Introduction

The problem of global warming and climate change is becoming increasingly serious, attracting the attention of researchers in many disciplines. Although the carbon sequestration capacity of the ocean remains stable (Borucke et al., 2013), the area of forest land has decreased by 50 million hectares every 10 years since the 20th century, according to Global Forest Resources Assessment. This leading to an overall diminishing ability of carbon storage. With the aggravation of greenhouse gas emissions (Xiang et al., 2023), the CO2 ecological footprint (CEF)Footnote 1 seriously exceeds the ecological carrying capacity (Fang et al., 2023), triggering multiple natural disasters, and ultimately endangering the social economy and human health (Ren et al., 2022; Tee et al., 2023; Zhang et al., 2017; Ren et al., 2023a). To alleviate climate problems, many countries pay for environmental protection costs through the public sector, especially the local fiscal expenditure on environmental protection under the financial decentralization system. For example, the proportion of LEPE to GDP in the United States and EU countries has shown an upward trend in the past decade. This is because the local government’s environmental protection expenditure can provide more information superiority and cost advantage, which can more effectively improve the regional environment than the central or federal government management (Esty, 1996). It should be noted that over half of the global population lives in cities (Moran et al., 2018; Zhang et al., 2022), which is responsible for about 75% of the total carbon emissions around the world. Thus, CEF reduction under the decentralized management of municipal government is the main goal in the future (Cheng et al., 2021a; Xu, 2022; Ren et al., 2023b).

With rapid urbanization, China has become the largest carbon emitter and energy consumer since 2006 (Liu et al., 2013; Zou et al., 2023). In this trend, China’s carbon emissions accounted for 31% of the world’s total emissions in 2019. It is imperative to calculate CEF in China after considering carbon sequestration. To cope with climate change, the Chinese government has delegated power, local governments officially set up a separate fiscal expenditure project for environmental protection in 2007, and then raised its budget as well. Local environmental protection expenditure (LEPE) mainly includes investment in environmental governance and prevention, regular expenditure for environmental protection, and subsidies or transfers for environmental protection activities (Broniewicz, 2011). It plays an important role in facilitating green investment and providing environmental governance funds (Fan et al., 2022). Moreover, it is a kind of market-oriented environmental regulation and a direct means of government energy conservation and emission reduction. Therefore, taking China, the largest carbon emitter, as the research object, it is worth studying how the intensity of LEPE can alleviate urban CEF under the fiscal decentralization in environmental protection expenditure.

Some papers used national or provincial data to study the effect of LEPE. Using the CS-ARDL method, Caglar and Yavuz (2023) found that LEPE would help the EU achieve a cleaner environment, including more renewable energy and better water quality. Huang (2018) conducted a cross-province study and found that LEPE can effectively reduce China’s sulfur dioxide emissions, and this effect would be stronger in more industrialized provinces. However, the negative externality of pollution, especially gas pollution with strong spillover, will lead to interactive competition on LEPE among local governments (Feng et al., 2023). Once the intensity of LEPE in a certain region increases, the pollution of surrounding regions will be reduced. Thus, in order to make better use of the budget, politicians are inclined to reduce LEPE in their own regions. From this perspective, using the samples with a higher administrative level fails to consider sufficiently the game competition of environmental expenditure between local governments. Although some scholars use city-level data to research the issue of LEPE (Hu et al., 2023), they still ignore the competitive behavior of LEPE in research methods or results analysis.

Research about urban carbon emissions has been a hot topic (Chen et al., 2023; Yan et al., 2023; Zhang et al., 2023; Wang et al., 2023). Many literatures use the nighttime light data to assess greenhouse gas emissions of countries or cities (Su et al., 2014; Chen et al., 2020; Ren et al., 2023c). However, such a method has limitations (Ghosh et al., 2010). For small-sized and medium-sized cities with a small population, there is a bias in using night light data to estimate carbon emissions (Meng et al., 2014). More importantly, it is limited to measure climate change only from the perspective of regional carbon emissions, because it does not take into account the carbon sequestration capacity of the earth, and it cannot directly reflect the impact of the region’s development on climate change. CEF is an ideal indicator, it can measure the greenhouse effect by considering both carbon emissions and carbon absorption and can directly reflect the land with biological productivity occupied by cities with different density gradients (metropolitan or small cities) due to carbon emissions (Minx et al., 2013).

In view of these problems, this paper measures the CEF of 253 cities in China from 2010 to 2019 and uses the spatial econometrics model to explore how the intensity of LEPE affects urban CEF. We use the dynamic spatial Durbin model (DSDM) to decompose the spatial effect into direct effect and indirect effect (including long-term and short-term), which aims to fully explore the spatial interaction among cities. The nonlinear regression results show that the relationship between LEPE and CEF presents an inverted U-shape and a specific fitting curve graph is drawn. The moderating effect of environmental rights trading schemes and low-carbon pilot policy are considered, respectively, it is to explore how these two policies affect the relationship between the intensity of LEPE and the urban CEF. Besides, regional heterogeneity and temporal heterogeneity are also analyzed. These conclusions also have reference value for other countries aiming at carbon emission reduction, especially for developing countries.

The contributions of this research are as follows. Firstly, in the context of fiscal decentralization about environmental expenditure, the negative externality of environmental pollution will drive local governments to compete with each other in environmental protection expenditure. Meanwhile, some studies have ignored this interaction. We use the DSDM model to decompose the spatial effect into indirect effect and direct effect. The result provides apparent evidence for the competition behavior of LEPE among local governments. Secondly, it is not enough to measure the greenhouse effect only by carbon emissions. Because it only considers the perspective of emissions, but ignores the carbon absorption, such as carbon fixation in the ocean and forest. This paper takes CEF as the research perspective to intuitively reflect the ecological occupation of cities on the earth’s land, reflects the role of environmental protection expenditure on climate change, and expands the research ideas. Thirdly, most studies focus more on the CEF of metropolises than on small-sized and medium-sized cities (Moran et al., 2018). We calculate the CEF of 253 cities (small-sized and medium-sized cities) in China after referring to the latest vegetation carbon storage (boreal, temperate, subtropical, tropical forests) and measuring the carbon sequestration capacity of the ocean (Harris et al., 2021). Fourthly, the ordinary panel models are applied by many studies about the urban greenhouse effect, which ignore the spatial spillover effect, spatial correlation, and difference in carbon emissions between cities (Cheng et al., 2021b). To fully estimate the spatial effect between urban agglomerations, we introduce a spatial econometrics model with dynamic panel data. Moreover, a series of further explorations have been conducted, including nonlinear regression, moderating effects, regional heterogeneity, and temporal heterogeneity. Significant reference value is shown in these results for policymakers.

The remainder of this paper is arranged as follows. Section “Data, theoretical methods, and models” is data sources, theoretical methods, and models. Section “Results and analysis” presents the analysis of empirical results. Section “Implications of the research for policymakers” discusses the impact of empirical results on policymakers’ decisions. The section “Conclusion and outlook” is the conclusion and outlook.

Data, theoretical methods, and models

Data source and specification

The dependent variable in this paper is CEF, it is calculated from carbon emission data. According to Wu and Guo (2016), the carbon emission data is obtained by all kinds of energy consumed in cities times the corresponding conversion factorFootnote 2. The explanatory variable is the intensity of LEPE, and the data is from the China City Statistical Yearbook. Control variables include population size, economic level, urbanization rate, industrial level, foreign direct investment, and green innovation. The green innovation data comes from China Research Data Service (CNRDS), and the rest comes from the China Statistical Yearbook, China Regional Statistical Yearbook, and other statistical bulletins. See Table 1 for descriptive statistics of each variable. The reasons for selecting data from 2010 to 2019 are as follows. Firstly, one of our goals is to explore the synergistic effect between LEPE and the low-carbon economic policy, but the low-carbon economic policy has been implemented since 2010. Secondly, although the LEPE policy began in 2007, the data with a lag of several years is conducive to observe a more stable internal relationship. Thirdly, the databases of some important variables are only updated until 2019.

Table 1 Descriptive statistics.
Full size table

Calculation of CO2 ecological footprint

The CEF is defined as the ecological land required which absorb the carbon dioxide emitted by a certain area, including cropland, grazing land, forest land, etc. However, most carbon storage is borne by forest, so we use the area of forest land that absorbs carbon dioxide to measure CEF (Borucke et al., 2013). The calculation formula is as follows:

$${{{\mathrm{CEF}}}}_{it} = \left[ {\frac{{C_{it}\left( {1 - S_{\rm {{absorb}}}} \right)}}{{\overline {\rm {{NEPc}}} }} \times {\rm {EQF}}_{\rm {c}}} \right]/P_{it}$$
(1)
$$\overline {\rm {{NEPc}}} = \frac{{{\rm {Gc}}}}{{{\rm {Ac}}}} = \frac{{\mathop {\sum}\nolimits_{j = 1}^n {{\rm {Gc}}_j} }}{{\mathop {\sum}\nolimits_{j = 1}^n {{\rm {Ac}}_j} }}$$
(2)

In Eq. (1), CEFit is per capita CO2 ecological footprint of city i in the t year, Cit is carbon emission of city i in the t year, Pit is the resident population of city i in the t year. Sabsorb represents carbon sequestration capacity of the oceanFootnote 3 EQFc means the equivalent factor for calculating CEF and is 1.26 based on the accounts of the Global Footprint Network version 2022. CEF is just one of the six types of ecological footprints, this means that different types of ecological footprints are measured by different types of bio-productive land. EQFc is introduced to standardize various types of bio-productive land and facilitate direct comparisons of CEF. The EQFc is 1.26 based on the accounts of the Global Footprint Network version 2022. \(\overline {{{{\mathrm{NEPc}}}}}\) is annual average carbon sequestration capacity of global forests (t/hm2), Eq. (2) reports its specific calculation method. Gc represents the carbon storage of global forests within a year, and Ac is the total area of global forests. The subscript j indicates the forest type, including boreal, temperate, subtropical, and tropical forests. The carbon absorption rate varies for each type of forest.

Spatial autocorrelation tests

The negative externality of pollution causes local governments may have a competitive effect on LEPE. Specifically, the intensity of LEPE in a local city will be affected by the intensity of LEPE in surrounding cities, and vice versa. We use spatial econometrics models to evaluate this spatial spillover effect, including direct and indirect effects. Firstly, spatial autocorrelation tests are conducted on CEF data to explore the systematic change rule in space. The global Moran’s index calculation formula is as follows:

$$I = \frac{{\mathop {\sum}\nolimits_{i = 1}^n {\mathop {\sum}\nolimits_{j = 1}^n {w_{ij}( {{\rm {CEF}}_i - \overline {{\rm {CEF}}} })} }( {{\rm {CEF}}_j - \overline {{\rm {CEF}}} } )}}{{S^2\mathop {\sum}\nolimits_{i = 1}^n {\mathop {\sum}\nolimits_{j = 1}^n {w_{ij}} } }}$$
(3)

The thought of Moran’s index is to seek the relationship between a variable and its spatial lag term. In Eq. (3), the value range of Moran’s I statistic is [− 1, 1]. The closer its absolute value to 1, the stronger the correlation is. When it is positive, the spatial correlation in CEF data is positive, that is, the closer the distance, the more similar the observed values are to each other. When it is negative, the spatial correlation in CEF data is negative, meaning that the closer the distance, the more discrepant the observed values are. n is the total number of spatial units and wij is the spatial weight. \(S^2 = \frac{1}{n}\mathop {\sum }\nolimits_{i = 1}^n ({{{\mathrm{CEF}}}}_i - \overline {{{{\mathrm{CEF}}}}} )^2\), CEFi is CO2 ecological footprint of the ith city, \(\overline {{{{\mathrm{CEF}}}}}\) is the arithmetic mean of CEF data in all cities.

Global spatial autocorrelation only reflects the overall spatial characteristics, which does not mean that there is a spatial correlation between each spatial individual and their surrounding cities. Therefore, it is necessary to analyze the spatial characteristics of each individual. The local indicator of spatial association (LISA) proposed by Anselin (1995) can decompose the global Moran’s index into the local Moran’s index.

$$I_i = \frac{{\left( {{\rm {CEF}}_i - \overline {{\rm {CEF}}} } \right)}}{{S^2}}\mathop {\sum}\limits_{j \ne i} {w_{ij}{\left( {{\rm {CEF}}_j - \overline {{\rm {CEF}}} }\right)}}$$
(4)

The meaning of each variable in Eq. (4) is the same as that in Eq. (3). When Ii is positive, it indicates that high CEF cities are surrounded by high CEF cities (H–H type), or low CEF cities are surrounded by low CEF cities (L–L type). When Ii is negative, it indicates that high CEF cities are surrounded by low CEF cities (H–L type), or low CEF cities are surrounded by high CEF cities (L–H type). Besides, LISA can also identify “cluster area” (clusters of high or low observations), as well as “hot spot area” (areas with completely different observations from surrounding cities).

For wij, this paper selects four types of spatial weight matrices to ensure the robustness of the results. The first is the Queen adjacency matrix (Queen), wij is assigned 1 if there is a common boundary or a connection point between two cities, otherwise it is 0. The second is the inverse distance matrix (Dist), wij is reciprocal of Euclidean distance between two cities. The third is the economic distance matrix (Econ), wij is the reciprocal of difference-value between economic levels of two cities. The fourth is the economic geography matrix (Econ-Dist), it not only considers economic connection between two cities, but also geographical distance. In this paper, it is the Hadamard product of the economic distance matrix and the inverse distance matrix.

Spatial Durbin model

If CEF data passes the spatial autocorrelation test, it indicates that there is a spatial spillover effect. In addition, as mentioned above, there may be competitive behavior in the LEPE. Therefore, it is appropriate to use spatial econometric models for research. We choose the spatial Dubin model that considers both spatial lag and spatial error. The fundamental formula is as follows:

$$\begin{array}{l}{{{\bf{ln}}}}{{{\bf{CEF}}}}_t = \rho {{{\boldsymbol{W}}}}{{{\bf{ln}}}}{{{\bf{CEF}}}}_t + {{{\boldsymbol{\alpha }}}}_{{{0}}} + \alpha _1{{{\bf{ln}}}}{{{\bf{ILEPE}}}}_t\\ \qquad \qquad \quad + \alpha _2{{{\boldsymbol{W}}}}{{{\bf{ln}}}}{{{\bf{ILEPE}}}}_t + {{{\boldsymbol{X}}}}_t{{{\boldsymbol{\beta }}}}_1 + {{{\boldsymbol{WX}}}}_t{{{\boldsymbol{\beta }}}}_2 + {{{\boldsymbol{\varphi }}}}_{{{i}}} + {{{\boldsymbol{\omega }}}}_{{{t}}} + {{{\boldsymbol{\varepsilon }}}}_{{{{it}}}}\end{array}$$
(5)

In Eq. (5), ρ is spatial lag parameter, it can reflect the spatial spillover effect if it is statistically significant. When the spatial lag parameter is positive, it denotes that the CEF of a city and the CEF of its neighboring cities are positively correlated. lnCEF = (lnCEF1, lnCEF2,⃛lnCEFN)′, lnILEPE = (lnILEPE1, lnILEPE2,⃛lnILEPEN)′, t = 1, 2, …, T. W is the spatial weight matrix, X represents a set of control variables, including population size, economic level, urbanization rate, industrial level, foreign direct investment, and green innovation. φi and ωt are, respectively, N × 1 vectors consisting of city-fixed effect and time-fixed effect. εit is N × 1 vector consisting of the spatial error term, each element in this vector is i.i.d. disturbance term and is a normal distribution with a mean of 0. ρ, α1, α2 are both scalars.

It can estimate long-term and short-term trends of spatial effects simultaneously by dynamic panel spatial Dubin models. Referring to Elhorst et al. (2013), the DSDM model constructed is as follows:

$$\begin{array}{l}{{{\bf{ln}}}}{{{\bf{CEF}}}}_t = \tau {{{\bf{ln}}}}{{{\bf{CEF}}}}_{t - 1} + \eta {{{\boldsymbol{W}}}}{{{\bf{ln}}}}{{{\bf{CEF}}}}_{t - 1} + \rho {{{\boldsymbol{W}}}}{{{\bf{ln}}}}{{{\bf{CEF}}}}_t + {{{\boldsymbol{\alpha }}}}_{{{0}}} + \alpha _1{{{\bf{ln}}}}{{{\bf{ILEPE}}}}_t\\ \qquad \qquad \quad + \,\alpha _2{{{\boldsymbol{W}}}}{{{\bf{ln}}}}{{{\bf{ILEPE}}}}_t + {{{\boldsymbol{X}}}}_t{{{\boldsymbol{\beta }}}}_1 + {{{\boldsymbol{WX}}}}_t{{{\boldsymbol{\beta }}}}_2 + {{{\boldsymbol{\varphi }}}}_{{{i}}} + {{{\boldsymbol{\omega }}}}_{{{t}}} + {{{\boldsymbol{\varepsilon }}}}_{{{{it}}}}\end{array}$$
(6)

Compared with Eq. (5), Eq. (6) has a one-period lag term for the dependent variable. τ and η are both scalars. π = 0 indicates that it does not have a time lag of the dependent variable. η = 0 indicates that it does not have a spatiotemporal lag of the dependent variable. Both of these lag terms exist if neither τ nor η is 0. Equation (6) can also be rewritten in the following matrix form (Elhorst, 2012):

$$\begin{array}{l}{{{\bf{ln}}}}{{{\bf{CEF}}}}_t = \left( {{{{\boldsymbol{I}}}} - \rho {{{\boldsymbol{W}}}}} \right)^{ - 1}\left( {\tau {{{\boldsymbol{I + }}}}\eta {{{\boldsymbol{W}}}}} \right){{{\bf{ln}}}}{{{\bf{CEF}}}}_{t - 1} + \left( {{{{\boldsymbol{I}}}} - \rho {{{\boldsymbol{W}}}}} \right)^{ - 1}\left( {{{{\boldsymbol{X}}}}_t{{{\boldsymbol{\alpha }}}} + {{{\boldsymbol{WX}}}}_t{{{\boldsymbol{\beta }}}}} \right)\\ \qquad \qquad \quad + \,\left( {{{{\boldsymbol{I}}}} - \rho {{{\boldsymbol{W}}}}} \right)^{ - 1}{{{\boldsymbol{V}}}}\end{array}$$
(7)

In Eq. (7), Xt represents a matrix consisting of all dependent variables at time t. V is N × 1 vector consisting of intercept, fixed effects, and error terms. \({\bf {\ln}} {\bf {CEF}}_{t - 1}\) does not exist for a specific time point (one cross-section). The subscript t is omitted to simplify the symbol, and the partial derivative matrix of lnCEF with respect to the kth explanatory variable of X is

$$\begin{array}{l}\left[ {\begin{array}{*{20}{c}} {\frac{{\partial {{{\bf{ln}}}}{{{\bf{CEF}}}}}}{{\partial x_{1k}}}} & \ldots & {\frac{{\partial {{{\bf{ln}}}}{{{\bf{CEF}}}}}}{{\partial x_{Nk}}}} \end{array}} \right]_t = \left[ {\begin{array}{*{20}{c}} {\frac{{\partial \ln CEF_1}}{{\partial x_{1k}}}} & \ldots & {\frac{{\partial \ln CEF_1}}{{\partial x_{Nk}}}} \\ \vdots & \ddots & \vdots \\ {\frac{{\partial \ln CEF_N}}{{\partial x_{1k}}}} & \ldots & {\frac{{\partial \ln CEF_N}}{{\partial x_{Nk}}}} \end{array}} \right]\\ \qquad \qquad\qquad \qquad\qquad \quad\,\, = \left( {{{{\boldsymbol{I}}}} - \rho {{{\boldsymbol{W}}}}} \right)^{ - 1}\left[ {\alpha _k{{{\boldsymbol{I + }}}}\beta _k{{{\boldsymbol{W}}}}} \right]\end{array}$$
(8)

Equation (8) is the short-term impact of the kth explanatory variable on CEF. If the time is considered long enough, it includes \({\bf {\ln}} {\bf {CEF}}_{t - 1}\). The partial derivative matrix of the kth explanatory variable in the whole Eq. (7) is the long-term impact of this variable on CEF:

$$\left[ {\begin{array}{*{20}{c}} {\frac{{\partial {{{\bf{ln}}}}{{{\bf{CEF}}}}}}{{\partial x_{1k}}}} & \ldots & {\frac{{\partial {{{\bf{ln}}}}{{{\bf{CEF}}}}}}{{\partial x_{Nk}}}} \end{array}} \right] = \left[ {\left( {1 - \tau } \right){{{\boldsymbol{I}}}} - \left( {\rho + \eta } \right){{{\boldsymbol{W}}}}} \right]^{ - 1}\left[ {\alpha _k{{{\boldsymbol{I + }}}}\beta _k{{{\boldsymbol{W}}}}} \right]$$
(9)

In Eqs. (8) and (9), αk and βk are both scalars. The expansion equation of matrix [αkI + βkW] is as follows:

$$\left[ {\alpha _k{{{\boldsymbol{I + }}}}\beta _k{{{\boldsymbol{W}}}}} \right] = \left[ {\begin{array}{*{20}{c}} {\alpha _k} & {w_{12}\beta _k} & \ldots & {w_{1N}\beta _k} \\ {w_{21}\beta _k} & {\alpha _k} & \ldots & {w_{2N}\beta _k} \\ \vdots & \vdots & \ddots & \vdots \\ {w_{N1}\beta _k} & {w_{N2}\beta _k} & \ldots & {\alpha _k} \end{array}} \right]$$
(10)

In Eq. (10), the average value of the main diagonal element represents the average direct effect of the kth explanatory variable on CEF. The average total effect is the average value of the sum of each column in the matrix, and the average indirect effect is the average total effect minus the average direct effect.

There are many kinds of spatial econometrics models, so a series of tests are needed to verify why the spatial Dubin model was chosen in this paper. The first is to use the LM test and robust LM test. If both of these tests reject the null hypothesis of “no spatial lag” and “no spatial error”, the spatial Dubin model can be preliminarily selected for estimation. However, Elhorst (2014) thinks that conclusions drawn from the LM test and robust LM test should still be used with caution. Therefore, the constraint tests can be used to determine the final model selection based on the estimated coefficients after using the spatial Durbin model. These two tests, which also incorporate spatial lag and spatial error tests, are the Wald and LR tests.

Results and analysis

Visualization of CO2 ecological footprint

The CEF data of 253 cities from 2010 to 2019 is shown in Fig. 1. The samples are divided into five urban agglomerations according to the size of latitude: the southern coastal urban agglomeration, the Yangtze River basin urban agglomeration, the capital urban agglomeration, the northern urban agglomeration, and the northeast urban agglomeration. The latitude of 253 cities is arranged from low to high on the horizontal axis. From a latitudinal perspective, there are significant geographical differences in the CEF per capita. The lowest level exists in the Yangtze River basin, while the capital and northern urban agglomerations appear relatively high. It should be noted that the northeast urban agglomeration has the highest per capita CEF; this is due to the fact that this urban agglomeration is mainly heavy industrial and energy-output cities, whose fossil fuel consumption is continuously high. Chronologically, the area of the blue block is decreasing, while the red area is increasing. This indicates that the per capita CEF of urban samples shows an overall increasing trend during 2010–2019.

Fig. 1: Spatial distribution of CO2 ecological footprint.
figure 1

The latitude of all cities is arranged from low to high on the horizontal axis.

Full size image

Spatial autocorrelation of CO2 ecological footprint

The results of the global spatial autocorrelation test are presented in Table A1 in Appendix A. No matter what kind of spatial weight matrix, the null hypothesis that there is no spatial autocorrelation is rejected within the time range studied. It demonstrates that CEF data are interdependent among the spatial units, thus it is appropriate to use the spatial econometrics models for research.

To understand specifically the correlation characteristics of CEF among every spatial individual, the local autocorrelation test is conducted. Taking the economic geography weight matrix as an example, the Moran scatter plot is shown in Fig. 2. X-axis is the observed value of CEF, and Y-axis is the spatial lag of CEF. The slope of the red regression line in each figure corresponds to the global Moran’s index in Table A1 in Appendix A. The figure is divided into four quadrants. Quadrant 1st to 4th corresponds to H–H (High High), L–H (Low High), L–L (Low Low), H–L (High Low) type regions, respectively. Most of the samples are located in the first and third quadrants, indicating that the CEF of most cities presents a high degree of aggregation around themselves.

Fig. 2: Moran scatter plot designed from the economic geography weight matrix.
figure 2

This figure reflects the spatial correlation between each spatial individual.

Full size image

In order to more intuitively present the spatial dynamic evolution of CEF observations, taking the Queen weight matrix as an example, Fig. 3 is the LISA map based on the local Moran’s index. The deep red is the H–H type region where local spatial autocorrelation is statistically significant at the 5% level. The area of H–H type regions, mainly distributed in central Inner Mongolia, has become increasingly wide over time. As China’s energy-exporting cities, these regions in Inner Mongolia hold higher CEF levels. Cities in the deep blue region (L–L type) are mainly located in the Yangtze River basin and the southern coastal area. From 2010 to 2019, the area of H–H type regions is gradually increasing, indicating the growth of CEF aggregation degree. The area of L–L type regions does not change obviously, but the degree of regions’ aggregation has increased, and there is a slight tendency to shift eastward. The light blue and light red regions are L–H and H–L type regions, respectively, which gradually decrease over time and almost disappear by 2019. The L–H type and H–L type indicate the dispersion degree and difference in the observed values, so the reduction of these regions also from the opposite side shows that CEF data in our research samples is increasingly aggregated.

Fig. 3: The LISA map designed from the Queen weight matrix.
figure 3

a–d The local Moran’s index for each city in the years 2010, 2013, 2016, and 2019, respectively. This figure depicts the visualization of the local indicator of spatial association (LISA).

Full size image

Baseline regressions analysis

After identifying the spatial autocorrelation, we first use the LM test and robust LM test to select model. The results are shown in Table A2 in Appendix A. No matter what kind of spatial weight matrix, LM-error, LM-lag, Robust LM-error, and Robust LM-lag tests reject the null hypothesis. This indicates that the model we constructed has both spatial error and spatial lag. Consequently, we preliminarily select the spatial Durbin model for estimation.

Table 2 shows baseline regression results with the economic geography weight matrix, including the fixed effect regression, and the main effect of spatial Durbin regression. The R2 of the spatial Durbin model is significantly improved compared to the ordinary fixed effect model. Moreover, the statistical significance of the explanatory variable and control variables has been increased. It suggests that it is appropriate to use spatial econometric models in our data. From the estimation results of the spatial Durbin model, LR and Wald tests reject the null hypothesis. This indicates that using the spatial Durbin model is robust. Regardless of the model, the estimated coefficient of lnILEPE is significantly negative. Thus, the current intensity of LEPE can inhibit CEF as a whole. The results are relatively robust because the estimation results of the three types of DSDM models are similar. The R2 of model 4 is significantly smaller than that of model 3 and model 5, mainly because the coefficient for the spatiotemporal lag term of lnCEF (W*lnCEF_L1.) is not significant. The coefficient for the time-lag term of lnCEF (lnCEF_L1.) is significantly 0.9360, indicating that the CEF of the local city in the previous period promotes the one in the current period. In Models 3 and 5, the coefficient of W*lnCEF_L1. is positive but not significant. In view of this, the CEF of surrounding cities in the previous period might not significantly increase the CEF of local cities in the current period.

Table 2 The results of baseline regressions.
Full size table

Dubé and Legros (2014) claimed that it was effective to avoid biased point estimation in spatial regression by decomposing spatial effects into direct and indirect effects. Model 5 with time lag and spatiotemporal lag of the dependent variable in Table 2 has the highest R-squared, so this model is used to decompose long-term and short-term spatial effects. The results are shown in Table 3. In the long term, the direct effect coefficient of lnILEPE is significantly negative, indicating that the current intensity of LEPE inhibits the local city’s CEF. The indirect effect coefficient of lnILEPE is significantly negative, and there may be two possible mechanisms for its impact. One is that the current intensity of LEPE in surrounding cities directly curbs CEF in the local city. And the other is that the current intensity of LEPE in surrounding cities inhibits their own CEF, which in turn lowers the local city’s CEF. In the short term, the direct effect coefficient of lnILEPE is significantly negative, indicating that the current intensity of LEPE has an immediate inhibitory effect on the local city’s CEF. However, the short-term indirect effect of LEPE is not statistically significant.

Table 3 Results of decomposition of spatial effects.
Full size table

The long-term indirect negative effect of LEPE on CEF provides clear evidence that there is game competition among local governments in terms of environmental expenditure. That is, due to the spatial spillover effect of CEF, the LEPE of city A will reduce CEF of surrounding city B. The neighboring city B will take a negative attitude towards LEPE if the intensity of LEPE in city A is high. The motivation behind this negative attitude is that LEPE of city A can be advantageous to reduce CEF of neighboring city B, allowing city B to reduce its LEPE and devote more fiscal spending to economic development. It is worth noting that this competition behavior of LEPE among local governments is not conducive to sustainable development (Feng et al., 2023).

We add the quadratic term of lnILEPE in the spatial Durbin model to explore dynamic change in the relationship between the intensity of LEPE and CEF. Table 4 shows the results. The estimation results of four spatial weight matrices are similar, indicating the robustness of the results. After adding the variable (lnILEPE)2, for both the main effect and the direct effect, the coefficient of (lnILEPE)2 shows negative, and the coefficient of lnILEPE changes from negative to positive. The coefficients of both are statistically significant at the 5% level. It reveals that there is an inverted U-shaped relationship between the intensity of LEPE and the CEF. When the intensity of LEPE gradually increases, it will not immediately inhibit CEF. Only when the intensity of LEPE increases to a certain extent, it will reduce CEF, and this inhibitory effect will become stronger as the intensity of LEPE continues to increase.

Table 4 Nonlinear relationship between the intensity of LEPE and the CEF.
Full size table

To visually display this inverted U-shaped dynamic change, ignoring the impact of control variables, a fitting curve of the non-linear relationship between the intensity of LEPE and the CEF is drawn based on the estimated coefficients in Table 4. The results are shown in Fig. 4. No matter what kind of spatial weight matrix, the inflection point of the nonlinear relationship is within the range of the studied sample. Because we define “LEPE/regional GDP” as the intensity of LEPE, this inverted U-shaped relationship can be explained by the EKC curve to a certain degree (Li et al., 2021). For an underdeveloped region, the CEF level is apparently low while the local government may devote little expenditure to environmental protection. With the preliminary development of the economic level, CEF will not decrease, even though the intensity of LEPE begins to increase. This is mainly because economic development at this stage often comes at the expense of the environment. When the economy reaches a certain level, the government begins to realize the importance of green development, which will forcefully expand the intensity of LEPE, and CEF will begin to gradually decrease.

Fig. 4: The curve fitting of nonlinear relationship.
figure 4

a–d The curve fitting of nonlinear relationship based on the regression results of matrices Queen, Dist, Econ, and Econ-Dist in Table 4, respectively. This figure is a visualization of nonlinear regression results.

Full size image

Heterogeneity analysis

In order to cope with the adverse effects of climate change, Chinese government has begun to implement the low-carbon economic policy, which can promote the construction of environmental-friendly and sustainable energy ecosystems in cities. Since 2010, China has been enforcing a low-carbon economic development model in some cities which are now known as “low-carbon pilot cities”. It is necessary for low-carbon pilot cities to control the total amount of greenhouse gas emissions in their respective regions and make allocation plans for greenhouse gas emissions. In the samples of this paper, a total of 89 cities have implemented the low-carbon economic policy. A dummy variable is introduced in the model to figure out the synergistic effect between this policy and LEPE in suppressing CEF. If a sample is the low-carbon pilot city, the dummy variable is set to 1, otherwise it is 0. The results are shown in Table 5, and the coefficient of the dummy variable is absorbed by city fixed effect. It can be seen that the coefficient of interaction term is present negative at a significance level of 5%, with the exception of the indirect effects in the Econ matrix and the Econ-Dist matrix. It demonstrates that LEPE performs more effective in suppressing CEF for those cities implementing low-carbon policy. The market economy means that local governments are unlikely to take too many special environmental policies to achieve the goal of energy conservation and emission reduction. The policy of low-carbon pilot cities may be a good solution, which can have a synergistic effect with LEPE, thereby producing the “1 + 1 > 2” effect in abating CEF.

Table 5 The moderating effect of the low-carbon pilot policy.
Full size table

Fig. 1 shows significant spatial difference in the CEF per capita of 253 cities, and Fig. 2 illustrates a high degree of spatial aggregation of CEF data. In addition, there are also significant differences in economic level, industrial structure, and population size among different regions in China. Therefore, the samples are divided into several urban agglomerations based on their geographic location to investigate regional heterogeneity. According to Fig. 1, the five urban agglomerations are divided into: the southern coastal urban agglomeration, the Yangtze River basin urban agglomeration, the capital urban agglomeration, the northern urban agglomeration, and the northeast urban agglomeration. Four dummy variables are introduced to re-estimate the model: Southern, the city belonging to the southern coastal urban agglomeration is 1, otherwise it is 0; Yangtze, the city belonging to the Yangtze River basin urban agglomeration is 1, otherwise it is 0; North, the city belonging to the northern urban agglomeration is 1, otherwise it is 0; Capital, the city belonging to the capital urban agglomeration is 1, otherwise it is 0.

Table 6 lists the estimated results of regional heterogeneity. Only the main and direct effects are discussed since most of the indirect effects’ coefficients are not significant. The coefficient of lnILEPE represents the impact of LEPE on CEF in the northeast urban agglomeration. The sum of the coefficients of lnILEPE plus South*lnILEPE is the impact of LEPE on CEF in the southern coastal urban agglomeration. However, because the coefficient of South*lnILEPE is not significant, it can be considered that there is no obvious regional difference between the southern coastal urban agglomeration and the northeast urban agglomeration. The coefficients of Yangtze*lnILEPE, North*lnILEPE, and Capital*lnILEPE are significantly positive, indicating that compared to northeast urban agglomeration, LEPE in the Yangtze River basin urban agglomeration, northern urban agglomeration, and capital urban agglomeration exhibit a weaker effect in reducing CEF. As a heavy industrial and energy export region in China, the northeastern cities hold the highest CEF per capita (see Fig. 1). The intensity of LEPE in these cities is relatively high, so the inhibitory effect on CEF is the strongest. It is worth noting that due to the developed economic level, large population size, and high intensity of LEPE, the inhibitory effect in the Yangtze River basin cities ranks second among the five urban agglomerations.

Table 6 Regional heterogeneity results.
Full size table

China has begun to enforce environmental protection tax since 2018. The tax law will replace the pollution fees previously which the Ministry of Environmental Protection charges from enterprises, meaning stricter penalties for pollution. Some scholars believe that China has effectively curbed its ecological footprint after imposing an environmental protection tax (Fang et al., 2023). In order to study the difference in the impact of LEPE on CEF before and after the levying of environmental protection tax, we divided research samples into two groups accordingly. After the environmental protection tax was levied in 2018, the dummy variable was assigned as 1, otherwise, it is 0. The estimated results are shown in Table 7, where the coefficient of the dummy variable is absorbed by the time-fixed effect. The coefficient of Tax*lnILEPE is not significant, regardless of the main effect, direct effect, or indirect effect. Consequently, there is no significant difference in the inhibitory effect of LEPE on CEF before and after the implementation of the environmental protection tax. This may be because China’s environmental protection tax is directly superseded by the pollutant fees, the tax rate has not significantly increased compared to the pollutant fees, signifying that actual punishment intensity for polluting enterprises has not changed significantly. From another point of view, the environmental protection tax has only been levied for two years within the time range of our study, so this difference before and after the enactment of the environmental protection tax has not yet been significantly reflected.

Table 7 Temporal heterogeneity results.
Full size table

Robustness and endogeneity

The above tables have all conducted constraint tests, indicating the robustness of using the spatial Durbin model. In addition, nonlinear relationship research, moderating effects analysis, and heterogeneity analysis have all reported regression results in four types of spatial weight matrices, ensuring the robustness of results. Since Table 3 only reports the spatial decomposition effect of the DSDM model in the economic geography weight matrix, the estimation of the other three spatial weight matrices is supplemented in Table A3 in Appendix A. The coefficients of lnILEPE are similar to those in Table 3, regardless of the long-term or short-term effects. The time lag and spatiotemporal lag terms of the dependent variable are similar to the estimation results of model 5 in Table 2. Thus, the robustness of the results is verified.

There may be endogenous problems for two reasons. One is sample selection bias because this paper selects samples based on the availability of data, and the other is that omitted variables may bring an endogenous problem. We choose the GMM estimation method to re-estimate the DSDM model in Eq. (6). Results are shown in Table A4 in Appendix A. No matter what kind of spatial weight matrix, the estimated results are similar to those of the DSDM model in Table 2. This indicates that the estimation results in this paper are reliable.

Implications of the research for policymakers

Our results illustrate the spatial spillover effect of CEF. Table 3 shows that there is a negative indirect effect of LEPE on CEF, that is, the environmental protection expenditure of a certain city will help to curb CEF of surrounding cities. Therefore, local governments may compete to adopt a negative attitude towards LEPE, which is not conducive to diminishing CEF. Although the power of government finance for environmental expenditure has been delegated, the central government should appropriately intervene to coordinate LEPE among cities and arrange for the rational intensity of LEPE between cities to fit their own urban economic development. It can turn competition into cooperation, not only achieving the common goal of reducing CEF but also ensuring the green and sustainable development of their own cities.

The results in Table 4 and Fig. 4 indicate that there is an inverted U-shaped relationship between the intensity of LEPE and the CEF. For those small-sized and medium-sized cities that are underdeveloped and not highly industrialized, the government should adopt a lower intensity of LEPE and invest more financial resources in economic development. Because only when increasing the intensity of LEPE to a certain extent, it can be advantageous to reducing CEF. However, it is obviously inappropriate for small-sized and medium-sized cities with poor economies to invest too much in environmental protection expenditure. For prosperous large cities with dense populations, it is necessary to appropriately increase the intensity of LEPE, which can be much more efficient in increasing their CEF. Another advantage of doing so is that it can help some small and medium-sized cities in the vicinity reduce their CEF.

Table 5 illustrates that it is necessary for the government to accelerate the promotion of low-carbon economic policy. For example, the relevant departments can make more cities become low-carbon pilot cities until the full implementation of the low-carbon strategy, in which case the effect of LEPE can be intensified. Figure 1 shows that the per capita CEF in the northern urban agglomeration and the capital urban agglomeration is at a relatively high level. However, the results of regional heterogeneity indicate that the LEPE of these two urban agglomerations is not effective enough in abating CEF relative to other urban agglomerations. Therefore, it is imperative to increase the intensity of LEPE to reduce CEF for these two urban agglomerations. The results of temporal heterogeneity show that the effectiveness of LEPE is not significantly different before and after the implementation of the environmental protection tax in China. The environmental protection tax should fully play its role as a government’s environment-related fiscal revenue. For example, more tax payments should be thrown into environmental protection expenditure to seek low-carbon development.

Conclusion and outlook

This paper calculates the CEF of 253 cities in China during the period 2010–2019 and finds that there is a significant spatial spillover effect in CEF. Under the context of fiscal decentralization in environmental expenditure, we find that LEPE has a significant inhibitory influence on CEF based on the spatial Durbin model. The spatial effect is decomposed into long-term and short-term effects. In the long run, the LEPE of a certain city presents the CEF inhibition influence not only on the local area but also on the surrounding cities. This result provides plain evidence for the game competition on environmental protection expenditure among local governments. In the short term, the LEPE of a certain city has the CEF inhibition influence on itself, but the influence on the surrounding cities is not significant.

Further research shows that there is an inverted U-shaped relationship between the intensity of LEPE and the CEF, and a fitting curve of this relationship is drawn. The moderating effects research indicates that executing a low-carbon pilot policy can strengthen the influence of LEPE in suppressing CEF. According to spatial location, the study samples are divided into five urban agglomerations. For the effect of LEPE on reducing CEF, the regional heterogeneity results show that there are significant differences among the Yangtze River basin urban agglomeration, the northern urban agglomeration, the capital urban agglomeration, and the northeast urban agglomeration, but the effect of LEPE is similar between the southern coastal urban agglomeration and the northeast urban agglomeration. The temporal heterogeneity results indicate that before and after the enforcement of the environmental protection tax in China, there has been no significant change in the effect of LEPE on curbing CEF.

We determine that the model has both a spatial error and a spatial lag by using LM and robust LM tests. Constrained LR and Constrained Wald tests are also used to ensure the robustness of model selection. For each model, four types of spatial weight matrices are utilized to estimate separately, ensuring the robustness of estimation results. Additionally, the DSDM model of GMM estimation is used, and the results demonstrate the reliability of the conclusions made in this study. Last but not least, our research gives an important reference for governments in formulating and adjusting market-oriented environmental policies.

The latest database of some key variables in this article is only updated to 2019. For temporal heterogeneity research, the implementation time of environmental protection tax is not long enough in our data, we cannot explore any further conclusions about the interaction between LEPE and environmental protection tax on CEF reduction. LEPE and environmental protection tax represent environment-related spending and revenue respectively, the influence of both on the ecological environment is very worthy of study. In the future, with more data about environmental protection tax, new conclusions can be plumbed about those two policies in China. In another way, we can treat developed countries as research objects to study potential interaction between them. For example, the synergy and complementarity between LEPE and environmental protection tax may exist in high-quality economic development.