revised the initial parts of the related empirical subsections to avoid restating the same explanation. These passages now focus more directly on clustering procedures, statistical evaluation, and the interpretation of exposure–vulnerability profiles, while referring back to the dedicated methodological clarification. Overall, these revisions remove redundant labelling, consolidate the overlapping passages, and clearly differentiate the introductory motivation from the methodological framework and empirical results.

 

 

Q5. Self-citation density is notably high. Works by Costantiello & Leogrande (2023a, 2023b), Drago et al. (2025a, 2025b), and Leogrande & Costantiello (2023) appear repeatedly across both the introduction and analytical sections. While not disqualifying, the editors may wish to flag potential over-reliance on authors' own prior output.

 

A5. Thank you for this important observation. We agree that the previous version contained an excessive density of self-citations, particularly in the Introduction and analytical sections. We revised the manuscript accordingly and removed the repeated self-citations to avoid any perception of over-reliance on the authors’ prior work. Specifically, we eliminated redundant references to Costantiello & Leogrande, Drago et al., and Leogrande & Costantiello where they were not strictly necessary for supporting the argument. We retained only essential references where directly relevant to the methodological or conceptual framing, and we replaced several self-citations with broader external literature where appropriate. Overall, this revision reduces the self-citation density, improves the balance of the reference base, and ensures that the manuscript is grounded more clearly in the wider literature rather than in the authors’ own prior output.

 

 

Q6. Several references in Table 1 (Literature Synthesis) cite works in the future tense (e.g., 'Drago et al., 2025a'; 'Yu et al., 2025c') without confirmed publication details. These should be verified and updated with full bibliographic information.

 

A6. We appreciate the value of the comment. The recent references quoted in the table of literature synthesis should indeed be verifiable and complete with full bibliographic details. As such, the table of literature synthesis and the list of references have been modified accordingly. References without verified information for publication have been either deleted from the table or replaced by fully verifiable ones. In addition, all recent references have been checked for completeness and updated if required with full bibliographic information. This includes the journal title, volume number, issue number, pagination/DOI, and DOI. Lastly, the in-text citation, table of literature synthesis, and list of references have been coordinated.

 

Q7. The discussion of 'Emerging Technologies and Digital ESG' (Section 2) reads as tangential to the core empirical contribution. Unless computational or digital ESG infrastructure is meaningfully connected to the clustering results, this sub-section should be condensed or removed.

A7. Thank you for this helpful comment. We agree that the previous discussion of emerging technologies and digital ESG was not sufficiently connected to the core empirical contribution of the paper. We revised this part of the literature review accordingly. The discussion has been substantially condensed and no longer appears as a standalone thematic stream. We retained only a brief reference to ESG data infrastructures and geospatial climate-risk assessment, limited to their relevance for data quality, comparability, and consistency in cross-country climate-risk analysis. We also revised the literature synthesis table by removing the separate stream on digital ESG and integrating the relevant point into the broader discussion of climatic exposure indicators and ESG frameworks. These revisions keep the literature review focused on the main contribution of the study: the construction of country-level ESG–HI35 exposure–vulnerability profiles through clustering.

Q8. The conceptual innovation of reframing HI35 as an endogenous, systemic vulnerability indicator is intellectually stimulating and well-articulated. However, the empirical design creates a significant tension with the conceptual claims.

 

A8. Thank you for this important observation. We agree that the previous version of the manuscript created a conceptual tension between the interpretation of HI35 and the empirical design adopted in the study. In response, we substantially revised the theoretical framing, methodological sections, and interpretative language throughout the manuscript in order to ensure full consistency between the conceptual claims and the unsupervised clustering framework. More specifically, the revised manuscript no longer presents HI35 as an endogenous indicator generated by ESG conditions. Instead, HI35 is now consistently defined as an exogenous climatic exposure indicator associated with geographical and atmospheric conditions, while ESG variables are interpreted as descriptors of environmental, social, and governance-related vulnerability, resilience, sensitivity, and adaptive capacity. We clarified this distinction explicitly in the Abstract, Introduction, Literature Review, and Methodological Framework. In addition, the manuscript now repeatedly emphasizes that the empirical strategy is entirely descriptive, associative, and classificatory rather than causal. No parametric regression models are estimated, and no directional or causal interpretation is inferred from the clustering results. To reinforce this conceptual clarification, we also introduced a dedicated methodological subsection explicitly distinguishing climatic exposure from ESG-based vulnerability conditions. Furthermore, we revised the interpretation of the clusters throughout the Results and Discussion sections to avoid any implication that ESG structures “generate,” “determine,” or “cause” HI35. Instead, the clusters are now interpreted as empirically observed configurations of climatic exposure and socio-institutional characteristics that coexist within different national exposure–vulnerability profiles. These revisions substantially improve the coherence between the conceptual framework and the empirical design of the paper.

 

Q9. The authors state explicitly that their framework is 'deliberately descriptive and associative' and that 'no assertion of pinpointing causality' is made. This is methodologically appropriate for unsupervised clustering, but it sits in tension with the repeated assertion in the introduction and conclusions that HI35 'results from' ESG conditions or that countries' ESG setups 'contribute to its formation'. The language of causality recurs throughout the manuscript while the methods are expressly non-causal. This inconsistency must be resolved — either the causal language should be removed throughout, or a qualifying caveat should be added wherever directional relationships are implied.

 

A9. Thank you for this very important point. It appears that there was indeed an inconsistency with regard to the non-causal nature of the clustering methodology used and some phrases contained in the manuscript that implied causality. Consequently, the manuscript was corrected in the following ways. First, all mentions of causality were either eliminated or modified. In particular, references to HI35 "resulting" from the presence of ESG features and ESG structures being factors contributing to HI35 formation were eliminated from the manuscript, with HI35 consistently described as an exogenous climatic exposure indicator that is related to geographical and atmospheric characteristics of the region. Second, the Abstract, Introduction, theoretical background, results, and discussion sections were adjusted with regard to their description of the ESG-HI35 link. Specifically, ESG factors were referred to as indicators representing the vulnerability, sensitivity, resilience, and adaptive capacity of various regions, but were not mentioned as determinants of the occurrence of extreme heat. Third, appropriate qualifying language was used to clarify the meaning of the cluster results obtained through the use of the ESG variables. In other words, while being described as descriptive, associative, and classificatory, the results of the clustering algorithm do not imply causation with regard to ESG conditions determining, generating, or modifying HI35.

 

 

Q10. The decision to treat HI35 as the 'dependent variable' in some sub-sections (Section 4) contradicts the unsupervised clustering framework in which no variable is designated as dependent or independent. The authors should clarify whether they are running any parametric regressions or whether all analyses are purely cluster-based.

 

A10. Thank you for this important clarification. We agree that, in an unsupervised clustering framework, no variable should be described as dependent or independent in an econometric sense. We revised the methodological sections accordingly. The manuscript now explicitly clarifies that no parametric regression models are estimated. HI35 is not treated as a dependent variable, and ESG indicators are not treated as independent explanatory variables. Instead, HI35 and ESG indicators jointly enter the clustering space as classification variables. Specifically, HI35 is defined as an exogenous climatic exposure indicator, while ESG variables are used as descriptors of environmental, social, and governance-related vulnerability and adaptive-capacity conditions. The analysis is therefore entirely unsupervised and cluster-based. We also revised the surrounding methodological language to avoid dependent–independent variable terminology and clarified that statistical indicators such as R², AIC, BIC, Silhouette, entropy, and Calinski–Harabasz are used only as clustering evaluation criteria, not as regression diagnostics. These revisions ensure consistency between the methodological framework and the interpretation of the empirical results.

 

Q11. The justification for using country-level ESG indicators to characterize vulnerability (rather than, for example, sub-national or population-weighted data) is not explicitly defended. Large federal states (China, USA, India, Brazil) are treated as single observations, which aggregates over enormous intra-national heterogeneity.

 

A11. Thank you for this insightful point. Indeed, the use of country-level ESG indicators needs to be justified in view of the fact that country aggregation may hide a lot of internal heterogeneity. This issue is addressed in the new version of the paper. Specifically, a discussion section on the unit of analysis and aggregation rationale was added, which explains why we used the country as the unit of analysis because we wanted to create a comparable cross-national taxonomy of ESG–HI35 exposure–vulnerability. The choice of the unit of analysis follows the design of the World Bank ESG database, where all ESG and HI35 variables are harmonized at the country level. At the same time, the limitations of our approach are clearly stated as well. Namely, we mention the possibility of the presence of different climate zones, levels of urbanization, distributions of infrastructure, institutional capacities, and social vulnerabilities within large and territorially diverse countries. Therefore, the interpretation of our findings should refer to a comparison between countries' exposure–vulnerability rather than to the subnational heat risk or vulnerability maps or estimates. Lastly, we added a statement about the further need to consider other units of analysis, such as subnational regions, metropoles, population exposure, and gridded climate data.

Q12. The data section (Section 3 and Table 2) provides a clear inventory of all variables. The sourcing from the World Bank ESG Portal is appropriate and publicly reproducible. Nonetheless, the following issues need attention.

 

A12. Thank you for this positive assessment of the data section and of the transparency of the World Bank ESG data sources used in the study. In response to the reviewer’s observations, we substantially revised Section 3 and the methodological framework in order to improve the clarity, reproducibility, and analytical consistency of the dataset construction and variable treatment.

More specifically, we added a dedicated subsection on the temporal scope and aggregation procedure, explicitly clarifying that the original dataset is an unbalanced annual country-level panel and that all variables were aggregated by computing country-level arithmetic means over the available years. We also introduced a new subsection on data cleaning and plausibility checks, where we explain how implausible observations, coding anomalies, and out-of-range values were identified and excluded before temporal aggregation and clustering. In addition, problematic variables such as the Under-5 Mortality Rate were removed from the Social specification after detecting physically impossible observations.

Furthermore, we clarified the rationale for adopting the country as the unit of analysis and explicitly discussed the limitations associated with national aggregation and intra-country heterogeneity. Finally, we improved the consistency between Table 2, the appendices, and the empirical sections, revised the variable framework accordingly, and clarified throughout the manuscript that the clustering analysis is based on temporally averaged country-level exposure–vulnerability profiles rather than on dynamic panel estimation or causal modeling. These revisions improve both the methodological transparency and the internal coherence of the empirical framework.

 

 

Q13. Temporal scope is underspecified in the main body. The authors aggregate observations by country across the entire available panel, which spans up to 64 years for some indicators. How aggregation is performed (mean, median, last available year?) is not stated in Section 3 or 4; it only appears obliquely in Section 6. This must be clarified explicitly in the methods section.

 

A13.  Thank you for this important clarification. We agree that the temporal scope and aggregation procedure needed to be stated explicitly in the main methodological sections. We revised the manuscript accordingly by adding a dedicated subsection  in the data and methods part of the paper. In this subsection, we now clarify that the original dataset is an unbalanced annual country-level panel, since temporal coverage differs across indicators and countries. We also specify the aggregation rule used in the analysis. For each country and each variable, annual observations were aggregated by computing the arithmetic mean over all available years. This procedure produces one temporally averaged value per country for HI35 and for each ESG indicator included in the clustering analysis. We further clarified that the clustering is applied to this temporally aggregated country-level dataset. Therefore, the empirical analysis does not model year-by-year variation, dynamic temporal relationships, or panel effects. The resulting clusters should be interpreted as long-run country-level exposure–vulnerability profiles, not as temporal trajectories or year-specific classifications. These revisions make the temporal scope, aggregation method, and interpretation of the clustering dataset explicit in the main body of the manuscript.

 

Q14. The Under-5 Mortality Rate (MR) variable in the social cluster analysis records a negative minimum value of -55.61 per 1,000 live births, which is physically impossible. The authors acknowledge this anomaly in Appendix B but do not address it analytically. A data cleaning protocol should be described, and observations with impossible values should be excluded or imputed before analysis.

 

A14.  Thank you for this important observation. We agree that the negative value recorded for the Under-5 Mortality Rate (MR) was physically impossible and required an explicit analytical correction rather than only an acknowledgment in the appendix. In response, we revised the data-cleaning procedure and added a dedicated subsection in the Methods section, entitled “Data Cleaning and Plausibility Checks.” This subsection now explains that all variables were screened before temporal aggregation and clustering in order to identify implausible values, coding anomalies, and scale inconsistencies. Values outside admissible conceptual or physical bounds were treated as data errors, set to missing before aggregation, and excluded from the computation of country-level averages. The final clustering dataset therefore includes only valid and interpretable observations. More specifically, we decided to exclude the Under-5 Mortality Rate (MR) from the Social cluster analysis because the presence of physically impossible negative observations raised concerns about the reliability of the variable for the purposes of clustering. The Social dimension was therefore re-estimated using only HI35, Fertility Rate (FR), access to safely managed drinking-water services (PUSM), and prevalence of undernourishment (POU). This change is reflected in the revised variable framework, where the Social pillar now contains only FR, PUSM, and POU. All related analyses, tables, figures, cluster profiles, and variable-importance results were updated accordingly. The revised Social clustering is now based on Random Forest clustering and identifies social exposure–vulnerability profiles using the corrected set of variables. The new results show that PUSM, FR, and POU are the main social indicators differentiating the clusters, while HI35 remains an exogenous climatic exposure indicator. This revision removes the problematic MR variable from the empirical analysis and ensures that the Social taxonomy is based only on conceptually valid and empirically reliable indicators. We believe that these revisions fully address the reviewer’s concern by both clarifying the data-cleaning protocol and correcting the Social cluster analysis through the exclusion of the problematic MR variable.

 

Q15. Missing data are handled using the missForest algorithm (Section 6), which is a reasonable choice. However, the extent of missingness — notably for EIN (2,388 missing of 4,623 total observations, >50%), RDE (2,580 missing), and PUSM (1,839 missing) — is substantial enough to warrant sensitivity analysis. Results derived from imputed majority-missing variables may be unstable.

 

A15. We thank the reviewer for this important and constructive observation. We agree that the extent of missingness in some variables, particularly EIN, RDE, and PUSM, requires additional caution and a robustness-oriented discussion. In response to this comment, we have revised the manuscript by adding a dedicated sensitivity analysis for the Social component, where PUSM plays a central role in the construction of the social exposure–vulnerability taxonomy. Specifically, we re-estimated the Social clustering on a reduced sample, excluding countries for which the relevant variables were completely missing over the observation period. The retained dataset therefore includes only countries with sufficient observed information for HI35, fertility rate, access to safely managed drinking-water services, and prevalence of undernourishment. This reduced-sample sensitivity analysis was introduced to test whether the main Social taxonomy remains statistically meaningful and substantively interpretable under stricter data-availability conditions. The results confirm that the main interpretive conclusion remains stable: HI35 and social vulnerability do not fully overlap, and the Social clusters continue to identify differentiated exposure–vulnerability profiles. We also report cluster numerosity and the Herfindahl–Hirschman Index (HHI), so that the robustness check evaluates not only standard clustering performance but also the balance and interpretability of the resulting cluster structure. We also acknowledge the reviewer’s concern regarding EIN and RDE, which belong respectively to the Environmental and Governance dimensions. Given the already extensive length and complexity of the empirical section, we did not add separate full reduced-sample clustering exercises for each ESG pillar. Instead, we revised the manuscript to make the limitation explicit and to qualify the interpretation of results involving variables with substantial missingness. In particular, we now clarify that Environmental and Governance results involving EIN and RDE should be interpreted as broad taxonomic classifications rather than as precise country-level assignments. We also identify data availability for these indicators as a limitation of the present study and as a priority for future robustness checks using alternative data sources or more complete ESG series. Overall, the revision addresses the reviewer’s concern by adding a concrete reduced-sample robustness check for the Social component and by explicitly qualifying the interpretation of results based on highly incomplete Environmental and Governance indicators.

 

Q16. The rationale for including Cooling Degree Days (CDD) as an 'environmental' ESG variable rather than as a climate exposure variable is not discussed. CDD is closely correlated with HI35 by construction, introducing potential collinearity in the environmental clustering step.

 

A16. We thank the reviewer for this important observation. We agree that the inclusion of Cooling Degree Days (CDD) alongside HI35 could introduce conceptual overlap and potential collinearity, since both variables are closely related to heat exposure by construction. In response to this comment, we removed CDD from the empirical framework and from all related sections of the manuscript, including the environmental clustering specification, tables, figures, and variable descriptions. The revised version retains HI35 as the sole climatic exposure indicator, while the Environmental pillar now includes only environmental-pressure, energy, and emissions-related variables. This revision improves the conceptual clarity and internal consistency of the ESG–HI35 exposure–vulnerability framework.

 

Q17. The methodological framework is ambitious and multi-layered. The use of six competing clustering algorithms with a multi-criteria selection procedure (Section 5) is a genuine strength and represents a more rigorous approach than single-algorithm studies. However, there are important weaknesses

A17. We thank the reviewer for recognizing the strength of the multi-algorithm clustering framework. We agree that the methodological design is ambitious and may appear complex. In response, we revised Section 5 to clarify that the use of six clustering algorithms is not intended to increase methodological complexity, but to reduce algorithm-specific bias and assess the robustness of the ESG–HI35 taxonomy across different classification logics. We also clarified that the final model is not selected mechanically on the basis of a single metric, but through a combined assessment of statistical performance, cluster balance, substantive interpretability, and consistency with the exposure–vulnerability framework. Finally, we now explicitly frame the clustering procedure as exploratory and classificatory rather than causal.

Q18. Algorithm selection across the three ESG pillars is inconsistent: Model-Based clustering is selected for both the Environmental and Governance dimensions, whereas Random Forest proximity clustering is selected for the Social dimension. The decision rule for this asymmetry is not made explicit. The authors state that structural balance (cluster size distribution) was used as a tiebreaker, but they do not provide a pre-registered or reproducible decision protocol. This ex post selection risks circularity.

 

A18.  We thank the reviewer for this important observation. We agree that the decision rule behind the selection of different algorithms across ESG pillars required clearer explanation. In the revised manuscript, we clarify that the same model-selection protocol was applied consistently to the Environmental, Social, and Governance dimensions. However, the preferred algorithm may differ across pillars because each clustering exercise is based on a different set of ESG indicators, while HI35 remains the common climatic exposure variable. Therefore, each ESG pillar generates a distinct clustering problem with different variance, density, separability, and cluster-balance characteristics. We also added an explicit reproducible decision sequence: exclusion of degenerate cluster solutions, comparison of normalized internal validity and fit metrics, evaluation of cluster-size balance using the HH index, and final selection based on substantive interpretability. This clarification reduces the risk of ex post circularity and makes the algorithm-selection procedure more transparent and replicable.

 

Q19. The PCA step in Section 6 is incompletely described. The number of principal components retained, the proportion of variance explained, loadings for each component, and the criterion used for component retention (e.g., eigenvalue > 1, scree plot, parallel analysis) are not reported. This information is essential for reproducibility and for interpreting the final cluster assignments.

A19.  Thank you very much for your significant insight. In the new manuscript, the discussion about the Principal Component Analysis (PCA) has been modified such that it includes the number of components chosen, percentage of variation explained, Scree Plot, and the PCA loadings for the selected components. We chose to retain only the first six principal components since they together explain about 74.3% of the total variance, which provides an optimal trade-off between data retention and dimensionality reduction. These six components’ values were then used as the input space to carry out the Hierarchical Clustering.

 

 

Q20. The choice of 20 clusters in the hierarchical taxonomy (Section 6) is determined via the gap statistic, but the gap statistic results themselves are not shown. A figure displaying the gap statistic as a function of the number of clusters should be included, along with the rationale for selecting k = 20 over other local optima.

 

A20.  We thank the reviewer for this important observation. In the revised manuscript, we have added a dedicated figure reporting the gap statistic as a function of the number of clusters for the hierarchical taxonomy. We also expanded Section 9.3 to provide a more detailed methodological discussion of the cluster-selection procedure. Specifically, the revised text now explains that the gap statistic increases progressively as the number of clusters rises, indicating improved separation between country groups. However, the slope of the curve becomes substantially flatter after approximately k = 20, suggesting diminishing marginal improvements from additional partitioning complexity. We further clarify that the selection of k = 20 was not based on the mechanical identification of the global maximum of the gap statistic, but on a combined assessment of statistical evidence, cluster stability, and substantive interpretability. In addition, we now explicitly discuss why alternative local optima associated with larger numbers of clusters were not retained. Although higher values of k generate slightly larger gap-statistic values, these solutions produce increasingly fragmented and less stable taxonomies characterized by very small country groups with limited comparative interpretability. Conversely, lower values of k produce excessively aggregated classifications that mask relevant heterogeneity across countries. The revised manuscript therefore clarifies that the 20-cluster specification was selected because it provides the most balanced and robust compromise between explanatory separation, classification stability, and interpretability of the ESG–HI35 exposure–vulnerability profiles.

 

Q21. The dendrogram from the hierarchical clustering (Ward's method) is not presented. Its inclusion would substantially aid interpretation of the cluster structure and provide evidence of separation between groups.

 

A21. Thank you for this valuable suggestion. We agree that the inclusion of the dendrogram associated with the hierarchical clustering procedure improves the transparency and interpretability of the clustering structure. In response, we revised the manuscript by adding the dendrogram generated from the hierarchical clustering analysis based on Ward’s linkage criterion. The figure is now included in the methodological and empirical results sections, together with an explanatory discussion of the hierarchical aggregation process and cluster separation patterns.

The dendrogram provides a visual representation of the nested structure of the country-level clusters and illustrates how observations are progressively merged according to similarity in the ESG–HI35 exposure–vulnerability space. This addition allows readers to better evaluate the degree of separation between clusters, the relative proximity of country groupings, and the rationale underlying the final cluster partition adopted in the analysis. We also added a brief methodological clarification explaining that Ward’s method was selected because it minimizes within-cluster variance and produces comparatively compact and interpretable hierarchical structures.

Furthermore, the discussion of the clustering results was revised to explicitly connect the dendrogram evidence with the final taxonomy interpretation and the PCA-based hierarchical clustering strategy described in Section 6. These revisions improve the interpretability, transparency, and robustness of the hierarchical clustering analysis presented in the manuscript.

 

Q22. The description 'pure, unintelligent machine learning methodology' in the abstract is imprecise and potentially misleading. The authors presumably mean 'unsupervised' (as opposed to 'supervised'). The term 'unintelligent' has no standard methodological meaning and should be replaced.

 

A22.  We thank the reviewer for their helpful recommendation. We have incorporated the hierarchical clustering dendrogram based on the Ward’s method in our revised manuscript. This dendrogram has been included in Section 9.3 along with the gap statistic. It is clear from the revision that the dendrogram represents the hierarchy of country classification, shows the chosen cluster solution of 20 clusters, and demonstrates group separation by showing different branch heights. It is also noted below the figure that this dendrogram has been created using Euclidean distance and Ward’s criterion for minimum variance on PCA-transformed data.

 

Q23. For each clustering analysis, standard validation measures (silhouette plots, within-cluster sum of squares curves) should be reported visually, not only as tabular normalized scores.

 

A23.  We would like to acknowledge the reviewer for this useful recommendation. In the updated manuscript, we extend our use of normalized measures of validity of the cluster analysis results in tabular form by providing visualization tools for evaluating the validity of clustering results. For example, we include silhouette plots for the chosen clustering solutions under the Environmental, Social, and Governance criteria. Also, the within-cluster sum of squares (WSS) and elbow-type validation curves have been used previously in the model selection panels for comparing various clustering solutions. These visualization tools have been included in the updated manuscript along with the analysis of the silhouettes for a thorough evaluation of the clustering results. In summary, the above changes enhance the transparency of the validation process and allow the readers to visually assess the balance between statistical fit, cluster separation, and interpretation.

 

Q24. The results are presented in considerable detail across Sections 5–7. The cluster mean tables (Tables 6, 10, 14) are informative. However, the narrative interpretation of results requires improvement: Cluster 17 (Section 6) groups China and the United States together, described as reflecting 'large economies with comparable exposure and adaptive capacity'. This is an unusual finding that warrants deeper scrutiny. The two countries differ dramatically across almost every ESG dimension. The authors should investigate whether this co-clustering is robust across different numbers of clusters or is an artefact of the Ward/PCA pipeline.

 

A24.  Thank you for this important observation. We agree that the co-clustering of China and the United States in Cluster 17 required further scrutiny, particularly because the two countries differ substantially across several ESG dimensions. To address this point, we conducted an additional robustness check by varying the number of clusters in the PCA–Ward hierarchical clustering procedure. Specifically, we examined whether China and the United States remained assigned to the same cluster across alternative solutions with k = 10, 12, 15, 18, 20, 22, 25, and 30 clusters. The robustness check shows that China and the United States are assigned to the same cluster for broader and intermediate classifications, namely k = 10, 12, 15, 18, and 20. However, once the clustering solution becomes more granular, they are no longer assigned to the same cluster. In particular, they separate at k = 22, k = 25, and k = 30. We have therefore revised the interpretation of Cluster 17. Rather than describing China and the United States as countries with comparable ESG profiles or comparable adaptive capacity, we now interpret their co-clustering at k = 20 as reflecting broad macro-structural similarities, such as very large economic scale, strong technological capacity, global integration, and substantial internal territorial heterogeneity. At the same time, we explicitly acknowledge that this co-clustering is not stable under more granular clustering specifications and should therefore not be interpreted as evidence of full ESG similarity. Accordingly, the revised manuscript now treats Cluster 17 as a specific result of the PCA–Ward representation at the selected 20-cluster level, rather than as a robust indication that China and the United States share similar ESG–HI35 vulnerability profiles. We also added the robustness table and graphical evidence to clarify this point and to avoid overstating the substantive similarity between the two countries.

 

Q25. The discussion of Cluster 6 — which records the highest HI35 with strong regulatory quality but weak accountability and legal rights — would benefit from naming which countries are included in this cluster. Azerbaijan alone is listed for Cluster 6 in Table 15, which raises questions about whether any substantive cross-national generalizations can be drawn from a single-country cluster.

 

A25. Thank you for this important observation. We agree that the interpretation of the previous Cluster 6 required clarification, particularly because the earlier version of the taxonomy included Azerbaijan as a single-country cluster, limiting the possibility of drawing broader cross-national conclusions. Following the revision of the clustering procedure, the composition and numbering of the clusters have changed. In the updated Table 15, Cluster 6 no longer contains Azerbaijan alone. Instead, Cluster 6 now includes Austria, Canada, Denmark, Estonia, Finland, France, Germany, Iceland, Israel, Japan, Sweden, Switzerland, and the United Kingdom. We therefore revised the interpretation of this cluster accordingly. The updated discussion now interprets Cluster 6 as a group of highly developed advanced economies characterized by strong institutional capacity, innovation systems, developed infrastructure, advanced public services, and greater adaptive resources within the ESG–HI35 framework. Azerbaijan is now assigned to Cluster 7 as a single-country cluster. Following the reviewer’s suggestion, we explicitly clarify in the revised manuscript that single-country clusters should not be interpreted as general cross-national regimes. Instead, they are treated as country-specific outlier profiles generated by the PCA–Ward clustering procedure. Accordingly, no broad institutional, economic, or policy generalization is drawn from these isolated clusters. We have revised both the narrative discussion and the interpretation of the cluster taxonomy to make this distinction clearer and to avoid overstating the substantive meaning of single-country classifications.

 

Q26. Several single-country clusters (Azerbaijan, Bahrain, Barbados, Belgium, Belize, Benin, Cambodia, China + USA) indicate that the 20-cluster solution may over-partition the data. The authors should discuss whether a more parsimonious solution (e.g., k = 10–12) would yield substantively similar taxonomy patterns with greater statistical stability.

A26. Thank you for this important observation. We agree that the presence of several single-country clusters in the original k = 20 taxonomy raised the possibility of over-partitioning and warranted additional robustness analysis. To address this point, we conducted an additional parsimony and sensitivity check by comparing the main k = 20 PCA–Ward taxonomy with more aggregated clustering solutions at k = 10 and k = 12. In particular, we examined whether the isolated country profiles identified in the k = 20 specification remained isolated under lower-k classifications. The results show that the more parsimonious k = 10 and k = 12 solutions generate only one single-country cluster, represented by Belgium. By contrast, the higher-resolution k = 20 solution produces seven isolated country profiles: Azerbaijan, Bahrain, Barbados, Belgium, Belize, Benin, and Cambodia. This evidence indicates that the k = 20 taxonomy captures finer country-level heterogeneity but also increases fragmentation. Importantly, Belgium remains isolated across all specifications, suggesting that it represents a stable outlier profile within the ESG–HI35 structure. Conversely, the additional single-country clusters emerging only at k = 20 appear to reflect the greater granularity of the selected classification rather than fully stable structural separation. Following the reviewer’s suggestion, the revised manuscript now explicitly discusses the trade-off between granularity and parsimony in the clustering procedure. We therefore interpret the k = 20 solution as a higher-resolution exploratory taxonomy, while the k = 10 and k = 12 specifications are presented as more parsimonious robustness benchmarks that preserve the main macro-structural country regimes with greater statistical stability.

 

Q27. Results are interpreted largely in terms of the cluster means of standardized variables. Unstandardized or back-transformed values (e.g., actual average HI35 days, actual mortality rates) would help readers assess real-world significance.

 

A27. Thank you for this helpful comment. We clarified in the revised manuscript that the clustering procedure was performed using the original, non-standardized input variables. Consequently, the reported cluster means are not standardized scores or back-transformed estimates, but actual cluster averages expressed in the original measurement scales of the variables. To improve interpretability and emphasize real-world significance, we revised the discussion of the cluster profiles by explicitly interpreting HI35 as the average annual number of days with apparent temperature above 35°C, while the ESG indicators are interpreted according to their respective original units (e.g., percentages, index values, energy-intensity measures, and emissions indicators). We also clarified that standardization/normalization was applied only to the comparison of model-selection statistics across clustering algorithms, and not to the input variables or to the reported cluster means.

 

Q28. The taxonomy table (Table 16) is a useful synthesis, but the 'Effect on HI35' column uses signs ('+', '-', 'mixed') without confidence intervals or test statistics. In an associative, non-causal framework, this notation risks being interpreted as directional causal claims.

 

A28. Thank you for this important observation. We agree that the original version of Table 16 could unintentionally suggest directional or causal interpretations through the use of symbols such as “+”, “−”, and “mixed”, despite the explicitly associative and non-causal nature of the clustering framework. To address this issue, we substantially revised Table 16 and the related discussion in the manuscript. First, we removed the “Effect on HI35” column and replaced the symbolic notation with descriptive textual interpretations based on observed co-occurrence patterns within the clustering structure. Second, we revised the table title, notes, and accompanying discussion to clarify that the taxonomy is purely descriptive and classificatory, rather than causal or inferential. More specifically, the revised version now emphasizes that:

• HI35 is treated as an exogenous climatic exposure indicator rather than a dependent variable;
• ESG variables are not interpreted as determinants or causes of HI35;
• the clustering procedure identifies exposure–vulnerability profiles based on the joint distribution of HI35 and ESG indicators;
• the reported taxonomy summarizes observed co-occurrences between heat exposure and ESG conditions, without estimating regression coefficients, directional impacts, confidence intervals, or hypothesis tests.

We also revised the narrative interpretation throughout the manuscript in order to replace potentially causal language (e.g., “effects”, “determinants”, “increases/decreases HI35”) with associative terminology such as “coexistence”, “configuration”, “profile”, and “co-occurrence”. These revisions improve the conceptual consistency of the paper and better align the interpretation of the taxonomy with the unsupervised and non-parametric nature of the methodological framework.

Q29. Regarding figures and tables, the authors must address the following questions. Figures A1–A5 and B1–B5 and C1–C5 are all placed in appendices and are described only in text within the appendix sections. Key distributional plots (e.g., Figure A1) and the scatterplot matrices should be referenced directly in the main analysis sections where they are most relevant.

 

A29. Thank you for this insightful remark. We have included explicit citations to the appendix material to improve the connection between our analysis and the appendix figures in the new draft of our paper. In particular, we cite to Figures A1–A5, B1-B5, and C1-C5 in the main clustering analysis section, where it is specified that such figures contain complementary distribution plots, density plots, and scatterplots of the Environmental, Social, and Governance clustering results. Namely, the revised text says that:

• Figures A1-A5 illustrate diagnostic plots of the Environmental data;
• Figures B1-B5 contain diagnostic plots of the Social data; and
• Figures C1-C5 are diagnostic plots of the Governance data.

As such, we incorporated all these citations to enhance the connection between our analysis and the appendix material to help readers better comprehend the distributional structure of the ESG-HI35 variables.

 

 

Q30. Table 15 (country-cluster assignments) contains a formatting error: Cuba appears in both Cluster 4 and Cluster 13. This duplication must be corrected.

 

A30. Thank you for pointing out this formatting error. We revised Table 15 and corrected the duplicate country entry. In the updated version, Cuba appears only once, in Cluster 18. We also checked the full country-cluster assignment table to ensure that all countries are listed only once and that the cluster counts are internally consistent.

 

Q31. Several figures (A2, B2, C2, C5) are dense scatterplot matrices that are difficult to read at journal resolution. The authors should either replace these with selective pair-plot panels or increase figure resolution and label clarity.

 

A31. It is much appreciated that you have provided such a useful comment. In the updated version of the manuscript, all the appendices' figures were recreated in a high-definition form for better presentation and visualization. To be more specific, all Figures A1-A5, B1-B5, and C1-C5 were recreated in much higher resolution with much better graphics, more readable labels, and overall better clarity in terms of axes' text presentation. A special emphasis should be put on scatterplot matrices depicted in Figures A2, B2, C2, and C5, which are better visible now. This significantly improves visual representation and quality of the figures both in their digital format and when printed in journal quality. Moreover, in this updated version of the paper, we provide references to all these appendices' figures in our discussions related to methodology and cluster analysis results.

 

Q32. Appendix D (national HI35 time series) is valuable but its link to the main analytical narrative is weak. The authors mention that it 'complements the main analytical outline' but do not integrate the temporal trends into any clustering or inference.

A32. Thank you for this important observation. We agree that the role of Appendix D in the original manuscript was not sufficiently connected to the main analytical narrative. To address this issue, we revised both the main text and Appendix D in order to clarify the analytical role of the national HI35 time series. Specifically, the manuscript now explicitly states that the time series reported in Appendix D are provided as complementary descriptive evidence documenting the temporal evolution and persistence of heat exposure across countries between 2000 and 2022. We also clarified that these temporal series are not used as direct inputs in the clustering algorithms, which are instead based on temporally averaged country-level indicators. Accordingly, the study does not perform temporal clustering, dynamic inference, or causal trend estimation using the Appendix D data. Rather, the purpose of the appendix is to provide contextual support for interpreting HI35 as a long-run climatic exposure indicator and to illustrate the historical variability of heat exposure across national contexts. These revisions strengthen the conceptual connection between Appendix D and the main analytical framework while preserving the descriptive and non-causal nature of the study design.

 

Q33. The English requires substantial revision. While the core ideas are discernible, the prose is often convoluted, repetitive, and at times ungrammatical. Specific issues include. Extremely long and syntactically complex sentences that impede comprehension (e.g., the paragraph beginning 'Such analysis explicitly diverges...' in Section 4).

 

A33. Thank you very much for pointing out this important observation. As suggested by the reviewer, the paper has been thoroughly edited and rewritten to improve its readability and simplify the syntactic construction of many sentences, especially those in Section 4. Long sentence fragments were broken up into more concise sentences, and long paragraphs were split up into shorter, thematic clusters. For example, the discussion on the difference between climatic exposure and ESG-vulnerability has been edited to clarify its meaning and provide greater continuity. In the new version, the empirical framework has been explained using simpler sentences and shorter paragraphs. Moreover, additional edits were made in order to make clear that the clustering framework used here is both non-causal and classificatory, minimizing redundancy in the text. The editing process focused on the section that explains the methodology, particularly on the paragraphs starting with “The empirical framework is non-causal and classificatory.” Furthermore, the editing includes the paragraphs discussing the ESG decomposition approach in Section 4.1.

 

Q34. Repetitive phrasing across sections — for example, the argument about relaxing local exogeneity is restated at least four times across Sections 1, 2, 4, and 9.

A34. We appreciate the insightful comment made by the reviewer. In our new version of the manuscript, we have meticulously ensured the avoidance of repetition regarding the conceptual difference between climatic exposure and vulnerability, as well as HI35 as the exogenous exposure variable. In particular, Section 1, Section 2, Section 4, and Section 9 have been condensed, where repetition has been curtailed, and the concepts have been expressed in a more concise form. In the new manuscript, we have made it a point to introduce the concept early on in the Introduction chapter and make subsequent mention only where it is methodologically relevant, especially in the Empirical Framework chapter. Moreover, we have consciously curtailed repetition of the phrase “relaxation of local exogeneity,” and refrained from stating explicitly in various sections that ESG factors do not causally influence HI35 unless it is essential for methodology. These changes have contributed toward improving the coherence of our manuscript without compromising its conceptual clarity.

Q35. Idiosyncratic word choices and apparent machine-translation artifacts: e.g., 'hedonic stress' (line 1003), 'extremurry' (line 136), 'favorite son metrics' (line 1082), 'decrement of GNI' used ambiguously.

A35. We thank the reviewer for this observation. We carefully checked the revised manuscript for the expressions indicated in the comment, including “hedonic stress”, “extremurry”, “favorite son metrics”, and ambiguous uses of “decrement of GNI”. These expressions are no longer present in the revised version. In addition, we carried out a broader language revision of the manuscript to remove possible machine-translation artifacts, improve terminology, and ensure that technical expressions are consistent with standard climate-risk and clustering terminology. In particular, terms related to heat exposure, heat stress, clustering metrics, and GNI-based indicators were reviewed for clarity and consistency.

 

Q36. Several section titles are grammatically non-parallel or redundant (compare 'Relaxing the common assumption of local exogeneity' in Section 2 with 'Relaxig the assumption of local exogeneity' [sic] in Section 4).

A36. The authors appreciate the reviewer’s comments. A review of the manuscript was done with respect to the section titles mentioned by the reviewer in the comments, that is, “Relaxing the common assumption of local exogeneity” and “Relaxig the assumption of local exogeneity.” These section titles are no longer used in the revised manuscript. Moreover, a detailed review of all the section titles used in the manuscript was done in order to improve grammatical accuracy, reduce redundancy, and maintain parallelism.

Q37. The references are generally appropriate and relevant to the research topic. However, several forward-dated references (2025) lack confirmed publication details and should be verified. Moreover, four of the five listed authors contribute co-authored works that are cited ten or more times in the manuscript. While the cited works may be relevant, the density of self-citation — particularly given that several cited papers appear to be preprints or working papers rather than peer-reviewed publications — warrants editorial scrutiny.

A37. We thank the reviewer for this important observation. The reference list was carefully revised and updated in order to improve bibliographic accuracy, consistency, and transparency. In particular, all forward-dated references (2025) were re-checked and corrected where necessary, with publication details verified and standardized throughout the manuscript. In addition, the manuscript was substantially revised to eliminate excessive self-citation. References associated with the authors’ previous co-authored works were critically reassessed, and redundant or non-essential self-citations were removed. As a result, the revised version no longer contains concentrated patterns of self-citation, and the literature review now relies more extensively on external peer-reviewed contributions from the broader ESG, climate-risk, and vulnerability literature. These revisions were implemented to improve the neutrality, balance, and scholarly robustness of the bibliographic framework.

Q38. In conclusion, with respect to the necessary adjustment to be made, as suggested above, the overall recommendation is to reconsider after a major revision.

A38. The authors would like to thank the reviewer for a detailed consideration of the manuscript and useful comments rendered throughout the peer-review procedure. The manuscript was significantly modified following the reviewers’ recommendations. Indeed, the current version of the manuscript is improved in many ways. Specifically, the manuscript is enhanced in terms of theoretical considerations, methodological explanation, description of clustering analysis outcomes, formatting tables, preparing figures, and other aspects. The specific modifications include: revision of theoretical justification for developing a taxonomy; elimination of unnecessary structures; highlighting the fact that the clustering analysis is not a causal analysis; more coherent linking between the appendices and the main body of the manuscript; citation of references; etc. The authors appreciate the contribution of the reviewer to improving the scientific merit of the manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This revision has addressed the concerns and suggestions I raised previously quite well.

Author Response

Point to Point Answers to Reviewer 1

 

Q1. This revision has addressed the concerns and suggestions I raised previously quite well.

A1. Thanks dear reviewer.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The revised manuscript entitled “Using Machine Learning Clustering to Build an ESG-Based Taxonomy of Heat Vulnerability” has been carefully re-evaluated in light of the comments provided in the first round of review. The authors have undertaken a substantial revision, and the new version shows clear and significant improvements in terms of conceptual clarity, methodological transparency, and overall coherence. Specifically, below I provide the assessment of how the main issues raised in the initial report have been addressed.

One of the most critical concerns in the first review was the inconsistency between the non-causal nature of the methodology and the repeated use of causal language linking ESG factors to HI35. In the revised manuscript, this issue has been convincingly resolved. HI35 is now consistently defined as an exogenous climatic exposure indicator, rather than as a measure of vulnerability or an endogenous outcome. The authors explicitly clarify that the analysis is descriptive and classificatory, and that no causal interpretation is intended.

Moreover, the earlier contradiction regarding the role of HI35 as a “dependent variable” has been entirely removed. The revised version clearly states that no variable is treated as dependent or independent, and that the empirical strategy is fully based on unsupervised clustering rather than regression modeling.

This represents a major improvement, effectively aligning the conceptual claims with the methodological approach.

About Temporal Scope, Aggregation, and Unit of Analysis, the original manuscript lacked transparency regarding temporal aggregation. The revised version now includes a dedicated section explaining that the dataset covers the period 1999–2023 and that country-level values are obtained by averaging annual observations to construct long-run profiles. This clarification is precise and well justified, particularly given the objective of identifying structural vulnerability regimes rather than short-term dynamics. In addition, the authors now explicitly justify the use of countries as the unit of analysis and acknowledge the limitations associated with intra-national heterogeneity (e.g., in large federal states).

 

Overall, the treatment of temporal and spatial aggregation is now clear, transparent, and appropriate.

The reviewer’s concerns regarding data anomalies—especially the presence of physically impossible values—have been appropriately addressed. The revised manuscript introduces a structured data cleaning and plausibility-check procedure, whereby out-of-range or inconsistent observations are removed prior to aggregation.  With respect to missing data, the authors provide a more explicit description of how missingness is handled and supplement the baseline results with sensitivity analyses that exclude countries with structural data gaps.

 

While the treatment of missing data is now clearly described and improved compared to the original version, the discussion could still be strengthened by a more formal assessment of the potential impact of imputation or missingness on cluster stability.

Continuing, the revised manuscript significantly improves the transparency of the methodological framework. The clustering procedure is now clearly described as a multi-step process involving several algorithms, evaluated through a combination of statistical performance metrics and structural criteria (e.g., cluster balance and interpretability). The authors also clarify that although a consistent selection protocol is applied across ESG dimensions, the final choice of algorithm may differ depending on the statistical structure of each dataset. Furthermore, the treatment of dimensionality reduction has been improved, with detailed reporting of principal component analysis (PCA), including eigenvalues and explained variance.

Passing to Results Interpretation and Cluster Structure, the presentation of results has been enriched through additional tables, figures, and diagnostic plots, which improve readability and transparency. The interpretation of clusters is generally clearer, and the distinction between climatic exposure (HI35) and ESG-based vulnerability is now consistently maintained. However, some issues raised in the initial report remain only partially addressed. In particular:

• The choice of a relatively large number of clusters (e.g., 20 clusters in the integrated taxonomy) is not thoroughly justified in terms of parsimony versus interpretability.
• The presence of small or single-country clusters is not fully discussed in terms of their substantive meaning and statistical robustness.
• Some unusual clustering outcomes (e.g., the grouping of structurally different countries) would benefit from deeper robustness analysis or sensitivity testing across alternative cluster solutions.

These aspects do not invalidate the findings but suggest room for further refinement.

Finally, the revised version includes a much more extensive and appropriate use of figures and tables, including silhouette plots and diagnostic visualizations that support the clustering results. This represents a clear improvement relative to the original submission. On the contrary, the issues related to self-citation density and forward-dated references appear only partially addressed. Although the literature review has been reorganized and clarified, further editorial attention may be warranted to ensure balance and accuracy in the reference list.

 

On “Language and Style”, the overall quality of English has been significantly improved. The text is now more consistent, technically precise, and easier to follow. While some verbosity and stylistic redundancy remain, the manuscript is now at an acceptable level for publication. If before submission, authors engage a professional academic English editing service, the manuscript will definitely acquire a high level.

Based on the above comments, I therefore recommend that the manuscript be accepted after minor revisions.

The remaining issues can be addressed through targeted revisions without requiring fundamental restructuring of the analysis.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The overall quality of English has been significantly improved. The text is now more consistent, technically precise, and easier to follow. While some verbosity and stylistic redundancy remain, the manuscript is now at an acceptable level for publication. If before submission, authors engage a professional academic English editing service, the manuscript will definitely acquire a high level.

Author Response

Point to point Answers to reviewer 2

Q1. The revised manuscript entitled “Using Machine Learning Clustering to Build an ESG-Based Taxonomy of Heat Vulnerability” has been carefully re-evaluated in light of the comments provided in the first round of review. The authors have undertaken a substantial revision, and the new version shows clear and significant improvements in terms of conceptual clarity, methodological transparency, and overall coherence. Specifically, below I provide the assessment of how the main issues raised in the initial report have been addressed.

A1. The authors appreciate the efforts taken by the reviewer in reconsidering the revised version of the paper with regard to improved conceptual focus, methodological clarity, and overall coherence. The authors are happy that their efforts in revising the paper have resolved the major criticisms posed initially. In this revision, the authors make an effort in resolving the remaining minor critiques made by the reviewer, specifically robustness to missing values, clustering analysis, small clusters, and the balance of references.

Q2. One of the most critical concerns in the first review was the inconsistency between the non-causal nature of the methodology and the repeated use of causal language linking ESG factors to HI35. In the revised manuscript, this issue has been convincingly resolved. HI35 is now consistently defined as an exogenous climatic exposure indicator, rather than as a measure of vulnerability or an endogenous outcome. The authors explicitly clarify that the analysis is descriptive and classificatory, and that no causal interpretation is intended. Overall, the revised manuscript now consistently distinguishes climatic exposure from ESG-based vulnerability conditions and avoids causal interpretations throughout the text.

A2. We sincerely thank the reviewer for recognizing that the revised manuscript has resolved the inconsistency between the non-causal nature of the methodology and the previous use of causal terminology linking ESG indicators and HI35. Following the reviewer’s suggestion, we further strengthened the conceptual and methodological clarification throughout the manuscript. In particular, HI35 is now consistently defined as an exogenous climatic exposure indicator rather than as a proxy for vulnerability or an endogenous outcome. We explicitly clarify that the study adopts a descriptive and classificatory unsupervised-learning framework and does not intend to infer causal relationships between ESG indicators and heat exposure. To reinforce this point, we revised several sections of the manuscript:

• In the Abstract, we explicitly state that no causal relationship is inferred between ESG indicators and HI35 and that the analysis is descriptive and cluster-based.
• In the Introduction, we replaced causal formulations such as “vulnerability determined by ESG factors” with non-causal expressions referring to “ESG-based vulnerability-related profiles.”
• In the Literature Review, we clarified that ESG variables are associated with vulnerability-related conditions rather than interpreted as causes of HI35.
• In the theoretical discussion of the ESG framework, we revised the language to describe ESG indicators as national characteristics associated with sensitivity, resilience, and adaptive capacity under climatic exposure.
• In Section 4 (Empirical Framework), we further clarified that no dependent–independent variable structure is imposed, that HI35 is treated as an exogenous exposure indicator, and that the resulting clusters should be interpreted as similarity patterns rather than causal mechanisms.
• In Section 4.1, we removed residual causal wording such as “impacts” and “determinants” and replaced it with terminology emphasizing associated conditions and vulnerability profiles.

Q3. Moreover, the earlier contradiction regarding the role of HI35 as a “dependent variable” has been entirely removed. The revised version clearly states that no variable is treated as dependent or independent, and that the empirical strategy is fully based on unsupervised clustering rather than regression modeling. This represents a major improvement, effectively aligning the conceptual claims with the methodological approach.

A3. We sincerely thank the reviewer for recognizing that the revised manuscript has resolved the previous inconsistency concerning the role of HI35 within the empirical framework. Following the reviewer’s observation, we further strengthened the methodological clarification throughout the manuscript. In particular, we explicitly clarify that the study does not estimate regression models and does not impose any dependent–independent variable structure. The empirical strategy is entirely based on unsupervised clustering techniques aimed at identifying exposure–vulnerability similarity patterns across countries. To reinforce this point, we revised Section 4 (“Empirical Framework and Methodological Rationale”) by explicitly stating that: (i) no parametric regression models are estimated; (ii) no variable is treated as dependent or independent in an econometric sense; (iii) HI35 is not used as a dependent, target, or outcome variable; (iv) HI35 is included only as an exogenous climatic exposure dimension within the clustering space together with ESG indicators; and (v) the analysis is fully descriptive, classificatory, and based on unsupervised learning. We also revised the surrounding methodological discussion to ensure consistency between the conceptual framework and the empirical implementation. The manuscript now consistently presents the clustering results as exposure–vulnerability classifications rather than causal or econometric relationships.

Q4. About Temporal Scope, Aggregation, and Unit of Analysis, the original manuscript lacked transparency regarding temporal aggregation. The revised version now includes a dedicated section explaining that the dataset covers the period 1999–2023 and that country-level values are obtained by averaging annual observations to construct long-run profiles. This clarification is precise and well justified, particularly given the objective of identifying structural vulnerability regimes rather than short-term dynamics. In addition, the authors now explicitly justify the use of countries as the unit of analysis and acknowledge the limitations associated with intra-national heterogeneity (e.g., in large federal states). Overall, the treatment of temporal and spatial aggregation is now clear, transparent, and appropriate.

A4. We sincerely thank the reviewer for recognizing the substantial improvement in the treatment of temporal scope, aggregation procedures, and unit of analysis in the revised manuscript. Following the reviewer’s previous concerns, we introduced dedicated methodological sections explicitly clarifying the temporal coverage, aggregation strategy, and spatial unit adopted in the study. In particular, Section 3.1 (“Temporal Scope and Aggregation Procedure”) now explains that the dataset covers the period 1999–2023 and that annual observations are aggregated into country-level means in order to construct long-run exposure–vulnerability profiles rather than analyze short-term fluctuations or year-by-year dynamics. We further clarified that the clustering framework is based on temporally aggregated country-level observations and therefore does not model dynamic panel relationships or temporal causality. In addition, Section 3.3 (“Unit of Analysis and Aggregation Rationale”) now explicitly justifies the use of countries as the unit of analysis, emphasizing that the World Bank ESG and HI35 databases are constructed at the national level and that the research objective is to develop a cross-country comparative taxonomy of exposure–vulnerability profiles. We also added a detailed discussion of the limitations associated with intra-national heterogeneity, particularly in large and geographically diverse countries such as China, the United States, India, and Brazil, where climatic exposure and institutional conditions may vary substantially across regions. To further reinforce transparency, the revised manuscript clarifies that the proposed taxonomy should not be interpreted as a subnational heat-risk map, but rather as a macro-level classification framework for comparative climate-risk assessment and ESG-based vulnerability profiling. Overall, these revisions were intended to improve methodological transparency and ensure consistency between the study’s conceptual objectives and its empirical implementation.

 

Q5. The reviewer’s concerns regarding data anomalies—especially the presence of physically impossible values—have been appropriately addressed. The revised manuscript introduces a structured data cleaning and plausibility-check procedure, whereby out-of-range or inconsistent observations are removed prior to aggregation.  With respect to missing data, the authors provide a more explicit description of how missingness is handled and supplement the baseline results with sensitivity analyses that exclude countries with structural data gaps.

A5. We sincerely thank the reviewer for recognizing the improvements made in the treatment of data anomalies, plausibility checks, and missing-data handling in the revised manuscript. Following the reviewer’s previous comments, we introduced a dedicated subsection entitled “Data Cleaning and Plausibility Checks” (Section 3.2), where we now describe in detail the data-validation procedure applied before temporal aggregation and clustering. In particular, we clarified that all environmental, social, governance, and HI35 indicators were screened separately in order to identify physically impossible observations, coding errors, out-of-range records, and inconsistencies in indicator scales. The revised manuscript now explicitly states that:

• admissible ranges were defined according to the substantive meaning and measurement scale of each variable;
• percentage indicators were checked for physical feasibility;
• counts, rates, and index variables were examined for negative or implausible values;
• physically impossible values, inconsistent observations, and missing records were excluded prior to aggregation; and
• country-level means were computed only from valid observations.

We also clarified that missing observations were not imputed before clustering. Instead, the aggregation procedure relied exclusively on available valid observations in order to reduce the risk that data anomalies or artificial imputations could distort the construction of long-run exposure–vulnerability profiles. In addition, we strengthened the robustness analysis presented in Section 6.1 by introducing a restricted-sample sensitivity analysis that excludes countries with structural data gaps in the ESG dimensions. The revised text now explicitly explains that this procedure was designed to assess whether structural missingness affects the stability, interpretability, and reliability of the baseline clustering taxonomy. The results confirm that the main exposure–vulnerability structures remain stable under the restricted sample. Overall, these revisions improve methodological transparency and reinforce the robustness of the ESG–HI35 clustering framework against data anomalies and missing-data issues.

Q6. While the treatment of missing data is now clearly described and improved compared to the original version, the discussion could still be strengthened by a more formal assessment of the potential impact of imputation or missingness on cluster stability.

A6. We sincerely thank the reviewer for this important observation regarding the potential influence of missingness and imputation choices on cluster stability. Following the reviewer’s suggestion, we substantially strengthened the discussion of missing-data handling and robustness assessment throughout the manuscript. In particular, we introduced an additional subsection entitled “Robustness to Missingness and Imputation Choices,” where we explicitly discuss the methodological implications of missing observations for the clustering framework. The revised manuscript now clarifies that the pillar-specific clustering analyses were conducted without imputing missing observations prior to clustering. Instead, country-level means were computed exclusively from valid observations after the exclusion of physically impossible, inconsistent, or out-of-range records. This strategy was adopted to avoid introducing artificially generated values that could distort distances between countries and alter cluster assignments. To assess the potential impact of structural missingness on cluster stability, we complemented the baseline analysis with restricted-sample sensitivity analyses excluding countries with complete absence of data for at least one ESG variable. The comparison between the baseline and restricted-sample results shows that the principal exposure–vulnerability configurations remain broadly stable and interpretable across specifications, suggesting that the proposed ESG–HI35 taxonomy is not primarily driven by missing-data structure. In addition, we clarified an important distinction concerning the integrated PCA-based taxonomy presented in Section 9. Unlike the pillar-specific clustering exercises, the PCA and hierarchical-clustering framework required a complete country-by-variable matrix. Therefore, a limited variable-specific mean imputation was applied only after temporal aggregation and only within the integrated PCA exercise in order to avoid excessive sample loss and preserve the full set of countries. This clarification has now been explicitly added to the manuscript to avoid ambiguity regarding the role of imputation in different parts of the empirical framework. Overall, these revisions strengthen the transparency and robustness of the proposed ESG–HI35 clustering methodology and provide a more formal assessment of the relationship between missing-data structure and cluster stability.

Q7. Continuing, the revised manuscript significantly improves the transparency of the methodological framework. The clustering procedure is now clearly described as a multi-step process involving several algorithms, evaluated through a combination of statistical performance metrics and structural criteria (e.g., cluster balance and interpretability). The authors also clarify that although a consistent selection protocol is applied across ESG dimensions, the final choice of algorithm may differ depending on the statistical structure of each dataset. Furthermore, the treatment of dimensionality reduction has been improved, with detailed reporting of principal component analysis (PCA), including eigenvalues and explained variance.

A7.  We sincerely thank the reviewer for recognizing the substantial improvements in the transparency and coherence of the methodological framework. Following the reviewer’s previous observations, we revised the methodological sections to provide a more explicit and structured description of the clustering strategy adopted in the manuscript. In particular, the revised version now clearly presents the empirical framework as a multi-step clustering procedure involving multiple unsupervised algorithms, including Density-Based, Fuzzy C-Means, Hierarchical, Model-Based, K-Means, Random Forest Proximity, and Neighborhood-Based clustering methods. The manuscript now explains that model selection was not based on a single statistical indicator, but on a multidimensional evaluation framework combining:

• clustering validity measures (e.g., Silhouette, Dunn index, Pearson’s γ);
• explanatory and separation metrics (e.g., R² and Calinski–Harabasz index);
• parsimony criteria (AIC and BIC);
• entropy measures; and
• structural considerations such as cluster balance, interpretability, and discriminatory capacity.

We also clarified that the same methodological selection logic was consistently applied across the Environmental, Social, and Governance dimensions, while allowing the preferred clustering algorithm to differ according to the statistical properties and internal structure of each ESG dataset. The revised text now explicitly explains that highly concentrated or weakly interpretable solutions were not preferred, even when they achieved strong values on isolated statistical metrics. In addition, we strengthened the discussion of dimensionality reduction and integrated taxonomy construction in Section 9. The revised manuscript now reports the principal component analysis (PCA) procedure in greater detail, including eigenvalues, explained variance, and the rationale for dimensionality reduction prior to hierarchical clustering. This clarification was intended to improve the transparency and reproducibility of the integrated ESG–HI35 taxonomy. Overall, these revisions substantially improve the methodological clarity, interpretability, and transparency of the clustering framework and ensure stronger alignment between the empirical implementation and the conceptual objectives of the study.

Q8. Passing to Results Interpretation and Cluster Structure, the presentation of results has been enriched through additional tables, figures, and diagnostic plots, which improve readability and transparency. The interpretation of clusters is generally clearer, and the distinction between climatic exposure (HI35) and ESG-based vulnerability is now consistently maintained. However, some issues raised in the initial report remain only partially addressed. In particular:

• The choice of a relatively large number of clusters (e.g., 20 clusters in the integrated taxonomy) is not thoroughly justified in terms of parsimony versus interpretability.
• The presence of small or single-country clusters is not fully discussed in terms of their substantive meaning and statistical robustness.
• Some unusual clustering outcomes (e.g., the grouping of structurally different countries) would benefit from deeper robustness analysis or sensitivity testing across alternative cluster solutions.

These aspects do not invalidate the findings but suggest room for further refinement.

A8. We sincerely thank the reviewer for this constructive comment. Following the reviewer’s suggestion, we further refined the interpretation of the cluster structure and strengthened the discussion of parsimony, small clusters, and robustness. First, we revised Section 9.3 to clarify that the 20-cluster solution should not be interpreted as the only possible country partition, but as a high-resolution exploratory taxonomy. We now explicitly explain that k = 20 is useful for capturing fine-grained heterogeneity across ESG–HI35 profiles, while more parsimonious solutions, especially k = 10 and k = 12, are used as robustness benchmarks. Second, we revised Section 9.4 to discuss small and single-country clusters more cautiously. The manuscript now states that such clusters may reflect distinctive ESG–HI35 configurations, but in some cases may also result from the greater granularity of the k = 20 taxonomy rather than fully stable country types. Third, we added robustness checks across alternative cluster solutions to assess unusual clustering outcomes. In particular, we examined the China–United States co-clustering across different values of k and showed that their joint classification is stable from k = 10 to k = 20 but disappears at higher resolutions. This clarifies that some apparently unusual groupings are meaningful at broader levels of aggregation but become separated in more granular taxonomies. Overall, these revisions improve the interpretation of the integrated taxonomy, clarify the trade-off between parsimony and interpretability, and provide a more cautious and robust reading of small or unusual clusters.

 

Q9. Finally, the revised version includes a much more extensive and appropriate use of figures and tables, including silhouette plots and diagnostic visualizations that support the clustering results. This represents a clear improvement relative to the original submission. On the contrary, the issues related to self-citation density and forward-dated references appear only partially addressed. Although the literature review has been reorganized and clarified, further editorial attention may be warranted to ensure balance and accuracy in the reference list.

A9. We sincerely thank the reviewer for recognizing the substantial improvements made in the presentation and visualization of the empirical results. Following the reviewer’s previous suggestions, the revised manuscript now includes a significantly expanded set of tables, figures, silhouette plots, diagnostic visualizations, and robustness analyses intended to improve the transparency, interpretability, and reproducibility of the clustering framework. Regarding the reviewer’s concerns about self-citation density and forward-dated references, we carefully revised the literature review and reference list in order to improve balance and editorial consistency. In particular, we reduced self-citations where possible and strengthened the integration of external literature across the different thematic dimensions of the study, including climatic exposure, ESG vulnerability, clustering methodologies, governance, and adaptation research. We also conducted a detailed verification of all forward-dated references included in the manuscript. The references cited as 2025–2026 correspond to articles that are already published, officially available online, or released as online-first publications by their respective journals. Therefore, these references were retained only after confirming their publication status and bibliographic validity. Overall, the revised manuscript now presents a more balanced and transparent literature framework while substantially improving the graphical and diagnostic presentation of the empirical analysis.

Q10. On “Language and Style”, the overall quality of English has been significantly improved. The text is now more consistent, technically precise, and easier to follow. While some verbosity and stylistic redundancy remain, the manuscript is now at an acceptable level for publication. If before submission, authors engage a professional academic English editing service, the manuscript will definitely acquire a high level.

A10. Our profound appreciation goes to the Reviewer for the positive appraisal of the modified manuscript, as well as acknowledgment of the remarkable improvement in the quality of the English language and writing style. According to the reviewer’s suggestion, we have undertaken one more step towards refining our paper by eliminating any excessive verbosity and stylistic repetition in our writing. In other words, several redundant sentences have been shortened, transitional devices have been used more effectively, and more precise terminology has been introduced. Moreover, the revisions of the Introduction, Literature Review, and Discussion sections have been carried out with special attention paid to making the manuscript more coherent, readable, and concise. As a result, we believe that our manuscript is even better than before. To conclude, before the manuscript is submitted for the final approval, we intend to submit it for one more academic editing session in English.

Q11. Based on the above comments, I therefore recommend that the manuscript be accepted after minor revisions.

A11. We sincerely thank the reviewer for the positive recommendation and for the constructive comments provided throughout the review process. We have carefully addressed all minor revisions requested and have revised the manuscript accordingly.

Q12. The remaining issues can be addressed through targeted revisions without requiring fundamental restructuring of the analysis.

A12. We would like to extend our thanks to the reviewer for their positive comments. All concerns were satisfactorily addressed via focused editing without changing the underlying structure of our analysis. This includes addressing our comments on parsimony, small clusters, missingness, literature cited, and writing style, respectively.

Author Response File: Author Response.pdf