Analysis of the Nonlinear Impact of Climate Policy Uncertainty on Total Factor Carbon Productivity in Chinese Cities

Abstract

With the frequent introduction of climate policies in China, the uncertainty surrounding these policies has gradually increased. However, the relationship between climate policy uncertainty and total factor carbon productivity remains unclear. To address this gap, the total factor carbon productivity of 277 Chinese cities from 2007 to 2022 is assessed using the EBM-GML model. Subsequently, a panel smooth transition model is employed to investigate the nonlinear relationship between climate policy uncertainty and total factor carbon productivity, incorporating economic growth, energy consumption structure, green finance, green innovation, and extreme climate as transition variables. Empirical analysis reveals that the impact of climate policy uncertainty on total factor carbon productivity is not uniformly positive or negative. When influenced by the five transition variables, higher levels of economic growth, the development of green finance, and advancements in green technology can shift the impact of climate policy uncertainty on total factor carbon productivity into a positive direction. Conversely, a higher reliance on coal consumption and frequent extreme weather events impede this positive influence. The heterogeneity analysis confirms significant regional and resource endowment heterogeneity in the observed nonlinear effects among cities. Furthermore, in most regions, the values of the transition variables do not exceed the threshold. Notably, under the influence of economic growth and green technology innovation, the potential for an increase in the impact coefficient of climate policy uncertainty on carbon productivity is substantial. Therefore, it is imperative to further enhance economic growth and green technology innovation. Additionally, specific climate policy targets must be established to address energy consumption structure and extreme weather, thereby improving carbon productivity.

1. Introduction

China’s remarkable economic growth has attracted global attention, but has also led to excessive carbon emissions. Currently, China’s green and low-carbon development faces significant imbalances. To address the challenge of balancing economic development and environmental protection, the Chinese government has continuously introduced and adjusted climate-related policies across different development stages, and the dynamic evolution of these policies inevitably generates climate policy uncertainty (CPU). Particularly in recent years, with the strengthening of climate governance efforts, the impact of CPU has become increasingly significant, and its spatial-temporal distribution characteristics are shown in Figure 1. Figure 1 shows that CPU levels in Chinese cities generally exhibit an upward trend during the study period, while there are certain regional differences in spatial distribution.

In this context, how to accurately assess urban low-carbon development performance becomes a key issue. Total Factor Carbon Productivity (TFCP), defined as the maximum economic output achievable per unit of carbon emissions under given input constraints [1], is a comprehensive indicator for measuring the degree of decoupling between economic growth and carbon emissions, providing an effective measurement tool for assessing urban low-carbon development. Therefore, understanding how CPU affects TFCP in Chinese cities is of great significance for advancing China’s “dual carbon” goals.

From a theoretical perspective, the impact mechanism of CPU on TFCP is complex and exhibits a dual nature. Based on real options theory, policy uncertainty generates “waiting value”, causing firms to delay irreversible green investment decisions [2], thereby suppressing TFCP improvement. Meanwhile, according to growth options theory, moderate policy uncertainty increases the option value of green technology investment, prompting firms to advance their green transformation layout to obtain the option value of future growth opportunities, avoiding the loss of competitive advantages after policy tightening and creating a “policy forcing” effect [3,4].

However, the impact of CPU on TFCP may vary significantly across different urban development conditions. Under different levels of economic growth, energy consumption structure, green finance development, green technology innovation capacity, and extreme climate event frequency, the direction and intensity of the impact of CPU on TFCP may undergo structural changes. This suggests that there may be a significant nonlinear relationship between CPU and TFCP, meaning that when these key development conditions cross specific thresholds, the CPU-TFCP relationship may shift from one operating pattern to another. Therefore, identifying these key threshold conditions is of great importance for understanding the complexity of the CPU-TFCP relationship.

Existing research on CPU mainly focuses on three aspects: First, regarding the impact of CPU on firm micro-behavior, related research finds that the impact of CPU on firm behavior exhibits heterogeneity: regarding negative effects, CPU reduces firm total factor productivity [5] and hinders green technology innovation [6]; regarding positive effects, research based on growth options theory finds that CPU can incentivize Chinese A-share listed companies to conduct green technology innovation [7]. Second, studies on the impact of CPU on macroeconomic variables. Existing literature explores the linear impacts of CPU on macroeconomic indicators such as carbon emissions [8], green total factor productivity [9], and carbon trading prices [10]. Third, CPU measurement and prediction research, mainly focusing on applications of CPU as a prediction indicator for energy and financial market prices [11,12,13].

Although existing research provides rich perspectives for understanding the economic impacts of CPU, several research gaps remain: First, existing research is mostly based on linear modeling frameworks, assuming that the marginal impact of CPU on dependent variables remains constant, failing to adequately consider the structural changes in the impact effects of CPU that may occur with changes in other conditions. However, given the complexity of the CPU mechanism, its impact on comprehensive indicators such as TFCP may exhibit threshold effects or regime-switching characteristics, which existing linear frameworks can hardly capture accurately. Second, research levels mainly focus on the national macro level or firm micro level, with insufficient attention to cities as the important policy implementation agents and carriers of low-carbon transformation. Finally, there is a lack of multi-dimensional systematic empirical testing of the specific transmission mechanisms through which CPU affects TFCP, with existing research mostly focusing on single mechanism pathways.

This study aims to fill the above research gaps by employing the Panel Smooth Transition Regression (PSTR) model to explore the nonlinear relationship between CPU and TFCP. The choice of the PSTR model is based on the following considerations. Compared to traditional threshold regression models, the PSTR model offers two distinct advantages. First, traditional threshold models assume that parameters undergo discrete jumps at threshold points—that is, coefficients shift instantaneously from one value to another when the transition variable crosses the threshold. This assumption is often too extreme in economic reality, as changes in economic growth, energy consumption structure, and other transition variables are typically gradual processes rather than abrupt shifts. The PSTR model addresses this limitation by introducing a smooth transition function that allows parameters to shift continuously between regimes, better capturing the gradual adjustment patterns observed in economic variables. Second, the PSTR model quantifies the speed of regime transition through the smoothness parameter: a larger value indicates a more abrupt transition, while a smaller value suggests a more gradual shift. This information is valuable for understanding the dynamic characteristics of regime switching, yet traditional threshold models cannot provide it [14]. This paper employs panel data covering 277 Chinese prefecture-level cities from 2007 to 2022, uses the EBM-GML model to measure cities’ TFCP, and systematically analyzes the transmission mechanisms through which CPU affects TFCP from five dimensions: economic growth, energy consumption structure, green finance, green technology innovation, and extreme climate events.

The main contributions of this study are reflected in the following three aspects:

First, in terms of methodological contribution, this study is the first to apply the PSTR model to examine the nonlinear relationship between CPU and TFCP. Traditional threshold models assume that parameters undergo discrete jumps at threshold points, whereas economic agents typically adjust their behavior gradually. The PSTR model, by employing a smooth transition function, allows parameters to shift continuously between regimes, thus better conforming to economic reality.

Second, in terms of mechanism analysis, the contribution of this study lies not in discovering new mechanism variables but in transforming and deepening the analytical perspective. Existing studies predominantly employ subsample regressions or interaction term approaches to examine the moderating effects of factors such as economic growth, energy structure, and green finance [15,16]. These methods implicitly assume that moderating effects are linear and continuous, making them incapable of identifying potential “critical points” or “turning points”. The innovation of this study lies in elevating these factors from conventional “moderating variables” to “transition variables” within the PSTR framework, systematically testing whether the impact of CPU on TFCP exhibits regime-switching points—that is, whether the effect of CPU reverses direction when these conditions cross specific thresholds. This shift in perspective enables us to address a key question: under what conditions does the CPU effect undergo a qualitative change? Furthermore, it allows us to quantify the precise threshold value for each condition, thereby addressing the limitation of existing research in identifying critical points of regime transition.

Third, in terms of policy contribution, the city-level empirical analysis not only identifies the key threshold points where CPU affects TFCP but also characterizes the current features of Chinese cities across various transmission mechanism dimensions, providing scientific evidence for cities at different development stages to formulate differentiated climate policies.

The structure of this paper is organized as follows: Section 2 provides a review of the relevant literature. Section 3 describes the theoretical foundation of the research methods employed. Section 4 presents a detailed explanation of the variables and data used in the study. Section 5 discusses the findings and offers an interpretation of the results. Section 6 concludes the paper and proposes potential policy recommendations.

2. Literature Review and Theoretical Framework

2.1. Related Literature Review

CPU refers to policy uncertainty arising from frequent adjustments of climate policies, changing targets, and unclear implementation paths, representing an important manifestation of transition risk. CPU reflects the complex trade-offs that governments face in formulating and implementing climate policies, as well as the dynamic adjustment process between policy targets and implementation measures. Although the literature specifically examining the relationship between CPU and carbon productivity remains relatively limited, a substantial body of research has explored the links between CPU and related indicators such as green total factor productivity (GTFP) and carbon emission efficiency [9,15,16]. Notably, GTFP and carbon productivity share the same theoretical foundation in their measurement approaches, as both treat carbon emissions as undesirable outputs [17,18]. Additionally, related research has explored the links between CPU and factors such as carbon emissions, green technology, green investment, carbon pricing, and energy prices [6,8,10,19,20].

Existing research shows that the impact of CPU on economic and environmental indicators exhibits significant dual characteristics. From the perspective of negative effects, based on real options theory, CPU hinders firms’ green investment decisions by increasing the “waiting value” of investment. CPU has been proven to reduce firm total factor productivity [5], suppress green technology innovation [21], and depress current firm valuations [22]. High CPU also increases firms’ financing difficulties, forcing high-emission firms to reduce their operating scale [23]. However, the positive effects of CPU cannot be ignored. From the growth options theory perspective, when firms view green technology investment as a potential future growth opportunity, policy uncertainty may prompt firms to advance their green investment to obtain the option value of future growth, avoiding the loss of competitive advantages after policy tightening [24]. Bai et al. [7] found that CPU can incentivize Chinese A-share listed companies to conduct green technology innovation. Tian and Li [8] used Chinese provincial data and found that higher CPU usage reduced the ratio of fossil fuels to clean energy, thereby reducing carbon emissions. He and Xu [25] also found that energy policy uncertainty can enhance urban green total factor productivity by indirectly stimulating corporate innovation and increasing the proportion of the tertiary sector.

2.2. Theoretical Framework for the Impact of CPU on Urban TFCP

The coexistence of the aforementioned positive and negative effects may manifest more complex patterns at the city level. Cities serve as agglomeration platforms for firms, financial institutions, and research organizations, and their carbon productivity performance essentially reflects the aggregated outcomes of decisions made by micro-level agents within them [26,27]. On the one hand, city governments possess a certain degree of autonomy in implementing climate policies and may adjust higher-level policies according to local development needs. On the other hand, the higher density of economic agents within cities enables the transmission effects of CPU to be more rapid and direct. Meanwhile, significant differences exist among cities in terms of industrial structure, technological level, and resource endowments [28], which may lead to stronger heterogeneity in the effects of CPU.

This study integrates real options theory and growth options theory to construct a city-level analytical framework. On the one hand, from the perspective of real options theory, Dixit and Pindyck point out that green investment is characterized by irreversibility, and “waiting” itself possesses option value [29]. When CPU rises, firms within cities tend to postpone green investments to acquire more policy information, while local governments may also delay the construction of green infrastructure [16]. This wait-and-see behavior of micro-level agents generates an aggregation effect at the city level, ultimately suppressing the improvement of TFCP. On the other hand, from the perspective of growth options theory, Myers argues that green investment can also be viewed as an option to capture future growth opportunities [30]. Faced with rising CPU, firms may interpret it as a signal that policies will become more stringent, thereby attaching greater importance to investment projects related to TFCP. By positioning their green investments ahead of time, firms can secure first-mover competitive advantages [7]. In this context, CPU instead stimulates investment motivation and promotes the improvement of urban TFCP.

Both effects coexist, and their relative strength depends on the specific conditions of each city. In cities with strong resource constraints and high transition costs, the real options effect dominates. In cities with abundant resources and strong technological capabilities, the growth options effect dominates. This implies that the impact of CPU on urban TFCP may exhibit threshold transition points—when certain key conditions of a city cross specific thresholds, the dominant effect shifts, and the direction of the impact of CPU changes accordingly.

2.3. Theoretical Mechanisms of Transition Variables

Based on the above framework, this study selects five city-level transition variables to characterize the relative strength of the two effects from the following dimensions: economic growth level, energy consumption structure, green finance development level, green technology innovation, and extreme climate events.

Economic growth level. The theoretical basis for selecting economic growth level as a transition variable is the Environmental Kuznets Curve hypothesis, which posits an inverted U-shaped relationship between economic growth and environmental degradation. In the early stages of economic development, governments typically prioritize economic expansion and often neglect environmental issues. At this stage, cities face tighter resource constraints, and increases in CPU strengthen the real options effect, leading to more conservative low-carbon investment strategies as firms tend to postpone green investments while waiting for policy clarity. Consequently, CPU may exert a negative impact on TFCP. As the economy reaches more advanced stages, the growth pattern shifts toward innovation and low-carbon strategies. Cities possess more abundant fiscal resources and stronger risk-bearing capacity, enabling them to perceive CPU as a “strategic opportunity window.” At this point, the growth options effect begins to dominate, and CPU demonstrates a positive impact on TFCP [31,32]. Therefore, in cities with lower economic development levels, the real options effect dominates and CPU has an inhibiting effect on TFCP, whereas in economically developed cities, the growth options effect dominates and CPU may promote TFCP improvement.

Energy consumption structure. Energy consumption structure is a key factor affecting urban carbon emission intensity, and its role in the CPU-TFCP relationship stems from transition cost differentials and path dependence mechanisms. Given that coal still accounts for a significant proportion of China’s total energy consumption and remains the primary contributor to carbon emissions [33,34], this study uses the proportion of coal consumption as a proxy for energy consumption structure. Cities with high coal dependence concentrate a large number of capital-intensive heavy and chemical industries with long asset life cycles [35]. Under such conditions characterized by high transition costs and strong path lock-in, combined with the irreversibility of investment, when CPU increases, the predictions of real options theory become more applicable: based on considerations of waiting value, firms typically postpone major investment decisions [36], and the aggregate effect of this micro-level behavior leads to a slow urban green transition and inhibits TFCP improvement. In contrast, low-coal cities possess well-developed clean energy infrastructure and advanced green technology reserves, with lower transition costs and weaker path dependence. Under these circumstances, the logic of growth options theory has greater explanatory power: CPU stimulates firms’ growth options value, prompting them to perceive CPU as a signal to seize green development opportunities. Firms tend to proactively position their green technology investments to secure future competitive advantages, and the aggregate effect of this micro-level investment behavior at the city level enhances overall TFCP.

Green finance development level. As a critical market mechanism for facilitating low-carbon transition, the development level of green finance directly influences cities’ capacity to respond to CPU through signaling mechanisms and risk-sharing mechanisms. Cities with well-developed green finance systems offer firms multiple financing channels—green bonds, green credit, green funds, and other instruments. These tools signal government commitment to low-carbon transitions, easing firms’ concerns about policy reversals and lowering financing costs through risk-sharing arrangements. Under CPU, firms in these cities can still access capital for green transitions through various channels, reducing financing constraints. Greater capital availability lowers the option value of waiting and raises the growth options value, making firms more likely to invest in green projects and innovation, which boosts urban TFCP [16]. In contrast, cities with weak green finance systems lack these channels. Firms must rely mainly on traditional financial institutions for capital. Without risk-sharing mechanisms, CPU makes banks more cautious about green projects, raising financing costs and limiting available capital. Here, the real options effect dominates: firms face tighter capital constraints and find it harder to invest in green technology upgrades, which can suppress urban TFCP [37].

Green technology innovation level. Green technology innovation capacity determines cities’ technological response capability to CPU through technology spillover effects and learning curve mechanisms. While green technology innovation is intuitively expected to reduce carbon emissions, its interaction effect with CPU exhibits complex characteristics that precisely reflect the competing relationship between the two option theories. On one hand, cities with low levels of green technology innovation lack knowledge spillover networks and technological learning capabilities, making it difficult to attract investors to perceive them as growth options. In such contexts, the real options effect dominates, leading CPU to potentially exert a negative impact on TFCP [2]. On the other hand, cities with high levels of green technology innovation possess extensive knowledge spillover networks formed by universities, R&D institutions, and innovative enterprises. In these settings, green technology innovation can achieve dual outcomes of carbon reduction and economic growth. Technological innovation reduces the cost curve of green transformation, enabling the growth options effect to maintain dominance even under elevated CPU levels. The characteristics of growth options become observable; consequently, CPU may facilitate a positive impact on TFCP [3,4].

Extreme climate events. Extreme climate events, as external environmental shocks, alter the impact mechanism of CPU on urban TFCP through risk perception intensification and urgency mechanisms. The essence of this moderation lies in changing the trade-off between the waiting value of real options and the opportunity value of growth options. Cities with fewer extreme weather events lack urgent pressure from tangible climate risks. When confronted with climate policy uncertainty, the logic of real options theory becomes more applicable: city governments and major economic agents tend to activate waiting option value, postponing irreversible decisions such as major green infrastructure construction and industrial transformation investments to await further policy clarity and avoid bearing substantial sunk cost risks. This city-level investment delay inhibits overall TFCP improvement [38,39]. Conversely, cities experiencing more frequent extreme weather events face direct threats from climate risks. In these contexts, the real threat of climate risk exceeds the waiting value induced by policy uncertainty, eroding the waiting option value while activating the growth options effect. The growth options value is activated in domains such as green infrastructure, adaptive capacity building, and clean technology industries. These cities perceive climate pressure as a signal to capture green development opportunities, prompting them to transition from “delay and wait” to “proactive planning”. This reallocation of urban option value significantly mitigates the negative impact of CPU [40].

3. Methodology

3.1. PSTR Model

This study employs a PSTR model, as used in prior research [14,41], to determine the nonlinear relationship between CPU and TFCP. The PSTR model for this relationship is constructed as follows:

$$TFC{P}_{it}={\alpha }_0+{\beta }_{1}CCP{U}_{it}+{\displaystyle \sum _{j=1}^r{\beta }_{2j}CCP{U}_{it}H(q_{it}^j;{\gamma }_j;c_{\mathrm{j}})}+{\mu }_{it}$$

(1)

In this framework, $TFC{P}_{it}$ denotes the total factor carbon productivity for city i in year t and serves as the study’s dependent variable. ${\alpha }_0$ signifies the function’s constant term, while $CCP{U}_{it}$ refers to the CPU of city i in year t, the main explanatory variable in this analysis. ${\beta }_1$ is the coefficient that measures the correlation between CPU and TFCP, considered the central parameter. ${\beta }_{2j}$ represents the coefficient associated with the nonlinear component. ${\mu }_{it}$ signifies the random disturbance. The parameter r indicates the number of transition function. $q_{it}^j$ denotes the transition variable, $H(q_{it}^j;{\gamma }_j;c_{\mathrm{j}})$ represents the transition function—a continuous and bounded function dependent on $q_{it}^j$. This function is normalized to fall within the range [0,1]. The slope parameter ${\gamma }_j>0$ dictates the transition speed within the model, where a higher value corresponds to a steeper transition slope, indicating a quicker shift. The m-dimensional vector serves as the location parameter, also known as the threshold, marking the point where the model transitions. Typically, the transition function $H(q_{it}^j;{\gamma }_{\mathrm{j}};c_{\mathrm{j}})$ is expressed as a logistic function.

$$H(q_{it}^j;{\gamma }_{\mathrm{j}};c_{\mathrm{j}})={\left[1+\exp (-{\gamma }_{\mathrm{j}}{\displaystyle \prod _{k=1}^{m_j}(q_{it}^j-c_{\mathrm{j},\mathrm{k}})})\right]}^{-1}$$

(2)

Additionally, H is a continuous function that varies smoothly within the [0,1] interval.

3.2. Linearity Test

Prior to building the PSTR model, conducting a linearity test is crucial. The model can only be established if the data series exhibits nonlinearity. This test is also vital in determining the appropriate value of m for the transition function $H(q_{it}^j;{\gamma }_{\mathrm{j}};c_{\mathrm{j}})$. The null hypothesis for the linearity test assumes either $H_0:r=0$ or $H_0:{\beta }_2=0$. However, the parameters in the PSTR model may become unidentifiable. To address this issue, the null hypothesis is instead set as $H_0:r=0$. Additionally, to avoid identification problems when r = 0, a Taylor series expansion is applied to $H(q_{it}^j;{\gamma }_{\mathrm{j}};c_{\mathrm{j}})$, expressed as:

$$y_{it}={\mu }_i+{\beta }_0^{\prime }{x}_{it}+{\beta }_1^{\prime }{x}_{it}{q}_{it}+\dots +{\beta }_m^{\prime }{x}_{it}{q}_{it}^m+{\mu }_{it}^{\prime }$$

(3)

In Equation (3), ${\beta }_0^{\prime }$, ${\beta }_1^{\prime }$,…, ${\beta }_m^{\prime }$ is generated by r, which are constants. ${\mu }_{it}^{\prime }={\mu }_{it}+R_m{\beta }_1^{\prime }{x}_{it}$. while $R_m$ represents the remaining terms in the Taylor series expansion. Therefore, the null hypothesis $H_0:{\beta }_1^{\prime }=\dots ={\beta }_m^{\prime }=0$ for the linearity test assumes the same form. Acceptance of the null hypothesis implies that constructing a PSTR model is not suitable.

3.3. Test for No Remaining Nonlinearity

The no remaining nonlinearity test in the PSTR model aims to determine if the residuals contain any significant nonlinear components. This test is conceptually similar to the linearity test. The null hypothesis for this test posits that no further nonlinearity is present, and it is expressed as $H_0:{\beta }_1^{\prime *}=\dots ={\beta }_m^{\prime *}=0$. Acceptance of this hypothesis implies that the PSTR model has effectively captured all nonlinear relationships in the data series.

4. Data

4.1. Dependent Variable: Measurement of TFCP

The dependent variable in this study is TFCP. Accurate measurement of TFCP is crucial for evaluating urban low-carbon development performance. Currently, there are two main methods for measuring carbon productivity: single-indicator measurement and multi-indicator measurement. The first method, single-indicator measurement, typically uses the ratio of carbon emissions to economic indicators. While this method is simple and direct, its limitation lies in its narrow measurement scope, which cannot fully reflect the comprehensive utilization efficiency of input factors. The second method improves upon the first by considering multiple factors from an input-output perspective, providing a more comprehensive assessment of regional carbon productivity. This method mainly employs two techniques: Data Envelopment Analysis (DEA) and Stochastic Frontier Analysis (SFA).

Compared to SFA, DEA has clear advantages in handling multiple-input, multiple-output problems: First, it avoids the subjectivity of functional form specification; Second, it can simultaneously handle both desirable and undesirable outputs; Third, it does not require prior assumptions about the specific form of the production function. The advantage of DEA lies in avoiding errors related to equation specification [42], but traditional DEA models lack the ability to distinguish environmental factors. The non-radial, non-oriented SBM model introduced by Tone [43] effectively addresses this problem. However, the SBM model has limitations in handling dynamic efficiency changes. To analyze dynamic efficiency more deeply, combining the SBM model with the Malmquist-Luenberger (ML) index is a widely used method for evaluating TFP changes over time [44,45].

Considering the specific characteristics of urban TFCP measurement, this study selects the EBM-GML model based on the following considerations: First, the EBM-GML model combines radial and non-radial distances, enabling it to handle both proportional improvements and slack improvements, which more accurately reflects urban flexibility in input-output adjustments. Second, this model employs the global GML index, which overcomes the potential infeasibility problems that may occur with traditional ML indices [46], ensuring comparability of TFCP across cities. Finally, this model is particularly suitable for handling production systems that include undesirable outputs, accurately reflecting the comprehensive performance of cities in pursuing economic growth while controlling carbon emissions.

The TFCP measurement framework constructed in this study includes three categories of input factors and two categories of output factors. Input variables include capital stock, energy consumption, and labor force; the desirable output is real GDP; the undesirable output is CO₂ emissions. This framework follows the basic characteristics of urban production systems and can comprehensively reflect the structure of factor inputs and outputs in urban economic activities, as listed in

Table 1. Evaluation Index System for TFCP.

Evaluation Index	Variable
Input Variables	Capital Stock (billion yuan)
	Labor Force (10,000 people)
	Energy Consumption (10,000 tons of standard coal)
Output Variables	Expected Variables	Regional Gross Domestic Product (billion yuan)
	Unexpected Variables	Carbon Dioxide Emissions (10,000 tons)

.

The specific quantification methods for input-output variables are as follows: Capital stock is calculated using the perpetual inventory method, starting with the total fixed asset investment in 2007 and applying a 9.6% depreciation rate to determine capital stock in subsequent years [47]. Labor input is measured by the employment numbers in the primary, secondary, and tertiary industries of each sample city. Energy consumption is calculated by converting the total usage of natural gas, liquefied petroleum gas, and electricity into standard coal equivalent using their respective conversion factors. Real GDP is measured annually, representing urban economic output, and is deflated using 2007 constant prices to eliminate the impact of price factors. CO₂ emissions are calculated using the method of Ren et al. [48], which estimates emissions by applying corresponding carbon emission factors to various types of urban energy consumption, covering major carbon emission sources such as manufactured gas, natural gas, liquefied petroleum gas, electricity consumption, and steam and hot water for heating.

4.2. Core Explanatory Variable: CPU

The key explanatory variable is CPU. Accurate measurement of CPU is a crucial component of this study. Existing literature shows that CPU indicators are mainly measured through text analysis techniques, but data sources vary across different regions. This variation primarily stems from differences in media environments, policy information dissemination channels, and linguistic characteristics across countries. For example, CPU data in the United States typically comes from major newspapers [49], while some studies in China use Twitter data [50]. However, Twitter usage is limited in China, and such data may not adequately reflect the characteristics of climate policy information dissemination in the country.

Given the high correlation between China’s climate policy objectives and economic development goals, as well as the unique role of Chinese media in policy information dissemination, this study argues that using Chinese mainstream media data can more accurately reflect the frequency of climate policy releases and the dynamic adjustments of policy objectives. Therefore, this study uses research data from Ma et al. [51], which provides city-level panel data with text data sourced from major Chinese newspapers, specifically including People’s Daily, Guangming Daily, and other authoritative media with national influence. The data construction method draws on the methodology of China’s Economic Policy Uncertainty index, ensuring the reliability and consistency of the indicator.

4.3. Transition Variables

This study selects five transition variables, and the measurement methods for each variable are as follows.

Economic Growth Level. This study uses per capita real GDP with 2007 as the base year as a proxy variable for economic development level.

Energy Consumption Structure. It should be noted that due to the difficulty in obtaining coal consumption data at the Chinese city level, this study adopts the estimation method of Ren et al. [48]. This method is based on the reality that China’s electricity mainly comes from coal-fired power generation, and estimates coal consumption through electricity consumption. Specifically, after converting electricity consumption using the standard coal conversion coefficient, combined with natural gas and liquefied petroleum gas consumption, the proportion of coal in total energy consumption is calculated. Although this method has certain measurement errors, in the absence of direct data, this method has been widely validated and applied in related studies.

Green Finance Development Level. Following the “Guidance on Building a Green Finance System” issued by the People’s Bank of China and six other ministries, and drawing on the research of Bai and Lin [52], this study constructs a comprehensive green finance index that includes seven dimensions: green credit, green investment, green insurance, green bonds, green support, green funds, and green equity.

Green Technology Innovation Level. This study selects the number of green invention patents obtained by each city as a proxy variable for green technology innovation, and this indicator is logarithmically transformed after adding 1 to alleviate the problem of skewed data distribution.

Extreme Climate Events. Due to the significant regional differences in climate characteristics across China, this study follows the relative threshold approach used by Pan et al. [53] to establish EW event thresholds, aligning with the percentile method proposed by Schär et al. [54].

4.4. Data Sources and Sample Description

Given the data availability and policy relevance, the research time window is selected as 2007–2022, based on the following considerations: 2007 was an important year for the implementation of China’s 11th Five-Year Plan, when the government issued the “National Climate Change Plan of China”, marking an important starting point for the systematic construction of China’s climate policy framework; 2020 was the key year when China proposed the “dual carbon goals”, and this time window can fully capture the process of China’s climate policy development from its initial stage to systematic development. The panel data used in the study covers 277 Chinese cities, with data sourced from the National Meteorological Administration and the EPS website. The descriptive statistics of variables are shown in

Table 2. The descriptive statistics of each variable.

Variable	Symbol	Obs	Mean	Std. Dev	Min	Max
Total factor carbon productivity	TFCP	4432	0.612	0.161	0.162	1
China’s climate policy uncertainty	CPU	4432	1.374	0.604	0	4.179
Real GDP per capita	PGDP	4432	4.767	2.626	0.625	15.634
Energy consumption structure	ECS	4432	0.810	0.150	0.084	0.997
Green finance	GF	4432	0.403	0.100	0.058	0.726
Green Technology Innovation	GTI	4432	2.507	1.732	0	8.731
Extreme weather	EW	4432	0.279	0.045	0.167	0.456

.

5. Results and Discussion

5.1. Estimation Results of TFCP

To visually present the spatiotemporal evolution of urban TFCP, this study selected six time periods: 2007, 2010, 2013, 2015, 2017, and 2020, and categorized the TFCP values into five classes. Figure 2 illustrates the spatial distribution and evolution of TFCP across Chinese cities. Over time, from 2007 to 2020, a pattern emerges where urban TFCP initially increases, then decreases, followed by a gradual improvement. Spatially, there are significant disparities in TFCP development among cities, characterized by a “higher in the east, lower in the west” distribution. This pattern may be attributed to the more advanced carbon reduction and energy utilization technologies in the eastern regions, and these technological advantages have increasingly highlighted the differences in carbon productivity between the economically developed east and the lagging west.

5.2. Stability Test

To evaluate the stationarity of each variable and prevent potential spurious regression, we employed panel unit root tests. In particular, we applied the LLC and IPS methods for this purpose. The test results, as presented in

Table 3. Panel unit root test.

Variable	LLC		IPS
	Statistic	p-Value	Statistic	p-Value
TFCP	−1.4179 *	0.0781	−8.2553 ***	0.0000
CPU	−18.9851 ***	0.0000	−24.6820 ***	0.0000
PGDP	−0.5589 **	0.0288	2.8419 ***	0.0022
ECS	−3.3242 ***	0.0004	−7.4483 ***	0.0000
GF	−12.2736 ***	0.0000	−24.8690 ***	0.0000
GTI	−14.7773 ***	0.0000	−27.5876 ***	0.0000
EW	−20.7419 ***	0.0000	−26.5800 ***	0.0000

The symbols ***, **, and * indicate significance at the 1%, 5%, and 10% confidence levels, respectively.

, indicate that TFCP, CPU, real GDP per capita, energy consumption structure, green finance, green technology innovation, and extreme climate events are all stationary at the 10% significance level. Therefore, it is appropriate to establish the PSTR model based on these variables.

5.3. Linearity Test

After all variables passed the unit root test, a linearity test was required to ensure the sufficient conditions for the nonlinear impact of CPU on TFCP. As shown in

Table 4. Linearity test results.

Model	Conversion Variable	H0: r = 0
		LM	LMF	LRT
1	PGDP	248.344 ***	246.584 ***	255.574 ***
		(0.000)	(0.000)	(0.000)
2	ECS	6.498 **	6.099 **	6.502 **
		(0.011)	(0.014)	(0.011)
3	GF	118.582 ***	114.200 ***	120.198 ***
		(0.000)	(0.000)	(0.000)
4	GTI	56.565 ***	53.702 ***	56.929 ***
		(0.000)	(0.000)	(0.000)
5	EW	4.530 **	4.250 **	4.532 **
		(0.033)	(0.039)	(0.033)

The values in brackets indicate the probability values. The symbols *** and ** indicate significance at the 1% and 5% confidence levels, respectively.

, the test statistics (LM, LMF, and LRT) for models 1 through 5 yield p-values that are all less than 1%, prompting us to reject the hypothesis of linearity. This finding confirms a nonlinear relationship between CPU and TFCP, influenced by the five transition variables.

5.4. Residual Nonlinearity Test

After establishing that the CPU exerts a nonlinear effect on TFCP in the presence of five selected transformation variables, a test for residual nonlinearity was conducted to determine the most appropriate number of transition functions (r) for the model.

Table 5. The remaining nonlinearity test results.

Model	Conversion Variable	H0: r = 1
		LM	LMF	LRT
1	PGDP	0.058	0.054	0.058
		(0.810)	(0.816)	(0.810)
2	ECS	1.222	0.572	1.222
		(−0.543)	(0.564)	(0.543)
3	GF	1.225	1.148	1.226
		(0.268)	(0.284)	(0.268)
4	GTI	0.021	0.020	0.021
		(0.884)	(0.887)	(0.884)
5	EW	2.523	2.365	2.524
		(0.112)	(0.124)	(0.112)

Notes: the values in brackets indicate the probability values.

reveals that the p-values for the three statistics associated with models 1 through 5 are all above 10%, which supports the acceptance of the null hypothesis that r equals 1. This result suggests that a transition function should be incorporated into the PSTR model.

5.5. Determining Location Parameters

After identifying the appropriate number of transition functions (r) within the model, the subsequent step involves identifying the correct number of location parameters (m) for each PSTR model. To accomplish this, estimations were carried out with m values of 1 and 2. The optimal value of m was then selected using the AIC and BIC criteria. As presented in

Table 6. Results of AIC and BIC for positional parameters determination.

Model	Conversion Variable	M Option	AIC	BIC	Optimal Value of m
1	PGDP	m = 1	−3.826	−3.820	m = 1
		m = 2	−3.836	−3.819
2	ECS	m = 1	−3.783	−3.778	m = 2
		m = 2	−3.809	−3.802
3	GF	m = 1	−3.789	−3.783	m = 1
		m = 2	−3.788	−3.781
4	GTI	m = 1	−3.774	−3.768	m = 1
		m = 2	−3.773	−3.766
5	EW	m = 1	−3.765	−3.759	m = 1
		m = 2	−3.764	−3.757

, the AIC and BIC outcomes reveal that, except for Model 2, which incorporates two location parameters, all other models are characterized by a single location parameter.

5.6. Findings from the Regression Analysis Using the PSTR Model

After selecting an appropriate transition function and optimal location parameters for the model, five PSTR models were constructed in this study. The regression results, shown in

Table 7. Regression results of PSTR model.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0436 ***	0.0087 *	−0.0683 ***	−0.0816 ***	−0.0178 ***
		t statistic	−9.5031	1.9773	−9.9090	−7.8942	−3.7770
	β2	Estimated value	0.1190 ***	−0.0982 ***	0.0749 ***	0.1227 ***	0.0153 ***
		t statistic	16.4399	−16.7544	10.7215	7.3457	4.091
Influence coefficient	β1 + β2		0.0754	−0.0895	0.0066	0.0411	−0.0025
Positional parameter	c		8.4826	0.9557	0.2296	2.504	0.2735
				0.4966
Slope coefficient	γ		0.6059	66.7938	15.7655	0.2534	147,250
Sum of squares of residuals	RSS		96.352	97.87	99.948	101.456	102.379

The symbols ***, and * indicate significance at the 1%, and 10% confidence levels, respectively.

, indicate that both the linear and nonlinear coefficients are significant at the 1% level.

5.7. Evaluating the Influence of Transitional Variables on the Connection Between CPU and TFCP

5.7.1. Assessing How Economic Growth Affects the Link Between CPU and TFCP

Model 1 examines how the impact coefficient of CPU on TFCP varies with changes in real GDP per capita. As shown in Table 7, Model 1 identifies a single threshold at c = 8.4826. The linear coefficient of CPU is β1 = −0.0436, while the coefficient of the nonlinear term is β2 = 0.1190. Consequently, the theoretical range of CPU impact on TFCP is [−0.0436, 0.0754].

Figure A1 displays the variation curves of the transition function and impact coefficient, revealing that the transition function shifts smoothly within the range of [0.0085, 0.9569] as real GDP per capita fluctuates. Moreover, the impact coefficient of CPU on TFCP transitions seamlessly between low and high states, with an actual range of [−0.0426, 0.0703]. When real GDP per capita falls below 84,826 RMB, CPU exerts a significant negative effect on TFCP, with the lowest impact coefficient recorded at −0.0426. However, when real GDP per capita surpasses 84,826 RMB, the negative effect of CPU on TFCP turns into a significant positive influence, reaching an impact coefficient of 0.0703, which remains well below the theoretical maximum.

These findings suggest that at lower levels of economic growth, increased CPU reduces TFCP. However, as economic growth improves, CPU begins to enhance TFCP. This result could be explained by the balancing act between achieving economic growth and meeting climate objectives. At lower levels of economic growth, governments tend to prioritize economic development, causing increases in CPU to manifest as more pronounced carbon risk shocks, thereby reducing TFCP. Once economic growth reaches a certain threshold, the focus shifts towards climate goals, and frequent climate policy enactments and the associated policy transition risks are perceived by the market as stricter climate policies, leading to an overall improvement in TFCP. This conclusion is consistent with previous findings regarding the separation of economic growth from carbon emissions. In developing countries, such as those in Asia, economic growth is often accompanied by increased carbon emissions. In contrast, in developed countries in Europe, governments place greater emphasis on promoting low-carbon economic growth [55,56]. As of 2022, out of 277 cities, only 40 had a real GDP per capita exceeding 84,826 RMB, leaving 85% of the cities below this threshold and thus subject to the adverse effects of increased CPU on TFCP.

5.7.2. Assessing How Energy Consumption Structure Affects the Link Between CPU and TFCP

Model 2 examines how the impact coefficient of CPU on TFCP varies with changes in the ECS. As shown in Table 7, Model 2 identifies two thresholds, 0.9557 and 0.4966, indicating that the impact of CPU on TFCP exhibits a double-threshold asymmetric characteristic. The linear coefficient of CPU is β1 = 0.0087, while the coefficient of the nonlinear term is β2 = −0.0982. Consequently, the theoretical range of CPU impact on TFCP is [−0.0982, 0.0087].

Figure A2 depicts the trends observed in the transition function and impact coefficient. The transition function ranges from a minimum value of G = 0.0288 to a maximum of G = 1, with a smooth transition between these extremes. Notably, the graph of the transition function exhibits a characteristic trend: it first declines and then rises. The function reaches its minimum value of 0.0288 at 0.72615 ((0.4966 + 0.9557)/2), where the nonlinear effect of the model is at its weakest. As the ECS changes, the impact coefficient of CPU on TFCP transitions smoothly between high and low states, with an actual range of [−0.0895, 0.0059]. When the ECS ranges from 0.4966 to 0.9557, the rate of change in the elasticity coefficient of CPU impact on TFCP continues to increase, with the maximum elasticity coefficient reaching 0.0052. When the ECS falls below 0.4966 or exceeds 0.9557, the elasticity coefficient of CPU impact on TFCP remains negative, reaching a minimum of −0.0895.

Although the theoretical analysis expected that low coal-dependent cities would be better able to convert CPU into improved TFCP, the empirical results show that both extremely low and extremely high coal dependence lead to negative effects of CPU on TFCP, while cities with moderate coal dependence demonstrate the best conversion performance. This may be because low coal consumption cities have already achieved highly optimized energy structures with relatively high TFCP levels. Under these circumstances, further green technology investments face diminishing marginal returns, and increases in CPU may cause firms to become overly cautious, reducing investments due to concerns about policy overshooting risks. Cities with moderate coal dependence are at a critical juncture of transition, having both sufficient room for improvement and certain technological and financial foundations. Moderate increases in CPU can effectively stimulate transformation incentives without creating excessive uncertainty burdens. The study’s findings align with those reported by Wang et al. [57], which classified resource-based cities into growth, mature, declining, and regenerating cities based on their development stages. The results indicate that urban sustainability planning tends to overly emphasize pollution reduction targets while neglecting environmental conditions, such as technological advancement and human capital, that foster clean production. Consequently, this policy has not facilitated an increase in GTFP in regenerating and growth cities; in some cases, it has even led to a decline in their green TFP. However, it has resulted in an improvement in GTFP for mature and declining cities [57]. As of 2022, 84% of cities were within the threshold range, while 16%, or 44 cities, fell outside this range. Among them, 22 cities exhibited an energy consumption share exceeding 0.9557, indicating that a subset of resource-dependent cities still require more clearly defined climate policy objectives to mitigate the adverse effects stemming from frequent policy changes.

5.7.3. Assessing How Green Finance Affects the Link Between CPU and TFCP

Model 3 examines how the impact coefficient of CPU on TFCP changes with variations in GF. As shown in Table 7, Model 3 identifies a single threshold at c = 0.2296. The linear coefficient of CPU is β1 = −0.0683, while the nonlinear coefficient is β2 = 0.0749. Therefore, the theoretical range of CPU impact on TFCP is [−0.0683, 0.0749].

Figure A3 depicts the trends observed in the transition function and impact coefficient. The transition function ranges from a low of G = 0.0623 to a high of G = 0.9984, with a gradual transition between these points. As GF progresses, the impact coefficient of CPU on TFCP transitions smoothly between high and low states, with a real-world range of [−0.0637, 0.0064]. When GF surpasses 0.2296, the positive influence of CPU on TFCP becomes significant, peaking at an impact coefficient of 0.0064, which closely aligns with the theoretical value. On the other hand, when GF is below 0.2296, the positive impact of CPU on TFCP diminishes and eventually turns negative, with the coefficient decreasing to −0.0637.

This outcome likely arises from the necessity of a more advanced GF sector to meet the “dual carbon goals” [47,58]. In contrast, at lower levels of GF, the heightened risks associated with CPU might prevent businesses from securing enough green funding to invest in low-carbon industries, thus increasing volatility in energy, financial, and carbon markets and causing a drop in TFCP. However, as GF increases, enterprises begin to have adequate funding for green projects, and the risks in energy, financial, and carbon markets gradually become more manageable. In this context, increased CPU actually boosts investor confidence in future green projects, thereby enhancing TFCP. This aligns with research suggesting that GF can serve as a tool to mitigate climate risks, as appropriately designed green financial policies can effectively address climate change [59]. Furthermore, raising the level of GF can strengthen energy resilience directly by increasing funding for renewable energy sources and supporting green initiatives. Indirectly, it can strengthen the capacity to withstand risks through GTI and the upgrading of industrial structures [60]. As of 2022, cities exceeding the GF threshold accounted for 87% of the 277 surveyed cities, indicating that the current level of GF in Chinese cities is sufficiently high to effectively mitigate the risks associated with policy transitions driven by increased CPU.

5.7.4. Assessing How Green Technology Innovation Affects the Link Between CPU and TFCP

Model 4 examines how the impact coefficient of CPU on TFCP varies with changes in GTI. As shown in Table 7, Model 4 identifies a single threshold at c = 2.504. The linear coefficient of CPU is β1 = −0.0816, while the nonlinear coefficient is β2 = 0.1227. Therefore, the theoretical range of CPU impact on TFCP is [−0.0816, 0.1227].

Figure A4 depicts the trends observed in the transition function and impact coefficient. The transition function ranges from a minimum value of G = 0.3465 to a maximum of G = 0.8212, with a smooth transition between these two points. As GTI advances, the impact coefficient of CPU on TFCP transitions smoothly between high and low regimes, spanning an actual range of [−0.0391, 0.0191]. When the level of GTI is less than 2.504, the impact of CPU on TFCP is significantly negative, with a coefficient as low as −0.0391. However, once GTI exceeds 2.504, the effect of CPU on TFCP shifts to a positive direction, increasing the coefficient to 0.0191, though it remains well below the expected lower limit. These findings imply that GTI is still at a relatively low level in most Chinese cities. In this research, green patents were used to gauge high-quality green innovation, indicating that most Chinese cities are lacking in this area.

The observed results may be attributed to the varying levels of GTI, which lead to different investment behaviors among investors. According to real options theory, investors may avoid investing due to fears of risks associated with climate policy transitions [21]. Conversely, the theory of growth options suggests that investors, optimistic about the future prospects of green projects, are more likely to invest in green innovation [7]. Consequently, at low levels of GTI, a rise in CPU generally has a detrimental effect on TFCP. In contrast, at a more advanced stage of GTI, a rise in CPU can result in enhanced carbon productivity. Furthermore, GTI has been identified as a key factor in enhancing TFCP and green total-factor productivity [61,62]. By 2022, only 34% of the 277 cities in the sample had surpassed the GTI threshold, while the other 66% remained below it. This highlights the overall low level of GTI across Chinese cities, with a limited number achieving high-quality green innovation. It is evident that there is a pressing need to elevate the level of high-quality GTI to better mitigate the risks associated with policy transitions.

5.7.5. Assessing How Extreme Weather Affects the Link Between CPU and TFCP

Model 5 analyzes how the impact coefficient of CPU on TFCP varies with changes in EW conditions. As shown in Table 7, Model 5 identifies a single threshold at c = 0.2735. The linear coefficient of CPU is β1 = −0.0178, while the nonlinear coefficient is β2 = 0.0153. Thus, the theoretical range of CPU impact on TFCP is [−0.0178, −0.0025].

Figure A5 depicts the trends observed in the transition function and impact coefficient. The transition function ranges from a minimum value of G = 1.82 × 10⁻¹⁹⁰ to a maximum of G = 1, with a smooth transition between these points. As the frequency of EW events increases, the impact coefficient of CPU on TFCP transitions smoothly between high and low regimes, spanning an actual range of [−0.0178, −0.0025]. When EW events occur with a frequency lower than 0.2735, CPU exerts a markedly negative influence on TFCP, with the impact coefficient dropping to a minimum of −0.0178. However, as the frequency of these events surpasses 0.2735, the negative influence of CPU on TFCP diminishes, and the impact coefficient reduces to −0.0025, which aligns with the theoretical maximum. Notably, when the frequency of extreme weather events increases, the negative impact decreases substantially, from −0.0178 to −0.0025.

The likely reasons for these results are not only that increased CPU can reduce corporate productivity and discourage investors from green project investments, but also that EW events can cause significant economic damage [63,64], further harming productivity. Specifically, there is an inverted U-shaped relationship between temperature and the TFP of Chinese firms [65]. Extreme heat can impair worker efficiency and machine performance, while prolonged low temperatures, accompanied by snow and strong winds, can disrupt freight efficiency [66], adversely affecting production. Additionally, increases in extreme rainfall not only reduce economic growth rates [67] but also diminish banks’ risk-taking capacity [53], leading to a contraction in lending, which is detrimental to economic development and productivity growth. Therefore, regardless of the frequency of EW events, the increased CPU exacerbates climate risks. This heightened risk deters investors from investing in green projects, resulting in a decrease in TFCP. Additionally, in cities experiencing frequent extreme weather events, the negative effect of CPU on TFCP is mitigated by the increase in growth options value. As of 2022, 72% of cities still had a probability of EW event occurrence below the threshold, leading to a further decline in TFCP under the impact of CPU.

5.8. Discussion on Endogeneity

To address the potential endogeneity bias arising from bidirectional causality between CPU and urban carbon productivity, this study employs the instrumental variable (IV) approach to re-estimate the PSTR model [68], constructing two instrumental variables to enhance the robustness of the results.

The first instrumental variable is constructed based on the following logic: global average temperature anomaly (Data source: https://berkeleyearth.org/data/ (accessed on 1 June 2025)), as exogenous natural climate information, influences the intensity of China’s climate policy formulation through international climate negotiation pressure and policy response mechanisms, but does not directly affect the production technology and carbon emission efficiency of specific cities. Given the dual time-series and cross-sectional dimensions of panel data, this study selects the interaction term between one-period-lagged CPU and global average temperature anomaly as the instrumental variable to capture the impact of exogenous climate shocks on policy uncertainty.

The second instrumental variable is constructed as follows: the product of the average CPU of other cities and U.S.CPU. Specifically, the average policy uncertainty of all other cities nationwide (excluding the target city and its geographically proximate cities [69,70]) reflects the national-level overall policy environment and is highly correlated with the target city’s policy uncertainty through central-local policy transmission mechanisms. Meanwhile, as one of the world’s largest economies, the U.S. climate policy changes constitute an exogenous international policy shock to China. The interaction term of these two components both ensures both strong correlation with the endogenous explanatory variable and satisfies the exogeneity requirement—by excluding geographically proximate cities, the average policy level of other cities does not directly affect the target city’s carbon productivity through channels such as industrial relocation, factor mobility, or technology spillovers; simultaneously, U.S. climate policy, as an exogenous international shock, only indirectly affects the target city by influencing China’s overall policy environment.

Table 8. PSTR Model Results with IV (Lagged CPU × Global Temperature Anomaly).

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0474 ***	0.0183 ***	−0.0635 **	−0.0771 ***	−0.0094 *
		t statistic	−10.1078	3.7676	−2.0645	−9.4118	−1.8885
	β2	Estimated value	0.1123 ***	−0.0903 ***	0.0613 **	0.1265 ***	0.0088 **
		t statistic	10.0005	−15.5391	2.1043	10.0824	2.3102
Influence coefficient	β1 + β2		9.9531	0.0720	0.0022	0.0494	−0.0006
Positional parameter	c		8.2730	0.9472	0.2145	−3.1297	0.2541
				0.5380
Slope coefficient	γ		0.5457	70.3143	2.0219	0.3260	3.0303 × 104
Sum of squares of residuals	RSS		88.585	91.983	93.082	94.131	96.431

The symbols ***, **, and * indicate significance at the 1%, 5%, and 10% confidence levels, respectively.

and

Table 9. PSTR Model Results with IV (Other Cities’ Average CPU × U.S. CPU).

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0527 ***	0.0101 *	−0.1115 ***	−0.0946 ***	−0.0224 ***
		t statistic	−9.4457	1.9631	−12.1873	−7.8351	−3.8908
	β2	Estimated value	0.1470 ***	−0.1198 ***	0.1184 ***	0.1410 ***	0.0189 ***
		t statistic	16.4714	−16.9909	12.2934	7.2551	4.1351
Influence coefficient	β1 + β2		0.0943	−0.1097	0.0069	0.0464	−0.0035
Positional parameter	c		8.4867	0.9554	0.4595	2.5756	0.2736
				0.4958
Slope coefficient	γ		0.6258	68.4842	24.6278	0.2684	8.4371 × 104
Sum of squares of residuals	RSS		96.250	97.802	100.470	101.470	102.356

The symbols ***, and * indicate significance at the 1%, and 10% confidence levels, respectively.

report the PSTR model results re-estimated using the two instrumental variables. Both sets of results show that the coefficients of the core explanatory variables remain statistically significant, with signs and magnitudes broadly consistent with the baseline regression results, thereby confirming the robustness of the main research findings.

5.9. Robustness Tests

To further verify the reliability of the impact of CPU on urban TFCP, this study conducts robustness checks from two dimensions: incorporating additional control variables, replacing model specifications, and extending the sample time span.

5.9.1. Additional Control Variables

Considering potential endogeneity concerns arising from omitted important explanatory variables, this study further incorporates three control variables that may affect TFCP into the baseline model. Specifically, these include: (1) Industrial structure, measured by the share of secondary industry value-added in urban GDP, to control for the impact of industrial composition on TFCP; (2) Foreign direct investment (FDI), measured by the logarithm of FDI amount, to capture technology spillover effects brought by foreign capital inflows; (3) Urbanization level, measured by the proportion of urban permanent residents in the total population, to reflect the impact of urbanization processes on TFCP.

The regression results after adding these control variables are presented in

Table 10. Regression results with additional control variables.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0434 ***	0.0058 **	−0.0506 ***	−0.0339 ***	−0.0234 ***
		t statistic	−9.2738	2.342	−8.7009	−6.7026	−4.7960
	β2	Estimated value	0.1264 ***	−0.1151 ***	0.0552 ***	0.0365 ***	0.0139 ***
		t statistic	13.4386	−16.5660	8.7205	6.7337	3.4081
Influence coefficient	β1 + β2		0.0830	−0.1093	0.0046	0.0026	−0.0095
Positional parameter	c		8.2966	0.9674	0.2477	2.9785	0.2736
				0.4809
Slope coefficient	γ		0.5815	56.6362	22.4823	2.1884	8.7259 × 104
Sum of squares of residuals	RSS		95.808	97.172	99.249	99.419	101.420

The symbols *** and ** indicate significance at the 1% and 5% confidence levels, respectively.

. The results demonstrate that after controlling for industrial structure, FDI, and urbanization level, the impact coefficient of CPU on urban TFCP, the significance level, and the form of the nonlinear relationship remain consistent with the baseline regression results, indicating that the research findings are robust.

5.9.2. Alternative Model Specification

Given that TFCP values are strictly bounded between 0 and 1, traditional linear panel models or standard PSTR models may inadequately account for the bounded nature of the dependent variable. To address this potential model specification bias, this study employs a fractional response panel model [71] as an alternative specification for robustness checks.

Specifically, based on the estimation results of the baseline PSTR model, this study partitions the sample into different regimes according to the location parameters of the five transition variables and separately estimates the impact effects of CPU within each regime. The regression results are presented in

Table 11. Regression results suing alternative model specifications (part I).

	(1)	(2)	(3)	(4)	(5)
	PGDP1	PGDP2	ECS1	ECS2	GF1
CPU	−0.0610 ***	0.0603 *	−0.0481 ***	0.0049 **	−0.0316 *
	(−4.7328)	(1.6905)	(−3.4245)	(2.5601)	(−1.8274)
cons	−0.6358 ***	0.2788 *	−0.9174 ***	0.3616 *	−0.9443 ***
	(−6.0617)	(1.9126)	(−8.0193)	(1.7876)	(−49.7359)
N	4051	381	3525	565	18

The symbols ***, **, and * indicate significance at the 1%, 5%, and 10% confidence levels, respectively.

and

Table 12. Regression results suing alternative model specifications (part II).

	(6)	(7)	(8)	(9)	(10)
	GF2	GTI1	GTI2	EW1	EW2
CPU	0.0603 ***	−0.0514 ***	0.0160 *	−0.0892 ***	−0.0347 *
	(4.7349)	(−3.1772)	(1.8844)	(−3.2554)	(−1.9735)
cons	−0.6637 ***	−0.5958 ***	−0.5704 ***	−1.0943 ***	−1.0601 ***
	(−6.7430)	(−5.1205)	(−8.4840)	(−8.3643)	(−4.2000)
N	4414	2777	1655	2293	1797

The symbols *** and * indicate significance at the 1% and 10% confidence levels, respectively.

, respectively. The findings reveal:

(1)

When PGDP serves as the transition variable, the coefficient of CPU is significantly negative in the low economic growth regime (PGDP ≤ 8.4826), whereas it turns positive in the high economic growth regime (PGDP > 8.4826), confirming the existence of a nonlinear relationship.

(2)

When ECS serves as the transition variable, the coefficient of CPU is negative in the intermediate coal consumption share regime (0.4966 ≤ ECS ≤ 0.9557), while it becomes positive outside this interval, exhibiting interval-dependent nonlinear characteristics.

(3)

When GF serves as the transition variable, the coefficient is negative in the low GF regime (GF ≤ 0.2296) and positive in the high GF regime (GF > 0.2296).

(4)

When GTI serves as the transition variable, the coefficient is negative in the low GTI regime (GTI ≤ 2.504) and positive in the high GTI regime (GTI > 2.504).

(5)

When EW serves as the transition variable, the coefficient remains negative in both the low regime (EW ≤ 0.2735) and high regime (EW > 0.2735), though the magnitude of impact differs across regimes.

These results demonstrate that even when employing alternative model specifications to address the bounded nature of the dependent variable, the nonlinear impact of CPU on TFCP remains significant and robust, providing further support for the core conclusions of this study.

5.10. Heterogeneity Analysis

5.10.1. Regional Heterogeneity

Given China’s vast territory and significant regional development disparities, this study partitions the full sample into three major regions—Eastern, Central, and Western—based on geographical location to conduct heterogeneity tests [72].

Table 13. Heterogeneity analysis by region: eastern China.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0491 ***	—	—	−0.6499 ***	—
		t statistic	−5.8605	—	—	−8.4566	—
	β2	Estimated value	0.1196 ***	—	—	0.6637 ***	—
		t statistic	16.4399	—	—	8.3668	—
Influence coefficient	β1 + β2		0.0705	—	—	0.0138	—
Positional parameter	c		8.9211	—	—	−3.2647	—
Slope coefficient	γ		0.3765	—	—	0.5198	—
Sum of squares of residuals	RSS		31.284	—	—	31.801	—

The symbols *** indicate significance at the 1% confidence levels, respectively.

,

Table 14. Heterogeneity analysis by region: central China.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	0.1428 ***	—	−0.0419 ***	—	−0.0420 ***
		t statistic	5.9139	—	−5.7030	—	−4.7962
	β2	Estimated value	−0.1798 ***	—	0.0295 ***	—	0.0220 **
		t statistic	−8.2438	—	4.4393	—	2.4896
Influence coefficient	β1 + β2		−0.0370	—	−0.0124	—	−0.0200
Positional parameter	c		1.2338	—	0.3097	—	0.2662
Slope coefficient	γ		4.3543	—	4.8707 × 104	—	89.8974
Sum of squares of residuals	RSS		30.8750	—	32.2320	—	32.6010

The symbols *** and ** indicate significance at the 1% and 5% confidence levels, respectively.

and

Table 15. Heterogeneity analysis by region: western China.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	—	—	—	−0.0101	—
		t statistic	—	—	—	−1.2109	—
	β2	Estimated value	—	—	—	−0.0460 ***	—
		t statistic	—	—	—	−7.5802	—
Influence coefficient	β1 + β2		—	—	—	−0.0561	—
Positional parameter	c		—	—	—	3.1323	—
Slope coefficient	γ		—	—	—	501.5138	—
Sum of squares of residuals	RSS		—	—	—	29.8860	—

The symbols *** indicate significance at the 1% confidence levels, respectively.

present the regression results for the three regions, respectively. The findings reveal pronounced heterogeneity across regions in terms of the selection of transition variables, the values of location parameters, and the nature of nonlinear relationships.

For the Eastern region, PGDP and GTI serve as significant transition variables, exhibiting nonlinear effects in the impact of CPU on TFCP, whereas ECS, GF, and EW do not demonstrate statistically significant nonlinear effects. For the Central region, PGDP, GF, and EW all demonstrate significant nonlinear effects in the impact of CPU on TFCP. For the Western region, only GTI exhibits significant nonlinear characteristics in the impact of CPU on TFCP.

The Eastern region’s PGDP location parameter significantly exceeds the full sample level, while the Central region’s falls below it. This means the East needs higher economic growth before CPU can boost TFCP. With more advanced industries and stronger firms, Eastern cities require a solid economic base to absorb policy shocks and channel them into innovation. Central and Western regions, having lower growth levels, may see firms respond to CPU with low-carbon initiatives even at lower economic thresholds.

GF shows unusual patterns in the Central region. CPU hurts TFCP at both high and low levels, unlike the national sample where effects become positive above the threshold. The region’s green finance remains underdeveloped, with poor resource allocation. When CPU hits, the GF system cannot buffer risks or allocate capital well, which may suppress TFCP through misallocated funding.

The Western region’s GTI location parameter is much higher than the Eastern region’s, meaning CPU only helps TFCP once GTI reaches higher levels. The Western region’s lower carbon productivity and less developed green industry explain this: firms there need more accumulated innovation capacity before they can capitalize on low-carbon opportunities under policy uncertainty.

For EW, the Central region displays a lower location parameter, with regime switching occurring at lower EW frequencies. The region’s below-average EW exposure creates heightened sensitivity: even moderate shocks trigger a wake-up effect that elevates climate risk awareness among firms and governments. CPU and EW thus interact to create a forcing mechanism that accelerates low-carbon transitions at relatively low EW thresholds.

5.10.2. Resource Endowment Heterogeneity

To account for heterogeneity in resource endowments, we classify sample cities as resource-based or non-resource-based following China’s National Sustainable Development Plan for Resource-Based Cities (2013–2020) [68].

Table 16. Heterogeneity analysis by city type: resource-based cities.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0249 ***	0.0421 ***	−2.1400 × 105 **	−0.0254 ***	−0.0130 *
		t statistic	−3.8701	4.7074	−2.1444	−3.6913	−1.7470
	β2	Estimated value	0.0747 ***	−0.0831 ***	9.0712 × 109 ***	0.1519 ***	0.0215 ***
		t statistic	9.4698	−8.4068	3.3688	11.1330	3.4008
Influence coefficient	β1 + β2		0.0498	−0.0410	9.0710 × 109	0.1265	0.0085
Positional parameter	c		8.5824	0.8689	4.8216	3.6843	0.2540
Slope coefficient	γ		2.0006	18.1529	5.9656	1.3369	3.1869 × 104
Sum of squares of residuals	RSS		36.4630	36.8870	38.3030	36.9410	38.3000

The symbols ***, **, and * indicate significance at the 1%, 5%, and 10% confidence levels, respectively.

and

Table 17. Heterogeneity analysis by city type: non-resource-based cities.

Model			1	2	3	4	5
Conversion variable			PGDP	ECS	GF	GTI	EW
Explanatory variable: Total factor carbon productivity TFCP	β1	Estimated value	−0.0629 ***	−0.1027 ***	−0.1246 ***	−0.0471 ***	—
		t statistic	−10.9264	−12.6829	−12.6513	−7.4489	—
	β2	Estimated value	0.0938 ***	0.0836 ***	0.1045 ***	0.0370 ***	—
		t statistic	15.2873	11.1900	11.5308	5.7896	—
Influence coefficient	β1 + β2		0.0309	−0.0191	−0.0201	−0.0101	—
Positional parameter	c		8.3123	0.5544	0.1913	5.2222	—
Slope coefficient	γ		1.9077	19.1916	17.7710	4.5746	—
Sum of squares of residuals	RSS		56.9360	59.7170	58.9400	60.9860	—

The symbols *** indicate significance at the 1% confidence levels, respectively.

report the results. PGDP, ECS, GF, and GTI exhibit significant nonlinear effects on the CPU-TFCP relationship in both city types, while EW displays nonlinearity only in resource-based cities.

For PGDP, resource-based cities show substantially higher location parameters than non-resource-based cities, indicating that positive CPU effects on TFCP require higher location parameters in resource-dependent regions. Their legacy of resource extraction and heavy industry concentration creates strong path dependence. These cities can only effectively navigate policy uncertainty and channel it toward low-carbon transitions after reaching sufficient economic development and accumulating adequate capital and technological capacity.

For ECS, resource-based cities exhibit higher location parameters than non-resource-based cities, meaning CPU negatively affects TFCP when coal consumption shares exceed the location parameter. Their pronounced fossil fuel dependence makes energy transitions costly and technologically demanding. High coal shares amplify policy uncertainty’s deterrent effect, inducing firms to delay low-carbon investments and thereby suppressing TFCP growth.

Resource-based cities show higher GF location parameters, with CPU only helping TFCP once green finance surpasses the threshold. These cities face stronger transformation pressures and environmental challenges, requiring substantial green finance to reduce transition risks and capital costs. Green finance needs to reach sufficient scale before it can create risk-sharing mechanisms that help firms handle policy uncertainty and invest in low-carbon transitions.

Non-resource cities are more sensitive to GTI thresholds. High GTI levels weaken CPU’s negative effects on TFCP considerably. These cities have flexible industries and innovation-driven economies where strong GTI makes firms more adaptable and competitive, letting them seize policy-created market opportunities that compensate for uncertainty costs.

Resource-based cities show a clear pattern with EW. CPU hurts TFCP when extreme weather is infrequent but helps when it is frequent. This shift appears to reflect learning: repeated extreme weather makes climate risks more visible, pushing governments and firms to increase low-carbon investments. In this context, policy uncertainty actually drives precautionary responses that boost TFCP.

6. Conclusions and Policy Implications

Within the framework of the “dual carbon goals”, it is crucial to assess how increasing uncertainty, represented by CPU, affects TFCP. The study utilized the EBM-GML model to assess the TFCP of 277 cities in China from 2007 to 2022. By employing the PSTR model, we investigated the nonlinear effects of CPU on TFCP by selecting real GDP per capita, energy consumption structure, green finance, green technology innovation, and extreme climate events as transition variables. Additionally, the research examined the threshold effects to determine the boundaries of CPU influence on TFCP. We addressed endogeneity concerns through instrumental variable estimation and conducted multiple robustness checks. Heterogeneity analyses across three geographic regions (Eastern, Central, and Western China) and by resource endowment (resource-based versus non-resource-based cities) revealed substantial variation in how CPU affects TFCP across regions and by resource endowment.

Main Findings:

(1)

Nonlinear effect of economic growth (threshold: RMB 84,826): High levels of economic growth can offset the negative impact of CPU. However, only 15% of cities exceed this threshold.

(2)

Nonlinear effect of the energy consumption structure (thresholds: 0.4966 and 0.9557): When the share of coal consumption is moderate (0.4966–0.9557), the effect of CPU turns positive; yet nearly 8% of cities rely excessively on coal (>0.9557), rendering them unable to effectively cope with policy transition risks.

(3)

Nonlinear effect of green finance (threshold: 0.2296): High levels of green finance can effectively guide cities to invest in green projects. Currently, 91% of cities exceed this threshold and therefore possess a relatively solid foundation.

(4)

Nonlinear effect of green technological innovation (threshold: 2.504): High-quality innovation can transform CPU into productivity gains, but only 30% of cities exceed this threshold, indicating a generally insufficient level of high-quality green innovation.

(5)

Nonlinear effect of extreme climate events (threshold: 0.2735): Regardless of the frequency of extreme weather events, CPU consistently generates negative impacts; however, the negative effect is relatively milder in cities with more frequent extreme events.

(6)

Significant regional heterogeneity: Eastern cities require higher levels of economic growth to convert CPU into low-carbon momentum; the underdevelopment of green finance in central regions weakens policy effectiveness; and western regions demand a higher level of green technological innovation.

(7)

Crucial role of resource endowment: Resource-based cities require higher thresholds for economic development, energy structure, and green finance to achieve positive effects, with particularly prominent path dependence.

Corresponding policy recommendations are as follows: First, deepen gradient-based climate policies tailored to different levels of economic development. For low-GDP cities (85%), it is necessary to further clarify medium- and long-term climate goals, reduce the frequency of policy adjustments, and promote low-carbon transition gradually while maintaining a reasonable rate of economic growth. For high-GDP cities (15%), climate policy standards should be raised in advance, with greater responsibility assumed in carbon quota allocation, and mechanisms for technology transfer and green investment should be established to support the transition of low-GDP cities.

Second, promote targeted transition pathways based on energy structure characteristics. For cities with moderate coal dependence (84%), clean energy substitution should be deepened and a phased coal-consumption control mechanism should be improved. For cities with high coal dependence (8%), policy expectations need to be further stabilized, the scale of special funds such as coal-substitution programs should be expanded, and the principle of “building the new before phasing out the old” should be consistently upheld to promote low-carbon transition.

Third, consolidate the advantages of green finance while addressing weak links. For cities with low levels of green finance (13%), efforts should be accelerated to establish pilot zones for green financial reform and innovation, and the scope of fiscal interest subsidies and risk-compensation mechanisms for green credit should be expanded. For cities with high levels of green finance (87%), the scale of green finance should continue to increase, carbon-finance products should be innovated, and paired-assistance mechanisms should be improved.

Fourth, substantially enhance high-quality green technological innovation capacity. For cities with low levels of innovation (66%), fiscal investment in science and technology should be significantly increased, with priority support for green invention patents, strengthened development of national-level research platforms, and increased recruitment of high-end talent. For cities with high levels of innovation (34%), efforts should focus on deepening breakthroughs in frontier green technologies and establishing sound platforms for technology transfer and patent-sharing mechanisms.

Fifth, strengthen differentiated response capacity to extreme climate risks. For cities with a low frequency of extreme weather events (72%), policy expectations should be further stabilized, investment in disaster-prevention infrastructure should be enhanced, early-warning and emergency-response systems should be improved, and the scale of climate-adaptation funds should be expanded. For cities with a high frequency of extreme weather events (28%), climate-adaptation capacity should be continuously consolidated, and climate-risk stress-testing mechanisms should be further developed.

Sixth, optimize the policy framework for region-specific measures based on regional heterogeneity. The eastern region should continue to consolidate its innovation-driven advantages and promote the development of central and western regions through technological spillovers and regional coordination mechanisms. The central region needs to deepen green finance reform, improve capital allocation efficiency, and reduce the negative impact of policy uncertainty on market participants. The western region should receive more substantial fiscal and technological support and benefit from more flexible policy constraints to compensate for its weak technological base and limited transition capacity.

Seventh, accelerate structural transformation and institutional breakthroughs in resource-based cities. Support for industrial transformation should be increased, and a mix of fiscal, tax, and financial policies should be used to weaken the path-locking effect of resource dependence. Phased and differentiated strategies for adjusting the energy structure should be continuously implemented, promoting coal substitution steadily while ensuring energy security. A dedicated green-finance mechanism for resource-based cities should be improved, with risk-sharing arrangements used to lower firms’ transition costs and enhance the financial system’s ability to support green transformation.

7. Research Limitations and Future Directions

This study provides empirical evidence on the CPU-TFCP relationship, but important limitations warrant attention.

First, we have not fully captured spatial dimensions. Cities do not operate in isolation—their TFCP trajectories are shaped by interactions with neighboring cities through technology diffusion, factor mobility, and policy demonstration effects, creating spatial spillover effects. Ignoring these spatial interdependencies could bias our parameter estimates and affect the precision of our policy assessments. Future work should use spatial panel econometric methods, such as the Spatial Durbin Model to separate direct and indirect effects and clarify how carbon productivity changes spread across urban networks.

Second, the mechanisms through which CPU affects carbon productivity may exhibit considerable spatial heterogeneity and nonlinear characteristics. Subsequent research could adopt the Spatial Panel Smooth Transition Regression models with different transition variables to investigate how the impact of CPU on carbon productivity varies across different spatial regimes. Such extensions would facilitate the identification of spatial nonlinear transition mechanisms in policy effects, thereby providing more robust theoretical foundations and empirical evidence for implementing targeted, differentiated regional climate governance strategies.