Green Innovation Efficiency in Manufacturing Under Carbon Constraints: Firm Heterogeneity and Resource Allocation in Chongqing
Abstract:
Manufacturing decarbonization requires firms to translate innovation expenditure and productive resources into economic output while limiting energy consumption and greenhouse gas emissions. This study investigates green innovation efficiency and its firm-level variation in Chongqing, with particular attention to frontier persistence, differences by factor intensity, and resource allocation. A balanced panel of 14 A-share-listed manufacturing firms covering 2022–2025 was examined using an epsilon-based efficiency measure incorporating energy inputs and undesirable greenhouse gas outputs. Distributional summaries, frontier transition analysis, and factor-intensity comparisons were combined with exploratory two-limit Tobit regressions. The pooled mean efficiency was 0.7259, and annual means increased from 0.6489 in 2022 to 0.8094 in 2024 before declining to 0.7515 in 2025. Six firms remained on the frontier throughout the period, and frontier observations accounted for 35 of the 56 firm-year observations. Across 42 adjacent-year comparisons, four frontier entries and two exits were recorded, while several firms remained persistently below the frontier. The two labor-intensive firms reached a group mean of one from 2023 onward, whereas technology- and capital-intensive firms exhibited substantial internal variation. Workforce educational attainment was positively associated with the latent efficiency outcome at the 10% significance level, while contemporaneous research and development (R&D) intensity was negatively associated at the 5% level. Both coefficient signs persisted when the final year was excluded. These findings indicate that higher average efficiency can coexist with persistent firm-level disparities. The study provides a basis for differentiated low-carbon transition support focused on workforce capabilities, R&D implementation, and the financing of production upgrades, while distinguishing relative efficiency gains from verified emissions reductions.
1. Introduction
Manufacturing firms face a joint economic and environmental allocation problem. They must finance innovation, maintain production, and control energy use and greenhouse gas emissions within the same operating system. Increasing research and development (R&D) expenditure or patent applications does not, by itself, show that this allocation has improved. A firm may expand innovation activity while its energy demand rises, or reduce emissions by reducing production rather than improving technology. Green innovation efficiency therefore concerns the relationship among resources, useful outputs, and environmental burdens. For Chongqing, this question has a clear regional relevance: earlier research identifies substantial industrial emissions-reduction potential and the importance of coordinated technological and policy responses [1].
This study examines 14 A-share-listed manufacturing firms in Chongqing over 2022–2025. It asks how relative efficiency changed, how much of the aggregate pattern reflects frontier participation rather than improvement among lagging firms, and how efficiency is associated with workforce and financial characteristics. Energy consumption is included as an input and greenhouse gas emissions as an undesirable output. Resource-based reasoning informs the discussion of capabilities, while an epsilon-based measure (EBM) provides the quantitative benchmark. Distributional summaries, frontier transitions, and group contributions are calculated from the firm-year efficiency panel.
The contribution is a firm-level assessment of low-carbon resource allocation in an industrial municipality. It connects innovation resources and economic outputs to energy use and emissions within a common evaluation framework. By distinguishing frontier participation from lower-tail performance, it identifies differences that aggregate means conceal and develops policy priorities around the conversion of resources into operating improvements. Section 2 develops the literature and analytical framework. Section 3 explains the sample, indicators, and methods. Section 4 reports the results. Section 5 discusses their economic interpretation, Section 6 sets out policy implications, and Section 7 concludes.
The policy relevance lies in how green-transition resources are allocated, not simply in whether manufacturing activity expands. A municipality can subsidize research, finance equipment, train workers, or improve access to shared technical services. These interventions operate at different points in the conversion of resources into useful low-carbon outcomes. An efficiency assessment helps identify where further diagnosis is needed, but the usefulness of that assessment depends on preserving the distinction between a relative benchmark and an environmental target. The same firm can be efficient relative to its peers and still face substantial opportunities for absolute emissions reduction.
Chongqing also provides a setting in which broad industrial diversity is concentrated within a common regional policy environment. Comparing listed firms within the municipality reduces some variation in location-level institutions while retaining differences in production processes and organizational capabilities. It does not eliminate differences in market exposure, ownership, product cycles, or environmental disclosure. The research questions are consequently framed at the level of the observed firms. The findings can inform regional policy design, but they do not estimate a municipality-wide efficiency level or describe the performance of firms excluded by the disclosure screen.
2. Literature and Analytical Framework
Green innovation links technological and organizational change to reduced environmental burdens. Rennings distinguishes eco-innovation from conventional innovation through its environmental externalities, while Beise and Rennings connect diffusion to regulation and lead-market conditions [2], [3]. Firm-level work further relates green product and process innovation to competitive advantage and clarifies the multidimensional nature of environmentally sustainable innovation [4], [5]. A bibliometric review documents the breadth of this research domain [6]. These foundations remain useful, but recent evidence calls for a clearer distinction between innovation activity, the efficiency of converting resources into outcomes, and realized environmental performance.
Efficiency studies increasingly incorporate undesirable outputs and the structure of innovation processes. Regional applications use green technology innovation indicators and Super-EBM benchmarking, while firm-level research employs two-stage network data envelopment analysis (DEA) to distinguish knowledge production from commercialization [7], [8], [9]. Directional efficiency and chain-network approaches offer complementary ways of linking economic performance to environmental burdens [10], [11]. These methods answer different questions: a network model locates losses between stages, whereas a single-stage model evaluates the combined resource-output bundle. The present study adopts the latter perspective and interprets its scores as an aggregate benchmark. It does not attribute a score gap to a particular internal stage without supporting process data.
Environmental indicator choice is central to that interpretation. Cross-country eco-technological efficiency work illustrates how environmental outcomes can be embedded in DEA [12], [13], and methodological research develops explicit treatments of undesirable outputs [14]. Evidence from European firms also shows that productivity assessments differ with the pollutant considered [15]. Incorporating greenhouse gases therefore sharpens the low-carbon focus, but it does not make a broad innovation-output measure equivalent to a green-patent measure. The current indicator set combines invention applications, business revenue, and fiscal contribution with energy use and greenhouse gases. It measures innovation-related environmental and economic efficiency rather than the productivity of a separately identified portfolio of green inventions.
The determinants literature identifies several channels through which resources may become useful innovation. Subsidies can alter financing conditions, regulation changes the returns to cleaner production, and customer demand influences environmental product innovation [16], [17], [18]. Work on Industry 4.0 and big data emphasizes the organizational capabilities required to put information and technology into use [19], [20]. Global value-chain participation can also change access to technology and environmental requirements [21], [22], [23]. Recent studies on digital transformation and the joint green and digital transition of manufacturing small and medium-sized enterprises (SMEs) reinforce the importance of complementary resources and institutional conditions [24], [25]. These findings motivate the five firm-level covariates in this study; they do not imply that this short panel can separately identify all of the corresponding channels.
Resource-based theory explains why firms with similar expenditure can differ in their ability to coordinate knowledge, assets, and routines [26]. This perspective motivates a distinction between resource availability and resource conversion. The empirical framework evaluates conversion through EBM scores and examines associations with compensation commitments, leverage, workforce education, R&D intensity, and subsidy intensity. Accounting and workforce indicators describe the observable resources and outcomes used in the efficiency assessment.
Recent research strengthens the case for distinguishing innovation inputs from environmental results. A systematic review of manufacturing green innovation identifies organizational, technological, policy, finance, and capability determinants, showing that no single resource explains the transition [27]. Meta-analytic evidence connects green innovation to environmental performance, but the existence of that general relationship does not imply that every patent or R&D expense has a measurable carbon effect [28]. Research on toxic emissions poses the related firm-level question of whether green innovative manufacturers actually pollute less [29]. For the present study, these contributions justify including an environmental burden in the evaluation and keeping the physical meaning of that burden separate from the innovation proxies.
Workforce capabilities are part of this conversion problem. Green human resource management can connect employee development to environmental strategy and innovation [30]. Organizational networks also shape access to information and the capacity to coordinate a proactive environmental strategy [31]. These studies support a distinction between the quantity of labor, the educational composition of the workforce, and the routines through which employees apply knowledge. The current dataset observes educational attainment but not green job content or training intensity. The resulting empirical association can therefore speak to workforce composition, while the explanation of learning and implementation remains a theoretically motivated interpretation.
Resource conversion also depends on the relationship between cost discipline and innovation capability. Research on emerging-economy SMEs examines how green brand knowledge and green innovation capability condition the relationship between cost efficiency and green product innovation [32]. A manufacturing study links organizational capabilities and environmental regulation to sustainable performance through green innovation [33]. Together, these contributions suggest that spending restraint and spending expansion are insufficient descriptions of effective green management. What matters is whether resources are coordinated around a feasible technical and commercial route. This consideration is especially relevant when contemporaneous R&D intensity and measured efficiency move in opposite directions.
Digitalization has become another major strand of the literature. Research on listed manufacturers examines the structure and direction of green innovation, while a mediated moderation framework connects enterprise digitalization with green innovation [34], [35]. Textual analysis of annual reports and corporate patent records provides further evidence on corporate digitalization [36]. Other work connects digital transformation to environmental performance and green economic growth [37], [38]. These studies are relevant because information systems can affect how firms monitor resource use and coordinate research, production, and reporting. Digitalization is not included as an estimated determinant here, however, so the present results should not be described as a test of a digital-transformation mechanism.
Institutional and competitive conditions help explain why technology adoption need not produce uniform outcomes. Green digitalization research examines institutional moderation, while manufacturing evidence places market competition alongside digital transformation and environmental performance [39], [40]. Case-based work traces the evolution of enterprise green strategies under digital empowerment [41]. In a different organizational setting, AI-related business model research examines the strategic conditions for carbon-neutral business activity among SMEs [42]. The common implication is that technology must be embedded in a working organization. This literature informs the interpretation of heterogeneity, but it does not establish that the sampled Chongqing firms adopted any particular digital or AI system.
Recent benchmarking research provides useful comparisons at several scales. A country-level study evaluates the efficiency of green innovations and renewable-energy activity against sustainability objectives [43]. Firm-level work examines financial misallocation and green innovation efficiency, connecting the allocation of finance with the conversion of innovation resources [44]. Studies of club convergence and correlation networks show that average efficiency can coexist with persistent differences across units [45]. These contributions support an analysis that reports both central tendency and distributional structure. They also caution against treating a change in the sample mean as evidence that all firms moved in the same direction.
Technology-specific research reaches a similar conclusion. AI adoption is examined in relation to green innovation efficiency among Chinese energy companies [46]. A three-stage Super-SBM application studies listed energy-intensive enterprises, separating environmental and operational considerations within its modeling framework [47]. Research on intelligent manufacturing investigates breakthroughs in green innovation [48]. The present analysis differs in its narrow regional sample and its retention of a common aggregate EBM benchmark. Its value lies in making the implications of that benchmark transparent, rather than claiming that a small panel can provide the technological detail or identification available in larger studies.
The relationship between economic and environmental outcomes also has a time dimension. Firm-level evidence links green innovation to financial performance through operating risks [49], while research on low-carbon development examines its relationship with digital transformation [50]. These outcomes may evolve over different horizons. A firm can incur research and implementation costs before the associated revenue or energy savings appear. Conversely, current revenue can improve without establishing a durable green technological advantage. The four-year panel captures a short segment of this process and therefore provides a contemporaneous allocation assessment. It cannot recover a full innovation lifecycle from annual scores alone.
Financing conditions connect innovation decisions to implementation. Chang et al. examine green finance policy and green innovation among heavily polluting listed enterprises [51]. Their policy-evaluation design addresses a different question from the descriptive benchmarking undertaken here, but it highlights the importance of distinguishing access to finance from the results achieved with that finance. In the present setting, subsidy intensity and leverage describe aspects of funding, whereas the efficiency panel describes the joint resource-output position. Connecting these perspectives helps frame project appraisal around an identifiable operating constraint and a measurable environmental objective.
Three analytical expectations guide the assessment. First, a rising aggregate score can coexist with a persistent lower tail when frontier and lagging firms follow different paths. Second, factor intensity can organize observed differences, while firm capabilities and implementation cycles generate variation within groups. Third, workforce education and R&D intensity may be associated with efficiency through capability and timing channels. Distributional and group comparisons examine the first two expectations; exploratory Tobit estimation examines the third.
3. Data Indicators and Methods
The study covers 14 A-share-listed manufacturing firms headquartered in Chongqing during 2022–2025. Screening began with 51 companies identified in the China Stock Market & Accounting Research (CSMAR) database using location and the China Securities Regulatory Commission’s industry classification. Firms designated ST or *ST and those without sufficient information on the selected financial, innovation, energy, and emissions indicators were excluded. The resulting balanced panel contains 56 firm-year observations. Table 1 identifies the firms and their stock codes. The sample is a disclosure-based panel of listed firms; it is not a representative sample of all manufacturers in Chongqing.
| Code | Stock Code | Company Abbreviation | Code | Stock Code | Company Abbreviation |
|---|---|---|---|---|---|
| 1 | 001696 | Zongshen Power | 8 | 603109 | Senci |
| 2 | 003006 | Baiya Corporation | 9 | 603601 | Zaisheng Technology |
| 3 | 300122 | Zhifei-Biol | 10 | 603697 | Youyou Foods |
| 4 | 300194 | Fuan Pharm | 11 | 603758 | QA |
| 5 | 300363 | Porton | 12 | 603766 | Loncin |
| 6 | 600847 | Wanli New Energy | 13 | 603976 | Zhengchuan Pharmapack |
| 7 | 601127 | Seres | 14 | 605122 | Sifang New Material |
Financial and R&D information was compiled from CSMAR and company annual reports. Invention patent applications were obtained from the patent-search system of the China National Intellectual Property Administration. Energy consumption and greenhouse gas information was compiled from corporate environmental disclosures and corporate social responsibility or sustainability reports, including the Chongqing enterprise environmental information disclosure platform. Linear interpolation was used for a small number of gaps in the underlying compilation. The interpolation procedure preserves a balanced panel but can smooth genuine discontinuities; this matters particularly when interpreting abrupt changes in efficiency. The empirical period ends in 2025, and no later-period observation enters the analysis.
The disclosure screen retains 27.45% of the initially identified firms. This proportion matters because environmental reporting is not randomly distributed across companies. Larger firms, firms with stronger reporting systems, and those facing particular regulatory or investor pressures may be more likely to provide usable records. The final panel therefore offers detailed information about a selected group while leaving the behavior of excluded manufacturers outside the empirical comparison. The balanced structure ensures that annual score changes are not produced by firms entering or leaving the sample, although reporting differences within retained firms can still affect measurement.
The dataset combines physical and monetary measures. Energy use and greenhouse gases describe resources and environmental burdens, while revenue, expenditure, and fiscal balances describe economic activity. DEA evaluates their joint relationship through normalized radial and slack components. Interpretation remains conditional on the reporting boundaries of the underlying records, particularly where consolidated financial statements and environmental disclosures cover different organizational units. Accordingly, the scores are used as innovation-related environmental-economic benchmarks, with direct emissions indicators retained as a separate requirement for project-level decarbonization assessment.
The indicator system links resources committed to production and innovation with economic, technological, and environmental outputs. Its five inputs are comprehensive energy consumption, R&D personnel, R&D expenditure, net fixed assets, and the cost-to-revenue ratio. Its three desirable outputs are main-business revenue, invention patent applications, and taxes and fees payable; greenhouse gas emissions form the undesirable output. This combination captures the economic and environmental dimensions of innovation-related resource allocation.
Resource-based reasoning provides the organizing perspective for Table 2. Energy represents an operating resource, R&D personnel and expenditure represent knowledge-development resources, and fixed assets describe productive capacity. Invention applications record inventive activity, while revenue and fiscal balances represent economic outcomes. These roles are observable in the dataset. Resource rarity, inimitability, and organizational effectiveness are broader strategic attributes that require separate measures and are therefore not assigned to individual accounting indicators.
Role | Indicator and Unit | Interpretation |
|---|---|---|
Input | Energy consumption (tce) | Operating energy resource |
Input | R&D personnel (persons) | Research workforce |
Input | R& expenditure (yuan) | Research expenditure |
Input | Net fixed assets (yuan) | Productive capital stock |
Input | Cost-to-revenue ratio (%) | Operating cost proxy |
Desirable output | Main-business revenue (yuan) | Economic output |
Desirable output | Invention patent applications (count) | Inventive activity |
Desirable output | Taxes and fees payable (yuan) | Fiscal accounting proxy |
Undesirable output | Greenhouse gas emissions (tonnes) | Environmental burden |
The environmental interpretation rests principally on energy consumption and greenhouse gas emissions. Energy enters as a resource whose use is evaluated against the joint output bundle; emissions enter as a burden to be reduced in the benchmark comparison. A firm is not penalized simply for operating in a large market, since DEA compares combinations of inputs and outputs. Nevertheless, the choice of constant returns to scale and a common reference technology means that scale and technological differences remain relevant. Economic and fiscal outputs can compensate within the composite score, so frontier status must be read alongside separate environmental indicators in practical decision-making.
Table 2 reports the nine indicators, their units, and their interpretation. The cost-to-revenue ratio is a dimensionless operating-cost proxy; a lower ratio represents a less costly revenue-generating process, other things equal. Taxes and fees payable are retained as a fiscal-contribution proxy, although a year-end liability is not identical to taxes actually paid. The patent variable counts applications rather than grants or citations. These distinctions constrain interpretation and prevent the composite measure from being presented as a direct carbon productivity or green patent-quality index.
The use of total invention applications also broadens the meaning of innovation. An application records an inventive effort at a particular date, but it may concern a product whose main purpose is unrelated to energy or emissions. Applications also precede examination and do not measure subsequent adoption. Including emissions in the same efficiency system imposes an environmental dimension on the joint evaluation, but it does not retrospectively classify all applications as green. The terminology used throughout the paper reflects this distinction: green innovation efficiency refers to the combined innovation-related economic and environmental benchmark, rather than a count-based measure of verified green inventions.
A similar distinction applies to economic outputs. Main-business revenue captures market activity and supports a comparison of resources with commercially realized output. It can increase because of higher sales volumes, prices, or changes in product mix. Taxes and fees payable add a fiscal dimension but can also change with settlement timing and tax rules. Neither variable measures environmental quality directly. For a low-carbon development interpretation, these outputs indicate the economic side of the trade-off; energy and greenhouse gases indicate the physical resource and environmental side. Practical project appraisal should preserve both sides instead of treating the composite score as a sufficient statistic for investment decisions.
The efficiency of green innovation in the manufacturing sector depends not only on the structure of R&D inputs and outputs but is also influenced by a combination of factors, including corporate governance, financial structure, human capital, technological transformation, and external policies. Based on existing research and data completeness, this study evaluates the potential determinants of green innovation efficiency in the manufacturing sector across five dimensions: corporate governance, financial structure, human capital, technological transformation, and external policies, as shown in Table 3.
Variable | Definition | Regression Scale |
|---|---|---|
Compensation payable ratio | Employee compensation payable/operating revenue | Decimal ratio |
Leverage | Total liabilities/total assets | Decimal ratio |
Workforce education | Employees with bachelor’s degree or higher/total employees | Percentage points |
R&D intensity | R&D expenditure/operating revenue | Decimal ratio |
Subsidy intensity | Government subsidies/operating revenue | Decimal ratio |
Compensation payable ratio is employee compensation payable divided by operating revenue. It describes the scale of the accounting obligation to employees relative to business activity. Compensation commitments may reflect incentives and workforce requirements, while also competing with other uses of working capital. Because the numerator is a payable balance, the variable is interpreted as a financial commitment rather than annual wages paid.
Leverage is total liabilities divided by total assets. Debt can expand the resources available for investment, while repayment obligations and risk constraints can restrict long-horizon research or equipment replacement. The sign of its association with contemporaneous efficiency therefore depends on how borrowing supports operating and innovation activities.
Workforce education is the percentage of employees holding a bachelor’s degree or higher. It describes educational composition and provides a proxy for the knowledge base available for interpreting technical information and implementing production changes. Its relationship with efficiency may depend on the allocation of qualified employees to research, engineering, and operating responsibilities.
R&D intensity is R&D expenditure divided by operating revenue. It captures current research spending relative to business scale. Research builds technological capability, but expenditure and realized outcomes can occur in different periods. Current R&D intensity can therefore be associated with lower contemporaneous efficiency even when the underlying projects have longer-term value.
Subsidy intensity is government subsidies divided by operating revenue. Public funding can ease financing constraints and share project risk, while program objectives and targeting affect its relationship with implementation. The variable covers aggregate subsidies and is therefore broader than funding specifically earmarked for green innovation.
The compensation variable is based on a payable balance rather than total annual compensation expense. It therefore captures the relationship between an accounting obligation to employees and revenue, not a complete measure of distributive fairness or incentive design. The term compensation payable ratio is used in the interpretation to avoid confusing a balance-sheet measure with wages actually paid. Likewise, government subsidies divided by revenue measures aggregate subsidy intensity. It does not identify whether the funding supports green projects, general innovation, employment, or other public objectives. A statistically insignificant coefficient on this broad ratio cannot establish that targeted green subsidies are ineffective.
The covariates also differ in their relationship with the first-stage inputs. R&D intensity includes R&D expenditure, which already enters the EBM input vector. This overlap makes its coefficient partly a relationship between a component of the efficiency construction and the resulting score. Leverage and subsidy intensity can be correlated with unobserved financial constraints, ownership, and policy targeting. Education can be correlated with firm technology and business strategy. The regression therefore supplies a compact conditional summary rather than an exogenous decomposition of the causes of inefficiency. This qualification is necessary before the coefficient estimates are translated into management implications.
EBM combines proportional resource-output adjustment with indicator-specific slacks [52]. It provides a bridge between radial DEA and the slack-based approach [53]. The efficiency panel is interpreted under a non-oriented, constant-returns-to-scale benchmark that includes greenhouse gases as an undesirable output. To distinguish the estimated scores from the subsequent distributional calculations, the measurement relationship is expressed compactly as follows:
In Eq. (1), $e$ denotes the relative efficiency score; $x$, $y$, and $b$ denote the input, desirable-output, and undesirable-output vectors; $g$ and $t$ index firms and years; and $T$ denotes the benchmark technology. The operator $E$ denotes the EBM evaluation, with $\varepsilon$ and $w$ representing its radial–slack balance and normalized indicator weights. Under constant returns to scale, the reference activity intensities are not constrained to sum to one. Frontier observations have a reported score of one. The EBM benchmark evaluates a joint input-output bundle, so its complement is not a physical estimate of feasible emissions reduction. Productivity and super-efficiency approaches provide related methodological context [54], [55].
The empirical analysis uses the firm-year scores reported in Table 4. Its distributional statistics, group contributions, and transitions are calculated directly from that panel. This distinction keeps the numerical evidence transparent: the additional summaries describe the dispersion and persistence of relative performance, while interpretation of the first-stage frontier remains conditional on its input-output definitions and benchmark specification.
| Stock Code | Company Abbreviation | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|
| 003006 | Baiya Corporation | 0.0008 | 1 | 1 | 1 |
| 300363 | Porton | 0.7549 | 0.4013 | 0.5209 | 0.0010 |
| 300194 | Fuan Pharm | 0.0018 | 0.0018 | 0.0024 | 0.0043 |
| 603766 | Loncin | 0.5622 | 1 | 1 | 1 |
| 603758 | QA | 0.0004 | 0.0011 | 1 | 0.7226 |
| 601127 | Seres | 1 | 1 | 1 | 1 |
| 603109 | Senci | 1 | 1 | 1 | 0.7912 |
| 605122 | Sifang New Material | 1 | 1 | 1 | 1 |
| 600847 | Wanli New Energy | 1 | 1 | 1 | 1 |
| 603697 | Youyou Foods | 1 | 1 | 1 | 1 |
| 603601 | Zaisheng Technology | 1 | 1 | 1 | 1 |
| 603976 | Zhengchuan Pharmapack | 0.2194 | 0.001 | 0.0019 | 0.0021 |
| 300122 | Zhifei-Biol | 1 | 1 | 1 | 1 |
| 001696 | Zongshen Power | 0.5444 | 0.3083 | 0.8058 | 1 |
Distributional analysis combines empirical cumulative distributions with a boundary-reflected Gaussian density. Exact frontier counts, medians, quartiles, and non-frontier means provide the principal summaries because each annual sample contains only 14 scores and a substantial point mass at one. The density uses a common bandwidth of 0.10 across all four years, allowing a consistent descriptive comparison. It illustrates the location of observations on the bounded scale without serving as a formal test of convergence or the number of population modes.
For a score $e$ in $[ 0,1]$, the reflected estimate adds Gaussian kernels centered at $e$, $-e$, and $2-e$, evaluates them over $[ 0,1]$, and normalizes the curve to unit area. Here $n$ is the annual sample size, $h$ is the bandwidth, $K$ is the standard Gaussian density, and $C_h$ is the integral of the unnormalized estimate over $[ 0,1]$. This boundary correction reduces kernel leakage but cannot remove the discrete point mass created by frontier observations. For that reason, the empirical distribution is shown separately and the graphical interpretation is supported by exact counts. Changes in a smoothed peak are treated as features of the display rather than independent evidence that firms became environmentally cleaner.
A two-limit Tobit model describes associations between the efficiency scores and the five covariates in Table 3. A latent outcome is specified as a linear function of the covariates, and the observation equation maps values outside $[ 0,1]$ to the nearest bound. The model is used for exploratory association analysis because DEA frontier observations arise from benchmarking rather than a directly observed censoring process. Simar and Wilson discuss the dependence and inference issues associated with second-stage efficiency regressions [56].
The indices $g$ and $t$ denote firm and year. $\mathrm{COMP}$, $\mathrm{LEV}$, $\mathrm{EDU}$, $\mathrm{RD}$, and $\mathrm{SUB}$ denote the compensation payable ratio, leverage, workforce education, R&D intensity, and subsidy intensity, respectively. The disturbance is assumed normal in the fitted Tobit likelihood. The reported estimates use observed-information-matrix standard errors. Firm and year effects are not separately estimated in this short panel. Coefficients refer to the latent outcome, not to marginal changes in observed efficiency; no average marginal effects are reported. Re-estimation after excluding 2025 assesses sensitivity to the observation period. It does not correct within-firm dependence, first-stage estimation uncertainty, endogeneity, or omitted common shocks. Those limitations delimit the inferential claims made below.
The distributional calculations use the 56 efficiency scores in Table 4 and the firm classifications in Section 4.3. Annual summaries include the arithmetic mean, median, sample standard deviation, quartiles, frontier share, and mean among observations below one. Quartiles use linear interpolation between ordered observations. Frontier membership is identified from a reported score of one, consistently with the four-decimal precision of Table 4.
For group $k$ in year $t$, the group mean is the arithmetic mean of the scores of its $n_k$ member firms. Its contribution to the full-sample mean is $(n_k/14)$ multiplied by that group mean. Because group membership and sample size are fixed, changes in these weighted contributions sum exactly to the change in the overall mean. This is an accounting decomposition of observed scores, not a causal decomposition of green development. It is useful precisely because it reveals how a small high-performing group and a larger mixed-performance group affect the same aggregate statistic.
Frontier transitions are computed for the 42 adjacent-year firm pairs. A firm can remain on the frontier, enter it, leave it, or remain below it. These four states summarize persistence and movement without requiring a smoothed density or a model of transition probabilities. With only three intervals and 14 firms, no long-run stationary distribution is estimated. The reported frequencies describe the observed period only. The four-year mean of each firm is also used to identify persistent performance patterns, while endpoint changes show how much each firm contributes arithmetically to the net sample change.
4. Results
Table 4 reports all 56 firm-year efficiency scores. Their pooled mean is 0.7259 and sample standard deviation is 0.4056. Firm-level four-year means range from 0.0026 for Fuan Pharm to 1 for the six persistent frontier firms. Nine of the 14 firms have a four-year mean above the pooled mean. This is a relative ranking within the selected benchmark, rather than evidence that the sector as a whole has already achieved low-carbon production. The coexistence of many frontier observations with very low scores makes the distribution more informative than its mean alone.
Annual means are 0.6489, 0.6938, 0.8094, and 0.7515 for 2022–2025, respectively. Frontier counts are 7, 9, 10, and 9. The total increase between the first and last years is 0.1027 score units; the largest annual increase is 0.1155 between 2023 and 2024, followed by a decline of 0.0578. These movements establish the temporal pattern in the scores. They do not isolate the contributions of technological change, revenue fluctuations, or emissions changes, because the composite measure combines all of these dimensions.
Seres, Sifang New Material, Wanli New Energy, Youyou Foods, Zaisheng Technology, and Zhifei-Biol have scores of one in every year. Their persistence distinguishes them from firms that reached the frontier only in selected periods. However, the EBM scores do not rank these six firms against one another. Nor does a score of one establish zero emissions, optimal process design, or the absence of additional abatement opportunities. The substantive finding is that no improvement in their joint input-output position is identified by the specified reference comparison.
Baiya Corporation moves from 0.0008 in 2022 to one in the remaining years. Loncin moves from 0.5622 to one over the same interval. Zongshen Power falls from 0.5444 to 0.3083 before increasing to 0.8058 and then one. QA increases from 0.0004 and 0.0011 to one in 2024, then declines to 0.7226. Senci remains on the frontier through 2024 and records 0.7912 in 2025. Figure 1 summarizes these contrasting firm-level trajectories.

The lower tail persists despite these transitions. Fuan Pharm records 0.0018–0.0043, and Zhengchuan Pharmapack records 0.0010–0.2194. Porton changes from 0.7549 to 0.4013 and 0.5209 before falling to 0.0010. These observations account for much of the distance between the mean and median. Extremely low scores can be sensitive to measurement boundaries, indicator scaling, and the mix of economic and environmental outputs. They warrant examination of the underlying resource-output records before being translated into firm-specific managerial diagnoses. The analysis therefore reports their arithmetic contribution to the distribution without claiming a verified causal explanation for each trajectory.
The distinction between the frontier and the rest of the sample is quantitatively important. The mean of the non-frontier firms is 0.2977 in 2022, 0.1427 in 2023, 0.3328 in 2024, and 0.3042 in 2025. Thus, the rise in the overall mean between 2022 and 2023 occurs alongside a decline in the non-frontier mean. Two firms enter the frontier over that interval, changing the composition of the remaining non-frontier group. This does not mean every lagging firm became worse. It means that the full-sample mean and the conditional non-frontier mean answer different questions and should be examined together.
The annual median tells a different story again. It reaches one once a majority of observations are at the frontier and cannot reflect subsequent differences among those frontier firms. The lower quartile rises substantially in 2024, then declines in 2025. Consequently, the panel does not support a simple account of continuous distributional improvement. The upper part of the distribution is compressed at the model boundary, while several lower-tail scores remain far from it. Reporting both quartiles and frontier shares helps avoid interpreting a bounded distribution as though it were an ordinary unbounded performance index.
Table 5 also shows why a near-zero score deserves a different interpretation from a moderate score. The former has a large effect on the mean distance from the frontier but reveals little about the source of that distance without the associated slacks and input-output records. A firm with a score of 0.001 is not thereby shown to have an achievable 99.9% emissions reduction. The score reflects the chosen joint objective, including economic and innovation outputs. Its policy use is to identify a case for closer diagnosis, not to generate a physical abatement target by subtraction from one.
Statistic | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|
Mean | 0.6489 | 0.6938 | 0.8094 | 0.7515 |
Median | 0.8775 | 1.0000 | 1.0000 | 1.0000 |
Standard deviation | 0.4239 | 0.4400 | 0.3669 | 0.4152 |
Lower quartile | 0.3007 | 0.3316 | 0.8544 | 0.7398 |
Upper quartile | 1.0000 | 1.0000 | 1.0000 | 1.0000 |
Non-frontier mean | 0.2977 | 0.1427 | 0.3328 | 0.3042 |
Frontier count | 7 | 9 | 10 | 9 |
Frontier share (%) | 50.00 | 64.29 | 71.43 | 64.29 |
Score below 0.01 | 3 | 3 | 2 | 3 |
Figure 2 presents the empirical cumulative distributions alongside the reflected density curves. The common bandwidth highlights the concentration of scores near the model boundary and the persistence of a small lower-tail group. The empirical curves show the observed cumulative proportions directly, while Table 5 supplies the exact counts and summary statistics. Together, these representations separate frontier concentration from changes in performance below the frontier.

Half of the firms were on the frontier in 2022, and the proportion exceeded one-half in each subsequent year. The median therefore increased from 0.8775 to one in 2023 and remained at that level. Several firms nevertheless retained scores close to zero, showing that frontier concentration coexisted with persistent lower-tail performance.
The evidence supports a persistent lower-tail group alongside a large frontier group. It does not establish statistical bimodality, distributional convergence, or a steadily narrowing gap. In 2025, three firms have scores below 0.01, while nine remain at one. This combination explains why the median can stay at the frontier even as the mean falls. The distributional problem is therefore one of uneven participation in the measured efficiency gains. A policy response based only on the median or on the number of frontier firms would overlook this lower tail.
Table 6 classifies firms by their dominant factor-intensity characteristics: eight technology-intensive, four capital-intensive, and two labor-intensive firms. The groups are descriptive categories applied to the study sample. They do not represent separately estimated technologies, and all group averages use the same underlying EBM scores. Their unequal sizes must be retained when interpreting aggregate changes. Figure 3 shows the resulting annual group means, and Table 7 reports their contributions to the sample mean.
| Stock Code | Company Abbreviation | Enterprise Type | Stock Code | Company Abbreviation | Enterprise Type |
|---|---|---|---|---|---|
| 001696 | Zongshen Power | Technology-Intensive | 603109 | Senci | Technology-Intensive |
| 003006 | Baiya Corporation | Labor-Intensive | 603601 | Zaisheng Technology | Technology-Intensive |
| 300122 | Zhifei-Biol | Technology-Intensive | 603697 | Youyou Foods | Labor-Intensive |
| 300194 | Fuan Pharm | Technology-Intensive | 603758 | QA | Capital-Intensive |
| 300363 | Porton | Technology-Intensive | 603766 | Loncin | Technology-Intensive |
| 600847 | Wanli New Energy | Capital-Intensive | 603976 | Zhengchuan Pharmapack | Capital-Intensive |
| 601127 | Seres | Technology-Intensive | 605122 | Sifang New Material | Capital-Intensive |

| Group | $\boldsymbol{n}$ | Mean 2022 | Mean 2025 | Change in Weighted Contribution |
|---|---|---|---|---|
| Technology intensive | 8 | 0.7329 | 0.7246 | $-$0.0048 |
| Capital intensive | 4 | 0.5550 | 0.6812 | $+$0.0361 |
| Labor intensive | 2 | 0.5004 | 1.0000 | $+$0.0714 |
| All firms | 14 | 0.6489 | 0.7515 | $+$0.1027 |
The technology-intensive group has the highest mean in 2022, while the labor-intensive group ranks first from 2023 onward. Both labor-intensive firms are frontier observations in the last three years. The technology-intensive group contains both persistent frontier firms and low-score firms, so its average conceals considerable variation. The capital-intensive group has the lowest mean in the final three years, although two of its four firms remain on the frontier throughout. These are within-sample contrasts, not formal tests of industry effects. Research on heterogeneous high-technology efficiency, manufacturing decarbonization, and innovation ecosystems helps explain why factor intensity is a useful discussion lens [57], [58], [59], but the current data do not independently identify asset lock-in or commercialization delays as the causes of the observed rankings.
The labor-intensive group contributes 0.0715 to the sample mean in 2022 and 0.1429 in 2025. Its endpoint contribution therefore rises by 0.0714. The corresponding contributions of the capital-intensive group are 0.1586 and 0.1946, an increase of 0.0361. The technology-intensive contribution declines slightly, from 0.4188 to 0.4140, or $-$0.0048. These changes sum to the overall increase of 0.1027. The labor-intensive group thus accounts for the largest positive arithmetic contribution despite containing only two firms. That result is driven by Baiya Corporation’s transition to the frontier, not by evidence of a broad improvement across many labor-intensive manufacturers.
Within the technology-intensive group, positive and negative changes offset one another. Zongshen Power and Loncin improve between the first and last years, while Porton and Senci decline. Several other firms remain on the frontier. The nearly unchanged group average therefore conceals substantial movement in opposite directions. Within the capital-intensive group, QA’s increase is partly offset by Zhengchuan Pharmapack’s decline, while the other two firms remain at one. The group decomposition complements the trajectory plot by showing why group-level stability should not be confused with uniform firm-level stability.
Table 8 and Figure 4 summarize the 42 adjacent-year transitions. There are 24 instances of remaining on the frontier, four entries, two exits, and 12 instances of remaining below it. The four entries occur for Baiya Corporation and Loncin in 2023, QA in 2024, and Zongshen Power in 2025. The two exits occur for QA and Senci in 2025. This sequence explains the annual frontier counts of seven, nine, ten, and nine. It also shows that the final-year decline includes both exits from the frontier and a large fall for a firm that was already below it. Across all intervals, 24 of the 26 pairs beginning on the frontier remain there, an observed retention frequency of 92.3%. Four of the 16 pairs beginning below the frontier enter it, an observed entry frequency of 25.0%. These frequencies describe the sample and period; they are not estimated long-run probabilities. Persistence can reflect durable capabilities, stable output profiles, or limited model discrimination. The transition analysis cannot distinguish these sources, but it makes the concentration of measured gains more explicit than an annual mean alone.
Transition | Remain on Frontier | Enter Frontier | Leave Frontier | Remain Below Frontier |
|---|---|---|---|---|
2022–2023 | 7 | 2 | 0 | 5 |
2023–2024 | 9 | 1 | 0 | 4 |
2024–2025 | 8 | 1 | 2 | 3 |
Total | 24 | 4 | 2 | 12 |

The largest positive endpoint change is Baiya Corporation’s 0.9992, followed by QA’s 0.7222, Zongshen Power’s 0.4556, and Loncin’s 0.4378. Porton records the largest negative endpoint change, $-$0.7539. These movements illustrate why the identity of changing firms matters. A regional average based on 14 observations per year can move appreciably because of one or two trajectories. The balanced panel avoids sample-entry effects, but it remains sensitive to firm-specific changes. Aggregate policy conclusions should therefore be checked against the individual paths before they are generalized.
Table 9 summarizes the efficiency scores and five covariates for the 56 firm-year observations. The covariates are compensation payable ratio, leverage, workforce education, R&D intensity, and subsidy intensity. Their units follow the definitions in Table 3, allowing the coefficients to be interpreted consistently with the regression scale.
Variable Name | Sample Size | Mean | Standard Deviation | Minimum | Maximum |
|---|---|---|---|---|---|
Green innovationefficiency | 56 | 0.7259 | 0.4056 | 0.0004 | 1.0000 |
Compensation payableratio | 56 | 0.0187 | 0.0093 | 0.0041 | 0.0529 |
Leverage | 56 | 0.3645 | 0.1760 | 0.0742 | 0.8738 |
Workforce education | 56 | 21.7853 | 15.5200 | 3.9709 | 63.4571 |
R&D intensity | 56 | 0.0357 | 0.0276 | 0.0021 | 0.1233 |
Subsidy intensity | 56 | 0.0083 | 0.0089 | 0.0002 | 0.0557 |
The variables differ in their measurement scales. Compensation payable, leverage, R&D intensity, and subsidy intensity are expressed as decimal ratios in the estimates, whereas workforce educational attainment is expressed in percentage points. Their standard deviations therefore cannot be compared directly as measures of relative variability. R&D intensity averages 0.0357, or 3.57% of revenue, and subsidy intensity averages 0.0083, or 0.83%. Workforce education averages 21.7853%, with a range of 3.9709–63.4571%. These differences describe the covariate variation available to the pooled regression.
Table 10 reports variance inflation factors. Their maximum is 2.57 and their arithmetic mean is 1.892, providing no indication of severe linear collinearity among the five observed covariates under conventional screening thresholds. This diagnostic does not establish exogeneity or remove conceptual overlap between R&D intensity and the R&D input used to construct efficiency. The reciprocal column is calculated directly as $1/\mathrm{VIF}$; it is a presentation of the same diagnostic rather than an additional test.
Variable | VIF | 1/VIF |
|---|---|---|
Compensation payable ratio | 2.22 | 0.4505 |
Leverage | 1.34 | 0.7463 |
Workforce education | 2.11 | 0.4739 |
R&D intensity | 2.57 | 0.3891 |
Subsidy intensity | 1.22 | 0.8197 |
Mean VIF | 1.892 | — |
Table 11 presents the exploratory Tobit estimates. Workforce education has a positive coefficient of 0.0246 with $p = $ 0.079, while R&D intensity has a negative coefficient of $-$15.4936 with $p = $ 0.041. The remaining coefficients are not statistically distinguishable from zero at conventional levels. These are conditional associations under the fitted likelihood and standard-error assumptions, considered alongside the concentration of frontier observations and the small number of firms.
Variable | Coefficient | Test Statistic | $p$-Value |
|---|---|---|---|
Compensation payable ratio | $-$25.0847 | $-$1.29 | 0.202 |
Leverage | 0.6451 | 0.66 | 0.514 |
Workforce education | 0.0246$^{*}$ | 1.79 | 0.079 |
R&D intensity | $-$15.4936$^{**}$ | $-$2.10 | 0.041 |
Subsidy intensity | $-$4.9798 | $-$0.35 | 0.724 |
Log likelihood | $-$41.2588 | — | — |
Observations | 56 | — | — |
The positive education coefficient is consistent with a role for workforce capabilities in resource conversion, but the evidence is weak at the 5% threshold. The negative R&D coefficient refers to contemporaneous expenditure intensity, not to the long-run return to innovation. A one-percentage-point increase in R&D intensity corresponds to a 0.01 change in its decimal-ratio regressor; multiplying by the coefficient gives $-$0.1549 on the latent scale. This is a scale interpretation, not an estimated reduction of 15.49 efficiency points in observed outcomes. The coefficient may reflect lags, allocation differences, or mechanical overlap with the first-stage R&D input. The current specification does not distinguish those explanations. Research on green finance and local regulation provides relevant institutional context [60]. In this sample, however, the subsidy estimate does not establish a clear conditional association with efficiency.
Table 12 reports the estimates after excluding 2025, leaving 42 observations. All five coefficient signs are unchanged. Education remains positive, with $p =$ 0.053, while R&D intensity remains negative, with $p =$ 0.013. The R&D coefficient changes from $-$15.4936 to $-$9.4699, a 38.9% reduction in absolute magnitude. The compensation coefficient changes more substantially, from $-$25.0847 to $-$1.4089, although it is imprecise in both specifications. The comparison therefore supports sign consistency for the two focal variables and demonstrates sensitivity in coefficient magnitude. It is a sample-period sensitivity check rather than comprehensive validation of the regression.
Variable | Coefficient | Test Statistic | $\boldsymbol{p}$-Value |
|---|---|---|---|
Compensation payable ratio | $-$1.4089 | $-$0.15 | 0.882 |
Leverage | 0.0917 | 0.26 | 0.798 |
Workforce education | 0.0130$^{*}$ | 2.00 | 0.053 |
R&D intensity | $-$9.4699$^{**}$ | $-$2.62 | 0.013 |
Subsidy intensity | $-$3.1005 | $-$0.48 | 0.636 |
Log likelihood | $-$15.6075 | — | — |
Observations | 42 | — | — |
The education association is positive at the 10% level in both estimation periods, while the R&D association is negative at the 5% level. Coefficient magnitudes nevertheless change. The shorter-period R&D estimate is 38.9% smaller in absolute value, and the education estimate falls from 0.0246 to 0.0130. Figure 5 displays the two pairs of coefficients on separate axes to preserve their different regressor scales. The comparison supports sign consistency while showing the importance of the observation period for estimated magnitudes.

For compensation payable, leverage, and subsidies, the imprecision of the estimates prevents a clear directional conclusion. A negative point estimate can coexist with a wide range of economically different underlying relationships. In particular, the sample does not contain the information required to determine whether subsidy targeting responds to poor prior performance or whether financing conditions are shaped by unobserved investment risk. The absence of a conventionally significant coefficient is therefore a limit on what the model establishes, rather than evidence that the corresponding policy or governance channel does not operate.
5. Discussion
The education association is economically plausible because skilled employees help interpret technical information, operate equipment, and coordinate research with production. Its practical meaning depends on job assignment and production routines, rather than academic credentials alone. The coefficient of 0.0246 is measured per percentage point of educational attainment on a latent scale. Its size cannot be judged by comparing it directly with coefficients on decimal ratios. Evidence linking green human resource management to environmental strategy and green innovation supports the importance of this organizational context [30]. The present panel, however, contains no direct measure of training content, engineering experience, or green occupations, so the result should guide capability assessment rather than justify a universal hiring target.
R&D intensity is negatively associated with contemporaneous efficiency in both reported specifications. An efficiency ratio can decline while a firm builds future technological capability: expenditure enters the current input bundle, whereas patents, commercial revenue, and energy savings may arrive later. A second possibility is that R&D supports product development with little immediate relevance to operating emissions. A third is a mechanical association arising because R&D expenditure appears in the first-stage score and in the second-stage regressor. These explanations are compatible with the observed sign but are not separately tested. The policy implication is to examine how projects reach implementation and measurable environmental outcomes, rather than infer that reducing R&D is an efficient decarbonization strategy.
The group results show why a common allocation rule is insufficient. Technology-intensive firms differ sharply despite sharing a broad factor-intensity label. Capital-intensive firms also include persistent frontier observations alongside firms near the lower tail. Labor-intensive performance rests on only two firms, one of which changes abruptly between the first two years. The relevant distinction is therefore not a deterministic hierarchy of technologies. It is the combination of factor requirements, firm-specific capabilities, and the performance path observed over time. Group averages can organize diagnosis, but they cannot establish the technological cause of an individual firm’s score.
This interpretation changes how the findings should inform regional governance. An efficient allocation mechanism needs to distinguish persistent frontier status, transition into or out of the frontier, and persistent lower-tail performance. These states indicate different questions for project appraisal: whether leading firms can demonstrate transferable operating practices, whether transitioning firms can sustain their gains, and whether low-score firms face measurement problems or genuine resource-conversion constraints. The current evidence supports this diagnostic ordering. Choosing a specific retrofit, subsidy rate, or lending condition still requires project-level information on costs, energy use, and expected emissions savings.
The framework joins innovation resources, economic outputs, and greenhouse gas burdens in a common relative comparison. Its advantage is this joint view of resource allocation. Its limitations follow from the same design: broad patents and revenue capture outcomes beyond green innovation, while fiscal liabilities are an imperfect proxy for social contribution. The indicator set is consequently best understood as innovation-related low-carbon efficiency. It should be complemented by direct measures of energy intensity, emissions intensity, and absolute emissions when evaluating decarbonization projects. A favorable composite score cannot substitute for those physical measures.
The large number of frontier observations also limits discrimination. Five inputs and four outputs are evaluated for a small firm sample, making it easier for specialized output profiles to appear efficient. Constant returns to scale and a common technology simplify comparison but can mix managerial efficiency with scale and sectoral differences. The Tobit stage then inherits uncertainty from the generated scores. Its conventional likelihood inference does not capture that uncertainty. These features are substantive limits of the assessment and explain why the paper emphasizes descriptive patterns, transparent arithmetic, and conditional associations rather than claims of causal identification or complete methodological validation.
The aggregate pattern is best described as uneven relative performance. The observed increase in the mean has a substantial contribution from frontier entry by a small number of firms. Persistent frontier firms anchor the upper part of the distribution, while several firms remain near the lower boundary. These features matter for regional resource allocation because the cost of helping a firm cross an implementation barrier can differ from the cost of improving an already effective operating system. A policy that rewards only the highest current score may overlook firms with feasible improvement projects; a policy that rewards only low scores may misallocate funds to cases where the measurement itself requires correction.
A useful policy sequence is therefore to verify comparable reporting, diagnose the operating constraint, appraise a specific intervention, and track its implementation. The efficiency score can help prioritize attention at the first two stages. Engineering estimates and financial appraisal are needed at the third, and physical monitoring is needed at the fourth. This sequence connects the benchmark to low-carbon development without assigning the model tasks it was not designed to perform. It also makes clear why a change in a composite score should be accompanied by evidence on production, energy use, and emissions before being described as a successful decarbonization outcome.
The same logic applies to learning from leading firms. Persistent frontier firms are potential sources of process knowledge, but the model alone does not establish which practice should be transferred. A specific operating routine must be documented and assessed for technical compatibility with the receiving firm. Learning across an industrial chain may be useful where firms share suppliers, energy systems, or process requirements, while transfer across unrelated production technologies may be limited. The regional contribution of an efficiency study is to structure those questions and identify contrasts for investigation, rather than to infer a universal best practice from a score of one.
6. Policy Implications
Regional support should begin with a firm-level diagnosis of innovation and environmental performance. The classification should combine score trajectories with energy and emissions information, distinguishing stable frontier firms, firms moving toward the frontier, firms moving away from it, and persistent lower-tail firms. This approach uses the observed heterogeneity without treating a relative score as a complete assessment of environmental quality. Support conditions should specify the operating change to be delivered and the environmental measure by which it will be evaluated.
For technology-intensive firms, project appraisal should connect research milestones to commercialization and implementation. A technically successful invention may still lack a route to lower operating energy use or emissions. Pilot facilities, production-line trials, and joint engineering teams can help establish that route. Public support should distinguish general product R&D from projects with a defined low-carbon objective, and review outcomes over a period consistent with the development cycle. Workforce programs should focus on the skills needed to operate and maintain the proposed technology rather than reward educational credentials without regard to production responsibilities.
For capital-intensive firms, financing should match the operating life and cash-flow profile of retrofits. Equipment upgrades often require coordination with maintenance shutdowns and production schedules. Project evaluation should therefore compare expected energy savings, emissions effects, downtime, and financing costs over the relevant horizon. The low group mean alone does not prove that all capital-intensive firms need identical assistance. Persistent frontier firms may offer useful process examples, while firms with low scores require separate verification of accounting and environmental boundaries before a costly investment prescription is made.
For labor-intensive firms, the small sample favors an emphasis on maintaining performance and improving access to practical green management. Shared energy audits, supplier guidance, and technical training may reduce the fixed costs of identifying feasible improvements. Support programs should combine documented implementation with comparable environmental reporting and account for smaller suppliers’ capacity to participate. These priorities follow from the resource-allocation problem identified by the assessment; their effectiveness should be evaluated at the project level.
A regional monitoring framework should retain three complementary outcomes. Absolute greenhouse gas emissions indicate the total environmental burden; emissions intensity relates that burden to a defined activity measure; and relative efficiency describes the joint resource-output position against a chosen benchmark. These measures can move differently. Production expansion may improve resource allocation while increasing total emissions, and a temporary reduction in production may lower absolute emissions without an improvement in technology. Monitoring all three provides a more informative basis for evaluating public support than a single composite target.
Reporting should also distinguish the expenditure, implementation, and outcome stages of a project. Expenditure documents that funds were committed. Implementation documents that equipment or a production practice was put into use. Outcomes establish what happened to energy demand, emissions, output quality, or other specified targets after implementation. This separation is relevant to the negative contemporaneous R&D association: current spending is an input, whereas a completed low-carbon operating change is an outcome. A review schedule aligned with those stages can maintain financial accountability without requiring an immediate environmental return from every research expense.
Workforce monitoring should connect training to operating tasks such as maintaining energy-efficient equipment, interpreting process data, reducing material losses, and applying consistent greenhouse gas accounting rules. Program evaluation should record whether trained employees perform the intended tasks and whether implementation problems decline. This makes the human-capital implication assessable through operational outcomes, extending the insight provided by workforce educational composition.
The interpretation of government support should follow the same discipline. Aggregate subsidy intensity combines programs with different objectives, so future monitoring should identify the program, recipient, project purpose, payment schedule, and intended environmental outcome. This would permit an assessment of whether support reaches firms facing a credible financing or implementation constraint. The current coefficient on total subsidies cannot answer that question. Differentiated policy design is justified by the heterogeneity of the resource-allocation problem, while a claim that a particular subsidy instrument works would require a separate evaluation design.
7. Conclusions and Research Limits
This study examined innovation-related low-carbon efficiency in 14 listed manufacturing firms in Chongqing during 2022–2025. It combined an undesirable-output EBM benchmark with exact distributional summaries, factor-intensity comparisons, frontier transitions, and exploratory Tobit associations. The main findings concern the uneven distribution of relative performance and its implications for resource allocation.
First, mean efficiency increased from 0.6489 in 2022 to 0.7515 in 2025, after peaking at 0.8094 in 2024. The pooled mean was 0.7259. Six firms remained on the frontier throughout, and frontier observations accounted for 35 of the 56 firm-years. High frontier participation coexisted with a persistent lower tail, so the aggregate mean cannot be taken as evidence of comprehensive low-carbon convergence.
Second, factor-intensity comparisons identified different group trajectories without establishing causal industry effects. The two labor-intensive firms reached a mean of one from 2023 onward. Technology-intensive firms contained both persistent frontier and persistently low-score observations. Capital-intensive firms had the lowest mean in the last three years, but half of that group remained on the frontier throughout. Policy differentiation should therefore combine group characteristics with individual performance histories.
Third, workforce education was positively associated with the latent Tobit outcome at the 10% level, and R&D intensity was negatively associated at the 5% level. Their signs persisted after excluding the last year, although coefficient magnitudes changed. These results support further attention to workforce capabilities and the timing and destination of innovation expenditure. They do not establish a causal benefit from higher educational attainment or a causal loss from R&D investment.
The contribution is a clearer connection between firm-level benchmarking and the allocation problems of a low-carbon manufacturing transition. Relative efficiency, absolute emissions reductions, and causal policy effectiveness are distinct outcomes. Keeping them separate permits a more useful policy interpretation: regional support should address the specific resource-conversion constraints of firms while tracking physical environmental results independently.
The sample is restricted to listed firms with sufficient disclosure, the observation period is short, and linear interpolation may smooth missing observations. The broad indicator set does not distinguish green from other invention applications or separately identify operational and value-chain emissions effects. A small reference sample and numerous inputs and outputs limit the discriminatory power of DEA. The second-stage results use conventional likelihood standard errors and omit firm and year effects; the sample-period comparison does not resolve these limitations. Further research should use a larger harmonized firm panel, report complete environmental accounting boundaries, and compare lagged innovation inputs, alternative frontier specifications, and inference procedures that address generated efficiency scores. These extensions would permit stronger claims about mechanisms and policy effectiveness than the present descriptive assessment supports. The first-stage interpretation also depends on the benchmark reference set and EBM parameter configuration. The reported score panel permits the distributional calculations to be reproduced, while independent reproduction of the frontier and regression requires the underlying firm-level records and complete estimation settings.
Conceptualization, D.M. and Z.Z.; methodology, C.H.; software, C.H.; validation, Z.Z., C.H., and C.Q.Z.; formal analysis, C.Q.Z.; investigation, C.Q.Z.; resources, D.M.; data curation, C.H.; writing—original draft preparation, Z.Z.; writing—review and editing, D.M.; visualization, C.Q.Z.; supervision, D.M.; project administration, Z.Z.; funding acquisition, D.M. All authors have read and agreed to the published version of the manuscript.
This work was supported by the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No.: KJZD-K202501111), titled “Research on the Improvement Paths and Governance Countermeasures of Green Innovation in Chongqing’s Manufacturing Industry under the ‘Dual Carbon’ Goals.”
The data used to support the research findings are available from the corresponding author upon request.
The authors declare no conflicts of interest.
