Following the analytical framework proposed by Moser and Voena, this study constructs a quasi-natural experiment based on the NEDC program to evaluate the effect of growing digital infrastructure on ULGUE. We determine the causal policy effects using a differences-in-differences (DID) model. In 2011, a collaborative policy document titled “Guidelines on Launching National E-commerce Demonstration Cities,” issued by five ministries, including the NDRC, laid the foundation for the three-phase implementation of the NEDC strategy across 70 cities. These phases were launched in 2011, 2014, and 2017, respectively, and show a typical staggered promotion pattern. The NEDC policy has strong exogeneity and institutional shock characteristics. First, pilot cities were designated by the central government through top-down administrative assignment, which eliminates self-selection by local governments and reduces sample selection bias. Second, the clear temporal structure of the policy rollouts facilitates the construction of a progressive DID model. Third, the division between treatment and control groups is explicit, so policy effects can be captured through interaction terms.

To estimate the effect of the development of the digital economy, as represented by the NEDC policy, on ULGUE, the baseline regression model (Equation 1) is created as follows: L g u e i t = α 0 + α 1 D i d i t + ∑ ρ c o n t r o l i t + u i + λ t + ε i t , (1) where i and t denote city and year, respectively; α 0 is a constant term; u i and λ t represent area fixed effects and time fixed effects, respectively; and ε i t refers to the residual term. The efficiency of land green usage is represented by the explanatory variable ( L g u e i t ), and the creation of digital infrastructure is represented by the core explanatory variable ( D i d i t ).

Using Baron and Kenny’s traditional mediation analysis approach, this study systematically developed multiple regression equation sets to examine the role of mediating factors in the policy influence process. To enhance the statistical credibility of the results, Sobel tests and bootstrap repeated sampling techniques were introduced to verify the significance of the mediating pathways. The specific analysis paths were divided into the following three groups of models (Equations 2–4): L g u e i t = α 0 + α 1 D i d i t + ∑ ρ c o n t r o l i t + u i + λ t + ε i t (2) M i t = β 0 + β 1 D i d i t + ∑ ρ c o n t r o l i t + u i + λ t + ε i t (3) L g u e i t = γ 0 + γ 1 D i d i t + γ 2 M i t + ∑ ρ c o n t r o l i t + u i + λ t + ε i t . (4)

In this study,   M i t is set as a mediating variable to reflect green technology innovation or resource use efficiency. When regression model (4) produces a statistically significant coefficient α₂ for the mediating variable, and the coefficient α₁ associated with the core explanatory variable D i d i t decreases substantially or loses significance relative to model (2), this suggests that the mediating mechanism exists and that the effect of the policy variable on the dependent variable may be partially or fully transmitted through the mediating path.

This study combines the spatial econometric method proposed by Corrado and Fingleton (2012) to extend the traditional DID Model (Equation 5) by embedding spatial lag variables and spatial weight matrices, thereby constructing the Spatial Durbin DID Model (SDM-DID). This approach systematically identifies regional linkage and spatial spillover effects resulting from policy implementation. L g u e i t = γ 0 + ρ ∑ j W i j L g u e j t + γ 1 D i d i t + γ 2 ∑ j W i j D i d j t + γ 3 c o n t r o l i t + γ 4 ∑ j W i j c o n t r o l j t + u i + λ t + ε i t , (5) where ρ is the spatial autoregressive coefficient of ULGUE, W i j L g u e j t , W i j , and  W i j c o n t r o l j t represent the spatially lagged values of the respective variables, and W i j indicates the element of the spatial weight matrix that captures the spatial linkage between the samples.

The study focuses on ULGUE as the key variable, aiming to show the correlation between the logical distribution of input elements and the efficiency of outputs in urban land resource development (Xu et al., 2025). Regarding Tone’s proposed super-efficiency SBM model, a system has been developed for practical measurement.

In terms of input specification, this study includes four categories of factor variables. First, labor input is quantified using annual year-end employment figures, which reflect the total workforce engaged in productive activities, following the methodology of Wang et al. (2021). This serves as an indicator of human capital utilization. Second, capital input is adapted from Xu et al. (2025), using total fixed asset investment to represent material capital expenditure within the city. Third, science and technology investment is measured as the aggregate of research and development expenditures and educational funding allocations, reflecting the level of knowledge capital investment. Fourth, the energy dimension is represented by annual total water and electricity consumption, which serve as proxy indicators for resource consumption.

For output indicators, this study adopts per capita GDP, following the approach of He et al. (2024) and Wang et al. (2021), as a proxy for urban economic development. In response to the unanticipated outcomes, and in line with the method of Wang and Han, environmental effect is measured using three types of industrial emissions: wastewater, sulfur dioxide, and smoke dust. All fluctuating data are sourced from the China Urban Statistical Yearbook, which guarantees reliability, uniformity in breadth, and thoroughness across the time series.

Drawing on the methodological framework of Liu et al. (2023), this study uses a DID estimation strategy to assess the causal effect of NEDC policies on ULGUE. The main identification strategy constructs interaction terms between binary treatment indicators (policy intervention group) and temporal dummy variables (pre/post implementation). Specifically, if a city initiates the policy pilot in a given year, the urban area is included in the treatment group, with the treatment group identification variable set to one. At the same time, the temporal indicator variable is coded as one for all years after policy implementation. Cities not selected for the pilot program serve as the control group, with both treatment and time dummies equal to zero. This interaction term specification allows for rigorous estimation of the policy’s causal effects using the difference-in-differences framework.

Resource use efficiency. Following Zheng and Chen (2024), resource use efficiency is measured by the amount of energy consumption that corresponds to economic output per person. The calculation method converts different types of energy consumption at the urban level—including total electricity consumption, as well as artificial, natural, and liquefied petroleum gas supplies—into a standard coal equivalent (measured in 100 tons). This value is then divided by the region’s per capita GDP (measured in yuan) to obtain the energy-to-per capita output ratio. A natural logarithmic transformation is applied to this ratio to represent “resource use efficiency” and to control for the effect of outliers.

Green technology innovation (GTI). Following Fu’s (2024) methodological framework, this study includes GTI as a core explanatory variable, measured using city-level counts of green patent applications. The calculation method is as follows: GTI = ln (1 + number of green invention patents + number of green utility model patents). This approach applies a logarithmic transformation to the annual sum of green invention and utility model patents declared by each city. This metric effectively captures regional innovation outputs in green technology, providing both practical applicability and strong representativeness for policy evaluation.

The analysis includes several control variables in three important dimensions to reduce the possibility of omitted variable bias and increase the accuracy of model estimation: (1) Economic Development Level (InPGDP): According to He et al. (2024) and Wang et al. (2021), the logarithm of a region’s per capita real GDP represents its economic output capabilities. (2) Industrial Structure (IS): According to Lu et al. (2020), the degree of service sector dominance is gauged by the tertiary industry’s added value as a percentage of the regional GDP. (3) Environmental Regulation Strength (ERS): Referring to Berman and Bui, (1998), an indicator was constructed by obtaining the frequency of occurrence of environment-related terms in policy documents through text analysis methods. (4) Population density (InDENS): It is defined as the natural logarithmic value of the number of permanent residents per unit of land area (Liu et al., 2022). (5) Financial development level (FinDev): Following the idea proposed by Beck et al. (2000), financial resource allocation capacity is proxied by the share of loan balances held by financial institutions relative to GDP. Variable definitions are listed in Table 1.

A balanced panel data set is constructed, including 282 prefecture-level and above cities in China, with continuous annual data from 2007 to 2022. The study period is selected to balance methodological requirements and data availability. Although the National E-commerce Demonstration City (NEDC) policy was formally introduced in 2011, the analysis extends the starting year to 2007 to provide an adequate pre-policy period, which meets the parallel trends assumption of the DID model. The end year is set as 2022 because it is the most recent year with complete and consistent city-level data, which allows for an assessment of the medium-to long-term policy effects.

To ensure data completeness and consistency in administrative divisions, the final sample includes 67 pilot cities with complete observations—23 from the first batch, 29 from the second, and 15 from the third—as well as 215 non-pilot cities that were never included in the policy. The main data sources are the China Urban Statistical Yearbook, the China Energy Statistical Yearbook, and statistical bulletins issued by cities at the prefecture level. Missing values for certain years are addressed using group mean imputation or linear interpolation, which aims to improve data integrity and reduce estimation bias. Table 2 displays descriptive statistics for the primary variables.

Additionally, Figure 2 shows the trajectory of ULGUE for various categories during 2007–2022. In terms of trend, ULGUE demonstrates a continuous upward development for the whole country, model cities, and non-model cities. Analyzing various urban clusters indicates a consistent superiority in the efficiency of green land use, with demonstration cities consistently outperforming non-demonstration cities throughout the research period. Between 2011 and 2022, the ULGUE of demonstration cities increased from 0.406 to 0.572, an increase of 40.726%, whereas the growth rate of this metric in cities without demonstrations during that period was 14.316%. The observed disparities suggest a causal linkage to the NEDC policy intervention. Pilot cities show significantly stronger performance gains in green land use intensification, while the divergence from non-pilot counterparts progressively widens.

We conduct robustness checks through variable substitution and model extension:

First, regarding the explanatory variables, we introduce three representative indicators of communication infrastructure: Internet penetration, 5G population density, and long-distance cable length per unit area. Although existing empirical studies rarely apply “5G base station population density” directly to urban land green use efficiency, Yang et al. (2018) from the communication engineering field show that base station density is a key factor affecting network performance, especially in urban deployment scenarios, where it determines the transition from noise-limited to interference-limited regimes. Based on this theoretical foundation, we treat 5G population density as a structural proxy for digital infrastructure and include it in the robustness analysis to strengthen the engineering logic behind our green land use model. Following the methodologies of Wang and Wang (2024) and El-Garaihy et al. (2022), we also use Internet penetration rate and fiber-optic cable density (km/km²) as key infrastructure indicators to examine the consistency of our core conclusions. As shown in Columns (1)–(3) of Table 4, all three infrastructure indicators have significantly positive coefficients at the 1% significance level (p < 0.01). The signs and magnitudes of the control variables remain consistent with theoretical expectations, and the model’s explanatory power (R ²) varies within ±0.03, which indicates the robustness of the baseline findings across alternative digital economy measures. Second, for the dependent variable, we extend the original Super-SBM model. In addition to the initial specification that includes only the “three industrial wastes” as undesirable outputs, CO₂ emissions are also included as an additional undesirable output to recalculate ULGUE. The regression results reported in Column (4) of Table 4 show that the effect of digital economy development remains positive and statistically significant at the 5% level, which provides further support for the external validity of the core model.

Taken together, these robustness checks, which include both the substitution of explanatory variables and the alternative estimation of the dependent variable, provide strong evidence that digital infrastructure construction significantly contributes to improving ULGUE. Therefore, these results reinforce the reliability and validity of our main conclusions.

This study adopts Guo and Zhang (2024)’s causal inference framework to address endogeneity, focusing on how e-commerce policies affect urban development, and applies the instrumental variable (IV) approach of Li et al. (2016) based on historical communication facilities. The two-stage least squares method (2 S LS) is used. Specifically, we employ the correlation between post office counts and Internet user numbers in 1984 as the key instrument, reflecting early infrastructure foundations of digitalization.

To justify exogeneity, we argue that postal density in 1984 mainly reflected administrative layout and communication demand, not current determinants of ULGUE. At that time, China’s land market was undeveloped and the notion of ULGUE did not exist, making a direct effect implausible. Our specification also controls for structural city characteristics such as economic development, population density, industrial structure, and financial development, reducing concerns that postal density captures persistent unobserved traits. From a theoretical standpoint, 1984 postal density is unlikely to correlate with pre-treatment ULGUE, supporting the exclusion restriction.

The instrument’s strength is confirmed by the first stage: the IV coefficient is 0.126 (p < 0.01), with an F-statistic of 158.049, well above the threshold of 10. The Kleibergen–Paaprk LM statistic (p = 0.000) rejects weak-instrument and under-identification concerns. In the second stage (Table 5, Column 2), the policy variable remains positive and significant (p < 0.01). After addressing endogeneity, digital economy development continues to enhance ULGUE, indicating robust results.

Overall, combined with strong statistical evidence and consistency with prior studies using historical infrastructure as instruments (Li et al., 2016; Guo and Zhang, 2024), the instrument is both relevant and exogenous, reinforcing the reliability of the baseline estimates.

To address potential non-randomness in sample selection, this paper adopts the Propensity Score Matching–Difference-in-Differences (PSM-DID) method following Moser and Voena (2012). A binary variable indicates city participation in the NEDC program (1 for pilot, 0 otherwise). Control variables are included as covariates in a Logit regression to estimate propensity scores. Two strategies are applied: kernel matching, which weights all controls by score distance, and caliper matching with a width of 0.1, conducted with replacement to improve comparability.

Balance tests show that covariate means between treatment and control groups are statistically indistinguishable, with standardized bias below 10%, confirming good match quality. Re-estimated DID regressions on the matched samples (Table 6) indicate that, at the 1% level, NEDC construction continues to exert a significantly positive effect on urban development, supporting the robustness of the baseline findings.

Four supplementary tests were conducted to ensure the robustness of the results. First, Propensity Score Matching (PSM) was used to reconstruct the sample by matching treated and untreated cities through kernel and caliper matching methods, which mitigates sample selection bias (see Columns (1) and (2) of Table 6). The findings confirm the reliability of the baseline results by showing that the beneficial effect of digital economy growth on ULGUE is statistically significant. Second, to prevent estimate bias because of inherent locational and resource advantages, key cities—municipalities directly under the central government, provincial capitals, and sub-provincial cities—were excluded from the sample (see Column (3)). Third, the core explanatory variable was lagged by one period to account for the delayed effect of policy implementation and to reduce potential reverse causality (see Column (4)). Lastly, dummy variables for the smart city policy and the low-carbon pilot policy were included to control for interference from concurrent policies that may also affect ULGUE (see Columns (5) and (6)).

The digital economy development coefficient consistently shows significant positive effects in all robustness tests, confirming the stability and reliability of the baseline results. This identification strategy aligns with previous studies. For example, Liu et al. (2025) adopted comparable strategies, such as sample exclusion and lagged variable settings, and found that the NEDC initiative notably enhanced urban green performance. Similarly, Zheng and Chen (2024) used a spatial DID model and placebo testing to verify the favorable effect of low-carbon policies on ULGUE. Liu, Feng, and Wang (2022) highlight the role of green transformation policies in increasing land use efficiency, which supports our inclusion of policy control variables. Together, these studies provide strong theoretical and empirical support for our estimation approach and conclusions.

According to Table 7, once matching was performed, no covariate differences between the treatment and control samples reached statistical significance (p > 0.05), suggesting that the groups achieved adequate balance through the matching process.

Based on the counterfactual idea of placebo testing, this article refers to the study by Wang and Wang, (2024) to construct false policy regions for robustness testing. This study conducts a spatial placebo test by randomly assigning pseudo-treatment status across potential pilot cities. Figure 4 shows the distribution of estimated coefficients and corresponding p-values using kernel density plots, which are derived from 500 random replications of the treatment group assignment. The results show two key findings: first, the mean regression coefficient (0.032) is close to zero, which indicates no systematic treatment effect under randomization; and second, more than 90% of the simulated p-values are above conventional significance thresholds (p > 0.1). These patterns confirm that the observed policy effects are unlikely to result from chance spatial correlations, thus supporting the validity of our causal inferences.

These findings are highly consistent with those of Zheng and Chen (2024), who use a spatial DID model to evaluate the effect of China’s low-carbon city pilot policy on ULGUE. Their analysis includes a placebo approach to validate the robustness of the identified causal relationship. The results show that when the treatment status is randomly assigned, the statistical significance of the policy effect disappears. This further confirms that the original estimation is not driven by model specification or sample selection bias, but instead reflects a genuine policy effect. Both studies demonstrate that the placebo test, as a counterfactual-based robustness check, is an effective tool for enhancing the credibility of causal inference and strengthening the reliability of empirical conclusions.

Building on Zhu et al. (2023), this study divides the sample into Eastern, Central, and Western regions to test for regional variation in the digital economy’s impact on urban land green use efficiency (ULGUE). Table 13 shows uniformly positive and significant coefficients, confirming that digitalization enhances ULGUE overall but with spatial gradients. The Eastern region records the strongest effect (0.068, p < 0.01), reflecting advantages in infrastructure, institutional support, and green innovation resources. The Central region also demonstrates a significant effect (0.049, p < 0.01), suggesting that digitalization contributes to greener land use even under weaker conditions. By contrast, the Western region’s effect is weaker (0.039, p < 0.05), constrained by limited infrastructure, inefficient resource allocation, and restricted factor mobility. These findings align with Zheng and Chen (2024), who report that low-carbon city pilots deliver stronger benefits in the East due to solid economic and governance foundations.

Figure 6 further illustrates these regional disparities. Standardized coefficients are 0.217, 0.280, and 0.223 for the Eastern, Central, and Western regions, respectively, underscoring the heterogeneity of policy effects across China’s economic zones.

To evaluate whether policy timing influences the green effects of digital economy development, this study incorporates batch-specific dummies within a triple difference (DDD) framework. As shown in Table 14, the estimated coefficients for the first and second batches are 0.084 and 0.058, both significant at the 1% level, while the third batch yields a coefficient of 0.010, significant only at the 10% level. These results indicate that the policy markedly enhanced ULGUE in the first two cohorts, but its impact in the third batch remains limited.

A Wald joint test is further performed on the coefficients of Digital Economy × 1st batch, 2nd batch, and 3rd batch. The joint statistic equals 17.840 and is significant at the 1% level, confirming that batch heterogeneity is statistically evident.

The weaker performance of later cohorts may stem from diminishing marginal returns, as the earlier pilots had already established critical digital infrastructure and institutional frameworks, reducing incremental benefits for subsequent cities. In addition, fiscal constraints, limited digital literacy, and weaker governance capacity hindered effective adoption of digital tools and enforcement of regulatory frameworks. These frictions along the chain from “digital investment → institutional embedding → green outcomes” further diluted policy effectiveness.

These findings accord with Liu et al. (2025), who show that the NEDC policy yields stronger impacts in early pilots with more advanced digital ecosystems. Similarly, Guo and Zhang (2024) reveal that initial batches facilitated efficient allocation of land and production inputs, supporting sustainable, low-carbon growth. Overall, the evidence indicates that the policy’s effectiveness in enhancing ULGUE depends on timing and local readiness, while later cohorts encounter diminishing returns and uneven outcomes from weaker implementation.

Several limitations should be noted. First, because pollution and energy statistics are published with a lag, the dataset ends in 2022 and does not capture emerging developments such as generative AI (AIGC). Second, although the DID approach satisfies the parallel trend assumption, potential interference from concurrent initiatives (e.g., smart city or low-carbon programs) and unobservable shocks cannot be ruled out. Third, the mechanism analysis relies on proxies—patent counts for green innovation and the energy-to-GDP ratio for resource efficiency. While widely used, these indicators may not fully reflect how digital technologies affect land-use practices: patents indicate innovation potential rather than application, and energy intensity may miss structural or behavioral changes. The lack of firm-level surveys or case studies also limits explanatory depth.

In addition, comparative fit statistics (e.g., log-likelihood, AIC) for the three spatial models are not reported; although useful for robustness, they would considerably increase the analytical burden. The findings are also shaped by China’s unique institutional context of state-led digital infrastructure and public land ownership, which restricts generalization to market-based or private land systems. Finally, the framework does not incorporate moderating factors such as institutional quality or urban density, nor account for feedback loops. Excluding these elements, though improving tractability, may reduce the framework’s comprehensiveness.

Future studies can explore the following directions. First, researchers can shift the analytical focus to the micro firm level. Although this study identifies green innovation and resource efficiency as mediators at the city level, firm-level data may provide deeper insights into how digital technologies affect production processes, supply chain management, and land use decisions. Firm surveys or case studies, in particular, could help verify whether patent-based indicators and energy-intensity measures accurately reflect changes in actual land-use efficiency. Second, cross-country comparative research can be conducted. To assess the external validity and institutional contingencies of the findings, future work may compare the digital economy’s effects on land-use efficiency across countries or regions with different governance structures. Third, the mechanism analysis can be refined and extended. Beyond technological and resource-based mechanisms, the digital economy may affect green land use through sociological channels, such as governance reforms, lifestyle shifts, and environmental preferences. In addition, subsequent research could integrate moderating factors such as institutional quality and urban density, as well as feedback loops within the theoretical framework, to better capture the complex interactions between digitalization and sustainable land use. Finally, future research should also pay attention to the negative externalities of digitalization, such as electronic waste and the rising energy consumption of data centers, to achieve a more balanced assessment of its role in sustainable urban transitions.