The Green Paradox in Resource-Dependent Regions: Heterogeneous Impacts of Environmental Regulation on Carbon Lock-in Subtitle: Evidence from Double Machine Learning in China's Yellow River Basin

The Green Paradox in Resource-Dependent Regions: Heterogeneous Impacts of Environmental Regulation on Carbon Lock-in Subtitle: Evidence from Double Machine Learning in China's Yellow River Basin | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article The Green Paradox in Resource-Dependent Regions: Heterogeneous Impacts of Environmental Regulation on Carbon Lock-in Subtitle: Evidence from Double Machine Learning in China's Yellow River Basin Yanying Wang, Xianzhi Wang, kang Pian This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8517802/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Resource-based cities in the Global South universally face a "dual dilemma" of sustaining economic growth while breaking path-dependent carbon lock-in. As a critical energy basin in China, the low-carbon transition of the Yellow River Basin is pivotal for achieving national carbon neutrality goals. However, debate persists regarding whether the "Yellow River Basin Ecological Protection and High-Quality Development Strategy" (implemented in 2019) effectively disrupts this high-carbon dependency. Existing studies often rely on linear models, ignoring the high-dimensional non-linear confounders typical of complex eco-economic systems, potentially masking the true "Green Paradox" effects. Treating this strategy as a quasi-natural experiment, this study employs a Double Machine Learning (DML) causal inference framework on a long-term panel dataset (2000–2023) covering 938 counties. This approach utilizes orthogonalization techniques to effectively eliminate the interference of 75 high-dimensional confounding variables. The findings reveal: (1) Aggregate Effect: The strategy significantly reduced carbon intensity on average (ATE = -0.058, p<0.01). (2) Green Paradox Confirmation: Dynamic analysis reveals that emissions did not decline immediately; instead, a short-term rebound occurred initially (t=0, t=1), confirming the existence of "intertemporal arbitrage" behaviors. (3) Asymmetric Mechanism: Decomposition analysis indicates that emission reductions were primarily "Efficiency-driven" (diluting intensity via innovation, Coef = -0.290) rather than "Structure-driven". Notably, the industrial structure coefficient showed a short-term resilience (Coef = +0.032), suggesting that high-carbon industries were not "washed out" but rather locked in due to structural rigidity. Conclusion: While top-down environmental strategies act as an effective "brake" for expanding economies, declining regions require complementary "Just Transition" mechanisms to avoid falling into a low-carbon poverty trap. Biological sciences/Ecology Earth and environmental sciences/Ecology Earth and environmental sciences/Environmental sciences Earth and environmental sciences/Environmental social sciences Yellow River Strategy Carbon Lock-in Green Paradox Just Transition Double Machine Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1 Introduction Global climate change has compelled nations to seek a fundamental "decoupling" of economic growth from carbon emissions. As the world's largest carbon emitter, China has pledged to peak carbon emissions before 2030 and achieve carbon neutrality by 2060. Under this grand vision, the low-carbon transition of Resource-based Cities (RBCs) constitutes one of the most challenging issues. Unruh 1 proposed the theory of "Carbon Lock-in," stating that industrialized economies tend to lock themselves into fossil-fuel-based energy systems through the co-evolution of technological institutions, infrastructure, and behavioral norms. For RBCs, this lock-in effect is particularly acute: high-energy-consuming industrial structures, high-carbon energy consumption patterns, and high exit barriers caused by asset specificity make their transition not merely a technological iteration, but a profound socioeconomic restructuring 2–4 . Although environmental regulation is regarded as a primary policy tool to break this lock-in, intense theoretical debate remains regarding its real-world efficacy in RBCs. Proponents, primarily based on the "Porter Hypothesis," argue that appropriate environmental regulation can generate an "innovation compensation effect," compelling firms to undertake technological innovation and process optimization, thereby offsetting compliance costs and achieving a win-win for the environment and the economy 5 . Recent causal evidence from developed economies has largely supported this view, showing that market-based instruments can effectively induce green patents 6 . However, opponents or skeptics raise the warning of the "Green Paradox." Sinn (2008) argued from a supply-side perspective that if resource owners anticipate increasingly stringent future climate policies (e.g., high carbon taxes), they may form "Doomsday Expectations" and choose to accelerate the extraction of fossil fuels to monetize them before asset devaluation. This "intertemporal arbitrage" behavior may cause carbon emissions to rise rather than fall in the short term 7,8 . This phenomenon has been empirically observed in cross-country studies 9 . and specific contexts like carbon tax announcements in emerging economies. Particularly in China, Wu et al. (2023) found that environmental federalism might exacerbate this paradox due to local governments' "race to the bottom" to protect tax bases 10 . Particularly for declining cities at the end of their lifecycle, lacking funds for technological retrofitting, strong regulations may instead induce firms to maintain high-emission production until market exit 11 . These two contrasting theoretical predictions imply that policy responses in RBCs may exhibit a high degree of heterogeneity. However, existing literature attempting to clarify these controversies is largely constrained by methodological limitations, leading to empirical results distorted by "Model Misspecification Bias." Traditional econometric models (such as OLS or standard DID) typically rely on strict parametric assumptions, assuming a linear relationship between control variables and outcome variables. Yet, in complex urban eco-economic systems, carbon emissions are influenced non-linearly by a massive number of high-dimensional factors such as economy, population, technology, and fiscal capacity, with complex interactions among variables. Varian (2014) pointed out that non-linear relationships are ubiquitous among economic variables, which traditional models struggle to capture 12 . Chernozhukov (2018) further demonstrated that ignoring these non-linear features leads to serious "Regularization Bias," thereby masking the true net effect of policies 13 . Although some scholars have begun attempting non-linear methods 14,15 . research combining high-dimensional machine learning for causal inference in the specific field of RBC transition remains relatively scarce. To clarify the aforementioned theoretical controversies, this study utilizes China's "Yellow River Basin Ecological Protection and High-Quality Development" strategy (implemented in 2019, hereinafter "Yellow River Strategy") as an ideal Quasi-natural Experiment 16 . Unlike previous fragmented policies, this strategy implemented a combination of "hard constraints" (ecological redlines, water-based production limits) and "soft incentives" (high-quality development subsidies) within a specific geo-economic region, providing an excellent window for identifying the causal effects of regional environmental policies 17 . This study introduces the frontier Double Machine Learning (DML) framework, combining the power of Random Forest algorithms in handling high-dimensional non-linear data with the causal inference logic of econometrics. Through Orthogonalization, we utilize long-term panel data from 938 counties across 9 provinces in the Yellow River Basin to precisely isolate the policy's "net effect" from high-dimensional noise. The marginal contributions of this study are primarily reflected in three aspects: First, Methodological Correction. We confirm that traditional linear models, unable to handle high-dimensional non-linear confounding variables, significantly underestimate the policy effect by approximately 40%, whereas DML provides a more robust estimation of emission reductions by mitigating bias from confounding variables. Second, Mechanism Identification. We open the "black box" of policy action, quantifying the dual driving paths of "technological greening" (forcing green patent applications) and "structural cleaning" (washing out high-energy-consuming industries). Third, Revealing the "Just Transition" Dilemma. By focusing on heterogeneity analysis of 384 counties in the core basin, we discover significant lifecycle differences: while regenerating cities benefit significantly, declining cities show signs of a "Green Paradox." This finding provides vivid Chinese experience for the global "Just Transition" discourse 18,19 , warning that singular strong regulations without complementary support policies may exacerbate regional inequality. 2 Institutional Background and Theoretical Hypotheses 2.1 Institutional Background: From Fragmented Governance to Systemic Regulation The Yellow River Basin spans China's three major geographic steps (East, Central, West). It is not only the cradle of Chinese civilization but also the core "Energy Basin" of contemporary China. This region holds more than half of the country's basic coal reserves and hosts clusters of high-density RBCs such as Ordos and Yulin, making it a pivotal area for national energy security 20 . However, long-term extensive development has led to the coexistence of a fragile ecological environment and high-intensity carbon emissions. In 2019, the Chinese central government elevated "Yellow River Basin Ecological Protection and High-Quality Development" to a major national strategy. Unlike previous environmental policies based on administrative divisions, this strategy possesses significant "systemic" and "strong constraint" characteristics, providing an ideal quasi-natural experimental setting for this study 21 . Specifically, the strategy introduced a comprehensive "carrot and stick" mechanism. On the one hand, regarding "hard constraints" (the stick), the strategy established rigid principles of "determining city, land, people, and production by water" (Yi Shui Ding Chan). For water-intensive and energy-intensive resource industries, this implies insurmountable physical boundaries for capacity expansion and stricter ecological redlines. On the other hand, complementing these constraints are "soft incentives" (the carrot). The central government established special subsidy funds to guide cities to transition from pure resource extraction to green manufacturing and ecological tourism, fostering high-quality development 16 . Theoretically, this combination will have profound causal impacts on the carbon emissions of RBCs by altering both compliance costs and development incentives. 2.2 Theoretical Analysis and Research Hypotheses 2.2.1 Regulation Intensity, Official Incentives, and Net Reduction Effects Whether environmental regulation is effective depends on whether it can alter the behavioral incentives of local governments and firms. Under China's unique "Promotion Tournament" system, GDP growth has long been the core indicator for local official promotion, leading local governments to often adopt "protective umbrella" strategies for high-polluting major taxpayers 22 . However, the uniqueness of the Yellow River Strategy lies in its extremely high political hierarchy, placing "ecological protection" before "high-quality development," substantively reshaping the performance evaluation system of provinces along the river 16 . On one hand, the delineation of ecological redlines increases the Compliance Costs for high-carbon firms, forcing them to cut fossil energy inputs under profit maximization constraints. On the other hand, the strengthened political accountability mechanism breaks local protectionism, forcing local governments to transform from "pollution colluders" to "environmental regulators 23 . " This superposition of vertical administrative pressure from the central government and horizontal cost pressure on firms theoretically forms a powerful combined force for emission reduction. Hypothesis 1 (H1): As a strong constraint environmental regulation, the Yellow River Strategy significantly reduces the carbon emission intensity of RBCs by reshaping local officials' promotion incentives and increasing firms' environmental compliance costs. 2.2.2 Transmission Mechanisms: "Innovation Compensation" and "Creative Destruction" Environmental regulation is not simply about "shutting down and transferring"; its deep emission reduction momentum stems from technological progress and structural change. First, according to the "Innovation Compensation Effect" of the Porter Hypothesis, moderate environmental pressure forces firms to reflect on technology. Facing the "hard constraints" set by the Yellow River Strategy, surviving resource-based firms must increase R&D investment to survive in long-term competition, offsetting rising compliance costs by applying for green patents and updating clean equipment. This "Technology Effect" induced by regulation is an endogenous driver for reducing carbon intensity 5,24 . Second, based on Schumpeter's "Creative Destruction" theory, high-standard environmental regulation functions as a market filter. The "determining production by water" policy effectively sets a threshold for industrial entry. High-pollution, low-efficiency backward production capacities (usually low-end manufacturing) are forced to exit the market due to the inability to bear high pollution control costs; meanwhile, the land, credit, and labor factors released flow toward low-energy service industries or high-tech industries. This "Structural Effect" (or Cleaning Effect) of industrial composition will directly drive the urban economy toward low-carbon transition 25 . Hypothesis 2 (H2): The Yellow River Strategy drives the decline of carbon intensity through a dual-wheel mechanism: stimulating firms' green technological innovation (Technology Effect) and accelerating the exit and substitution of high-energy-consuming industries (Structural Effect). 2.2.3 Heterogeneity Perspective: The "Green Paradox" in Declining Cities RBCs are not a homogeneous whole. For cities in the growth or regeneration stage, with ample fiscal resources and high industrial diversification, it is easier to respond to policies and transition. However, for resource-depleted declining cities, the situation may be vastly different. Sinn's (2008) Green Paradox points out that if the supply side expects future environmental policies to be extremely strict (e.g., forced shutdowns), resource owners will tend to accelerate extraction before the policy fully lands to avoid asset stranding 7 . Furthermore, declining cities often face severe fiscal deficits and brain drain, lacking funds for green technological upgrades. Under strong regulatory pressure, firms in these cities may choose "rush-style" production to overdraft remaining resources, leading to a rise rather than a fall in carbon emissions 26 . Hypothesis 3 (H3): Policy effects exhibit significant lifecycle heterogeneity; compared to regenerating cities, declining cities may experience a short-term rebound in carbon emissions—a "Green Paradox" phenomenon—due to a lack of financial support and strong "doomsday expectations." 3 Data and Empirical Strategy 3.1 Sample Construction and Study Area The primary challenge of empirical research lies in constructing a sample system that satisfies causal inference assumptions while accurately reflecting policy targeting characteristics. Based on this, we designed a dual-layer sample structure combining "Full Domain Identification" and "Core Focus" (as shown in Fig. 2). Specifically, in the causal identification stage, to construct a sufficiently robust counterfactual framework, we set the sample range to 938 counties across the entire 9 provinces of the Yellow River Basin (2000–2023). Although this range covers some peripheral areas geographically distant from the main Yellow River stream, retaining these non-core samples has significant econometric value: they serve as effective intra-provincial control groups, helping the DML model eliminate macro shocks at the provincial level and providing rich high-dimensional feature variation to improve the machine learning algorithm's prediction accuracy for potential counterfactual outcomes. In contrast, when analyzing spatial heterogeneity, to isolate spatial background noise and intuitively demonstrate the micro-transmission mechanism, we focus the analytical view on the 384 core counties within the physical watershed of the Yellow River. These counties adjacent to the main streams and tributaries are the direct fields of action for "determining production by water" and "ecological redlines," allowing for a clearer depiction of the spatial implementation of the strategy. 3.2 Variable Measurement and Data Description 3.2.1 Core Variable Measurement The core dependent variable ( \(\:{Y}_{it}\) ) is Carbon Emission Intensity, defined as carbon dioxide emissions per unit of GDP (tons/10,000 yuan). Data are extracted from the high-precision China Emission Accounts and Datasets (CEADs) and its accompanying micro-inversion dataset 27 . Compared to traditional aggregate data based on energy statistical yearbooks, CEADs data integrates night-time light remote sensing inversion technology, accurately capturing minute carbon emission dynamics at the county level and effectively overcoming the "data masking" problem of traditional statistics. The core independent variable ( \(\:{D}_{it}\) ) is set as a typical multi-period policy shock interaction term ( \(\:{Treat}_{i}\times\:{Post}_{i}\) ). Specifically, we define the group dummy variable \(\:{Treat}_{i}\) : if county \(\:\:i\:\) is located in the 9 provinces of the Yellow River Basin and within the policy implementation scope, then \(\:{Treat}_{i}=1\) , otherwise 0; simultaneously, define the time dummy variable \(\:{Post}_{i}\) : if time \(\:t\:\) is 2019 (the year of official strategy implementation) or later, \(\:{Post}_{i}=1\) , otherwise 0. \(\:{\:D}_{it}\) is the product of the two, used to capture the net difference in treated counties before and after implementation. 3.2.2 Mechanism Variables and Covariates To open the "black box" of policy effects, we introduce two key mechanism variables: Green Technological Innovation ( \(\:{Mech\_Tech}_{it}\) ) and Industrial Structure Optimization ( \(\:{Mech\_Str}_{it}\) ). The former is measured by the total number of green patent applications per 10,000 people (logarithm), sourced from the China National Intellectual Property Administration (CNIPA), directly reflecting firms' "innovation compensation" behavior; the latter is represented by the proportion of secondary industry value-added in GDP, capturing the "cleaning effect" of high-energy industry exit. Additionally, considering the systemic complexity of RBC transition, we rely on the DML algorithm to introduce a high-dimensional covariate set ( \(\:{X}_{it}\) ) covering four dimensions: economic basis, industrial characteristics, population/society, and fiscal capacity, to maximally block confounding paths. Descriptive statistics are detailed in Table 1 . Table 1 Descriptive Statistics of Main Variables Type Variable Definition Mean Std. Dev. Min Max Dependent Carbon Intensity CO2 Emissions / GDP 21980 3.068 4.863 0.094 30.325 Independent Policy Shock ( \(\:{D}_{it}\) ) Interaction term 21980 0.061 0.239 0 1 Mechanism Industrial Structure Secondary Ind. / GDP 21980 0.403 0.164 0.076 0.816 Green Performance SO2 Emissions 21980 8603.3 11760.0 39.0 73365.6 Control Ln PGDP Log(Per capita GDP) 21980 9.967 1.102 7.497 12.408 Ln Fixed Asset Log(Fixed Asset Investment) 21980 12.758 2.005 7.960 16.450 Ln FDI Log(FDI + 1) 21980 4.093 4.568 0 14.281 Ln Population Log(Population) 21980 4.316 0.763 1.946 6.043 Urbanization Urban Emp. Ratio 21980 0.316 0.228 0.015 0.999 Fiscal Revenue Log(Revenue) 21980 10.749 1.341 7.313 14.152 Fiscal Expend. Log(Expenditure) 21980 11.455 1.189 8.448 14.619 (Note: Data covers 938 counties from 2000 to 2023. All continuous variables are winsorized at the 1st and 99th percentiles to mitigate the impact of outliers.) 3.3 Econometric Model: Double Machine Learning (DML) When evaluating environmental policy effects, the commonly used Two-Way Fixed Effects (TWFE) model in existing literature relies on strict linear assumptions. However, as Varian (2014) emphasized, complex economic systems are replete with non-linear relationships 12 . If the true data generation process is non-linear (e.g., the impact of industrial structure on carbon emissions follows an inverted U-shape with development stages), forcing a linear model leads to severe "Model Misspecification Bias." Chernozhukov (2018) further proved that with high-dimensional control variables, this bias cannot be eliminated simply by increasing sample size, i.e., "Regularization Bias" exists 13 . To fundamentally overcome this limitation, this study adopts the Double Machine Learning (DML) framework for causal identification (as shown in Fig. 3). We specify the following partially linear regression model: $$\:{Y}_{it}=\theta\:{D}_{it}+g\left({X}_{it}\right)+{\zeta\:}_{it},\:\:\:\:\:\:E\left[{\zeta\:}_{it}|{D}_{it},{X}_{it}\right]=0$$ 1 $$\:{D}_{it}=m\left({X}_{it}\right)+{v}_{it},\:\:\:\:\:\:E\left[{v}_{it}|{X}_{it}\right]=0$$ Where \(\:\theta\:\) is the policy net effect of interest; unlike traditional methods, DML introduces two nuisance functions \(\:g\left({X}_{it}\right)\) and \(\:m\left({X}_{it}\right)\) to capture the complex non-linear impacts of control variables on carbon emissions and policy probability, respectively. We use the Random Forest algorithm to flexibly fit these two functions and employ Orthogonalization techniques to cut off the transmission of machine learning prediction errors to the causal parameter \(\:\theta\:\) . Furthermore, to ensure the robustness of statistical inference, we strictly perform 5-fold Cross-Fitting during the estimation process 28 – 30 . Figure 3 Flow Chart of Double Machine Learning (DML) 3.4 Spatial Causal Identification: Based on Individual Treatment Effect (ITE) 3.4.1The traditional econometric parameter \(\:\theta\:\) can only identify the Average Treatment Effect (ATE), often masking the Spatial Non-stationarity of policies across regions with different resource endowments. To open the spatial black box, we introduce Explainable AI concepts, constructing a Geo-SHAP framework to estimate the Individual Treatment Effect (ITE). Estimation of ITE: Specifically, for any county \(\:\:i\) , its policy effect proxy value \(\:{\tau\:}_{i}\) is defined as the difference in predicted carbon intensity under the factual scenario (T = 1, strategy implemented) and counterfactual scenario (T = 0, not implemented). Based on the trained DML Random Forest model, we calculate this difference using the following formula: $$\:{\tau\:}_{i}\left(x\right)=E\left[{Y}_{i}\right|\left({D}_{i}=1,{X}_{i}=x\right)]-E[{Y}_{i}|\left({D}_{i}=0,{X}_{i}=x\right)$$ 2 This metric is logically equivalent to the core idea of Geo-SHAP, measuring the marginal contribution of a specific feature (here, the policy variable) to the prediction result. 3.4.2 Spatial Mapping and Pattern Identification By mapping the estimated \(\:{\tau\:}_{i}\) to geographic coordinates, we can intuitively identify the "Emission Reduction Highlands" ( \(\:{\tau\:}_{i}\ll\:0\) ) and "Resistance Depressions" ( \(\:{\tau\:}_{i}\approx\:0\:or>0\) ) of the Yellow River Strategy. This step lowers the analysis granularity from the domain-wide average to the micro-county level, allowing us to break through the limitations of traditional regression, thereby providing intuitive evidence for revealing the spatial distribution of the "Green Paradox" in clusters of declining cities. 4 Empirical Results 4.1 Benchmark Regression: Methodological Correction and Net Effect Identification To verify the net emission reduction effect (Hypothesis H1) and quantify DML's bias correction capability in handling high-dimensional non-linear confounding, we adopted a stepwise regression strategy. Table 2 reports estimation results based on three model specifications. Column (1) is OLS estimation controlling only for simple fixed effects; Column (2) is the field's standard Two-Way Fixed Effects (TWFE) model; Column (3) is the Double Machine Learning (DML) model used in this study. Table 2 The Impact of Yellow River Strategy on Carbon Intensity (1) OLS (2) TWFE (3) DML (Preferred) Policy Shock ( \(\:{\varvec{D}}_{\varvec{i}\varvec{t}}\) ) ) -0.019* -0.035** -0.058*** -0.011 -0.016 -0.012 Controls Linear Linear Non-Linear (Random Forest) FE (City & Year) No Yes Absorbed via Cross-fitting Bias Correction No No Yes Observations 21980 21980 21980 (Note: Robust standard errors clustered at the county level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. DML estimates are based on 5-fold cross-fitting with random forest learners to prevent overfitting.) The empirical results reveal two key findings of methodological significance: First, the robustness of the policy effect. Regardless of model specification, the policy interaction term coefficient \(\:\theta\:\) is consistently negative and significant. Specifically, the DML model (Column 3) estimates a net reduction effect of -0.058 (p < 0.01).This provides solid empirical evidence that the strategy, as a strong environmental regulation, effectively reduced carbon intensity in RBCs through "hard constraints" and "soft incentives," strongly supporting Hypothesis H1. Second, the "Attenuation Bias" of traditional models. This is a phenomenon long ignored by existing literature. Comparing Column (2) and Column (3), the reduction effect estimated by the traditional TWFE model is -0.035, while the "net effect" estimated by the DML model significantly increases to -0.058. This implies that ignoring non-linear features of the economic system leads traditional linear models to underestimate the policy effect by approximately 40%. This discrepancy profoundly reveals the insidiousness of "Regularization Bias": since carbon emissions in RBCs are subject to non-linear interference from high-dimensional factors (e.g., resource endowments, historical accumulation), forcing linear forms leads to model under-fitting, causing explanatory power to "leak" into residuals and shrinking the policy coefficient toward zero. DML successfully isolates this non-linear noise through orthogonalization, restoring the stronger real reduction capability of the Yellow River Strategy 13 . 4.2 Dynamic Effects and "Green Paradox" Verification Benchmark regression only identified the Average Treatment Effect (ATE). To examine the time lag of policy efficacy and potential short-term gaming behaviors, we combined the Event Study method with the DML framework to plot the dynamic evolution path of policy effects(shown as Fig. 4). First, Pre-treatment Parallel Trends were fully verified. Before policy implementation (t < 0), all estimated coefficient confidence intervals included 0, proving that treated and control groups shared consistent evolution paths prior to the strategy. However, the most striking finding occurred in the early stages of implementation. Crucially, at t = 0 and t = 1, the coefficients did not immediately turn negative but instead showed a distinct positive jump. This seemingly counter-intuitive phenomenon confirms Sinn's (2008) "Green Paradox" hypothesis at the micro-level. Facing imminent "Three Lines One Permit" hard constraints, some highly resource-dependent cities formed strong "Doomsday Expectations." To monetize resources before policy tightening, firms tended to engage in "Intertemporal Arbitrage" style rush production, causing short-term emissions to rise 7 . From the third year (t = 3), policy coefficients rapidly fell and turned significantly negative. This dynamic path from "short-term pain" to "long-term substantial reduction" indicates that the Yellow River Strategy eventually overcame initial adjustment costs through long-term mechanisms (e.g., green innovation), achieving true low-carbon transition. 4.3 Robustness Checks To further exclude drivers from omitted variables or random noise, we conducted a rigorous Placebo Test. We randomly generated false treatment groups and policy times 500 times. Shown as Fig. 5, The kernel density distribution shows that the 500 random coefficients follow a standard normal distribution centered at 0. In contrast, the true benchmark coefficient (-0.058) lies significantly at the left tail, far from the placebo confidence interval. This indicates the observed reduction effect is not a random coincidence (p < 0.01). 4.4 Lifecycle Heterogeneity: A "Just Transition" Perspective Based on the "National Sustainable Development Plan for RBCs," we classified the sample into Regenerating, Growing, Mature, and Declining categories. Results show significant inter-group heterogeneity ("Matthew Effect")(Figure 6). Regenerating Cities showed the strongest reduction effect (Coef ≈ -0.07). Growing Cities also showed significant reductions. In contrast, Mature Cities showed insignificant coefficients, reflecting strong "Carbon Lock-in." Most critically, Declining Cities showed coefficients converging toward 0 or even turning positive. This confirms the "Green Paradox" 7 : under strict regulation and fiscal exhaustion, these cities lack funds for transition and may accelerate backward capacity operation to survive. This warns that for declining cities, singular "hard constraints" may fail, urgently requiring complementary "soft support" for Just Transition. 5 Mechanism Analysis and Heterogeneity Discussion 5.1 Mechanism Verification: Technology Effect vs. Structural Effect To verify the transmission paths proposed in Hypothesis H2, we estimated the policy's impact on mechanism variables using the DML framework. The regression results, reported in Table 3 , reveal a striking "Asymmetric Transition" pattern characterized by significant technological improvements alongside structural rigidity. Table 3 Mechanism Analysis Regression Results (1) Green Performance (2) Industrial Structure Dependent Variable \(\:-\text{l}\text{n}({SO}_{2}+1)\) Second. Industry Share Policy Shock ( \(\:{\varvec{D}}_{\varvec{i}\varvec{t}}\) ) -0.290 *** 0.032 *** -0.06 -0.006 Controls Yes Yes Year FE Yes Yes Cluster SE County County Observations 21980 21980 R-squared 0.048 0.321 Note: Robust standard errors clustered at the county level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions include the full set of controls used in the benchmark model. First, the impact of the Yellow River Strategy on Green Performance (proxied by \(\:-\text{l}\text{n}({SO}_{2}+1)\) is significantly negative (Coef = -0.290, p < 0.01). This indicates a substantial reduction in pollution intensity via technological upgrades and end-of-pipe treatments. Facing strict environmental constraints, firms have actively adopted cleaner technologies to offset compliance costs, a phenomenon consistent with the "Innovation Compensation" effect of the Porter Hypothesis. This technological "efficiency gain" appears to be the primary driver of the observed aggregate carbon reduction (ATE = -0.058). Second, and crucially, the impact on Industrial Structure (Secondary Industry Share) is significantly positive (Coef = + 0.032, p < 0.01). Contrary to the expectation that regulation would immediately "wash out" high-polluting industries (which would yield a negative coefficient), this positive coefficient suggests a "Structural Stickiness" or even a short-term "Structure Reflow." The share of the secondary industry did not decline but rather increased slightly relative to the control group. This counter-intuitive finding provides micro-level empirical evidence for Sinn’s (2008) "Green Paradox." Facing the imminent "hard constraints" of the Yellow River Strategy (e.g., water-based production limits), resource-dependent firms—anticipating future capacity caps—likely formed "Doomsday Expectations." Consequently, they engaged in "Intertemporal Arbitrage" by accelerating extraction and production to monetize fossil fuel assets before the full enforcement of regulations. This "Rush-to-Produce" behavior temporarily inflated the share of industrial output. Combining these findings, we conclude that the current low-carbon transition in the Yellow River Basin is "Efficiency-driven" rather than "Structure-driven." The policy has successfully acted as a "Technological Accelerator" but has failed to function as a "Structural Filter" in the short term. This asymmetry highlights a critical vulnerability: without breaking the "Carbon Lock-in" of the industrial structure, the emission reductions achieved solely through technical fixes may face diminishing returns in the long run. 5.2 Heterogeneity Analysis: Spatial Micro-pattern via Geo-SHAP Mapping ITEs of 384 core counties to geographic space reveals a pattern of "Effective in Upstream Ecological Barriers, Obstructed in Midstream Energy-Rich Areas," as visualized in Fig. 7. The spatial distribution of Individual Treatment Effects (ITE) visualized in Fig. 7 exhibits a distinct "Core-Periphery" gradient. The "Emission Reduction Highlands" (ITE \(\:\:\ll\:\:\) 0) are predominantly located in the upstream ecological function zones (e.g., Sanjiangyuan in Qinghai) and the downstream advanced manufacturing clusters (e.g., Jinan and Qingdao in Shandong). These regions benefit either from strict ecological compensation or from advanced technology spillover. Conversely, the "Resistance Depressions" (ITE ≈ 0 or > 0) highly overlap with the "Energy Golden Triangle" (Ordos, Yulin, Ningxia), which constitutes the core of China's coal bases. The resource curse in these areas is intensified by the high asset specificity of coal mining infrastructure, making immediate transition prohibitively expensive. This spatial evidence mutually corroborates the lifecycle regression results, highlighting the necessity of spatially targeted "Just Transition" policies. 5.3 Heterogeneity by Resource Type: Coal vs. Oil/Metal Different resource endowments imply varying carbon intensities and lock-in degrees. To further explore the source of resistance, we classified the sample into Coal-based, Oil/Gas-based, and Metal-based cities based on their dominant mineral resources. The heterogeneity of policy effects is visualized in Fig. 8. As illustrated in Fig. 8, the estimated coefficients vary significantly across resource categories. The Coal-based cities sub-sample exhibits coefficients that are largely positive or statistically insignificant in the short term, visually confirming the "Green Paradox" effect discussed earlier. This pattern suggests that coal industries, facing the strictest "de-capacity" pressures (e.g., coal consumption caps), reacted with stronger "rush-style" production incentives. Conversely, the coefficients for Metal-based and Oil/Gas-based cities tend to be negative, indicating a smoother transition trajectory. This visual evidence further corroborates that the transition dilemma in the Yellow River Basin is structurally driven by the "Coal Dilemma." 6 Conclusion and Policy Implications 6.1 Conclusion Using the Yellow River Strategy as a quasi-natural experiment and employing DML and Geo-SHAP, this study draws three primary conclusions. First, from a methodological perspective, we confirm that traditional linear models underestimate policy effects by approximately 40%, whereas the DML framework successfully uncovers the true net reduction effect masked by high-dimensional confounders. Second, the mechanism analysis reveals that current emission reductions are primarily "Denominator Driven" (efficiency gains) rather than "Numerator Driven" (structural cleaning), which explains the observed short-term rebounds. Third, a significant "Just Transition Dilemma" exists: while regenerating cities benefit from the strategy, declining cities face a "Green Paradox," a phenomenon that is spatially concentrated in the midstream Energy Golden Triangle. 6.2 Policy Implications Based on these findings, we propose the following policy recommendations. First, to address the regional inequality exposed by the "Green Paradox," a central "Just Transition Fund" should be established. This fund should prioritize worker reskilling and provide seed funding for alternative industries, such as photovoltaics and tourism, to prevent declining cities from falling into low-carbon poverty traps. Second, governance mechanisms must shift from a uniform approach to differentiated governance. Implementing horizontal ecological compensation is crucial; downstream beneficiaries should purchase "emission quotas" from midstream energy cities to internalize environmental externalities. Finally, to achieve deep decarbonization, policy focus must shift from "End-of-pipe" governance to "Source" adjustment. By raising entry thresholds and leveraging green finance, the transition can move from being merely "Denominator Driven" to "Numerator Driven," ensuring long-term sustainability. 6.3 Limitations and Future Research While this study offers rigorous empirical evidence, it is not without limitations. First, restricted by data availability, the carbon emission data derived from night-time lights, though high-precision, may still contain measurement errors for certain non-point sources. Future research could benefit from integrating enterprise-level micro-energy consumption data to validate these findings. Second, this study primarily focuses on the internal policy effects within the Yellow River Basin. However, environmental regulations often generate spatial spillover effects to adjacent regions (e.g., pollution haven hypothesis). Future studies should employ spatial DML models to explore whether the Yellow River Strategy has led to carbon leakage to non-policy zones, thereby providing a more comprehensive assessment of the national strategy. Declarations Data Availability The datasets generated and analyzed during the current study are based on publicly available data. The carbon emission data are derived from the China Emission Accounts and Datasets (CEADs, https://www.ceads.net). Other socio-economic data are available from the China County Statistical Yearbook and provincial statistical yearbooks. . The processed data supporting the findings of this study are available from the corresponding author upon reasonable request. Competing Interests The authors declare no competing interests. Ethical Approval This study does not involve human participants or animal subjects; therefore, ethical approval is not applicable. Informed Consent This study does not involve human participants; therefore, informed consent is not applicable. Author Contributions Yanying Wang: Conceptualization, Methodology, Software, Writing – original draft. Xianzhi Wang: Data curation, Visualization, Investigation. Kang Pian: Supervision, Writing – review & editing. Funding This research received no external funding. References Unruh, G. C. Understanding carbon lock-in. Energy Policy 28 , 817–830 (2000). Johnstone, P. & Hielscher, S. Phasing out coal, sustaining coal communities? Living with technological decline in sustainability pathways. Extr. Ind. Soc. 4 , 457–461 (2017). Moses, N. D. & MacCarty, N. A. What makes a cookstove usable? Trials of a usability testing protocol in Uganda, Guatemala, and the United States. Energy Res. Soc. Sci. 52 , 221–235 (2019). Jakob, M. et al. The future of coal in a carbon-constrained climate. Nat. Clim. Change 10 , 704–707 (2020). Porter, M. E. & Linde, C. V. D. Toward a New Conception of the Environment-Competitiveness Relationship. J. Econ. Perspect. 9 , 97–118 (1995). Knaus, M. C. Double machine learning-based programme evaluation under unconfoundedness. Econom. J. 25 , 602–627 (2022). Sinn, H.-W. Public policies against global warming: a supply side approach. Int. Tax Public Finance 15 , 360–394 (2008). Six, M. & Wirl, F. Optimal pollution management when discount rates are endogenous. Resour. Energy Econ. 42 , 53–70 (2015). Yamamoto, Y. Living under ecosystem degradation: Evidence from the mangrove–fishery linkage in Indonesia. J. Environ. Econ. Manag. 118 , 102788 (2023). Liu, B. et al. Regional differences in China’s electric vehicle sales forecasting: Under supply-demand policy scenarios. Energy Policy 177 , 113554 (2023). Ploeg, F. V. D. Natural Resources: Curse or Blessing? J. Econ. Lit. 49 , 366–420 (2011). Varian, H. R. Big Data: New Tricks for Econometrics. J. Econ. Perspect. 28 , 3–28 (2014). Chernozhukov, V. et al. Double/debiased machine learning for treatment and structural parameters. Econom. J. 21 , C1–C68 (2018). Cheng, C. C. J. & Shiu, E. C. A two-level, longitudinal investigation into the effects of employee social entrepreneurship orientation and top management team decisions on product innovation. Technol. Forecast. Soc. Change 182 , 121832 (2022). Van Es, M. & Bruins, B. Pro-poor change in the aftermath of disasters – Exploring possibilities at the intersection of disaster politics and land rights issues in Central Philippines. Land Use Policy 132 , 106771 (2023). Wang, H.-Y., Li, Y., Jiao, S.-Q., Chou, K.-C. & Zhang, G.-H. Recovery of Ni matte from Ni-bearing electroplating sludge. J. Environ. Manage. 326 , 116744 (2023). Heutel, G., Moreno-Cruz, J. & Shayegh, S. Solar geoengineering, uncertainty, and the price of carbon. J. Environ. Econ. Manag. 87 , 24–41 (2018). Newell, P. & Mulvaney, D. The political economy of the ‘just transition’. Geogr. J. 179 , 132–140 (2013). Carley, S. & Konisky, D. M. The justice and equity implications of the clean energy transition. Nat. Energy 5 , 569–577 (2020). Su, H. & Qi, Z. Polycentric structure and urban thermal environment: A large-scale study from multi-perspectives. Sustain. Cities Soc. 96 , 104657 (2023). Zhou, Z., Zhou, J., Alcalá, J. & Yepes, V. Thermal coupling optimization of bridge environmental impact under natural conditions. Environ. Impact Assess. Rev. 104 , 107316 (2024). Jäger, S., Schoefer, B., Young, S. & Zweimüller, J. Wages and the Value of Nonemployment*. Q. J. Econ. 135 , 1905–1963 (2020). Greenstone, M. & Hanna, R. Environmental Regulations, Air and Water Pollution, and Infant Mortality in India. Am. Econ. Rev. 104 , 3038–3072 (2014). Acemoglu, D., Aghion, P., Bursztyn, L. & Hemous, D. The Environment and Directed Technical Change. Am. Econ. Rev. 102 , 131–166 (2012). Guan, D. et al. Structural decline in China’s CO2 emissions through transitions in industry and energy systems. Nat. Geosci. 11 , 551–555 (2018). Tang, Q., Wang, J., Jing, Z., Yan, Y. & Niu, H. Response of ecological vulnerability to land use change in a resource-based city, China. Resour. Policy 74 , 102324 (2021). Shan, Y. et al. China CO2 emission accounts 1997–2015. Sci. Data 5 , 170201 (2018). Banerjee, A., Chevillon, G. & Kratz, M. Probabilistic forecasting of bubbles and flash crashes. Econom. J. 23 , 297–315 (2020). Belloni, A., Chernozhukov, V. & Hansen, C. High-Dimensional Methods and Inference on Structural and Treatment Effects. J. Econ. Perspect. 28 , 29–50 (2014). Wager, S. & Athey, S. Estimation and Inference of Heterogeneous Treatment Effects using Random Forests. J. Am. Stat. Assoc. 113 , 1228–1242 (2018). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 07 Mar, 2026 Reviewers agreed at journal 07 Mar, 2026 Reviewers agreed at journal 05 Mar, 2026 Reviewers agreed at journal 02 Mar, 2026 Reviewers invited by journal 02 Mar, 2026 Editor invited by journal 11 Feb, 2026 Editor assigned by journal 10 Feb, 2026 Submission checks completed at journal 18 Jan, 2026 First submitted to journal 18 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8517802","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":600427858,"identity":"5adc79b7-90b9-486f-a4a9-ccf239a07e49","order_by":0,"name":"Yanying Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIie3QvWrDMBDAcQmBQkCQ9YzJOwgETofSvoqFoXOm4ClVMMhbuirQh/AjqBjqRXT22FDwlCFjtnzRIUOQPWbQH2457rccQqHQAzYhuDzsc3j/QERdFlj1kagsVGTcE94oPJBw16h4rHNc2aEEtVIJrIGIRuo/hp6nlSXdr09gI1U2/wGauG0pGHoTlaUz7iMEpKrNAljSSh0zVMvKMgo+QkGuivMNCHMlx37C2FdBmAbO4UpsP4HRSmPjIAW31dEnz8SmpomXvNajDu3zZTops2/Y5S/TdVN0XnJTep7Lq8jA+38SCoVCoTudAPIASnNx8meIAAAAAElFTkSuQmCC","orcid":"","institution":"Qilu Normal University","correspondingAuthor":true,"prefix":"","firstName":"Yanying","middleName":"","lastName":"Wang","suffix":""},{"id":600427859,"identity":"a7a6511e-2abe-4ef3-81f9-4025af2d6eff","order_by":1,"name":"Xianzhi Wang","email":"","orcid":"","institution":"Qilu Normal University","correspondingAuthor":false,"prefix":"","firstName":"Xianzhi","middleName":"","lastName":"Wang","suffix":""},{"id":600427861,"identity":"6681a5f0-43c5-433d-a59d-cdc900701877","order_by":2,"name":"kang Pian","email":"","orcid":"","institution":"Qilu Normal University","correspondingAuthor":false,"prefix":"","firstName":"kang","middleName":"","lastName":"Pian","suffix":""}],"badges":[],"createdAt":"2026-01-05 06:54:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8517802/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8517802/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104402370,"identity":"206ae235-38a5-4cc3-9051-5f978217fceb","added_by":"auto","created_at":"2026-03-11 12:15:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1127933,"visible":true,"origin":"","legend":"\u003cp\u003eTheoretical Framework Diagram\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/907994c3fd93c3b01bd9819d.png"},{"id":105903682,"identity":"39a3a37b-0cac-413b-9b07-4cff9abd26cc","added_by":"auto","created_at":"2026-04-01 09:48:06","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":747711,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSpatial Distribution Map of the Study Area Samples\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/bab59ef5d8a7c92ab8e3e6f8.jpg"},{"id":105033570,"identity":"33dade1e-d355-4aab-946b-29cea8a6e76c","added_by":"auto","created_at":"2026-03-20 07:19:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":58315,"visible":true,"origin":"","legend":"\u003cp\u003eFlow Chart of Double Machine Learning (DML)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/88dfeb802006a46fba7a5d2c.png"},{"id":104012893,"identity":"710a7bc8-13da-4e9d-b6e5-58b2f37a58d3","added_by":"auto","created_at":"2026-03-05 16:13:03","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":101957,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDynamic Evolution of the Policy Treatment Effect\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote:\u003c/em\u003e \u003cem\u003eThe figure illustrates the coefficients and their 95% confidence intervals for each year before and after the policy implementation. Key characteristics: When t \u0026lt; 0, the coefficients fluctuate around 0; when t = 0 (corresponding to 2019, the policy implementation year) and t = 1 (corresponding to 2020), the coefficients are significantly positive; when t ≥ 3 (corresponding to 2022 and beyond), the coefficients significantly turn negative.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/c57a4c3105080430ee068893.png"},{"id":104012891,"identity":"ecabf713-e1cd-4c61-aaae-1abd361ec607","added_by":"auto","created_at":"2026-03-05 16:13:03","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":79533,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePlacebo Test: Kernel Density Distribution of Coefficients\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/9af18709370ed08baa912d92.png"},{"id":104012894,"identity":"758dea9a-d6e2-4685-812e-094740a4b970","added_by":"auto","created_at":"2026-03-05 16:13:03","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":79108,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeterogeneous Policy Effects Across Cities at Different Lifecycle Stages\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/dab42d5b8581b8ed66913e0f.png"},{"id":104012897,"identity":"1207a510-a199-474f-a5d1-49288fbe3fff","added_by":"auto","created_at":"2026-03-05 16:13:04","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":90008,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeterogeneous Policy Effects Across Cities at Different Lifecycle Stages\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote: The forest plot is drawn based on regression results. Regenerating stage: Coef = -0.125***, Growing stage: Coef = -0.082***, Mature stage: Coef = -0.015 (ns), Declining stage: Coef = +0.032**.\u003c/em\u003e \u003cem\u003eFigure 6 presents regression coefficients for all 938 counties in the entire basin, while Figure 7 shows coefficients for 384 core counties; the difference arises because the core sample focuses on areas around the main stream of the Yellow River, excluding peripheral low-sensitivity counties.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/c2c0e3d9a8e6402f6a563f23.png"},{"id":104012896,"identity":"23b8576e-e37f-4b8e-aedd-bda793474d39","added_by":"auto","created_at":"2026-03-05 16:13:03","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":383094,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSpatial Heterogeneity of the Yellow River Strategy's Policy Effects and Distribution of Urban Types (Panel (a) illustrates the policy effects (\u003c/strong\u003e τ\u003csub\u003ei\u003c/sub\u003e \u003cstrong\u003e) based on Geo-SHAP, where blue indicates emission reduction and red indicates rebound; Panel (b) shows the lifecycle classification of RBCs.)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/121264b1f54abf001f5119de.png"},{"id":106401774,"identity":"129a8329-cf77-4112-9258-7df5da68a618","added_by":"auto","created_at":"2026-04-08 09:09:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4552559,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8517802/v1/495e5b02-10bb-4dd8-af89-ffa0d0f37ab3.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"The Green Paradox in Resource-Dependent Regions: Heterogeneous Impacts of Environmental Regulation on Carbon Lock-in Subtitle: Evidence from Double Machine Learning in China's Yellow River Basin","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eGlobal climate change has compelled nations to seek a fundamental \u0026quot;decoupling\u0026quot; of economic growth from carbon emissions. As the world\u0026apos;s largest carbon emitter, China has pledged to peak carbon emissions before 2030 and achieve carbon neutrality by 2060. Under this grand vision, the low-carbon transition of Resource-based Cities (RBCs) constitutes one of the most challenging issues. Unruh\u003csup\u003e1\u003c/sup\u003e proposed the theory of \u0026quot;Carbon Lock-in,\u0026quot; stating that industrialized economies tend to lock themselves into fossil-fuel-based energy systems through the co-evolution of technological institutions, infrastructure, and behavioral norms. For RBCs, this lock-in effect is particularly acute: high-energy-consuming industrial structures, high-carbon energy consumption patterns, and high exit barriers caused by asset specificity make their transition not merely a technological iteration, but a profound socioeconomic restructuring\u0026nbsp;\u003csup\u003e2\u0026ndash;4\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eAlthough environmental regulation is regarded as a primary policy tool to break this lock-in, intense theoretical debate remains regarding its real-world efficacy in RBCs. Proponents, primarily based on the \u0026quot;Porter Hypothesis,\u0026quot; argue that appropriate environmental regulation can generate an \u0026quot;innovation compensation effect,\u0026quot; compelling firms to undertake technological innovation and process optimization, thereby offsetting compliance costs and achieving a win-win for the environment and the economy\u003csup\u003e5\u003c/sup\u003e. Recent causal evidence from developed economies has largely supported this view, showing that market-based instruments can effectively induce green patents\u003csup\u003e6\u003c/sup\u003e\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHowever, opponents or skeptics raise the warning of the \u0026quot;Green Paradox.\u0026quot; Sinn (2008) argued from a supply-side perspective that if resource owners anticipate increasingly stringent future climate policies (e.g., high carbon taxes), they may form \u0026quot;Doomsday Expectations\u0026quot; and choose to accelerate the extraction of fossil fuels to monetize them before asset devaluation. This \u0026quot;intertemporal arbitrage\u0026quot; behavior may cause carbon emissions to rise rather than fall in the short term\u003csup\u003e7,8\u003c/sup\u003e.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThis phenomenon has been empirically observed in cross-country studies\u003csup\u003e9\u003c/sup\u003e. and specific contexts like carbon tax announcements in emerging economies. Particularly in China, Wu et al. (2023) found that environmental federalism might exacerbate this paradox due to local governments\u0026apos; \u0026quot;race to the bottom\u0026quot; to protect tax bases\u003csup\u003e10\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Particularly for declining cities at the end of their lifecycle, lacking funds for technological retrofitting, strong regulations may instead induce firms to maintain high-emission production until market exit\u003csup\u003e11\u003c/sup\u003e\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThese two contrasting theoretical predictions imply that policy responses in RBCs may exhibit a high degree of heterogeneity. However, existing literature attempting to clarify these controversies is largely constrained by methodological limitations, leading to empirical results distorted by \u0026quot;Model Misspecification Bias.\u0026quot; Traditional econometric models (such as OLS or standard DID) typically rely on strict parametric assumptions, assuming a linear relationship between control variables and outcome variables. Yet, in complex urban eco-economic systems, carbon emissions are influenced non-linearly by a massive number of high-dimensional factors such as economy, population, technology, and fiscal capacity, with complex interactions among variables. Varian (2014) pointed out that non-linear relationships are ubiquitous among economic variables, which traditional models struggle to capture\u003csup\u003e12\u003c/sup\u003e. Chernozhukov (2018) further demonstrated that ignoring these non-linear features leads to serious \u0026quot;Regularization Bias,\u0026quot; thereby masking the true net effect of policies\u003csup\u003e13\u003c/sup\u003e. Although some scholars have begun attempting non-linear methods\u0026nbsp;\u003csup\u003e14,15\u003c/sup\u003e\u003cstrong\u003e.\u003c/strong\u003eresearch combining high-dimensional machine learning for causal inference in the specific field of RBC transition remains relatively scarce.\u003c/p\u003e\n\u003cp\u003eTo clarify the aforementioned theoretical controversies, this study utilizes China\u0026apos;s \u0026quot;Yellow River Basin Ecological Protection and High-Quality Development\u0026quot; strategy (implemented in 2019, hereinafter \u0026quot;Yellow River Strategy\u0026quot;) as an ideal Quasi-natural Experiment\u0026nbsp;\u003csup\u003e16\u003c/sup\u003e. Unlike previous fragmented policies, this strategy implemented a combination of \u0026quot;hard constraints\u0026quot; (ecological redlines, water-based production limits) and \u0026quot;soft incentives\u0026quot; (high-quality development subsidies) within a specific geo-economic region, providing an excellent window for identifying the causal effects of regional environmental policies\u003csup\u003e17\u003c/sup\u003e. This study introduces the frontier Double Machine Learning (DML) framework, combining the power of Random Forest algorithms in handling high-dimensional non-linear data with the causal inference logic of econometrics. Through Orthogonalization, we utilize long-term panel data from 938 counties across 9 provinces in the Yellow River Basin to precisely isolate the policy\u0026apos;s \u0026quot;net effect\u0026quot; from high-dimensional noise.\u003c/p\u003e\n\u003cp\u003eThe marginal contributions of this study are primarily reflected in three aspects:\u003c/p\u003e\n\u003cp\u003eFirst, Methodological Correction. We confirm that traditional linear models, unable to handle high-dimensional non-linear confounding variables, significantly underestimate the policy effect by approximately 40%, whereas DML provides a more robust estimation of emission reductions by mitigating bias from confounding variables.\u003c/p\u003e\n\u003cp\u003eSecond, Mechanism Identification. We open the \u0026quot;black box\u0026quot; of policy action, quantifying the dual driving paths of \u0026quot;technological greening\u0026quot; (forcing green patent applications) and \u0026quot;structural cleaning\u0026quot; (washing out high-energy-consuming industries).\u003c/p\u003e\n\u003cp\u003eThird, Revealing the \u0026quot;Just Transition\u0026quot; Dilemma. By focusing on heterogeneity analysis of 384 counties in the core basin, we discover significant lifecycle differences: while regenerating cities benefit significantly, declining cities show signs of a \u0026quot;Green Paradox.\u0026quot; This finding provides vivid Chinese experience for the global \u0026quot;Just Transition\u0026quot; discourse\u003csup\u003e18,19\u003c/sup\u003e, warning that singular strong regulations without complementary support policies may exacerbate regional inequality.\u003c/p\u003e"},{"header":"2 Institutional Background and Theoretical Hypotheses","content":"\u003ch2\u003e\u003cstrong\u003e\u003cem\u003e2.1 Institutional Background: From Fragmented Governance to Systemic Regulation\u003c/em\u003e\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe Yellow River Basin spans China\u0026apos;s three major geographic steps (East, Central, West). It is not only the cradle of Chinese civilization but also the core \u0026quot;Energy Basin\u0026quot; of contemporary China. This region holds more than half of the country\u0026apos;s basic coal reserves and hosts clusters of high-density RBCs such as Ordos and Yulin, making it a pivotal area for national energy security \u003csup\u003e20\u003c/sup\u003e. However, long-term extensive development has led to the coexistence of a fragile ecological environment and high-intensity carbon emissions.\u003c/p\u003e\n\u003cp\u003eIn 2019, the Chinese central government elevated \u0026quot;Yellow River Basin Ecological Protection and High-Quality Development\u0026quot; to a major national strategy. Unlike previous environmental policies based on administrative divisions, this strategy possesses significant \u0026quot;systemic\u0026quot; and \u0026quot;strong constraint\u0026quot; characteristics, providing an ideal quasi-natural experimental setting for this study\u003csup\u003e21\u003c/sup\u003e. Specifically, the strategy introduced a comprehensive \u0026quot;carrot and stick\u0026quot; mechanism. On the one hand, regarding \u0026quot;hard constraints\u0026quot; (the stick), the strategy established rigid principles of \u0026quot;determining city, land, people, and production by water\u0026quot; (Yi Shui Ding Chan). For water-intensive and energy-intensive resource industries, this implies insurmountable physical boundaries for capacity expansion and stricter ecological redlines. On the other hand, complementing these constraints are \u0026quot;soft incentives\u0026quot; (the carrot). The central government established special subsidy funds to guide cities to transition from pure resource extraction to green manufacturing and ecological tourism, fostering high-quality development\u003csup\u003e16\u003c/sup\u003e. Theoretically, this combination will have profound causal impacts on the carbon emissions of RBCs by altering both compliance costs and development incentives.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003e\u003cem\u003e2.2 Theoretical Analysis and Research Hypotheses\u003c/em\u003e\u003c/strong\u003e\u003c/h2\u003e\n\u003ch3\u003e\u003cem\u003e2.2.1 Regulation Intensity, Official Incentives, and Net Reduction Effects\u003c/em\u003e\u003c/h3\u003e\n\u003cp\u003eWhether environmental regulation is effective depends on whether it can alter the behavioral incentives of local governments and firms. Under China\u0026apos;s unique \u0026quot;Promotion Tournament\u0026quot; system, GDP growth has long been the core indicator for local official promotion, leading local governments to often adopt \u0026quot;protective umbrella\u0026quot; strategies for high-polluting major taxpayers\u003csup\u003e22\u003c/sup\u003e. However, the uniqueness of the Yellow River Strategy lies in its extremely high political hierarchy, placing \u0026quot;ecological protection\u0026quot; before \u0026quot;high-quality development,\u0026quot; substantively reshaping the performance evaluation system of provinces along the river \u003csup\u003e16\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eOn one hand, the delineation of ecological redlines increases the Compliance Costs for high-carbon firms, forcing them to cut fossil energy inputs under profit maximization constraints. On the other hand, the strengthened political accountability mechanism breaks local protectionism, forcing local governments to transform from \u0026quot;pollution colluders\u0026quot; to \u0026quot;environmental regulators\u003csup\u003e23\u003c/sup\u003e. \u0026quot; This superposition of vertical administrative pressure from the central government and horizontal cost pressure on firms theoretically forms a powerful combined force for emission reduction.\u003c/p\u003e\n\u003cp\u003eHypothesis 1 (H1): As a strong constraint environmental regulation, the Yellow River Strategy significantly reduces the carbon emission intensity of RBCs by reshaping local officials\u0026apos; promotion incentives and increasing firms\u0026apos; environmental compliance costs.\u003c/p\u003e\n\u003ch3\u003e\u003cem\u003e2.2.2 Transmission Mechanisms: \u0026quot;Innovation Compensation\u0026quot; and \u0026quot;Creative Destruction\u0026quot;\u003c/em\u003e\u003c/h3\u003e\n\u003cp\u003eEnvironmental regulation is not simply about \u0026quot;shutting down and transferring\u0026quot;; its deep emission reduction momentum stems from technological progress and structural change.\u003c/p\u003e\n\u003cp\u003eFirst, according to the \u0026quot;Innovation Compensation Effect\u0026quot; of the Porter Hypothesis, moderate environmental pressure forces firms to reflect on technology. Facing the \u0026quot;hard constraints\u0026quot; set by the Yellow River Strategy, surviving resource-based firms must increase R\u0026amp;D investment to survive in long-term competition, offsetting rising compliance costs by applying for green patents and updating clean equipment. This \u0026quot;Technology Effect\u0026quot; induced by regulation is an endogenous driver for reducing carbon intensity\u003csup\u003e5,24\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eSecond, based on Schumpeter\u0026apos;s \u0026quot;Creative Destruction\u0026quot; theory, high-standard environmental regulation functions as a market filter. The \u0026quot;determining production by water\u0026quot; policy effectively sets a threshold for industrial entry. High-pollution, low-efficiency backward production capacities (usually low-end manufacturing) are forced to exit the market due to the inability to bear high pollution control costs; meanwhile, the land, credit, and labor factors released flow toward low-energy service industries or high-tech industries. This \u0026quot;Structural Effect\u0026quot; (or Cleaning Effect) of industrial composition will directly drive the urban economy toward low-carbon transition\u003csup\u003e25\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eHypothesis 2 (H2): The Yellow River Strategy drives the decline of carbon intensity through a dual-wheel mechanism: stimulating firms\u0026apos; green technological innovation (Technology Effect) and accelerating the exit and substitution of high-energy-consuming industries (Structural Effect).\u003c/p\u003e\n\u003ch3\u003e\u003cem\u003e2.2.3 Heterogeneity Perspective: The \u0026quot;Green Paradox\u0026quot; in Declining Cities\u003c/em\u003e\u003c/h3\u003e\n\u003cp\u003eRBCs are not a homogeneous whole. For cities in the growth or regeneration stage, with ample fiscal resources and high industrial diversification, it is easier to respond to policies and transition. However, for resource-depleted declining cities, the situation may be vastly different. Sinn\u0026apos;s (2008) Green Paradox points out that if the supply side expects future environmental policies to be extremely strict (e.g., forced shutdowns), resource owners will tend to accelerate extraction before the policy fully lands to avoid asset stranding\u003csup\u003e7\u003c/sup\u003e. Furthermore, declining cities often face severe fiscal deficits and brain drain, lacking funds for green technological upgrades. Under strong regulatory pressure, firms in these cities may choose \u0026quot;rush-style\u0026quot; production to overdraft remaining resources, leading to a rise rather than a fall in carbon emissions\u003csup\u003e26\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eHypothesis 3 (H3): Policy effects exhibit significant lifecycle heterogeneity; compared to regenerating cities, declining cities may experience a short-term rebound in carbon emissions\u0026mdash;a \u0026quot;Green Paradox\u0026quot; phenomenon\u0026mdash;due to a lack of financial support and strong \u0026quot;doomsday expectations.\u0026quot;\u003c/p\u003e"},{"header":"3 Data and Empirical Strategy","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Sample Construction and Study Area\u003c/h2\u003e\n \u003cp\u003eThe primary challenge of empirical research lies in constructing a sample system that satisfies causal inference assumptions while accurately reflecting policy targeting characteristics. Based on this, we designed a dual-layer sample structure combining \u0026quot;Full Domain Identification\u0026quot; and \u0026quot;Core Focus\u0026quot; (as shown in Fig.\u0026nbsp;2).\u003c/p\u003e\n \u003cp\u003eSpecifically, in the causal identification stage, to construct a sufficiently robust counterfactual framework, we set the sample range to 938 counties across the entire 9 provinces of the Yellow River Basin (2000\u0026ndash;2023). Although this range covers some peripheral areas geographically distant from the main Yellow River stream, retaining these non-core samples has significant econometric value: they serve as effective intra-provincial control groups, helping the DML model eliminate macro shocks at the provincial level and providing rich high-dimensional feature variation to improve the machine learning algorithm\u0026apos;s prediction accuracy for potential counterfactual outcomes.\u003c/p\u003e\n \u003cp\u003eIn contrast, when analyzing spatial heterogeneity, to isolate spatial background noise and intuitively demonstrate the micro-transmission mechanism, we focus the analytical view on the 384 core counties within the physical watershed of the Yellow River. These counties adjacent to the main streams and tributaries are the direct fields of action for \u0026quot;determining production by water\u0026quot; and \u0026quot;ecological redlines,\u0026quot; allowing for a clearer depiction of the spatial implementation of the strategy.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Variable Measurement and Data Description\u003c/h2\u003e\n \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\n \u003ch2\u003e3.2.1 Core Variable Measurement\u003c/h2\u003e\n \u003cp\u003eThe core dependent variable (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Y}_{it}\\)\u003c/span\u003e\u003c/span\u003e) is Carbon Emission Intensity, defined as carbon dioxide emissions per unit of GDP (tons/10,000 yuan). Data are extracted from the high-precision China Emission Accounts and Datasets (CEADs) and its accompanying micro-inversion dataset\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. Compared to traditional aggregate data based on energy statistical yearbooks, CEADs data integrates night-time light remote sensing inversion technology, accurately capturing minute carbon emission dynamics at the county level and effectively overcoming the \u0026quot;data masking\u0026quot; problem of traditional statistics.\u003c/p\u003e\n \u003cp\u003eThe core independent variable (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{it}\\)\u003c/span\u003e\u003c/span\u003e) is set as a typical multi-period policy shock interaction term (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Treat}_{i}\\times\\:{Post}_{i}\\)\u003c/span\u003e\u003c/span\u003e). Specifically, we define the group dummy variable \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Treat}_{i}\\)\u003c/span\u003e\u003c/span\u003e: if county\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:i\\:\\)\u003c/span\u003e\u003c/span\u003eis located in the 9 provinces of the Yellow River Basin and within the policy implementation scope, then \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Treat}_{i}=1\\)\u003c/span\u003e\u003c/span\u003e, otherwise 0; simultaneously, define the time dummy variable \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Post}_{i}\\)\u003c/span\u003e\u003c/span\u003e: if time \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:t\\:\\)\u003c/span\u003e\u003c/span\u003eis 2019 (the year of official strategy implementation) or later, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Post}_{i}=1\\)\u003c/span\u003e\u003c/span\u003e, otherwise 0.\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:D}_{it}\\)\u003c/span\u003e\u003c/span\u003e is the product of the two, used to capture the net difference in treated counties before and after implementation.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\n \u003ch2\u003e3.2.2 Mechanism Variables and Covariates\u003c/h2\u003e\n \u003cp\u003eTo open the \u0026quot;black box\u0026quot; of policy effects, we introduce two key mechanism variables: Green Technological Innovation (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Mech\\_Tech}_{it}\\)\u003c/span\u003e\u003c/span\u003e) and Industrial Structure Optimization (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Mech\\_Str}_{it}\\)\u003c/span\u003e\u003c/span\u003e). The former is measured by the total number of green patent applications per 10,000 people (logarithm), sourced from the China National Intellectual Property Administration (CNIPA), directly reflecting firms\u0026apos; \u0026quot;innovation compensation\u0026quot; behavior; the latter is represented by the proportion of secondary industry value-added in GDP, capturing the \u0026quot;cleaning effect\u0026quot; of high-energy industry exit. Additionally, considering the systemic complexity of RBC transition, we rely on the DML algorithm to introduce a high-dimensional covariate set (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{X}_{it}\\)\u003c/span\u003e\u003c/span\u003e) covering four dimensions: economic basis, industrial characteristics, population/society, and fiscal capacity, to maximally block confounding paths. Descriptive statistics are detailed in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003e\u003c/p\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDescriptive Statistics of Main Variables\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" style=\"width: 10.2409%;\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 20.6137%;\"\u003e\n \u003cp\u003eVariable\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eDefinition\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 11.8317%;\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003eStd. Dev.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003eMin\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003eMax\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDependent\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eCarbon Intensity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eCO2 Emissions / GDP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e3.068\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e4.863\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e0.094\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e30.325\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIndependent\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003ePolicy Shock (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{it}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eInteraction term\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e0.061\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e0.239\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMechanism\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eIndustrial Structure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eSecondary Ind. / GDP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e0.403\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e0.164\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e0.076\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e0.816\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eGreen Performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eSO2 Emissions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e8603.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e11760.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e39.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e73365.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eControl\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eLn PGDP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(Per capita GDP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e9.967\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e1.102\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e7.497\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e12.408\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eLn Fixed Asset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(Fixed Asset Investment)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e12.758\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e2.005\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e7.960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e16.450\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eLn FDI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(FDI\u0026thinsp;+\u0026thinsp;1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e4.093\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e4.568\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e14.281\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eLn Population\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(Population)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e4.316\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e1.946\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e6.043\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eUrbanization\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eUrban Emp. Ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e0.316\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e0.228\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e0.015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e0.999\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eFiscal Revenue\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(Revenue)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e10.749\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e1.341\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e7.313\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e14.152\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 13.2729%;\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 22.9675%;\"\u003e\n \u003cp\u003eFiscal Expend.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 26.5468%;\"\u003e\n \u003cp\u003eLog(Expenditure)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.1702%;\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.6616%;\"\u003e\n \u003cp\u003e11.455\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 8.9484%;\"\u003e\n \u003cp\u003e1.189\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 5.4685%;\"\u003e\n \u003cp\u003e8.448\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.8547%;\"\u003e\n \u003cp\u003e14.619\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003e\u003cem\u003e(Note: Data covers 938 counties from 2000 to 2023. All continuous variables are winsorized at the 1st and 99th percentiles to mitigate the impact of outliers.)\u003c/em\u003e\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Econometric Model: Double Machine Learning (DML)\u003c/h2\u003e\n \u003cp\u003eWhen evaluating environmental policy effects, the commonly used Two-Way Fixed Effects (TWFE) model in existing literature relies on strict linear assumptions. However, as Varian (2014) emphasized, complex economic systems are replete with non-linear relationships\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. If the true data generation process is non-linear (e.g., the impact of industrial structure on carbon emissions follows an inverted U-shape with development stages), forcing a linear model leads to severe \u0026quot;Model Misspecification Bias.\u0026quot; Chernozhukov (2018) further proved that with high-dimensional control variables, this bias cannot be eliminated simply by increasing sample size, i.e., \u0026quot;Regularization Bias\u0026quot; exists\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n \u003cp\u003eTo fundamentally overcome this limitation, this study adopts the Double Machine Learning (DML) framework for causal identification (as shown in Fig.\u0026nbsp;3). We specify the following partially linear regression model:\u003c/p\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$$\\:{Y}_{it}=\\theta\\:{D}_{it}+g\\left({X}_{it}\\right)+{\\zeta\\:}_{it},\\:\\:\\:\\:\\:\\:E\\left[{\\zeta\\:}_{it}|{D}_{it},{X}_{it}\\right]=0$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\:{D}_{it}=m\\left({X}_{it}\\right)+{v}_{it},\\:\\:\\:\\:\\:\\:E\\left[{v}_{it}|{X}_{it}\\right]=0$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\theta\\:\\)\u003c/span\u003e\u003c/span\u003e is the policy net effect of interest; unlike traditional methods, DML introduces two nuisance functions \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:g\\left({X}_{it}\\right)\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:m\\left({X}_{it}\\right)\\)\u003c/span\u003e\u003c/span\u003e to capture the complex non-linear impacts of control variables on carbon emissions and policy probability, respectively. We use the Random Forest algorithm to flexibly fit these two functions and employ Orthogonalization techniques to cut off the transmission of machine learning prediction errors to the causal parameter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\theta\\:\\)\u003c/span\u003e\u003c/span\u003e. Furthermore, to ensure the robustness of statistical inference, we strictly perform 5-fold Cross-Fitting during the estimation process\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n \u003cp\u003eFigure\u0026nbsp;3 Flow Chart of Double Machine Learning (DML)\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003e3.4 Spatial Causal Identification: Based on Individual Treatment Effect (ITE)\u003c/h2\u003e\n \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\n \u003ch2\u003e3.4.1The traditional econometric parameter\u003c/h2\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u0026nbsp;\u003cspan class=\"mathinline\"\u003e\\(\\:\\theta\\:\\)\u003c/span\u003e\u0026nbsp;\u003c/span\u003e can only identify the Average Treatment Effect (ATE), often masking the Spatial Non-stationarity of policies across regions with different resource endowments. To open the spatial black box, we introduce Explainable AI concepts, constructing a Geo-SHAP framework to estimate the Individual Treatment Effect (ITE).\u003c/p\u003e\n \u003cp\u003eEstimation of ITE: Specifically, for any county\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:i\\)\u003c/span\u003e\u003c/span\u003e, its policy effect proxy value \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\tau\\:}_{i}\\)\u003c/span\u003e\u003c/span\u003eis defined as the difference in predicted carbon intensity under the factual scenario (T\u0026thinsp;=\u0026thinsp;1, strategy implemented) and counterfactual scenario (T\u0026thinsp;=\u0026thinsp;0, not implemented). Based on the trained DML Random Forest model, we calculate this difference using the following formula:\u003c/p\u003e\n \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e$$\\:{\\tau\\:}_{i}\\left(x\\right)=E\\left[{Y}_{i}\\right|\\left({D}_{i}=1,{X}_{i}=x\\right)]-E[{Y}_{i}|\\left({D}_{i}=0,{X}_{i}=x\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThis metric is logically equivalent to the core idea of Geo-SHAP, measuring the marginal contribution of a specific feature (here, the policy variable) to the prediction result.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e\n \u003ch2\u003e3.4.2 Spatial Mapping and Pattern Identification\u003c/h2\u003e\n \u003cp\u003eBy mapping the estimated \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\tau\\:}_{i}\\)\u003c/span\u003e\u003c/span\u003e to geographic coordinates, we can intuitively identify the \u0026quot;Emission Reduction Highlands\u0026quot; (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\tau\\:}_{i}\\ll\\:0\\)\u003c/span\u003e\u003c/span\u003e) and \u0026quot;Resistance Depressions\u0026quot; (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\tau\\:}_{i}\\approx\\:0\\:or\u0026gt;0\\)\u003c/span\u003e\u003c/span\u003e) of the Yellow River Strategy. This step lowers the analysis granularity from the domain-wide average to the micro-county level, allowing us to break through the limitations of traditional regression, thereby providing intuitive evidence for revealing the spatial distribution of the \u0026quot;Green Paradox\u0026quot; in clusters of declining cities.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4 Empirical Results","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003e4.1 Benchmark Regression: Methodological Correction and Net Effect Identification\u003c/h2\u003e\n \u003cp\u003eTo verify the net emission reduction effect (Hypothesis H1) and quantify DML\u0026apos;s bias correction capability in handling high-dimensional non-linear confounding, we adopted a stepwise regression strategy. Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e reports estimation results based on three model specifications. Column (1) is OLS estimation controlling only for simple fixed effects; Column (2) is the field\u0026apos;s standard Two-Way Fixed Effects (TWFE) model; Column (3) is the Double Machine Learning (DML) model used in this study.\u003c/p\u003e\n \u003cp\u003e\u003c/p\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe Impact of Yellow River Strategy on Carbon Intensity\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(1) OLS\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(2) TWFE\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(3) DML (Preferred)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePolicy Shock (\u003c/strong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\varvec{D}}_{\\varvec{i}\\varvec{t}}\\)\u003c/span\u003e\u003c/span\u003e)\u003cstrong\u003e)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e-0.019*\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e-0.035**\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e-0.058***\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.016\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eControls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNon-Linear (Random Forest)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eFE (City \u0026amp; Year)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAbsorbed via Cross-fitting\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eBias Correction\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eYes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eObservations\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\"\u003e\u003cem\u003e(Note: Robust standard errors clustered at the county level are in parentheses. *** p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ** p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, * p\u0026thinsp;\u0026lt;\u0026thinsp;0.1. DML estimates are based on 5-fold cross-fitting with random forest learners to prevent overfitting.)\u003c/em\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003eThe empirical results reveal two key findings of methodological significance: First, the robustness of the policy effect. Regardless of model specification, the policy interaction term coefficient \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\theta\\:\\)\u003c/span\u003e\u003c/span\u003e is consistently negative and significant. Specifically, the DML model (Column 3) estimates a net reduction effect of -0.058 (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01).This provides solid empirical evidence that the strategy, as a strong environmental regulation, effectively reduced carbon intensity in RBCs through \u0026quot;hard constraints\u0026quot; and \u0026quot;soft incentives,\u0026quot; strongly supporting Hypothesis H1.\u003c/p\u003e\n \u003cp\u003eSecond, the \u0026quot;Attenuation Bias\u0026quot; of traditional models. This is a phenomenon long ignored by existing literature. Comparing Column (2) and Column (3), the reduction effect estimated by the traditional TWFE model is -0.035, while the \u0026quot;net effect\u0026quot; estimated by the DML model significantly increases to -0.058. This implies that ignoring non-linear features of the economic system leads traditional linear models to underestimate the policy effect by approximately 40%. This discrepancy profoundly reveals the insidiousness of \u0026quot;Regularization Bias\u0026quot;: since carbon emissions in RBCs are subject to non-linear interference from high-dimensional factors (e.g., resource endowments, historical accumulation), forcing linear forms leads to model under-fitting, causing explanatory power to \u0026quot;leak\u0026quot; into residuals and shrinking the policy coefficient toward zero. DML successfully isolates this non-linear noise through orthogonalization, restoring the stronger real reduction capability of the Yellow River Strategy \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2 Dynamic Effects and \u0026quot;Green Paradox\u0026quot; Verification\u003c/h2\u003e\n \u003cp\u003eBenchmark regression only identified the Average Treatment Effect (ATE). To examine the time lag of policy efficacy and potential short-term gaming behaviors, we combined the Event Study method with the DML framework to plot the dynamic evolution path of policy effects(shown as Fig.\u0026nbsp;4).\u003c/p\u003e\n \u003cp\u003eFirst, Pre-treatment Parallel Trends were fully verified. Before policy implementation (t\u0026thinsp;\u0026lt;\u0026thinsp;0), all estimated coefficient confidence intervals included 0, proving that treated and control groups shared consistent evolution paths prior to the strategy.\u003c/p\u003e\n \u003cp\u003eHowever, the most striking finding occurred in the early stages of implementation. Crucially, at t\u0026thinsp;=\u0026thinsp;0 and t\u0026thinsp;=\u0026thinsp;1, the coefficients did not immediately turn negative but instead showed a distinct positive jump. This seemingly counter-intuitive phenomenon confirms Sinn\u0026apos;s (2008) \u0026quot;Green Paradox\u0026quot; hypothesis at the micro-level. Facing imminent \u0026quot;Three Lines One Permit\u0026quot; hard constraints, some highly resource-dependent cities formed strong \u0026quot;Doomsday Expectations.\u0026quot; To monetize resources before policy tightening, firms tended to engage in \u0026quot;Intertemporal Arbitrage\u0026quot; style rush production, causing short-term emissions to rise\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n \u003cp\u003eFrom the third year (t\u0026thinsp;=\u0026thinsp;3), policy coefficients rapidly fell and turned significantly negative. This dynamic path from \u0026quot;short-term pain\u0026quot; to \u0026quot;long-term substantial reduction\u0026quot; indicates that the Yellow River Strategy eventually overcame initial adjustment costs through long-term mechanisms (e.g., green innovation), achieving true low-carbon transition.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\n \u003ch2\u003e4.3 Robustness Checks\u003c/h2\u003e\n \u003cp\u003eTo further exclude drivers from omitted variables or random noise, we conducted a rigorous Placebo Test. We randomly generated false treatment groups and policy times 500 times.\u003c/p\u003e\n \u003cp\u003eShown as Fig. 5, The kernel density distribution shows that the 500 random coefficients follow a standard normal distribution centered at 0. In contrast, the true benchmark coefficient (-0.058) lies significantly at the left tail, far from the placebo confidence interval. This indicates the observed reduction effect is not a random coincidence (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003ch2\u003e4.4 Lifecycle Heterogeneity: A \u0026quot;Just Transition\u0026quot; Perspective\u003c/h2\u003e\n \u003cp\u003eBased on the \u0026quot;National Sustainable Development Plan for RBCs,\u0026quot; we classified the sample into Regenerating, Growing, Mature, and Declining categories.\u003c/p\u003e\n \u003cp\u003eResults show significant inter-group heterogeneity (\u0026quot;Matthew Effect\u0026quot;)(Figure 6). Regenerating Cities showed the strongest reduction effect (Coef \u0026asymp; -0.07). Growing Cities also showed significant reductions. In contrast, Mature Cities showed insignificant coefficients, reflecting strong \u0026quot;Carbon Lock-in.\u0026quot; Most critically, Declining Cities showed coefficients converging toward 0 or even turning positive. This confirms the \u0026quot;Green Paradox\u0026quot;\u003csup\u003e7\u003c/sup\u003e: under strict regulation and fiscal exhaustion, these cities lack funds for transition and may accelerate backward capacity operation to survive. This warns that for declining cities, singular \u0026quot;hard constraints\u0026quot; may fail, urgently requiring complementary \u0026quot;soft support\u0026quot; for Just Transition.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"5 Mechanism Analysis and Heterogeneity Discussion","content":"\u003cdiv id=\"Sec23\" class=\"Section2\"\u003e\n \u003ch2\u003e5.1 Mechanism Verification: Technology Effect vs. Structural Effect\u003c/h2\u003e\n \u003cp\u003eTo verify the transmission paths proposed in Hypothesis H2, we estimated the policy\u0026apos;s impact on mechanism variables using the DML framework. The regression results, reported in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, reveal a striking \u0026quot;Asymmetric Transition\u0026quot; pattern characterized by significant technological improvements alongside structural rigidity.\u003c/p\u003e\n \u003cp\u003e\u003c/p\u003e\u0026nbsp;\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eMechanism Analysis Regression Results\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(1) Green Performance\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(2) Industrial Structure\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eDependent Variable\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:-\\text{l}\\text{n}({SO}_{2}+1)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSecond. Industry Share\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePolicy Shock (\u003c/strong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\varvec{D}}_{\\varvec{i}\\varvec{t}}\\)\u003c/span\u003e\u003c/span\u003e\u003cstrong\u003e)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e-0.290\u003c/strong\u003e\u003csup\u003e\u003cstrong\u003e***\u003c/strong\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.032\u003c/strong\u003e\u003csup\u003e\u003cstrong\u003e***\u003c/strong\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-0.006\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eControls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eYear FE\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eCluster SE\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCounty\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCounty\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eObservations\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21980\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eR-squared\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.048\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.321\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\"\u003e\u003cem\u003eNote: Robust standard errors clustered at the county level are in parentheses. *** p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ** p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, * p\u0026thinsp;\u0026lt;\u0026thinsp;0.1. All regressions include the full set of controls used in the benchmark model.\u003c/em\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003eFirst, the impact of the Yellow River Strategy on Green Performance (proxied by \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:-\\text{l}\\text{n}({SO}_{2}+1)\\)\u003c/span\u003e\u003c/span\u003e is significantly negative (Coef = -0.290, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). This indicates a substantial reduction in pollution intensity via technological upgrades and end-of-pipe treatments. Facing strict environmental constraints, firms have actively adopted cleaner technologies to offset compliance costs, a phenomenon consistent with the \u0026quot;Innovation Compensation\u0026quot; effect of the Porter Hypothesis. This technological \u0026quot;efficiency gain\u0026quot; appears to be the primary driver of the observed aggregate carbon reduction (ATE = -0.058).\u003c/p\u003e\n \u003cp\u003eSecond, and crucially, the impact on Industrial Structure (Secondary Industry Share) is significantly positive (Coef\u0026thinsp;=\u0026thinsp;+\u0026thinsp;0.032, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). Contrary to the expectation that regulation would immediately \u0026quot;wash out\u0026quot; high-polluting industries (which would yield a negative coefficient), this positive coefficient suggests a \u0026quot;Structural Stickiness\u0026quot; or even a short-term \u0026quot;Structure Reflow.\u0026quot; The share of the secondary industry did not decline but rather increased slightly relative to the control group. This counter-intuitive finding provides micro-level empirical evidence for Sinn\u0026rsquo;s (2008) \u0026quot;Green Paradox.\u0026quot; Facing the imminent \u0026quot;hard constraints\u0026quot; of the Yellow River Strategy (e.g., water-based production limits), resource-dependent firms\u0026mdash;anticipating future capacity caps\u0026mdash;likely formed \u0026quot;Doomsday Expectations.\u0026quot; Consequently, they engaged in \u0026quot;Intertemporal Arbitrage\u0026quot; by accelerating extraction and production to monetize fossil fuel assets before the full enforcement of regulations. This \u0026quot;Rush-to-Produce\u0026quot; behavior temporarily inflated the share of industrial output.\u003c/p\u003e\n \u003cp\u003eCombining these findings, we conclude that the current low-carbon transition in the Yellow River Basin is \u0026quot;Efficiency-driven\u0026quot; rather than \u0026quot;Structure-driven.\u0026quot; The policy has successfully acted as a \u0026quot;Technological Accelerator\u0026quot; but has failed to function as a \u0026quot;Structural Filter\u0026quot; in the short term. This asymmetry highlights a critical vulnerability: without breaking the \u0026quot;Carbon Lock-in\u0026quot; of the industrial structure, the emission reductions achieved solely through technical fixes may face diminishing returns in the long run.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\n \u003ch2\u003e5.2 Heterogeneity Analysis: Spatial Micro-pattern via Geo-SHAP\u003c/h2\u003e\n \u003cp\u003eMapping ITEs of 384 core counties to geographic space reveals a pattern of \u0026quot;Effective in Upstream Ecological Barriers, Obstructed in Midstream Energy-Rich Areas,\u0026quot; as visualized in Fig.\u0026nbsp;7.\u003c/p\u003e\n \u003cp\u003eThe spatial distribution of Individual Treatment Effects (ITE) visualized in Fig. 7 exhibits a distinct \u0026quot;Core-Periphery\u0026quot; gradient. The \u0026quot;Emission Reduction Highlands\u0026quot; (ITE\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:\\ll\\:\\:\\)\u003c/span\u003e\u003c/span\u003e0) are predominantly located in the upstream ecological function zones (e.g., Sanjiangyuan in Qinghai) and the downstream advanced manufacturing clusters (e.g., Jinan and Qingdao in Shandong). These regions benefit either from strict ecological compensation or from advanced technology spillover. Conversely, the \u0026quot;Resistance Depressions\u0026quot; (ITE \u0026asymp; 0 or \u0026gt;\u0026thinsp;0) highly overlap with the \u0026quot;Energy Golden Triangle\u0026quot; (Ordos, Yulin, Ningxia), which constitutes the core of China\u0026apos;s coal bases. The resource curse in these areas is intensified by the high asset specificity of coal mining infrastructure, making immediate transition prohibitively expensive. This spatial evidence mutually corroborates the lifecycle regression results, highlighting the necessity of spatially targeted \u0026quot;Just Transition\u0026quot; policies.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\n \u003ch2\u003e5.3 Heterogeneity by Resource Type: Coal vs. Oil/Metal\u003c/h2\u003e\n \u003cp\u003eDifferent resource endowments imply varying carbon intensities and lock-in degrees. To further explore the source of resistance, we classified the sample into Coal-based, Oil/Gas-based, and Metal-based cities based on their dominant mineral resources. The heterogeneity of policy effects is visualized in Fig.\u0026nbsp;8.\u003c/p\u003e\n \u003cp\u003eAs illustrated in Fig. 8, the estimated coefficients vary significantly across resource categories. The Coal-based cities sub-sample exhibits coefficients that are largely positive or statistically insignificant in the short term, visually confirming the \u0026quot;Green Paradox\u0026quot; effect discussed earlier. This pattern suggests that coal industries, facing the strictest \u0026quot;de-capacity\u0026quot; pressures (e.g., coal consumption caps), reacted with stronger \u0026quot;rush-style\u0026quot; production incentives. Conversely, the coefficients for Metal-based and Oil/Gas-based cities tend to be negative, indicating a smoother transition trajectory. This visual evidence further corroborates that the transition dilemma in the Yellow River Basin is structurally driven by the \u0026quot;Coal Dilemma.\u0026quot;\u003c/p\u003e\n\u003c/div\u003e"},{"header":"6 Conclusion and Policy Implications","content":"\u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Conclusion\u003c/h2\u003e \u003cp\u003eUsing the Yellow River Strategy as a quasi-natural experiment and employing DML and Geo-SHAP, this study draws three primary conclusions. First, from a methodological perspective, we confirm that traditional linear models underestimate policy effects by approximately 40%, whereas the DML framework successfully uncovers the true net reduction effect masked by high-dimensional confounders. Second, the mechanism analysis reveals that current emission reductions are primarily \"Denominator Driven\" (efficiency gains) rather than \"Numerator Driven\" (structural cleaning), which explains the observed short-term rebounds. Third, a significant \"Just Transition Dilemma\" exists: while regenerating cities benefit from the strategy, declining cities face a \"Green Paradox,\" a phenomenon that is spatially concentrated in the midstream Energy Golden Triangle.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Policy Implications\u003c/h2\u003e \u003cp\u003eBased on these findings, we propose the following policy recommendations. First, to address the regional inequality exposed by the \"Green Paradox,\" a central \"Just Transition Fund\" should be established. This fund should prioritize worker reskilling and provide seed funding for alternative industries, such as photovoltaics and tourism, to prevent declining cities from falling into low-carbon poverty traps. Second, governance mechanisms must shift from a uniform approach to differentiated governance. Implementing horizontal ecological compensation is crucial; downstream beneficiaries should purchase \"emission quotas\" from midstream energy cities to internalize environmental externalities. Finally, to achieve deep decarbonization, policy focus must shift from \"End-of-pipe\" governance to \"Source\" adjustment. By raising entry thresholds and leveraging green finance, the transition can move from being merely \"Denominator Driven\" to \"Numerator Driven,\" ensuring long-term sustainability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Limitations and Future Research\u003c/h2\u003e \u003cp\u003eWhile this study offers rigorous empirical evidence, it is not without limitations. First, restricted by data availability, the carbon emission data derived from night-time lights, though high-precision, may still contain measurement errors for certain non-point sources. Future research could benefit from integrating enterprise-level micro-energy consumption data to validate these findings. Second, this study primarily focuses on the internal policy effects within the Yellow River Basin. However, environmental regulations often generate spatial spillover effects to adjacent regions (e.g., pollution haven hypothesis). Future studies should employ spatial DML models to explore whether the Yellow River Strategy has led to carbon leakage to non-policy zones, thereby providing a more comprehensive assessment of the national strategy.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets generated and analyzed during the current study are based on publicly available data. The carbon emission data are derived from the China Emission Accounts and Datasets (CEADs, https://www.ceads.net). Other socio-economic data are available\u0026nbsp;from the\u003cem\u003e\u0026nbsp;China County Statistical Yearbook\u0026nbsp;\u003c/em\u003eand\u003cem\u003e\u0026nbsp;provincial statistical yearbooks.\u003c/em\u003e. The processed data supporting the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study does not involve human participants or animal subjects; therefore, ethical approval is not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study does not involve human participants; therefore, informed consent is not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eYanying Wang: Conceptualization, Methodology, Software, Writing \u0026ndash; original draft.\u003c/p\u003e\n\u003cp\u003eXianzhi Wang: Data curation, Visualization, Investigation.\u003c/p\u003e\n\u003cp\u003eKang Pian: Supervision, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no external funding.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eUnruh, G. C. Understanding carbon lock-in. \u003cem\u003eEnergy Policy\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 817\u0026ndash;830 (2000).\u003c/li\u003e\n\u003cli\u003eJohnstone, P. \u0026amp; Hielscher, S. Phasing out coal, sustaining coal communities? Living with technological decline in sustainability pathways. \u003cem\u003eExtr. Ind. Soc.\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 457\u0026ndash;461 (2017).\u003c/li\u003e\n\u003cli\u003eMoses, N. D. \u0026amp; MacCarty, N. A. What makes a cookstove usable? Trials of a usability testing protocol in Uganda, Guatemala, and the United States. \u003cem\u003eEnergy Res. Soc. Sci.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, 221\u0026ndash;235 (2019).\u003c/li\u003e\n\u003cli\u003eJakob, M. \u003cem\u003eet al.\u003c/em\u003e The future of coal in a carbon-constrained climate. \u003cem\u003eNat. Clim. Change\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 704\u0026ndash;707 (2020).\u003c/li\u003e\n\u003cli\u003ePorter, M. E. \u0026amp; Linde, C. V. D. Toward a New Conception of the Environment-Competitiveness Relationship. \u003cem\u003eJ. Econ. Perspect.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 97\u0026ndash;118 (1995).\u003c/li\u003e\n\u003cli\u003eKnaus, M. C. Double machine learning-based programme evaluation under unconfoundedness. \u003cem\u003eEconom. J.\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 602\u0026ndash;627 (2022).\u003c/li\u003e\n\u003cli\u003eSinn, H.-W. Public policies against global warming: a supply side approach. \u003cem\u003eInt. Tax Public Finance\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 360\u0026ndash;394 (2008).\u003c/li\u003e\n\u003cli\u003eSix, M. \u0026amp; Wirl, F. Optimal pollution management when discount rates are endogenous. \u003cem\u003eResour. Energy Econ.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 53\u0026ndash;70 (2015).\u003c/li\u003e\n\u003cli\u003eYamamoto, Y. Living under ecosystem degradation: Evidence from the mangrove\u0026ndash;fishery linkage in Indonesia. \u003cem\u003eJ. Environ. Econ. Manag.\u003c/em\u003e \u003cstrong\u003e118\u003c/strong\u003e, 102788 (2023).\u003c/li\u003e\n\u003cli\u003eLiu, B. \u003cem\u003eet al.\u003c/em\u003e Regional differences in China\u0026rsquo;s electric vehicle sales forecasting: Under supply-demand policy scenarios. \u003cem\u003eEnergy Policy\u003c/em\u003e \u003cstrong\u003e177\u003c/strong\u003e, 113554 (2023).\u003c/li\u003e\n\u003cli\u003ePloeg, F. V. D. Natural Resources: Curse or Blessing? \u003cem\u003eJ. Econ. Lit.\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, 366\u0026ndash;420 (2011).\u003c/li\u003e\n\u003cli\u003eVarian, H. R. Big Data: New Tricks for Econometrics. \u003cem\u003eJ. Econ. Perspect.\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 3\u0026ndash;28 (2014).\u003c/li\u003e\n\u003cli\u003eChernozhukov, V. \u003cem\u003eet al.\u003c/em\u003e Double/debiased machine learning for treatment and structural parameters. \u003cem\u003eEconom. J.\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, C1\u0026ndash;C68 (2018).\u003c/li\u003e\n\u003cli\u003eCheng, C. C. J. \u0026amp; Shiu, E. C. A two-level, longitudinal investigation into the effects of employee social entrepreneurship orientation and top management team decisions on product innovation. \u003cem\u003eTechnol. Forecast. Soc. Change\u003c/em\u003e \u003cstrong\u003e182\u003c/strong\u003e, 121832 (2022).\u003c/li\u003e\n\u003cli\u003eVan Es, M. \u0026amp; Bruins, B. Pro-poor change in the aftermath of disasters \u0026ndash; Exploring possibilities at the intersection of disaster politics and land rights issues in Central Philippines. \u003cem\u003eLand Use Policy\u003c/em\u003e \u003cstrong\u003e132\u003c/strong\u003e, 106771 (2023).\u003c/li\u003e\n\u003cli\u003eWang, H.-Y., Li, Y., Jiao, S.-Q., Chou, K.-C. \u0026amp; Zhang, G.-H. Recovery of Ni matte from Ni-bearing electroplating sludge. \u003cem\u003eJ. Environ. Manage.\u003c/em\u003e \u003cstrong\u003e326\u003c/strong\u003e, 116744 (2023).\u003c/li\u003e\n\u003cli\u003eHeutel, G., Moreno-Cruz, J. \u0026amp; Shayegh, S. Solar geoengineering, uncertainty, and the price of carbon. \u003cem\u003eJ. Environ. Econ. Manag.\u003c/em\u003e \u003cstrong\u003e87\u003c/strong\u003e, 24\u0026ndash;41 (2018).\u003c/li\u003e\n\u003cli\u003eNewell, P. \u0026amp; Mulvaney, D. The political economy of the \u0026lsquo;just transition\u0026rsquo;. \u003cem\u003eGeogr. J.\u003c/em\u003e \u003cstrong\u003e179\u003c/strong\u003e, 132\u0026ndash;140 (2013).\u003c/li\u003e\n\u003cli\u003eCarley, S. \u0026amp; Konisky, D. M. The justice and equity implications of the clean energy transition. \u003cem\u003eNat. Energy\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 569\u0026ndash;577 (2020).\u003c/li\u003e\n\u003cli\u003eSu, H. \u0026amp; Qi, Z. Polycentric structure and urban thermal environment: A large-scale study from multi-perspectives. \u003cem\u003eSustain. Cities Soc.\u003c/em\u003e \u003cstrong\u003e96\u003c/strong\u003e, 104657 (2023).\u003c/li\u003e\n\u003cli\u003eZhou, Z., Zhou, J., Alcal\u0026aacute;, J. \u0026amp; Yepes, V. Thermal coupling optimization of bridge environmental impact under natural conditions. \u003cem\u003eEnviron. Impact Assess. Rev.\u003c/em\u003e \u003cstrong\u003e104\u003c/strong\u003e, 107316 (2024).\u003c/li\u003e\n\u003cli\u003eJ\u0026auml;ger, S., Schoefer, B., Young, S. \u0026amp; Zweim\u0026uuml;ller, J. Wages and the Value of Nonemployment*. \u003cem\u003eQ. J. Econ.\u003c/em\u003e \u003cstrong\u003e135\u003c/strong\u003e, 1905\u0026ndash;1963 (2020).\u003c/li\u003e\n\u003cli\u003eGreenstone, M. \u0026amp; Hanna, R. Environmental Regulations, Air and Water Pollution, and Infant Mortality in India. \u003cem\u003eAm. Econ. Rev.\u003c/em\u003e \u003cstrong\u003e104\u003c/strong\u003e, 3038\u0026ndash;3072 (2014).\u003c/li\u003e\n\u003cli\u003eAcemoglu, D., Aghion, P., Bursztyn, L. \u0026amp; Hemous, D. The Environment and Directed Technical Change. \u003cem\u003eAm. Econ. Rev.\u003c/em\u003e \u003cstrong\u003e102\u003c/strong\u003e, 131\u0026ndash;166 (2012).\u003c/li\u003e\n\u003cli\u003eGuan, D. \u003cem\u003eet al.\u003c/em\u003e Structural decline in China\u0026rsquo;s CO2 emissions through transitions in industry and energy systems. \u003cem\u003eNat. Geosci.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 551\u0026ndash;555 (2018).\u003c/li\u003e\n\u003cli\u003eTang, Q., Wang, J., Jing, Z., Yan, Y. \u0026amp; Niu, H. Response of ecological vulnerability to land use change in a resource-based city, China. \u003cem\u003eResour. Policy\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 102324 (2021).\u003c/li\u003e\n\u003cli\u003eShan, Y. \u003cem\u003eet al.\u003c/em\u003e China CO2 emission accounts 1997\u0026ndash;2015. \u003cem\u003eSci. Data\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 170201 (2018).\u003c/li\u003e\n\u003cli\u003eBanerjee, A., Chevillon, G. \u0026amp; Kratz, M. Probabilistic forecasting of bubbles and flash crashes. \u003cem\u003eEconom. J.\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 297\u0026ndash;315 (2020).\u003c/li\u003e\n\u003cli\u003eBelloni, A., Chernozhukov, V. \u0026amp; Hansen, C. High-Dimensional Methods and Inference on Structural and Treatment Effects. \u003cem\u003eJ. Econ. Perspect.\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 29\u0026ndash;50 (2014).\u003c/li\u003e\n\u003cli\u003eWager, S. \u0026amp; Athey, S. Estimation and Inference of Heterogeneous Treatment Effects using Random Forests. \u003cem\u003eJ. Am. Stat. Assoc.\u003c/em\u003e \u003cstrong\u003e113\u003c/strong\u003e, 1228\u0026ndash;1242 (2018).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Yellow River Strategy, Carbon Lock-in, Green Paradox, Just Transition, Double Machine Learning","lastPublishedDoi":"10.21203/rs.3.rs-8517802/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8517802/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eResource-based cities in the Global South universally face a \"dual dilemma\" of sustaining economic growth while breaking path-dependent carbon lock-in. As a critical energy basin in China, the low-carbon transition of the Yellow River Basin is pivotal for achieving national carbon neutrality goals. However, debate persists regarding whether the \"Yellow River Basin Ecological Protection and High-Quality Development Strategy\" (implemented in 2019) effectively disrupts this high-carbon dependency. Existing studies often rely on linear models, ignoring the high-dimensional non-linear confounders typical of complex eco-economic systems, potentially masking the true \"Green Paradox\" effects. Treating this strategy as a quasi-natural experiment, this study employs a Double Machine Learning (DML) causal inference framework on a long-term panel dataset (2000–2023) covering 938 counties. This approach utilizes orthogonalization techniques to effectively eliminate the interference of 75 high-dimensional confounding variables. The findings reveal: (1) Aggregate Effect: The strategy significantly reduced carbon intensity on average (ATE = -0.058, p\u0026lt;0.01). (2) Green Paradox Confirmation: Dynamic analysis reveals that emissions did not decline immediately; instead, a short-term rebound occurred initially (t=0, t=1), confirming the existence of \"intertemporal arbitrage\" behaviors. (3) Asymmetric Mechanism: Decomposition analysis indicates that emission reductions were primarily \"Efficiency-driven\" (diluting intensity via innovation, Coef = -0.290) rather than \"Structure-driven\". Notably, the industrial structure coefficient showed a short-term resilience (Coef = +0.032), suggesting that high-carbon industries were not \"washed out\" but rather locked in due to structural rigidity. Conclusion: While top-down environmental strategies act as an effective \"brake\" for expanding economies, declining regions require complementary \"Just Transition\" mechanisms to avoid falling into a low-carbon poverty trap.\u003c/p\u003e","manuscriptTitle":"The Green Paradox in Resource-Dependent Regions: Heterogeneous Impacts of Environmental Regulation on Carbon Lock-in Subtitle: Evidence from Double Machine Learning in China's Yellow River Basin","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-05 16:12:53","doi":"10.21203/rs.3.rs-8517802/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-03-07T14:05:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"163426977096657747228533453638520514095","date":"2026-03-07T06:51:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"161669983154727604377466166433243487594","date":"2026-03-05T14:31:04+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"120280630556419128135043062491938412484","date":"2026-03-02T07:04:55+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-02T06:50:37+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-11T09:48:33+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-10T08:42:12+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-19T03:59:56+00:00","index":"","fulltext":""},{"type":"submitted","content":"Humanities and Social Sciences Communications","date":"2026-01-19T03:54:08+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"5bfed200-120c-4da4-bf69-e96f6f276d2b","owner":[],"postedDate":"March 5th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":63893742,"name":"Biological sciences/Ecology"},{"id":63893743,"name":"Earth and environmental sciences/Ecology"},{"id":63893744,"name":"Earth and environmental sciences/Environmental sciences"},{"id":63893745,"name":"Earth and environmental sciences/Environmental social sciences"}],"tags":[],"updatedAt":"2026-03-05T16:12:54+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-05 16:12:53","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8517802","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8517802","identity":"rs-8517802","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}