Environmental Information Disclosure Quality, Governance Structure Characteristics, and the Input–Output Efficiency of Green Innovation: A Hierarchical Linear Model Investigation of Chinese Listed Firms

Abstract

This study examines how the quality of corporate environmental information disclosure, jointly with governance structure characteristics, shapes the input–output efficiency of green innovation in Chinese A-share listed firms, and whether industry- and region-level conditions moderate that relationship. Drawing on a panel of 2847 firms spanning 2014–2024, we construct a multidimensional disclosure quality index through content analysis across completeness, verifiability, quantification depth, and forward-looking commitment, and measure green innovation efficiency through a super-efficiency slacks-based DEA model accommodating undesirable outputs. A three-level hierarchical linear model partitions variance across firm, industry, and provincial layers and permits the disclosure–efficiency slope to vary with industry regulation intensity and provincial marketization. The results indicate that higher disclosure quality is associated with greater green innovation efficiency, a link we attribute to financing-constraint relief, reputational accumulation, and intensified external monitoring, offered as interpretive channels rather than as separately tested mediators. Board independence, environmentally experienced executives, and institutional shareholding amplify the conversion, while ownership concentration dampens it. Cross-level evidence shows that industry regulation intensity and regional marketization further steepen the firm-level slope. Findings remain stable across alternative measurement, restricted sampling, propensity score matching, and instrumental variable identification. The analysis offers a multilevel reframing of disclosure–innovation research and informs the design of mandatory disclosure rules, governance reform, and green finance infrastructure in transitioning economies.

1. Introduction

The intensification of climate-related risks since the Paris Agreement has pressured firms worldwide to widen the scope and depth of their environmental reporting, with disclosure now treated less as voluntary citizenship and more as a quasi-regulatory expectation [1]. Stock exchanges across more than sixty jurisdictions have issued sustainability reporting guidance, and the consolidation of ISSB standards in 2023 marked a turning point toward comparable, decision-useful environmental data [2]. Within this evolving landscape, Chinese listed firms occupy a distinctive position. They face both the national “dual-carbon” pledge—peaking carbon emissions before 2030 and reaching neutrality by 2060—and a fragmented disclosure regime that until recently relied on voluntary CSR reports [3]. The Ministry of Ecology and Environment’s 2021 measures, together with the 2024 mandatory sustainability disclosure rules issued by the three major stock exchanges, have begun to reshape what counts as credible environmental information. Evidence from China’s earlier 2008 mandatory CSR reporting policy indicates that such mandates can measurably raise firms’ substantive environmental response, including green innovation performance [4].

Green innovation has emerged as the principal mechanism through which firms reconcile growth ambitions with decarbonization commitments. Unlike conventional R&D, green innovation embeds environmental constraints into the production function itself, generating dual externalities—knowledge spillovers and pollution abatement—that markets price imperfectly [5]. For Chinese manufacturers in particular, the transition demands not only more green R&D spending but also higher conversion efficiency from inputs to patentable outcomes [6]. The efficiency dimension is where most firms struggle: aggregate green patent filings have surged, yet output-to-input ratios remain uneven across industries and provinces [7].

Three strands of literature speak to this puzzle but rarely converge. The first examines whether environmental disclosure quality, measured by GRI compliance, third-party assurance, or text-based readability, alters firms’ real environmental behavior; findings lean positive but vary considerably with the measurement instrument chosen [8]. The second strand probes governance characteristics—board independence, the presence of environmental committees, executive green-pay sensitivity, ownership concentration—and links them to innovation outcomes [9]. Dual-class structures, state ownership, and institutional investor horizons all surface as moderators of green strategy [10]. A third strand quantifies green innovation efficiency through DEA, SFA, or Malmquist–Luenberger indices, often comparing regions or industries without modeling firm-level heterogeneity explicitly [11]. What is striking is how little these strands agree once their findings are placed side by side. The disclosure strand reports effect signs that swing from strongly positive to null depending on whether disclosure is captured as a binary report dummy or a graded content score, a divergence that is almost certainly a measurement artifact rather than a substantive disagreement. The governance strand splits along institutional lines: studies set in common-law markets tend to read concentrated ownership as patient stewardship, whereas China-based work more often reads it as entrenchment, so the moderating sign reverses across settings. The efficiency strand, by operating at whichever single level the data happen to be aggregated, produces region- and industry-level conclusions that need not hold at the firm level at all. Seen together, these are not three parallel literatures, but three sources of unresolved tension, and it is precisely the cross-level, measurement-graded design adopted here that lets us adjudicate among them rather than add a fourth disconnected result.

What unsettles us about the current body of work is its treatment of measurement and data structure. Disclosure quality is frequently proxied by a single binary indicator—whether a CSR report is issued—collapsing substantial variation in content credibility [12]. Green innovation efficiency studies tend to operate at one analytic layer (firm, industry, or region) even though the data-generating process is unambiguously nested: firms are embedded in industries shaped by competitive dynamics and regulatory pressure, while industries are themselves embedded in regional institutional environments [13]. Pooled OLS or single-level fixed-effects estimators absorb but do not decompose this clustering, leaving cross-level interactions—how regional environmental enforcement amplifies the disclosure–innovation link, for instance—largely unidentified. Recent contributions have flagged this concern, yet few move beyond two-way clustering toward formal multilevel specifications [14].

A second gap concerns mechanism identification. Most empirical work documents an average association between disclosure and innovation outcomes without separating the variance attributable to firm-level governance from that absorbed by industry technology cycles or provincial green-finance pilots. When the dependent variable is itself a ratio—innovation output over innovation input—failing to model random effects at each level can bias both slope estimates and their standard errors [15]. We see this as an opening rather than a flaw to be patched.

Against this background, our study asks how the quality of corporate environmental information disclosure, jointly with governance structure characteristics, shapes the input–output efficiency of green innovation in Chinese listed firms and whether industry- and region-level conditions moderate that relationship. Hierarchical linear modeling (HLM) is brought in deliberately: it partitions variance across firm, industry, and provincial layers; permits random slopes for disclosure quality across industries; and accommodates cross-level interactions between, say, firm-level disclosure scores and provincial environmental court density. A sample of A-share manufacturing firms covering the full 2014–2024 window supports the empirical work, with disclosure quality scored through a multidimensional text-and-content rubric and green innovation efficiency computed through a super-SBM model that accounts for undesirable outputs. We retain this eleven-year window consistently across the abstract, methods, and conclusions; where a shorter 2017–2024 horizon appears later, it denotes a deliberately restricted robustness subsample rather than the main panel.

Three contributions follow, and we want to be explicit that the value added is theoretical rather than a mere recombination of familiar tools. First, the paper does not simply propose another disclosure index; it reconceptualizes disclosure quality as a multidimensional signal whose efficiency payoff is theoretically contingent on the governance and institutional layers through which it must travel, a proposition that binary or count-based proxies cannot even articulate. Second, the three-level specification is not adopted for methodological novelty but because it operationalizes a claim that the prior literature could not test: that the disclosure–efficiency slope is itself a random quantity governed by industry regulation and regional marketization, so that cross-level interaction terms become the formal expression of a nested-institutions theory rather than a modeling convenience. Third, by relocating the dependent variable from innovation volume to conversion efficiency, the framework reframes the theoretical question itself—away from whether disclosure induces more green activity and toward whether it reshapes the internal selection of which green projects survive. The combination, in short, yields a conditional theory of disclosure value that none of the three constituent methods generates on its own. The remainder proceeds as follows: Section 2 develops the theoretical framework and hypotheses; Section 3 details data, variables, and the multilevel specification; Section 4 reports estimation results and robustness checks; Section 5 discusses mechanisms and boundary conditions; Section 6 closes with policy implications.

2. Theoretical Foundations and Research Hypotheses

2.1. Theoretical Linkage Between Environmental Disclosure Quality and Green Innovation Efficiency

Three theoretical lenses anchor our reasoning about why disclosure quality should bear on the efficiency—not merely the volume—of green innovation. Signaling theory, in its Spence-derived form, treats high-quality environmental information as a costly, hard-to-imitate signal that separates committed adopters from greenwashers [16]. Legitimacy theory, drawing on Suchman’s tripartite distinction among pragmatic, moral, and cognitive legitimacy, situates disclosure as a discursive instrument through which firms negotiate their license to operate with regulators and civil society [17], a framing echoed in evidence on standalone sustainability reporting [18]. Stakeholder theory closes the triangle by recasting disclosure as a relational good that recalibrates expectations across creditors, suppliers, customers, and employees [19]. Rather than invoking these lenses as separate justifications, we read them as one mechanism observed from three angles: a credible signal (signaling) earns a renewed license to operate (legitimacy) by realigning the expectations of resource-controlling parties (stakeholders), and it is this single chain—not the simple accumulation of theories—that we expect to surface as the financing, reputation, and monitoring channels developed below.

The first channel runs through financing constraints. Credible environmental signals shrink the informational wedge between insiders and outside capital providers, lowering the risk premium attached to green projects whose payoffs are long-dated and technologically uncertain [20]. Empirical work on China’s green credit policy confirms that credit is reallocated toward firms with stronger environmental credentials and that this reallocation raises their green innovation output—conditions that matter for capital-intensive abatement R&D [21]. When financing frictions ease, the same yuan of green R&D input can be allocated toward higher-risk but higher-payoff projects rather than incremental tweaks—raising the output-to-input ratio that defines efficiency. We capture this relationship via the conventional disclosure–cost-of-capital mapping (standard formulation):

$${Ke}_{i,t}=r_f+{\beta }_i·{MRP}_t-\delta ·{EDQ}_{i,t}+{ϵ}_{i,t}$$

(1)

where ${Ke}_{i,t}$ denotes the firm’s cost of equity, ${EDQ}_{i,t}$ the environmental disclosure quality score, and $\delta >0$ the disclosure premium discount; the sign of $\delta$ encodes the financing channel.

A second channel operates through reputational capital. Disclosure that is detailed, forward-looking, and externally assured accumulates as an intangible asset on which firms draw when entering green procurement networks, attracting environmentally screened institutional investors, or recruiting scientifically trained personnel [22]. Reputational rents do not appear directly in standard financial statements, yet they shift the marginal product of green R&D by improving collaboration quality with universities and downstream buyers willing to co-invest. The mechanism is asymmetric: damage from a disclosure scandal travels faster than the slow build-up from consistent reporting, which explains why firms with sustained high-quality disclosure show steeper innovation efficiency gains than recent converts [23].

A third channel works through intensified external monitoring. High-quality environmental disclosure equips analysts, environmental NGOs, and local regulators with the granularity needed to detect resource misallocation inside the firm [24]. Under stronger monitoring, managers internalize the cost of slack—idle green R&D personnel, duplicative pilot projects, abandoned patents—and reallocate effort towards projects with measurable environmental and commercial returns. This sharpening of internal discipline pushes the firm closer to its production-possibility frontier in green innovation space, which we represent through a standard efficiency-decomposition expression (adapted from the directional distance function literature):

$${GIE}_{i,t}=f({Input}_{i,t},{EDQ}_{i,t}\mid Z_{i,t})={\displaystyle \frac{{Output}_{i,t}^{green}}{{Input}_{i,t}^{green}}}·exp(\lambda ·{EDQ}_{i,t})$$

(2)

with ${GIE}_{i,t}$ the green innovation efficiency, $Z_{i,t}$ a vector of firm-level controls, and $\lambda$ capturing the disclosure-induced efficiency shift.

The three channels are not strictly additive. Financing relief without reputational accumulation can divert resources to superficial projects; monitoring without financing access yields disciplined but underfunded R&D pipelines. We expect their interaction to be reinforced in firms where governance permits credible signaling in the first place, which prefigures the moderating role assigned to governance characteristics in later sections [12]. Combining these arguments, our first hypothesis follows:

H1. 

Environmental information disclosure quality is positively associated with the input–output efficiency of green innovation in Chinese listed firms. We advance the financing-constraint, reputational, and external-monitoring channels as the theoretical rationale for this association; because the present design estimates the overall relationship rather than a formal mediation system, these channels are offered as candidate explanations to be probed in future mediator-explicit work rather than as empirically confirmed transmission paths.

2.2. Moderating Mechanisms of Governance Structure Characteristics

Disclosure quality does not translate into innovation efficiency through a single uniform pathway; the internal architecture of corporate control conditions how the signal is received, interpreted, and acted upon. Agency theory and upper-echelons theory together supply the analytical scaffolding here. The former, in its Jensen–Meckling formulation, frames governance as a set of devices that constrain managerial opportunism over discretionary investments such as green R&D [25]. The latter, following Hambrick and Mason, treats observable executive characteristics as proxies for cognitive filters that shape strategic interpretation under uncertainty [26]. Four governance dimensions appear most consequential for the disclosure–efficiency link.

Board independence comes first. Independent directors with no operational ties to management are better placed to scrutinize the credibility of disclosed environmental information and to challenge resource allocation decisions in green innovation projects whose payoffs are diffuse [27]. Their presence raises the cost of disclosure that is decoupled from real abatement effort, which in turn pushes managers toward disclosure–action consistency. We expect the slope of efficiency on disclosure quality to steepen as the independent director ratio rises. Yet the relationship may not be strictly monotonic—excessive independence without industry-specific expertise can produce overcautious vetoes on novel green technologies [28]. The moderated relationship is captured as

$${GIE}_{i,t}={\alpha }_0+{\alpha }_1{EDQ}_{i,t}+{\alpha }_2{IND}_{i,t}+{\alpha }_3({EDQ}_{i,t}\times {IND}_{i,t})+X_{i,t}\gamma +{\epsilon }_{i,t}$$

(3)

with ${IND}_{i,t}$ denoting the proportion of independent directors and ${\alpha }_3$ the interaction coefficient of interest.

Ownership concentration moves in the opposite direction. Highly concentrated ownership in Chinese listed firms often coincides with controlling-shareholder entrenchment, where minority-investor protections weaken and green disclosure risks being repurposed as a legitimacy veneer rather than a commitment device [29]. Concentrated owners may also prefer short-cycle commercial R&D over green innovation whose returns are uncertain and partially externalized. We thus anticipate a dampening effect: the higher the largest shareholder’s stake, the weaker the conversion from disclosure quality into innovation efficiency.

Executives with environmental backgrounds—prior experience in environmental agencies, sustainability consultancies, or green technology firms—enter through the upper-echelons logic. Their interpretive schemata privilege long-horizon environmental payoffs and reduce the cognitive cost of redirecting R&D portfolios toward abatement-relevant trajectories [30], a reading that rests on the reflexive foundations of managerial cognition emphasized in the strategic-management literature [31]. Such executives also tend to invest in internal carbon accounting infrastructure, which makes disclosure information more actionable inside the firm rather than purely outward-facing. The amplification effect on the disclosure–efficiency relationship is modeled as

$${GIE}_{i,t}={\beta }_0+{\beta }_1{EDQ}_{i,t}+{\beta }_2{ENVEXE}_{i,t}+{\beta }_3({EDQ}_{i,t}\times {ENVEXE}_{i,t})+X_{i,t}\eta +{\nu }_{i,t}$$

(4)

where ${ENVEXE}_{i,t}$ is a binary indicator for environmentally experienced executives in the top management team.

Institutional ownership constitutes the fourth lever, though its sign is theoretically ambiguous and likely contingent on investor horizon. Long-horizon institutional investors—pension funds, social security funds, green-mandated ETFs—reward disclosure-consistent green strategies and absorb the short-term earnings volatility that abatement R&D can introduce [32]. Transient institutional holders behave differently, often pressing management towards quarterly performance targets that crowd out green investment efficiency gains. Without distinguishing horizon explicitly, the average moderating effect should still be positive, since pressure-resistant institutional investors have been shown to raise green innovation in Chinese listed firms while pressure-sensitive holders do not [33]. The composite interaction is written as

$${GIE}_{i,t}={\theta }_0+{\theta }_1{EDQ}_{i,t}+{\theta }_2{INST}_{i,t}+{\theta }_3({EDQ}_{i,t}\times {INST}_{i,t})+X_{i,t}\varphi +{\xi }_{i,t}$$

(5)

with ${INST}_{i,t}$ measuring institutional shareholding proportion.

The four mechanisms operate in parallel rather than sequentially, and their relative strength is itself an empirical question. We formalize the resulting set of moderation hypotheses:

H2a. 

Board independence positively moderates the relationship between environmental disclosure quality and green innovation efficiency, strengthening the conversion from disclosure into efficiency gains.

H2b. 

Ownership concentration negatively moderates the disclosure–efficiency relationship, weakening the efficiency translation of high-quality disclosure.

H2c. 

The presence of executives with environmental backgrounds positively moderates the disclosure–efficiency relationship, amplifying the efficiency gains derived from credible disclosure.

H2d. 

Institutional investor shareholding positively moderates the disclosure–efficiency relationship, reinforcing the link between disclosure quality and green innovation efficiency.

2.3. Cross-Level Nesting Effects and Multilevel Hypothesis Construction

Firm-level mechanisms operate inside contexts that are themselves variable. Resource dependence theory, following Pfeffer and Salancik, reminds us that firms calibrate their environmental behavior against the external actors controlling critical resources—regulators, capital providers, technology partners—whose composition varies by industry [34]. Institutional theory adds a second layer: organizational fields embedded in different regional institutional matrices face heterogeneous coercive, normative, and mimetic pressures that selectively reward or punish disclosure-driven strategies [35]. Together, these perspectives imply that the firm-level disclosure–efficiency slope is itself a random variable, with industry and provincial conditions as candidate predictors of its variation.

Two contextual conditions matter most for our setting. Industry-level environmental regulation intensity—captured through pollution discharge standards, emission permit stringency, and sector-specific carbon quotas—alters the marginal value of a credible environmental signal [36]. In heavily regulated industries such as thermal power or steel, disclosure carries higher informational weight because the regulatory cost of misrepresentation is steep, and the disclosure–efficiency relationship should tilt upward accordingly. In lightly regulated industries, the same signal blurs into corporate communications noise, and its efficiency payoff thins out. Regional marketization, measured through the Fan Gang–Wang Xiaolu marketization index (hereafter the Fan–Wang index) or its successors, captures the depth of factor markets, the protection of property rights, and the maturity of intermediary services [37]. Provinces scoring high on marketization channel disclosure benefits more efficiently into green credit, technology transactions, and labor reallocation; provinces with thinner markets dampen the transmission [38].

The nested structure—firm-year observations grouped within industries grouped within provinces—motivates a three-level random-slope specification. At Level 1, firm-year green innovation efficiency depends on disclosure quality and governance moderators. At Level 2, the disclosure slope varies across industries as a function of regulatory intensity. At Level 3, both the intercept and the slope absorb provincial marketization variation. The compact reduced form for the cross-level interaction is

$$\begin{gathered} {GIE}_{ijk,t}={\gamma }_{000}+{\gamma }_{100}{EDQ}_{ijk,t}+{\gamma }_{010}{REG}_{jk}+{\gamma }_{001}{MKT}_k+{\gamma }_{110}({EDQ}_{ijk,t}\times {REG}_{jk}) \\ +{\gamma }_{101}({EDQ}_{ijk,t}\times {MKT}_k)+u_{00k}+r_{0jk}+e_{ijk,t} \end{gathered}$$

(6)

where $i$, $j$, $k$, $t$ index firm, industry, province, and year; ${REG}_{jk}$ denotes industry-level regulation intensity, ${MKT}_k$ provincial marketization, and the ${\gamma }_{110}$, ${\gamma }_{101}$ terms encode the cross-level interactions [39]. The error components $u_{00k}$, $r_{0jk}$, and $e_{ijk,t}$ partition variance across the three layers. The intraclass correlation coefficients are derived in the conventional manner:

$${ICC}_{province}={\displaystyle \frac{{\sigma }_u^2}{{\sigma }_u^2+{\sigma }_r^2+{\sigma }_e^2}}, {ICC}_{industry}={\displaystyle \frac{{\sigma }_r^2}{{\sigma }_u^2+{\sigma }_r^2+{\sigma }_e^2}}$$

(7)

with non-trivial ICC values providing the empirical license for hierarchical estimation [40].

Two cross-level hypotheses follow, completing the multilevel system that runs from H1 through H3:

H3a. 

Industry-level environmental regulation intensity positively moderates the firm-level relationship between environmental disclosure quality and green innovation efficiency; the disclosure–efficiency slope steepens in industries facing tighter regulatory pressure.

H3b. 

Provincial marketization positively moderates the firm-level disclosure–efficiency relationship; the efficiency conversion of disclosure quality is stronger in regions with more developed market institutions.

Together with H1 (the direct effect) and H2a–H2d (firm-level governance moderators), these cross-level propositions form a coherent three-tier hypothesis system in which firm conduct, industry context, and regional institutions jointly shape the productivity of green innovation pipelines.

3. Research Design and Model Construction

3.1. Sample Selection and Data Sources

The empirical analysis draws on Chinese A-share listed firms in the Shanghai and Shenzhen Stock Exchanges across the 2014–2024 window. This period brackets several institutional shifts that are central to our research question: the 2015 revision of the Environmental Protection Law, the launch of the national carbon market in 2021, and the issuance of the 2024 mandatory sustainability disclosure rules by the three major exchanges [41]. Beginning the panel in 2014 also gives us one pre-policy year as a baseline against which subsequent disclosure-quality trajectories can be benchmarked.

Sample construction followed a sequence of screening filters. Firms designated as ST or *ST during the observation window were dropped because financial distress confounds the disclosure–innovation relationship. The asterisk in *ST is part of the exchange’s special-treatment label, not a footnote marker. Financial sector firms (CSRC code J) were removed given their distinct regulatory regime and the conceptual difficulty of applying manufacturing-oriented green innovation metrics to them. Firms with fewer than three consecutive years of observations were excluded to preserve the panel structure required by the three-level model. Observations missing values on key variables—disclosure quality scores, green patent counts and applications, R&D inputs, or core governance characteristics—were eliminated through listwise deletion at the variable-construction stage.

Four data streams feed the analysis. Environmental disclosure quality scores were constructed by hand-coding annual reports, standalone CSR reports, and ESG reports against a multidimensional rubric covering content completeness, third-party verification, forward-looking commitments, and quantification depth, following the content-analytic approach to Chinese corporate environmental disclosure of Meng et al. [42], with raw documents pulled from the CNINFO (Juchao) platform. Governance structure variables—board independence ratio, ownership concentration, executive backgrounds, and institutional shareholding—came from the CSMAR database, supplemented by manual cross-checking of director biographies for environmental experience tagging. Green innovation outputs and inputs were assembled from the CNRDS green patent database and the CSMAR R&D expenditure module, with patents classified as green using the IPC Green Inventory mapping issued by WIPO [43]. Industry-level environmental regulation intensity was constructed from the China Statistical Yearbook on Environment, while provincial marketization scores follow the updated Fan–Wang marketization index series [44], a provincial institutional measure widely used in the regional green-innovation efficiency literature [45].

The matching procedure linked datasets through stock codes for firm-level variables, two-digit CSRC industry codes for industry-level moderators, and provincial codes derived from registered office locations for regional moderators. Cross-database inconsistencies—particularly around mid-year industry reclassifications and corporate restructurings—were resolved by referring back to original SSE and SZSE announcements. Figure 1 traces the full screening and matching workflow from raw data sources to the analytical panel.

After applying the filters, the analytical panel comprises 2847 unique firms and 21,536 firm-year observations spanning eleven calendar years, distributed across 28 two-digit manufacturing and related industry codes and 30 provincial-level administrative units.

Table 1. Sample screening process and final sample distribution.

Screening Step	Firms Removed	Firm-Years Removed	Remaining Firms	Remaining Firm-Years
Initial A-share pool (2014–2024)	—	—	4612	38,724
Exclude ST/*ST firms	412	3128	4200	35,596
Exclude financial sector (CSRC code J)	218	1896	3982	33,700
Exclude firms with fewer than 3 consecutive years	487	4103	3495	29,597
Exclude observations missing disclosure scores	326	3847	3169	25,750
Exclude observations missing governance variables	198	2512	2971	23,238
Exclude observations missing green patent or R&D data	124	1702	2847	21,536
Final analytical sample	—	—	2847	21,536

summarizes screening counts at each step together with the final sample distribution.

3.2. Variable Definitions and Measurement

3.2.1. Environmental Disclosure Quality

Building the disclosure quality index occupied the most labor-intensive part of our measurement work. Following the content-analysis tradition refined by Clarkson and colleagues, we constructed a scoring rubric organized along four dimensions—content completeness, verifiability, quantification depth, and forward-looking commitment—each decomposed into three items rated on a 0–3 ordinal scale, where 0 denotes silence, 1 a boilerplate qualitative mention, 2 a substantive narrative disclosure, and 3 a quantified and verifiable disclosure [46]. To take one concrete decision rule, a firm that states only that it “values environmental protection” scores 1 on the emissions item, a firm reporting emission categories in prose scores 2, and a firm tabulating tonnage by pollutant against a prior-year baseline scores 3; carbon-neutrality language without a dated target is capped at 2 on the forward-looking items. The aggregate is formed by summing the twelve items with equal weights—a deliberate choice, since data-driven weighting schemes (entropy or principal-component loadings) reproduced the same firm rankings with a Spearman correlation above 0.97, so equal weighting buys transparency at no cost to ordering—and the raw 0–36 total is then linearly rescaled to a 0–100 continuum. On reliability, the two coders were postgraduate researchers who completed a calibration phase on 300 pilot reports before scoring began, with the rubric frozen only after their pilot kappa cleared 0.75; the full corpus was then scored over roughly nine months, blind to the firms’ innovation outcomes to forestall hindsight bias. The reported Cohen’s kappa of 0.83 is computed on the production set, item-level disagreements were reconciled by joint adjudication with a third senior researcher breaking ties, and a 10 percent random audit re-coded after a two-month interval reproduced 94 percent of item scores exactly, which we take as adequate evidence of intra-coder stability at this scale.

Table 2. Environmental disclosure quality scoring framework.

Dimension	Indicator Items	Score Range	Source of Evidence
Content completeness	Environmental policy statement	0–3	Annual/CSR report
Content completeness	Pollutant emission types and volumes	0–3	CSR/ESG report
Content completeness	Resource consumption disclosure	0–3	Annual report appendix
Verifiability	Third-party assurance presence	0–3	Assurance statement
Verifiability	Quantitative vs. narrative ratio	0–3	Full-text analysis
Verifiability	Comparable prior-year data	0–3	CSR report tables
Quantification depth	Carbon emission inventory (Scope 1–3)	0–3	ESG report
Quantification depth	Environmental investment monetary figures	0–3	Financial notes
Quantification depth	Energy intensity metrics	0–3	Operational disclosure
Forward-looking commitment	Carbon neutrality timeline	0–3	Strategic section
Forward-looking commitment	Quantified abatement targets	0–3	Sustainability strategy
Forward-looking commitment	Climate-related risk scenarios	0–3	TCFD-aligned content

sets out the full scoring scheme.

3.2.2. Green Innovation Efficiency

For the dependent variable, we adopted a super-efficiency slacks-based measure DEA model with undesirable outputs—commonly labeled the super-SBM approach in environmental productivity literature [47]. The choice is not incidental: radial DEA models cannot handle slack in inputs and undesirable outputs simultaneously, which matters when green R&D pipelines generate both desired patents and unwanted by-products such as failed pilot waste. The underlying SBM efficiency score for decision-making unit $d$ is given by the standard formulation:

$${\rho }^{*}=min{\displaystyle \frac{1-\frac{1}{m}{\sum }_{i=1}^m\frac{s_i^{-}}{x_{i0}}}{1+\frac{1}{q_1+q_2}({\sum }_{r=1}^{q_1}\frac{s_r^g}{y_{r0}^g}+{\sum }_{l=1}^{q_2}\frac{s_l^b}{y_{l0}^b})}}$$

(8)

where $x_{i0}$ denotes inputs, $y_{r0}^g$ desirable outputs, $y_{l0}^b$ undesirable outputs, and $s_i^{-}$, $s_r^g$, $s_l^b$ the corresponding slack variables [48]. Because the standard SBM cannot discriminate among units that attain a score of unity, we report the super-efficiency variant of Equation (8), which excludes the evaluated unit from its own reference set and therefore admits scores above one. Inputs comprise green R&D expenditure and the headcount of R&D personnel engaged in green projects; desirable outputs are green patent grants and green patent citation-weighted counts; undesirable outputs cover industrial wastewater discharge and SO₂ emissions normalized by output value. We should be candid about how the two undesirable outputs were measured at the firm level, since firm-specific pollution figures are indeed patchy. Where a firm reported discharge and emission volumes directly in its environmental or ESG disclosure, those figures were used; where they were absent, we imputed firm emissions by allocating the relevant two-digit industry total from the China Statistical Yearbook on Environment in proportion to the firm’s share of industry output value, an approach that recovers cross-firm intensity differences while acknowledging that the absolute levels for non-reporting firms are estimated rather than observed. Firms for which neither a reported value nor a credible output-share allocation could be constructed were excluded at the missing-data screening stage rather than assigned zeros, so the efficiency scores are not contaminated by spurious clean records. One further caveat belongs here: green patent counts proxy innovation output imperfectly, because a granted green patent can reflect either a genuine technological advance or strategic, subsidy-driven filing, and our citation-weighting only partially separates the two; the efficiency scores should therefore be read as upper bounds on real abatement-relevant productivity.

3.2.3. Governance Structure Variables

Four governance attributes operationalize the moderators developed in Section 2.2. Board independence (IND) is the ratio of independent directors to total board members. Ownership concentration (OC) takes the shareholding percentage of the largest shareholder. Environmental executive background (ENVEXE) is a binary marker equal to one if any member of the top management team has prior service in an environmental regulatory agency, an environmental NGO, or a green technology firm, based on biographical data screened from CSMAR and cross-checked manually [49]. Institutional shareholding (INST) measures the aggregate proportion held by institutional investors at year-end.

3.2.4. Industry- and Region-Level Variables

Industry-level regulation intensity (REG) is derived from the ratio of pollution treatment investment to industry value added at the two-digit CSRC industry code level, log-transformed to reduce skewness. Provincial marketization (MKT) follows the Fan–Wang index, which scores regional market development on government–market relations, non-state economic share, factor market sophistication, and legal environment [44]. Both moderators enter the model grand-mean centered to ease the interpretation of cross-level interaction terms.

3.2.5. Controls

Firm-level controls cover size (Size, natural log of total assets), leverage (Lev, debt-to-asset ratio), profitability (ROA), firm age (Age), state ownership (SOE, binary), and Tobin’s Q. Industry-level controls include industry concentration (HHI). Provincial controls absorb GDP per capita (lnGDP) and environmental court density per million population [50].

Table 3. Variable definitions and measurement methods.

Variable	Symbol	Definition/Measurement
Green innovation efficiency	GIE	Super-SBM score with undesirable outputs
Environmental disclosure quality	EDQ	Content-analysis composite score (0–100)
Board independence	IND	Independent directors/total board members
Ownership concentration	OC	Largest shareholder’s stake (%)
Environmental executive background	ENVEXE	Dummy = 1 if TMT contains environmental experience
Institutional shareholding	INST	Year-end institutional holding ratio (%)
Industry regulation intensity	REG	Log of pollution treatment investment/industry value added
Provincial marketization	MKT	Fan–Wang marketization index score
Firm size	Size	Natural logarithm of total assets
Leverage	Lev	Total debt/total assets
Profitability	ROA	Net income/total assets
Firm age	Age	Years since incorporation
State ownership	SOE	Dummy = 1 if ultimate controller is state
Tobin’s Q	TobinQ	(Market value + total debt)/total assets
Provincial GDP per capita	lnGDP	Log of per capita GDP

consolidates the full variable inventory.

3.3. Multilevel Linear Model Specification and Estimation

3.3.1. Reasons for Selecting HLM over OLS

The case for hierarchical estimation in our setting rests less on statistical preference than on the structure of the data itself. Firm-year observations in our panel are not independent draws: a chemical manufacturer in Jiangsu shares unobserved exposure to sector-specific emission standards with its industry peers, and simultaneously faces a provincial regulatory environment distinct from one in Sichuan. Pooled OLS ignores this clustering and produces downward-biased standard errors on slope estimates, while two-way clustered errors correct the variance but cannot recover the cross-level variance components themselves [51]. Hierarchical linear modeling, originally formalized by Raudenbush and Bryk, decomposes total variance into firm-, industry-, and province-level components and admits random slopes that vary across higher-level units [39]. For our research question, which explicitly asks how regional and industry conditions reshape firm-level disclosure–efficiency slopes, this decomposition is non-negotiable. Two features of our data structure deserve a frank statement, since they complicate any naive “Russian-doll” reading of the hierarchy. First, industry and province are not strictly nested: the same two-digit industry appears across many provinces and each province hosts firms from many industries, so the two contexts are more accurately treated as cross-classified higher-level factors rather than one cleanly contained within the other. We therefore specify industry and province as crossed random effects on the intercept and on the EDQ slope, and we verified that a fully cross-classified estimator returns variance components within rounding distance of the strictly nested version, which is why the substantive conclusions are insensitive to the choice. Second, because the panel stacks repeated firm-year observations, the truly lowest grouping unit is the firm, not the firm-year; we accordingly add a firm-level random intercept nested under the crossed industry and province effects, together with year fixed effects and cluster-robust standard errors at the firm level, so that within-firm serial dependence is absorbed rather than mistaken for signal [39].

3.3.2. Three-Level Nested Architecture

Figure 2 traces the nesting logic graphically: firm-year observations (Level 1) sit inside industries (Level 2), which in turn sit inside provinces (Level 3).

Table 4. Multilevel model hierarchy and variable assignment.

Level	Unit of Analysis	Predictors	Random Component
Level 1	Firm-year	EDQ, IND, OC, ENVEXE, INST, Size, Lev, ROA, Age, SOE, TobinQ	$e_{ijk,t}$
Level 2	Industry	REG, HHI	$r_{0jk}$$, r_{1jk}$
Level 3	Province	MKT, lnGDP, court density	$u_{00k}$$, u_{10k}$

maps variables to their corresponding analytic layer.

3.3.3. Stepwise Model Construction

We build the specification in four stages, each progressively relaxing constraints. The null model partitions variance with no predictors:

$${GIE}_{ijk,t}={\gamma }_{000}+u_{00k}+r_{0jk}+e_{ijk,t}$$

(9)

from which the intraclass correlations are computed as

$${ICC}_3={\displaystyle \frac{{\sigma }_{u_{00}}^2}{{\sigma }_{u_{00}}^2+{\sigma }_{r_0}^2+{\sigma }_e^2}},{ICC}_2={\displaystyle \frac{{\sigma }_{r_0}^2}{{\sigma }_{u_{00}}^2+{\sigma }_{r_0}^2+{\sigma }_e^2}}$$

(10)

Following Hox’s threshold guidance, ICC values above 0.05 at any level justify proceeding to multilevel estimation [52]. The random-intercept model then adds firm-level predictors while letting intercepts vary across industries and provinces:

$${GIE}_{ijk,t}={\beta }_{0jk}+{\beta }_1{EDQ}_{ijk,t}+{\sum }_{p=2}^P{\beta }_p{X}_{p,ijk,t}+e_{ijk,t}$$

(11)

$${\beta }_{0jk}={\gamma }_{00k}+r_{0jk},{\gamma }_{00k}={\gamma }_{000}+u_{00k}$$

(12)

The random-slope model permits the disclosure–efficiency slope itself to vary across higher-level units, which is the structural feature that distinguishes our analysis from conventional fixed-effects panels:

$${\beta }_{1jk}={\gamma }_{100}+r_{1jk}+u_{10k}$$

(13)

The full cross-level interaction model incorporates industry regulation and provincial marketization as predictors of both intercepts and slopes:

$${\beta }_{0jk}={\gamma }_{000}+{\gamma }_{010}{REG}_{jk}+{\gamma }_{001}{MKT}_k+r_{0jk}+u_{00k}$$

(14)

$${\beta }_{1jk}={\gamma }_{100}+{\gamma }_{110}{REG}_{jk}+{\gamma }_{101}{MKT}_k+r_{1jk}+u_{10k}$$

(15)

Substituting (14) and (15) into (11) yields the reduced form:

$$\begin{array}{l}{GIE}_{ijk,t}={\gamma }_{000}+{\gamma }_{100}{EDQ}_{ijk,t}+{\gamma }_{010}{REG}_{jk}+{\gamma }_{001}{MKT}_k+{\gamma }_{110}(EDQ\times REG)\\ \hspace{1em}\hspace{1em}\hspace{1em} +{\gamma }_{101}(EDQ\times MKT)+X^{\prime }\gamma +r_{0jk}+r_{1jk}EDQ+u_{00k}\\ \hspace{1em}\hspace{1em}\hspace{1em} +u_{10k}EDQ+e_{ijk,t}\end{array}$$

(16)

Table 5. Model specification and equation correspondence.

Stage	Model Name	Corresponding Equation
1	Null model	Equation (9)
2	Random-intercept model	Equations (11)–(12)
3	Random-slope model	Equation (13)
4	Cross-level interaction model	Equations (14)–(16)
5	Full model with controls	$\mathrm{Equation} \left(16\right) \mathrm{with} X^{\prime }\gamma$

maps each modeling stage to the corresponding equation.

3.3.4. Estimation and Fit Assessment

Parameters are estimated through restricted maximum likelihood (REML), which is preferred over full maximum likelihood when variance components are the inferential target, since REML yields less biased estimates in small upper-level samples [53,54]. The choice matters here precisely because our upper levels are modest—28 industries and 30 provinces—a regime in which full maximum likelihood is known to understate between-group variance; for the same small-sample reason we test fixed effects through Wald statistics with the Kenward–Roger degrees-of-freedom adjustment, which inflates the effective denominator degrees of freedom to guard against the anti-conservative inference that plagues unbalanced multilevel panels. Model fit improvements across nested specifications are assessed through the deviance statistic:

$$\Delta D=-2ln{L}_{restricted}+2ln{L}_{full}~{\chi }_k^2$$

(17)

where $k$ is the difference in free parameters between models [55]. Complementary fit indices include Akaike’s Information Criterion and the pseudo-$R^2$ at each level, calculated as the proportional reduction in residual variance relative to the null model:

$$R_{Level1}^2={\displaystyle \frac{{\sigma }_{e,null}^2-{\sigma }_{e,full}^2}{{\sigma }_{e,null}^2}}$$

(18)

All estimation was carried out in HLM 8.0 (Scientific Software International, Skokie, IL, USA) with cross-validation against the mixed command in Stata 19 (StataCorp LLC., College Station, TX, USA) to confirm computational stability. To show that the hierarchical machinery earns its keep rather than merely complicating the analysis, we re-estimated the disclosure–efficiency relationship with a conventional two-way fixed-effects panel carrying firm and year dummies, the workhorse of corporate-finance research. That estimator returns an EDQ coefficient of comparable sign and magnitude but, by construction, sweeps the industry- and province-level variance into the fixed effects and so cannot recover the cross-level interaction terms that constitute our actual research question; it also delivers a worse information-criterion fit than the random-slope specification. The comparison, summarized alongside the variance components in

Table 6. Variance decomposition and model comparison across estimation stages.

Specification	Province ICC	Industry ICC	Deviance	AIC	Pseudo-${\mathit{R}}^2$ (L1)
Null model	0.143	0.182	12,884	12,892	—
Random intercept	0.121	0.156	12,597	12,615	0.18
Random slope	0.118	0.149	12,463	12,489	0.24
Cross-level interaction	0.104	0.131	12,387	12,421	0.29
Full model + controls	0.092	0.118	12,205	12,259	0.34
Two-way FE (firm + year)	absorbed	absorbed	—	12,498	0.27

Lower AIC indicates better fit. Under the two-way fixed-effects benchmark, the industry and province variance is absorbed into the fixed effects and cannot be recovered as a component.

, is the concrete evidence that HLM here improves identification of the conditioning effects rather than simply offering an alternative bookkeeping of the same average slope.

4. Empirical Results and Analysis

4.1. Descriptive Statistics and Baseline Regression

Across the 21,536 firm-year observations, the disclosure quality index averages 42.7 with a standard deviation of 18.3, indicating wide dispersion that the panel structure preserves rather than smooths over. The interquartile range runs from 28.4 to 56.9, and a non-trivial right tail above 80 reflects the small group of firms—mostly large central state-owned enterprises in energy and materials—that produce GRI-aligned, externally assured sustainability reports. Green innovation efficiency scores range from 0.12 to 1.34, with a mean of 0.518, consistent with the modest aggregate productivity figures reported for Chinese manufacturing green R&D pipelines [56]. Pairwise correlations between key explanatory variables stay below 0.45 in absolute value, and variance inflation factors range from 1.21 to 3.67 with a mean of 2.04, well below the conventional threshold of 10, so multicollinearity is not a serious concern.

Before fitting any predictor-augmented specification, we ran the null model defined in Equation (9) to recover the variance components. The intraclass correlation at the provincial level lands at 0.143, while the industry-level ICC is 0.182, leaving 67.5 percent of total variance at the firm-year layer. Both higher-level ICCs comfortably exceed the 0.05 cutoff suggested by methodologists working with nested organizational data [40], confirming that pooling all observations into a single-level OLS would obscure systematic between-cluster differences and bias inference on the disclosure–efficiency slope. Table 6 reports the variance components alongside the deviance, AIC, and level-specific pseudo-$R^2$ for each modeling stage and for the two-way fixed-effects benchmark.

The grouped comparison in Figure 3 makes the disclosure–efficiency relationship visible before any conditioning takes place. Firms in the top quartile of disclosure quality post mean efficiency scores roughly 38 percent above those in the bottom quartile, and the gap widens monotonically across the four bands.

Pooling individual firm-years rather than quartile means, Figure 4 traces the conditional relationship in a scatterplot with a LOESS fit overlaid. The upward slope is unmistakable, though the relationship flattens at the upper end of the disclosure distribution, hinting at diminishing marginal returns once disclosure quality moves beyond the regulatory frontier.

Temporal dynamics deserve a separate look. Figure 5 plots sample-wide mean efficiency by year and shows a gentle upward drift from 0.43 in 2014 to 0.58 in 2024, with a visible acceleration after 2020 that coincides with the dual-carbon policy announcement and the green credit guideline rollout [20,57].

Heterogeneity across two-digit industries warrants attention before the regressions interpret it away. Figure 6 displays the boxplot distribution by industry code, with pollution-intensive sectors—steel, chemicals, thermal power—exhibiting both lower medians and wider dispersions than electronics or pharmaceuticals.

Baseline random-intercept estimation, with EDQ entered as the sole Level-1 predictor alongside firm controls, returns a coefficient of ${\gamma }_{100}=0.0046$ ($SE=0.0008$, $p<0.01$). A one-standard-deviation rise in disclosure quality is associated with a 0.084 increase in the efficiency score—about 16 percent of the sample mean—an association whose economic size mirrors the visual pattern in Figure 3 and Figure 4. We phrase this deliberately as an association: even with the robustness battery that follows, the baseline rests on observational variation, and the magnitude should be read as the conditional difference between high- and low-disclosure firms rather than as the effect of an exogenous change in disclosure. The deviance test against the null model returns $\Delta D=287.4$ with two degrees of freedom, decisively rejecting the no-effect specification. Firm size and ROA enter with the expected positive signs, while leverage suppresses efficiency, consistent with the financing-constraint mechanism developed in Section 2.1 and broadly aligned with panel evidence linking financing constraints and green innovation in Chinese firms [58]. The baseline coefficient supplies a benchmark against which the random-slope and cross-level interaction models in subsequent subsections will be evaluated.

Table 7. Hierarchical linear model estimates of green innovation efficiency.

Variable	Baseline (RI)	Governance	Cross-Level	Full Model
$\mathrm{EDQ} ({\gamma }_{100}$)	0.0046 *** (0.0008)	0.0044 *** (0.0008)	0.0045 *** (0.0008)	0.0043 *** (0.0009)
EDQ × IND	—	0.021 *** (0.007)	—	0.019 ** (0.008)
EDQ × OC	—	−0.014 *** (0.005)	—	−0.013 ** (0.005)
EDQ × ENVEXE	—	0.0034 *** (0.0010)	—	0.0031 *** (0.0011)
EDQ × INST	—	0.009 ** (0.004)	—	0.008 ** (0.004)
$\mathrm{EDQ} \times \mathrm{REG} ({\gamma }_{110}$)	—	—	0.0018 *** (0.0006)	0.0017 *** (0.0006)
$\mathrm{EDQ} \times \mathrm{MKT} ({\gamma }_{101}$)	—	—	0.0015 *** (0.0005)	0.0014 *** (0.0005)
$\mathrm{REG} ({\gamma }_{010}$)	—	—	0.012 ** (0.005)	0.011 ** (0.005)
$\mathrm{MKT} ({\gamma }_{001}$)	—	—	0.009 ** (0.004)	0.008 ** (0.004)
Firm controls	Yes	Yes	Yes	Yes
Constant	0.402 *** (0.031)	0.398 *** (0.030)	0.405 *** (0.032)	0.399 *** (0.033)
Observations	21,536	21,536	21,536	21,536

Standard errors in parentheses. *** p < 0.01, ** p < 0.05. Firm controls: Size, Lev, ROA, Age, SOE, Tobin’s Q. Estimation by REML with Kenward–Roger adjustment.

collects the full set of fixed-effect coefficients, standard errors, and significance levels for the baseline, governance-moderation, cross-level, and fully controlled specifications side by side, so the progression can be read directly rather than reconstructed from the text.

4.2. Moderating Effects of Governance Structure Characteristics

Testing the four governance moderators developed under H2a–H2d proceeds through sequential introduction of interaction terms into the random-intercept specification, with all continuous moderators grand-mean centered to keep the conditional EDQ slope interpretable at sample means [59]. We resist the temptation to enter all four interactions simultaneously in a single equation, since collinearity between governance variables would muddy the marginal effect estimates; instead, each moderator enters in its own augmented model with the others held as Level-1 controls.

The board independence interaction produces a positive and statistically reliable coefficient. Estimated from the specification corresponding to Equation (3) in Section 2.2, the interaction term returns ${\alpha }_3=0.021$ ($SE=0.007$, $p<0.01$), with the marginal effect of EDQ on GIE expressed as

$${\displaystyle \frac{\partial {GIE}_{i,t}}{\partial {EDQ}_{i,t}}}={\alpha }_1+{\alpha }_3·{IND}_{i,t}$$

(19)

Substituting the 25th and 75th percentile values of IND (0.33 and 0.45) yields conditional EDQ slopes of 0.0038 and 0.0063, respectively, a 66 percent increase across the interquartile range. The amplification is consistent with the monitoring logic developed in H2a: independent boards raise the cost of decoupled disclosure and steer green R&D toward higher-conversion projects [60].

Ownership concentration cuts the other way. The interaction coefficient is negative and significant (${\beta }_3^{OC}=-0.014$, $SE=0.005$, $p<0.01$), and the conditional slope expression follows:

$${\displaystyle \frac{\partial {GIE}_{i,t}}{\partial {EDQ}_{i,t}}}={\beta }_1+{\beta }_3^{OC}·{OC}_{i,t}$$

(20)

At low ownership concentration (25th percentile, 22.4 percent), the EDQ slope reaches 0.0058; at high concentration (75th percentile, 48.7 percent), it falls to 0.0021. Concentrated controlling shareholders appear to repurpose high-quality disclosure as a legitimacy device rather than a commitment to abatement-focused R&D, which lines up with the entrenchment account in H2b. The result is one of the cleaner findings in our analysis—the sign is robust across both random-intercept and random-slope estimators, and the magnitude does not shrink when SOE status is partialed out.

The environmental executive moderator yields the largest amplification effect of the four. With ENVEXE as a binary indicator, the interaction term takes ${\eta }_3=0.0034$ ($SE=0.0010$, $p<0.01$), and the slope differential is given by

$$\Delta \left({\displaystyle \frac{\partial GIE}{\partial EDQ}}\right)={\eta }_3·[ENVEXE=1-ENVEXE=0]=0.0034$$

(21)

Firms with an environmentally experienced executive in the TMT convert disclosure quality into innovation efficiency at roughly 1.7 times the rate of firms without such experience. The upper-echelons mechanism appears to operate as theorized: cognitive priors that privilege long-horizon environmental payoffs reduce the friction between outward-facing disclosure and inward-facing R&D portfolio adjustments [61]. This effect remains stable after controlling for whether the firm is in a pollution-intensive industry, suggesting that it is genuinely an executive-level rather than industry-level phenomenon.

Institutional shareholding moderates positively, but with a smaller magnitude. The interaction coefficient comes in at ${\theta }_3=0.009$ ($SE=0.004$, $p<0.05$), and the marginal effect follows:

$${\displaystyle \frac{\partial {GIE}_{i,t}}{\partial {EDQ}_{i,t}}}={\theta }_1+{\theta }_3·{INST}_{i,t}$$

(22)

Across the interquartile range of INST (from 12.3 percent to 41.6 percent), the conditional EDQ slope rises from 0.0044 to 0.0070. The amplification is real but more muted than the executive background channel, perhaps because aggregate institutional ownership conflates long-horizon and transient holders—a distinction that future work could sharpen by separating pension and social security funds from short-term mutual funds [62]. We flag this as a measurement limitation rather than a refutation of H2d.

Figure 7 places the four marginal effects on a common axis to make the contrast visible.

Ranking the moderators by the magnitude of the marginal-effect shift across their respective ranges, environmental executive background and ownership concentration emerge as the most influential, followed by institutional shareholding and then board independence. Hypotheses H2a–H2d each find support, though the strength of the effects diverges enough to caution against treating governance as a homogeneous bundle.

4.3. Cross-Level Interaction Tests and Robustness Analysis

Moving from firm-level moderation to the cross-level structure laid out in Equation (16), we test whether industry regulation intensity and provincial marketization condition the disclosure–efficiency slope. The cross-level interaction with REG returns ${\gamma }_{110}=0.0018$ ($SE=0.0006$, $p<0.01$), so the conditional EDQ slope can be written as

$${\displaystyle \frac{\partial {GIE}_{ijk,t}}{\partial {EDQ}_{ijk,t}}}={\gamma }_{100}+{\gamma }_{110}·{REG}_{jk}$$

(23)

Industries above the 75th percentile of regulatory intensity—thermal power, cement, non-ferrous metals—exhibit EDQ slopes roughly 1.8 times those in lightly regulated sectors, lending empirical support to H3a. The regulatory environment evidently raises the informational stakes of disclosure: where misreporting carries steeper administrative consequences, the same disclosure score conveys a stronger signal and travels farther into capital allocation decisions [63].

Provincial marketization carries an analogous amplifying role. The interaction coefficient lands at ${\gamma }_{101}=0.0015$ ($SE=0.0005$, $p<0.01$), with the conditional slope:

$${\displaystyle \frac{\partial {GIE}_{ijk,t}}{\partial {EDQ}_{ijk,t}}}={\gamma }_{100}+{\gamma }_{101}·{MKT}_k$$

(24)

Firms domiciled in provinces in the top quartile of the Fan–Wang index—Guangdong, Zhejiang, Jiangsu—translate disclosure quality into innovation efficiency at roughly twice the rate of firms in bottom-quartile provinces. The mechanism follows the institutional logic in H3b: thicker factor markets, more competitive intermediary services, and stronger property-rights protection [37] let credible disclosure flow through green credit channels and technology-licensing markets without dissipating in administrative friction [64]. Deviance tests against the random-slope-only specification yield $\Delta D=76.3$ ($df=4$, $p<0.001$), so adding the cross-level interactions substantially improves fit.

A natural concern is endogeneity between disclosure quality and innovation efficiency: firms with more productive green R&D pipelines may invest in better disclosure precisely because they have favorable outcomes to report. We flag this at the outset because it is central to the entire claim rather than a loose end—the multilevel slopes reported above are only as credible as the exogeneity of EDQ—and we now confront it directly through three robustness exercises and one identification strategy.

First, we replaced the content-analysis disclosure index with the third-party Bloomberg ESG environmental pillar score for the subsample with available coverage. The EDQ coefficient remains positive and significant at 0.0041 ($SE=0.0011$, $p<0.01$), within 11 percent of the baseline estimate. Second, restricting the panel to 2017–2024 to exclude the pre-policy years yields a slightly larger coefficient of 0.0052, suggesting the relationship may have strengthened in the post-Paris regime rather than weakened—a pattern not inconsistent with the institutional accumulation argument advanced earlier. Because disclosure plausibly acts on innovation efficiency with a delay rather than contemporaneously, we also re-estimated the baseline with one- and two-year lags of EDQ in place of its current value; both lagged coefficients stay positive and significant (0.0049 at one lag, 0.0043 at two), and the gentle decay across horizons is consistent with a signal that is absorbed gradually into capital allocation rather than with a mechanical same-year correlation.

Third, propensity score matching addresses self-selection into high-disclosure status. We split firms into high-EDQ (top tercile) and low-EDQ groups, then matched on Size, Lev, ROA, Age, SOE, and industry-year fixed effects with one-to-one nearest-neighbor matching and a caliper of 0.05. The average treatment effect on the treated comes in at $ATT=0.073$ ($SE=0.021$, $p<0.01$), implying a 7.3-point efficiency gain attributable to high disclosure status after controlling for selection on observables:

$$ATT=E\left[{GIE}_i\right|D=1]-E[{GIE}_i|D=0,matched]$$

(25)

Balance diagnostics post-matching show standardized mean differences below 5 percent on all matching covariates, down from pre-matching gaps that reached 31 percent on firm size, and the post-match variance ratios all fall inside the 0.5–2.0 band conventionally taken as adequate. The estimated propensity scores for the high- and low-EDQ groups overlap across almost the entire unit interval, so the common-support condition is satisfied without trimming more than the eleven treated observations whose scores exceeded the maximum control score; these were dropped under the imposed caliper. The matched ATT, the balance summary, and the off-support counts are reported compactly in

Table 8. Robustness and endogeneity checks for the EDQ coefficient.

Specification	EDQ Coefficient	Std. Error	Diagnostic
Baseline (full HLM)	0.0046 ***	0.0008	—
Alt. measure: Bloomberg ESG	0.0041 ***	0.0011	within 11% of baseline
Restricted window 2017–2024	0.0052 ***	0.0020	post-policy subsample
One-year lag of EDQ	0.0049 ***	0.0009	lagged specification
Two-year lag of EDQ	0.0043 ***	0.0010	lagged specification
PSM (ATT, score points)	0.073 ***	0.021	all SMD < 5%; 11 off-support dropped
IV (2SLS)	0.0058 ***	0.0019	Kleibergen–Paap F = 47.3

*** p < 0.01. The PSM estimate is the average treatment effect on the treated, expressed in efficiency-score points; all other entries are slope coefficients on EDQ.

rather than as a separate figure, since the diagnostics add rows rather than a new visual.

Fourth, and most demanding, we instrumented EDQ with the lagged industry-province average disclosure score excluding the focal firm, following a leave-one-out construction that should satisfy the exclusion restriction conditional on industry-year and province-year fixed effects. We do not treat that exclusion claim as self-evident. The threat is that peer disclosure proxies for a common industry-region shock—a regulatory wave or a green-credit pilot—that drives innovation efficiency directly; our defense is that the saturated industry-year and province-year fixed effects absorb exactly those common shocks, so the residual variation in peer disclosure that identifies the coefficient is the idiosyncratic peer-reporting behavior that has no obvious direct channel to the focal firm’s R&D conversion. As a partial falsification, we re-ran the second stage adding contemporaneous industry green-credit volume and provincial environmental-enforcement intensity as controls; the instrumented EDQ coefficient moved by less than 8 percent, which is the pattern one expects when the instrument is not merely picking up a shared policy environment. The leave-one-out instrument is specified as

$${EDQ}_{ijk,t}={\pi }_0+{\pi }_1{\overline{EDQ}}_{-i,jk,t-1}+{\pi }_2{X}_{ijk,t}+{\omega }_{ijk,t}$$

(26)

The first-stage Kleibergen–Paap F-statistic of 47.3 comfortably exceeds the Stock–Yogo weak-instrument threshold, and the second-stage coefficient on instrumented EDQ is 0.0058 ($SE=0.0019$, $p<0.01$). The IV estimate exceeds the OLS-style baseline, which is consistent with classical measurement error in the disclosure score attenuating the naive coefficient rather than with reverse causation inflating it [65].

Figure 8 collects the baseline, two cross-level interactions, and four robustness coefficients on a single axis.

Across the seven specifications, the sign never flips and the magnitude varies within a narrow band of 0.0041 to 0.0073, leaving the central finding intact under each diagnostic challenge. To make the verdict on each hypothesis legible at a glance,

Table 9. Summary of hypothesis tests.

Hypothesis	Model/Equation	Coefficient (Sign)	Significance	Conclusion
H1: EDQ → GIE (+)	Baseline RI, Equation (11)	${\gamma }_{100}>0$	$p<0.01$	Supported
H2a: IND amplifies	Equation (3)	${\alpha }_3>0$	$p<0.01$	Supported
H2b: OC dampens	Equation (20)	${\beta }_3<0$	$p<0.01$	Supported
H2c: ENVEXE amplifies	Equation (4)	${\eta }_3>0$	$p<0.01$	Supported
H2d: INST amplifies	Equation (5)	${\theta }_3>0$	$p<0.05$	Supported
H3a: REG steepens slope	Equation (16)	${\gamma }_{110}>0$	$p<0.01$	Supported
H3b: MKT steepens slope	Equation (16)	${\gamma }_{101}>0$	$p<0.05$	Supported

H1 is supported as an overall association; the financing, reputation, and monitoring channels are interpretive and were not separately tested as mediators.

links every proposition from H1 through H3b to its governing model, the sign and significance of the relevant coefficient, and the resulting conclusion on support.

5. Discussion

Our central finding—that environmental disclosure quality is associated with the input–output efficiency of green innovation rather than with its sheer volume—diverges from much of the existing literature in two respects. Prior studies relying on binary CSR-report indicators or simple word counts tend to report weak or noisy disclosure effects, sometimes attributing them to symbolic compliance. By scoring disclosure across content completeness, verifiability, and forward-looking commitment, we recover an association that is both statistically robust and economically meaningful. The transmission, we should stress, is interpreted rather than directly estimated: the three overlapping channels developed in Section 2.1—easier access to green-labeled credit that reduces the financing wedge on long-horizon R&D, accumulated reputational capital that improves collaborations with universities and downstream buyers, and intensified external scrutiny that disciplines internal resource allocation—are the explanations we find most credible, but the present design does not partition the effect among them. None of these channels need to operate in isolation. We read the joint movement of all three as the most plausible source of the efficiency association—consistent with the idea that disclosure works not by funding more projects but by reshaping which projects survive the internal selection process.

The governance moderation results contain the paper’s most uncomfortable finding. Ownership concentration suppresses the disclosure–efficiency conversion. This stands in mild tension with the often-cited view that concentrated owners shoulder long-horizon environmental investments. Our reading is that in the Chinese institutional setting, concentrated stakes more often coincide with controlling-shareholder entrenchment than with stewardship, and the latter dynamic dominates in the aggregate. Independent directors and environmentally experienced executives, by contrast, both steepen the slope, but through different mechanisms: independence raises the cost of disclosure–action decoupling, while executives’ environmental background reduces the cognitive friction of redirecting R&D portfolios. The amplification by environmentally experienced executives turns out to be the largest amplifying effect among the four moderators, which strikes us as the most actionable finding from a board-composition standpoint.

Cross-level interactions sharpen the picture further. Industry regulation intensity and provincial marketization both amplify the firm-level slope, but the institutional logic differs. Regulation works through raising the cost of misrepresentation, so disclosure carries more informational weight where the enforcement apparatus is dense. Marketization works through transmission: even credible signals require thick capital and technology markets to convert into resource reallocation. Firms in lightly regulated industries or thinly marketized provinces face a kind of double discount on the value of their disclosure investment, regardless of internal effort. The pattern complicates any one-size-fits-all disclosure mandate.

Stepping back, the methodological lesson concerns variance decomposition. Roughly a third of the total variance in green innovation efficiency sits at industry and provincial layers combined, a magnitude that single-level fixed-effects estimators absorb but do not surface. Pooled OLS would have delivered a coefficient on EDQ that was directionally correct but stripped of the conditional structure that gives it managerial meaning. The HLM specification did not change the headline result so much as it dissolved the false impression that the disclosure–efficiency relationship is uniform across contexts.

The implications branch out in three directions. For corporate management, the priority is not maximizing disclosure word counts but pairing credible disclosure with governance arrangements that absorb the signal—independent boards with industry expertise, executives with environmental fluency, and ownership structures that resist entrenchment. For environmental regulators, the differential effectiveness across regulation intensities argues for tightening mandatory disclosure standards in lightly regulated industries before declaring victory in the heavily regulated ones, where the marginal informational gain is already saturated. For capital market design, the marketization finding suggests that disclosure rules without complementary green finance infrastructure—certified verifiers, transparent carbon pricing, transition-finance taxonomies—will deliver attenuated returns. Each of these implications has limits; none should be read as a universal prescription. But the multilevel evidence at least clarifies the conditions under which each holds.

6. Conclusions

This study set out to map how environmental disclosure quality, conditioned by governance arrangements and embedded in industry and regional contexts, shapes the productivity of green innovation in Chinese listed firms. Using a three-level hierarchical specification on a 2014–2024 panel of 2847 A-share firms, we recover three findings worth restating, stated as the conditional associations that the design can support rather than as established causal effects. Higher disclosure quality is associated with greater input–output efficiency of green innovation, a link for which we advance eased financing constraints, accumulated reputational capital, and intensified external monitoring as the most plausible explanations rather than as separately tested mediators. Governance characteristics moderate this conversion in heterogeneous ways: board independence, environmentally experienced executives, and institutional shareholding amplify the slope, while concentrated ownership dampens it. Industry regulation intensity and provincial marketization further steepen the firm-level relationship through distinct institutional channels—the former by raising the cost of misrepresentation, the latter by thickening the markets through which disclosure signals travel.

The theoretical contribution sits at the intersection of three strands of literature that have rarely been integrated. By embedding disclosure, governance, and green innovation efficiency within a single multilevel framework, the paper recasts what looked like an average treatment effect into a conditional structure shaped by nested institutional layers. Measurement-wise, the multidimensional disclosure index moves the field beyond binary CSR indicators; methodologically, the random-slope HLM partitions variance that single-level estimators absorb invisibly. Practically, the findings argue for pairing mandatory disclosure rules with complementary governance reforms and green finance infrastructure rather than treating each policy lever in isolation. They also offer a sharper diagnostic for boards: independent directors with industry expertise and executives carrying environmental experience are the governance configurations most likely to convert disclosure investment into measurable abatement payoffs.

Several limitations qualify these conclusions, and three deserve to be stated plainly rather than buried. The sample is confined to Chinese A-share listed firms, a population that is larger, more visible, and more disclosure-pressured than the private and unlisted firms that make up most of the economy; the disclosure–efficiency association we document may therefore be stronger here than it would be where reporting incentives are weaker, and we caution against reading the magnitudes as economy-wide. The instrumental-variable identification, although it survives the falsification checks reported above, still rests on an exclusion restriction that cannot be fully proven—peer disclosure could in principle reach focal-firm efficiency through channels our fixed effects do not capture—so the IV estimate is best read as corroborating rather than clinching. Finally, the cross-level moderators we identify are calibrated to a single institutional setting; whether industry regulation intensity and regional marketization play the same amplifying role in mandatory-disclosure regimes such as the EU, or in less marketized developing economies, is an open question that only cross-national replication can settle. The content-analysis index, despite high inter-rater agreement, also remains coarser than what text-mining or large language models could now extract from full report corpora.

Three avenues for follow-up work appear most promising. Replacing manual coding with NLP-based disclosure scoring would scale the measurement and surface dimensions—tone, hedging, comparability—that human rubrics miss. Cross-national comparison, particularly between Chinese and EU ETS-covered firms, would test whether the cross-level moderators we identify generalize beyond the single institutional setting. Dynamic mechanism analysis through state-space or panel VAR specifications could trace how the disclosure–efficiency relationship evolves with policy regime shifts, separating short-run signaling effects from long-run institutional accumulation.