The Impact of the Digital Economy on Urban-Rural Income Inequality: A Meta-Analysis of 63 Empirical Studies in China ()
1. Introduction
In the course of China’s modernization, the digital economy has become an important force reshaping production, factor allocation, and income distribution [1] [2]. Under the national goals of building a “Digital China” and promoting common prosperity, understanding whether and how digital development affects the urban-rural income gap has gained increasing theoretical and practical significance. Although China has made substantial progress in digital infrastructure, digital industrialization, and institutional reform, income disparities across urban and rural areas, regions, and social groups remain pronounced. In particular, differences in digital access, digital skills, platform participation, and factor mobility continue to limit the equal distribution of digital dividends.
Existing research has not reached a consistent conclusion on the relationship between the digital economy and income inequality. Some studies argue that digital development narrows the urban-rural income gap by reducing information and transaction costs, expanding employment and entrepreneurial opportunities, and improving access to public services [3]-[5]. Others suggest that technology bias, platform concentration, and unequal digital capabilities may instead widen inequality or generate nonlinear effects. These divergent findings are closely related to differences in measurement strategies, administrative levels, samples, and econometric methods. They also reflect a broader reality: improved digital access does not automatically translate into effective use, and increased use does not necessarily lead to equal income gains.
Against this backdrop, this study conducts a meta-analysis of empirical research on the digital economy and income inequality in China. Based on 63 studies and 76 effect sizes, we standardize reported estimates and synthesize them under a random-effects framework. We further examine publication bias, heterogeneity, robustness, subgroup differences, and meta-regression moderators. By integrating fragmented evidence within a unified framework, this study aims to provide a more reliable estimate of the average effect of the digital economy on the urban-rural income gap, identify major sources of heterogeneity, and offer evidence-based support for policies aimed at common prosperity and high-quality development.
2. Literature Review
2.1. Theoretical Foundations
Existing studies mainly explain the relationship between the digital economy and income inequality from two perspectives: digital inclusion and technology bias. The digital inclusion perspective argues that improvements in digital infrastructure, digital inclusive finance, and platform-based economic activities can reduce information and transaction costs, enhance resource allocation efficiency, and expand disadvantaged groups’ access to markets and public services, thereby helping narrow income disparities [6]-[8]. In contrast, the technology bias perspective emphasizes that the digital economy is often capital- and skill-biased, which may reinforce platform concentration and “winner-takes-all” effects, and thus widen income gaps [9]. From the perspective of the “access-use-benefit” framework, improved access does not automatically lead to effective use, nor does increased use necessarily translate into equal income gains. Therefore, the digital economy may generate both “digital dividends” and “digital divides,” and its actual distributional effect remains an empirical question.
2.2. Empirical Evidence to Date
Empirical studies on China can generally be divided into three strands. The first strand consists of macro-level studies, which mainly use provincial or cross-regional panel data to examine the impact of the digital economy on the urban-rural income gap. For example, Jiang et al. [10] find that the digital economy helps narrow the gap by promoting industrial upgrading and improving public service equalization. In contrast, Guo [11] argues that regional differences in digital development may instead intensify spatial disparities [12] [13].
The second strand focuses on meso-level analyses at the city and county levels. These studies often find a stronger inequality-reducing effect. For instance, Deng and Jin [14] show that digitalization improves factor allocation efficiency and labor-market matching, thereby helping reduce income disparities. However, Tu et al. [15] point out that in some less-developed countries, weak infrastructure and talent outflows may limit this effect [16]-[19].
The third strand includes micro-level studies based on household or firm data. Most of them suggest that digital participation can increase rural income and expand employment and entrepreneurial opportunities. For example, Li et al. [20] find that internet access and digital use raise rural household income, although the benefits vary across different groups [21] [22].
Overall, while the literature has increasingly adopted methods such as instrumental variables, difference-in-differences, and spatial econometric models, conclusions on the direction and magnitude of the digital economy’s effect remain inconsistent.
2.3. Synthesis of the Literature
Overall, the existing literature has provided important insights into the relationship between the digital economy and income inequality, but several limitations remain. First, empirical findings are still highly inconsistent, making it difficult to draw clear policy implications. Second, most studies focus on a single administrative level or a specific region, and systematic cross-level comparison is limited. Third, insufficient attention has been paid to the sources of heterogeneity, such as differences in measurement strategies, spatial scales, and identification methods. Fourth, publication bias and the overall robustness of the accumulated evidence have rarely been examined in a systematic way. In light of these limitations, a meta-analysis using unified effect-size measures is necessary to synthesize existing findings, estimate the average effect size more reliably, and identify the main sources of heterogeneity across studies.
3. Research Design
3.1. Data Sources and Screening Criteria
To ensure both reproducibility and broad coverage, this study constructed its sample by combining Chinese- and English-language sources, journal articles, and grey literature, and database searches with backward and forward citation tracing. Chinese-language studies were mainly collected from CNKI and Wanfang Data, while English-language studies were identified through the Web of Science Core Collection and Google Scholar. Additional relevant studies were located through reference-list checks and citation tracing.
The search strategy covered three dimensions: topic domain, outcome variables, and methodological features. Topic-related keywords included digital economy, digital transformation, digital infrastructure, industrial digitalization, digital inclusive finance, and digital rural development. Outcome-related terms included income inequality, urban-rural income gap, rural-urban income disparity, and common prosperity. Methodological keywords included empirical, impact, mechanism, regression, panel, threshold, spatial, endogeneity, instrumental variables, 2SLS, GMM, DID, and PSM. After merging and removing duplicates, we conducted a multi-stage screening process including title-abstract screening and full-text review.
Studies were included if they: 1) focused on China or specific regions within China; 2) directly examined the relationship between the digital economy or its subdimensions and income inequality, the urban-rural income gap, or common prosperity; and 3) reported sufficient statistical information for effect-size conversion. Studies were excluded if they were purely theoretical, lacked usable quantitative information, were inconsistent with the research topic, or duplicated the same dataset.
3.2. Coding and Variables
Based on the studies that passed the screening process, we constructed a database in which each entry corresponds to a single study-effect size pair, ensuring statistical independence across observations. The dependent variable was uniformly defined as various measures of income inequality, and we recorded the specific metric used in each study, including the Gini coefficient, Theil index, urban-rural income ratio, quantile income gaps, and logarithmic income differences. The key independent variables were standardized into four categories—composite digital economy indices, digital infrastructure, digital application measures, and digital finance—while also documenting their construction methods and data sources in detail.
To account for sources of heterogeneity, we systematically coded the following study characteristics: research level (province, city, county, or micro-individual), model type (OLS, fixed effects, two-way fixed effects, IV/2SLS, GMM, DID, spatial econometric models, threshold models, matching methods, etc.), whether endogeneity was addressed (endo), whether the independent or dependent variables were log-transformed (logx, logy), sample coverage period, and categories of control variables. Where available, we additionally recorded mechanism variables—such as human capital, employment structure, factor allocation efficiency, and digital divide indicators—for potential extended analyses.
For effect-size computation, the main analysis employed Fisher’s z-transformed correlation coefficients. For studies directly reporting correlation coefficients (r), we applied Fisher’s transformation and weighted observations using v = 1/(n − 3). For studies reporting t-statistics and degrees of freedom (df), we first converted the estimates into correlation coefficients using the standard formula
, and subsequently applied Fisher’s z-transformation to obtain a unified effect-size metric. For studies reporting only regression coefficients (β) and their standard errors, we treated them separately in a “β-based” subset for robustness comparison rather than combining them with z-based estimates. When sample sizes were not explicitly reported, but the variance of z was available, we inferred sample sizes using n ≈ 1/v + 3. Studies for which essential information could not be recovered were excluded or incorporated solely into qualitative discussions. In cases where multiple estimates were reported from the same underlying dataset, we selected one representative effect for the main analysis and used clustered robust standard errors when appropriate to account for within-study dependence.
3.3. Analytical Methods
The primary analysis was conducted within a random-effects framework. We first used Cochran’s Q and the I2 statistic to assess cross-study heterogeneity and estimated the between-study variance τ2 using restricted maximum likelihood (REML). We then computed the pooled effect size and its 95% confidence interval, complemented by a prediction interval to capture the plausible range of effects in future comparable studies. To further strengthen inferential robustness, we reported Knapp-Hartung-adjusted confidence intervals as an additional benchmark. The full analytical sequence followed standard meta-analytic procedures: heterogeneity testing, random-effects estimation, uncertainty quantification, and extended robustness checks.
Robustness analyses were carried out following standard procedures. We conducted leave-one-out and blockwise exclusion tests to examine the sensitivity of the pooled effect size and τ2 to influential studies; influence diagnostics were used to identify observations that disproportionately contributed to Cochran’s Q or to shifts in the pooled estimate. We further compared alternative estimators—such as REML and DerSimonian-Laird—to assess estimator stability. Publication bias was evaluated via Egger’s regression test and the trim-and-fill correction procedure, supplemented by numerical distribution inspection of effect sizes and standard errors; PET-PEESE regressions were further adopted as an extended robustness check to address small-sample and selective-reporting biases. Heterogeneity exploration relied on both subgroup analyses and meta-regression: the former compared pooled effects across indicator types (composite indices, infrastructure, applications, digital finance), administrative levels (provincial, municipal, county), and model types (OLS, fixed effects, IV, DID, spatial models, threshold models, matching methods), while the latter incorporated variables such as endo, region level, index type, log transformations, and sample period into the REML estimation, using CR2 cluster-robust standard errors for inference and, where appropriate, multilevel models or robust variance estimators to account for dependence among multiple effect sizes from the same study. Standard outputs—quantitative influence diagnostics results and a variable dictionary—were generated to ensure transparency and reproducibility.
4. Results
4.1. Data Overview and Descriptive Statistics
The final sample for this study includes 76 effect sizes suitable for quantitative synthesis (Table 1), extracted from multiple empirical studies examining the relationship between the digital economy and income inequality in China. All effect sizes were converted into Fisher’s z values and aggregated under a random-effects framework. Descriptive statistics show that the distribution of effect sizes centers around zero with moderate dispersion, and that sample sizes are markedly right-skewed. As shown in Table 1, the median sample size falls within a wide interquartile range, covering studies with both relatively small and very large samples, indicating substantial cross-study variation in sample scale. For studies that did not report sample sizes directly, we inferred them using n ≈ 1/v + 3 based on the variance of the z estimates. Overall, the number of effect sizes with available sample-size information is sufficient to support reliable weighting and uncertainty estimation in subsequent analyses.
Table 1. Descriptive statistics.
Statistic |
Value |
Number of effect sizes (k) |
76 |
Mean of yi (Fisher’s z) |
−0.021 |
Standard deviation of yi (Fisher’s z) |
0.301 |
Median variance vi |
0.00276 |
Median sample size n_median |
366 |
Interquartile range of sample size ([Q1, Q3]) |
[300, 2492] |
Sample size range [min, max] |
[11, 72,356] |
Sample size available/missing (number of effects) |
76/0 |
Effect sizes distribute roughly symmetrically around zero without extreme outliers, while sample sizes follow a right-skewed distribution, a typical feature of meta-analysis integrating multi-source empirical literature. Overall, effect values cluster near zero, which preliminarily implies modest average treatment effects and substantial cross-study heterogeneity. This pattern arises from multi-dimensional differences in digital economy measurement, research context, and variable construction across studies, laying a foundation for subsequent heterogeneity diagnosis and pooled effect estimation. The overall data quality satisfies the standard prerequisites for meta-analytic modeling.
4.2. Heterogeneity Tests
To assess the consistency of estimates across studies, we first calculated Cochran’s Q statistic and its
significance using fixed-effect weights wi = 1/vi, and subsequently derived the heterogeneity indices I2 and H2. The Q statistic is substantially larger than its degrees of freedom, with p < 0.001, indicating pronounced heterogeneity (Table 2). The associated I2 value also suggests a high proportion of total variance attributable to between-study differences rather than sampling error. Given this high cross-study heterogeneity, we conduct additional influence diagnostic analyses in Section 4.5, alongside subgroup analyses and meta-regression in Sections 4.6 and 4.7 to further explore potential sources of heterogeneity.
Table 2. Heterogeneity test.
Statistic |
Value |
k |
76 |
Q |
2136.911 |
df |
75 |
p-value |
0 |
I2 (%) |
96.5 |
H2 |
28.492 |
tau2_DL |
0.016 |
4.3. Overall Effect Size
Given the presence of significant heterogeneity, we estimated the pooled effect size using a random-effects model (REML). The results show that the combined correlation between the digital economy and income inequality is r = −0.030, with a 95% confidence interval of [−0.090, 0.030] (Table 3). The corresponding prediction interval (PI) is considerably wider, ranging from [−0.496, 0.450], indicating substantial variability in the potential effect sizes of future comparable studies. Applying the Knapp-Hartung adjustment yields a confidence interval of [−0.093, 0.034], which is consistent in both direction and magnitude with the baseline estimate. Individual effect coefficients extracted from each study fluctuate widely and center near zero, consistent with the high cross-study heterogeneity identified in Section 4.2. In sum, the pooled estimate under the main specification does not provide statistically significant evidence that the digital economy systematically reduces income inequality. However, the direction and magnitude of the effects vary across different samples and model settings—a pattern that will be explored in greater detail in Sections 4.5 - 4.7 through robustness checks, subgroup analyses, and meta-regression.
Table 3. Overall pooled effect.
Statistic |
Value |
Pooled method |
Random effects (REML) |
Effect-size metric |
Correlation coefficient (r) |
Number of effect sizes (k) |
76 |
Overall pooled effect (r̂) |
−0.030 |
95% confidence interval |
[−0.090, 0.030] |
Knapp-Hartung 95% CI |
[−0.093, 0.034] |
Prediction interval (PI) |
[−0.496, 0.450] |
Although the overall effect points in a negative direction, its magnitude is small and statistically insignificant. This pattern aligns with the mixed findings in existing research, where some studies report inequality-reducing effects, others find inequality-widening effects, and still others detect no significant relationship. [23]-[26] According to our comprehensive analysis, these discrepancies stem primarily from differences in spatial scale, measurement strategies, and identification methods rather than from fundamental contradictions in effect direction. The near-zero average effect observed after harmonizing effect sizes reflects the layered and threshold-dependent nature of the digital economy’s influence: improvements in digital access do not automatically translate into effective usage, and increased usage does not necessarily map linearly onto income gains. Breakdowns in this chain—whether due to digital skills, human capital, or platform governance—can weaken or offset expected benefits. Thus, the small but directionally consistent average effect is theoretically plausible.
4.4. Publication Bias Tests
We conducted publication bias tests through formal statistical testing and correction procedures. Preliminary inspection of effect size and standard error distributions shows no obvious asymmetric clustering, providing initial evidence against severe small-study effects or publication bias.
Next, we conducted Egger’s regression test to formally detect asymmetric distribution between effect sizes and standard errors. The results show z = 1.449 and p = 0.147, indicating that the test is not statistically significant at conventional levels. The estimated intercept as the standard error approaches zero is b = −0.0867 (95% CI: [−0.1843, 0.0110]), which is directionally consistent with the overall negative effect but not statistically distinguishable from zero. We further applied the trim-and-fill procedure to adjust for potential missing studies. The number of imputed studies was zero, and the adjusted pooled effect remained at r = −0.030 (95% CI: [−0.090, 0.030]), identical to the baseline estimate.
Overall, the publication bias tests do not indicate the presence of significant bias, and the trim-and-fill adjusted results remain stable. This suggests that the small magnitude of the pooled effect is not driven by selective reporting or small-sample bias but is instead attributable to structural differences across studies. The stability of the adjusted estimates enhances confidence in the validity and reliability of the overall findings.
4.5. Robustness Analysis
To assess the sensitivity of the overall conclusions to extreme observations and model specifications, we conducted a series of robustness checks, including 1) influence diagnostics to identify high-influence studies; 2) re-estimating the pooled effect after removing influential observations; and 3) additional leave-one-out tests and parallel analyses based on alternative effect-size metrics. Diagnostic tests show several individual studies exert disproportionate impacts on the overall heterogeneity statistic Q and pooled correlation coefficient; quantitative influence diagnostics further screen out three high-leverage observations with large residual values, corresponding to study serial numbers 71, 74, and 61. After excluding these three studies, the random-effects model yields a pooled correlation of r = −0.047 with a 95% confidence interval of [−0.091, −0.002], compared with the baseline estimate of r = −0.030 (95% CI: [−0.090, 0.030]) (Table 4). This suggests that the direction of the effect remains unchanged but becomes slightly stronger and statistically significant once influential studies are removed, although residual heterogeneity still requires further investigation through subgroup analyses and meta-regression.
Table 4. Summary of robustness checks.
Test |
r̂ |
95% Confidence Interval |
REML |
−0.030 |
[−0.090, 0.030] |
Excluding high-influence studies {71, 74, 61} |
−0.047 |
[−0.091, −0.002] |
The results of the robustness checks demonstrate that the direction of the pooled effect remains stable across all sensitivity tests, with no reversals or contradictory findings emerging from the leave-one-out analyses. Parallel estimations using alternative metrics (such as β/SE-based effect sizes) or alternative interval estimation approaches also yield consistent directions and significance patterns, showing only minor and theoretically interpretable variations in magnitude. These findings collectively indicate that the overall effect of the digital economy on income inequality is relatively stable and not driven by isolated studies or methodological artifacts. The variations observed across models mainly reflect differences in sample composition and econometric specifications rather than randomness in estimation.
After excluding these three high-impact studies, the negative pooled effect becomes statistically significant. Parallel estimations with alternative effect-size metrics yield consistent directional results, confirming the robustness of our core findings. Cross-model discrepancies mainly stem from differences in sample composition and empirical design instead of random estimation errors.
4.6. Subgroup Analysis
Given the substantial overall heterogeneity, this section conducts subgroup analyses from two analytical perspectives—research design and measurement specifications—to explore potential sources of variation in effect sizes. First, when comparing results across administrative levels (province, city, county), the pooled estimates reveal that city-level studies produce a significantly negative combined effect (r = −0.066, 95% CI: [−0.113, −0.018]), whereas county-level (r = −0.030, 95% CI: [−0.114, 0.053]) and provincial-level estimates (r = −0.007, 95% CI: [−0.075, 0.062]) remain statistically insignificant. A fixed-effects test of between-group differences shows (QB = 52.358), df = 2, p < 0.001, indicating that effect magnitudes differ systematically across administrative scales. This suggests that city-level units—reflecting labor mobility, industrial clustering, and digital infrastructure more directly—are more likely to capture the inequality-reducing effects of the digital economy, whereas effects at broader or narrower spatial scales may be diluted due to aggregation or data limitations.
Next, the subgroup results based on whether endogeneity was addressed indicate that studies employing strategies to correct for potential reverse causality or omitted-variable bias exhibit a weak but statistically significant negative association (r = −0.037, 95% CI: [−0.071, −0.002]). In contrast, studies that do not address endogeneity report an insignificant estimate (r = −0.022, 95% CI: [−0.105, 0.061]). The between-group difference test yields (QB = 8.672), df = 1, p = 0.003, confirming a statistically meaningful difference between the two groups. These findings suggest that once endogeneity concerns are mitigated, the inequality-reducing effect of the digital economy becomes more evident, consistent with the results observed after removing high-influence studies in Section 4.5. This also reinforces the notion that identification quality materially affects effect-size estimates.
The subgroup results reveal that the negative effect of the digital economy is statistically significant primarily at the city level, whereas studies at the provincial and county levels do not yield significant findings. This pattern suggests that the influence of the digital economy is more detectable at spatial scales that align closely with labor market dynamics and industrial clustering, while aggregation at larger scales or data limitations at smaller scales may attenuate the observable effects.
Additionally, studies that account for endogeneity produce more stable negative estimates, indicating that controlling for potential reverse causality and omitted-variable bias sharpens the identification of the digital economy’s impact on income inequality. Together, these subgroup patterns explain a substantial portion of the heterogeneity observed in the literature and highlight the role of spatial granularity and identification strategies in shaping empirical conclusions.
4.7. Meta-Regression Analysis
To further explain the sources of substantial heterogeneity in effect magnitudes, we conducted meta-regression analyses using Fisher’s z-transformed effect sizes as the dependent variable, incorporating whether endogeneity was addressed (endo) and the administrative level of the study (region_level) as moderators. In the model specification, the reference group is defined as provincial-level studies that do not address endogeneity (endo = 0). Thus, the estimated coefficients capture the incremental effect relative to this baseline. To account for within-study clustering and potential inconsistencies in model specifications, we report both the REML estimates and cluster-robust (CR2) standard errors with Satterthwaite-adjusted degrees of freedom, supplemented by joint significance tests. Due to insufficient variation in indicator types (index_type) in this dataset, this variable was excluded from the final model.
The conventional REML output shows that the coefficient for endogeneity treatment (endo = 1) is −0.012 (SE = 0.072, p = 0.868), while the coefficients for city-level and county-level studies relative to the provincial baseline are −0.055 (SE = 0.066, p = 0.409) and −0.022 (SE = 0.118, p = 0.853), respectively; none of these are statistically significant. The residual heterogeneity estimates are τ2 = 0.0703 and I2 = 98.87%, and the joint test of moderators yields QM(3) = 0.765, p = 0.858, indicating that the model explains approximately 0% of between-study heterogeneity. This inconsistency between subgroup significance and insignificant meta-regression coefficients mainly arises from unmeasured confounding factors (such as indicator measurement methods and estimation strategies for digital economy indices) that cannot be fully captured by the two moderator variables used in this model.
Using CR2 robust standard errors produces consistent conclusions in both directions and significance: the joint significance test yields F(3, 20.69) = 0.221, p = 0.881. This indicates that, under the current moderator specification and sample coverage, endo and region_level do not significantly explain cross-study variation in effect sizes. When considered together with the subgroup findings in Section 4.6—where the city-level group and the endo = 1 group exhibit weak but statistically significant negative effects—it can be inferred that the within-group structural differences may be related to other unobserved study characteristics (such as measurement choices, model specifications, or instrument quality) that are not captured by the present meta-regression framework.
Based on these results, the core conclusions are as follows. The overall association between the digital economy and income inequality is mildly negative and statistically insignificant in the full sample (pooled r = −0.030, 95% CI [−0.090, 0.030]). The negative effect becomes more pronounced in city-level studies and in studies that employ stricter identification strategies by addressing endogeneity. At the meta-regression level, the linear combination of endo and region_level does not significantly reduce residual heterogeneity, suggesting that differences in effect magnitude stem more from research context and measurement choices than from changes in the overall direction of the effect. These findings are consistent with the strengthened results observed after removing high-influence studies (r = −0.047, 95% CI [−0.091, −0.002]) and with the absence of significant publication bias. Overall, the direction of the digital economy’s impact on income inequality is relatively stable, but the magnitude of the effect is highly context-dependent and cannot be fully explained by any single study characteristic.
Integrating the overall effect, subgroup differences, and meta-regression results reveals that the digital economy exerts a more distinct negative influence on income inequality at spatial scales closely aligned with labor market dynamics and industrial agglomeration. At larger or smaller statistical units, this influence tends to be diluted. This finding carries meaningful policy implications: from optimizing the spatial layout of digital infrastructure and improving the supply of digital skills to strengthening the digital application environment for small and medium-sized actors, policy priorities should be aligned with each region’s industrial structure and labor-mobility characteristics. Furthermore, incorporating income-gap indicators into the evaluation frameworks for digital-development initiatives, together with targeted attention to the digital-use capabilities of disadvantaged groups, can enhance the effectiveness of the digital economy in improving income distribution. These implications, grounded in the empirical evidence of this chapter, underscore the dependence of the digital economy’s impact on spatial scale and institutional conditions.
5. Conclusions
This study systematically synthesizes empirical evidence on the relationship between the digital economy and income inequality in China. The meta-analysis shows that the digital economy tends to reduce income inequality, but the average effect is modest and not statistically significant in the full sample. The effect becomes clearer after excluding influential studies and is more evident in city-level research and in studies with stronger identification strategies.
The substantial heterogeneity across studies suggests that differences in sample coverage, spatial scale, measurement, and methodology jointly affect the magnitude of estimated effects. Publication bias tests do not indicate serious bias, which supports the credibility of the overall findings. Taken together, the results suggest that the digital economy has a relatively stable inequality-reducing direction, but its actual impact depends heavily on specific institutional and spatial conditions.
These findings imply that policies should not only promote digital infrastructure and digital finance, but also improve digital skills, governance capacity, and equitable access to digital opportunities. Only by strengthening these complementary conditions can digital development be translated into more substantial improvements in income distribution and common prosperity.
Author Contributions
Wanxin Zhang designed the study, conducted the literature screening and data extraction, performed the meta-analysis and robustness checks, and drafted the original manuscript. Hongpeng Zhang and Xian’en Wang provided overall supervision and methodological guidance, refined the analytical framework and interpretation of results, secured project support, and reviewed and edited the manuscript. All authors contributed to the final version of the paper, approved it for publication, and agree to be accountable for all aspects of the work.
Funding
1) 2025 University-level College Students’ Scientific Research and Innovation Program Project: Research on Government-Enterprise Green Collaborative Governance Mechanism under the Dual Carbon Goal—Evidence from the Yangtze River Delta Region; 2) Zhejiang Culture Research Project: Science-Driven Marine Development—Key Bottlenecks and Strategic Breakthroughs for Zhejiang’s Marine Innovation Competitiveness (Project No.: 24WH02-2Z).
Acknowledgements
The authors would like to thank the editors and anonymous reviewers for their constructive comments, which helped improve the quality and clarity of this manuscript. We also acknowledge the support and research environment provided by Zhejiang Ocean University, as well as colleagues who offered helpful discussions during the development of this study. Any remaining errors are the responsibility of the authors.