Not all innovations are created equal: environmental taxes, strategic R&D accounting, and green patent quality in China

1 Introduction

Environmental regulation is widely expected to encourage firms to develop cleaner technologies and increase green innovation (Jaffe et al., 2002; Popp, 2002; Yan et al., 2024). A large body of research, often related to the Porter Hypothesis, argues that well-designed environmental policies can offset compliance costs by stimulating innovation and improving productivity (Jaffe and Palmer, 1997; Lei et al., 2022; Porter and Vanderlinde, 1995). However, measured green innovation does not always represent substantive technological progress (Hall et al., 2005; Wang et al., 2024). In many institutional settings, especially in emerging markets, regulatory evaluation, public legitimacy, and external stakeholder assessment often rely on observable indicators such as R&D expenditure and patent counts. When these indicators become targets, firms may have incentives to improve measurable signals rather than underlying innovation quality (Goodhart, 1984; Peng and Kong, 2024). Thus, an increase in green patenting after environmental regulation may reflect either genuine technological upgrading or strategic compliance.

This article examines whether environmental taxation induces substantive green innovation or strategic compliance. Specifically, it asks three related questions. First, does environmental taxation affect firms' R&D accounting recognition, such as the choice between expensing and capitalizing R&D? Second, does environmental taxation increase green patenting? Third, if green patents increase, is the response concentrated in higher-quality invention patents or in lower-threshold utility model patents? These questions are important because treating all green patents as equivalent may overstate the innovation effect of environmental regulation and obscure firms' strategic responses to policy pressure.

To answer these questions, I exploit China's 2018 Environmental Fee-to-Tax Reform as a quasi-natural experiment (Niu et al., 2025; Wang and Wheeler, 2005). The reform replaced the previous pollution discharge fee system with a formal environmental tax. Compared with the earlier fee-based system, the tax reform strengthened the legal basis of pollution charges, improved enforcement rigidity, and increased the compliance salience of environmental regulation for heavily polluting firms. I compare heavily polluting listed firms with other A-share listed firms before and after the reform using an entropy-balanced difference-in-differences design (Hainmueller, 2012). The entropy-balancing weights are constructed from firm-level pre-treatment covariates during 2013–2017, ensuring that the weighted control group is comparable to the treated group before the policy shock.

The empirical analysis uses Chinese A-share listed firms from 2013 to 2024. I combine financial and governance data from the China Stock Market & Accounting Research Database (CSMAR) with green patent data from the Chinese Research Data Services Platform (CNRDS). The main outcomes capture both input-side accounting choices and output-side innovation responses. On the accounting side, I examine R&D expensing, capitalized R&D, the R&D capitalization ratio, and the ratio of expensed to capitalized R&D (Cho and Kim, 2024; Oswald, 2008). On the innovation side, I distinguish total green patents, green utility model patents, and green invention patents (Dang and Motohashi, 2015). This distinction is central to the article because utility model patents are generally easier, faster, and less substantively demanding than invention patents, making them more likely to capture low-threshold patenting responses.

I find four main results. First, after the Environmental Fee-to-Tax Reform, heavily polluting firms significantly increase R&D expensing while reducing capitalized R&D and the R&D capitalization ratio. This pattern suggests a shift in R&D recognition toward more immediate expense reporting rather than long-term capitalization. I interpret this as evidence consistent with strategic R&D accounting adjustment, rather than direct proof of fraudulent manipulation. Second, the reform increases green patenting, but the increase is concentrated in green utility model patents rather than green invention patents. The coefficient on utility model patents is significantly larger than that on invention patents, and the utility model share rises after the reform. Third, the patent results are robust to alternative estimators for zero-inflated outcomes, including Poisson pseudo-maximum likelihood, inverse hyperbolic sine transformation, and extensive-margin models. These tests address concerns about using the log(1+x) transformation when patent applications contain many zeros (Chen and Roth, 2024). Fourth, ownership heterogeneity reveals different strategic responses: state-owned enterprises exhibit stronger increases in R&D expensing and utility model patenting, consistent with symbolic compliance, whereas privately owned enterprises mainly reduce capitalized R&D, consistent with defensive contraction.

The findings are robust to disclosure-selection tests, same-sample estimates, event-study diagnostics, alternative estimators for zero-inflated patent outcomes, placebo tests, and several alternative specifications. These tests support the interpretation that the observed responses are not driven by MD&A disclosure selection, sample-composition differences, or the log(1+x) transformation of patent outcomes.

This article contributes to three strands of literature. First, it contributes to the literature on patent greenwashing. Recent studies show that green finance policies may motivate firms to engage in greenwashing through strategic patenting behavior (Huang et al., 2025). This article differs by examining a tax-based environmental regulation that directly increases compliance costs for heavily polluting firms. The setting allows me to show that environmental taxation, not only green finance incentives, can induce strategic increases in low-threshold green patents. Second, it contributes to research on environmental regulation and green innovation by distinguishing patent quantity from patent quality. The results qualify the Porter Hypothesis by showing that environmental taxation increases measured green patenting, but the increase is concentrated in utility model patents rather than invention patents. Third, the article links firms' innovation-output responses with accounting-input responses. The evidence suggests that firms may adjust both R&D recognition and green patenting behavior as part of a broader strategic compliance portfolio.

The remainder of the article proceeds as follows. Section 2 presents the institutional background and develops the hypotheses. Section 3 describes the data, variables, and empirical design. Section 4 reports the main results. Section 5 examines ownership heterogeneity and strategic portfolio evidence. Section 6 presents robustness checks and discusses the implications. Section 7 concludes.

2 Institutional background and hypothesis development

2.1 China's environmental fee-to-tax reform

China's Environmental Protection Tax Law came into effect on January 1, 2018 (Niu et al., 2025). The reform replaced the previous pollution discharge fee system with a statutory environmental tax. Under the earlier fee-based regime, local environmental protection authorities collected pollution discharge fees from firms that emitted regulated pollutants. Although the fee system had existed for many years, it was often criticized for weak enforcement, discretionary implementation, local protectionism, and insufficient deterrence (van Rooij et al., 2017). In practice, firms could sometimes bargain over fee payments, delay payments, or face uneven enforcement across regions.

The 2018 reform converted pollution discharge fees into a formal tax instrument. This change was institutionally important for three reasons. First, the reform strengthened the legal status of pollution charges. Unlike administrative fees, taxes are backed by statutory authority and are subject to more standardized collection procedures. Second, the reform increased enforcement rigidity. The involvement of tax authorities made collection more rule-based and reduced the discretion previously associated with local fee collection. Third, the reform improved transparency and compliance salience. Firms became more clearly subject to legally specified tax liabilities based on pollutant emissions, making environmental compliance more visible to regulators, investors, and other stakeholders.

The Environmental Protection Tax applies to taxable pollutants, including air pollutants, water pollutants, solid waste, and noise. The tax burden depends on pollutant emissions and local tax rates within the statutory range. Because emission intensity differs substantially across industries, the reform did not affect all firms equally. Firms in heavily polluting industries were more directly exposed to the reform because they were more likely to emit taxable pollutants and to face higher compliance costs. Therefore, the reform provides a useful quasi-natural experiment for examining how firms respond to a stronger and more formalized environmental tax regime.

Compared with command-and-control regulation, the fee-to-tax reform changed the institutional environment in which pollution-related costs were assessed, collected, and monitored (Chu and Wang, 2025). Once pollution charges became formal tax liabilities, heavily polluting firms faced stronger incentives to manage both real compliance costs and observable environmental responsiveness.

2.2 Treatment group: heavily polluting firms

The treatment group consists of listed firms operating in heavily polluting industries (Cai et al., 2020). These firms are more directly exposed to the Environmental Fee-to-Tax Reform because their production activities are more likely to generate taxable pollutants and attract environmental tax scrutiny. In the empirical analysis, Treat equals one for firms in heavily polluting industries and zero otherwise. Post equals one for 2018 and subsequent years. The DID variable is the interaction between Treat and Post.

The treatment definition is industry-based rather than firm-specific emission-based. This approach is appropriate because industry classification is predetermined relative to the reform and less likely to be affected by firms' post-reform strategic behavior. The estimated coefficient should therefore be interpreted as the average differential response of heavily polluting firms relative to other listed firms after the reform.

2.3 Strategic compliance under environmental taxation

The Environmental Fee-to-Tax Reform may affect firms through two related channels. The first is a real cost-pressure channel. By converting pollution discharge fees into a statutory tax and strengthening enforcement, the reform increased expected compliance costs for heavily polluting firms (Lei and Kocoglu, 2025). Higher environmental tax pressure may affect firms' resource allocation, investment horizon, and willingness to commit to uncertain long-term R&D projects.

The second is an observable compliance-pressure channel. Environmental regulation also creates incentives for firms to demonstrate environmental responsiveness to regulators, investors, creditors, and local governments. These stakeholders often evaluate firms using measurable indicators, such as R&D expenditure and green patent counts. When such indicators become targets, firms may improve observable signals without generating proportional improvements in underlying technological quality (Marquis et al., 2016). This logic is consistent with Goodhart's Law: when a measure becomes a target, it may cease to be a reliable measure of the underlying construct (Goodhart, 1984).

In this setting, firms may respond on both the input and output sides. On the input side, firms may adjust R&D accounting recognition by increasing expensed R&D and reducing capitalized R&D. On the output side, firms may increase green patent applications, especially utility model patents, which are easier, faster, and less costly to obtain than invention patents. These two responses are not mutually exclusive. Firms may simultaneously adjust R&D recognition and green patenting as part of a broader strategic compliance portfolio.

Figure 1 summarizes the conceptual framework.

2.4 Environmental taxation and R&D accounting recognition

The first implication of this framework concerns R&D accounting recognition. Environmental taxation may affect how heavily polluting firms report and recognize R&D expenditures. After the Environmental Fee-to-Tax Reform, these firms face higher expected compliance costs and greater uncertainty about future cash flows. Such pressure may reduce their willingness to commit to long-term development projects whose benefits are uncertain and realized over a longer horizon.

R&D capitalization generally reflects development expenditures that are expected to generate future economic benefits and satisfy recognition criteria. By contrast, R&D expensing recognizes expenditures in the current period (Oswald, 2008). A decline in capitalized R&D may indicate that firms reduce long-term development commitments or shift away from projects that satisfy capitalization criteria. An increase in R&D expensing may indicate a stronger emphasis on current-period innovation expenditure and more immediately observable R&D effort.

From a strategic compliance perspective, R&D expensing can also serve as a visible signal of innovation input. Because R&D expenditure is commonly used by external stakeholders as an indicator of innovation effort, increasing expensed R&D may help firms display responsiveness after an environmental tax shock. This does not imply fraudulent accounting manipulation. Rather, it suggests that firms may strategically adjust R&D recognition under heightened compliance pressure.

Therefore, the Environmental Fee-to-Tax Reform is expected to increase R&D expensing while reducing capitalized R&D and the R&D capitalization ratio among heavily polluting firms.

H1a. The Environmental Fee-to-Tax Reform increases R&D expensing among heavily polluting firms.

H1b. The Environmental Fee-to-Tax Reform decreases R&D capitalization and the R&D capitalization ratio among heavily polluting firms.

2.5 Environmental taxation and green patent quality

The second implication concerns green patenting behavior. Environmental taxation increases firms' incentives to demonstrate green innovation. Green patents are observable, quantifiable, and widely used as indicators of environmental innovation (Johnstone et al., 2010). As a result, firms facing stronger environmental tax pressure may increase green patent applications to signal compliance and technological responsiveness.

However, green patents differ in quality and technological substance. In China, invention patents generally require higher novelty and inventiveness, involve substantive examination, and take longer to obtain. Utility model patents are typically easier, faster, and less costly to apply for and obtain (Dang and Motohashi, 2015; Zhang and Jiang, 2025). They may therefore be more attractive when firms seek to generate visible green innovation outputs quickly.

If environmental taxation induces substantive green innovation, the reform should generate a robust increase in green invention patents. If firms instead respond strategically to observable compliance pressure, the increase in green patenting may be concentrated in green utility model patents. In that case, total green patent counts may rise even though higher-quality green invention patents do not increase significantly.

This distinction is important because substantive innovation and strategic patenting are not necessarily mutually exclusive. Firms may engage in both. The empirical question is whether the marginal response to environmental taxation is disproportionately concentrated in lower-threshold patent types. A disproportionate increase in green utility model patents without a corresponding robust increase in green invention patents would be consistent with patent greenwashing or strategic low-quality green patenting (Huang et al., 2025).

Accordingly, I propose the following hypotheses:

H2a. The Environmental Fee-to-Tax Reform increases green utility model patents among heavily polluting firms.

H2b. The Environmental Fee-to-Tax Reform does not generate a robust increase in green invention patents among heavily polluting firms.

2.6 Ownership heterogeneity and strategic portfolio

The previous hypotheses examine R&D accounting recognition and green patenting separately. Yet firms may adjust these indicators jointly. Under observable compliance pressure, managers may seek to improve multiple measurable signals at the same time. R&D expensing signals innovation input, while green utility model patents signal green innovation output. Together, these indicators can create a broader appearance of environmental responsiveness.

This joint response is consistent with a strategic compliance portfolio. The idea is not that R&D expensing mechanically causes utility model patenting. Rather, both may arise from the same underlying incentive to demonstrate compliance under environmental tax pressure. If firms optimize observable indicators, then accounting-side R&D recognition and output-side low-threshold patenting should co-move within firms.

This relationship is especially relevant for understanding patent greenwashing. A firm that increases R&D expensing while also increasing green utility model patenting can appear more innovation-oriented and environmentally responsive, even if there is no robust increase in high-quality invention patents. Therefore, the within-firm association between R&D expensing and utility model patenting provides supplementary evidence on whether firms adopt coordinated compliance strategies.

Because this association is not designed to identify a causal effect of R&D expensing on patenting, I treat it as supplementary evidence rather than a primary causal hypothesis. The expected relationship is:

H3. R&D expensing and green utility model patenting are positively associated within firms after absorbing firm and year fixed effects.

Firms' responses to environmental taxation may differ by ownership structure. In China, state-owned enterprises and privately owned enterprises face different objectives, constraints, and stakeholder pressures. These differences can shape whether firms respond through symbolic compliance or defensive contraction.

State-owned enterprises are more closely connected to government objectives and are subject to stronger political and legitimacy pressures (Marquis and Qian, 2014; Zhang et al., 2022). They are often expected to align with public policy priorities and demonstrate responsiveness to environmental regulation. Because of their visibility and political obligations, SOEs may have stronger incentives to produce observable compliance signals. They may therefore respond to the Environmental Fee-to-Tax Reform by increasing R&D expensing and green utility model patenting. Such a pattern would be consistent with symbolic compliance: improving visible indicators of environmental effort without necessarily producing a robust increase in substantive invention patents.

Privately owned enterprises face a different set of constraints. They often have weaker access to policy resources and external financing than SOEs (Huang and Liu, 2025). When environmental tax pressure increases costs and uncertainty, private firms may respond more defensively by reducing long-term and risky development commitments. This response should be reflected in lower capitalized R&D and a lower R&D capitalization ratio. Compared with SOEs, private firms may have less capacity or incentive to produce symbolic compliance signals through patenting.

Therefore, ownership structure is expected to moderate the effect of environmental taxation. SOEs are expected to show stronger symbolic compliance through R&D expensing and utility model patenting, whereas privately owned enterprises are expected to show stronger defensive contraction through reduced R&D capitalization.

H4a. State-owned enterprises exhibit stronger increases in R&D expensing and green utility model patents after the Environmental Fee-to-Tax Reform.

H4b. Privately owned enterprises exhibit stronger reductions in R&D capitalization after the Environmental Fee-to-Tax Reform.

3 Data, variables, and empirical design

3.1 Data sources and sample construction

This study uses a panel of Chinese A-share listed firms from 2013 to 2024. The financial, accounting, governance, and ownership data are obtained from the China Stock Market & Accounting Research Database (CSMAR). Green patent data are obtained from the Chinese Research Data Services Platform (CNRDS). The patent data allow me to distinguish between different types of green patent applications, including green invention patents and green utility model patents.

The sample period begins in 2013 to provide a sufficiently long pre-reform window before the 2018 Environmental Fee-to-Tax Reform. The post-reform period starts in 2018, when the Environmental Protection Tax Law formally came into effect. The full sample period ends in 2024, which allows the analysis to capture both short-run and medium-run firm responses after the reform.

Financial firms are excluded because their accounting structure, leverage, regulation, and R&D reporting differ substantially from those of non-financial firms. I also exclude ST, ^*ST, and PT firms because these firms are subject to abnormal listing conditions and may have unusual financial reporting behavior. All continuous variables are winsorized at the 1st and 99th percentiles to reduce the influence of extreme values.

Table 1 reports the full sample construction process. The initial sample contains 5,627 A-share listed firms and 55,845 firm-year observations from 2013 to 2024. After excluding financial firms, ST/^*ST/PT firms, and observations with missing core financial controls, the cleaned panel contains 4,580 firms and 46,825 firm-year observations. I then merge the green patent data from CNRDS. Firm-years without matched green patent records are coded as zero patent applications rather than being excluded from the sample.

Table 1. Sample construction flow.

The sample spans 2013–2024. Financial firms, ST/^*ST/PT firms, and observations with missing core controls are excluded. Firm-years without matched green patent records are coded as zero patent applications. The cleaned panel contains 46,825 observations. The EB-DID sample contains 31,090 observations because entropy-balancing weights require valid pre-treatment covariates during 2013–2017. This reduction is not outcome-based. The MD&A R&D expensing disclosure subsample contains 9,400 observations.

For the entropy-balanced difference-in-differences estimation, I further require firms to have valid pre-treatment covariates during 2013–2017. This restriction is necessary because the entropy-balancing weights are constructed from firm-level pre-treatment covariate averages. Firms listed after 2017 or firms with incomplete pre-reform financial information cannot be assigned valid entropy-balancing weights. After merging the entropy-balancing weights, the final baseline estimation sample contains 2,866 firms and 31,090 firm-year observations.

It is important to emphasize that the reduction from 46,825 observations in the cleaned panel to 31,090 observations in the EB-DID estimation sample is not an outcome-based restriction. It reflects the requirement that entropy-balancing weights can only be constructed for firms with valid pre-treatment covariates. Therefore, the estimation sample is determined by the availability of pre-reform firm characteristics rather than by post-reform outcomes.

The R&D expensing variable disclosed in the management discussion and analysis section is available for a smaller subsample. The MD&A R&D expensing disclosure subsample contains 9,400 observations, and 9,323 observations remain in the corresponding regressions after singleton deletion. Because this variable is not available for all firm-years, I explicitly test whether the reform affects the probability of R&D expensing disclosure and further conduct same-sample regressions to ensure that the main results are not driven by differences in sample composition.

3.2 Variable definitions

3.2.1 R&D accounting variables

The first group of dependent variables captures firms' R&D accounting recognition. These variables measure whether firms respond to environmental tax pressure by adjusting the way R&D expenditures are recognized in financial reporting (Canace et al., 2022; Dinh et al., 2016).

The first variable is R&D expensing, denoted as lnRD_Exp_Mgmt. It is measured as the logarithm of R&D expenses disclosed in the management discussion and analysis section. This variable captures current-period R&D expenditure recognized as an expense. Because MD&A R&D expensing is not available for all firms, I separately examine whether the Environmental Fee-to-Tax Reform changes the likelihood of disclosure and re-estimate the main results using the same disclosure sample.

The second variable is R&D capitalization, denoted as lnRD_Cap_Acc. It is measured as the logarithm of capitalized R&D reported on the balance sheet. This variable captures R&D expenditures recognized as development assets and is available for a broader sample than MD&A R&D expensing.

The third variable is the R&D capitalization ratio, defined as:

RD CapRatioit=CapRDitCapRDit+ExpRDit(1)

where CapRD denotes capitalized R&D and ExpRD denotes expensed R&D.

This ratio directly measures the relative importance of capitalized R&D in total recognized R&D expenditure. A decline in this ratio indicates a shift away from capitalization toward expensing.

The fourth variable is the logarithm of the ratio of expensed to capitalized R&D, denoted as ln(Exp/Cap). This variable provides an additional measure of the relative shift between expensing and capitalization.

Together, these variables allow the analysis to examine whether the reform affects not only the level of R&D reporting but also the composition of R&D accounting recognition.

3.2.2 Green patent variables

The second group of dependent variables captures firms' green innovation outputs. I construct several patent variables using green patent data from CNRDS.

The first variable is total green patents, measured as the logarithm of total green patent applications. The second variable is green utility model patents, measured as the logarithm of green utility model patent applications. The third variable is green invention patents, measured as the logarithm of green invention patent applications.

The distinction between utility model patents and invention patents is central to the analysis. Invention patents generally require higher novelty and inventiveness, involve substantive examination, and take longer to obtain. Utility model patents are typically easier, faster, and less costly to apply for and obtain. Therefore, an increase in green utility model patents without a corresponding increase in green invention patents may indicate a low-threshold patenting response rather than a robust improvement in substantive green innovation.

I also construct two share-based patent variables. Utility model share is defined as the number of green utility model patents divided by total green patents. Invention share is defined as the number of green invention patents divided by total green patents. These share variables are computed for firm-years with at least one green patent application.

Firm-years without matched green patent records are coded as zero patent applications rather than treated as missing. This coding choice is important because the absence of a matched patent record indicates no observed green patent application in that firm-year, not missing financial information. Since patent outcomes contain many zeros, I later use alternative estimators, including PPML, inverse hyperbolic sine transformation, and extensive-margin models, to address concerns about the log(1+x) transformation (Silva and Tenreyro, 2006).

3.2.3 Treatment variable

The treatment group consists of firms in heavily polluting industries. The treatment indicator, Treat, equals one if a firm belongs to a heavily polluting industry and zero otherwise. The post-reform indicator, Post, equals one for years 2018 and after and zero for years before 2018. The main difference-in-differences variable is:

DIDit=Treati×Postt(2)

The coefficient on this interaction term captures the differential change in outcomes for heavily polluting firms after the Environmental Fee-to-Tax Reform relative to non-heavy-polluting firms.

3.2.4 Control variables

The baseline regressions include a set of firm-level control variables commonly used in studies of corporate innovation, accounting choices, and environmental regulation. These controls account for differences in firm size, financial condition, profitability, growth opportunities, operating performance, firm maturity, asset structure, valuation, and governance.

The control variables include:

Size: the natural logarithm of total assets;

Leverage: total liabilities divided by total assets;

ROA: return on assets;

Growth: revenue growth;

CashFlow: operating cash flow scaled by total assets;

Age: the natural logarithm of firm age;

Tangibility: tangible assets scaled by total assets;

TobinQ: Tobin's Q, capturing market valuation and growth opportunities;

Board: the natural logarithm of board size;

Indep: the proportion of independent directors on the board.

All baseline models include firm fixed effects and year fixed effects. Firm fixed effects control for time-invariant firm characteristics, while year fixed effects absorb common macroeconomic, regulatory, and market-wide shocks. Standard errors are clustered at the firm level.

Descriptive statistics are reported in Appendix Table A1. The EB estimation sample contains 31,090 firm-year observations from 2,866 firms. Patent outcomes and capitalized R&D are highly zero-inflated, which motivates the use of alternative estimators for patent outcomes in Section 4.4.

3.3 Baseline DID model

The baseline empirical strategy uses a difference-in-differences design that compares heavily polluting firms with non-heavy-polluting firms before and after the 2018 Environmental Fee-to-Tax Reform. The estimating equation is:

Yit=α+βTreati×Postt+γXit+μi+λt+ϵit(3)

where Y_it denotes the outcome variable for firm i in year t. Treat_i equals one for heavily polluting firms, and Post_*t* equals one for 2018 and subsequent years. X_it is the vector of firm-level controls. Firm fixed effects, μ_*i*, absorb time-invariant firm characteristics, while year fixed effects, λ_*t*, absorb common shocks. Standard errors are clustered at the firm level. The coefficient of interest is β, which captures the differential post-reform change in heavily polluting firms relative to non-heavy-polluting firms.

3.4 Entropy balancing and parallel trends

Although the difference-in-differences design controls for time-invariant firm characteristics and common year shocks, heavily polluting and non-heavy-polluting firms may differ in observable pre-treatment characteristics. Such differences may affect firms' innovation behavior, accounting choices, and patenting outcomes. To improve comparability between the treatment and control groups, I use entropy balancing (Hainmueller, 2012).

Because heavily polluting and non-heavy-polluting firms differ in observable characteristics, I construct firm-level entropy-balancing weights using pre-treatment covariate averages from 2013 to 2017. The covariates used for balancing include firm size, leverage, return on assets, revenue growth, operating cash flow, firm age, tangibility, Tobin's Q, board size, and the independent director ratio.

The entropy-balancing procedure is implemented at the firm level. Specifically, I first calculate each firm's pre-treatment average for the covariates during 2013–2017. I then construct weights for control firms so that the weighted distribution of pre-treatment covariates in the control group matches the corresponding distribution in the treated group. These firm-level weights are then merged back into the firm-year panel and used in the baseline DID regressions.

Because the weights are constructed only from pre-treatment variables, the procedure avoids conditioning on post-treatment outcomes or post-treatment firm characteristics. The firm-level weights are merged back into the panel and used in the DID regressions, making the weighted control group comparable to the treated group before the reform.

Table 2 reports the covariate balance before and after entropy balancing. Before weighting, treated and control firms differ along several observable dimensions, including cash flow, tangibility, growth, and Tobin's Q. After entropy balancing, all standardized biases are reduced substantially. The post-weighting standardized biases range from 0.001 to 0.003, indicating that the weighted control group closely resembles the treated group in terms of pre-treatment observable characteristics.

Table 2. Covariate balance: entropy balancing based on pre-treatment characteristics (2013–2017).

Entropy balancing weights are constructed from firm-level pre-treatment (2013–2017) averages of all listed covariates. This design avoids conditioning on post-treatment variables and ensures the weighted control group resembles the treatment group before the 2018 reform. Post-EB standardized biases range from 0.001 to 0.003, with bias reductions of 95.8%−99.5%. Firms without valid pre-treatment data are excluded (see Table 1, Step 9).

The validity of the difference-in-differences design relies on the parallel-trends assumption. That is, absent the Environmental Fee-to-Tax Reform, heavily polluting and non-heavy-polluting firms should have followed similar trends in the outcome variables. To assess this assumption, I estimate an event-study specification that traces the dynamic effects of the reform before and after 2018 (Roth et al., 2023).

The event-study model is specified as follows:

Yit=α+∑k≠-1βkTreati×1[EventTimet=k] +γXit+μi+λt+ϵit(4)

where EventTime_*t* denotes the year relative to the reform. The year 2017 is used as the reference year, corresponding to k = −1, because it is the last year before the Environmental Protection Tax Law came into effect. The coefficients β_*k* capture the difference between treated and control firms in each event year relative to the omitted reference year.

The pre-treatment coefficients are used to assess the parallel-trends assumption. If the treated and control groups followed similar trends before the reform, the coefficients for the pre-reform years should be statistically indistinguishable from zero. I report the joint pre-trend p-values in the main regression tables and plot the dynamic coefficients in Figure 2.

4 Main results

This section reports the main empirical results. I first examine whether the Environmental Fee-to-Tax Reform affects firms' R&D accounting choices. I then address the potential disclosure-selection concern related to MD&A R&D expensing and provide same-sample estimates. Next, I examine whether the reform affects green patent quantity and quality. Finally, I use alternative estimators to address concerns about zero-inflated patent outcomes and the use of the log(1+x) transformation.

4.1 R&D accounting choices

Table 3 reports the baseline EB-DID estimates for R&D accounting choices. The dependent variables include R&D expensing disclosed in MD&A, capitalized R&D reported on the balance sheet, the R&D capitalization ratio, and the ratio of expensed to capitalized R&D. All specifications include firm fixed effects, year fixed effects, entropy-balancing weights, and the full set of firm-level controls. Standard errors are clustered at the firm level.

Table 3. Environmental tax reform and R&D accounting choices.

All models include firm and year fixed effects with entropy-balancing weights (pre-treatment 2013–2017 covariates). Standard errors clustered at the firm level in parentheses. Column (1): MD&A R&D expensing disclosure sample. Column (2): full EB estimation sample (balance-sheet capitalization data, mandatory disclosure). Column (3): RD Cap Ratio = Capitalized R&D / (Capitalized R&D + Expensed R&D), requiring both components non-missing and non-zero (N = 4,715). Column (4): ln(Expensed/Capitalized), same sample as Column (1). Joint pre-trend p-values are reported for all columns. Event-study plots for the ratio outcomes (Cols 3–4) are provided in the Appendix;pre-treatment coefficients are jointly insignificant (p > 0.10), providing support for the parallel-trends assumption. ^**p < 0.05, ^***p < 0.01.

Table 3 shows that the reform significantly changes R&D accounting recognition among heavily polluting firms. The DID coefficient for R&D expensing is 0.251 and significant at the 1% level, indicating an increase in expensed R&D after the reform. In contrast, the coefficient for capitalized R&D is −0.622 and significant at the 1% level. The R&D capitalization ratio also declines significantly, while the ratio of expensed to capitalized R&D increases.

These results support H1a and H1b. The reform increases R&D expensing and reduces capitalized R&D, suggesting a shift away from capitalization toward expensing. This pattern is consistent with strategic R&D recognition under heightened environmental compliance pressure. It should not be interpreted as direct evidence of fraudulent accounting manipulation.

4.2 R&D expensing disclosure selection and same-sample estimates

A potential concern is that R&D expensing disclosed in MD&A is not observed for all firm-years. If the Environmental Fee-to-Tax Reform changes which firms disclose R&D expensing, the baseline expensing result could be affected by disclosure selection. To address this concern, Table 4 provides two additional tests.

Table 4. R&D expensing disclosure selection and same-sample estimates.

^aDependent variable = 1 if firm discloses R&D expensing in MD&A. Firm FE, year FE, EB weights, and full controls included. Standard errors clustered at firm level. DID coefficient is small and statistically insignificant (p = 0.503), indicating that the environmental tax reform did not systematically alter which firms disclose R&D expenses in MD&A. This mitigates concerns that the expensing results are driven by sample selection. ^bAll estimates use the same 9,323 observations with non-missing MD&A R&D expensing. Firm FE, year FE, EB weights, full controls. Standard errors at firm level. Results are consistent with full-sample estimates: expensing increases (Col 1), capitalization declines (Col 2), green utility model patents increase (Col 5), while green invention patents remain insignificant (Col 6, p = 0.535). ^**p < 0.05, ^***p < 0.01.

Table 4 addresses the disclosure-selection concern. Panel A shows that the DID coefficient for MD&A R&D expensing disclosure is small and insignificant, indicating that the reform does not systematically affect whether firms disclose R&D expensing. Panel B re-estimates the key outcomes on the same 9,323-observation disclosure sample. The results remain consistent: R&D expensing increases, capitalized R&D and the capitalization ratio decline, green utility model patents increase, and green invention patents remain insignificant. These tests suggest that the main findings are not driven by MD&A disclosure selection or by differences in estimation samples.

4.3 Green patent quantity and quality

Table 5 examines whether the Environmental Fee-to-Tax Reform affects green patenting behavior. The dependent variables include total green patents, green utility model patents, green invention patents, utility model share, and invention share. The first three outcomes are measured using log-transformed patent applications, while the share variables are computed for firm-years with at least one green patent application.

Table 5. Environmental tax reform and green patent quantity and quality.

All models include firm and year FEs with EB weights. Standard errors clustered at firm level. Columns (4)–(5): estimated on the subsample with GreenTotal > 0 (N = 17,250; 55.5% of the full sample has at least one green patent). Column (3): green invention patents are not statistically significant (p = 0.352). Column (2): green utility model patents are strongly significant (p < 0.001). The Wald test rejects the null that the two coefficients are equal (p < 0.001). The utility model share rises significantly (Col 4), while the invention share does not increase and shows a negative but statistically insignificant coefficient (Col 5, −0.022, p = 0.116). This asymmetric pattern is consistent with strategic patenting or patent greenwashing. ^**p < 0.05, ^***p < 0.01.

Table 5 shows that the reform increases total green patenting, but the increase is concentrated in utility model patents. The DID coefficient is 0.114 for total green patents and 0.185 for green utility model patents, both significant at the 1% level. In contrast, the coefficient for green invention patents is 0.028 and statistically insignificant. The Wald test rejects equality between the utility model and invention coefficients.

The composition results reinforce this pattern. The utility model share increases significantly, while the invention share does not increase. These findings support H2a and H2b. They suggest that the green innovation response is concentrated in low-threshold utility model patents rather than higher-quality invention patents.

4.4 Alternative estimators for patent outcomes

Patent outcomes contain many zeros, and the use of log(1+x) transformations can be problematic in zero-inflated data because coefficients may be difficult to interpret and estimates may be sensitive to the choice of transformation (Chen and Roth, 2024). To address this concern, I estimate the patent results using three alternative approaches: Poisson pseudo-maximum likelihood (PPML), inverse hyperbolic sine (IHS) transformation, and extensive-margin linear probability models (Bellemare and Wichman, 2020). Table 6 reports the results.

Table 6. Alternative estimators for zero-inflated patent outcomes.

All panels include firm FE, year FE, EB weights, and full controls (N = 31,090). Standard errors clustered at firm level. Panel A: PPML with raw count outcomes, estimated via ppmlhdfe absorbing firm and year FEs (Chen and Roth, 2024); all models converged. Panel B: inverse hyperbolic sine transformation. Panel C: linear probability model. Across all three estimators, green utility model patents are consistently significant while green invention patents are consistently not (PPML p = 0.332, IHS p = 0.289, LPM p = 0.161). This suggests that the asymmetric quality pattern is not driven by the log(1+x) transformation. ^**p < 0.05, ^***p < 0.01.

Table 6 addresses concerns about zero-inflated patent outcomes and the log(1+x) transformation. Across PPML, IHS, and extensive-margin specifications, the reform significantly increases total green patents and green utility model patents, while green invention patents remain statistically insignificant. The asymmetric pattern is therefore not driven by the functional form used in the baseline patent regressions.

5 Heterogeneity and strategic portfolio evidence

This section examines heterogeneity by ownership structure and whether accounting-side and patent-side responses co-move within firms. The ownership analysis distinguishes between symbolic compliance and defensive contraction, while the strategic portfolio analysis provides supplementary evidence on the association between R&D expensing and green utility model patenting.

5.1 Ownership heterogeneity

Table 7 reports heterogeneous effects by ownership type. The analysis separates privately owned enterprises from state-owned enterprises because these firms face different regulatory pressures, financing constraints, and political incentives. Private firms may be more sensitive to the cost and uncertainty created by environmental taxation, whereas SOEs may face stronger political and legitimacy pressure to demonstrate environmental responsiveness.

Table 7. Ownership heterogeneity: state-owned enterprises vs. privately owned enterprises.

Notes: All models include firm FE, year FE, EB weights, and full controls. Standard errors clustered at firm level. “POE effect” = DID coefficient when SOE = 0. “SOE difference” = incremental effect for SOEs (DID × SOE). “SOE Total Effect” = linear combination DID + DID × SOE (p-value from delta-method test). POEs exhibit defensive contraction: significant capitalization reduction (Col 2, −0.915^***) with no significant response in any patent outcome (Cols 4–6, all ns for POE). SOEs exhibit symbolic compliance: strong expensing surge (Col 1, total = 0.396^***) and utility model expansion (Col 4, total = 0.325^***). The SOE differential for invention patents is weakly positive (Col 5, DID × SOE = 0.098^*), but the SOE total effect on invention patents is statistically insignificant (0.046, p = 0.294), indicating no robust substantive innovation response even among SOEs. ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

Table 7 shows clear ownership heterogeneity. Private firms mainly respond by reducing capitalized R&D. The DID coefficient for private firms is −0.915 and significant at the 1% level, while their patent responses are statistically insignificant. This pattern is consistent with defensive contraction under higher compliance-cost pressure.

SOEs display a different response. The SOE total effect is positive and significant for R&D expensing and green utility model patents, indicating stronger increases in visible compliance indicators. However, the SOE total effect on green invention patents is statistically insignificant. Thus, even among SOEs, the patent response is concentrated in utility model patents rather than invention patents. These results are consistent with symbolic compliance among SOEs and defensive contraction among private firms.

5.2 Strategic portfolio evidence

I next examine whether financial-side R&D expensing and output-side utility model patenting co-move within firms. Both variables are residualized by firm and year fixed effects. Appendix Table A2 shows that residualized R&D expensing is positively associated with residualized green utility model patenting. The association remains positive after controlling for the DID indicator and is stronger among SOEs than among private firms. Appendix Figure A2 provides a residualized binscatter visualization.

This evidence is correlational and should not be interpreted as causal. It does not show that R&D expensing causes utility model patenting, or vice versa. Rather, it provides supplementary evidence that accounting-side and patent-side responses may co-move as part of a broader strategic compliance portfolio.

6 Robustness checks and discussion

This section reports robustness checks and discusses the theoretical interpretation of the findings. The robustness tests focus on the four core outcomes: R&D expensing, R&D capitalization, green utility model patents, and green invention patents. The central pattern remains stable: R&D expensing increases, R&D capitalization decreases, green utility model patenting increases, and green invention patenting is not robustly significant.

6.1 Alternative specifications and placebo tests

Table 8 reports the DID coefficients under several alternative specifications. The first row reproduces the baseline EB-DID estimates. The following rows examine whether the results are robust to unweighted DID estimation, lagged controls, excluding the transition year 2018, using a shorter sample window, excluding real estate firms, and adding industry-specific linear trends.

Table 8. Robustness checks: DID coefficients across alternative specifications.

This table reports the DID coefficient (Treat × Post) across alternative specifications. All models include firm and year fixed effects and full control variables unless otherwise noted. Standard errors clustered at the firm level are reported in parentheses. The specifications include unweighted DID, lagged controls, excluding 2018, a shorter 2014–2022 window, excluding real estate firms, and industry-specific linear trends. The core pattern—positive expensing response, negative capitalization response, positive utility model response, and largely insignificant invention patent response—is robust. The lagged-controls specification yields a marginally significant Green Inv coefficient (0.037^*), but the effect is small in magnitude and not robust across alternative estimators (see Table 6). ^*p < 0.10, ^***p < 0.01.

Table 8 shows that the main results are robust to alternative weighting choices, lagged controls, excluding the transition year 2018, using a shorter 2014–2022 sample window, excluding real estate firms, and adding industry-specific linear trends. Across these specifications, the core pattern remains: R&D expensing increases, capitalized R&D decreases, and green utility model patents increase.

The evidence for green invention patents is weaker. One lagged-control specification yields a marginally significant invention coefficient at the 10% level. However, the magnitude is small, and the result is not robust across other specifications or alternative estimators. Therefore, I interpret the invention-patent response as not robustly significant.

To further assess whether the main findings could arise from random assignment of treatment status, I conduct placebo tests based on 500 random treatment assignments. In each placebo iteration, treatment status is randomly assigned to firms, and the DID model is re-estimated. The distribution of placebo estimates is then compared with the actual estimated DID coefficient.

The placebo tests reinforce the main findings (Figure 3). The actual estimates for R&D expensing, R&D capitalization, and green utility model patents are unlikely to arise from random treatment assignment. In contrast, the green invention estimate does not stand out from the placebo distribution, consistent with the conclusion that invention patents do not respond robustly to the reform.

6.2 Discussion

6.2.1 Environmental regulation and the porter hypothesis

The findings do not imply that environmental regulation never induces innovation. Recent evidence shows that environmental and energy-saving policies can stimulate green innovation under certain institutional and policy conditions (Jaffe and Palmer, 1997; Porter and Vanderlinde, 1995; Su et al., 2025). Instead, the present findings qualify the Porter Hypothesis by showing that the marginal innovation response depends on innovation quality. Environmental taxation increases measured green patenting, but the increase is concentrated in utility model patents rather than invention patents. Thus, aggregate patent counts may overstate the substantive innovation effect of environmental regulation when increases are driven by lower-threshold patent types.

6.2.2 Substantive innovation and strategic patenting

I do not assume that substantive innovation and strategic patenting are mutually exclusive. Firms may engage in both (Marquis and Qian, 2014). The empirical question is whether the marginal response to the reform is disproportionately concentrated in low-threshold patent types. The evidence shows that utility model patents increase significantly, while invention patents do not increase robustly. This pattern suggests that the marginal patenting response contains a stronger strategic or symbolic component, although it does not imply that all utility model patents are meaningless or that all firms engage in greenwashing.

6.2.3 Interpretation of R&D accounting responses

The accounting results should be interpreted as evidence of strategic R&D recognition rather than direct evidence of fraudulent manipulation. R&D capitalization and expensing involve managerial judgment and depend on project uncertainty, expected future benefits, and recognition criteria. A decline in capitalized R&D may reflect reduced long-term development commitments or more conservative recognition under higher compliance-cost pressure. Therefore, the evidence shows that firms adjust R&D recognition in response to environmental taxation, but it does not prove accounting fraud.

7 Conclusion

This article examines how China's 2018 Environmental Fee-to-Tax Reform affects firms' R&D accounting choices and green patenting behavior. Using Chinese A-share listed firms from 2013 to 2024 and an entropy-balanced DID design, I find that the reform changes both innovation-related inputs and outputs among heavily polluting firms.

The reform increases R&D expensing and reduces capitalized R&D, leading to a lower R&D capitalization ratio. On the innovation-output side, it increases total green patenting, but the increase is concentrated in green utility model patents. Green invention patents do not show a robust increase across the main specifications and alternative estimators. Ownership heterogeneity further shows that SOEs exhibit stronger increases in R&D expensing and utility model patenting, consistent with symbolic compliance, whereas private firms mainly reduce capitalized R&D, consistent with defensive contraction.

These findings suggest that environmental taxation can increase visible innovation indicators without generating a proportional increase in higher-quality green innovation. The results qualify the Porter Hypothesis by showing that the quality and composition of innovation responses matter. They also highlight the importance of interpreting R&D accounting responses cautiously as strategic recognition rather than direct evidence of manipulation.

The policy implication is that regulators should not rely solely on aggregate green patent counts or total R&D amounts when evaluating firms' environmental innovation performance. Regulators should complement quantitative green innovation metrics with quality-sensitive indicators, such as invention patents, patent citations, patent maintenance, commercialization outcomes, and actual environmental performance outcomes. Evaluation systems that emphasize innovation quality rather than only innovation quantity can reduce incentives for symbolic compliance and encourage more substantive green innovation.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this work, the author used Gemini to translate the original draft from Chinese to English and for language polishing to improve readability. After using this tool, the author thoroughly reviewed and edited the content to ensure accuracy and coherence, and takes full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frevc.2026.1894951/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Madhu Verma, Iora Ecological Solutions, India

Reviewed by: Qihui Chen, China Agricultural University, China

Qi Zhang, Harbin Institute of Technology, China