Review Reports - Digital Spatial Governance and Built-Up Land Carbon Intensity: Evidence from China’s Smart City Pilot Program

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript would benefit from a clearly defined Methods section. At present, the methodology is spread across the Research Design section and several later subsections, making it difficult to understand the full analytical approach and sequence of methods used. For example, why section 4.1 include the terms policy treatment, next to sample and data??
The overall structure and flow of the manuscript need considerable improvement. The paper moves between theory, research design, econometric methods, spatial analysis, robustness tests, heterogeneity analysis, and machine learning without a sufficiently clear progression. Reorganizing these sections would make the study much easier to follow.
The study uses a large number of analytical methods, including DID, spatial models, PSM, entropy balancing, heterogeneity analysis, RF-RFECV, and SHAP. The authors should provide a clearer explanation of why each method is included and how these different analyses complement one another.
Several variables and methodological terms are not defined with enough detail. For reproducibility, the authors should clearly report how each variable was calculated, including its formula, unit, transformation, data source, and interpretation.
Table 3 reports Variance Inflation Factors (VIFs), but it is not clear how these values were calculated. The authors should explain the VIF calculation, identify which variables and regression specification were used, and clarify whether the reported values correspond to the baseline model.
The interpretation of the VIF results also needs more discussion. The VIF values for lnGDP and green patent authorizations are 5.84 and 5.81, respectively. Simply stating that these values are below 10 may not be sufficient, as values above 5 can still indicate meaningful multicollinearity. The authors should examine whether this affects the stability of the regression estimates.
There appears to be an important inconsistency in the use of the variable “Smart.” In the policy-treatment description, Smart appears to be a binary variable taking a value of 0 or 1, whereas the descriptive statistics report Smart as a digital-intelligence integration index ranging from 1.32 to 9.41. This needs to be carefully checked and corrected, as Smart is central to the analysis.
The construction of the digital-intelligence integration index needs to be explained much more clearly. The manuscript refers to the use of “entropy weighting and/or PCA.” These are different approaches, so the authors should specify exactly which method was used and provide enough information for the index to be reproduced.
The calculation of the main dependent variable, built-up land carbon intensity (CLCI), should be presented more transparently. The manuscript should provide the complete calculation, including the energy categories considered, emission factors, built-up land definition, units, data sources, and any transformations applied.
A particularly serious issue is the statement that it remains unclear whether CLCI was logarithmically transformed during the original data-processing procedure. The transformation of the main dependent variable cannot remain uncertain, as this directly affects the interpretation of the regression coefficients. The authors need to verify and clearly report this before the results can be properly assessed.
The in-text citation formatting needs careful revision throughout the manuscript. References are not consistently positioned relative to punctuation: some citations appear before the period, others after it, and in several places there is no space between the citation and the surrounding text. A consistent journal style should be applied throughout.
The reference list also needs to be carefully reformatted. DOI presentation is inconsistent, with some references providing complete DOI links and others giving only DOI numbers. Journal titles and other bibliographic elements should also be checked for consistency.
There are many spacing and formatting problems throughout the manuscript. In several places, words, sentences, and citations run together. A thorough formatting and proofreading pass is needed before publication.
I recommend adding a dedicated Data and Variable Construction subsection. This would allow the authors to explain sample selection, data sources, missing-data handling, transformations, policy coding, and the construction of derived indicators in one clear location.
The terminology used for the main variables should be standardized. Terms such as “Smart,” “Smart City implementation,” “Smart City Pilot,” “digital-intelligence integration index,” and “index” are used throughout the paper and can easily be confused. Each concept should have a unique and consistently used name.
Some of the definitions in Table 1 remain too vague to reproduce the analysis. For example, the interpretation of INS is said to depend on its “coding direction,” but the actual coding direction is not clearly specified. The authors should provide the exact definition and interpretation.
The causal econometric analysis and the predictive machine-learning analysis should be separated more clearly. Although the authors state that SHAP results are predictive rather than causal, the organization of the manuscript could still lead readers to mix the two types of evidence. A clearer structural separation would help avoid this confusion.
The machine-learning methodology needs enough detail to allow replication. The authors should report the train/test split, random seed, cross-validation procedure, hyperparameter tuning, selected parameters, RF-RFECV feature-selection procedure, and final selected features.
The statement that training and testing samples “should maintain city-level independence” is too vague. The authors should confirm whether city-level independence was actually enforced and explain exactly how cities were separated during training, validation, and testing.
The spatial analysis also needs stronger methodological explanation. In particular, the authors should explain why inverse distance was selected as the main spatial-weighting approach and clearly describe how the spatial matrices were constructed and normalized.
Methods and results are not always clearly separated. In several places, methodological choices are introduced alongside their findings and interpretation. The model specifications, assumptions, diagnostic tests, and robustness procedures should ideally be established before presenting the corresponding results.
The paper contains many hypotheses, robustness tests, spatial analyses, mechanisms, heterogeneity analyses, and machine-learning results, which makes the main message difficult to identify. The authors should streamline the presentation and clearly distinguish the primary analyses from supporting or supplementary analyses.
Definitions should be provided when terms first appear. Several abbreviations, indices, constructed variables, and methodological concepts are introduced before they are fully explained, which makes the manuscript unnecessarily difficult to follow.
The equations need clearer explanation. Every symbol used in the models should be defined immediately after the relevant equation, including subscripts, fixed effects, spatial terms, treatment variables, control vectors, and error terms.
The manuscript needs a thorough language and technical edit. There are several grammatical issues, inconsistent terminology, capitalization problems, missing spaces, and awkward transitions. These issues sometimes make technically important sections harder to understand.
The Results section could be reorganized to make the main findings much clearer. I suggest presenting the baseline estimates first, followed by identification and robustness checks, spatial analysis, mechanisms and heterogeneity, and finally the machine-learning analysis.
A simple methodological flowchart would greatly improve the manuscript. Showing the progression from data and variable construction to treatment definition, baseline DID, robustness tests, spatial analysis, channel/heterogeneity analysis, and machine learning would help readers understand how the different parts of the study fit together.

Author Response

Response to Reviewer 1

Dear Reviewer 1,

We sincerely thank you for taking the time to review our manuscript so carefully and for providing detailed and constructive comments. Your suggestions helped us recognize more clearly where the manuscript still needed improvement, particularly in its overall organization, variable definitions, research design, and discussion of reproducibility. We carefully reconsidered each point you raised and revised the analysis, presentation of results, and structure of the manuscript accordingly. Where the available records do not allow a complete reconstruction, we have stated these limitations openly in the revised manuscript and narrowed the interpretation of the findings to avoid claims that go beyond what the data and documentation can support. Below, we respond to each of your comments in turn. For ease of checking, each response identifies the relevant section, table, or figure in the revised manuscript, together with a searchable text location and the corresponding revision color. We are very grateful for your thoughtful guidance. Your comments have been highly valuable in improving the clarity, rigor, and overall quality of the manuscript.

The combined revised manuscript uses three colors to distinguish revisions made in response to the three reviewers: green for Reviewer 1, blue for Reviewer 2, and red for Reviewer 3. When a revision addresses comments raised by more than one reviewer, one primary color is used in the manuscript and is identified in the corresponding location note. Because pagination may change during revision, locations are identified mainly by section, table, or figure number together with searchable text. The color used in each response letter distinguishes the reviewer, while the color stated in the location note indicates the corresponding revision in the manuscript. We again sincerely appreciate your careful review and helpful guidance, and we hope that the revised manuscript now responds more fully to the concerns you raised.

Comment 1

The manuscript would benefit from a clearly defined Methods section. At present, the methodology is spread across the Research Design section and several later subsections, making it difficult to understand the full analytical approach and sequence of methods used. For example, why section 4.1 include the terms policy treatment, next to sample and data??

Response: Thank you for identifying this organizational problem. We agree that the previous structure mixed sample construction, treatment definition, and estimation choices in a way that obscured the analytical sequence. We have therefore rebuilt Section 4 as a self-contained Methods section. Section 4.1 now focuses on the sample and data sources; Section 4.2 brings together treatment coding and variable construction; Section 4.3 reports descriptive statistics and diagnostics; and Section 4.4 establishes the empirical strategy before any corresponding results are presented. This change also separates methodological choices from their empirical findings more clearly.

Location in revised manuscript: 4.1 Sample and Data Sources; paragraph beginning “The balanced panel covers 40 prefecture-level cities from 2011 to…” [red]; Figure 1 shows the location of the 40-city research area and places the regional sample within China. The map is descriptive and also provides the spatial frame used to construct the geographic weight matrices; the preferred policy analysis subsequently excludes the three sub-prefectural pilot cases described in Section 4.1. SCP equals one from the year of valid whole-city designation onward and zero otherwise. Cities with a zero or missing adoption year are treated as never treated. According to the compiled policy records, 13 cities entered the program in 2013 and five in 2015. These dates follow the announcement-year convention used in the archive and indicate policy designation rather than the intensity of implementation. As a sensitivity check, a broader treatment definition assigns the three sub-prefectural pilots to their corresponding prefecture-level cities. The preferred definition retains only whole-city designations so that the policy unit is consistent with the city-level outcome. This restriction does not exclude the possibility that sub-city pilots in comparison cities generated some local exposure. [red]; 4.2 Data and Variable Construction; paragraph beginning “The benchmark controls include the natural logarithms of fixed-asset investment,…” [green]; 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red].

Comment 2

The overall structure and flow of the manuscript need considerable improvement. The paper moves between theory, research design, econometric methods, spatial analysis, robustness tests, heterogeneity analysis, and machine learning without a sufficiently clear progression. Reorganizing these sections would make the study much easier to follow.

Response: We appreciate this broader structural comment. We reorganized the manuscript so that the argument now proceeds from the literature and hypotheses to Methods, Empirical Results, Discussion, Conclusions, and Limitations. Within the Results section, the sequence is now: main policy estimates, dynamic evidence, outcome/design sensitivity, spatial exposure, related outcomes, pre-policy heterogeneity, and finally the predictive machine-learning extension. The intention was to make the primary policy analysis visible before the supporting and predictive analyses.

Location in revised manuscript: 4 Methods; paragraph beginning “4 Methods…”; 5 Empirical Results; heading “5. Empirical Results…”; 6 Discussion; paragraph beginning “6 Discussion…”; 7 Conclusions; paragraph beginning “7 Conclusions…”; 8 Limitations and Future Research; paragraph beginning “8 Limitations and Future Research…”.

Comment 3

The study uses a large number of analytical methods, including DID, spatial models, PSM, entropy balancing, heterogeneity analysis, RF-RFECV, and SHAP. The authors should provide a clearer explanation of why each method is included and how these different analyses complement one another.

Response: Thank you for asking us to clarify the role of each method. We now distinguish the methods by the question they answer. The Callaway–Sant’Anna group-time estimator is the primary policy analysis, while TWFE is retained only as a conventional benchmark. The lnCO₂/lnBUA decomposition clarifies the numerator and denominator of lnCLCI. Concurrent-policy controls and pre-policy matching are sensitivity checks for observed selection and policy overlap; exact entropy balancing is reported as infeasible rather than presented as a successful robustness exercise. Spatial exposure is treated as a supporting association, pooled interactions test heterogeneity using pre-treatment groups, and Random Forest/RFECV/SHAP are confined to prediction. Figure 4 summarizes these roles and keeps the predictive branch separate from policy identification.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “Selection adjustment is used only as a supplementary check. Matching…” [red]; 4.4.2 Spatial Policy Exposure; paragraph beginning “The spatial analysis examines whether policy exposure in other cities…” [red]; 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “Related-outcome regressions use the same sample, controls, and city and…” [blue]; 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The prediction sample covers all 40 cities from 2013 to…” [green]; 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The full grouped procedure is repeated with seeds 7, 123,…” [blue]; Figure 4. Analytical workflow separating policy estimation, supporting diagnostics, and city-grouped prediction.

Comment 4

Several variables and methodological terms are not defined with enough detail. For reproducibility, the authors should clearly report how each variable was calculated, including its formula, unit, transformation, data source, and interpretation.

Response: We are grateful for this reproducibility comment. Table 1 has been expanded to consolidate each variable’s measurement, transformation, unit where preserved, and available data provenance. We also corrected the population-per-built-land measure, disclosed the 2011–2012 DII backcasting, and excluded INS from estimation because its original construction cannot be recovered. For lnCLCI, we verified directly that the stored series equals lnCO₂ − lnBUA. Where the archived materials do not preserve upstream details—such as the complete emissions inventory or original DII weights—we state the limitation explicitly rather than reconstructing an undocumented procedure. We recognize that disclosing these gaps does not replace the missing source documentation, and that complete upstream reproducibility remains unresolved.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The dependent variable, lnCLCI, is the natural logarithm of city-level…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The benchmark controls include the natural logarithms of fixed-asset investment,…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The DII workbook provides source-based values for 2013–2022 and identifies…” [green]; Table 1. Variable Definitions and Data Sources [green]; 8 Limitations and Future Research; paragraph beginning “Some source records are also incomplete. The archived materials do…” [green].

Comment 5

Table 3 reports Variance Inflation Factors (VIFs), but it is not clear how these values were calculated. The authors should explain the VIF calculation, identify which variables and regression specification were used, and clarify whether the reported values correspond to the baseline model.

Response: Thank you for requesting a more precise description of the VIF diagnostic. We recalculated VIFs using SCP and the six benchmark controls in the preferred 37-city sample, with an intercept and without city and year dummy variables. For each regressor j, VIF is defined as 1/(1 − R_j²), where R_j² is obtained by regressing that regressor on the remaining regressors in the pooled covariate matrix. We now state explicitly that this is a pooled-matrix diagnostic rather than the VIF of a within-transformed fixed-effects design.

Location in revised manuscript: 4.3 Descriptive Statistics and Diagnostics; paragraph beginning “Note: VIFs use SCP and the six benchmark controls in…” [green]; 4.3 Descriptive Statistics and Diagnostics; paragraph beginning “As shown in Table 3, green patent authorizations and GDP…” [green].

Comment 6

The interpretation of the VIF results also needs more discussion. The VIF values for lnGDP and green patent authorizations are 5.84 and 5.81, respectively. Simply stating that these values are below 10 may not be sufficient, as values above 5 can still indicate meaningful multicollinearity. The authors should examine whether this affects the stability of the regression estimates.

Response: We agree that simply invoking a threshold of 10 was not sufficient. The revised text describes the VIFs for green patent authorizations (5.786) and GDP (5.666) as moderate overlap that may reduce precision. We therefore added coefficient-stability checks: dropping green patents yields an SCP coefficient of 0.008 (p = 0.912), while dropping GDP yields 0.018 (p = 0.806). We interpret these checks cautiously; they show that the near-zero policy estimate is not being generated by either variable alone, but they do not imply that multicollinearity is absent.

Location in revised manuscript: 4.3 Descriptive Statistics and Diagnostics; paragraph beginning “As shown in Table 3, green patent authorizations and GDP…” [green].

Comment 7

There appears to be an important inconsistency in the use of the variable “Smart.” In the policy-treatment description, Smart appears to be a binary variable taking a value of 0 or 1, whereas the descriptive statistics report Smart as a digital-intelligence integration index ranging from 1.32 to 9.41. This needs to be carefully checked and corrected, as Smart is central to the analysis.

Response: Thank you for catching this important naming inconsistency. We now reserve SCP exclusively for the binary Smart City Pilot treatment indicator and DII exclusively for the continuous digital–intelligence integration series. Table 2 has been recalculated for the preferred sample, and DII is summarized separately for its 2013–2022 observed-source period. The former use of ‘Smart’ for two different concepts has been removed.

Location in revised manuscript: Figure 1 shows the location of the 40-city research area and places the regional sample within China. The map is descriptive and also provides the spatial frame used to construct the geographic weight matrices; the preferred policy analysis subsequently excludes the three sub-prefectural pilot cases described in Section 4.1. SCP equals one from the year of valid whole-city designation onward and zero otherwise. Cities with a zero or missing adoption year are treated as never treated. According to the compiled policy records, 13 cities entered the program in 2013 and five in 2015. These dates follow the announcement-year convention used in the archive and indicate policy designation rather than the intensity of implementation. As a sensitivity check, a broader treatment definition assigns the three sub-prefectural pilots to their corresponding prefecture-level cities. The preferred definition retains only whole-city designations so that the policy unit is consistent with the city-level outcome. This restriction does not exclude the possibility that sub-city pilots in comparison cities generated some local exposure. [red]; Table 2 reports descriptive statistics for the preferred 37-city policy sample. SCP and DII are listed as separate variables, and DII is summarized only over its 2013–2022 observed-source period. The table is descriptive; differences between treated and untreated cities are not policy effects. [green]; Table 2. Descriptive Statistics [green].

Comment 8

The construction of the digital-intelligence integration index needs to be explained much more clearly. The manuscript refers to the use of “entropy weighting and/or PCA.” These are different approaches, so the authors should specify exactly which method was used and provide enough information for the index to be reproduced.

Response: We appreciate this comment because the previous wording implied a level of reproducibility that the archived materials do not support. We rechecked the DII files and found a supplied 2013–2022 composite series and a separate record showing that 2011–2012 values were extrapolated backward from city-specific linear trends. The indicator-level weights and subsystem formula are not preserved. We therefore removed the phrase ‘entropy weighting and/or PCA’ rather than choosing one method without evidence. DII is no longer used to identify a causal channel and is used in the prediction analysis only for 2013–2022, where the supplied source values are available. We recognize that disclosing these gaps does not replace the missing source documentation, and that complete upstream reproducibility remains unresolved.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The DII workbook provides source-based values for 2013–2022 and identifies…” [green]; 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The prediction sample covers all 40 cities from 2013 to…” [green]; 8 Limitations and Future Research; paragraph beginning “Some source records are also incomplete. The archived materials do…” [green].

Comment 9

The calculation of the main dependent variable, built-up land carbon intensity (CLCI), should be presented more transparently. The manuscript should provide the complete calculation, including the energy categories considered, emission factors, built-up land definition, units, data sources, and any transformations applied.

Response: Thank you for emphasizing the need for transparent construction of the dependent variable. We verified that lnCLCI equals lnCO₂ − lnBUA for all 480 archived observations. The component files label CO₂ in tonnes and built-up land area in km², and the land worksheet states that linear interpolation was applied. We now interpret the unlogged ratio as tonnes of city-level CO₂ per km² of built-up land. The archived land worksheet labels BUA as built-up land area, but the available records do not preserve a more detailed operational definition or classification rule for what is included as built-up land; this limitation is stated explicitly in the revised manuscript. The archive mentions coal, oil, and natural gas and contains a standard-coal carbon-factor example, but it does not preserve the complete fuel-specific factors, carbon-to-CO₂ conversion, electricity treatment, full inventory boundary, or observation-level interpolation flags. These missing upstream records are disclosed explicitly rather than filled by assumption. We recognize that disclosing these gaps does not replace the missing source documentation, and that complete upstream reproducibility remains unresolved.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The dependent variable, lnCLCI, is the natural logarithm of city-level…” [green]; Table 1. Variable Definitions and Data Sources [green]; Table 6 shows that the estimated coefficient for lnCO₂ is −0.042, with a standard error of 0.047 and a p-value of 0.382, while the coefficient for lnBUA is −0.050, with a standard error of 0.049 and a p-value of 0.316. Their difference reproduces the lnCLCI coefficient to numerical precision, as required by the accounting identity lnCLCI = lnCO₂ − lnBUA under the common sample and specification. The confidence intervals in Table 6 include zero for both components, so the estimates do not determine whether the small lnCLCI response reflects lower emissions, a change in built-up land area, or a combination of both. [blue]; 8 Limitations and Future Research; paragraph beginning “Some source records are also incomplete. The archived materials do…” [green].

Comment 10

A particularly serious issue is the statement that it remains unclear whether CLCI was logarithmically transformed during the original data-processing procedure. The transformation of the main dependent variable cannot remain uncertain, as this directly affects the interpretation of the regression coefficients. The authors need to verify and clearly report this before the results can be properly assessed.

Response: We fully agree that the transformation of the main dependent variable cannot remain uncertain. We therefore checked the stored outcome against its components and verified lnCLCI = lnCO₂ − lnBUA to numerical precision. The revised manuscript consistently labels the outcome as lnCLCI and no longer treats its logarithmic status as unknown.

Comment 11

The in-text citation formatting needs careful revision throughout the manuscript. References are not consistently positioned relative to punctuation: some citations appear before the period, others after it, and in several places there is no space between the citation and the surrounding text. A consistent journal style should be applied throughout.

Response: Thank you for drawing attention to the citation presentation. We revised in-text citation placement and spacing throughout the manuscript so that citation brackets are separated consistently from surrounding words and placed consistently relative to punctuation. We also renumbered citations after removing duplicated reference entries.

Location in revised manuscript: 1 Introduction; paragraph beginning “Urban carbon policy concerns not only energy use but also…”; References; entry beginning “1. Batty, M.; Axhausen, K. W.; Giannotti, F.; Pozdnoukhov, A.;…” [green].

Comment 12

The reference list also needs to be carefully reformatted. DOI presentation is inconsistent, with some references providing complete DOI links and others giving only DOI numbers. Journal titles and other bibliographic elements should also be checked for consistency.

Response: We appreciate this bibliographic comment. We rechecked the reference list for duplicate records, journal-title consistency, bibliographic punctuation, and DOI presentation. Duplicate entries were removed, several titles were corrected against the retained metadata, and DOI formatting was standardized where a DOI could be verified. We did not invent a DOI for records for which one could not be established from the supplied materials.

Location in revised manuscript: References; entry beginning “1. Batty, M.; Axhausen, K. W.; Giannotti, F.; Pozdnoukhov, A.;…” [green]; References; entry beginning “21. Du, M.; Huang, Z.; Hou, X. The impact of…” [blue].

Comment 13

There are many spacing and formatting problems throughout the manuscript. In several places, words, sentences, and citations run together. A thorough formatting and proofreading pass is needed before publication.

Response: Thank you for noting the formatting problems. We carried out a manuscript-wide technical edit covering spacing between words and citations, capitalization, punctuation, terminology, table and figure numbering, and paragraph flow. We also reduced repeated stock transitions and rewrote several dense result passages so that statistical uncertainty is stated directly rather than buried in formatting-heavy summaries.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “Related-outcome regressions use the same sample, controls, and city and…” [blue]; 4 Methods; paragraph beginning “4 Methods…”.

Comment 14

I recommend adding a dedicated Data and Variable Construction subsection. This would allow the authors to explain sample selection, data sources, missing-data handling, transformations, policy coding, and the construction of derived indicators in one clear location.

Response: Thank you for this helpful suggestion. We added a dedicated Section 4.2, ‘Data and Variable Construction,’ so that treatment coding, lnCLCI construction, benchmark controls, derived land measures, DII provenance, transformations, units, and unrecoverable source limitations are explained in one place. Table 1 was revised to support this subsection.

Location in revised manuscript: Figure 1 shows the location of the 40-city research area and places the regional sample within China. The map is descriptive and also provides the spatial frame used to construct the geographic weight matrices; the preferred policy analysis subsequently excludes the three sub-prefectural pilot cases described in Section 4.1. SCP equals one from the year of valid whole-city designation onward and zero otherwise. Cities with a zero or missing adoption year are treated as never treated. According to the compiled policy records, 13 cities entered the program in 2013 and five in 2015. These dates follow the announcement-year convention used in the archive and indicate policy designation rather than the intensity of implementation. As a sensitivity check, a broader treatment definition assigns the three sub-prefectural pilots to their corresponding prefecture-level cities. The preferred definition retains only whole-city designations so that the policy unit is consistent with the city-level outcome. This restriction does not exclude the possibility that sub-city pilots in comparison cities generated some local exposure. [red]; 4.2 Data and Variable Construction; paragraph beginning “The dependent variable, lnCLCI, is the natural logarithm of city-level…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The benchmark controls include the natural logarithms of fixed-asset investment,…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The DII workbook provides source-based values for 2013–2022 and identifies…” [green].

Comment 15

The terminology used for the main variables should be standardized. Terms such as “Smart,” “Smart City implementation,” “Smart City Pilot,” “digital-intelligence integration index,” and “index” are used throughout the paper and can easily be confused. Each concept should have a unique and consistently used name.

Response: We agree that the previous terminology made different concepts unnecessarily easy to confuse. The revised manuscript uses SCP only for the binary Smart City Pilot treatment, DII only for the continuous digital–intelligence integration series, and W × SCP only for neighboring policy exposure. The terms used in tables, equations, results, and the prediction section have been standardized accordingly.

Comment 16

Some of the definitions in Table 1 remain too vague to reproduce the analysis. For example, the interpretation of INS is said to depend on its “coding direction,” but the actual coding direction is not clearly specified. The authors should provide the exact definition and interpretation.

Response: Thank you for pointing out the ambiguity surrounding INS. We could not recover the original formula or coding direction from the archived materials. Rather than assigning an interpretation that cannot be verified, we removed INS from the channel regressions and the predictive feature set and withdrew the earlier industrial-upgrading interpretation. The remaining limitation is stated explicitly. We recognize that disclosing these gaps does not replace the missing source documentation, and that complete upstream reproducibility remains unresolved.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The DII workbook provides source-based values for 2013–2022 and identifies…” [green]; 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “Related-outcome regressions use the same sample, controls, and city and…” [blue]; 8 Limitations and Future Research; paragraph beginning “Some source records are also incomplete. The archived materials do…” [green].

Comment 17

The causal econometric analysis and the predictive machine-learning analysis should be separated more clearly. Although the authors state that SHAP results are predictive rather than causal, the organization of the manuscript could still lead readers to mix the two types of evidence. A clearer structural separation would help avoid this confusion.

Response: We appreciate this concern and have separated the two evidence types more sharply. The policy estimators and their identifying assumptions are presented in Sections 4.4.1–4.4.3 and Results Sections 5.1–5.6. The machine-learning analysis is placed in a separate Section 4.4.4/5.7 and is described only as cross-city prediction. SHAP feature contributions are not used to validate the DID estimates, establish mediation, or define policy thresholds.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The prediction sample covers all 40 cities from 2013 to…” [green]; 5.7 Prediction in Held-Out Cities; paragraph beginning “5.7 Prediction in Held-Out Cities…”.

Comment 18

The machine-learning methodology needs enough detail to allow replication. The authors should report the train/test split, random seed, cross-validation procedure, hyperparameter tuning, selected parameters, RF-RFECV feature-selection procedure, and final selected features.

Response: Thank you for asking for a reproducible description of the machine-learning procedure. The revised Methods reports the 2013–2022 prediction sample, the city-level train/test split, seed 42, five-fold GroupKFold within training cities, the RFECV procedure, the nine candidate predictors, the Random Forest settings, the absence of a separate hyperparameter search, and an OLS benchmark evaluated on the same held-out cities. The implementation uses GroupShuffleSplit with a test fraction of 0.20; the Random Forest contains 100 trees, uses squared-error splitting, has no depth restriction, and uses min_samples_split = 2, min_samples_leaf = 1, max_features = 1.0, and bootstrap = True. The complete grouped procedure is repeated with seeds 7, 123, 2026, and 31415 to assess sensitivity to the particular cities held out. In the Results, the seed-42 RFECV retains seven predictors—DII and the six benchmark controls—while SCP and EI are not selected. For seed 42, Random Forest has a held-out-city R² of 0.568, compared with 0.767 for OLS; the post-selection GroupKFold diagnostic for Random Forest has a mean R² of −0.040 and a standard deviation of 1.075. Figure 6 reports prediction sensitivity across the five city splits, Figure 7 reports the RFECV profile, and Table 13/Figure 8 report the held-out SHAP feature contributions. Post-selection fold scores are described as internal diagnostics rather than independent nested-validation estimates.

Location in revised manuscript: 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The prediction sample covers all 40 cities from 2013 to…” [green]; 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The full grouped procedure is repeated with seeds 7, 123,…” [blue]; 5.7.1 Predictive Performance and Feature Selection; paragraph beginning “Note: Seed 42 assigns 32 cities (320 observations) to training…” [blue]; Table 12. Predictive Performance under City-Grouped Validation [blue]; Figure 6. Prediction sensitivity across five city-grouped train/test splits; held-out-city R² for Random Forest and OLS; Figure 7. RFECV performance using training-city GroupKFold; error bars show one standard deviation across folds; Table 13. Feature Contributions in the Seed-42 Forest [blue]; Figure 8. Mean absolute SHAP contributions for held-out cities in the seed-42 model.

Comment 19

The statement that training and testing samples “should maintain city-level independence” is too vague. The authors should confirm whether city-level independence was actually enforced and explain exactly how cities were separated during training, validation, and testing.

Response: We agree that ‘should maintain city-level independence’ was too vague. City-level independence is now actually enforced: with seed 42, 32 complete cities (320 observations) are assigned to training and eight different cities (80 observations) are held out for testing, so no city appears in both sets. Five-fold GroupKFold is applied only within the training sample and keeps cities intact across folds. The manuscript therefore states the implemented split rather than a recommended procedure. The full grouped procedure is also repeated under four additional seeds, and Figure 6 reports how held-out-city R² varies across the five city splits.

Comment 20

The spatial analysis also needs stronger methodological explanation. In particular, the authors should explain why inverse distance was selected as the main spatial-weighting approach and clearly describe how the spatial matrices were constructed and normalized.

Response: Thank you for requesting a clearer spatial specification. The revised manuscript explains that inverse geographic distance is the main matrix because the hypothesized cross-city exposure is assumed to weaken with geographic separation; contiguity is used as a discrete geographic alternative. The matrices are aligned by city ID, restricted to the 37-city policy sample, and row-normalized after exclusions. The archived ratio matrix is described only as a diagnostic because it is not a pairwise economic-distance matrix and uses full-period information. We also state that these estimates are conditional associations rather than fully identified causal spillovers. We also corrected an inherited equation that still contained a spatial lag of the dependent variable; the revised equation now matches the linear exposure specification described in the text and Table 9.

Location in revised manuscript: 4.4.2 Spatial Policy Exposure; paragraph beginning “The spatial analysis examines whether policy exposure in other cities…” [red]; 4.4.2 Spatial Policy Exposure; paragraph beginning “W denotes a fixed spatial-weight matrix with zeros on the…” [red]; 5.4 Spatial Policy Exposure; paragraph beginning “Note: Table 9 reports coefficients from the linear spatial-exposure model…” [red]; 6.4 Spatial Exposure and Regional Interpretation; paragraph beginning “The spatial estimates in Table 9 are sensitive to how…” [red].

Comment 21

Methods and results are not always clearly separated. In several places, methodological choices are introduced alongside their findings and interpretation. The model specifications, assumptions, diagnostic tests, and robustness procedures should ideally be established before presenting the corresponding results.

Response: We agree that the previous version blurred methods and findings. Section 4 now establishes the sample, variable definitions, diagnostic procedures, model specifications, assumptions, and validation design before Section 5 reports the corresponding estimates. Results subsections now refer back to those established specifications rather than introducing new methodological choices alongside the findings.

Location in revised manuscript: 4 Methods; paragraph beginning “4 Methods…”; 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “4.4.1 Group-Time Treatment Effects and TWFE Benchmark…”; 5 Empirical Results; heading “5. Empirical Results…”.

Comment 22

The paper contains many hypotheses, robustness tests, spatial analyses, mechanisms, heterogeneity analyses, and machine-learning results, which makes the main message difficult to identify. The authors should streamline the presentation and clearly distinguish the primary analyses from supporting or supplementary analyses.

Response: Thank you for encouraging us to streamline the manuscript. We now treat the group-time estimator as the primary policy analysis and demote the remaining analyses to clearly defined supporting roles. We no longer present infeasible entropy balancing as a robustness estimate, no longer frame related land outcomes as mediation, and keep spatial exposure and machine learning separate from the main causal interpretation. Figure 4 makes this hierarchy explicit, and the Conclusions now focus on the uncertainty of the main estimates rather than accumulating claims across methods.

Location in revised manuscript: Figure 4. Analytical workflow separating policy estimation, supporting diagnostics, and city-grouped prediction.; 5.1 Main Policy Estimates; paragraph beginning “As reported in Table 5, the unadjusted group-time ATT is…” [red]; 7 Conclusions; paragraph beginning “In the preferred 37-city sample, Table 5 reports an unadjusted…” [red].

Comment 23

Definitions should be provided when terms first appear. Several abbreviations, indices, constructed variables, and methodological concepts are introduced before they are fully explained, which makes the manuscript unnecessarily difficult to follow.

Response: We appreciate this readability point. Section 4.2 and Table 1 consolidate the definitions of lnCLCI, BUA, SCP, DII, EI, and the land-based derived measures before the empirical models are introduced. Group-time ATT, fixed effects, spatially weighted exposure, and the city-grouped prediction split are then defined in Section 4.4 before their results appear.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The dependent variable, lnCLCI, is the natural logarithm of city-level…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The benchmark controls include the natural logarithms of fixed-asset investment,…” [green]; 4.2 Data and Variable Construction; paragraph beginning “The DII workbook provides source-based values for 2013–2022 and identifies…” [green]; 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red].

Comment 24

The equations need clearer explanation. Every symbol used in the models should be defined immediately after the relevant equation, including subscripts, fixed effects, spatial terms, treatment variables, control vectors, and error terms.

Response: Thank you for requesting clearer notation. The revised manuscript now defines the symbols immediately after both the TWFE benchmark and the spatial-exposure specification. For TWFE, the text defines the city and year indices, SCP, the six-control vector, city and year fixed effects, the error term, and the city-clustered inference convention. For the spatial-exposure equation, SCPᵢₜ is defined as the local Smart City Pilot treatment indicator, SCPⱼₜ as the treatment status of city j used to construct neighboring policy exposure, and wᵢⱼ, Xᵢₜ, Xⱼₜ, the fixed effects, the error term, and the coefficient vectors are defined in the accompanying paragraph. The revised text also distinguishes the local-treatment and neighboring-exposure coefficients and states explicitly that the specification does not include a spatial lag of the dependent variable. Because city-clustered inference permits within-city dependence over time but does not address residual dependence across cities, the spatial coefficients are interpreted as conditional associations rather than fully identified direct and spillover treatment effects.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “Here, i indexes cities and t indexes years. SCP is…” [red]; 4.4.2 Spatial Policy Exposure; paragraph immediately following “The spatial-exposure regression is specified as,” beginning “Here, i and j index cities and t indexes years…” [red].

Comment 25

The manuscript needs a thorough language and technical edit. There are several grammatical issues, inconsistent terminology, capitalization problems, missing spaces, and awkward transitions. These issues sometimes make technically important sections harder to understand.

Response: We appreciate the request for a thorough language and technical edit. We revised grammar, capitalization, terminology, spacing, and transitions throughout the manuscript. We also replaced several overly compressed result statements with connected prose and consistently distinguish ‘statistically imprecise’ from ‘no effect.’ The editing was intended to make the technical sections easier to follow without strengthening the substantive claims.

Location in revised manuscript: 4 Methods; paragraph beginning “4 Methods…”; 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “Related-outcome regressions use the same sample, controls, and city and…” [blue].

Comment 26

The Results section could be reorganized to make the main findings much clearer. I suggest presenting the baseline estimates first, followed by identification and robustness checks, spatial analysis, mechanisms and heterogeneity, and finally the machine-learning analysis.

Response: Thank you for suggesting a clearer Results sequence. We adopted that logic. Section 5 now begins with the main policy estimates and dynamic evidence, moves to outcome/design sensitivity, then spatial exposure, related outcomes and heterogeneity, and ends with the predictive machine-learning analysis. This ordering places the primary estimates before all supporting extensions.

Location in revised manuscript: 5 Empirical Results; heading “5. Empirical Results…”; 5.1 Main Policy Estimates; paragraph beginning “5.1 Main Policy Estimates…”; 5.2 Dynamic Effects: Pre-policy Trends and Post-policy Changes; paragraph beginning “5.2 Dynamic Effects: Pre-policy Trends and Post-policy Changes…”; 5.3 Outcome and Design Sensitivity; paragraph beginning “5.3 Outcome and Design Sensitivity…”; 5.4 Spatial Policy Exposure; paragraph beginning “5.4 Spatial Policy Exposure…”; 5.5 Related Outcomes and Channel Limitations; paragraph beginning “5.5 Related Outcomes and Channel Limitations…”; Table 11. Pooled Interaction Tests with Pre-Policy Groups [blue]; 5.7 Prediction in Held-Out Cities; paragraph beginning “5.7 Prediction in Held-Out Cities…”.

Comment 27

A simple methodological flowchart would greatly improve the manuscript. Showing the progression from data and variable construction to treatment definition, baseline DID, robustness tests, spatial analysis, channel/heterogeneity analysis, and machine learning would help readers understand how the different parts of the study fit together.

Response: Thank you for this practical suggestion. We added Figure 4 as a methodological workflow. It separates data/treatment reconstruction, the main policy-estimation branch, supporting diagnostics and sensitivity analyses, and the city-grouped prediction branch. The figure deliberately does not depict a confirmed causal mechanism, because the revised manuscript does not claim one.

Location in revised manuscript: Figure 4. Analytical workflow separating policy estimation, supporting diagnostics, and city-grouped prediction.

Thank you again for your careful review and thoughtful suggestions. Your comments have been very helpful in improving the clarity, rigor, and overall presentation of the manuscript. We have carefully revised the paper in response to your concerns and have also made the remaining limitations more explicit where the available data and records do not allow a complete resolution. We sincerely appreciate the time and effort you devoted to evaluating our work. We hope that the revisions have adequately addressed your concerns and that the manuscript is now clearer and more robust. We would be very grateful if you could consider the revised version favorably, and we sincerely hope that it is now suitable for publication.

Sincerely,

The Authors

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I appreciate the authors’ effort. To confirm the value of their work, I have several minor concerns about the paper.

The channel variables share their denominator with the outcome: land_POP = population/built-up land and land_expand is built-up land growth, so their "support" for H3c is partly an identity rather than new evidence; EI is significant only at 10% (t = −1.74) and land_GDP is insignificant. Please either run a formal mediation with bootstrap confidence intervals for the indirect effect or relabel Table 7 as reduced-form associations and drop/reframe the denominator-overlapping indicators.
The heterogeneity splits use time-varying lndens and land_expand, which can be affected by the pilot, and no formal test of the difference between subsample coefficients is reported (Table 8 only lists separate t-statistics). Please define the high/low groups from pre-treatment values, estimate interaction models, and report a permutation or interaction-based test of whether the −0.161 (high-density) and −0.022 (low-density) coefficients actually differ.
Because CLCI = emissions / built-up land, a decline can arise mechanically from land expansion when emissions grow more slowly, yet the paper reports no separate effects on emissions and on land area. The authors should add more recently published related papers about low-carbon development as complementary references, such as: doi.org/10.1016/j.jenvman.2026.130707. lease decompose the result by estimating the policy effect on log emissions and log built-up land separately, so that readers can judge whether the pilot cut emissions or merely diluted them over a larger denominator.
The machine-learning section concedes that it is unknown whether the train/test split respected city-level independence, so with repeated city-year observations the reported test R² of 0.8191 and CV R² of 0.8931 are likely inflated by leakage. Please re-estimate with GroupKFold or leave-one-city-out, report the resulting performance alongside an OLS benchmark, and note that the exclusion of Smart from the RFECV feature set is consistent with a modest average effect and should not be used to reinforce the mediation narrative.
index is included as a predictor even though it is itself significantly raised by the pilot (0.333, t = 6.05), so SHAP importance for a post-treatment channel cannot be combined with the DID result; the SHAP "turning points" in Figures 10–11 are also discussed in language close to causal thresholds. Please add stability checks across random seeds and restate these findings strictly as in-sample predictive associations with no policy-threshold reading.

Author Response

Response to Reviewer 2

Dear Reviewer 2,

We sincerely thank you for taking the time to review our manuscript so carefully and for providing thoughtful and constructive comments. Your suggestions have been particularly helpful in clarifying the distinction between related land outcomes and mediation, avoiding the use of post-treatment subgroup definitions, and addressing potential leakage in the prediction analysis. These comments also prompted us to reconsider several aspects of the empirical design and the way the results are presented.We fully acknowledge the weaknesses you identified and have revised both the analysis and the corresponding discussion with great care. Where the available records do not permit a complete reconstruction, we have stated the remaining limitations explicitly and narrowed the interpretation of the findings rather than implying that the issue has been fully resolved. We are very grateful for your careful attention to these points, as they have helped us improve the rigor, transparency, and overall clarity of the manuscript.Below, we reproduce your comments and respond to each of them individually. For ease of checking, each location note identifies the relevant section, table, or figure in the revised manuscript, together with a searchable text anchor and the corresponding revision color. We sincerely appreciate the time and effort you devoted to evaluating our work and are grateful for the opportunity to revise the manuscript in response to your suggestions.

Revision colors in the combined manuscript are as follows: Reviewer 1, green; Reviewer 2, blue; Reviewer 3, red. When a revision addresses comments raised by more than one reviewer, one primary reviewer color is used and identified in the corresponding location note. To ensure that the revisions remain easy to locate even if pagination changes, location notes refer to section, table, or figure numbers together with searchable text anchors. The response-text color identifies the reviewer, while the color stated in the location note indicates the corresponding revision in the manuscript.

Comment 1

The channel variables share their denominator with the outcome: land_POP = population/built-up land and land_expand is built-up land growth, so their "support" for H3c is partly an identity rather than new evidence; EI is significant only at 10% (t = −1.74) and land_GDP is insignificant. Please either run a formal mediation with bootstrap confidence intervals for the indirect effect or relabel Table 7 as reduced-form associations and drop/reframe the denominator-overlapping indicators.

Response: We appreciate this important point and agree that variables sharing the built-up-land denominator with lnCLCI should not be presented as independent mediation evidence. Following your suggestion, we removed the causal mediation interpretation and recast the analysis as reduced-form related-outcome evidence. DII is excluded from channel identification because observed pre-policy values are unavailable for the first treatment cohort, and INS is excluded because its original formula cannot be recovered. We also reconstructed resident population per built-up land area from the component data instead of using the archived land_POP series. The resulting coefficient is 0.047 (p = 0.348). The land-based ratios and land-growth measure are now explicitly described as denominator-related outcomes, not mediators. The table that served this purpose in the previous version has been reorganized and is now Table 10. We have followed the reduced-form option in your comment; no bootstrap mediation result is claimed. The updated EI estimate is 0.002 (p = 0.892), and the built-up-land-growth estimate is −0.042 (p = 0.054), so these results are not presented as confirmed mechanisms.

Location in revised manuscript: 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “Related-outcome regressions use the same sample, controls, and city and…” [blue]; Table 10. Related-Outcome Regressions [blue]; 6.1 The Policy Result and the Land-Based Outcome; paragraph beginning “The component results in Table 6 do not show a…” [blue].

Comment 2

The heterogeneity splits use time-varying lndens and land_expand, which can be affected by the pilot, and no formal test of the difference between subsample coefficients is reported (Table 8 only lists separate t-statistics). Please define the high/low groups from pre-treatment values, estimate interaction models, and report a permutation or interaction-based test of whether the −0.161 (high-density) and −0.022 (low-density) coefficients actually differ.

Response: Thank you for identifying the post-treatment grouping problem. We agree that using time-varying density or land expansion to define high/low subsamples could condition the analysis on variables affected by the policy. We therefore define both groups from 2011–2012 city means and keep group membership fixed over time. We then estimate pooled interaction models and test the interaction coefficient directly. The high-minus-low density contrast is −0.123 (p = 0.180), and the corresponding land-growth contrast is 0.059 (p = 0.498). Neither contrast provides clear evidence of systematic heterogeneity, so the earlier subgroup-significance comparison and density-specific policy recommendation have been removed. The former heterogeneity table has been reorganized and is now Table 11.

Location in revised manuscript: 4.4.3 Related Outcomes and Pre-policy Heterogeneity; paragraph beginning “For each variable, cities above the median of the preferred…” [blue]; 5.6 Heterogeneity Using Pre-Policy City Conditions; paragraph beginning “Note: Groups use 2011–2012 city means and the preferred-sample median.…” [blue]; 5.6 Heterogeneity Using Pre-Policy City Conditions; paragraph beginning “As shown in Table 11, the high-minus-low density difference is…” [blue]; 6.3 City Conditions and Prediction; paragraph beginning “The pooled interaction estimates in Table 11 do not provide…” [blue].

Comment 3

Because CLCI = emissions / built-up land, a decline can arise mechanically from land expansion when emissions grow more slowly, yet the research reports no separate effects on emissions and on land area. The authors should add more recently published related research about low-carbon development as complementary references, such as: doi.org/10.1016/j.jenvman.2026.130707. lease decompose the result by estimating the policy effect on log emissions and log built-up land separately, so that readers can judge whether the pilot cut emissions or merely diluted them over a larger denominator.

Response: We are grateful for this suggestion because it changes how the ratio outcome should be interpreted. We estimated the policy response separately for lnCO₂ and lnBUA using the same preferred sample and benchmark TWFE regressors. The lnCO₂ coefficient is −0.042 (SE = 0.047, p = 0.382), while the lnBUA coefficient is −0.050 (SE = 0.049, p = 0.316). Their difference reproduces the lnCLCI coefficient to numerical precision, as implied by lnCLCI = lnCO₂ − lnBUA. Because both component estimates are statistically imprecise, the revised manuscript does not claim either absolute emission reduction or denominator dilution. We also added the suggested 2026 Journal of Environmental Management paper (DOI: 10.1016/j.jenvman.2026.130707) as Reference 21 and discuss it as related low-carbon digitalization evidence rather than as direct evidence on the same policy/outcome.

Location in revised manuscript: 2.1 Smart Cities and Low-Carbon Development; paragraph beginning “Related evidence also comes from China’s national smart-tourism pilot, where…” [blue]; References; entry beginning “21. Du, M.; Huang, Z.; Hou, X. The impact of…” [blue]; 5.3.1 Emissions and Land-Area Decomposition; paragraph beginning “The ratio decomposition estimates lnCO₂ and lnBUA using the same…” [blue]; Table 6. Policy Responses in the Ratio and Its Components [blue]; Table 6 shows that the estimated coefficient for lnCO₂ is −0.042, with a standard error of 0.047 and a p-value of 0.382, while the coefficient for lnBUA is −0.050, with a standard error of 0.049 and a p-value of 0.316. Their difference reproduces the lnCLCI coefficient to numerical precision, as required by the accounting identity lnCLCI = lnCO₂ − lnBUA under the common sample and specification. The confidence intervals in Table 6 include zero for both components, so the estimates do not determine whether the small lnCLCI response reflects lower emissions, a change in built-up land area, or a combination of both. [blue]; 6.1 The Policy Result and the Land-Based Outcome; paragraph beginning “The component results in Table 6 do not show a…” [blue].

Comment 4

The machine-learning section concedes that it is unknown whether the train/test split respected city-level independence, so with repeated city-year observations the reported test R² of 0.8191 and CV R² of 0.8931 are likely inflated by leakage. Please re-estimate with GroupKFold or leave-one-city-out, report the resulting performance alongside an OLS benchmark, and note that the exclusion of Smart from the RFECV feature set is consistent with a modest average effect and should not be used to reinforce the mediation narrative.

Response: Thank you for pointing out the leakage risk in the earlier prediction design. We reran the predictive analysis with city-level separation throughout. Under seed 42, all observations from 32 cities are assigned to the training sample and all observations from eight different cities are held out for testing, so no city appears in both sets. Five-fold GroupKFold is applied only within the training cities, and RFECV is likewise confined to the training sample. Under this grouped design, Random Forest achieves a held-out-city R² of 0.568, with an RMSE of 0.400 and an MAE of 0.328. The OLS benchmark evaluated on the same held-out cities performs better, with R² = 0.767, RMSE = 0.294, and MAE = 0.251. The post-selection GroupKFold diagnostic for Random Forest has a mean R² of −0.040 and a standard deviation of 1.075, indicating substantial variation across city folds. The seed-42 RFECV retains seven predictors: DII and the six benchmark controls; SCP and EI are excluded. Their exclusion is not used to infer a policy effect or mediation. The earlier observation-level test and cross-validation results have therefore been withdrawn. Figure 6 reports the sensitivity of held-out-city R² across five city splits, and Figure 7 reports the RFECV profile. Post-selection fold scores reuse the feature-selection folds and are treated as internal diagnostics, not independent nested-validation estimates.

Location in revised manuscript: 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The prediction sample covers all 40 cities from 2013 to…” [green]; 5.7.1 Predictive Performance and Feature Selection; paragraph beginning “Note: Seed 42 assigns 32 cities (320 observations) to training…” [blue]; Table 12. Predictive Performance under City-Grouped Validation [blue]; Figure 6. Prediction sensitivity across five city-grouped train/test splits; held-out-city R² for Random Forest and OLS; Figure 7. RFECV performance using training-city GroupKFold; error bars show one standard deviation across folds.

Comment 5

index is included as a predictor even though it is itself significantly raised by the pilot (0.333, t = 6.05), so SHAP importance for a post-treatment channel cannot be combined with the DID result; the SHAP "turning points" in Figures 10–11 are also discussed in language close to causal thresholds. Please add stability checks across random seeds and restate these findings strictly as in-sample predictive associations with no policy-threshold reading.

Response: We agree that SHAP importance for a predictor that may respond to policy designation cannot establish a mechanism. The complete city-grouped procedure is repeated using seeds 42, 7, 123, 2026, and 31415. Across these five city splits, held-out R² for Random Forest ranges from 0.305 to 0.673, while OLS ranges from 0.201 to 0.767; Figure 6 reports this sensitivity directly. SHAP values use only the held-out cities in the seed-42 forest. DII has the largest mean absolute contribution at 0.538, followed by lnGDP at 0.163, lndens at 0.121, and fiscal at 0.107. SCP and EI are not retained in that forest and therefore have no estimated SHAP contribution. These quantities are contemporaneous predictive associations conditional on correlated predictors, not mediation evidence, causal mechanisms, or policy thresholds. The previous turning-point interpretation has been removed. Figure 7 reports the RFECV profile and Figure 8 reports the held-out SHAP ranking.

Location in revised manuscript: 4.4.4 Predictive Machine-Learning Analysis; paragraph beginning “The full grouped procedure is repeated with seeds 7, 123,…” [blue]; 5.7.1 Predictive Performance and Feature Selection; paragraph beginning “Figure 6 shows the sensitivity of held-out-city R² to the city split…” [blue]; Figure 6. Prediction sensitivity across five city-grouped train/test splits; held-out-city R² for Random Forest and OLS; Figure 7. RFECV performance using training-city GroupKFold; error bars show one standard deviation across folds; Table 13. Feature Contributions in the Seed-42 Forest [blue]; Figure 8. Mean absolute SHAP contributions for held-out cities in the seed-42 model; 5.7.2 Feature Contributions [blue]; 5.7.3 Scope of Predictive Interpretation [blue].

Sincerely,

The Authors

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript addresses a relevant and timely research question: whether Smart City Pilot programs affect the carbon intensity of built-up land in Chinese cities.

My concerns are:

The main TWFE-DID specification is problematic when treatment occurs at different times and treatment effects may be heterogeneous. The authors should consider estimators specifically designed for staggered adoption, such as Callaway–Sant’Anna or Sun–Abraham, and report cohort- and period-specific treatment effects.
The manuscript appears to classify an entire prefecture-level city as treated when the pilot program concerns a new town, district, or county. This assumption requires stronger justification and robustness analysis.
The evidence for spatial spillovers is not sufficiently stable. The estimated spatial effects vary considerably depending on the spatial-weight matrix. The authors should therefore be cautious about claiming a general spatial spillover effect and provide stronger justification for the selected spatial specification.
The manuscript should clearly report the level at which standard errors are clustered, the number of clusters, the variance estimator used, whether small-cluster corrections are applied and how spatial and serial dependence are handled.
The paper should specify the emission sources, energy categories, emission factors, electricity treatment, definition of built-up land, units, missing-value treatment, and whether CLCI is transformed logarithmically.
The interpretation of the reported coefficient CLCI =−0.0745 is not sufficiently clear.
Smart is defined elsewhere as a binary treatment variable, but Table 2 reports a mean of 3.96, with a minimum of 1.32 and a maximum of 9.41. These values appear to correspond to the digital-intelligence integration index rather than the binary treatment variable. The variable name should be checked and corrected.
PSM and entropy balancing improve balance on observed characteristics but do not eliminate the problems associated with staggered treatment, policy overlap, and spatial interference. Moreover, the results are not fully stable: the PSM-DID estimate becomes statistically insignificant, whereas the entropy-balancing estimate remains significant.
The Random Forest results demonstrate predictive performance, not causal effects. More importantly, the manuscript should clearly explain how training, validation, and test observations were divided.

Author Response

Response to Reviewer 3

Dear Reviewer 3,

We are sincerely grateful for your careful reading of our manuscript and for the rigorous methodological comments you provided. Your review encouraged us to re-examine several key aspects of the empirical design, particularly the treatment of staggered policy adoption, the definition of the treated unit, the interpretation of spatial exposure, and the reporting of statistical inference. These issues are central to the credibility of the analysis, and your comments helped us identify places where the original manuscript required more precise justification and a more cautious presentation of the evidence.

In response, we revisited the relevant specifications, clarified the scope of the treatment definition, strengthened the explanation of the identification strategy, and revised the interpretation of the spatial and inferential results. We have also made the remaining data and documentation limitations more explicit. Where the available records do not permit a complete reconstruction, we state this directly and avoid drawing conclusions beyond what the evidence can support. We greatly appreciate your emphasis on these points, as it has helped us make the manuscript more transparent and methodologically disciplined.Below, we respond to each of your comments in detail. For ease of checking, each response identifies the relevant section, table, or figure in the revised manuscript, together with a searchable text anchor and the corresponding revision color. We are very thankful for the time and expertise you devoted to this review and for the opportunity to improve the manuscript through your suggestions.

Revision colors in the combined manuscript are as follows: Reviewer 1, green; Reviewer 2, blue; Reviewer 3, red. When one revision addresses comments from more than one reviewer, a single primary reviewer color is used and identified in the corresponding location note. To make the revisions easy to trace even if pagination changes, the location notes refer to section, table, or figure numbers together with searchable text anchors. The response-text color identifies the reviewer, while the color stated in the location note indicates the corresponding revision in the manuscript.

Comment 1

The main TWFE-DID specification is problematic when treatment occurs at different times and treatment effects may be heterogeneous. The authors should consider estimators specifically designed for staggered adoption, such as Callaway–Sant’Anna or Sun–Abraham, and report cohort- and period-specific treatment effects.

Response: Thank you for raising this important identification issue. We agree that conventional TWFE should not be treated as the preferred estimator under staggered adoption when treatment effects may vary across cohorts and time. We therefore use Callaway–Sant’Anna group-time ATT(g,t) as the primary policy estimator and retain TWFE only as a conventional benchmark. The unadjusted aggregate ATT is 0.026 (SE = 0.072, bootstrap p = 0.708), the pre-treatment-covariate-adjusted estimate is −0.040 (SE = 0.098, p = 0.673), and the TWFE benchmark is 0.008 (p = 0.909). Table 5 reports the overall average and separate averages for the 2013 and 2015 treatment cohorts. Figure 5 provides the period-specific relative-time dynamics from the Sun–Abraham interaction-weighted event study with year −1 as the reference. The pointwise intervals include zero in all reported pre-treatment and post-treatment periods, and we state explicitly that the short common pre-policy window limits the strength of the pre-trend evidence.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; Table 5. DID Estimates under Alternative Treatment Definitions and Staggered Adoption [red]; 5.1 Main Policy Estimates; paragraph beginning “As reported in Table 5, the unadjusted group-time ATT is…” [red]; 5.2 Dynamic Effects: Pre-policy Trends and Post-policy Changes; paragraph beginning “Figure 5 presents the Sun–Abraham event-time estimates…” [red]; Figure 5. Sun–Abraham interaction-weighted event-time estimates with city-clustered pointwise 95% intervals; relative year −1 is the reference; 6.2 Identification and Concurrent Policies; paragraph beginning “Smart City Pilot designation is not randomly assigned. Reconstructing city-specific…” [red].

Comment 2

The manuscript appears to classify an entire prefecture-level city as treated when the pilot program concerns a new town, district, or county. This assumption requires stronger justification and robustness analysis.

Response: We are grateful for this comment because it prompted us to re-audit the policy matching at the administrative-unit level. We found three cases—Luoyang, Lüliang, and Baotou—where the compiled designation refers to a district, county, or new town rather than the whole prefecture-level city. The preferred policy sample therefore excludes these three cases and contains 37 cities (444 observations), including 18 treated and 19 never-treated cities. We also report strict and broad all-city coding as sensitivity checks. This improves alignment between the policy unit and the city-level outcome, while the revised Limitations section still acknowledges that sub-city exposure may exist within comparison cities.

Location in revised manuscript: 4.1 Sample and Data Sources; paragraph beginning “The balanced panel covers 40 prefecture-level cities from 2011 to…” [red]; Figure 1 shows the location of the 40-city research area and places the regional sample within China. The map is descriptive and also provides the spatial frame used to construct the geographic weight matrices; the preferred policy analysis subsequently excludes the three sub-prefectural pilot cases described in Section 4.1. SCP equals one from the year of valid whole-city designation onward and zero otherwise. Cities with a zero or missing adoption year are treated as never treated. According to the compiled policy records, 13 cities entered the program in 2013 and five in 2015. These dates follow the announcement-year convention used in the archive and indicate policy designation rather than the intensity of implementation. As a sensitivity check, a broader treatment definition assigns the three sub-prefectural pilots to their corresponding prefecture-level cities. The preferred definition retains only whole-city designations so that the policy unit is consistent with the city-level outcome. This restriction does not exclude the possibility that sub-city pilots in comparison cities generated some local exposure. [red]; Table 4. Whole-City Treatment Cohorts in the Preferred Sample [red]; Table 5. DID Estimates under Alternative Treatment Definitions and Staggered Adoption [red]; 8 Limitations and Future Research; paragraph beginning “This study has several limitations. The common pre-treatment window is…” [red].

Comment 3

The evidence for spatial spillovers is not sufficiently stable. The estimated spatial effects vary considerably depending on the spatial-weight matrix. The authors should therefore be cautious about claiming a general spatial spillover effect and provide stronger justification for the selected spatial specification.

Response: Thank you for emphasizing the instability of the spatial results. We agree that the earlier interpretation was too strong. The revised manuscript uses inverse geographic distance as the main exposure matrix because the hypothesized cross-city exposure is assumed to weaken with geographic separation, while contiguity is retained as a discrete geographic alternative. The archived ratio network is reported only as a diagnostic because it is not a pairwise economic-distance matrix and incorporates information from the full sample period. Under both geographic matrices, the coefficient on neighboring policy exposure is statistically imprecise. The revised manuscript therefore does not claim a general causal spillover effect. We also revised the spatial-exposure specification and its notation so that the equation and its interpretation are fully aligned. SCPᵢₜ denotes the local Smart City Pilot treatment indicator, whereas SCPⱼₜ denotes the treatment status of city j used to construct neighboring policy exposure. The local and spatially weighted control variables are defined separately, and the specification explicitly excludes a spatial lag of the dependent variable. Because the design does not separately identify spatial interference and city-clustered covariance does not address residual cross-city dependence, the coefficients on local treatment and neighboring policy exposure are interpreted as conditional associations rather than fully identified direct and spillover treatment effects.

Location in revised manuscript: Section 4.4.2, Spatial Policy Exposure; paragraphs beginning “The spatial analysis examines whether policy exposure in other cities…” and “Here, i and j index cities and t indexes years…” [red]; Table 9. Conditional Policy Exposure under Alternative Spatial Matrices [red]; Section 5.4, Spatial Policy Exposure [red]; Section 6.4, Spatial Exposure and Regional Interpretation [red].

Comment 4

The manuscript should clearly report the level at which standard errors are clustered, the number of clusters, the variance estimator used, whether small-cluster corrections are applied and how spatial and serial dependence are handled.

Response: Thank you for requesting a more complete account of statistical inference. The revised manuscript states that the preferred TWFE model uses 37 city clusters, city and year fixed effects, a CR1 cluster-robust covariance estimator with finite-sample adjustment, and t inference with 36 degrees of freedom. The matched model uses 27 city clusters and 26 degrees of freedom. Group-time inference uses 999 city-cluster multiplier-bootstrap replications with seed 20261003, and the event-study intervals are reported as pointwise rather than simultaneous bands. We also distinguish what clustering does and does not address: city clustering permits serial dependence within cities but does not account for residual dependence across cities, and the spatial-exposure analysis is not presented as a covariance correction for cross-city dependence.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “Here, i indexes cities and t indexes years. SCP is…” [red]; 5.1 Main Policy Estimates; note beginning “Group-time estimates use never-treated cities and 999 city-cluster multiplier-bootstrap replications…” [red]; 5.3.3 Selection Adjustment Using Pre-Policy Covariates; note beginning “Matching uses city means of the six controls over 2011–2012…” [red]; 5.4 Spatial Policy Exposure; note beginning “Table 9 reports coefficients from the linear spatial-exposure model…” [red]; 8 Limitations and Future Research; paragraph beginning “This study has several limitations. The common pre-treatment window is…” [red].

Comment 5

The research should specify the emission sources, energy categories, emission factors, electricity treatment, definition of built-up land, units, missing-value treatment, and whether CLCI is transformed logarithmically.

Response: We appreciate this request for fuller source documentation. We verified that lnCLCI = lnCO₂ − lnBUA across all 480 archived observations; the component files label CO₂ in tonnes and BUA in km², and the land worksheet states that linear interpolation was applied. The archived land worksheet labels BUA as built-up land area, but the available records do not preserve a more detailed operational definition or classification rule for what is included as built-up land; this remaining limitation is now stated explicitly. The archive refers to coal, oil, and natural gas and contains a standard-coal carbon-factor example. However, it does not preserve the complete fuel-specific emission factors, the carbon-to-CO₂ conversion procedure, electricity accounting, the full inventory boundary, or observation-level interpolation flags. Rather than reconstructing these missing elements from assumption, the revised manuscript reports them as upstream reproducibility limitations. The dependent variable is consistently identified as logarithmic. We recognize that disclosing these gaps does not replace the missing source documentation, and that complete upstream reproducibility remains unresolved.

Location in revised manuscript: 4.2 Data and Variable Construction; paragraph beginning “The dependent variable, lnCLCI, is the natural logarithm of city-level…” [green]; Table 1. Variable Definitions and Data Sources [green]; 8 Limitations and Future Research; paragraph beginning “Some source records are also incomplete. The archived materials do…” [green].

Comment 6

The interpretation of the reported coefficient CLCI =−0.0745 is not sufficiently clear.

Response: Thank you for asking us to clarify the reported coefficient. After reconstructing the treatment coding and replacing TWFE as the primary estimator, the earlier coefficient of −0.0745 is no longer part of the revised results and its previous percentage interpretation has been withdrawn. The preferred revised estimates are the group-time effects reported above, while TWFE is retained only as a benchmark. Because all of these estimates are statistically imprecise, we do not translate the benchmark coefficient into a claimed percentage reduction.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “The main policy analysis follows Callaway and Sant’Anna [38] and…” [red]; 5.1 Main Policy Estimates; paragraph beginning “As reported in Table 5, the unadjusted group-time ATT is…” [red]; 7 Conclusions; paragraph beginning “In the preferred 37-city sample, Table 5 reports an unadjusted…” [red].

Comment 7

Smart is defined elsewhere as a binary treatment variable, but Table 2 reports a mean of 3.96, with a minimum of 1.32 and a maximum of 9.41. These values appear to correspond to the digital-intelligence integration index rather than the binary treatment variable. The variable name should be checked and corrected.

Response: Thank you for catching this inconsistency. We now use SCP only for the binary policy-treatment indicator and DII only for the continuous digital–intelligence integration index. Table 2 is recalculated for the preferred sample, SCP is shown with the expected 0–1 range, and DII is reported separately for the 2013–2022 observed-source period. The former ‘Smart’ row that mixed the two concepts has been removed.

Location in revised manuscript: Figure 1 shows the location of the 40-city research area and places the regional sample within China. The map is descriptive and also provides the spatial frame used to construct the geographic weight matrices; the preferred policy analysis subsequently excludes the three sub-prefectural pilot cases described in Section 4.1. SCP equals one from the year of valid whole-city designation onward and zero otherwise. Cities with a zero or missing adoption year are treated as never treated. According to the compiled policy records, 13 cities entered the program in 2013 and five in 2015. These dates follow the announcement-year convention used in the archive and indicate policy designation rather than the intensity of implementation. As a sensitivity check, a broader treatment definition assigns the three sub-prefectural pilots to their corresponding prefecture-level cities. The preferred definition retains only whole-city designations so that the policy unit is consistent with the city-level outcome. This restriction does not exclude the possibility that sub-city pilots in comparison cities generated some local exposure. [red]; Table 2 reports descriptive statistics for the preferred 37-city policy sample. SCP and DII are listed as separate variables, and DII is summarized only over its 2013–2022 observed-source period. The table is descriptive; differences between treated and untreated cities are not policy effects. [green]; Table 2. Descriptive Statistics [green].

Comment 8

PSM and entropy balancing improve balance on observed characteristics but do not eliminate the problems associated with staggered treatment, policy overlap, and spatial interference. Moreover, the results are not fully stable: the PSM-DID estimate becomes statistically insignificant, whereas the entropy-balancing estimate remains significant.

Response: We agree with the reviewer that matching and reweighting do not solve staggered timing, policy overlap, unobserved selection, or spatial interference. We therefore withdrew the previous pooled PSM and entropy-balancing estimates. The revised matching exercise uses 2011–2012 city-level covariate means, nearest-neighbor matching with replacement, and a weighted city-and-year fixed-effects regression; the resulting SCP estimate is −0.081 (SE = 0.108, p = 0.458). Exact nonnegative entropy balancing of all six pre-policy covariate means is infeasible because the treated target lies outside the feasible convex support of comparison cities, so no replacement entropy-balancing coefficient is reported. Concurrent-policy indicators are examined separately, and these exercises are described only as supplementary sensitivity checks.

Location in revised manuscript: 4.4.1 Group-Time Treatment Effects and TWFE Benchmark; paragraph beginning “Selection adjustment is used only as a supplementary check. Matching…” [red]; 5.3.2 Concurrent Policies; paragraph beginning “Note: Each policy is added separately. Indicators are rebuilt from…” [red]; 5.3.2 Concurrent Policies; paragraph beginning “As shown in Table 7, adding the low-carbon city, Broadband…” [red]; 5.3.3 Selection Adjustment Using Pre-Policy Covariates; paragraph beginning “Note: Matching uses city means of the six controls over…” [red]; Table 8 reports the selection-adjusted TWFE estimate after city-level nearest-neighbor matching on pre-policy covariates. The estimated SCP coefficient is −0.081, with a standard error of 0.108 and a p-value of 0.458, so the matched estimate remains statistically imprecise. [red]; 5.3.3 Selection Adjustment Using Pre-Policy Covariates; paragraph beginning “The matching procedure underlying Table 8 improves balance on observed…” [red]; 6.2 Identification and Concurrent Policies; paragraph beginning “Controls for concurrent policies and matching based on pre-policy characteristics…” [red].

Comment 9

The Random Forest results demonstrate predictive performance, not causal effects. More importantly, the manuscript should clearly explain how training, validation, and test observations were divided.

Response: Thank you for this important clarification. We reran the machine-learning analysis so that complete cities, rather than individual city-year observations, are assigned to training or testing. With seed 42, 32 cities (320 observations) are used for training and eight different cities (80 observations) are held out for testing. Five-fold GroupKFold and RFECV are confined to the training cities, so no held-out city contributes to model fitting or feature selection. Random Forest achieves a held-out R² of 0.568, with an RMSE of 0.400 and an MAE of 0.328. The OLS benchmark evaluated on the same held-out cities gives R² = 0.767, RMSE = 0.294, and MAE = 0.251. The mean post-selection cross-validated R² for Random Forest is −0.040, with a standard deviation of 1.075, indicating substantial variation across the training-city folds. The seed-42 RFECV retains seven predictors—DII and the six benchmark controls—while SCP and EI are not selected. Figure 6 reports prediction sensitivity across five city splits, Figure 7 reports the RFECV profile, and Table 13/Figure 8 report held-out SHAP contributions. The revised manuscript treats these quantities strictly as predictive-performance and feature-contribution measures rather than causal estimates, policy channels, or mediation evidence. Post-selection fold scores are treated as internal diagnostics rather than independent nested-validation estimates.

Thank you again for your careful review and for the thoughtful and rigorous comments you provided. We are sincerely grateful for the time, attention, and expertise you devoted to evaluating our manuscript. Your suggestions prompted us to reconsider several important aspects of the empirical design and to present the remaining limitations more clearly and cautiously.We have tried our best to address each of your concerns and to revise the manuscript in a more transparent and methodologically careful manner. We recognize that some limitations cannot be fully resolved with the available data and records, and we have therefore stated them explicitly rather than overstating the strength of the evidence. We deeply appreciate your guidance, which has helped us improve both the rigor of the analysis and the clarity of the manuscript.We sincerely hope that the revisions have responded adequately to your concerns. We would be very grateful if you could consider the revised manuscript favorably, and we respectfully hope that it is now suitable for publication.

Sincerely,

The Authors

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

now Ok

Reviewer 2 Report

Comments and Suggestions for Authors

Accept in present form.

Reviewer 3 Report

Comments and Suggestions for Authors

I think that the authors have adequately addressed all my concerns in the
revised version of the manuscript. Therefore, I have no further comments.