This study utilizes panel data from Chinese A-share listed companies for the period 2013–2023. The sample is carefully filtered through several steps to ensure data quality and consistency. First, firms labeled as ST, *ST, or PT are excluded due to potential financial distress or abnormal trading status. Second, companies in the financial sector are removed because of their distinct regulatory environment. Third, firms that shift between treatment and control groups during the sample period are excluded to maintain group stability. Fourth, treatment group firms with fewer than one observation prior to policy implementation are omitted to ensure adequate pre-treatment data. Fifth, firms with only a single observation are excluded. Finally, all continuous variables are winsorized at the 1st and 99th percentiles to reduce the influence of outliers. The data are drawn from the CSMAR database and supplemented with information from the annual reports of listed companies.

To evaluate the proposed hypotheses, we construct quantitative models using the variables and data outlined earlier. As a first step, we assess whether environmental regulation influences firm innovation by applying a difference-in-differences (DID) approach. The empirical model is specified as follows:

Model 1

The Effect of Environmental Tax Reform on Firm Innovation P a t e n t i , t = α 0 + α 1 D I D i , t + α 2 S i z e i , t + α 3 L e v i , t + α 4 R E C i , t + α 5 I N V i , t + α 6 C a s h f l o w i , t + α 7 T o p 10 i , t + α 8 B a l a n c e i , t + α 9 B M i , t + α 10 T M T P a y i , t + α 11 M s h a r e i , t + λ i + y e a r t + ε i , t (1)

Here, the dependent variable Patent represents firm innovation level, with i and t denoting firm and year, respectively, α 0 is the intercept, α 1 − α 11 is the coefficients of each variable, λ i is the individual fixed effects, y e a r t is the year fixed effects, and ε i , t is the randomized disturbance term.

Firms often rely on research and development (R&D) investment to drive innovation. This raises a key question: Does environmental regulation promote innovation in heavily polluting firms by encouraging increased R&D spending? To answer this, it is necessary to identify and test the specific channels through which such regulations influence innovation outcomes. This study addresses the question by constructing a mediation model to examine the potential role of R&D investment in this process.

Model 2

The Mediating Role of R&D Investment in the Effect of Environmental Tax Reform on Firm Innovation L R D i n c i , t = α 0 + α 1 D I D i , t + α 2 S i z e i , t + α 3 L e v i , t + α 4 R E C i , t + α 5 I N V i , t + α 6 C a s h f l o w i , t + α 7 T o p 10 i , t + α 8 B a l a n c e i , t + α 9 B M i , t + α 10 T M T P a y i , t + α 11 M s h a r e i , t + λ i + y e a r t + ε i , t (2)

In Equation 2, the dependent variable LRDinc represents firm’s R&D investment, the meaning of each symbol is the same as Equation 1. P a t e n t i , t = α 0 + α 1 D I D i , t + α 2 L R D i n c i , t + α 3 S i z e i , t + α 4 L e v i , t + α 5 R E C i , t + α 6 I N V i , t + α 7 C a s h f l o w i , t + α 8 T o p 10 i , t + α 9 B a l a n c e i , t + α 10 B M i , t + α 11 T M T P a y i , t + α 12 M s h a r e i , t + λ i + y e a r t + ε i , t (3)

In Equation 3, α 1 − α 12 is the coefficient of each variable, and the rest of the symbols have the same meaning as in Equation 1.

Table 2 presents the descriptive statistics for the main variables. The mean value of the enterprise innovation variable (Patent) is 2.820, with a median of 2.996 and a standard deviation of 1.735. The slightly lower mean relative to the median suggests a concentration of firms with lower innovation levels, alongside a subset of firms with notably high patent activity. The values range from 0 to 6.967, indicating that while some firms submitted no patent applications during the year, others recorded substantial innovation output.

The mean R&D investment intensity (LRDinc) is 0.053, with a median of 0.039. The distribution reflects relatively modest R&D spending among most firms, though a few exhibit significantly higher investment levels. The variable ranges from 0 to 0.345, indicating that while some firms allocated no resources to R&D, others invested up to 34.5% of their prior year’s operating revenue.

The treatment group indicator (Treat) has a mean value of 0.195, showing that 19.5% of the sample firms are classified as heavily polluting. The average value of the DID interaction term is 0.117, suggesting that 11.7% of the observations correspond to treated firms in the post-policy period.

To ensure the validity of the linear regression model, it is important to confirm that multicollinearity among independent variables is not severe. High multicollinearity can distort coefficient estimates and weaken the reliability of the results. Therefore, before proceeding with the regression analysis, a multicollinearity test is conducted to evaluate the degree of correlation among explanatory variables. Table 3 presents the results, the average variance inflation factor (VIF) is 1.64, with all individual VIF values below the commonly accepted threshold of 5. This suggests that multicollinearity among the explanatory variables is not a concern and does not compromise the reliability of the regression estimates.

The benchmark regression results, based on the model specified in this study, are summarized in Table 4. Column (1) presents estimates without control variables, revealing a positive and statistically significant coefficient on the DID term, indicating an initial link between the environmental tax reform and increased firm innovation. After including a comprehensive set of control variables in Column (2), the DID coefficient remains positive and significant at the 1% level, with a magnitude of 0.249. This finding suggests that, controlling for firm-specific factors, the environmental tax reform correlates with a 24.9% increase in innovation activity among heavily polluting firms.

The validity of the Difference-in-Differences (DID) approach rests on the assumption that, in the absence of the policy intervention, the treatment and control groups would have followed similar trends. To assess this assumption, a parallel trends test is performed by comparing the pre-policy evolution of the outcome variable across the two groups. This is implemented using a dynamic DID model, specified as Equation 4: P a t e n t i t = α 0 + ∑ s = − 4 − 1   β s p r e t r e a t i × I t − T D = s + ∑ s = 0 5   β s l a s t r e a t i × I t − T D = s + γ X i , t + λ i + y e a r t + ε i , t (4) where β s p r e and β s l a s denote the regression coefficients of the dummy variables in each period before and after the policy implementation respectively, t r e a t i is the identification variable of whether the sample firms deal with the group, which is quantified as 1 if the sample firms are affected by the policy in the data period, and 0 otherwise; I · is the schematic function, t − T D = s denotes the period before and after the policy implementation, s∈ [-5,5], and the rest of the symbols have the same meaning as that of the baseline regression model. Taking the first period of the data cycle as the base period, the test results are shown in Figure 1.

Figure 1 reports the results of the parallel trends test. During the pre-policy period (years −4 to −1), the confidence intervals of the estimated coefficients include zero and are statistically insignificant, indicating no meaningful difference in innovation levels between the treatment and control groups before the policy was enacted. This confirms that the parallel trends assumption holds.

In contrast, the estimated coefficients for the policy implementation year and the following 5 years are positive and statistically significant, as their confidence intervals exclude zero. This suggests that the environmental tax reform led to a notable increase in innovation among firms in heavily polluting industries relative to those in less polluting sectors. These findings support the existence of a dynamic treatment effect, demonstrating that the policy had a sustained positive impact on innovation performance in the targeted firms.

To verify that the observed effects are driven by the environmental tax reform rather than external factors, this study conducts a placebo test focused on innovation outcomes among heavily polluting firms. Following the methodology of Ferrara, the analysis employs 500 iterations of random sampling to construct a pseudo-treatment variable (Ferrara et al., 2012). The resulting distribution of estimated coefficients, p-values, and kernel density curves from these placebo regressions is illustrated in Figure 2.

Figure 2 presents the results of the placebo test. The distribution of regression coefficients from 500 random samples falls within the range of approximately [–0.1, 0.1], which is notably distant from the baseline estimate of 0.252. Additionally, most of the placebo coefficients are statistically insignificant, with p-values exceeding 0.1. The results follow a normal distribution centered around zero, suggesting no systematic effect in the absence of actual policy intervention. These findings confirm the robustness of the main results and support the conclusion that the observed increase in innovation among heavily polluting firms is attributable to the environmental tax reform, rather than to random external shocks.

In this study, heavily polluting firms constitute 19.5% of the sample, while non-heavy polluting firms account for 80.5%. The relatively small size of the treatment group raises concerns about potential sample selection bias, which may lead to endogeneity and bias the regression results. To mitigate this issue, propensity score matching (PSM) is applied before conducting regression analysis.

Three matching approaches are utilized: mixed matching on the full sample, year-by-year matching to control for temporal variations, and individual matching using pre-policy data (2013–2017) transformed into a wide panel format. The individual matching method improves upon the other approaches by addressing discontinuities in the control group sample and enhancing comparability between treated and control firms. This, in turn, increases the reliability of the difference-in-differences estimates. The results from these matched samples are presented in Table 5.

Due to space limitations, the results of the balance tests are reported in the Appendix Table A1. After matching, no statistically significant differences are observed in the covariates between the treatment and control groups, indicating that the balance condition is largely satisfied. This suggests that propensity score matching effectively improves the comparability of the original sample by approximating a randomized assignment, thereby enhancing the credibility and reliability of the estimated results.

The PSM-DID regression results consistently show a significant positive impact of the policy on patenting across all three matching approaches. This confirms that, even after correcting for sample selection bias, the policy’s effect on firm innovation remains robust. These findings reinforce the credibility of the study’s core conclusions.

To ensure the robustness of the results, this study incorporates province-by-year and city-by-year fixed effects, thereby accounting for region-specific policy environments and time-varying local shocks that may affect innovation among heavily polluting firms across the broad sample of A-share listed companies. In addition, to further address concerns regarding industry-level heterogeneous trends and other contemporaneous policy shocks, we include industry-specific linear time trends in the robustness analysis, which helps absorb unobserved industry-wide dynamics such as cyclical fluctuations, structural adjustments, or sector-specific reforms around the policy implementation period.

Furthermore, recognizing the significant disruption caused by the COVID-19 pandemic—particularly widespread production shutdowns in 2020—the analysis excludes observations from that year to avoid confounding effects. Additionally, patents are categorized into invention patents, utility model patents, and design patents. Compared to the latter two, invention patents involve higher R&D costs and greater technical complexity, with stricter application procedures and protection regulations, thereby better reflecting substantive innovation capabilities. A higher proportion of utility model and design patents may indicate the presence of a “patent bubble,” and treating these patent types as homogeneous could introduce measurement bias. To address this issue, we construct a quality-weighted alternative innovation output indicator. Specifically, drawing on scholarly consensus regarding the relative technical content and innovation value of the three patent types, we assign differentiated weights of 3:2:1 to invention patents, utility model patents, and design patents, respectively. We then re-estimate the baseline models using the weighted patent index as the dependent variable. The results from these robustness checks are reported in Table 6.

Table 6 presents robustness tests that incorporate province-by-year and city-by-year fixed effects, excluding data from 2020, replacing core explanatory variables, and including industry-specific trends. The results show that the DID variable maintains a significant positive impact on patent activity, further supporting the robustness and validity of the study’s main findings.

To examine heterogeneity in firms’ responses to the environmental tax reform, we conduct subgroup regressions along three dimensions: ownership structure, firm size, and technological intensity. The results are reported in Table 7. First, firms are classified into state-owned enterprises (SOEs) and non–state-owned enterprises (non-SOEs). As shown in Columns (1) and (2) of Table 7, the estimated DID coefficient is 0.311 for SOEs and 0.166 for non-SOEs, both statistically significant at the 1% level, indicating that the innovation-enhancing effect of the environmental tax reform is substantially stronger for SOEs. This difference may reflect the institutional advantages enjoyed by SOEs in China, including preferential access to fiscal support, policy information, and key production factors. Such advantages allow SOEs to better absorb compliance costs and transform regulatory pressure into incentives for technological upgrading. Moreover, relatively stable resource allocation and organizational structures facilitate long-term investment in research and development.

Second, firms are divided into large and small- and medium-sized enterprises based on the median firm size. Columns (3) and (4) of Table 7 show that the estimated coefficient is 0.303 for large firms and 0.155 for small and medium-sized firms, with both estimates significant at the 1% level. Compared with smaller firms, large firms typically possess stronger financing capacity, more developed R&D infrastructure, and greater risk tolerance. These characteristics enable them to respond to environmental tax constraints by pursuing sustained innovation to achieve both regulatory compliance and competitive advantage. In contrast, smaller firms face tighter financial and market constraints and are therefore more likely to adopt short-term, cost-minimizing strategies.

Finally, firms are grouped by technological intensity into high-tech and non–high-tech enterprises based on industry classification. As reported in Columns (5) and (6) of Table 7, the DID coefficient is 0.361 for high-tech firms and 0.183 for non–high-tech firms, both significant at the 1% level. This disparity reflects the stronger innovation orientation and absorptive capacity of technology-intensive firms, which are more inclined to increase R&D investment in response to environmental tax pressure. By contrast, non–high-tech firms are more constrained by technological path dependence and limited innovation capabilities, resulting in a weaker policy response. Overall, the heterogeneity analysis suggests that the innovation-stimulating effect of the environmental tax reform is more pronounced among firms with stronger resource endowments and institutional advantages, underscoring the critical role of firm characteristics in shaping the transmission of environmental policy.

Using the mediation model developed in this study, we conducted a regression analysis, with the results summarized in Table 8. As shown in column (1), the estimated coefficient of the difference-in-differences (DID) variable on R&D investment intensity is 0.00125 and statistically significant at the 10% level. This suggests that the implementation of the environmental tax reform policy has encouraged heavier polluting firms to increase their R&D spending. Column (2) further shows that both the policy variable (DID) and R&D intensity are positively associated with patent output. According to the standard three-step mediation test (Baron and Kenny, 1986), these findings confirm a mediating effect: the policy boosts innovation partly by encouraging greater investment in R&D.

This relationship supports established views in the innovation literature, which emphasize the central role of R&D in driving technological progress. By investing in research and development, firms are better positioned to introduce new products, improve production processes, and maintain competitiveness in evolving markets. These improvements not only enhance firm-level performance but also contribute to broader industrial transformation and long-term economic growth.

Beyond firm performance, R&D investment also plays a critical role in sustainable development. From this perspective, innovation is not only a path to higher productivity but also a means to address environmental and social challenges. Through targeted R&D, companies can develop cleaner technologies, reduce resource consumption, and minimize environmental impact. In doing so, they align economic objectives with broader societal goals. This approach reflects a growing recognition that business innovation should contribute to sustainability and that long-term value creation increasingly depends on a company’s ability to address both economic and environmental imperatives.

The preceding analysis shows that the environmental protection tax promotes firm innovation by increasing R&D investment. This section further examines the quality and efficiency of innovation. Following Bradley et al. (2016), innovation quality is measured using the number of forward citations received by firms’ patents (Bradley et al., 2017). Specifically, we construct two measures of innovation quality. Overall innovation quality (LnCit) is defined as the natural logarithm of one plus the total number of forward citations received by patents applied for in the following year. Average innovation quality (LnCit2) is measured as the natural logarithm of one plus the average number of forward citations per patent applied for in the following year. Innovation efficiency (IE) is proxied by the number of patent applications per unit of R&D expenditure, calculated as ln (1 + Patent) divided by ln (1 + RD).

The regression results are reported in Table 9. Column (1) presents the results for overall innovation quality (LnCit), Column (2) for average innovation quality (LnCit2), and Column (3) for innovation efficiency (IE). In Column (1), the coefficient on the core explanatory variable is 0.09 and statistically significant at the 1% level. In Column (2), the corresponding coefficient is 0.035, also significant at the 1% level. These findings indicate that the environmental tax reform significantly improves the quality of firm innovation. In Column (3), the estimated coefficient is 0.01 and significant at the 1% level, suggesting that the reform also enhances innovation efficiency.

Compared with other types of innovation, green innovation exhibits dual externalities related to both knowledge creation and environmental benefits. As a result, firms’ green innovation activities are jointly influenced by resource endowments and environmental regulation Song et al. (2020). To further distinguish the effects of the environmental protection tax across different types of innovation, we classify firms’ patents into green and non-green patents.

Green patents are identified based on the “Green Inventory” introduced by the World Intellectual Property Organization (WIPO) in 2010, which is constructed in accordance with the classification standards of the United Nations Framework Convention on Climate Change and mapped to International Patent Classification (IPC) codes. The Green Inventory covers seven broad categories: transportation, waste management, energy conservation, alternative energy production, administrative regulatory or design aspects, agriculture or forestry, and nuclear power generation. We aggregate firms’ patent applications in these categories, add one, and take the natural logarithm to construct the measure of green innovation (LnGreen), with larger values indicating higher levels of green innovative activity. The corresponding regression results are reported in Columns (4) and (5) of Table 9, where Column (4) presents the results for green innovation and Column (5) for non-green innovation. The estimated coefficients on the core explanatory variable are positive and statistically significant at the 1% level in both regressions, indicating that the environmental protection tax reform not only stimulates firms’ green innovation but also promotes non-green innovation.