An Empirical Analysis of the Effectiveness of Local Industrial Policies for China’s New Energy Vehicle Sector

Abstract

Despite China’s success in its new energy vehicle (NEV) transition, significant regional imbalances persist, raising the question of why provincial policy effectiveness is so context-dependent. To investigate this, this study develops a novel framework to measure policy “quality” and “style”, systematically quantifying 2455 provincial policy documents from 2013 to 2023. Our empirical analysis reveals that policy quality—encompassing its authoritativeness, instrument strength, and resource commitment—is a far more decisive determinant of effectiveness than sheer policy quantity. We identify three primary policy styles with distinct impacts: substantive-driving policies are crucial for stimulating market demand, whereas coordinative-programmatic policies are more effective in guiding industrial supply, revealing a significant goal-mismatch. Conversely, high-level authoritative policies can unexpectedly inhibit infrastructure development. Crucially, the study finds that provincial policies act more as “catalysts” than “creators”, their effectiveness being highly contingent on local economic, fiscal, and industrial fundamentals. The findings of this research offer direct implications for policymaking: decision-makers should shift their focus from pursuing policy quantity to enhancing policy quality and design targeted, “precision-irrigation” policy instrument portfolios tailored to the specific contexts and development objectives (e.g., promoting sales or guiding production) of different regions.

1. Introduction

1.1. Research Background and Strategic Context

The profound decarbonization of the global transportation sector has evolved from a purely environmental issue into a significant industrial revolution and a matter of geo-economic competition [1,2]. Against this backdrop, the new energy vehicle (NEV) industry has not only become a cornerstone for achieving the temperature control targets of the Paris Agreement [3] but has also emerged as a central arena for industrial policy rivalry among nations [4]. For many countries, particularly emerging economies like China, the motivation for advancing the electric mobility transition extends far beyond environmental commitments; it is regarded as a comprehensive national strategy serving multiple state interests [5]. This strategy aims to secure energy independence, reshape the global automotive industry value chain, and ultimately pursue leadership in critical global technology sectors [6,7]. This dual-driven transition, propelled by both climate and industrial policies, highlights the complexity of ensuring an equitable evolution during the pursuit of net-zero emissions, demanding particular attention to the developmental needs and imbalances within different economies [8].

Within this global competitive landscape, China has rapidly elevated its NEV industry to a position of national strategic priority. Its policy approach is not a simple market intervention but is considered a model of combining an “effective government” with market forces. The goal is to leverage its potential comparative advantages in manufacturing and resource endowments to overcome coordination failures, thereby achieving a “lane-changing and overtaking” maneuver in the industry [9,10]. This systematic infusion of state power has catalyzed explosive market growth. By 2023, China had been the world’s largest NEV market for several consecutive years, accounting for nearly two-thirds of global sales [11], while its domestic market penetration rate also far exceeded expectations [12]. This achievement places China’s policy model—one characterized by state dominance and an emphasis on full-lifecycle management—in stark contrast to the models of European and American countries, which rely more heavily on market mechanisms and consumer incentives [13].

However, the macroeconomic success story at the national level masks a reality of profound regional disparities. China’s NEV revolution is not unfolding on a homogenous “blank slate” but rather exhibits a distinct “spatial mosaic” pattern across its geography [14]. NEV sales and charging infrastructure are highly concentrated in a few economically developed coastal provinces [15,16], while the vast inland and less-developed regions lag significantly, creating a substantial regional divide. Recent research further confirms that different regions feature markedly different policy objective portfolios and varying internal mechanisms of competition and cooperation, which exacerbates this regional heterogeneity [14]. This imbalance leads to the central question of this study: within China’s unique governance framework of “central-local relations” [17], where provincial governments are key actors in policy innovation [18], to what extent have the varied policy instruments they adopted effectively promoted market development? And why do these policy effects “vary by place”? Answering this question is crucial for unlocking the “black box” of local policy effectiveness and for providing valuable empirical evidence for sub-national governance in the global transition toward net-zero transportation.

1.2. Literature Review and Research Gap

Evaluating the effectiveness of sub-national policies in promoting clean technology transitions is a central theme in international public policy research. In the United States, a substantial body of research has focused on the actual effects of state-level policies like the renewable portfolio standards (RPS), yet the conclusions are markedly contested. Some studies confirm their positive impact on the installed capacity of renewable energy [19], while others argue their effects are marginal and far less effective than federal-level policy interventions [20,21]. These discrepancies underscore the complexity of policy evaluation, revealing that policy outcomes are deeply embedded in the administrative implementation capacity, market structure, and complex political-economic dynamics of their context [22]. International experience suggests that policy success depends not only on the technical design of the instruments but, more critically, on overcoming the “implementation gap” between policy text and on-the-ground practice [23] and on building sufficiently powerful political coalitions to counteract resistance from incumbent interest groups [24]. This global research context provides a reference framework for examining China’s case and highlights the necessity of deeply analyzing the effectiveness of local-level policies and their underlying mechanisms.

Empirical research focusing on China’s NEV policies can be broadly categorized into two streams. The first concentrates on evaluating the effects of single policy instruments, such as fiscal subsidies or license plate exemptions [25,26]. While such studies offer precise causal identification for the impact of specific interventions, their fundamental limitation lies in treating policy instruments as isolated variables. This methodological reductionism prevents them from capturing the complex synergistic or offsetting effects that arise from the coexistence of multiple policies in the real world; that is, it fails to effectively assess the overall architecture of the “policy mix” [27].

The second stream of research adopts a “policy mix” perspective, acknowledging the combined impact of multiple policies [28], which represents a paradigm advancement. However, these studies are generally constrained by a fundamental methodological challenge: how to scientifically and objectively quantify the “strength” of the policies themselves. Existing research suffers from significant deficiencies in policy quantification, with most studies employing overly simplistic methods such as a binary approach (0/1) [29] or policy counting [30]. The binary method leads to severe information loss by completely ignoring the qualitative differences in the authoritativeness, scope, and intensity of policies [31]. The policy counting method [32] erroneously equates the “quantity” of policy documents with the “quality” of policy signals, operating on an implicit assumption that “all policies are equivalent,” which starkly conflicts with reality [33]. Although some cutting-edge research has begun to apply text-mining techniques [34], it often focuses on a single dimension (such as policy topics) and fails to capture the multi-dimensional nature of policy strength, which is constituted by the issuing body’s authority, the document’s legal force, and resource commitments [35]. This crudeness in measuring the policy “input” fundamentally constrains the accuracy of assessing the policy “output”.

In summary, a clear, cascading research gap exists in the current literature. First, at the methodological level, there is a lack of a multi-dimensional, comprehensive framework for quantifying policy strength that integrates both the extrinsic authority and intrinsic content of policies. Second, this methodological limitation leads to a narrow evaluation scope: unable to measure comprehensive policy strength, existing studies tend to examine its impact on a single output indicator (e.g., sales) [36], while neglecting the multifaceted and differential impacts of policy on the entire industrial ecosystem, which encompasses industrial supply, end-user sales, and infrastructure. Finally, this also results in a weak explanatory power for the mechanisms at play. Although it is widely acknowledged that policy effects “vary by place” [37], the inability to precisely measure the differences in provincial policy characteristics has prevented existing research from providing a robust explanation of the driving factors behind regional heterogeneity—that is, why specific policies are more effective under specific regional conditions. This study aims to fill these gaps by constructing an innovative policy quantification framework.

1.3. Research Questions and Hypotheses

Drawing upon the aforementioned research gaps, this study aims to construct an innovative policy quantification framework to systematically evaluate the effectiveness of China’s provincial-level NEV policies and their underlying mechanisms. The core objective of this research is to answer the following questions: Do different dimensions of provincial policy strength have a statistically significant impact on the key output indicators of the NEV market (encompassing industry, sales, and infrastructure)? Do these impacts exhibit significant heterogeneity contingent on the regional characteristics of the provinces?

To investigate these questions, this study proposes the following hypotheses for empirical testing:

(1)

H1: A comprehensive policy strength index, constructed based on multi-dimensional characteristics, explains the development of the NEV market more significantly than the mere quantity of policies. This hypothesis directly challenges the fundamental flaw of the “policy counting method” prevalent in existing research [33].

(2)

H2: Different types of policies exert stronger targeted effects on their directly related output indicators. Here, a “targeted effect” is operationally defined as a policy mix designed for a specific objective (e.g., promoting sales, guiding production, or developing infrastructure) having a more significant impact on the corresponding output indicator for that objective (e.g., sales volume, production output, or number of charging piles).

(3)

H3: The impact of policy strength is subject to significant regional heterogeneity. Here, “regional heterogeneity” is operationally defined as the magnitude, and even the direction, of policy effects showing statistically significant differences contingent on key contextual factors of a province, such as its geographical location, level of economic development, and industrial base [38,39].

1.4. Contributions and Structure

This study aims to make contributions to the relevant field at the theoretical, methodological, and practical levels. Methodologically, the core contribution of this paper is the proposition and application of a multi-dimensional analytical framework, the policy strength index (PSI), designed to overcome the numerous limitations of existing research in policy quantification [40]. Theoretically, by empirically examining the heterogeneous effects of provincial policies, this research provides new micro-level evidence for understanding China’s unique central-local relations, policy implementation, and green transition governance, engaging in a dialogue with theories such as “fiscal federalism” and the “promotion tournament” [41,42]. Practically, the findings of this study can offer policymakers more targeted insights that go beyond a “one-size-fits-all” approach, thereby helping to optimize the allocation efficiency of public resources in promoting the green transition.

The overall research design and analytical path of this paper are systematically presented in Figure 1 (the technical roadmap of this study). The remainder of this paper is structured as follows: Section 2 details the research design, data sources, and methodology. Section 3 presents the empirical results. Section 4 provides an in-depth discussion of these results. Section 5 concludes the paper, clarifies its limitations, and offers prospects for future research.

2. Research Design, Data, and Methods

This section elaborates on the overall design and methodology of the study. Its core task is to construct an analytical framework capable of effectively quantifying the strength of provincial-level NEV policies in China. Based on this framework, an econometric model is specified to test the actual effects of policy strength on the development of the NEV industry.

2.1. Data Sources and Variable Definitions

This study constructs a panel dataset covering 30 provinces, municipalities, and autonomous regions in China (excluding Hong Kong, Macao, Taiwan, and Tibet due to data availability). The data primarily consist of three components: policy text data, NEV market development data, and provincial-level macroeconomic data.

2.1.1. Policy Text Data

The policy text data for this study were sourced from the “PKULaw” legal and regulatory database. To ensure the comprehensiveness of the policy collection, we used “new energy vehicle”, “electric vehicle”, and “charging pile” as keywords to search for all policy documents issued by central and local governments between 1 January 2013, and 31 December 2023. The year 2013 was chosen as the starting point because it is widely regarded as the beginning of the rapid development phase for China’s NEV industry. Extending the data to the end of 2023 provides support for testing time-lag effects in subsequent models.

After obtaining the initial policy texts, we conducted a rigorous data screening process. This involved removing non-policy documents such as news reports, meeting minutes, and leadership speeches; merging identically re-issued policies; and excluding policies not directly relevant to the research topic. Following this screening and organization, a final sample of 2455 valid policy documents was obtained. These texts form the foundation for the subsequent construction of the policy strength index (PSI).

2.1.2. Dependent Variables and Control Variables

To comprehensively measure the development level of provincial NEV markets, this study selects four key indicators as the core dependent variables, covering four dimensions: industrial supply, demand, infrastructure, and stock. These indicators are also commonly used in existing research when evaluating policy effects [43,44,45]. The detailed definitions and data sources for each dependent variable are presented in

Table 1. Definitions and data sources of dependent variables.

Variable	Definition	Time Span	Data Source
NEV Sales	Annual insured sales volume of new energy vehicles in each province	2016–2023	Souche Intelligence Cloud
NEV Production	Annual production volume of new energy vehicles in each province	2018–2023	Provincial Bureaus of Statistics, Statistical Yearbooks, government press releases, and authoritative media reports
Public Charging Pole Stock	Cumulative number of public charging poles in each province at year-end	2016–2023	China Electric Vehicle Charging Infrastructure Promotion Alliance (EVCIPA)
NEV Stock	Cumulative stock of new energy vehicles in each province at year-end	2017–2023	Provincial Bureaus of Statistics, Statistical Yearbooks, government press releases, and authoritative media reports

.

To more accurately isolate the net effect of policies and to avoid estimation bias arising from omitted variables, this study incorporates a series of provincial-level control variables into the model. The selection of these variables is based on established practices in existing NEV policy evaluation and regional economic studies [46]. All control variable data were obtained from the official website of the National Bureau of Statistics of China, covering the period from 2013 to 2023. The specific variable definitions are provided in

Table 2. Definitions and data sources of control variables.

Variable	Definition	Unit
Per Capita GDP	Gross regional product/Year-end permanent population	Yuan/person
Fiscal Expenditure	General budgetary expenditure of the local government, reflecting its intervention capacity and public service level	100 million Yuan
Secondary Industry Share	Value added of the secondary industry/Gross regional product	%
Tertiary Industry Share	Value added of the tertiary industry/Gross regional product	%
Electricity Generation	Reflects the regional energy supply capacity	100 million kWh
Highway Mileage	Reflects the level of regional transportation infrastructure	10,000 km
Per Capita Disposable Income	Reflects the purchasing power of residents	Yuan
Permanent Population	Year-end permanent population, reflecting the regional market size	10,000 persons
Household Car Ownership	Average number of household cars owned per 100 households at year-end	Vehicles/100 households

.

2.2. Analytical Framework of the Policy Strength Index (PSI)

To overcome the limitations of traditional policy quantification methods, such as the “all-or-nothing” binary classification or simple counting [47], this study constructs a multi-dimensional policy strength index (PSI). This framework adheres to the core principles of scientific rigor, objectivity, and multi-dimensionality, aiming to open the “black box” of policy effectiveness evaluation and provide a structured analytical tool for assessing the potential efficacy of policies.

2.2.1. Theoretical Framework and Indicator System for Index Construction

This study deconstructs policy strength into two core dimensions: A. extrinsic attributes of the policy and B. intrinsic attributes of the policy. Extrinsic attributes assess the “inherent” authoritativeness and legal force of a policy at the time of its issuance, features that do not depend on an in-depth reading of the specific text content [48]. Intrinsic attributes, conversely, delve into the policy text to evaluate the internal strength of its content, instruments, language, and goal-setting [49].

The strength of an effective policy stems not only from the hierarchical status of its issuing body but also from the rigor of its content design and the credibility of its resource commitments. Therefore, we have constructed a comprehensive indicator system comprising seven dimensions. Their specific definitions, theoretical underpinnings, and significance are detailed in

Table 3. PSI indicator system, definitions, and theoretical basis.

Dimension	Indicator Name	Definition & Measurement Logic	Theoretical Basis & Importance
A. Extrinsic Attributes	A1. Issuing Body Authority	Measures the administrative rank of the policy-issuing body. For documents jointly issued by multiple departments, the rank of the highest-level body is adopted.	The efficacy of a policy is rooted in the “legal-rational authority” of its issuing body. Directives from higher-level authorities possess greater mobilization capacity and binding force within the bureaucratic system, forming the fundamental source of policy strength [50].
	A2. Joint Issuance Breadth	Measures the number of departments participating in the joint issuance of a policy, processed using a logarithmic transformation.	The breadth of joint issuance reflects the degree of inter-departmental coordination and the consensus basis of the policy. It embodies the advantages of “collaborative governance” and can effectively overcome departmental obstacles during implementation to form policy synergy [51].
	A3. Document Type Efficacy	Measures the legal force of the policy document “genre,” e.g., “Measures” have a higher efficacy than “Circulars”.	In China’s administrative system, different document types carry varying degrees of legal binding force, which directly determines the policy’s level of compulsion and the seriousness of its enforcement.
B. Intrinsic Attributes	B1. Policy Instrument Strength	Measures the coerciveness of the core intervention tools adopted by the policy. For policies containing multiple instruments, the score of the strongest instrument is adopted.	Drawing on the classic framework of McDonnell & Elmore, policy instruments can be categorized into types such as mandates, inducements, and capacity-building, which differ significantly in their intrinsic strength and directly determine the policy’s intervention intensity [52].
	B2. Measure Intensity	Calculates the density of decisive words expressing compulsion, command, or prohibition within the policy text through text mining.	Policy language is a “symbolic political” act that conveys government resolve. Strong, decisive wording sends a high-cost, non-negotiable “policy signal,” enhancing the credibility of the policy [53].
	B3. Degree of Quantification	Calculates the density of words related to specific numerical values, deadlines, and other assessable targets within the policy text through text mining.	Clear, measurable objectives are the foundation of policy success and a prerequisite for subsequent monitoring and accountability. Quantified targets significantly increase the government’s accountability costs by setting explicit “performance pledges” [54].
	B4. Resource Assurance	Calculates the density of words within the policy text that explicitly promise fiscal, human, or organizational support and resources through text mining.	Policy implementation is impossible without resource support. Explicit resource commitments represent the “sunk costs” the government is willing to invest, serving as a vital signal of the policy’s feasibility and the government’s determination [55].

.

2.2.2. Indicator Quantification and Data Processing

To ensure the objectivity and replicability of the quantification process, this study assembled an expert panel consisting of three researchers familiar with Chinese industrial policy. The panel members first collaboratively developed a detailed coding manual and then independently conducted cross-coding of the 2455 policy texts. For any items with inconsistent codes, a consensus was reached through group discussion. The inter-coder reliability, tested using Kendall’s W, was 0.982, indicating a high degree of reliability in the coding results.

For A1 (Issuing Body Authority), we assigned scores as a 1-to-5 ordinal variable based on the administrative rank. Specifically, policies issued by the provincial-level Party Committee or Government (and their general offices) received the highest score of 5; provincial government departments (e.g., Development and Reform Commission, Department of Industry and Information Technology) received a 4; municipal-level Party Committees or Governments received a 3; municipal government departments received a 2; and policies issued by county-level and lower governments and departments were assigned the lowest score of 1.

For A2 (Joint Issuance Breadth), we quantify this dimension by calculating the total number of co-issuing departments, $N$, and applying a logarithmic transformation formula $A2\_Score=\ln (1+N)$. This approach is designed to capture the diminishing marginal influence of joint issuance, whereby the impact of increasing from one to two departments is far greater than that of increasing from ten to eleven.

For A3 (Document Type Efficacy), we assigned scores as a 1-to-5 ordinal variable based on the legal force of the document “genre” within the Chinese administrative system. “Legal/Regulatory/Mandatory” documents (e.g., “Measures”, “Provisions”) with the highest binding force were scored a 5; “Programmatic/Decision-making” documents (e.g., “Plans”, “Decisions”) were scored a 4; “Deployment/Executive” documents (e.g., “Implementation Plans”, “Circulars”) were scored a 3; “Guidance/Responsive” documents (e.g., “Guidelines”, “Catalogs”) were scored a 2; and “Informational” documents with the weakest force (e.g., “Announcements”, “Press Releases”) were scored a 1.

For B1 (Policy Instrument Strength), drawing on the framework of McDonnell & Elmore [52], we classified policy instruments into four categories based on their degree of coerciveness and assigned scores as a 1-to-5 ordinal variable. “Mandatory tools” (e.g., regulations, prohibitions) were scored a 5; “Incentive tools” (e.g., subsidies, tax breaks) were scored a 4; “Capacity-building tools” (e.g., technical support, talent development) were scored a 3; and “Symbolic and hortatory tools” (e.g., publicity campaigns, calls to action) were scored a 1. For policies containing multiple instruments, we adopted the score of the highest-strength instrument.

For the three text-based dimensions—B2 (Measure Intensity), B3 (Degree of Quantification), and B4 (Resource Assurance)—we measured their strength by calculating the density of specific keywords within the text to ensure objectivity. First, based on a pilot reading of policy texts and a review of relevant literature, the expert panel constructed a specialized keyword dictionary for each dimension. This dictionary underwent multiple rounds of discussion and revision to ensure its comprehensiveness and accuracy (the complete keyword list is detailed in Appendix A). Second, we used the jieba tokenization tool, loading a custom dictionary containing all keyword lists and domain-specific terms, along with a standard Chinese stop-word list, to segment the policy texts. Finally, the density score for each dimension was calculated using the formula

$$\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{f}\mathrm{r}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{y}\mathrm{o}\mathrm{f}\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{k}\mathrm{e}\mathrm{y}\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{s}}/{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}\mathrm{c}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}.$$

2.2.3. Index Synthesis: Objective Weighting Based on Principal Component Analysis (PCA)

After obtaining the baseline scores for the seven dimensions, this study employs principal component analysis (PCA) for objective weighting. The selection of PCA is predicated on three primary advantages. First, it ensures objective weighting. By determining the weights for combining the seven PSI dimensions into a composite index based on the data’s inherent statistical properties, PCA entirely eliminates researcher subjectivity, surpassing simpler methods like equal-weight averaging [56]. Second, it addresses multicollinearity. The seven constructed dimensions may be intercorrelated (e.g., policies issued by higher-level authorities may be more likely to use stronger language), and including them all directly in a regression model would introduce significant multicollinearity issues. By generating mutually orthogonal principal components, PCA effectively resolves this problem [57]. Third, it provides structured dimensionality reduction and academic validation. It condenses multiple correlated indicators into a few highly explanatory latent dimensions, which helps to structure the complex concept of policy and thereby provides a clearer framework for analyzing policy interventions [58].

PCA is a dimensionality reduction technique widely used in social science research. Its core objective is to transform a set of correlated original variables into a smaller set of uncorrelated composite variables, known as principal components, while retaining most of the information from the original variables [59]. The fundamental principle is to identify a new coordinate system wherein the first dimension (the first principal component) captures the maximum variance in the data. Subsequent components are orthogonal to the preceding ones and capture the next largest amount of variance. Each principal component is a linear combination of the original variables, and the associated weights (loadings) are determined entirely by the intrinsic structure of the data—specifically, by the contribution of each original variable to the total variance explained by that component.

In this study, the use of PCA offers three primary advantages. First, it provides objective weighting. It determines the weights for combining the seven PSI dimensions into a composite index based on the statistical properties inherent in the data, thereby entirely eliminating researcher subjectivity. Second, it addresses multicollinearity. The seven dimensions we constructed may be intercorrelated (e.g., policies issued by higher-level authorities may be more inclined to use stronger language), and including them all directly in a regression model would lead to significant multicollinearity issues. By generating mutually orthogonal principal components, PCA effectively resolves this problem. Third, it has strong academic acceptance. PCA is considered the “gold standard” for constructing composite indices in the social sciences and is widely used to measure complex, multidimensional concepts such as institutional quality and socioeconomic status, with its scientific validity and appropriateness having been fully endorsed by the academic community [60].

The procedural steps for implementing PCA in this study are as follows: First, the original matrix, which consists of 2455 policy samples and their scores across the seven dimensions, is standardized using Z-scores to eliminate differences in scale and value ranges among the dimensions. Second, prior to execution, the Kaiser-Meyer-Olkin (KMO) test and Bartlett’s test of sphericity will be conducted to assess the suitability of the data for factor analysis. Finally, PCA is performed on the standardized data matrix to extract the principal components. Based on their eigenvalues, explained variance ratios, and the loading matrix, a final composite index of policy strength is constructed for the econometric analysis.

2.3. Model Specification and Empirical Strategy

Having constructed the policy strength index (PSI), this section details the econometric model specified to evaluate its policy effects, along with a series of empirical strategies designed to enhance the robustness of the conclusions.

2.3.1. Benchmark Model: Two-Way Fixed Effects (TWFE)

To quantify the impact of provincial policy strength on the various output indicators of NEV industry development, this study constructs the following two-way fixed effects (TWFE) panel data econometric model:

$$Y_{it}={\beta }_0+\sum _{k=1}^K{\beta }_k\cdot {\mathrm{P}\mathrm{C}}_{k,i,t-l}+\gamma \prime Z_{it}+{\mu }_i+{\lambda }_t+{\epsilon }_{it}$$

where $i$ denotes the province and $t$ denotes the year. $Y_{it}$ is the dependent variable, representing the NEV market development level of province $i$ in year $t$. The core independent variable, ${\mathrm{P}\mathrm{C}}_{k,i,t-l }(k=\mathrm{1,2},\dots ,\mathrm{K})$, represents the score of the k-th principal component of policy strength for province $i$ in year t − l. This variable is calculated by summing the respective PC scores of all policies issued within a specific policy category (e.g., sales-side, production-side) in that year. The choice of a 1 to 3-year lag (l) is based on findings from related empirical studies, which suggest that a policy lag of 2–3 years is common for environmental and industrial policies to translate into measurable innovation or structural change [61,62]. To examine the time-lag characteristics of policy effects, we investigate various lag periods $l (l=\mathrm{0,1},\mathrm{2,3})$. $Z_{it}$ is a vector of time-varying provincial-level control variables.

The selection of the TWFE model as the benchmark is primarily based on the following considerations:

(1)

Province-fixed effects (${\mu }_i$): This is a core advantage of the model. It effectively controls for all time-invariant intrinsic characteristics of a province, such as geographical location, resource endowments, political traditions, and cultural atmosphere. By including province-fixed effects, the model’s basis of comparison shifts from being “between provinces” to “within each province over time” (i.e., comparing a province to its own history), which significantly mitigates the endogeneity problem arising from omitted variables.

(2)

Year-fixed effects (${\lambda }_t$): This effect controls for all factors that have a common impact on all provinces in the same year, such as national-level macroeconomic shocks, uniform national policy changes (e.g., the phase-out of national subsidies), or major technological breakthroughs. This ensures that the coefficients we estimate are attributable to differential provincial-level policies rather than to nationwide common trends.

Furthermore, to address potential issues of serial autocorrelation and heteroskedasticity within provinces, this study employs cluster-robust standard errors at the province level in all regressions, which is a standard practice in contemporary empirical research. We acknowledge that one of the core assumptions of this model, strict exogeneity, may be difficult to fully satisfy in reality. We will partially address this potential issue in subsequent endogeneity tests using strategies such as the instrumental variable approach.

2.3.2. Robustness and Endogeneity Test Strategy

To ensure the reliability of the benchmark regression conclusions, this study designs a systematic strategy for robustness and endogeneity checks.

Robustness checks: To verify that the benchmark regression results are not an artifact of specific methodological choices, this study will conduct systematic robustness checks on the model specification from multiple angles. These include replacing the aggregation method by using the mean instead of the sum of policy strength scores, replacing the index construction method with an equal-weighting approach that does not rely on PCA, and replacing the core explanatory variable with the traditional policy count metric for a comparative analysis [63,64]. Furthermore, to test whether our core findings on the impacts of different policy styles are robust to the number of principal components extracted, particularly if the KMO statistic is near the threshold, we will also perform a key sensitivity analysis. This involves re-running the benchmark regression models after forcing the PCA to extract a different number of principal components (specifically, two and four).

Endogeneity test: Although the TWFE model can address certain endogeneity issues, the estimates may still be affected by time-varying omitted variables or reverse causality. To address this potential problem, this study will employ the instrumental variable (IV) method, estimated using two-stage least squares (2SLS). Acknowledging the practical difficulty of finding strong instruments and the potential for estimation bias from weak ones, we will also conduct a comparative analysis using the limited information maximum likelihood (LIML) method, which is more robust to weak instruments, to more cautiously assess the impact of endogeneity. The selection rationale for the instrumental variables, the process of their validity testing, and the complete estimation results will be detailed in the results section.

2.3.3. Robustness and Endogeneity Test Strategy

To explore the “contextual dependency” of policy effects and to answer one of this study’s central questions—why policy effects “vary by place”—this research will employ subgroup analysis for an in-depth heterogeneity analysis. This method involves partitioning the total sample into several subsamples based on specific criteria and running the benchmark TWFE model independently within each subsample. This approach allows for a direct comparison of the differences in policy effects across different groups.

This study plans to partition the sample along core dimensions such as geographical region, level of economic development, and industrial base to systematically test whether policy effects are moderated by these key contextual factors.

3. Empirical Results

This chapter presents the core empirical findings of this study. We first report the descriptive statistics for the dimensions of the policy strength index (PSI) and present the results of the PCA to reveal the underlying structure of provincial NEV policies. Subsequently, we report the results from the benchmark regression, robustness checks, and heterogeneity analysis. All figures and tables herein are from the author’s original analysis.

3.1. Descriptive Statistics and PCA Results

3.1.1. Descriptive Statistics

After quantifying the 2455 policy texts, we obtained the base scores for the seven dimensional indicators within the PSI framework.

Table 4. Descriptive statistics for the PSI dimensional indicators.

Indicator	N	Mean	Sth. Dev.	Min	Max
A1. Issuing Body Authority	2455	2.99	1.03	1.00	5.00
A2. Joint Issuance Breadth	2455	0.86	0.37	0.69	3.40
A3. Document Type Efficacy	2455	3.18	0.81	1.00	5.00
B1. Policy Instrument Strength	2455	4.13	0.90	1.00	5.00
B2. Measure Intensity	2455	0.10	0.04	0.00	0.32
B3. Degree of Quantification	2455	0.12	0.08	0.00	0.61
B4. Resource Assurance	2455	0.10	0.05	0.00	0.37

presents the descriptive statistics for these indicators.

The mean values show that B1 (Policy Instrument Strength) has the highest score (mean = 4.13), indicating that the use of strong instruments such as incentives (score of 4) and mandates (score of 5) is a common practice in NEV policy. The mean scores for A1 (Issuing Body Authority) and A3 (Document Type Efficacy) are approximately 3.0 and 3.2, respectively, suggesting that policies are predominantly issued by municipal-level governments and are of a deployment/executive nature. The three text-mining-based density indicators—B2 (Measure Intensity), B3 (Degree of Quantification), and B4 (Resource Assurance)—have lower mean values, but their relatively large standard deviations reveal significant variation among policies in terms of linguistic intensity, target specificity, and resource commitment.

3.1.2. Principal Component Analysis

Before conducting the PCA, we first tested the suitability of the data for this method. The Kaiser-Meyer-Olkin (KMO) measure was 0.577, and the p-value for Bartlett’s test of sphericity was less than 0.001. These results indicate that a certain degree of correlation exists among the seven dimensional indicators, making the data suitable for PCA.

Figure 2a displays a heatmap of the correlation matrix among the seven dimensional indicators. It is evident that low-to-moderate correlations exist between the indicators. For instance, B4 (Resource Assurance) shows a significant positive correlation with A3 (Document Type Efficacy), B1 (Policy Instrument Strength), and B2 (Measure Intensity), with correlation coefficients of 0.27, 0.16, and 0.23, respectively. This suggests that policies with higher efficacy levels, stronger instruments, and more intense measures also tend to provide greater resource assurance. In contrast, B3 (Degree of Quantification) exhibits a negative correlation with most other dimensions, implying a potential trade-off between “target quantification” and other policy strength characteristics. The existence of these relationships provides a strong rationale for using PCA to address multicollinearity and extract composite dimensions.

Based on the classic Kaiser criterion (i.e., retaining principal components with eigenvalues greater than 1) and in conjunction with the position of the “elbow” in the scree plot in Figure 2b, this study decided to extract the first three principal components (PC1, PC2, and PC3). As shown in

Table 5. Eigenvalues and variance explained from principal component analysis.

Principal Component	Eigenvalue	Variance Explained (%)	Cumulative Variance (%)
PC1	1.695	24.21	24.21
PC2	1.161	16.58	40.79
PC3	1.060	15.14	55.93
PC4	0.953	13.60	69.54
PC5	0.802	11.45	80.99
PC6	0.729	10.41	91.40
PC7	0.603	8.60	100.00

, the eigenvalues for these three components are all greater than 1, and they collectively explain 55.93% of the total variance of the original seven dimensions, thus representing the majority of the information in the original data.

Table 6. Principal component loading matrix (rotated component matrix).

Indicator	PC1	PC2	PC3
A1. Issuing Body Authority	0.016	0.663	−0.334
A2. Joint Issuance Breadth	0.050	0.445	0.745
A3. Document Type Efficacy	0.580	0.124	0.384
B1. Policy Instrument Strength	0.625	0.111	−0.097
B2. Measure Intensity	0.635	−0.149	−0.386
B3. Degree of Quantification	−0.479	−0.456	0.179
B4. Resource Assurance	0.578	−0.516	0.238

Note: Loadings with an absolute value greater than 0.4 are shown in bold to highlight the main contributing factors.

displays the principal component loading matrix, which reveals the relationship between each principal component and the original seven dimensional indicators. This is key to understanding the substantive meaning of each component.

Based on the loading matrix, we interpret and name the three extracted principal components as follows:

(1)

PC1: Substantive-Driving Policy

This principal component has high positive loadings on A3 (Document Type Efficacy), B1 (Policy Instrument Strength), B2 (Measure Intensity), and B4 (Resource Assurance), while also having a significant negative loading on B3 (Degree of Quantification). This depicts a distinct policy “style”: these policies tend to use high-level, legally potent document types, employ strong and substantive intervention instruments, and are supported by decisive language and explicit resource guarantees. Simultaneously, however, they place less emphasis on setting specific, quantifiable performance targets. PC1 captures the “hardness” and “substance” of policy content. To qualitatively validate this interpretation, a typical example is the “Implementation Rules for the Hefei Municipal Bureau of Economy and Information Technology’s Executive Clauses on Several Policies for Further Promoting the Application of New Energy and Intelligent Connected Vehicles”. As an “implementation rule”, its core function is to specify higher-level policies. Its content includes a large number of directly operational subsidy amounts and application procedures (high policy instrument strength) and clearly defines the executive departments and funding sources (strong resource assurance), precisely embodying the characteristics of a substantive-driving policy.

(2)

PC2: High-Level Authoritative Policy

This principal component is predominantly driven by the two extrinsic attributes, A1 (Issuing Body Authority) and A2 (Joint Issuance Breadth), both of which have high positive loadings. This clearly points to policies that are led by high-ranking government bodies and formulated after extensive inter-departmental coordination. The strength of such policies lies in the authority of their “origin.” PC2 captures the “issuance level” and “political standing” of a policy. A typical example is the “Notice of the Anhui Provincial Committee of the Communist Party of China and the Anhui Provincial People’s Government on the Issuance of the ‘Anhui Province Carbon Peak Implementation Plan’.” This document was jointly issued by the province’s highest decision-making bodies—the Provincial Committee and the Provincial Government (extremely high issuing body authority). As a programmatic “implementation plan”, its content aims to unify thinking across the province and set strategic direction, serving as a prime example of a high-level authoritative policy.

(3)

PC3: Coordinative-Programmatic Policy

The interpretive structure of this principal component is relatively complex. Its most prominent features are the highest positive loading on A2 (Joint Issuance Breadth) and a significant negative loading on B2 (Measure Intensity). This portrays a type of policy that emphasizes cross-departmental collaboration but is not expressed in forceful terms regarding its content and measures. This likely represents programmatic policies, such as “plans”, “outlines”, or “guiding opinions”, which focus on coordinating the actions of various parties rather than on direct coercion. PC3 captures the “coordinative” and “programmatic” characteristics of a policy. For example, the “Notice of 15 Departments including the Department of Industry and Information Technology of Inner Mongolia Autonomous Region on the Issuance of the ‘Inner Mongolia Autonomous Region New Energy Vehicle High-Quality Development Implementation Plan (2023–2025)’” is representative of this type of policy. The document was jointly issued by 15 departments, reflecting an exceptionally high degree of coordination (high joint issuance breadth). Its nature as an “implementation plan” focuses on task decomposition and pathway planning, aiming to orchestrate the actions of all parties rather than to directly compel them, which aligns perfectly with the characteristics of a coordinative-programmatic policy.

In the subsequent analysis, we will use these three principal component indices (PC1, PC2, and PC3), now imbued with clear substantive meanings, as the core independent variables to test the impact of different types of policy strength on the development of the NEV market.

3.2. Baseline Regression Analysis

To systematically test the actual effects of the three types of policy strength indices constructed in this study (PC1: Substantive-Driving, PC2: High-Level Authoritative, and PC3: Coordinative-Programmatic) on the various output indicators of the provincial NEV market, and to deeply explore their time-lag characteristics, this section employs a TWFE panel data model for the baseline regression analysis.

3.2.1. Analytical Strategy

The analytical strategy of this study adheres to two main principles: “precision matching” and “time-lag testing”. First, to test hypothesis H2—that different policy instrument portfolios have stronger targeted effects on their directly related domains—we categorize the policy texts into “production-side”, “sales-side”, and “infrastructure-side” based on their core content. We then test their respective impacts on NEV production, sales, and charging infrastructure stock. Second, considering the time delay between policy enactment and its full effect, we examine the impact of the policy variables for each model in the concurrent period (t = 0) as well as with lags of one to three years (t − 1, t − 2, t − 3).

Before presenting a detailed analysis of the results, we first assessed the overall validity of all regression models. The results show that the F-statistic for every model is highly significant at the 1% level, indicating that the overall model specification is valid and that the independent variables jointly explain a substantial portion of the variation in the dependent variables.

3.2.2. Analysis of the Impact of Policy Strength on Domain-Specific Outputs

Next, we will report and interpret the regression results of the policy strength indices for each domain. It is important to note that since our core independent variables (PC1, PC2, PC3) are standardized principal components (mean = 0, standard deviation = 1), their regression coefficients have a straightforward economic interpretation. The magnitude of a coefficient measures the impact of a meaningful, typical change in policy strength. Specifically, when a province’s annual composite strength for a certain policy type increases by one standard deviation from the mean level, the corresponding dependent variable is expected to change, on average, by a certain percentage. This interpretation allows us to assess the real-world effect of a non-marginal, practically significant policy push.

(1) Impact on NEV Sales: Substantive-Driving Policy as the Core Driver

Table 7. Regression results of sales-side policies on NEV sales.

	(1) Contemporaneous	(2) Lag 1	(3) Lag 2	(4) Lag 3
Variables	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)
PC1_sum_Sales-Side	0.005 (0.007)	0.015 * (0.007)	0.010 *** (0.004)	0.004 (0.004)
PC2_sum_Sales-Side	−0.001 (0.007)	−0.005 (0.007)	−0.005 (0.006)	−0.003 (0.006)
PC3_sum_Sales-Side	0.003 (0.008)	0.003 (0.008)	0.003 (0.008)	0.002 (0.008)
Control Variables	Yes	Yes	Yes	Yes
Observations	248	248	248	248
R-squared	0.152	0.156	0.164	0.149
Province Fixed Effects	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes

Note: Robust standard errors clustered at the province level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Control variables are the same as in Table 3.

presents the complete regression results of sales-side policy strength on NEV sales (log_Sales). The core finding is that substantive-driving policy (PC1) is the key variable stimulating sales. Its impact exhibits a clear time lag, reaching its peak effect in the second lag period (t − 2), where it is highly significant at the 1% level. In terms of magnitude, a one-standard-deviation increase in the sales-side PC1 index is associated with an average increase in NEV sales of approximately 1.0% two years later. Although a 1.0% increase may seem modest numerically, given China’s vast NEV market base, this still translates to tens of thousands of additional vehicle sales annually, signifying substantial economic and policy relevance for achieving policy goals and shaping market expectations. This indicates that policies containing substantive content, such as strong subsidies, are most effective at driving market demand. In contrast, high-level authoritative (PC2) and coordinative-programmatic (PC3) policies show no statistically significant effect on sales.

(2) Impact on NEV Production: The Industrial Guidance Power of Coordinative-Programmatic Policy

Table 8. Regression results of production-side policies on NEV production.

	(1) Contemporaneous	(2) Lag 1	(3) Lag 2	(4) Lag 3
Variables	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)
PC1_sum_Production-Side	−0.003 (0.012)	−0.001 (0.012)	0.002 (0.012)	0.002 (0.013)
PC2_sum_Production-Side	0.003 (0.011)	0.003 (0.011)	0.003 (0.011)	0.003 (0.012)
PC3_sum_Production-Side	0.024 (0.016)	0.027 (0.016)	0.039 ** (0.015)	0.028 (0.016)
Control Variables	Yes	Yes	Yes	Yes
Observations	186	186	186	186
R-squared	0.176	0.181	0.190	0.182
Province Fixed Effects	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes

Note: Robust standard errors clustered at the province level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Control variables are the same as in Table 3.

presents the regression results of production-side policy strength on NEV production (log_Production). In contrast to sales, the significant driving force for NEV production comes from production-side coordinative-programmatic policy (PC3). Its effect is significantly positive at the 5% level in the second lag period (t − 2), with a coefficient of 0.039. This implies that a one-standard-deviation increase in the production-side PC3 index is associated with an average increase in NEV production of approximately 3.9% two years later. This finding suggests that guiding and expanding local NEV production capacity does not rely on short-term direct incentives but rather on forward-looking industrial planning. By coordinating key factors of production such as land, energy, talent, and supply chains, this type of policy provides a stable institutional environment and predictable policy expectations for firms to invest in building factories and expanding capacity. The effects of such policies require a longer period for planning, construction, and eventual commissioning, which is perfectly consistent with the time-lag characteristics observed in the results.

(3) Impact on Charging Infrastructure: The Unexpected Ineffectiveness of Authority and Planning

Table 9. Regression results of infrastructure-side policies on the stock of charging poles.

	(1) Contemporaneous	(2) Lag 1	(3) Lag 2	(4) Lag 3
Variables	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)
PC1_sum_Infrastructure-Side	0.002 (0.005)	0.002 (0.005)	0.002 (0.005)	0.002 (0.005)
PC2_sum_Infrastructure-Side	−0.013 ** (0.006)	−0.005 (0.006)	−0.004 (0.006)	−0.005 (0.006)
PC3_sum_Infrastructure-Side	−0.004 (0.007)	−0.009 (0.007)	−0.013 * (0.007)	−0.013 ** (0.007)
Control Variables	Yes	Yes	Yes	Yes
Observations	248	248	248	248
R-squared	0.209	0.203	0.200	0.200
Province Fixed Effects	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes

Note: Robust standard errors clustered at the province level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Control variables are the same as in Table 3.

presents the regression results of infrastructure-side policies on the stock of public charging poles (log_Chargers), revealing one of the most surprising findings of this study. High-level authoritative policy (PC2) has a significant negative impact on the number of charging poles in the concurrent period (p < 0.05). Meanwhile, coordinative-programmatic policy (PC3) exhibits a sustained and significant negative impact in the second and third lag periods.

(4) Impact on NEV Stock: The Long-Term Cumulative Effect of Multiple Policies

As the ultimate stock indicator, the total stock of NEVs is an accumulation of annual increments, and the factors influencing it are consequently more complex. The regression results in

Table 10. Regression results of all policies on NEV stock (selected variables).

	(1) Contemporaneous	(2) Lag 1	(3) Lag 2	(4) Lag 3
Variables	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)	Coef. (Std. Err.)
L2.PC1_sum_Sales-Side	…	…	0.011* (0.006)	…
PC2_sum_Sales-Side	−0.011 * (0.006)	…	…	…
L1_PC2_sum_Infrastructure-Side	…	0.008 * (0.004)	…	…
L2_PC3_sum_Sales-Side	…	…	0.015 * (0.008)	…
L3_PC3_sum_Sales-Side	…	…	…	0.014 * (0.008)
Control Variables	Yes	Yes	Yes	Yes
Observations	217	217	217	217
R-squared	0.277	0.305	0.293	0.281
Province Fixed Effects	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes

Note: To highlight the core findings, only selected significant policy variables are listed. Robust standard errors clustered at the province level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Control variables are the same as in Table 3.

show that NEV stock is affected by multiple influences from different policy types, different domains, and with different time lags. Both sales-side substantive-driving (PC1) and coordinative-programmatic (PC3) policies show a significant positive impact on NEV stock after a lag of 2–3 years. This clearly indicates that the effects of policies promoting current-period sales accumulate over time, ultimately manifesting as growth in the total stock.

Taken together, the baseline regressions reveal a “goal-matching” phenomenon between different policy styles and market outputs: substantive-driving policies (PC1) primarily affect the demand side (sales), whereas coordinative-programmatic policies (PC3) mainly influence the supply side (production). However, the counter-intuitive negative effects observed in the charging infrastructure domain expose a potential gap between policy intent and actual outcomes. This result may stem from several mechanisms, such as a “planning fallacy” disconnected from market realities, an “authority failure” where high-level directives encounter local resistance during implementation, or a misallocation of resources caused by a mismatch between policy instruments and local governance capacity. Furthermore, this anomalous result could also be heavily influenced by potential endogeneity issues. We will conduct a more in-depth analysis of these possibilities in the discussion in Section 4.

3.3. Robustness Checks

To ensure that the conclusions drawn from the baseline regressions are reliable and stable, rather than being artifacts of a specific index construction method or model specification, this section implements a series of rigorous robustness checks. Our core strategy is to systematically alter the construction of the core explanatory variables or the model specification for all policy effects that were significant in the baseline regressions, and then observe whether the original conclusions still hold.

3.3.1. Design of the Robustness Checks

We designed the following six types of robustness checks to challenge the baseline model’s specification from different perspectives:

(1) Alternative aggregation rule: The baseline model uses the annual sum of the policy strength index scores to measure the “total policy effort.” This check replaces the sum with the annual mean to investigate whether market development is driven by the “total effort” or the “average intensity” of policies.

(2) Alternative index construction methods (Equal-weighted PSI and Hierarchical Equal-weighted PSI): To test the dependency of the results on the weighting method, we construct an “equal-weighted PSI” by taking a simple average of the seven standardized dimensions, completely independent of PCA, and re-run the regressions. The hierarchical equal-weighted PSI is a theoretical refinement of this approach. We first group the seven dimensions into two layers, “A. Extrinsic Attributes” and “B. Intrinsic Attributes,” calculate the mean within each layer, and then create the “hierarchical equal-weighted PSI” by taking a weighted sum of the two layer scores.

(3) Quantifying the explanatory contribution of PSI: To directly address the core motivation of this study and quantify the explanatory advantage of the PSI, we replace our composite PSI entirely with the policy count metric, which is widely used in the literature. We then compare the change in the adjusted R-squared between the two models to visually demonstrate the additional explanatory power provided by the PSI beyond mere policy “quantity.”

(4) Disentangling policy “quality” and “quantity”: This is the most stringent test. We include both the PSI and the policy count variable simultaneously in the baseline model to clearly isolate the independent effects of policy “quality” and “quantity.”

(5) PCA sensitivity analysis: To test the robustness of our core findings to the number of principal components extracted, especially given that our KMO statistic was borderline, we conduct a sensitivity analysis. This test is implemented by forcing the PCA to extract a different number of principal components (specifically, two and four) and re-running the baseline regression models.

(6) Controlling for province-specific time trends: To rule out potential confounding from unobserved, linearly time-varying factors unique to each province (such as distinct industrial upgrading paths or continuous evolution of policy styles), we add province-specific linear trends (an interaction term between province dummies and a time trend) to the baseline model. If the core findings remain robust after controlling for these heterogeneous trends, the credibility of our conclusions will be substantially enhanced.

3.3.2. Robustness Check Results and Analysis

We applied the aforementioned robustness checks to all 11 significant policy effects identified in the baseline regressions.

Table 11. Summary table of robustness check results for key findings.

Baseline Finding (Independent → Dependent Variable)	(0) Baseline Model	(1) Replace with Mean Aggregation	(2) Equal-Weighted PSI	(3) Hierarchical-Weighted PSI	(4) Replace with Policy Count	(5) PSI + Policy Count
Panel A: Impact on NEV Sales
PC1 → Sales (Lag 1)	0.015 *	0.029	0.021	0.021	−0.014	0.011 *
PC1 → Sales (Lag 2)	0.010 ***	0.044	0.043 **	0.045 **	−0.009	0.010 ***
Panel B: Impact on NEV Production
PC3 → Production (Lag 2)	0.039 **	0.044	−0.041	−0.030	0.015	0.051 ***
Panel C: Impact on Charging Infrastructure
PC2 → Chargers (Contemporaneous)	−0.013 **	0.007	0.007	0.001	−0.002	−0.014 **
PC3 → Chargers (Lag 2)	−0.013 *	−0.067 *	−0.040 **	−0.039 **	−0.004	−0.011
PC3 → Chargers (Lag 3)	−0.013 **	−0.076*	−0.021	−0.018	0.001	−0.015 **
Panel D: Impact on NEV Stock
PC1 → Stock (Lag 2)	0.011 *	—	—	—	—	—
PC2 → Stock (Contemp., Sales-Side)	−0.011 *	0.003	−0.014	−0.017	−0.007	−0.010 *
PC2 → Stock (Lag 1, Infra-Side)	0.008 *	0.019	0.017	0.018	0.002	0.006
PC3 → Stock (Lag 2, Sales-Side)	0.015 *	0.002	0.022	0.022	−0.016 **	0.007
PC3 → Stock (Lag 3, Sales-Side)	0.014 *	0.017	0.025 ***	0.026 ***	−0.003	0.008

Note:—indicates that the test was not performed or is not applicable. Cells contain the regression coefficients for the key policy variables. *** p < 0.01, ** p < 0.05, * p < 0.1.

systematically summarizes the core results of all these tests.

First, a series of checks based on alternative indicator construction and aggregation rules (detailed in Table 11) indicates that the core findings of this study are highly robust. As seen in columns (2), (3), and (5) of Table 11, the policy effects we identified (such as the positive effect of PC1 on sales and PC3 on production) persist regardless of whether an equal-weighted or hierarchical equal-weighted PSI is used, or when controlling for policy quantity in the model. The coefficient signs and significance levels remain largely unchanged.

The most critical robustness check in this study is the one that distinguishes between policy “quality” and “quantity.” As shown in column (4) of Table 11, when using only the traditional policy count as the core explanatory variable, nearly all the significant results from the baseline model disappear. However, in the model in column (5), where both the PSI and the policy count are included, the magnitude and significance of the PSI coefficients from the baseline model remain fundamentally stable. This provides compelling evidence that the policy “quality” and “style” captured by our constructed PSI are the truly effective factors influencing market outcomes, independent of mere policy “quantity,” thus offering strong validation for Hypothesis H1.

Furthermore, regarding the impact of the aggregation method, the results in column (1) of Table 11 show that when the “mean” of policy strength is used instead of the “sum,” the significance of some results diminishes. This suggests that, for the market, the total policy effort within a year (sum) may be more important than the average intensity of individual policies (mean).

To further enhance the reliability of our conclusions, we conducted three additional, more stringent robustness checks, the results of which are reported below.

(1) PCA Sensitivity Analysis

To test the robustness of our core findings to the number of principal components extracted, especially given that our KMO statistic was borderline, we conducted a sensitivity analysis. This test was implemented by forcing the PCA to extract a different number of principal components (specifically, two and four) and re-running the baseline regression models. As shown in

Table 12. Summary of PCA sensitivity analysis results.

Dependent Variable	Model Specification	Core Explanatory Variable	Coefficient	p-Value	Conclusion
Panel A: Impact on NEV Sales	Baseline Model (3 PCs)	Sales-Side Substantive-Driving Policy (PC1)	0.010	(<0.01)	Baseline Finding
	Robustness Check (2 PCs)	Sales-Side Substantive-Driving Policy (PC from 2-PC Model)	0.014	(<0.01)	Robust
	Robustness Check (4 PCs)	Sales-Side Substantive-Driving Policy (PC from 4-PC Model)	0.013	(<0.01)	Robust
Panel B: Impact on NEV Production	Baseline Model (3 PCs)	Production-Side Coordinative-Programmatic Policy (PC3)	0.039	(<0.05)	Baseline Finding
	Robustness Check (2 PCs)	Production-Side Substantive-Driving Policy (PC from 2-PC Model)	−0.004	(0.874)	Effect attenuated
	Robustness Check (4 PCs)	Production-Side Coordinative-Programmatic Policy (PC from 4-PC Model)	0.041	(<0.01)	Robust

Note: This table summarizes the regression results of log(NEV Sales) and log(NEV Production) on the core policy principal components, lagged by two periods. Baseline model results are from Table 7 and Table 8.

, our core findings exhibit a high degree of robustness. Regarding the impact of policy on NEV sales (Panel A), the conclusion is very stable: under both the 2-component and 4-component specifications, the core principal component representing “substantive incentives” consistently has a statistically significant (p < 0.01) positive impact on NEV sales with a two-year lag. Regarding the impact on NEV production (Panel B), the conclusion is also strongly supported: under the more granular 4-component specification, a new principal component perfectly replicates the positive effect of the “coordinative-programmatic” policy on production found in the baseline model.

(2) Incremental Explanatory Power of the PSI

To quantify the superiority of the PSI constructed in this study compared to the traditional “policy count” metric, we conducted a nested model comparison test, using the within R-squared as the core measure of model fit. As shown in

Table 13. Incremental within R-squared contribution of the PSI.

Model Description	R2 (within) (Policy Count Only)	R2 (within) (PSI + Policy Count)	Incremental R2 (2)−(1)	Relative Increase in Explanatory Power (3)/(1)
Model for Sales	0.0369	0.0676	+0.0307	+83.2%
Model for Production	0.0122	0.0579	+0.0457	+374.6%

Note: This table compares models of the impact of policy variables, lagged by two periods, on NEV sales and production. The only control variable included is the logarithm of the permanent population.

, our PSI significantly enhances the model’s explanatory power. In the model for sales, the inclusion of the PSI variable leads to a relative increase of 83.2% in the within R-squared. This effect is even more pronounced in the model for production, where the model’s explanatory power increases by more than 3.7-fold. This result again demonstrates that the PSI, which measures policy “quality” and “style,” explains the driving factors of NEV market development across China’s provinces much more effectively than the crude “policy count” metric.

(3) Controlling for Province-Specific Time Trends

For the most stringent test, we augmented the baseline model by incorporating province-specific time trends to control for the distinct, linearly time-varying developmental trajectories of each province. As shown in

Table 14. Robustness check results after controlling for province-specific trends.

Dependent Variable	Model Specification	Core Explanatory Variable	Coefficient	p-Value	Conclusion
Panel A: Impact on NEV Sales	(1) Baseline Model	Sales-Side Substantive-Driving Policy (PC1)—Lag 2	0.010	(<0.01)	Baseline Finding
	(2) + Province-Specific Trends	Sales-Side Substantive-Driving Policy (PC1)—Lag 2	0.0079	0.246	No longer significant
Panel B: Impact on NEV Production	(1) Baseline Model	Production-Side Coordinative-Programmatic Policy (PC3)—Lag 2	0.039	(<0.05)	Baseline Finding
	(2) + Province-Specific Trends	Production-Side Coordinative-Programmatic Policy (PC3)—Lag 2	0.0171	0.200	No longer significant

Note: This table compares the results of the baseline model with the robustness check model that includes province-specific trends.

, after including this strict control, the previously significant core policy variables are no longer statistically significant. This result does not negate our prior findings but rather provides a more profound interpretation: the policy effects we identified are largely intertwined with each province’s unique development trends and growth “inertia.” In other words, policy effectiveness may be highly dependent on a locality’s existing development path, with policy playing more of a role in accommodating or accelerating these trends rather than acting as an independent, new driver of growth. This finding is itself highly insightful, as it reveals the complex interplay between policy and local heterogeneity, providing strong evidence for understanding the “contextual dependency” of policy effects.

In summary, through a series of rigorous robustness checks, the core conclusions of this study are strongly supported. The empirical results clearly demonstrate that policy “quality”—encompassing its content, instruments, and authoritativeness—rather than mere “quantity”, is the key determinant of policy effectiveness. These checks significantly enhance the internal validity and credibility of the study’s conclusions.

3.4. Endogeneity Test

Although the TWFE model can control for time-invariant provincial heterogeneity and time-varying common shocks, it may still face endogeneity problems arising from time-varying omitted variables or reverse causality. Strong development momentum in a province’s NEV market (the dependent variable) might, in turn, prompt the local government to introduce more and stronger supportive policies (the independent variable), leading to an overestimation of the policy’s effect. To address this potential issue, this section employs the instrumental variable (IV) method for in-depth diagnosis and analysis.

3.4.1. Construction and Theoretical Basis of Instrumental Variables

Finding an ideal instrumental variable in policy evaluation research is exceptionally challenging. This study adopts a “wide-net” strategy, systematically constructing a total of ten instrumental variables across four major categories based on different theoretical logics, with the aim of screening for relatively effective IVs from this set.

(1)

Geographical and spatial spillover effects: This class of IVs is based on the theories of “policy diffusion” and “yardstick competition,” which posit that a province’s policymaking is influenced by its geographical neighbors [42]. We construct the average of neighboring provinces’ policies (IV_NeighborAvg), the sum of neighboring provinces’ policies (IV_NeighborSum), the lagged average of neighboring provinces’ policies (IV_NeighborLag), and the policy of the best-performing neighbor (IV_BestNeighbor). It must be acknowledged that the exogeneity of such spatial instruments is not perfect. The core challenge lies in the “exclusion restriction”, as it is difficult to completely rule out the possibility that neighboring provinces’ policies directly affect the home province’s market through channels other than policy diffusion, such as regional supply chain integration or cross-province consumption. Nevertheless, this remains one of the most commonly used and widely accepted identification strategies in current macroeconomic policy evaluation.

(2)

Central-local interactions and institutional arrangements: This class of IVs is based on China’s unique central-local relations and institutional context. We construct a classic Bartik instrument (IV_Bartik) (an interaction term between a province’s share of the national automobile industry output value in 2010 and the number of central-level NEV policies in a given year), the count of central-level policies (IV_National_Count), and a trend variable for pilot provinces based on the “Ten Cities, One Thousand Vehicles” demonstration program (IV_Pilot_Trend) [65].

(3)

Intra-provincial political structure: We construct instruments based on the policy strength of the provincial capital city (IV_Capital_PC1, IV_Capital_PC2, IV_Capital_PC3). The logic is that the policy direction of the provincial capital, as the political center, often leads or reflects the policy orientation of the entire province.

3.4.2. First-Stage Test: The Pervasive Weak Instrument Problem

The validity of the instrumental variable (IV) method primarily depends on the correlation between the IV and the endogenous variable. This is typically assessed using the F-statistic from the first-stage regression. The widely accepted rule of thumb in academia is that the F-statistic should be greater than 10 to consider the instrument sufficiently strong [66].

Table 15. Summary of first-stage F-statistics for all instrumental variable strategies.

IV Strategy	No. of Tests	Mean F-Statistic	Max F-Statistic	No. of Tests with F > 10	No. of Tests with F > 5
Neighboring Provinces’ Policy Average (NeighborAvg)	11	3.07	9.16	0	2
Neighboring Provinces’ Policy Sum (NeighborSum)	11	2.76	6.70	0	3
Pilot Province Trend (Pilot_Trend)	11	1.02	4.39	0	0
Best Neighbor’s Policy (BestNeighbor)	11	0.54	2.18	0	0
Lagged Neighboring Policy (NeighborLag)	11	1.02	2.12	0	0
Bartik Instrument (Bartik)	11	1.13	4.61	0	0
Provincial Capital’s Policy (Capital_PC1/2/3)	33	1.32	3.00	0	0
National Policy Count (National_Count)	11	0.24	0.83	0	0
Total	110	1.51	9.16	0	5

Note: Detailed first-stage regression results for the instrumental variable method are provided in Appendix B.

summarizes the first-stage F-statistics for all instrumental variable strategies. The results clearly indicate that among the 110 tests conducted to match instrumental variables for the 11 significant findings in the baseline model, not a single instrumental variable’s first-stage F-statistic meets the conventional threshold of 10. The highest F-value obtained was merely 9.16.

This result provides strong evidence that finding a powerful instrumental variable that simultaneously satisfies the relevance and exogeneity requirements is extremely difficult in the complex domain of provincial new energy vehicle policy. Nevertheless, we note that the F-statistics for five models exceeded the threshold of 5. Although these still fall into the category of weak instruments, we decided to conduct a more in-depth robustness diagnosis on these five relatively best-performing models to investigate the specific impact of the weak instrument problem on the estimation results.

3.4.3. Addressing Weak Instruments: Contradictory Results from 2SLS and LIML Estimations

In the presence of weak instruments, the standard 2SLS estimator can be severely biased. To obtain a more reliable diagnosis, we employ the limited information maximum likelihood (LIML) method, which is theoretically superior in weak instrument situations, for a comparative analysis.

Table 16. Comparison of 2SLS and LIML estimation results with weak instruments.

Dependent Variable	Log (Public Charging Pole Stock)	Log (Public Charging Pole Stock)	Log (NEV Stock)	Log (NEV Stock)	Log (NEV Stock)
Endogenous Variable (Policy Type)	Coordinative-Programmatic (PC3)—Lag 3	Coordinative-Programmatic (PC3)—Lag 3	High-Level Authoritative (PC2)—Contemporaneous	Coordinative-Programmatic (PC3)—Lag 2	Coordinative-Programmatic (PC3)—Lag 2
Instrumental Variable	Neighboring Provinces’ Policy Average	Neighboring Provinces’ Policy Sum	Neighboring Provinces’ Policy Sum	Neighboring Provinces’ Policy Average	Neighboring Provinces’ Policy Sum
First-Stage F-statistic	9.16	5.73	5.32	9.07	6.70
2SLS Coefficient (p-value)	0.058 (0.080)	0.059 (0.227)	−0.073 (0.065)	0.058 (0.011)	0.062 (0.062)
LIML Coefficient (p-value)	−0.626 (0.177)	−0.726 (0.177)	−0.161 (0.270)	−0.329 (0.285)	−0.621 (0.285)

presents the detailed regression results for the five models that passed the F >5 test.

The results are interpreted as follows:

(1)

Highly unstable and contradictory estimates: A striking phenomenon is observed in Table 16: in all five models, the signs of the 2SLS and LIML estimated coefficients are completely opposite. For instance, in Model (1) and Model (4), which have the highest F-statistics, 2SLS yields weakly significant or significant positive coefficients, whereas the LIML estimates—which are more robust to weak IVs—are negative and much larger in absolute value.

(2)

Significant p-values are illusory: In Model (4), 2SLS produces a result that is highly statistically significant (p = 0.011). However, given that the corresponding LIML result has the opposite sign and is entirely insignificant, this significant p-value is almost certainly a case of spurious inference caused by a weak instrument and carries no meaningful economic interpretation.

(3)

Overall failure of the IV strategy: The vast differences and complete contradiction in direction between the 2SLS and LIML results are classic symptoms of a total failure in estimation caused by weak instruments. This indicates that even the relatively best-performing instrumental variables are far from adequate for making reliable causal inferences. We cannot select any single result from this analysis to either support or refute the findings of the baseline model.

3.4.4. Summary of the Endogeneity Test

In summary, this exhaustive instrumental variable analysis clearly indicates that we were unable to find a valid instrument to resolve the potential endogeneity problem. All constructed IVs suffer from a severe weak correlation problem, rendering the second-stage estimates unreliable, highly unstable, and even leading to the extreme case where the 2SLS and LIML results are in complete opposition.

Therefore, this study cannot make stronger claims about the causal effects of the policies based on the IV method. The results from the baseline model (TWFE) should be interpreted with caution as correlations rather than as strict causal relationships. This “null result” is, in itself, an important and cautionary finding for the entire field of macroeconomic policy evaluation. It provides detailed evidence demonstrating the profound difficulty of finding a valid instrumental variable that meets textbook standards in the complex policy environment at the provincial level. This suggests that future research must exercise extreme caution when heavily relying on the IV method for causal inference and may need to shift focus toward research designs that offer clearer quasi-natural experimental settings.

3.5. Heterogeneity Analysis

The baseline regression analysis revealed the average effects of different policy types on the development of the NEV market. However, this average effect may mask significant variations in how policies perform across different regions and under different conditions. To investigate the “contextual dependency” of policy effects and to answer one of this study’s central questions—why policy effects “vary by place”—this section conducts an in-depth heterogeneity analysis.

3.5.1. Dimensions and Methods for Heterogeneity Analysis

We group the sample along three core dimensions to test whether policy effects are moderated by key regional characteristics. Our primary analytical method is subgroup analysis, which involves re-running the baseline two-way fixed effects model within each subsample. By comparing the magnitude and significance of the regression coefficients for the same policy variable across different subsamples, we can determine whether policy effects exhibit significant heterogeneity. The specific grouping criteria are as follows:

(1)

Geographic region: Following the official standards of China’s National Bureau of Statistics (NBS), we divide the 31 provincial-level administrative units into four major regions: Eastern, Central, Western, and Northeastern. This is intended to test whether policy effects are influenced by macro-level factors such as national regional development strategies, degree of marketization, level of foreign exposure, and geographical location.

(2)

Economic development level: To avoid endogeneity (i.e., the current year’s economic level could be influenced by the current year’s policies), we use the per capita gross regional product at the beginning of the sample period (2016) as the basis for division. We rank the 31 provinces by this indicator and divide them into three equal groups (terciles): high-income, middle-income, and low-income. This aims to test whether policy effectiveness is contingent on the local economic base, residents’ purchasing power, market maturity, and governmental fiscal capacity.

(3)

Industrial foundation: Based on historical data of each province’s automobile industry output value in 2010, we classify the provinces into three groups: those with a strong, medium, or weak automobile industry. This dimension is used to examine whether a locality’s industrial endowment and historical path dependence moderate policy effectiveness.

3.5.2. Heterogeneity Analysis Results

Table 17. Detailed results of the heterogeneity analysis of policy effects.

Dependent Variable	Policy Variable (Lag)	Grouping Type	Group	Coefficient	Std. Err.	N
Panel A: NEV Sales (log_Sales)	PC1_Sales-Side (Lag 1)	Economic Level	Middle-Income	0.019 *	(0.011)	88
	PC1_Sales-Side (Lag 2)	Geographic Region	Western	0.012 *	(0.007)	96
		Economic Level	High-Income	0.008 **	(0.004)	80
		High-Income	Strong Auto Industry	0.018 **	(0.007)	80
Panel B: NEV Production (log_Production)	PC3_Production-Side (Lag 2)	Geographic Region	Western	0.075 *	(0.042)	72
Panel C: Charging Poles (log_Chargers)	PC2_Infrastructure-Side (Current)	Geographic Region	Central	−0.024 ***	(0.007)	48
		Industrial Foundation	Medium Auto Industry	−0.015 *	(0.009)	80
	PC3_Infrastructure-Side (Lag 2)	Geographic Region	Central	−0.015 **	(0.007)	48
		Economic Level	Middle-Income	−0.031 *	(0.018)	88
	PC3_Infrastructure-Side (Lag 3)	Geographic Region	Eastern	−0.012 **	(0.006)	80
		Economic Level	High-Income	−0.013 *	(0.007)	80
Panel D: NEV Stock (log_NEV_Stock)	PC2_Sales-Side (Current)	Economic Level	Low-Income	−0.020 ***	(0.006)	70
	PC2_Infrastructure-Side (Lag 1)	Industrial Foundation	Weak Auto Industry	0.011 *	(0.006)	77
	PC1_Sales-Side (Lag 2)	Geographic Region	Eastern	0.005 *	(0.003)	70
		Economic Level	High-Income	0.005 *	(0.003)	70
		Industrial Foundation	Strong Auto Industry	0.009 ***	(0.002)	70
	PC3_Sales-Side (Lag 3)	Economic Level	Low-Income	−0.019 **	(0.008)	70
		Economic Level	High-Income	0.014 *	(0.008)	70
		Industrial Foundation	Weak Auto Industry	−0.018 *	(0.010)	77

Note: This table displays only the results from the subgroup regressions that are significant at least at the 10% level. Robust standard errors clustered at the province level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

reports the detailed results of the subgroup regressions for the 11 significant baseline models. To facilitate a more intuitive understanding of these differences, we also provide visualizations of the results using coefficient plots and heatmaps.

Based on the coefficient plot in Figure 3 and the heatmap in Figure 4, we derive the following core findings:

(1)

The negative effect on charging poles is concentrated in the Central and Eastern regions. The heterogeneity analysis clarifies the source of the negative effect of coordinative-programmatic policies (PC3) on charging pole construction that was identified in the baseline regression. As shown in the coefficient plot (Figure 3), this negative effect is primarily concentrated in the Eastern region (coefficient = −0.012, p < 0.05). Similarly, the negative effect of high-level authoritative policies (PC2) on charging poles is most pronounced in the Central region (coefficient = −0.024, p < 0.01). This suggests that for infrastructure like charging poles, which have long investment return cycles, grand plans or authoritative directives may easily become detached from market realities, thereby producing an inhibitory effect. This is particularly true in the Central region, which faces significant economic transition pressure but has a lower degree of marketization than the East, and in the Eastern region, where factor costs are extremely high.

(2)

The positive effects on NEV stock and sales are more pronounced in high-income provinces and those with a strong automobile industry. The promotional effect of policy on market development clearly exhibits a “Matthew effect” where the strong grow stronger. The positive impact of substantive-driving policies (PC1) on NEV stock is almost entirely concentrated in provinces with a strong auto industry (coefficient = 0.009, p < 0.01), while being insignificant in other provinces. Likewise, the effect of this policy on sales is most significant in high-income regions (coefficient = 0.008, p < 0.05) and in provinces with a strong auto industry (coefficient = 0.018, p < 0.01). This indicates that in regions with a solid industrial foundation and stronger consumer purchasing power, powerful incentive policies can be more effectively translated into growth in both the stock and flow of the market.

(3)

Insights from the Heatmap: A “Central Collapse” and Western Opportunities. The heatmap in Figure 4 provides a more comprehensive regional picture of policy effectiveness. The large red areas in the map (especially in the Central and middle-income regions) indicate that many policies that appear effective at the national level are substantially less effective, or even have negative effects, in these specific regions. For example, regarding charging pole construction, the effectiveness of nearly all policy types is significantly worse than the national average in Central and low-to-middle-income regions. In stark contrast, the few blue areas on the map identify the “sweet spots” where policies achieve their maximum utility. Most notably, the positive effect of production-side coordinative-programmatic policies (PC3) on production is greatly enhanced in the Western region (a 121% increase in the coefficient relative to the baseline). This may be highly consistent with the strategic need of the Western region to undertake industrial transfers and develop emerging industries.

3.5.3. Testing Heterogeneity Mechanisms: Interaction Effect Analysis

The preceding subgroup analysis described the heterogeneity of policy effects across different regions. To deepen the analysis from “description” to “explanation”, this section directly tests the driving mechanisms behind this heterogeneity. We introduce interaction terms between policy strength and provincial characteristics into the baseline model, aiming to test whether key contextual factors at the provincial level—such as economic development, industrial structure, and fiscal capacity—systematically moderate the effectiveness of policies.

(1) Interaction Term Construction and Mechanism Variables

Building upon the baseline two-way fixed effects model, we add an interaction term between the core policy variable and a provincial-level moderating variable. The model is specified as follows:

$$Y_{it}={\beta }_1\cdot {\mathrm{P}\mathrm{o}\mathrm{l}\mathrm{i}\mathrm{c}\mathrm{y}}_{i,t-l}+{\beta }_2\cdot {\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r}}_{it}+{\beta }_3\cdot ({\mathrm{P}\mathrm{o}\mathrm{l}\mathrm{i}\mathrm{c}\mathrm{y}}_{i,t-l}\times {\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r}}_{it})+\gamma \prime Z_{it}+{\mu }_i+{\lambda }_t+{ϵ}_{it}$$

Here, the coefficient of the interaction term ($\mathrm{P}\mathrm{o}\mathrm{l}\mathrm{i}\mathrm{c}\mathrm{y}$×$\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r}$), ${\beta }_3$, is our primary focus. A significant ${\beta }_3$ indicates that the provincial characteristic ($\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r}$) has a significant moderating effect on the policy’s impact. We selected the following six proxy variables to serve as moderators:

• Per Capita GDP (log_PerCapitaGDP): The logarithm of per capita gross regional product is used to directly measure the local level of economic development and the average purchasing power of residents.
• Industrial Structure (Share_SecondaryIndustry): The share of the value-added of the secondary industry in the gross regional product is used as a proxy for a region’s level of industrialization and traditional industrial base.
• Commercial Factor Cost (log_LandCost_Proxy): The logarithm of the “average sales price of commercial business buildings” is used as a proxy variable to reflect the factor costs of local commercial activities.
• Provincial Fiscal Space (Fiscal_Expenditure): The logarithm of local government general budgetary expenditure is used as a proxy for the local government’s capacity to translate policy intentions into actual public resource investment.
• Regional Electricity Load (log_PowerConsumption_Proxy): The logarithm of the “total electricity consumption of the whole society” is used as a proxy for the grid capacity, reflecting the actual electricity demand and load level in the region.
• Share of SOE Assets in Industry (SOE_AssetShare_Proxy): Calculated by dividing the “total assets of state-controlled industrial enterprises” by the “total assets of industrial enterprises above a designated size,” this serves as a proxy for the influence of the state-owned economy in the local industrial system.

(2) Interaction Effect Analysis Results

After conducting a total of 66 sets of interaction term regressions (11 significant baseline models and 6 mechanism variables), we identified 6 statistically significant interaction effects.

Table 18. Summary of significant interaction effect regression results.

Model (Dependent Var.~Policy Var.)	Moderating Variable	Interaction Term (Policy × Moderator)	Interaction Coefficient	p-Value
Panel A: Moderating Role of Fiscal Space
(1) log (Chargers)~PC2 (Contemp.)	Log (Fiscal Expenditure)	PC2 × log (Fiscal Expenditure)	0.009 **	0.048
(2) log (Chargers)~PC3 (Lag 3)	Log (Fiscal Expenditure)	PC3 × log (Fiscal Expenditure)	0.008 **	0.030
(3) log (Stock)~PC3 (Lag 2)	Log (Fiscal Expenditure)	PC3 × log (Fiscal Expenditure)	0.005 *	0.083
Panel B: Moderating Role of Economic and Industrial Foundation
(4) log (Stock)~PC1 (Lag 2)	Log (Per Capita GDP)	PC1 × log (Per Capita GDP)	0.018 *	0.056
(5) log (Sales)~PC1 (Lag 1)	Secondary Industry Share	PC1 × Secondary Industry Share	−0.117 *	0.066
(6) log (Stock)~PC2 (Contemp.)	Secondary Industry Share	PC2 × Secondary Industry Share	0.163 **	0.018

Note: This table reports only the coefficients for the interaction terms. ** p < 0.05, * p < 0.1. For ease of interpretation, the coefficients on the interaction terms involving fiscal expenditure have been multiplied by 10,000.

summarizes these key findings.

(3) Interpretation of Results and Discussion of Mechanisms

These significant interaction effects provide powerful mechanistic evidence for understanding the “contextual dependency” of policy effectiveness. They reveal how a locality’s fiscal health, level of economic development, and industrial structure systematically moderate the effectiveness of provincial NEV policies.

Fiscal capacity is key to mitigating “policy failure” and enhancing long-term effects. The most consistent findings are centered on provincial fiscal capacity (Panel A). First, for the negative impacts of high-level authoritative (PC2) and coordinative-programmatic (PC3) policies on charging pole construction, we find that local fiscal expenditure plays a significant buffering role. The positive and significant interaction terms imply that the stronger a province’s fiscal capacity, the smaller the negative impact of these two policy types becomes (and may even turn positive). This offers a fiscal-dimension explanation for the observed “authority failure” or “planning fallacy”: in provinces with tight budgets, high-level directives or plans may become “unfunded mandates” without sufficient financial backing, thus producing negative outcomes. Conversely, in fiscally affluent provinces, the government has the capacity to translate plans into real financial investments, effectively mitigating these negative effects. Second, fiscal expenditure also acts as an enhancer for the long-term positive effect of sales-side coordinative-programmatic policies (PC3) on NEV stock. The significant positive interaction term indicates that the greater a province’s fiscal capacity, the stronger this positive long-term impact. This suggests that the effectiveness of programmatic policies aimed at fostering a favorable market environment depends on sustained public resource investment.

The complex moderating role of economic and industrial foundations. The results confirm that the level of economic development acts as an “amplifier” of policy effectiveness (Panel B, Model 4). The long-term positive effect of substantive-driving policies (PC1) on NEV stock is significantly stronger in provinces with higher per capita GDP. This clearly reveals that the effectiveness of consumption-stimulus policies (such as vehicle purchase subsidies) is highly dependent on the purchasing power of local residents, providing direct mechanistic evidence for the “Matthew effect” in policy outcomes. The moderating role of industrial structure (proxied by the share of the secondary industry) is more complex. On the one hand, the short-term promotional effect of substantive-driving policies (PC1) on NEV sales is actually weakened in provinces with a higher share of secondary industry (Model 5). This may reflect greater inertial resistance to market transition in regions with a strong traditional industrial base, or that local protectionism is more inclined to safeguard the conventional fuel vehicle industry. On the other hand, for the negative impact of high-level authoritative policies (PC2) on NEV stock, a higher share of secondary industry can significantly mitigate this negative effect (Model 6). This could imply that although such “slogan-like” policies are ineffective in the general consumer market, they might be more effectively received and implemented by large state-owned or joint-venture enterprises in industrially strong provinces (e.g., through corporate fleet procurement), thereby partially offsetting their negative impact on the consumer market.

3.5.4. Summary of the Heterogeneity Analysis

Synthesizing the results from the subgroup regressions and the interaction effect analysis, this study profoundly reveals the complex picture of NEV policy effectiveness in China and provides robust, multi-level evidence for Hypothesis H3 (that policy effects exhibit significant regional heterogeneity). Our findings deepen the analysis from “describing differences” to “explaining mechanisms”, with the main conclusions as follows:

First, the “contextual dependency” of policy effects is a universal phenomenon; no “one-size-fits-all” policy exists. The results of the subgroup analysis clearly show that policy effects are significantly moderated by geographical location, economic level, and industrial base. For instance, policies aimed at promoting consumption and expanding market stock are more effective in economically developed regions with strong industrial foundations, whereas the Western region shows greater potential for benefiting from industrial policies.

Second, the interaction effect analysis provides direct mechanistic evidence for why these differences exist. By introducing interaction terms, we find that a locality’s fiscal capacity, economic level, and industrial structure systematically moderate policy effectiveness, serving as the key to understanding the heterogeneity of policy outcomes:

Fiscal capacity is a core “buffer” and “enhancer”. Our results consistently show that stronger fiscal capacity (proxied by local fiscal expenditure) can significantly mitigate the potential negative impacts of “high-level authoritative” (PC2) and “coordinative-programmatic” (PC3) policies on charging pole construction. This provides a fiscal-dimension explanation for the “authority failure” or “planning fallacy” we observed. At the same time, it also enhances the long-term positive effect of programmatic policies on NEV stock, highlighting the importance of sustained public resource investment for realizing policy effects.

Economic development level is an “amplifier” of policy effects. The interaction term results confirm that the higher a province’s per capita GDP, the stronger the long-term promotional effect of “substantive-driving policies” (PC1) on NEV stock. This clearly reveals that the effectiveness of consumption-stimulus policies (such as vehicle purchase subsidies) is highly dependent on the purchasing power of local residents, providing direct mechanistic evidence for the “Matthew effect” in policy outcomes.

Industrial structure is a complex “double-edged sword”. The moderating role of a traditional industrial base (proxied by the share of the secondary industry) exhibits complexity. On the one hand, it weakens the short-term promotional effect of substantive-driving policies on NEV sales, possibly reflecting greater inertial resistance to market transition in traditional industrial strongholds. On the other hand, it can mitigate the negative impact of high-level authoritative policies on NEV stock, which may imply that industrially strong provinces can more effectively respond to and implement policies through large enterprises (e.g., corporate fleet procurement), thereby offsetting their shortcomings in the general consumer market.

In conclusion, the heterogeneity analysis not only validates the existence of regional differences but, more importantly, successfully unlocks the “black box” behind these differences through the testing of interaction effects. These findings provide a solid, mechanism-based foundation for shifting China’s NEV policy from a “one-size-fits-all” approach toward precision governance that is “tailored to local conditions”.

4. Discussion

This chapter aims to integrate the preceding empirical findings into a coherent academic narrative. We will begin by consolidating the core findings, then delve into the theoretical contributions, and finally, draw out practical policy implications to showcase the analytical and speculative value of the research.

4.1. Consolidation and Interpretation of Core Findings

The core contribution of this study lies in shifting the analytical perspective from the traditional assessment of policy “quantity” to an inquiry into policy “quality” and “style” by constructing a multi-dimensional PSI. Here, we must first clarify our core concepts: the seven dimensions within the PSI framework collectively constitute the intrinsic and extrinsic “quality” of a policy. PCA then objectively aggregates these quality characteristics into three distinct policy “styles” or models: substantive-driving (PC1), high-level authoritative (PC2), and coordinative-programmatic (PC3). This distinction demonstrates that high-quality policy can be achieved through different styles, and the central aim of this study is to investigate which “style” is more effective for which objective [67].

Our empirical results provide robust quantitative evidence for the core assertion that “quality trumps quantity”. The incremental R-squared test (see Section 3.3) clearly shows that, after controlling for policy quantity, our constructed PSI still adds substantial explanatory power to the models, with relative increases ranging from 83% (for the sales model) to 374% (for the production model). This confirms this study’s central hypothesis (H1): that policy “quality” and “style”, rather than the mere number of documents issued, are the key drivers explaining the development of China’s NEV market.

Building on this, the study’s most central finding is the significant “target mismatch” phenomenon present in provincial NEV policy. The empirical results consistently show that substantive-driving policies (PC1) are the core engine for igniting demand-side outcomes (sales), whereas coordinative-programmatic policies (PC3) have a comparative advantage in guiding supply-side outcomes (production). This finding reveals the differential effects of various policy styles when applied to different segments of the industrial value chain, providing precise empirical evidence for the selection of industrial policy instruments.

However, the most thought-provoking finding of this research is the unexpected negative effect observed in the domain of charging infrastructure construction. The baseline regressions show that both high-level authoritative (PC2) and coordinative-programmatic (PC3) policies exhibit an inhibitory effect. Our mechanism tests (see Section 3.5) offer a preliminary explanation: provincial “fiscal space” is the key moderating variable, with the negative effect being smaller in provinces with greater fiscal capacity. This strongly suggests that the negative effect may stem from a “policy-capacity mismatch” [68]. For infrastructure like charging poles, which require long investment return cycles and are difficult to coordinate, high-level authoritative directives or grand coordinative plans, if not backed by sufficient local fiscal and implementation capacity, may not only become “unfunded mandates” that cannot be realized but could even produce inhibitory effects by distorting market expectations or causing resource misallocation. This phenomenon resonates with a core tenet of governance research: every governance model has its fatal vulnerability, or “Achilles’ heel”, and when the chosen policy instrument happens to strike at the weakest point of the implementation system’s capacity, policy failure becomes inevitable [69].

4.2. Theoretical Contributions and Academic Dialogue

The empirical findings of this study offer a contextualized, non-binary answer to the classic debate in industrial policy concerning the “relationship between government and market”. Traditional industrial policy theory has been shaped by a persistent tension between the “effective government” argument, which emphasizes overcoming market failures [70], and the “government failure” argument, which warns against resource misallocation [71]. Our results show that both theories have explanatory power in the transition of China’s NEV industry, but their applicability is highly dependent on the policy’s target domain and design details. On the one hand, the strong promotional effect of substantive-driving policies (PC1) on sales corroborates the necessity for an “effective government” to overcome coordination failures through direct incentives during the early stages of market cultivation [72]. On the other hand, the negative effects of authoritative and programmatic policies observed in the charging infrastructure domain clearly reveal the risks of “government failure”, especially when the chosen policy instruments are mismatched with local implementation capacity [68,69].

More importantly, this study transforms a seemingly negative robustness check result into a more profound theoretical contribution. When we added “province-specific time trends” to control for the inherent long-term development trajectories of each province, the policy effects in the baseline model lost their statistical significance (see Section 3.3). This result does not imply that policy is ineffective but rather reveals a deeper mechanism of its operation: provincial NEV policies act more as a “catalyst” than a “creator” in local development. The effectiveness of policy does not arise from creating new growth paths out of thin air but is highly dependent on its synergy with a locality’s pre-existing development trends and growth “inertia”. In other words, the function of policy is to identify, accommodate, and accelerate positive regional development dynamics that are already in motion, rather than to serve as an independent driver of new growth.

This “policy as catalyst” perspective perfectly integrates the heterogeneity findings of our study. If policy is a catalyst, its efficacy must depend on the quality of the “reaction conditions.” Our interaction effect analysis (see Section 3.5) precisely identifies these critical “reaction conditions”. The study finds that local characteristics such as economic development level (per capita GDP), industrial structure (share of secondary industry), and fiscal capacity (fiscal space) are the key moderating variables that determine whether the policy “catalyst” can take effect and how strong that effect will be. For example, the promotional effect of substantive-driving policies (PC1) on NEV stock is significantly stronger in provinces with higher per capita GDP, indicating that the “catalyst” of consumption incentives achieves its maximum potency in a “reaction environment” with stronger purchasing power. This series of findings collectively points to a single conclusion: effective industrial policy lies not only in the design of the instrument itself but, more critically, in its ability to achieve a “virtuous match” with the unique institutional and economic context of a locality, thereby catalyzing the intended chemical reaction. This provides new micro-level evidence for understanding how local governments in China engage in strategic policy adaptation under the incentives of the “promotion tournament” [73,74].

4.3. Practical and Policy Implications

Based on the in-depth discussion in the preceding sections, the findings of this study offer specific and actionable implications for optimizing the policy design of future green industry transitions.

Recommendation 1: A Shift in focus—from pursuing “quantity” to enhancing “quality and style”. One of the clearest conclusions of this study is that the “quality” and “style” of a policy are far more decisive for its effectiveness than its mere “quantity.” This provides a fundamental guideline for policymakers: rather than demonstrating commitment to an issue by frequently issuing documents, efforts should be focused on optimizing the design of individual policies. The PSI framework constructed in this study (Table 3) can effectively serve as a “self-checklist” for policy design. When drafting documents, policymakers should consciously aim to enhance the policy’s authoritativeness (e.g., by securing issuance from a higher-level body or a broader coalition of core departments), select more binding instruments, use more decisive language, set clearer quantitative targets, and provide more credible resource guarantees. Only by systematically optimizing these design details can “hard policies” that effectively influence market outcomes be forged.

Recommendation 2: Precision governance—achieving a precise match between policy style and objective. The “target mismatch” phenomenon discovered among different policy styles serves as a warning to policymakers to avoid a “one-key-fits-all-locks” mindset. If the policy objective is to “promote consumption and expand demand” in the short term, the focus should be on “substantive-driving” (PC1) policies. This means direct fiscal incentives (e.g., purchase subsidies) and non-fiscal incentives (e.g., right-of-way priority) remain the most effective tools. If the objective is the long-term “guidance of industry and promotion of production,” the focus should shift to “coordinative-programmatic” (PC3) policies. That is, to attract large-scale industrial investment, providing a forward-looking and stable industrial development plan co-endorsed by multiple departments is more important than short-term direct subsidies. In particular, for infrastructure construction like charging poles, a “policy-capacity mismatch” must be avoided. This means that when formulating plans, they must be accompanied by substantive measures that can solve the practical profitability dilemmas of operators, such as financial support and land-use guarantees, to avoid a “planning fallacy”.

Recommendation 3: Context is king—designing policies that are tailored to local conditions and align with existing trends. The “policy as catalyst” theory and the heterogeneity findings of this study profoundly reveal that no universally “optimal policy” exists. Policy design must shift from a “one-size-fits-all” approach to “precision irrigation” tailored to local conditions. For Eastern and high-income provinces, where the market is relatively mature, the policy focus should shift from universal consumption subsidies to resolving deeper market bottlenecks, leveraging their strong economic and industrial foundations to amplify policy effects. For late-developing regions like the West, their potential to benefit from industrial policy should be fully utilized by actively formulating high-quality “coordinative-programmatic” (PC3) policies to attract external investment and build new industrial growth poles by creating a stable institutional environment. For regions like the Central provinces, which may be caught in a “policy effectiveness collapse zone”, the most urgent task may not be to issue more new policies but to first conduct an in-depth diagnosis of the specific obstacles to local market development and, on that basis, design “small but beautiful” policies oriented toward solving specific problems.

5. Conclusions

5.1. Research Summary

This study aimed to open the “black box” of the effectiveness of China’s provincial NEV policies, responding to the paradox of the industry’s tremendous national success coexisting with severe regional imbalances. To achieve this objective, this research introduced a methodological innovation by constructing and applying a multi-dimensional PSI framework. This framework moves beyond the traditional assessment of policy “quantity” to a deeper inquiry into policy “quality” and “style”. Our core finding is that policy effectiveness depends on the organic integration of quality, style, and context. First, the empirical results clearly demonstrate that policy “quality” and “style”, rather than the mere number of documents issued, are the key drivers explaining the development of China’s NEV market. Second, we identified a significant “target mismatch” phenomenon, wherein different policy styles have differential effects on different segments of the industrial chain (demand-side vs. supply-side). Finally, this study proposes a “policy as catalyst” theory, arguing that provincial policies act more as a “catalyst” that accommodates and accelerates a locality’s pre-existing development trends rather than as an independent “creator” of new growth, with the catalytic effect being highly dependent on local “reaction conditions” such as economic, fiscal, and industrial foundations.

5.2. Limitations and Future Research

Although this study strived for rigor, it is subject to certain limitations, which in turn point to directions for future research.

First is the challenge of endogeneity. Despite our systematic exploration of multiple instrumental variable strategies, the pervasive problem of weak instruments prevents this study from making strong causal inferences. Therefore, the conclusions of this study should be cautiously interpreted as revealing strong correlations between policy styles and market development, rather than strict causal relationships. Future research could more clearly identify causal effects through more refined quasi-experimental designs, such as using policy boundaries for a regression discontinuity design (RDD) or constructing a counterfactual for specific provinces using the synthetic control method (SCM).

Second is the gap between “policy on paper” and “policy in practice”. The PSI constructed in this study measures the “intended strength” embodied in policy texts, not the actual investment and efficiency of local governments at the implementation level. As related literature has pointed out, a “policy-capacity mismatch” often exists between policy intent and implementation capacity [69]. Therefore, a highly promising direction for future research is to integrate this study’s PSI of textual strength with data on “implementation strength”—such as actual local fiscal expenditures, personnel allocation, or regulatory enforcement—to more comprehensively assess the entire chain of policy from intention to outcome.

Third is the limitation of data granularity. While the analysis at the provincial level reveals macroscopic regional differences, it may mask even more significant micro-level heterogeneity among cities within the same province. Extending this study’s analytical framework down to the city level would enable the revelation of more detailed policy interactions and market dynamics. Particularly for testing the effects of policies that are highly related to spatial layout, such as charging infrastructure, city-level analysis would provide more precise insights. Furthermore, applying the PSI analytical framework to evaluate policy effectiveness in other strategic emerging industries (such as renewable energy or semiconductors) would be another important step in deepening the theoretical contributions of this research. Finally, how China’s vast domestic market lays the foundation for the global expansion of its NEV industry, including exports, is also a frontier research topic worthy of further exploration.