Green Technology Innovation, Productivity, and ESG Performance: Asymmetric Associations and Empirical Channels in Chinese A-Share Firms

Abstract

This study examines whether green technology innovation is associated differently with firm productivity and ESG performance. Using panel data on Chinese A-share listed firms from 2013 to 2022, we estimate firm and year fixed-effects models and compare within-firm and pooled associations. In the primary corrected revenue-mode specification, green patent applications are negatively associated with short-run within-firm revenue productivity but positively associated with ESG ratings; the productivity result is construct- and specification-bounded. Alternative productivity measures further show that pooled and within-firm specifications capture materially different relationships. Firm age strengthens the ESG association more than the productivity association, whereas temporally ordered tests do not support environmental attention as a mediating channel, and the asset-turnover diagnostics are null. These findings qualify the conventional win–win view of green innovation and highlight the importance of distinguishing external rating responsiveness from internal operational outcomes. The productivity measure is revenue-based and does not represent physical efficiency.

1. Introduction

The transition to a sustainable economy depends critically on the ability of firms to translate environmental investments into both economic value and social legitimacy [1,2]. Among the various levers available to managers, green technology innovation (GTI)—defined as the development of patented technologies that reduce environmental burdens during production, usage, or disposal [3,4]—is widely promoted as a “win–win” instrument capable of simultaneously improving competitive positions and stakeholder evaluations [5,6]. Empirical research has separately examined the drivers and adoption of eco-innovation [7,8] and its financial consequences [9]. Yet most of this literature treats green-innovation outcomes as a homogeneous bundle, implicitly assuming that the same underlying mechanisms—for example, regulatory pressure, resource reallocation, or stakeholder signaling—drive the various dimensions of firm performance with comparable intensity.

This homogeneity assumption is increasingly difficult to reconcile with three streams of evidence. First, the 2013 and 2014 meta-analyses of environmental and financial performance document substantial effect-size heterogeneity across outcome domains [10,11]. Second, research on ESG measurement shows that ratings depend jointly on observable firm actions, disclosure, stakeholder attention, and provider methodology [12,13]. Recent evidence further links ESG practice to sustainable-business-model innovation [14] and green innovation to reduced greenwashing [15], underscoring that visible sustainability activity can have both substantive and evaluative consequences. Third, research on innovation returns emphasizes that measured productivity gains can require absorptive capacity and complementary investment [16,17,18]. Recent firm-level work likewise found that media conditions and green innovation jointly shape corporate productivity [19]. These strands motivate a direct comparison of the two associations, but they do not justify assuming that either represents a causal mechanism.

Failure to separate these associations can lead investors, managers, and policymakers to read a rating change as evidence of contemporaneous operating-efficiency improvement, or to assume that visible innovation and productivity respond on the same timetable. Existing studies have rarely compared productivity and ESG outcomes within the same empirical framework. A common design is therefore needed to test, rather than presume, whether the two associations have similar magnitudes and survive the same identification checks.

Despite the rapid expansion of research on green innovation, relatively little is known about whether its productivity and ESG rating associations survive the same identification tests. This distinction matters because productivity and ESG ratings are generated through different processes. Productivity indices are intended to capture internal productive capability, which depends on complementary assets, production routines, and absorptive capacity; the revenue-based index used here is not a direct physical-efficiency measure. ESG ratings are external evaluations that can respond to observable environmental actions and disclosure while also reflecting provider-specific methodology. Treating the two outcomes as interchangeable may therefore obscure whether an observed coefficient represents within-firm change, persistent between-firm heterogeneity, or rating responsiveness.

We address this gap using a main panel of 2933 Chinese A-share listed firms over 2013–2022 (23,047 firm-year observations). China provides an informative setting because environmental regulation, green innovation, and national carbon-policy salience have intensified together [20,21,22]. For the present revision, we obtained a Mark Data package containing provider-supplied listed-company financial statement data and construction code and independently reconstructed LP, OP, and OLS productivity indices. Because the supplied LP command used the software’s default value-added mode with revenue output, we also executed the documented revenue option. This corrected pooled revenue-mode LP series replaces both the legacy variable and the direct default-mode reconstruction in every primary productivity analysis; the default-mode, OP, and OLS variants are retained only as benchmarks or sensitivity checks, while the industry-specific variant is a partial diagnostic. No authoritative deflator was available. ESG is measured with the Sino-Securities Index Huazheng rating [23]. Firm and year fixed effects form the principal specification. Industry-by-year and province-by-year fixed effects absorb common sectoral and regional policy shocks; a corrected IV analysis is retained only in Appendix C because instrument relevance and exclusion remain inadequate for causal interpretation. A matched comparison around the 2020 carbon neutrality announcement is likewise framed as a policy-salience heterogeneity test rather than a quasi-experiment.

This study makes three contributions. First, it compares revenue productivity and ESG ratings within a common identification hierarchy and shows that the primary within-firm associations have opposite signs rather than a uniform win–win pattern. Second, it makes the pooled-to-firm-FE estimand transition itself analytically informative: direct within–between models and OP/OLS/FE sensitivity tests distinguish short-run changes within a firm from persistent differences across firms, instead of treating a pooled coefficient as within-firm evidence. Third, it formally compares standardized outcome slopes and age interactions while auditing construct and channel evidence. Correcting the LP software mode changes the productivity conclusion, the earlier same-year EnvAtt result does not survive temporal ordering, and the ATO term remains null in a strict revenue-and-asset-overlap diagnostic. These contributions qualify the conventional win–win view without claiming that green patents causally change either outcome.

The primary research question is therefore: are one-year-lagged green patent applications associated with within-firm changes in productivity and Huazheng ESG ratings with comparable magnitude? A complementary, non-directional question asks how the GTI–productivity association differs between pooled and within-firm estimands and across reconstructed productivity indices. Specifically, we (i) estimated both outcomes under firm and year fixed effects; (ii) added industry-by-year and province-by-year shocks; (iii) compared standardized slopes and age interactions formally; (iv) estimated variance components and within–between decomposition slopes; and (v) reassessed proposed channels using temporal ordering and accounting-overlap diagnostics. These objectives address a gap in studies that estimate one outcome at a time or label a pooled association as a within-firm mechanism without establishing where the identifying variation resides.

Overall, five findings emerged. First, under the primary firm- and year-fixed-effects design, one-year-lagged green patent applications are negatively associated with the corrected nominal revenue-productivity index (β = −0.0036, 95% CI [−0.0070, −0.0002]) but positively associated with Huazheng ESG ratings (β = 0.0441, 95% CI [0.0183, 0.0699]); accordingly, H1a is not supported and H1b is supported only in the primary application-based one-year-lag model. Second, the standardized ESG slope exceeds the productivity slope by 0.0508 outcome SD per one pooled-sample SD of GTI (p < 0.001). Third, the within–between decomposition and alternative productivity estimators show that pooled and within-firm specifications capture materially different relationships. Fourth, the standardized GTI-by-age interaction is larger for ESG than for productivity (difference = 0.0287, p = 0.008), supporting H3. Fifth, the temporally ordered channel tests do not support H2 or an ATO mechanism, while the industry and policy-salience results remain exploratory. Taken together, the evidence supports construct- and estimand-bounded asymmetric associations rather than causal pathways.

The remainder of this paper is organized as follows. Section 2 develops the theoretical framework and hypotheses. Section 3 describes the materials and methods. Section 4 reports the baseline, robustness, channel, moderation, heterogeneity, and policy-salience results. Section 5 discusses the findings, implications, and limitations. Section 6 provides the conclusions.

2. Theoretical Framework and Hypotheses

2.1. Green Technology Innovation as a Sustainability Lever

The economics-of-innovation tradition treats environmental innovation as a special case of directed technical change: when firms face explicit or implicit pricing of environmental externalities, R&D effort reallocates toward technologies that abate emissions, save resources, or substitute cleaner inputs [24,25]. Patent filings—particularly those classified under the World Intellectual Property Organization’s “Green Inventory”—provide a measurable, legally bounded indicator of such directed effort [26,27]. In the strategic management tradition, the natural-resource-based view (NRBV) recasts environmental capability as a source of sustainable competitive advantage when it is rare, valuable, and difficult to imitate [2,28]. Together, these traditions suggest that GTI should improve firm performance, but they leave unspecified the relative magnitudes and mediating channels through which different performance dimensions are affected.

2.2. Direct Associations: GTI with Productivity and ESG

Productivity association: resource-based logic. Environmental practices may reduce material, process, and regulatory compliance costs [29,30]. The resource-based view is the principal theory [2,28,31]: complementary assets and routines condition whether an invention becomes productive capability, absorptive capacity supports implementation [16], and complementary investment can delay measured productivity gains [18]. The competing selection explanation is that firms with durable capabilities may both patent more and remain more productive. A pooled positive coefficient can therefore reflect productive integration, capability-based selection, or both; firm fixed effects test whether a short-run within-firm association remains. Thus,

H1a. 

Green technology innovation is positively associated with subsequent firm productivity.

ESG rating association: legitimacy logic. Legitimacy theory is the principal account: visible environmental action can support external acceptance [32], and we propose patent signaling as an observable implication by analogy to general signaling theory [33]. ESG ratings vary with provider scope and measurement choices [12,13], while the current Sino-Securities methodology draws on public information and company reports [23]. These features create an evaluative link that can operate even when a patent has not yet altered production efficiency. Because the outcome is a provider rating rather than independently measured environmental impact, the prediction concerns rating responsiveness:

H1b. 

Green technology innovation is positively associated with subsequent ESG performance.

These theories generate distinct empirical predictions. The productivity coefficient should attenuate if stable capabilities account for the cross-firm pattern. The ESG rating coefficient may survive firm fixed effects if evaluators respond to changes in visible green activity. We therefore compared the standardized coefficients formally under a common firm- and year-fixed-effects system rather than inferring asymmetry from coefficients measured on different native scales. We also asked, without imposing a post hoc directional hypothesis, whether pooled and within-firm productivity associations differ across reconstructed LP, OP, OLS, and FE indices.

2.3. Environmental Attention and Accounting-Overlap Diagnostics

The earlier version of this study treated environmental attention and asset turnover as parallel mediators. The revised framework separates a temporally testable ESG hypothesis from a productivity-side accounting diagnostic. Environmental-disclosure attention is theoretically relevant because it is observable to stakeholders and rating providers. Asset turnover, by contrast, shares operating revenue and asset-scale components with the LP productivity measure; it cannot be treated as an independent mechanism without first addressing that mechanical overlap.

2.3.1. ESG-Side Mediator (Environmental-Disclosure Attention)

Environmental attention (EnvAtt). Cognitive accounts of corporate sustainability emphasize that attention to environmental issues can precede environmental action [34,35]. We observe disclosure intensity, not managerial cognition: EnvAtt is the environment-related word count in the annual MD&A, with the archived variable attributed to Wu and Hua [36], but without the underlying dictionary or text-processing materials. We propose that a filing in year t − 1 may be followed by richer environmental disclosure in year t, which may then be reflected in the rating observed in year t + 1. This temporal sequence is stricter than the same-year covariance decomposition used in the previous version, but the available sources do not establish it as a validated mechanism.

H2. 

Environmental attention positively mediates the association between prior green technology innovation and subsequent ESG performance.

H2 was tested as a temporally ordered statistical indirect association rather than causal mediation.

2.3.2. Asset Turnover as an Accounting-Overlap Diagnostic

Asset turnover is defined neutrally as operating revenue divided by total assets. The LP production-function output is also revenue-based, and the resulting productivity index remains highly correlated with asset size. The ATO and TFP_LP therefore share accounting components. We retained the ATO only as a diagnostic: in addition to the raw index, we used an overlap-residualized outcome obtained by projecting TFP_LP on log total assets and reconstructed log operating revenue (Size + ln ATO), with year and industry effects. This is a strict accounting diagnostic, not a new production-function estimate. We did not state an ATO hypothesis, calculate a mediation proportion, or interpret the term as an independent economic mechanism.

2.3.3. Unobserved Alternatives

Earlier checks of management cost, operating cash flow, and institutional ownership remain reported in Appendix B.4 for transparency, but null results for selected variables cannot establish that the ATO is dominant. Unobserved channels such as patent quality, implementation expenditure, supply-chain redesign, energy-intensity changes, and workforce adaptation remain plausible. The revised analysis therefore treats all channel evidence as exploratory.

2.4. Moderation Hypothesis: Firm Age

We developed two competing, study-specific predictions for firm age from broader theory rather than treating either as an established empirical fact. General legitimacy theory [32] suggests that accumulated relational capital and regulatory familiarity could make a new green patent a credible extension of an established commitment in external evaluation. Conversely, legacy routines and production assets may delay the conversion of a new patent into productive capability, while accumulated environmental liabilities could also lead evaluators to discount a new filing [37]. These arguments predict a stronger age-contingent association for the ESG rating than for revenue productivity. We tested that difference directly after placing both outcomes, GTI, and firm age on a common standardized scale:

H3. 

Firm age strengthens the association between green technology innovation and ESG performance more than it strengthens the corresponding association with revenue productivity.

The parallel industry-intensity contingency is developed in the next subsection as an exploratory boundary-condition analysis.

2.5. Boundary Condition: Industry Type

General technology and productivity theory treats substitution, complementary investment, and organizational environment as potential conditions on innovation returns [18,38,39,40]. Building on that logic, we propose—rather than treat as an established empirical ranking—that capital- and technology-intensive industries may have stronger capital, engineering, and skilled-labor complements for integrating green patents, whereas some labor-intensive settings may face greater financing, workforce, or process-redesign constraints.

Complementarity predicts a stronger GTI–productivity slope in capital- and technology-intensive settings. ESG providers, however, evaluate visible environmental action across sectors, so the rating association need not depend on the same production complementarities. This is an evaluative-decoupling argument: a rating can change without a contemporaneous change in the production function. We tested the cross-industry slope differences directly in pooled interaction models. The productivity equality test is null, whereas the ESG test yields exploratory evidence of different slopes across the documented classes. Because industry type is observational and was not a pre-specified hypothesis, neither result is interpreted causally.

2.6. Supplementary Policy-Salience Setting: The 2020 Carbon Neutrality Announcement

In September 2020, China announced its dual-carbon goals, committing to peak carbon emissions by 2030 and achieve carbon neutrality by 2060 [22]. The announcement increased environmental-policy salience for Chinese firms, especially in industries covered by the official 2008 environmental verification catalog [41]. Firm-level evidence from Chinese A-share listed companies shows that economic policy uncertainty has an inverted U-shaped association with green-innovation quantity and a positive association with innovation quality [42], reinforcing the need to treat policy context as a boundary condition rather than assume a uniform corporate response. Our comparison is not a treatment effect of GTI: the announcement was nationwide, heavy-pollution status is a differential-exposure marker, and parallel trends must be assessed outcome by outcome. Because most of 2020 predates the September announcement, 2020 is treated as a transition year and 2021–2022 as the full post-announcement period.

2.7. Positioning Relative to ESG-to-Productivity Studies

Recent studies often place ESG upstream of innovation, productivity, or sustainable-business-model change. Ding, Han, and Wang [43] modeled ESG as an antecedent of firm productivity through financing constraints and R&D; Zhang et al. [14] examined ESG-to-business-model innovation; and Ma et al. [15] connected green innovation with reduced greenwashing. Our design differs in analytical ordering and purpose. The GTI is the common observable antecedent, while the revenue-productivity index (TFP_LP) and HZESG are parallel outcomes. The revised contribution is not a new sequential mechanism but a comparison of whether the two outcome associations survive the same within-firm identification and temporal tests.

A related earlier article by Shen and Kim [44] used a substantially overlapping 2013–2022 Chinese A-share sample to examine green management and green technology innovation in relation to ESG performance, including contemporaneous environmental-attention mediation and R&D-intensity moderation. The present study is not a translation or republication of that article. It adds revenue productivity as a parallel outcome, reconstructs and audits multiple productivity measures, makes firm and year fixed effects the primary estimand, formally compares standardized outcome slopes and age interactions, and re-estimates the environmental-attention pathway using a temporally ordered design. Under the present firm-and-year-fixed-effects and temporal specification, the environmental-attention indirect association is not supported at the 5% level. The overlapping ESG-side sample and constructs are therefore treated explicitly as prior related work.

Figure 1 summarizes the conceptual framework and distinguishes the proposed ESG-side channel from the productivity-side accounting diagnostic.

Figure 1. Theoretical framework for asymmetric GTI associations with revenue productivity and ESG ratings. Solid arrows denote direct association paths. Green dashed arrows denote the proposed temporal channel test (H2); amber dashed arrows denote the accounting-overlap diagnostic; and gray dashed connectors denote firm-age moderation of the GTI–outcome slopes (H3). Green solid paths denote the external-evaluation branch, and blue solid paths denote the internal-efficiency branch. Industry type and the 2020 announcement are exploratory boundary contexts.

3. Materials and Methods

3.1. Sample and Data Sources

The source panel comprises Chinese A-share listed firms over 2013–2022. We constructed the panel by merging four data sources matched on stock code (six-digit, zero-padded) and year:

Firm financials, governance, and patent counts from the China Stock Market and Accounting Research (CSMAR) database, supplemented by the Chinese Research Data Services (CNRDS) green patent module.

ESG ratings from the Sino-Securities Index (Huazheng ESG (Shanghai, China), henceforth HZESG), which provides nine-grade ratings (AAA → C) coded numerically. The current provider methodology is cited for provider identity and rating structure, not as proof that every historical year used the current-version rules [23].

Productivity-construction inputs and code from a Mark Data package containing provider-supplied listed-company financial statement data. The provider’s public page describes annual reports as its source and does not establish that this package is CSMAR-derived [45]. We reconstructed LP, OP, and OLS indices during the present revision; the reconstructed LP series is primary, and reconstructed OP and OLS indices are sensitivity outcomes [46,47,48,49,50,51]. The legacy provider series is retained only for validation comparison.

Environmental attention scores (EnvAtt) from textual analysis of management discussion and analysis (MD&A) sections in annual reports. The archived documentation attributes the environmental-keyword dictionary underlying EnvAtt to Wu and Hua [36]; the dictionary and source-level text-processing materials are unavailable for independent reconstruction.

At the data-construction stage, we excluded firms that were already designated special-treatment (ST/PT) and firms in financial or real estate industries, required complete corrected revenue productivity and green patent applications, required at least three valid years, and winsorized continuous variables at the 1st and 99th percentiles. An audit identified that an earlier internal build had also required nonmissing GreenMI even though GreenMI appears in no retained hypothesis, regression, table, or conclusion. We therefore removed that analytically unnecessary condition and rebuilt every variable, exact lag, model, table, and figure from the source. This correction prevented sample selection on an unused variable; GreenMI was not used in panel inclusion, winsorization, lag construction, or estimation. Of 24,644 source firm-years, 24,300 had complete corrected productivity and GreenApply; the three-valid-year rule yielded the final 2013–2022 panel of 23,047 observations for 2933 firms. Exact calendar-year lag matching and complete-case control requirements yielded 19,708 observations for the primary productivity model and 19,696 for HZESG, representing 2932 firms; the 12-observation difference reflects ESG missingness. A sensitivity analysis reported in Section 4.2 shows that ending the window in 2021 preserved both primary conclusions. A stricter firm-level exclusion removed every observation for any firm entering ST/PT status during the observed period; under that definition, H1a remains unsupported, but the negative productivity coefficient’s confidence interval spans zero, while the ESG coefficient remains positive. The construction-stage screen and the stricter sensitivity address different criteria: the former removes firms already designated ST/PT, whereas the latter removes the complete panel history of firms entering ST/PT during the study window. LP estimation-sample eligibility is a separate observation-level production-function criterion assessed in the robustness analysis and Appendix A.

Table 1 summarizes the sample-construction flow.

Table 1. Sample construction for the GreenMI-free 2013–2022 main panel.

Step	Firm-Year Observations
Raw merged firm-year source	24,644
Restricted to 2013–2022	24,644
Complete corrected revenue productivity and GreenApply	24,300
Firms with at least three valid years	23,047
Continuous variables winsorized at 1%/99%	23,047
Final 2013–2022 main panel	23,047

Note: GreenMI is not part of a retained hypothesis or model and is not used as an inclusion condition. Complete corrected revenue productivity and GreenApply are required before the three-valid-year rule.

The sample and the ESG-side research design build on Chapter 4 of the first author’s doctoral dissertation [52]. The present study extends that work with the productivity outcome, a common firm-fixed-effects hierarchy, standardized cross-outcome tests, within/between variance decomposition, temporal channel diagnostics, industry comparisons, and the policy-salience analysis.

3.2. Variables

Dependent variables. The LP-based nominal revenue-productivity index (TFP_LP) is the corrected revision-stage reconstruction of the Mark Data LP inputs. Following Levinsohn and Petrin [46] and the Stata/MP 18 implementation documented by Petrin, Poi, and Levinsohn [48], the method uses intermediate input as a proxy for unobserved productivity. The provider describes its five-method listed-company dataset as following the Chinese firm-level comparison by Lu and Lian [49], while the exact variable definitions used here are those in the supplied workbook and code [45]. Output is the operating revenue divided by 10,000; capital is the net fixed assets divided by 10,000; labor is the employee count; and proxy intermediate input is the operating cost plus selling, administrative, and financial expenses, less depreciation/amortization, and cash paid to employees, divided by 10,000. The supplied code applies sample restrictions, transforms inputs as ln(x + 1), winsorizes log inputs annually at 1%/99%, and estimates one pooled eligible-firm production function. It uses no price deflator and omits levpet’s revenue option even though the dependent variable is the operating revenue. We first reproduced that default-mode benchmark and then executed the methodologically aligned pooled specification in Stata/MP 18 with levpet 1.0.1 and the documented revenue option. The revenue-mode re-estimation returned finite coefficients, and the resulting series was mapped to the rebuilt study panel. Appendix A reports estimation-sample eligibility, the command-generated prediction scope, the group-count reconciliation, and the e(sample)-restricted sensitivity. This corrected revenue-mode series is the primary productivity outcome; the provider default-mode reconstruction is retained only in Appendix Table A1. An industry-specific revenue-mode diagnostic produced usable estimates for too few industries and only 7837 matched study-panel observations, so it is reported as a partial diagnostic rather than an alternative primary measure. No deflated variant was estimated because the archive contains no authoritative industry-year deflator series or documented concordance linking such deflators to the analytical inputs. As Van Beveren [50] emphasizes, firm-level productivity estimates depend on input and output definitions, deflation, sample construction, and implementation; the primary index is therefore not interpreted as a physical TFP or validated scale-free technical-efficiency residual. Reconstructed OP and OLS provider-specification indices are reported only as sensitivity outcomes; the OP check follows Olley and Pakes [47] and its Stata implementation [51]. Some Chinese firm-level studies operationalize high-quality enterprise development using TFP [53]. In this study, however, the revenue-productivity index is treated only as a productivity-related outcome and is not equated with the broader policy concept. Appendix A documents the full decision. ESG (HZESG) is the Sino-Securities Index Huazheng rating. The archived annual series maps AAA through C to 9 through 1 and stores a firm-year mean; the source archive does not contain the quarterly transformation and missing-quarter log. It is a provider assessment, not a direct measure of environmental impact. Because the licensed extract lacked separate E, S, and G pillar scores and an alternative-provider series, we could not perform a pillar-level or provider-comparison test. In the sample, HZESG is effectively continuous on [1.25, 6.25] (mean 4.09, SD 1.05). The supplementary ordered logit rounds the annual mean to the nearest integer grade and estimates a different, pooled ordinal-link benchmark rather than the primary within-firm estimand.

Main independent variable. The GTI is measured as the natural log of one plus the number of green-invention and green-utility-model patent applications filed in year t, classified per the WIPO Green Inventory IPC codes [27]; the variable GreenApply carries this logged quantity throughout the data and tables (Table 2 and Table 3). For robustness analyses, we additionally used GreenGrant, defined analogously as the log of one plus the count of granted green patents in year t; using grants rather than applications captures the quality-filtered subset of green innovation that has passed substantive review. Because grants are recorded in the year of authorization, a firm-year’s granted count aggregates applications filed over multiple preceding years, which is why mean GreenGrant exceeds mean GreenApply in Table 2. Green-patent application counts capture codified, publicly observable green-innovation activity and commitment rather than commercialized output; in particular, they do not necessarily measure implementation, commercialization, or realized environmental impact. After acknowledging this as a measurement-scope choice, OECD guidance was used to bound patent-indicator interpretation [26] and granted patents (GreenGrant) were used as a robustness check.

Table 2. Descriptive statistics for the 2013–2022 main panel.

Variable	N	Mean	SD	Min	p25	Median	p75	Max
TFP_LP	23,047	0.867	0.245	0.417	0.721	0.813	0.943	1.894
GTI	23,047	0.495	0.895	0.0	0.0	0.0	0.693	3.829
GTI-grant	23,047	0.84	1.097	0.0	0.0	0.0	1.386	4.506
ESG	23,028	4.09	1.054	1.25	3.5	4.0	5.0	6.25
Size	23,047	22.245	1.246	20.026	21.36	22.07	22.932	26.213
Lev	23,047	0.414	0.195	0.062	0.259	0.407	0.554	0.898
ROE	23,047	0.057	0.141	−0.699	0.026	0.068	0.118	0.365
TobinQ	23,047	2.15	1.387	0.849	1.29	1.712	2.471	8.806
Mfee	23,047	0.087	0.065	0.01	0.045	0.071	0.109	0.377
ATO	23,047	0.623	0.364	0.098	0.383	0.552	0.769	2.234
Growth	23,047	0.158	0.36	−0.496	−0.027	0.103	0.258	2.104
Board	23,047	2.112	0.195	1.609	1.946	2.197	2.197	2.639

Note: Continuous variables were winsorized at the 1st and 99th percentiles. N differs for ESG because of 19 missing firm-year ratings.

Table 3. Correlation matrix and variance inflation factors.

Variable	1	2	3	4	5	6	7	8	9	10	11	12	VIF
1. TFP_LP	1.0	0.036	−0.003	0.193	−0.002	−0.369	0.458	0.193	0.072	−0.182	0.16	−0.046	—
2. GTI	0.01	1.0	0.529	0.213	0.226	0.151	0.075	−0.107	−0.116	0.079	0.033	0.048	1.11
3. GTI-grant	−0.04	0.612	1.0	0.195	0.424	0.278	0.051	−0.234	−0.183	0.051	0.032	0.09	—
4. ESG	0.162	0.214	0.211	1.0	0.187	−0.089	0.248	−0.076	−0.126	0.088	0.079	0.036	—
5. Size	−0.006	0.302	0.512	0.225	1.0	0.474	0.144	−0.518	−0.407	0.093	0.038	0.25	1.74
6. Lev	−0.359	0.161	0.287	−0.114	0.464	1.0	−0.129	−0.376	−0.291	0.147	−0.01	0.129	1.54
7. ROE	0.337	0.071	0.05	0.248	0.126	−0.247	1.0	0.191	−0.232	0.35	0.388	0.025	1.41
8. TobinQ	0.215	−0.096	−0.186	−0.075	−0.367	−0.27	0.108	1.0	0.327	0.006	0.12	−0.155	1.27
9. Mfee	0.071	−0.111	−0.176	−0.16	−0.36	−0.233	−0.247	0.274	1.0	−0.551	−0.124	−0.087	1.59
10. ATO	−0.175	0.058	0.035	0.06	0.099	0.141	0.233	−0.003	−0.469	1.0	0.232	0.02	1.37
11. Growth	0.098	0.004	0.017	0.016	0.039	0.006	0.282	0.076	−0.115	0.161	1.0	−0.011	1.11
12. Board	−0.036	0.068	0.105	0.047	0.267	0.136	0.043	−0.127	−0.097	0.01	−0.002	1.0	1.08

Note: Pearson correlations are below the diagonal and Spearman correlations above it. VIFs use the focal GTI regressor and eight baseline controls only; the outcomes and grant variant are omitted.

Channel variables. EnvAtt is the archived natural log of one plus the count of environment-related keywords in the firm’s annual MD&A. The archived documentation attributes the environmental-keyword dictionary underlying EnvAtt to Wu and Hua [36]; the dictionary and source-level text-processing materials are unavailable for independent reconstruction. It is a textual proxy for environmental-disclosure attention, not managerial cognition. Because the current archive lacks the Chinese dictionary, text-processing script, MD&A corpus, and document-length measure, the temporal EnvAtt test is explicitly exploratory and cannot assess length normalization or alternative missing-text rules. The ATO is the operating revenue divided by total assets. Because both the ATO and reconstructed LP index embed revenue and scale information, the ATO was excluded from the relevant control vector and used only in an accounting-overlap diagnostic. The strict variant residualizes TFP_LP on log assets and reconstructed log revenue before estimating the outcome equation.

Moderators and historical diagnostics. FirmAge is the natural log of one plus years since founding and is the pre-specified moderator in the revised analysis. GTI and FirmAge were mean-centered before their interaction was formed. With firm and year fixed effects, the age main effect is not a cross-sectional age comparison; the interaction is identified from whether the within-firm GTI slope changes as logged age evolves differentially across firms. INST is the share of equity held by institutional investors, and Intang is the ratio of intangible assets to total assets; their earlier moderation and contemporaneous-channel specifications are retained only in the archived benchmark tables.

Controls. The baseline control vector includes eight variables: firm size (Size, log of total assets), leverage (Lev), return on equity (ROE), Tobin’s Q, management-fee ratio (Mfee, administrative expenses over operating revenue), asset-turnover ratio (ATO; excluded from the outcome equation when the ATO is evaluated as a channel diagnostic), revenue growth (Growth, year-on-year operating revenue growth rate), and board size (Board, the natural log of the number of directors). The cash-flow ratio (Cashflow, net operating cash flow over total assets) appears only in the historical alternative-channel diagnostics reported in Appendix B.4.

Heterogeneity variables. HighTech is a binary indicator for high-technology firms following Ministry of Science and Technology certification. IndType is a source-based annual three-class industry-intensity variable (1 = labor-intensive, 2 = asset-/capital-intensive, and 3 = technology-intensive). The classification follows Yin, Sheng, and Li’s SPSS 19.0 Ward hierarchical-cluster solution [54], which uses the fixed-asset ratio (net fixed assets divided by average total assets) and the R&D-to-payroll ratio (R&D expenditure divided by employee compensation payable) and reports a mutually exclusive 67-code crosswalk under the 2012 CSRC industry system [55]. We merged the source record by stock code and year and used its contemporaneous annual industry code and class; thus, a firm is reclassified when its source-record industry code moves across class boundaries. All 23,047 study-panel firm-years matched the source record. Of these, 22,916 were covered by the published 67-code mapping: 6181 labor-intensive, 4228 asset-/capital-intensive, and 12,507 technology-intensive observations. The remaining 131 firm-years were left unclassified and excluded from the IndType analyses. Appendix B.2 and the retained industry-type classification audit record document the full mapping, source audit, exclusions, and re-estimation results.

Policy-salience variables. Treat is a binary indicator equal to one for firms classified as heavy-pollution by the 2008 Ministry of Environmental Protection list [41]. FullPost equals one in 2021–2022, the full calendar years after the September 2020 announcement. The clean contrast excludes 2020 as a transition year; the event study reports its separate transition-year coefficient. The 2008 catalogue is used here as a fixed historical heavy-pollution industry classification convention; it is not assumed to have remained legally operative throughout the 2013–2022 study window.

Notation. Throughout the text and tables, GTI (t − 1) and GTI (t − 2) denote exact one- and two-calendar-year lags of GTI, the log-transformed green-patent application count (variable GreenApply); GTI-grant denotes the granted-patent variant based on GreenGrant; and (c) marks mean-centered variables in the moderation specifications. Lags were constructed by matching firm i in year t to the same firm’s record in year t − 1 or t − 2. If that exact calendar-year record was absent, the lag was left missing rather than filled from the previous observed row.

Software environment. Productivity reconstruction was performed using Stata/MP 18 (StataCorp LLC, College Station, TX, USA) with levpet 1.0.1, opreg 2.0.0, and winsor2 1.1. Downstream fixed-effects, within–between, cross-equation, matching, wild-bootstrap, figure, and table routines were conducted using Python 3.13.5 (Python Software Foundation, Beaverton, OR, USA) with NumPy 2.3.5, pandas 2.3.3, SciPy 1.17.1, linearmodels 7.0, and scikit-learn 1.8.0, as recorded in the retained execution environment.

Generative artificial intelligence use. OpenAI Codex (OpenAI, San Francisco, CA, USA; https://openai.com/zh-Hans-CN/ accessed on 17 August 2026) was used to assist with code review, reproducibility checks, numerical and structural consistency review, reference verification, English-language editing, and document formatting. It was not used to create or alter source data. All code, numerical outputs, modeling decisions, interpretations, and final text were independently reviewed and approved by the authors.

Table 2 reports descriptive statistics for the main panel, and Table 3 presents the correlation matrix and variance-inflation-factor diagnostics.

3.3. Empirical Specification

Our primary specification is:

$$\mathrm{Y}(\mathrm{i},\mathrm{t})=\mathsf{\alpha }\left(\mathrm{i}\right)+\mathsf{\mu }\left(\mathrm{t}\right)+\mathsf{\beta }·\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i},\mathrm{t}-1)+\mathsf{\lambda }·\mathrm{X}(\mathrm{i},\mathrm{t})+\mathsf{\epsilon }(\mathrm{i},\mathrm{t})$$

(1)

where Y is either the reconstructed productivity index (TFP_LP) or HZESG, α_i and μ_t are the firm and year fixed effects, respectively, and X is the control vector. Standard errors are clustered by firm. The coefficient β is therefore identified from within-firm changes after absorbing economy-wide annual shocks. The exact one-calendar-year lag reduces same-year simultaneity but does not eliminate time-varying omitted variables, so we describe β as a within-firm association rather than a causal effect.

We estimated a fixed-effects hierarchy rather than selecting one specification after observing significance. The hierarchy comprised (i) firm and year fixed effects, our primary model; (ii) firm and industry-by-year fixed effects; and (iii) firm, industry-by-year, and province-by-year fixed effects. The latter models absorb common sector-year and province-year shocks. A pooled year-and-industry-fixed-effects model, which retains both within- and cross-firm variation, is reported only as a descriptive benchmark; it is not a between estimator. For firm-FE specifications, we report the within R² after fixed-effect absorption; for the pooled benchmark, we labeled the corresponding partial R² separately.

We first decomposed the variance of the lagged GTI and both outcomes into exact within- and between-firm sums of squares. Because variance shares do not decompose a regression coefficient, we then estimated a within–between decomposition model with firm means and firm-mean deviations:

$$\begin{gathered} \mathrm{Y}(\mathrm{i},\mathrm{t})=\mathsf{\mu }\left(\mathrm{t}\right)+\mathsf{\nu }\left(\mathrm{j}\right)+\mathsf{\beta }\mathrm{W}·\left[\mathrm{G}\mathrm{T}\mathrm{I}\right(\mathrm{i},\mathrm{t}-1)-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i}\left)\right] \\ +\mathsf{\beta }\mathrm{B}·\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{G}\mathrm{T}\mathrm{I}\left(\mathrm{i}\right)+\mathsf{\Sigma }\mathrm{k}\{\mathsf{\gamma }\mathrm{W}\mathrm{k}·[\mathrm{X}\mathrm{k}(\mathrm{i},\mathrm{t})-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{X}\mathrm{k}\left(\mathrm{i}\right)] \\ +\mathsf{\gamma }\mathrm{B}\mathrm{k}·\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{X}\mathrm{k}\left(\mathrm{i}\right)\}+\mathsf{\epsilon }(\mathrm{i},\mathrm{t}) \end{gathered}$$

(2)

This is an absorbed OLS decomposition, not a random-intercept or correlated-random-effects estimator. For each outcome, firm means were calculated on that model’s complete-case estimation sample. One-year-lagged green patent applications and every time-varying control enter jointly as a firm-mean deviation and a firm mean, together with year and industry fixed effects. Standard errors are clustered by firm. The deviation coefficient βW estimates the conditional within-firm association, while the firm-mean coefficient βB estimates the conditional between-firm association. A cluster-robust Wald test was used to evaluate the equality of the two slopes. Neither coefficient is interpreted causally.

Finally, on common estimation samples, we standardized the GTI and both outcomes using their pooled-sample standard deviations, residualized the direct-association equations on firm and year fixed effects, and estimated them as a cluster-robust system. A standardized slope is therefore measured in outcome SDs per one pooled-sample SD increase in the GTI. For H3, we additionally standardized firm age, entered the standardized GTI, age, and their product in both firm-and-year-FE outcome equations, and used the firm-clustered cross-equation covariance to test the equality of the interaction coefficients. An algebraically equivalent regression using the standardized ESG-minus-productivity outcome provided a verification check. These procedures compare slopes and interactions directly rather than inferring a difference from separate significance decisions [56].

The appendix-only instrumental-variable sensitivity specifications instrument GTI (t − 1) with GTI (t − 2), the lagged leave-one-out industry-by-year mean of peer GTI, or both. They include the full control vector and year and industry fixed effects, with standard errors clustered by firm:

$$\begin{gathered} \mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i},\mathrm{t}-1)=\mathsf{\pi }1·\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i},\mathrm{t}-2)+\mathsf{\pi }2·\mathrm{p}\mathrm{e}\mathrm{e}\mathrm{r}\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{j},\mathrm{t}-1) \\ +\mathsf{\pi }·\mathrm{X}(\mathrm{i},\mathrm{t})+\mathsf{\mu }\left(\mathrm{t}\right)+\mathsf{\nu }\left(\mathrm{j}\right)+\mathsf{\eta }(\mathrm{i},\mathrm{t}) \end{gathered}$$

(3)

The persistence instrument is strongly relevant, whereas the lagged peer instrument is weak. Relevance alone would not establish validity in any case: GTI (t − 2) can affect current outcomes through persistent capabilities, while peer GTI can proxy for industry-year demand, regulation, or technology shocks. The joint specification did not reject the overidentifying restrictions, but that diagnostic cannot establish the exclusion restrictions. We therefore retained the IV estimates only as a transparent appendix audit; they do not corroborate the primary result, identify a defensible local average treatment effect, or support causal language.

The channel analysis is temporally ordered:

$$\begin{gathered} \mathrm{S}\mathrm{t}\mathrm{e}\mathrm{p}1:\mathrm{M}(\mathrm{i},\mathrm{t})=\mathrm{a}·\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i},\mathrm{t}-1)+\mathsf{\delta }·\mathrm{X}(\mathrm{i},\mathrm{t}) \\ +\mathsf{\alpha }\left(\mathrm{i}\right)+\mathsf{\mu }\left(\mathrm{t}\right)+\mathrm{u}(\mathrm{i},\mathrm{t}) \\ \mathrm{S}\mathrm{t}\mathrm{e}\mathrm{p}2:\mathrm{Y}(\mathrm{i},\mathrm{t}+1)={\mathrm{c}}^{\mathrm{\prime }}·\mathrm{G}\mathrm{T}\mathrm{I}(\mathrm{i},\mathrm{t}-1)+\mathrm{b}·\mathrm{M}(\mathrm{i},\mathrm{t})+\mathsf{\delta }·\mathrm{X}(\mathrm{i},\mathrm{t}) \\ +\mathsf{\alpha }\left(\mathrm{i}\right)+\mathsf{\mu }(\mathrm{t}+1)+\mathrm{v}(\mathrm{i},\mathrm{t}+1) \\ \mathrm{I}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{a}\mathrm{s}\mathrm{s}\mathrm{o}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{a}·\mathrm{b} \end{gathered}$$

(4)

Both equations use the same path-specific complete-case sample, absorb firm and outcome-year fixed effects, and cluster by firm; confidence intervals for a·b use the delta method and are descriptive rather than causal. Controls are measured at t. Environmental attention (EnvAtt) is evaluated as a pre-specified but exploratory ESG channel. Asset turnover is retained as an accounting-overlap diagnostic, not a mediator hypothesis, because both the ATO and the reconstructed LP productivity series incorporate operating revenue. We therefore report the productivity index (TFP_LP) and an outcome residual obtained by projecting that index on log assets and reconstructed log operating revenue, and we avoided mediation-share calculations.

The policy-salience analysis estimates propensity scores from the eight controls measured in 2019 and applies 1:1 nearest-neighbor matching with replacement within a caliper of 0.05 standard deviations of the logit propensity score. Repeated control draws enter the matched panel through frequency weights. Covariate balance was assessed using weighted standardized mean differences (|SMD| < 0.10 [57]). The clean full-post contrast omits 2020 and estimates:

$$\begin{gathered} \mathrm{Y}(\mathrm{i},\mathrm{t})=\mathsf{\alpha }\left(\mathrm{i}\right)+\mathsf{\mu }\left(\mathrm{t}\right)+\mathsf{\beta }·\left[\mathrm{T}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}\right(\mathrm{i})\times \mathrm{F}\mathrm{u}\mathrm{l}\mathrm{l}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{t}(\mathrm{t}\left)\right] \\ +\mathsf{\lambda }·\mathrm{X}(\mathrm{i},\mathrm{t})+\mathsf{\epsilon }(\mathrm{i},\mathrm{t}) \end{gathered}$$

(5)

where Treat denotes heavy-pollution-industry membership and FullPost equals one in 2021–2022. Firm and year fixed effects are included. Because exposure was assigned through industry status, standard errors and confidence intervals clustered on 54 detailed industries in the full sample and 50 detailed industries in the matched sample, and p-values used a 999-draw Webb-weight wild-cluster bootstrap. Because the announcement applied nationwide and Treat is a time-invariant exposure category, the design is a differential policy-salience comparison, not a direct intervention on GTI. Event-study coefficients for 2014–2022 use 2019 as the reference year, display 2020 separately as the transition year, apply the matching frequency weights, and report detailed-industry-clustered 95% confidence intervals and joint pre-trend tests.

4. Results

4.1. Baseline Associations: GTI with Productivity and ESG

Table 4 reports the fixed-effects hierarchy using the corrected pooled revenue-mode index. In the primary firm-and-year-FE model, one-year-lagged green patent applications are negatively associated with the revenue-productivity index (β = −0.0036, SE = 0.0017, p = 0.038; 95% CI [−0.0070, −0.0002]; N = 19,708). Because H1a predicted a positive association, H1a is not supported. The point estimate remains negative but is marginal after absorbing industry-by-year shocks (β = −0.0032, p = 0.057) and both industry-by-year and province-by-year shocks (β = −0.0031, p = 0.070). The pooled year-and-industry-FE benchmark is also negative and statistically null (β = −0.0013, p = 0.657). We therefore describe the primary negative estimate as specification-bounded and associational rather than as evidence of a causal productivity loss.

Table 4. Fixed-effects hierarchy for revenue productivity and ESG ratings.

Outcome	Specification	GTI (t − 1)	95% CI	p-Value	N	Firms	R2 Definition	R2 After FE Absorption
TFP_LP	Firm + year FE (primary)	−0.004 ** (0.002)	[−0.0070, −0.0002]	0.038	19,708	2932	Within R2	0.247
TFP_LP	Firm + industry × year FE	−0.003 * (0.002)	[−0.0066, 0.0001]	0.057	19,708	2932	Within R2	0.245
TFP_LP	Firm + industry × year + province × year FE	−0.003 * (0.002)	[−0.0064, 0.0003]	0.070	19,708	2932	Within R2	0.243
TFP_LP	Year + industry FE (pooled benchmark)	−0.001 (0.003)	[−0.0070, 0.0044]	0.657	19,708	2932	Partial R2 after year/industry FE	0.332
ESG (HZESG)	Firm + year FE (primary)	0.044 *** (0.013)	[0.0183, 0.0699]	<0.001	19,696	2932	Within R2	0.056
ESG (HZESG)	Firm + industry × year FE	0.041 *** (0.013)	[0.0162, 0.0660]	0.001	19,696	2932	Within R2	0.057
ESG (HZESG)	Firm + industry × year + province × year FE	0.044 *** (0.013)	[0.0193, 0.0692]	<0.001	19,696	2932	Within R2	0.057
ESG (HZESG)	Year + industry FE (pooled benchmark)	0.170 *** (0.014)	[0.1413, 0.1978]	<0.001	19,696	2932	Partial R2 after year/industry FE	0.175

Note: The corrected pooled revenue-mode index is the primary productivity outcome. Firm/year FE is primary. Within R² is reported after firm-effect absorption; the pooled row instead labels partial R² after year and industry effects. Confidence intervals are reported to four decimals. *** p < 0.01, ** p < 0.05, * p < 0.10.

For Huazheng ESG ratings, the primary one-year-lagged application coefficient is positive and statistically significant (β = 0.0441, SE = 0.0132, p < 0.001; 95% CI [0.0183, 0.0699]; N = 19,696). It remains positive after industry-by-year fixed effects (β = 0.0411, p = 0.001) and after adding province-by-year fixed effects (β = 0.0442, p < 0.001). The pooled year-and-industry-FE benchmark is larger (β = 0.1695, p < 0.001), consistent with the added cross-firm variation but not decomposing its coefficient by source. H1b is supported in the primary application-based one-year-lag model, but later robustness checks show that this association is not invariant to the granted-patent proxy or the two-period lag.

Panel A of Table 5 reports the exact sums-of-squares decomposition: 74.7% of GTI variance, 76.4% of corrected revenue-productivity variance, and 64.6% of ESG variance is between firms. Panel B estimates the coefficient decomposition that variance shares alone cannot provide. For the productivity index, the within–between model estimates a negative within slope of −0.0051 (SE = 0.0018, p = 0.004; 95% CI [−0.0087, −0.0016]) and a statistically null between slope of −0.0016 (SE = 0.0040, p = 0.685; 95% CI [−0.0095, 0.0062]); their difference is 0.0035 (SE = 0.0045, p = 0.432). For ESG, the within slope is 0.0375 (SE = 0.0131, p = 0.004; 95% CI [0.012, 0.063]), the between slope is 0.2088 (SE = 0.0201, p < 0.001; 95% CI [0.169, 0.248]), and the difference is 0.1713 (SE = 0.0239, p < 0.001). These terms remain conditional observational associations; the productivity result does not establish a causal within-firm loss, and the larger ESG between slope does not establish selection by stable capabilities.

Table 5. Exact variance components and within–between decomposition estimates.

Panel	Variable/Outcome	Between Variance Share	Within GTI Slope	Between GTI Slope	Between − Within	Difference p	N	Firms
A. Exact variance decomposition	GTI (t − 1)	74.7%	—	—	—	—	19,708	2932
A. Exact variance decomposition	TFP_LP	76.4%	—	—	—	—	23,047	2933
A. Exact variance decomposition	ESG (HZESG)	64.6%	—	—	—	—	23,028	2933
B. Within–between decomposition model	TFP_LP	76.4%	−0.005 *** (0.002)	−0.002 (0.004)	0.004 (0.004)	0.432	19,708	2932
B. Within–between decomposition model	ESG (HZESG)	64.6%	0.038 *** (0.013)	0.209 *** (0.020)	0.171 (0.024)	<0.001	19,696	2932

Note: Panel A uses exact sums of squares. Panel B is the absorbed OLS, not a random-intercept or correlated-random-effects model. On each model-specific complete-case sample, the GTI and every time-varying control enter as a firm-mean deviation and a firm mean; year and industry fixed effects are absorbed, and inference clusters by firm. Difference p tests the equality of the GTI slopes. *** p < 0.01.

Table 6 formally compares the within-firm standardized slopes on a common sample using a cluster-robust system. The revenue-productivity slope is −0.0135 (SE = 0.0065, p = 0.038; 95% CI [−0.0262, −0.0007]), whereas the ESG slope is 0.0373 (SE = 0.0112, p < 0.001; 95% CI [0.0155, 0.0592]). Their difference is 0.0508 (cluster-robust SE = 0.0126, p < 0.001; 95% CI [0.0261, 0.0755]; N = 19,696; 2932 firms). H1a was decided from the native-scale primary model, not from this standardized comparison. On the common complete-case sample used for the standardized comparison, one pooled-sample SD of GTI is 0.909; the corresponding full-panel SD reported in Table 2 is 0.895. On the native scale, that common-sample SD corresponds to 0.040 HZESG points (95% CI [0.017, 0.063]), or about 0.037 of the common-sample HZESG SD. These are associational magnitudes, not causal effects.

Table 6. Cluster-robust comparison of standardized GTI slopes.

Comparison	Estimate	SE	95% CI	p-Value	Observations	Firms
TFP standardized slope	−0.013	0.006	[−0.026, −0.001]	0.038	19,696	2932
ESG standardized slope	0.037	0.011	[0.015, 0.059]	<0.001	19,696	2932
Difference (ESG − TFP)	0.051	0.013	[0.026, 0.076]	<0.001	19,696	2932

Note: The outcomes and GTI use pooled-sample standard deviations on one common complete-case sample. The GTI SD used for standardization is 0.909 on this common sample; Table 2 reports 0.895 for the full 23,047-observation panel. The system residualizes on firm and year effects and uses firm-clustered SUR covariance. H1a was decided from the native-scale primary model, not this comparison.

Because Huazheng ratings are ordinal in origin, Table 7 reports an ordered-logit benchmark for the discretized 1–9 grade. The firm-clustered coefficient is positive (log-odds β = 0.341, SE = 0.028, p < 0.001, N = 19,696). This secondary year-and-industry-FE result provides sign robustness to an ordinal link; it does not replace the firm-FE linear model used for the primary within-firm inference.

Table 7. Ordered-logit benchmark for the GTI–ESG association.

Term	Estimate	Clustered SE	z	p-Value
GTI (t − 1)	0.341 ***	0.028	12.05	<0.001
Observations	19,696
Outcome	HZESG (discretized 1–9)
Controls/Year FE/Industry FE	Yes

Note: HZESG is discretized to the nearest integer grade (1–9). The provider coding is defined on a 1–9 scale; the realized discretized grades in this sample occupy only a subset of that scale. This secondary model includes full controls and year/industry fixed effects with firm-clustered inference; it is not the primary within-firm model. *** p < 0.01.

4.2. Primary-Estimand Robustness

Table 8 and Table 9 align the sensitivity analyses with the primary firm- and year-fixed-effects estimand. The reconstructed alternative OP (β = −0.0046, p = 0.135) and OLS (β = −0.0009, p = 0.707) indices are statistically null. Appendix Table A2 shows why substituting a pooled model changes the empirical question rather than providing an interchangeable robustness result. In the year-and-industry-FE pooled benchmarks, the OP (β = 0.0139, p = 0.022), OLS (β = 0.0110, p = 0.005), and conventional FE-index productivity (β = 0.0101, p = 0.010) are positive. In the direct within–between models, the OP conditional between slope exceeds its within slope (difference p = 0.005), whereas the OLS difference is marginal (p = 0.056) and the FE-index difference is not significant (p = 0.167). These estimator-dependent contrasts indicate that pooled positive results can be concentrated in persistent cross-firm differences; they do not establish a positive within-firm association. In contrast, the corrected revenue-mode LP coefficient remains negative for granted patents (β = −0.0062, p < 0.001), the two-period application lag (β = −0.0040, p = 0.014), and the exclusion of heavy-pollution observations (β = −0.0043, p = 0.031), but is marginal when 2020 is excluded (p = 0.051) or variables are winsorized at 5%/95% (p = 0.075). Thus, the productivity conclusion is explicitly conditional on both the estimand and the reconstructed index.

Table 8. Firm-and-year-fixed-effects robustness checks for revenue productivity.

Specification	Focal GTI Coefficient	95% CI	p-Value	N	Firms	Within R2
(R1) Outcome = TFP_OP	−0.005 (0.003)	[−0.0107, 0.0014]	0.135	19,708	2932	0.694
(R2) Outcome = TFP_OLS	−0.001 (0.002)	[−0.0056, 0.0038]	0.707	19,708	2932	0.863
(R3) GTI = granted patents	−0.006 *** (0.002)	[−0.0095, −0.0029]	<0.001	19,708	2932	0.248
(R4) GTI (t − 2)	−0.004 ** (0.002)	[−0.0073, −0.0008]	0.014	16,879	2932	0.248
(R5) Exclude 2020	−0.004 * (0.002)	[−0.0073, 0.00002]	0.051	17,040	2930	0.248
(R6) Winsorize model variables 5%/95%	−0.002 * (0.001)	[−0.0051, 0.0002]	0.075	19,708	2932	0.27
(R7) Exclude heavy-pollution observations	−0.004 ** (0.002)	[−0.0082, −0.0004]	0.031	13,680	2600	0.233
(R8) Restrict to LP e(sample)-eligible source rows	−0.0035 ** (0.0017)	[−0.0069, −0.0001]	0.045	19,353	2930	0.251

Note: The corrected revenue-mode LP index was used except where the OP or OLS is named. Rows use the 2013–2022 window, firm/year fixed effects, and firm-clustered standard errors. R6 winsorizes the outcome, focal lag, and controls at 5%/95%. R8 restricts the primary model to source rows eligible for the LP production-function e(sample). Confidence intervals are shown to four decimals; a bound that would otherwise round to zero is shown with one additional digit. *** p < 0.01, ** p < 0.05, * p < 0.10.

Table 9. Firm-and-year-fixed-effects robustness across outcomes.

Outcome	Specification	Focal GTI Coefficient	95% CI	p-Value	N	Firms	Within R2
TFP_LP	(R3) GTI = granted patents	−0.006 *** (0.002)	[−0.0095, −0.0029]	<0.001	19,708	2932	0.248
TFP_LP	(R4) GTI (t − 2)	−0.004 ** (0.002)	[−0.0073, −0.0008]	0.014	16,879	2932	0.248
TFP_LP	(R5) Exclude 2020	−0.004 * (0.002)	[−0.0073, 0.00002]	0.051	17,040	2930	0.248
TFP_LP	(R6) Winsorize model variables 5%/95%	−0.002 * (0.001)	[−0.0051, 0.0002]	0.075	19,708	2932	0.27
ESG (HZESG)	(R3) GTI = granted patents	0.021 * (0.012)	[−0.0030, 0.0440]	0.087	19,696	2932	0.055
ESG (HZESG)	(R4) GTI (t − 2)	0.007 (0.014)	[−0.0205, 0.0344]	0.618	16,872	2932	0.065
ESG (HZESG)	(R5) Exclude 2020	0.045 *** (0.014)	[0.0184, 0.0723]	<0.001	17,028	2930	0.058
ESG (HZESG)	(R6) Winsorize model variables 5%/95%	0.040 *** (0.013)	[0.0159, 0.0650]	0.001	19,696	2932	0.039
Tobin’s Q	(R3) GTI = granted patents	0.062 *** (0.014)	[0.0354, 0.0885]	<0.001	19,708	2932	0.069
Tobin’s Q	(R4) GTI (t − 2)	0.038 *** (0.015)	[0.0091, 0.0664]	0.010	16,879	2932	0.079
Tobin’s Q	(R5) Exclude 2020	0.043 *** (0.015)	[0.0143, 0.0714]	0.003	17,040	2930	0.077
Tobin’s Q	(R6) Winsorize model variables 5%/95%	0.019 * (0.011)	[−0.0023, 0.0400]	0.081	19,708	2932	0.098

Note: Productivity rows use the corrected revenue-mode LP index except where the OP or OLS is explicitly named. Every row uses firm and year fixed effects with firm-clustered standard errors. Confidence intervals are shown to four decimals; a bound that would otherwise round to zero is shown with one additional digit. *** p < 0.01, ** p < 0.05, * p < 0.10.

For ESG, excluding 2020 (β = 0.045, p < 0.001) and 5%/95% winsorization (β = 0.040, p = 0.001) preserved the positive within-firm coefficient. The grant-based coefficient (β = 0.021, p = 0.087) is marginal only at 10%, and the two-period-lag coefficient (β = 0.007, p = 0.618) is statistically null. H1b is therefore supported in the primary application-based one-period-lag model and several specification checks, but it is not invariant to proxy or lag horizon. The supplementary Tobin’s Q results are reported for scope rather than used to support either primary hypothesis.

Table 8 also reports the source-row eligibility audit. Restricting the primary productivity regression to observations eligible for the LP production-function e(sample) removes 355 documented full-panel predictions and leaves the negative coefficient and 5% decision unchanged (β = −0.0035, SE = 0.0017, p = 0.045; 95% CI [−0.0069, −0.0001]; N = 19,353). Table 10 treats 2013–2022 as the main window, shows the corresponding sample ending in 2021, and applies the stricter ever-ST/PT firm exclusion. Ending the window in 2021 leaves the productivity estimate negative (β = −0.0036, p = 0.043) and the ESG estimate positive (β = 0.0419, p = 0.004), so the inclusion of 2022 does not determine either primary hypothesis decision. The ever-ST/PT rule attenuates the productivity coefficient to −0.0029 (p = 0.107), while the ESG coefficient remains positive (β = 0.0412, p = 0.002). H1a remains unsupported under every sample definition, but statistical evidence for a negative productivity association is not invariant to the distress exclusion.

Table 10. Sensitivity to ending the window in 2021 and later ST/PT entry.

Sample	Outcome	GTI (t − 1)	95% CI	Observations	Firms
Main panel: 2013–2022	Productivity index	−0.004 ** (0.002)	[−0.0070, −0.0002]	19,708	2932
Main panel: 2013–2022	ESG (HZESG)	0.044 *** (0.013)	[0.0183, 0.0699]	19,696	2932
Ending in 2021	Productivity index	−0.004 ** (0.002)	[−0.0070, −0.0001]	16,901	2931
Ending in 2021	ESG (HZESG)	0.042 *** (0.014)	[0.0136, 0.0702]	16,889	2931
2013–2022, excluding firms ever entering ST/PT	Productivity index	−0.003 (0.002)	[−0.0064, 0.0006]	17,917	2693
2013–2022, excluding firms ever entering ST/PT	ESG (HZESG)	0.041 *** (0.014)	[0.0147, 0.0677]	17,905	2693

Note: The GreenMI-free 2013–2022 panel is primary. Ending the window in 2021 preserves both hypothesis decisions. The ever-ST/PT rule removes all observations for any firm entering ST/PT during the period. Confidence intervals are reported to four decimals. *** p < 0.01, ** p < 0.05.

Appendix C (Table A3) retains the corrected revenue-mode IV estimates solely as a diagnostic audit. The lagged peer instrument remains weak (first-stage χ² = 0.05), while the persistent-lag instrument is strongly relevant but has an ambiguous exclusion restriction. The persistence-only and combined second-stage estimates are negative but statistically null (β = −0.0038, p = 0.331, and p = 0.335, respectively); the combined overidentification p-value is 0.146. These results do not corroborate a causal interpretation of the primary association and support no local-average-treatment-effect conclusion [58,59,60].

4.3. Temporally Ordered Channel Tests

Table 11 reports the temporally ordered firm-and-year-FE channel tests. For environmental attention, the indirect term for GTI (t − 1) → EnvAtt(t) → ESG(t + 1) is 0.0010 (delta-method 95% CI [−0.0001, 0.0020], p = 0.071; N = 16,350). The interval includes zero at the pre-specified 5% level, so H2 is not supported under temporal ordering and within-firm identification. The estimate is at most marginal evidence and is not interpreted as mediation.

Table 11. Temporally ordered channel tests and ATO diagnostics.

Path	Total c	a: GTI → M	b: M → Y|GTI	Direct c′	Indirect a·b	Delta 95% CI	p-Value	N	Firms
EnvAtt → ESG	0.0142	0.0147	0.0655	0.0132	0.001	[−0.0001, 0.0020]	0.071	16,350	2916
Raw ATO → productivity index	−0.0052	−0.0029	−0.0334	−0.0053	0.0001	[−0.0001, 0.0003]	0.348	16,521	2924
ATO → overlap-residualized productivity index	−0.0058	−0.0064	0.0058	−0.0058	−0.00004	[−0.0002, 0.0001]	0.557	16,521	2924

Note: Models use GTI (t − 1), mediator(t), outcome(t + 1), firm/year fixed effects, and firm-clustered inference. Delta-method intervals describe noncausal indirect associations. The overlap-residualized revenue-productivity outcome is an accounting diagnostic, not a new production-function estimate.

For revenue productivity, the raw-ATO indirect term is 0.00010 (95% CI [−0.00011, 0.00030], p = 0.348). In the stricter diagnostic, the productivity index was first residualized on log assets and reconstructed log operating revenue; the ATO indirect term is likewise null (−0.00004, 95% CI [−0.00016, 0.00009], p = 0.557). Neither result supports an independent ATO mechanism. We make no mediated-percentage claim and do not privilege ATO over unobserved operational, governance, or supply-chain channels.

4.4. Firm-Age Moderation Results

Table 12 presents native-scale firm-and-year-FE moderation models and a formal standardized common-sample comparison. On the native scales, the GTI × FirmAge interaction was imprecisely estimated for the productivity index (β = −0.0030, SE = 0.0052; p = 0.572) and positive for ESG (β = 0.1077, SE = 0.0417; p = 0.0098). Figure 2 plots those outcome-specific marginal associations with 95% confidence intervals. On the common standardized scale, the interaction is −0.0031 for productivity (SE = 0.0055, p = 0.573) and 0.0256 for ESG (SE = 0.0099, p = 0.0098). The direct ESG-minus-productivity difference is 0.0287 (SE = 0.0108, 95% CI [0.0075, 0.0500], p = 0.008; N = 19,696; 2932 firms), and the outcome-difference regression yields the same result. H3 is therefore supported by a direct cross-outcome test, not by comparing separate significance decisions. The interaction remains observational and does not establish causal moderation.

Table 12. Firm-age moderation and standardized cross-outcome interaction test.

Term or Test	Productivity Index	ESG (HZESG)
Panel A. Native-scale firm-and-year-FE models
GTI (centered)	−0.003 ** (0.002)	0.040 *** (0.013)
Firm age (centered)	0.017 (0.037)	−0.646 *** (0.248)
GTI × firm age	−0.003 (0.005)	0.108 *** (0.042)
Controls	Yes	Yes
Firm fixed effects	Yes	Yes
Year fixed effects	Yes	Yes
Observations	19,708	19,696
Firm clusters	2932	2932
Within R2	0.248	0.058
Panel B. Standardized common-sample comparison
Standardized GTI × standardized age	−0.003 (0.006)	0.026 *** (0.010)
ESG − productivity interaction	—	0.0287 *** (SE 0.0108; p = 0.008)

Note: Panel A reports native-scale firm-and-year-FE models with firm-clustered standard errors. Panel B uses one common sample and pooled-sample SD scaling (N = 19,696; 2932 firms). The ESG-minus-productivity interaction difference uses firm-clustered cross-equation covariance; its 95% CI is [0.0075, 0.0500]. All interactions are associational. *** p < 0.01, ** p < 0.05.

Figure 2. Firm-age moderation of GTI associations with revenue productivity and ESG ratings under firm and year fixed effects. The dark-blue fitted line shows the estimated marginal association, the light-blue shaded band shows the 95% confidence interval, and the colored points mark firm age at −1 SD (blue), the mean (gray), and +1 SD (red).

4.5. Industry Heterogeneity

Table 13 and Table 14 retain the HighTech and factor-intensity split samples as descriptive benchmarks. Under year and industry fixed effects, the corrected revenue-productivity index is statistically null for HighTech firms (β = 0.000) and non-HighTech firms (β = −0.005); their estimated difference is also null (p = 0.528). Using the documented source-based IndType classification, the factor-intensity productivity slopes are −0.001 for labor-intensive firms, approximately zero for asset-/capital-intensive firms, and 0.003 for technology-intensive firms; none is statistically significant. The corresponding ESG slopes are positive in all three separate models (0.192, 0.136, and 0.165; all p < 0.01). These separate-model patterns do not themselves establish different slopes [56].

Table 13. HighTech versus non-HighTech descriptive benchmarks.

Outcome	Group	GTI (t − 1)	N	R2
TFP_LP	HighTech	0.000 (0.003)	11,324	0.477
TFP_LP	Non-HighTech	−0.005 (0.008)	6389	0.43
TFP_LP	Diff (HT−NH)	Δ = 0.0052 (z = 0.631, p = 0.528)	17,713
ESG (HZESG)	HighTech	0.167 *** (0.018)	11,319	0.166
ESG (HZESG)	Non-HighTech	0.179 *** (0.026)	6382	0.243
ESG (HZESG)	Diff (HT−NH)	Δ = −0.0112 (z = −0.355, p = 0.723)	17,701

Note: Split-sample models use year and industry fixed effects and firm-clustered inference. Difference rows compare subgroup slopes. *** p < 0.01.

Table 14. Factor-intensity descriptive benchmarks under the documented source classification.

Outcome	Industry Group	GTI (t − 1)	Observations	Firms	R2
Productivity index	Labor-intensive	−0.001 (0.008)	5310	820	0.447
Productivity index	Asset-/capital-intensive	0.000 (0.009)	3605	553	0.451
Productivity index	Technology-intensive	0.003 (0.003)	10,677	1683	0.48
ESG rating	Labor-intensive	0.192 *** (0.033)	5300	820	0.221
ESG rating	Asset-/capital-intensive	0.136 *** (0.038)	3605	553	0.16
ESG rating	Technology-intensive	0.165 *** (0.018)	10,675	1683	0.205

Note: The annual classification follows the 67-code Yin–Sheng–Li crosswalk under the 2012 CSRC system. Separate models absorb year and industry effects and cluster by firm. The 131 observations outside the published crosswalk were left unclassified and excluded. Separate significance decisions do not establish different slopes. Firm counts across separate industry groups are not mutually exclusive over the full panel because annual industry-code changes can place the same firm in different intensity classes in different years; classification remains mutually exclusive at the firm-year level. *** p < 0.01.

Table 15 reports the corresponding formal pooled equality tests using firm and year fixed effects, the same source-based classification, and the sample excluding unclassified firm-years. The estimated revenue-productivity slopes are −0.001, −0.004, and −0.004 for labor-, asset-/capital-, and technology-intensive industries, respectively; the joint equality test did not reject equality (χ² = 0.38, p = 0.826; N = 19,592; 2916 firms). The corresponding ESG slopes are 0.098, 0.073, and 0.017, and their joint equality test rejected equality at 5% (χ² = 7.30, p = 0.026; N = 19,580; 2916 firms). This is exploratory evidence that the within-firm GTI–ESG association differs across the documented factor-intensity classes, not a pre-specified moderator hypothesis or a causal contrast. The transparent reclassification resolved the construct-provenance issue; inference remains bounded because classes are observational and 131 firm-years fall outside the published mapping.

Table 15. Formal equality tests under the documented source-based industry classification.

Outcome	Labor Slope	Asset/Capital Slope	Technology Slope	Joint χ2	p-Value	Observations	Firms
Productivity index	−0.001	−0.004	−0.004	0.38	0.826	19,592	2916
ESG rating	0.098	0.073	0.017	7.3	0.026	19,580	2916

Note: Joint tests used pooled firm/year-FE interaction models with firm-clustered inference and exclude unclassified observations. The productivity equality test is null; the ESG test rejects equality at 5% and is interpreted as exploratory observational heterogeneity.

4.6. Supplementary Policy-Salience Evidence Around the 2020 Carbon Neutrality Announcement

Table 16 excludes the 2020 transition year and compares the 2021–2022 full post-announcement period with 2013–2019. With firm and year fixed effects, detailed-industry clustered standard errors, and Webb-weight wild-cluster p-values, the full-sample GTI contrast is −0.050 (SE = 0.033, wild p = 0.227) and the matched-weighted contrast is −0.034 (SE = 0.042, p = 0.469). The corrected revenue-productivity contrasts are imprecise (full β = 0.001, wild p = 0.902; matched β = 0.007, p = 0.473). The full-sample ESG contrast is positive (β = 0.093, SE = 0.046, p = 0.039), while the matched-weighted estimate is statistically null (β = 0.067, SE = 0.059, p = 0.282). Matching with replacement produced 843 treated-control draws containing 843 treated firms and 587 unique controls (1430 unique firms). Each matched draw was counted once in the balance calculation; repeated controls entered the matched panel with the corresponding frequency weights. All eight weighted post-match covariates satisfy |SMD| < 0.10 (largest |SMD| = 0.052; Table 17).

Table 16. Policy-salience comparison for the full post-announcement period.

Outcome	Sample	Treat × FullPost	95% CI	Wild p	Observations	Firms	Industry Clusters
Green patent applications	Full sample	−0.050 (0.033)	[−0.115, 0.015]	0.227	20,189	2933	54
Green patent applications	Matched weighted	−0.034 (0.042)	[−0.119, 0.051]	0.469	10,688	1430	50
TFP_LP	Full sample	0.001 (0.010)	[−0.019, 0.022]	0.902	20,189	2933	54
TFP_LP	Matched weighted	0.007 (0.010)	[−0.013, 0.027]	0.473	10,688	1430	50
ESG (HZESG)	Full sample	0.093 ** (0.046)	[0.001, 0.184]	0.039	20,170	2933	54
ESG (HZESG)	Matched weighted	0.067 (0.059)	[−0.051, 0.185]	0.282	10,671	1430	50

Note: Treat marks heavy-pollution-industry exposure; FullPost equals one in 2021–2022, and 2020 was excluded as a transition year. Matched models used frequency weights. Inference clusters on detailed industries; Wild p used 999 Webb-weight draws. The nationwide announcement and diagnostic limitations preclude causal interpretation. ** p < 0.05.

Table 17. Covariate balance for the policy-salience matched sample.

Variable	Treat (Pre-Match)	Ctrl (Pre-Match)	SMD pre	Treat (Matched)	Ctrl (Weighted)	SMD Post	Pass (|SMD| < 0.1)
Size	22.442	22.13	0.247	22.421	22.396	0.019	✓
Lev	0.402	0.419	−0.089	0.401	0.394	0.036	✓
ROE	0.066	0.043	0.155	0.066	0.062	0.034	✓
TobinQ	1.774	1.965	−0.162	1.757	1.811	−0.052	✓
Mfee	0.066	0.079	−0.249	0.066	0.068	−0.04	✓
ATO	0.67	0.604	0.183	0.661	0.669	−0.022	✓
Growth	0.101	0.118	−0.054	0.101	0.091	0.034	✓
Board	2.123	2.089	0.177	2.122	2.121	0.007	✓

Note: SMD is the standardized mean difference for 2019 covariates. Matching is 1:1 with replacement; repeated controls received frequency weights. All eight post-match covariates satisfy |SMD| < 0.10 (largest |SMD| = 0.052).

Figure 3 reports weighted 2014–2022 event-study estimates with detailed-industry-clustered 95% confidence intervals and displays 2020 separately as the transition year. Industry-clustered joint pre-policy tests reject the null for GTI at 5% (p = 0.041) but do not reject for corrected revenue productivity (p = 0.113) or ESG (p = 0.090; Table 18). The pre-policy GTI instability, nationwide announcement, nonrandom heavy-pollution classification, and COVID−19 overlap precluded a causal policy interpretation. We therefore treated the full-sample ESG contrast only as contextual policy-salience evidence, especially because the matched estimate is null.

Figure 3. Weighted policy-salience event study for GTI, revenue productivity, and ESG ratings. Blue dots and solid lines denote estimated event-time coefficients; light-blue vertical intervals denote 95% confidence intervals; the red vertical dashed line marks the 2020 transition; and the gray dotted horizontal line marks zero (the reference).

Table 18. Industry-clustered joint pre-policy tests for the weighted event study.

Outcome	Pre-Policy Joint F	df	p-Value	Industry Clusters	Matched Firms
Green patent applications	2.53	5, 49	0.041	50	1430
Revenue productivity	1.89	5, 49	0.113	50	1430
ESG (HZESG)	2.04	5, 49	0.09	50	1430

Note: The joint null set all 2014–2018 Treat × year coefficients to zero relative to 2019. Tests used detailed-industry clustered covariance; 2020 is the transition year.

5. Discussion

5.1. Summary of Findings

Table 19 summarizes the hypothesis decisions under the designated primary tests. The evidence is asymmetric and inconsistent with a uniform win–win claim. H1a is rejected because the primary within-firm revenue-productivity coefficient is negative rather than positive. H1b is accepted with qualification because the one-year-lag application coefficient is positive, while the granted-patent estimate is only marginal and the two-period-lag estimate is null. For H2, the decision is “Not supported at the 5% level (rejected under the designated criterion).” The temporally ordered environmental-attention indirect-effect confidence interval includes zero. H3 is accepted because the standardized GTI-by-age interaction is significantly larger for ESG than for productivity. These decisions indicate empirical support under the specified tests; they do not establish causal effects or causal mechanisms.

Table 19. Summary of hypothesis tests and decisions.

Hypothesis	Primary Prediction and Test	Key Evidence	Decision
H1a	GTI is positively associated with subsequent firm productivity; primary test: GTI (t − 1) in the firm/year-FE revenue-productivity model.	β = −0.0036, p = 0.038; the sign is opposite to the prediction, and the result is estimator- and specification-bounded.	Rejected (not supported).
H1b	GTI is positively associated with subsequent ESG performance; primary test: GTI (t − 1) in the firm/year-FE HZESG model.	β = 0.0441, p < 0.001; grant-based p = 0.087 and two-period-lag p = 0.618.	Accepted with qualification.
H2	Environmental attention positively mediates the association between prior GTI and subsequent ESG performance.	Indirect = 0.0010; 95% CI [−0.0001, 0.0020]; p = 0.071.	Not supported at the 5% level (rejected under the designated criterion).
H3	Firm age strengthens the GTI–ESG association more than the GTI–revenue-productivity association.	Standardized ESG-minus-productivity interaction difference = 0.0287; 95% CI [0.0075, 0.0500]; p = 0.008.	Accepted (supported).

Note: “Accepted” and “rejected” denote support or non-support under the designated primary 5% test; they do not imply causal proof. For H2, the decision is “Not supported at the 5% level (rejected under the designated criterion).” This means that the indirect-effect interval includes zero at the pre-specified threshold. H1b is qualified by proxy and lag sensitivity, and all moderation and channel estimates remain observational.

The productivity result contrasts with Sun et al. [19], who report a positive relationship between green technological innovation and corporate productivity among Chinese listed firms, but the two findings address different estimands and productivity constructions. In our data, the OP, OLS, and conventional FE productivity indices were positive in pooled benchmarks and null under firm fixed effects, whereas only the corrected nominal revenue-mode LP index was negative in the primary within-firm model. This internal pattern favors an estimand-and-timing explanation over a physical-efficiency-loss mechanism. It is also consistent with the productivity J-curve [18], under which complementary intangible investment, organizational adaptation, and commercialization can delay measured operating gains after an innovation is recorded. The negative coefficient is therefore interpreted as a short-run, construct-specific revenue-productivity association rather than evidence that green innovation destroys productive efficiency.

The positive GTI–ESG association is consistent with Huang, Sun, and Zhang [61], who found that green technology innovation improves ESG performance among Chinese listed companies, and with research linking visible sustainability activity to business-model renewal and reduced greenwashing [14,15]. Our result is deliberately narrower. The Huazheng coefficient remains positive after firm, industry-by-year, and province-by-year fixed effects, but it is strongest for recent patent applications and weaker for granted patents or a two-period lag. This timing pattern is consistent with signaling and legitimacy: current applications are public and timely indicators that can enter external evaluation before productive deployment is complete. Because ESG providers differ in scope, measurement, and weighting [12,13], the coefficient is interpreted as Huazheng rating responsiveness rather than verified environmental improvement.

The channel and contingency results further delimit this interpretation. The EnvAtt indirect-effect confidence interval crosses zero, so the archived MD&A word-count measure is not a sufficient transmission channel. The persistence of the direct ESG coefficient alongside the null EnvAtt path indicates that the rating process may draw on a broader information set—including patent records, company reports, regulatory information, and other public signals—than this single disclosure proxy captures [23]; the data do not identify which component is decisive. The larger age interaction for ESG than for productivity is consistent with accumulated legitimacy making a new patent signal more credible for older firms, whereas age does not accelerate operational conversion. The exploratory industry result—ESG slope differences but no productivity slope differences—is likewise consistent with sector-specific signal materiality rather than production complementarity, but it is not treated as a causal mechanism. The policy-salience estimates remain contextual because the GTI pre-trend fails and the matched ESG contrast is null.

5.2. Theoretical Contributions

First, this study separates an external evaluation outcome from an internal efficiency outcome under the same firm-and-year-FE design. This makes the comparison identified from common within-firm variation and formally tests both the standardized main-slope difference and the standardized age-interaction difference, advancing work that often studies only one outcome or compares coefficients informally [9,14,43,62].

Second, the results qualify signaling and stakeholder accounts. Observable green patent activity is associated with rating improvement within firms, but the absence of a temporally ordered environmental-attention indirect effect means that disclosure attention cannot be claimed as the verified transmission mechanism. This distinction matters in light of rating disagreement and possible greenwashing [12,13,15].

Third, this study makes estimand and construct sensitivity part of the theoretical interpretation. In the submitted design, the positive productivity conclusion came from a pooled model, while the accompanying firm-FE check was already null. Making firm FE primary therefore changed the hypothesis decision before the revenue-mode correction; the subsequent correction changed the primary point estimate from null to negative. The OP, OLS, and conventional FE productivity indices further show positive pooled but null within-firm coefficients. This layered evidence does not establish that green innovation reduces physical efficiency; it redirects the theoretical question toward persistent capability sorting, short-run adjustment costs, implementation horizons, revenue composition, prices, markups, and complementary capabilities [16,17,18,19].

5.3. Practical Implications

5.3.1. Evidence-Based Implications

Managers should not treat additional patent filings as a near-term productivity improvement. The combination of a negative corrected-LP coefficient, null OP and OLS within-firm estimates, and attenuation under stronger fixed effects and alternative preprocessing shows that the short-run finding is specific to the nominal revenue-productivity construction and implementation horizon. This pattern is consistent with adjustment and commercialization lags, not evidence of a physical-efficiency loss. A practical, non-validated screening matrix can keep three axes separate: (i) innovation quality and deployment, including grants, citations, and commercialization milestones; (ii) operational evidence, including asset-utilization trends and realized process changes; and (iii) external evaluation and impact, including provider-specific ESG ratings and verified emissions. A firm with rating improvement but little deployment evidence would be classified as evaluation-responsive rather than efficiency-validated. The matrix is a decision aid, not an index estimated by this study. Policy evaluators should likewise report innovation, environmental, and productivity outcomes separately.

5.3.2. Theory-Motivated Recommendations

The following recommendations follow from theory but were not directly estimated. Labor-intensive firms may prioritize workforce training, modular process redesign, and lower-capital deployment pathways; capital-intensive firms may pair patents with equipment renewal and integration engineering; technology-intensive firms may emphasize scale-up, commercialization, and intellectual property quality. Investors could request patent citation, grant, deployment, and emissions data alongside ratings. Policymakers could combine disclosure standards with commercialization support and skills programs. These segmented strategies are testable propositions, not effects established by the null productivity equality test or the exploratory ESG equality result.

5.4. Within-Firm Identification and the Pooled Benchmark

The 74.7% between-firm share of GTI variance shows that cross-firm information is substantial, but variance shares do not decompose coefficients. The primary revenue-mode within–between model directly estimates a negative within slope (β = −0.0051) and a statistically null conditional between slope (β = −0.0016); the two slopes are not significantly different. It likewise estimates a smaller positive within ESG slope (β = 0.0375) and a larger positive between ESG slope (β = 0.2088). For OP productivity, the direct decomposition instead produces a positive between slope that exceeds the null within slope; the analogous OLS difference is marginal. The difference across productivity indices is itself substantively informative: the original pooled result cannot be generalized to within-firm change, and the revenue-mode LP result cannot be generalized to every productivity construction. These terms remain conditional associations, not causal effects of within-firm innovation changes or stable capabilities.

This contrast does not prove that green innovation reduces productivity. The corrected LP application, grant, and two-period-lag coefficients are negative, but the OP, OLS, and conventional FE alternatives are null under firm FE, higher-dimensional fixed effects attenuate conventional significance, and the ever-ST/PT exclusion also spans zero. This manuscript therefore treated the primary firm-FE estimate as the H1a test, retained pooled estimates only to diagnose why the submitted conclusion changed, used alternative estimators and sample definitions as robustness boundaries, and interpreted every coefficient as noncausal. The contribution of this research is sharper identification of where an association resides, not preservation of the largest or most favorable coefficient.

5.5. Limitations

The corrected revision-stage LP outcome is independently reconstructable from the identified Mark Data inputs, but it is not validated as physical TFP. We corrected the supplied command’s mode mismatch by using levpet’s revenue option, yet the implementation still uses nominal operating revenue without a price deflator and estimates one pooled production function rather than industry-specific technologies. The resulting index can retain price, markup, scale, and accounting effects. This distinction matters because revenue productivity can diverge from physical productivity when producer prices and demand conditions vary [63]. In addition, the functional-dependence identification concern highlighted by Ackerberg, Caves, and Frazer [64] applies to this LP implementation; extending the audit to an ACF estimator is left for future research. The industry-specific revenue-mode diagnostic covers only 7837 of 23,047 study-panel rows and is unsuitable as a primary replacement; a deflated variant could not be estimated because no authoritative deflator series or documented deflator-to-input concordance was available. Reconstructed OP and OLS indices are sensitivity outcomes, not independent construct validation. The corrected pooled revenue-mode index is nearly unrelated to the legacy analytical variable (r = −0.060), demonstrating that the prior stability narrative cannot be retained. Downstream standard errors condition on the estimated productivity index and do not propagate uncertainty from the production-function estimation stage. Appendix Table A1 reports the full decision.

Huazheng ratings are a single provider’s composite measure. The licensed archive contains the overall score but not a complete, consistently defined E/S/G pillar panel, and no second-provider rating is available. We therefore cannot distinguish environmental-pillar improvement from rater responsiveness or rating disagreement. Patent applications similarly measure codified innovation activity, not implementation, commercial success, or realized emissions reduction.

The 2013–2022 sample covers listed Chinese firms, so results may not generalize to private firms or other institutional settings. Firm fixed effects do not remove time-varying confounding, the channel tests are observational, the Appendix C sensitivity contains one plainly weak peer instrument and an exclusion-ambiguous persistence instrument, and the nationwide policy-salience design lacks an untreated policy group. The GTI event-study pre-policy test rejects, and the ever-ST/PT exclusion attenuates the primary negative coefficient. Future work needs validated deflators, sufficiently covered industry technologies, longer post-innovation horizons, verified commercialization and emissions measures, green managerial and organizational innovation measures, multi-provider ESG data, and genuinely exogenous innovation shocks.

6. Conclusions

Using 23,047 firm-year observations from 2933 Chinese A-share firms during 2013–2022, this study finds a negative primary within-firm association between one-year-lagged green patent applications and a corrected LP-based nominal revenue-productivity index, alongside a positive association with subsequent Huazheng ESG ratings. H1a is not supported because it predicted a positive productivity association. H1b is supported in the primary application-based one-year-lag model but is not invariant to the granted-patent proxy or the two-period lag. A direct standardized test shows that firm age strengthens the ESG association more than the productivity association, supporting H3. The pooled-to-firm-FE contrast is analytically meaningful rather than a robustness failure: alternative OP, OLS, and conventional FE productivity indices are positive in pooled benchmarks but null within firms, with a statistically significant OP between–within difference and a marginal OLS difference. Robustness is therefore not uniform across estimands or productivity constructions. Channel and policy analyses remain exploratory or contextual; the industry equality test suggests exploratory ESG heterogeneity but no productivity heterogeneity. The primary productivity index uses the correct revenue mode but remains nominal, pooled, and undeflated; it is not a physical TFP.

Green technology innovation should therefore not be presented as an automatic dual-performance lever, nor should the negative revenue-productivity coefficient be presented as a causal efficiency loss. Visible green innovation may be reflected positively in external ratings while short-run revenue productivity moves in the opposite direction under the primary specification. Managers, investors, and policymakers should evaluate innovation inputs, ESG assessments, commercialization, operational efficiency, prices, and environmental outcomes as distinct objects. The central contribution of this study is a disciplined, construct-bounded asymmetric-association result and a transparent account of what the available panel can—and cannot—identify.

Appendix B.1. Environmental Attention (EnvAtt)

The archived documentation defines EnvAtt as the natural log of one plus the count of environment-related keywords in the firm’s annual MD&A and states that the measure uses a 64-term Chinese-language dictionary reported by Wu and Hua [36]. The archived analytical panel contains the resulting firm-year variable, but the current archive does not include the dictionary itself, the text-processing script, the underlying MD&A corpus, a document-length measure, or a missing-text handling log. The precise term composition, tokenization, length normalization, and missing-text treatment therefore cannot be independently verified or reconstructed from the available files. Exact reconstruction requires those author-provided source materials; no unverified imputation rule or missing-text rate is reported.

Appendix B.2. Industry-Type Classification (IndType)

We replaced the archived IndType assignments with a documented source-based annual classification. The 2000–2024 source dataset contains 84,336 firm-year rows, annual two-digit CSRC codes, and mutually exclusive labor-, asset-/capital-, and technology-intensive indicators. The Excel range Sheet1!A1:AG84337 contains one header row and 84,336 data rows; its Stata/MP 18 counterpart contains the same 33 variables. The firm-year keys, industry codes, and three intensity indicators agree record by record across the two formats, with no duplicate keys, cell-level discrepancies in the audited fields, or multiple-class assignments. The indicator values reproduced all 67 code assignments in Table 5 of Yin, Sheng, and Li [54], who used SPSS 19.0 Ward hierarchical clustering with the fixed-asset ratio and the R&D-to-payroll ratio under the 2012 CSRC industry system [55].

We merged the source classification to the 23,047-row study panel by stock code and year; all records matched. The annual source-record code and its accompanying published class indicator were used together rather than recombining values from two sources, preserving the source’s contemporaneous classification rule. Of 23,047 firm-years, 22,916 were classified as the following: 6181 labor-intensive, 4228 asset-/capital-intensive, and 12,507 technology-intensive. The 131 unclassified firm-years—B10 (5), C42 (30), E47 (15), O80 (4), and Q83 (77)—fall outside the published 67-code mapping and were excluded from all IndType models rather than assigned by inference.

The complete published crosswalk is as follows. Labor-intensive (40 codes): A01, A02, A03, A05, B06, B08, B09, C13, C14, C15, C17, C18, C19, C20, C21, C23, C24, C32, C34, D46, E48, E49, E50, F51, F52, G53, G54, G58, G59, I63, I64, K70, L72, M73, M75, N78, P82, R85, R87, and S90. Asset-/capital-intensive (14 codes): A04, B07, B11, C22, C25, C26, C28, C30, C31, D44, D45, G55, G56, and R86. Technology-intensive (13 codes): C27, C29, C33, C35, C36, C37, C38, C39, C40, C41, I65, M74, and N77.

Among 21,019 classified observations for which the archived assignment is also available, the archived and source-based assignments agree for 78.84%. Sixteen of 49 overlapping codes change class: B07, C23, C24, C29, C32, C33, C34, D44, D45, I63, I64, L72, M73, M75, N77, and N78. We therefore did not retain the archived assignments as the operative measure. Table 14 and Table 15 were re-estimated from the source-based classification, and the supplementary record reports the complete crosswalk, unclassified codes, source hashes, merge audit, scripts, log, and model results. The revised classification resolves the provenance limitation. The formal productivity equality test remains null, while the exploratory ESG equality test rejects equal slopes; neither result establishes causal industry moderation.

Appendix B.3. ATO (Asset Turnover)

ATO = Operating revenue/Total assets, computed from firm financial statement data. When the ATO was evaluated as a channel diagnostic, it was removed from the outcome equation’s control vector to avoid double-counting. Because the reconstructed LP index is revenue-based, ATO and TFP_LP share accounting components. Table 11 first used the productivity index (TFP_LP) and then applied a stricter outcome-side diagnostic: log operating revenue was reconstructed as Size + ln(ATO) for positive ATO, and TFP_LP was projected on reconstructed log revenue and log total assets with year and industry effects. The resulting productivity-index residual was used as the t + 1 outcome in the second ATO path. This residual is not a new production-function estimate, and the indirect term is not interpreted as causal mediation.

Appendix B.4. Historical Contemporaneous Channel Diagnostics

The previous manuscript tested three additional candidate productivity mediators using contemporaneous OLS with year and industry dummies and 200 firm-cluster bootstrap replications. These 200 draws apply only to the historical Mfee/Cashflow/INST screening. They are not the 1000-draw firm-cluster bootstrap used for the v7 core EnvAtt/ATO paths, as verified from the preserved configuration and revision record. The two counts therefore describe different analyses rather than alternative provenance for the same interval. These historical calculations used the legacy provider productivity variable and were not rerun as part of the reconstructed-outcome analysis. They are retained only as an explicitly labeled audit trail, are excluded from all revised inferences, and do not enter any table, figure, hypothesis decision, or conclusion. Their contemporaneous timing does not match the revised lagged design, and null results for three observed variables cannot establish that another variable is the dominant mechanism.

Management cost (Mfee) fails the a path (a = 0.0001, p = 0.84; indirect = −0.00027; bootstrap 95% CI [−0.0025, 0.0025]). Operating cash flow fails both paths (a = 0.0004, p = 0.60; b = −0.0430, p = 0.42; indirect = −0.00002; bootstrap 95% CI [−0.0002, 0.0000]). Institutional ownership has sign-inconsistent a and b paths (a = −0.0075, p = 0.04; b = −0.0549, p = 0.002) and only marginal indirect evidence (a·b = 0.00041; Sobel p = 0.09).

None of the three candidates shows the predicted positive two-path pattern. This does not validate the ATO: the temporally ordered ATO terms in Table 11 are statistically null, and unobserved operational, technological, governance, and supply-chain channels remain possible.