Breaking through ingrained beliefs: revisiting the impact of the digital economy on carbon emissions

Global energy and climate challenges have intensified, with carbon emissions growth peaking in the past decade and a widening gap between reduction needs and measures. The digital economy, while often associated with energy-intensive infrastructures (e.g., data centers, blockchain), may also generate system-wide efficiencies. The paper distinguishes the digital economy (digital industrialization plus industrial digitalization) from the narrower digital industry. It addresses the observed “Digital Economy Paradox” in developing contexts, where regions with higher digital development show higher emissions. The research question is whether, and through which mechanisms, the digital economy reduces carbon emissions at the regional level. The study proposes four hypotheses: H1, the digital economy reduces carbon emissions; H2, it reduces emissions by enhancing low-carbon technological innovation; H3, it reduces emissions via industrial diversification; H4, digital industrialization may impede the mitigation effects of the digital economy. Using county-level data from China, the study constructs a dataset leveraging big data to test these hypotheses and to provide a comprehensive framework linking digital economy development to carbon outcomes.

Prior research often equates the digital economy with the digital industry, emphasizing direct energy and emissions from ICT infrastructure, devices, and data centers. Micro-level LCAs highlight significant emissions during network transmission and from PCs in data centers. Meso-level evidence indicates a considerable and growing carbon footprint of ICT. Macro-level work finds limited or non-linear (e.g., inverted U-shaped) effects, with some evidence of emissions promotion in developing contexts. However, this overlooks digital spillovers into non-digital sectors. Scholars differentiate digital industrialization (production of digital technologies) from industrial digitalization (application of digital tech across sectors). While producing digital devices raises energy use, digitalization facilitates renewable integration, efficiency, and system optimization; the overall reduction can outweigh direct ICT energy use. Hence, when considering both production and application segments, the digital economy may mitigate emissions. The study formalizes: H1 the digital economy reduces emissions; H2 via low-carbon technological innovation; H3 via industrial diversification; H4 digital industrialization may dampen the mitigation effect, contributing to the paradox.

Data: Unbalanced county/district panel for China covering 1579 units from 2004–2017 (19,766 observations) across 30 provinces. Sources: satellite-based measures for CO2 emissions and terrestrial vegetation carbon sequestration; regional digital economy measures from digital patents; Digital Financial Inclusion Index for robustness; low-carbon innovation from IncoPat patents; controls from China County Statistical Yearbook. Variables: Explained variable (CO2) at county level constructed using harmonized DMSP/OLS and NPP/VIIRS nighttime lights (scale harmonization, unit root testing, top-down weighted allocation, artificial neural network to align sensors). Core explanatory variable (Digital): intensity of digital economy measured by counts of digital technology patents (AI, blockchain, big data, IoT, cloud), capturing regional capabilities in digital R&D and application. Mediators: Low-carbon technological innovation (Li) from IncoPat, using WIPO Green Inventory (2010) categories; three patent groups extracted: primary classification in alternative energy production or energy conservation; secondary waste management (waste disposal); secondary administrative regulation/design (carbon/emissions trading). Industrial diversification (Div) measured as inverse Herfindahl-Hirschman index of employment shares across industries: Div_it = 1 / Σ_j S_ijt^2 (higher indicates more diversified and balanced industrial structure). Controls: administrative area (Area), total power of agricultural machinery (Machine), registered population (Popu), value added of primary industry (Str), general public budget expenditure (Budget), number of industrial enterprises above designated size (Industry). Estimation strategy: Causal mediation model based on an expanded structural equation framework with individual and year fixed effects. Main equations: (i) CO2_it on Digital_it and controls; (ii) mediator Me_it (Li or Div) on Digital_it and controls; (iii) CO2_it on Digital_it, mediator, and controls. Quasi-Bayesian Monte Carlo approximation (counterfactual framework) used to compute Average Causal Mediation Effect (ACME), Average Direct Effect (ADE), and Average Treatment Effect (ATE), with 1000 simulations. Robustness checks include propensity score matching (high vs low digital economy counties using KNN caliper matching, K=4, caliper=0.01), substitution of outcome with carbon sequestration (CO2seq), and substitution of Digital with the Digital Financial Inclusion Index (Dfiic). Endogeneity addressed via instrumental variables: historical fixed telephone lines per 100 people and post offices per million people in 1984, each interacted with annual internet users to form time-varying IVs; 2SLS with individual and year fixed effects and Kleibergen-Paap statistics reported. Moderation analysis of digital industrialization (Din), proxied by the sum of software business revenue and industrial value-added; Din=1 if above mean, else 0; interaction Din×Digital identifies the “Digital Economy Paradox.”

- Direct effect: The digital economy significantly reduces carbon emissions. In the baseline, Digital’s coefficient on CO2 is negative and significant at 1% (e.g., −0.123***), confirming H1. - Mechanisms: Digital significantly promotes mediators (Li and Div). When mediators enter, their coefficients on CO2 are negative and significant (e.g., Li −0.277*** to −0.449***), indicating that higher low-carbon innovation and greater diversification reduce emissions, supporting H2 and H3. - Mediation magnitudes (Table 5): - Via low-carbon technological innovation (Li): ACME = −0.007 (95% CI [−0.013, −0.002]); ADE = −0.938 ([−1.100, −0.834]); ATE = −0.945 ([−1.107, −0.840]); proportion mediated ≈ 0.008. - Via industrial diversification (Div): ACME = −0.010 ([−0.014, −0.005]); ADE = −0.838 ([−0.949, −0.767]); ATE = −0.848 ([−0.959, −0.776]); proportion mediated ≈ 0.011. - Robustness (Table 3): Results hold when (i) using PSM between high/low digital economy groups; (ii) replacing outcome with carbon sequestration (Digital → CO2seq: −0.059***); (iii) replacing Digital with Dfiic (negative and significant effects persist). - Endogeneity (Table 4): Historical telecom and postal infrastructure (interacted with internet users) provide strong, relevant IVs (e.g., Kleibergen-Paap rk statistics 237.795, 91.521). Second-stage estimates show Digital retains a significantly negative effect on CO2, reinforcing causality claims. - Digital economy paradox (Table 6): Digital remains negative (e.g., −0.039*** to −0.121***), but the interaction Din×Digital is positive and significant (≈ 0.029*** to 0.033***). In regions where digital industrialization dominates (Din=1), the carbon-mitigation effect of the digital economy is weakened, supporting H4 and explaining the paradox. - Controls: Larger administrative area, population, value added in primary industry, and number of large industrial firms tend to raise emissions; greater agricultural machinery power and higher public budget expenditures are associated with lower emissions in several specifications.

The findings confirm that considering the digital economy as a system—encompassing both digital industrialization and the digitalization of traditional sectors—resolves the apparent contradiction in earlier studies that focused only on direct ICT emissions. Digital technologies reduce emissions in traditional sectors through efficiency gains, optimized resource allocation, spillovers, and cross-industry integration (e.g., smart logistics, industrial internet, precision agriculture). These indirect reductions outweigh the direct footprint of digital infrastructures, yielding a net inhibitory effect on emissions. Mediation results show statistically significant, albeit modest, channels via low-carbon technological innovation and industrial diversification, indicating the digital economy facilitates green innovation and broader, more interconnected industrial structures that curb emissions. The moderation analysis reveals why high-digital regions may display high emissions: when digital industrialization dominates the digital economy’s composition, the direct energy demands of data centers and ICT manufacturing can mask system-wide benefits, dampening net mitigation. Overall, the results address the core research question by identifying and quantifying both the direct and mediated impacts, and by clarifying the conditions under which mitigation gains are obscured.

The study contributes a comprehensive causal mediation framework demonstrating that the digital economy reduces regional carbon emissions directly and indirectly via low-carbon technological innovation and industrial diversification. It clarifies the “Digital Economy Paradox” by showing that a composition skewed toward digital industrialization can attenuate mitigation effects. Policy implications include accelerating the digital transformation of traditional industries; leveraging large-scale, coordinated computing and data infrastructure (e.g., interregional data center networks) to improve energy allocation efficiency; and adopting dynamic, region-specific digital economy strategies reflecting resource endowments and energy structures. Governments should also use digital tools to strengthen environmental data collection and oversight. Future research should examine heterogeneity across industries and entities as the digital economy evolves, providing more granular evidence on sector-specific pathways to decarbonization.

- Composition and evolution: The digital economy is rapidly changing, with emerging industries and applications potentially altering the balance between digital industrialization and industrial digitalization; effects may vary over time and across regions. - Measurement constraints: Digital economy intensity is proxied by digital patents, and low-carbon innovation by selected patent categories; these measures may not capture all relevant dimensions of digital adoption or green innovation quality. - Data scope: Despite extensive coverage, the panel is unbalanced and limited to China’s counties/districts; generalizability to other developing contexts requires caution. - Identification assumptions: While fixed effects, IVs, and quasi-Bayesian mediation help address endogeneity and causal pathways, unobserved shocks or measurement error could remain. - Confidentiality: Full datasets are not publicly shareable due to agreements and sensitivity, limiting replication to provided supplementary materials.