China’s pilot to integrate informatization and industrialization boosted green outcomes and productivity at listed manufacturers, driven by intelligent-manufacturing upgrades; effects are largest for big firms, polluting sectors and in the Yangtze River Economic Belt.

Main Finding

China’s Integration of Informatization and Industrialization Pilot (IIP) policy materially improves green development in publicly listed manufacturing firms. Using Difference‑in‑Differences (DID) and Double Debiased Machine Learning (DDML) frameworks, the authors find a modest but statistically significant increase in firms’ environmental efficiency and green productivity after IIP adoption. The positive effect operates in large part through firms’ intelligent transformation (digitalization / intelligent manufacturing), and it is stronger for firms in the Yangtze River Economic Belt, heavy‑polluting industries, and larger firms. Results are robust to PSM‑DID, DDML specifications, an IV approach (Terrain Ruggedness Index), and alternative dependent variables (e.g., Green Total Factor Productivity).

Key Points

• Policy tested: Integration of Informatization and Industrialization Pilot (IIP) initiated by MIIT (China); quasi‑experimental roll‑out from 2014.
• Primary result: IIP adoption → higher environmental efficiency and green productivity in manufacturing firms.
• Mechanism: Intelligent transformation (multi‑dimensional index) mediates the policy → green development link.
• Heterogeneity: Larger effect sizes in the Yangtze River Economic Belt, in heavily polluting sectors, and among larger firms.
• Robustness: Findings hold under:
  • Parallel trend checks,
  • Propensity Score Matching DID (nearest neighbour),
  • DDML causal estimates (adjusting for high‑dimensional confounders),
  • Instrumental variable strategy using Terrain Ruggedness Index,
  • Alternative DV: Green Total Factor Productivity.
• Measurement innovations:
  • Intelligent transformation measured as a composite index using multiple sub‑indicators and entropy weighting (dynamic weights).
  • Use of ML (random forests, neural nets) within DDML to handle high‑dimensional controls and reduce model‑specification bias.

Data & Methods

• Data: Panel of Chinese publicly listed manufacturing firms, 2007–2022; 24,081 firm‑year observations from 2,880 firms; 551 treated firms (IIP pilots after 2014). Sources: Wind database, firm annual reports, city yearbooks, and hand‑collected policy data.
• Outcome measures: Environmental efficiency and green productivity (also Green Total Factor Productivity for robustness).
• Independent variable: DID treatment dummy (firm included in IIP × post‑policy period).
• Estimation approaches:
  • Standard DID with parallel trends testing.
  • Propensity Score Matching DID (nearest neighbour).
  • Double Debiased Machine Learning (DDML) to flexibly control high‑dimensional covariates and reduce bias from model misspecification.
  • Instrumental variable (Terrain Ruggedness Index) to address endogeneity concerns.
• Mechanism analysis: Mediating role of intelligent transformation tested within DDML framework.
• Index construction: Intelligent transformation index assembled from multiple indicators (investment, applications, innovation measures) with entropy weighting to assign objective, data‑driven weights.

Implications for AI Economics

• Micro‑level evidence that AI/digitalization policy (IIP) can simultaneously raise environmental efficiency and productivity in manufacturing — supporting the notion that intelligent manufacturing can be pro‑environment and pro‑growth.
• Intelligent transformation is a key transmission channel: policies that directly reduce adoption costs, improve digital skills, or subsidize intelligent equipment/software are likely to magnify green benefits.
• Heterogeneous returns imply targeted policy design: heavier support or tailored instruments for smaller firms, non‑Yangtze regions, and non‑polluting sectors may be necessary to equalize benefits and maximize aggregate environmental gains.
• Methodological lesson: DDML and other modern causal‑machine‑learning tools are valuable in policy evaluation where many potential confounders and non‑linearities exist — improving causal inference about AI adoption effects.
• Measurement insight: Composite, entropy‑weighted indices better capture multi‑dimensional AI/automation adoption than single proxies (e.g., intangible assets or text counts); useful for future empirical work in AI economics.
• Policy scaling: The shift (noted by authors) from firm‑level pilots to city‑level “5G + Industrial Internet” pilots suggests evaluations should address spatial spillovers and infrastructure complementarities (5G, data platforms) in future AI economics research.
• Future research directions suggested: long‑run labor and distributional impacts of intelligent transformation, cost‑benefit analysis combining environmental and productivity gains, spillover effects across supply chains and regions, and generalizability beyond publicly listed firms and China.