AI for Green: Data on Changes in the Green Technology Network Position of New Energy Vehicle Enterprises

The temporal resolution is annual, spanning 24 years. The spatial scope covers the registered addresses and operational locations of the listed firms in mainland China, with spatial information disaggregated mainly to the city and regional levels. In the empirical models, we control for firm, year, industry, and city fixed effects.The dataset is constructed by matching multiple data sources, including patents, annual reports, financial and governance data, regional institutional environment indicators, and derived network metrics.

Green patent data and joint application information are obtained from the China National Intellectual Property Administration (CNIPA), the IncoPat patent database, and the CNRDS database. Firm financial, governance, regional, and other operational information are primarily sourced from Wind, the CSMAR database, and Juchao Information (巨潮资讯网). Data processing begins with standardising stock codes, firm names, and annual identifiers, while handling firm name changes, abbreviation variations, and duplicate records.

Next, we screen invention patents that simultaneously possess NEV‑related and green attributes, based on the patent classification systems for new energy vehicle technologies and green technologies. Firms that jointly apply for green patents are treated as network nodes, and the co‑application relationships between firms serve as network edges. The edge weights comprehensively reflect the number of joint applications and patent citation counts.

To reduce annual fluctuations, we construct dynamic weighted undirected cooperation networks year by year using a three‑year moving window.The firm‑level green technology network position is captured by three dimensions: weighted degree centrality, weighted betweenness centrality, and structural holes constraint. Weighted degree centrality reflects the number and intensity of a firm’s direct cooperative relationships; weighted betweenness centrality captures the firm’s ability to control knowledge‑flow paths in the green technology network; and structural holes constraint reflects the firm’s capacity to connect non‑redundant innovation partners.

After aligning the negative indicator (structural holes constraint) in the same direction, we apply principal component analysis (PCA) to extract the first principal component, which forms a composite index of a firm’s green technology network position. A higher index value indicates that the firm is closer to the network core and possesses stronger advantages in resource connectivity and knowledge control. The intensity of AI application is measured by constructing a multi‑level dictionary of AI‑related keywords based on the full text of listed firms’ annual reports, counting the frequency of AI‑related keywords, and standardising it by the total number of words in each annual report to obtain a relative word frequency density.

Control variables include firm size, R&D intensity, ownership concentration, proportion of independent directors, government subsidy intensity, and equity balance. These are measured respectively by the natural logarithm of total assets, the ratio of R&D personnel to total employees, the shareholding ratio of the largest shareholder, the proportion of independent directors on the board, government subsidies as a share of total assets, and the difference between the top ten shareholders’ shareholding ratio and the largest shareholder’s ratio.

Proportional variables are expressed as decimals; firm size is in logarithmic form; and the composite network position and PCA‑based indices are unit‑less.The data tables are organised with the “firm‑year” as the basic record unit. The main panel table contains 9,640 valid observations. The first row contains column headers, and each row corresponds to one record for a given firm in a given year.

The recommended submitted data files include: (1) a basic firm information table, recording firm codes, firm names, industries, cities, and sample years; (2) a firm‑year panel data table, containing AI application density, the composite green technology network position, control variables, moderating variables, heterogeneity grouping variables, and economic consequence variables; (3) a detailed green patent table, recording patent application numbers, application years, applicants, green technology categories, co‑application relationships, and citation information; (4) an annual network edge table, recording years, firm‑node pairs, and edge weights; (5) an annual network structural characteristics table, recording the average network position, number of nodes, number of network ties, network density, and average degree for 2001–2024; and (6) a variable description and data dictionary, specifying field names, variable symbols, definitions, data types, units of measurement, value ranges, and sources.

The units for the number of network nodes and network ties are “firms” and “ties”, respectively. Network density is defined as the ratio of actual cooperative relationships to all potential relationships, and average degree is the average number of direct connections per node.In the raw data, there may be cases where firms are not listed, annual reports are unavailable, patent information is incomplete, key financial indicators are missing, or database coverage and definitions differ.

The study forms the final analytical sample by deleting observations with missing or abnormal key variables, resulting in an unbalanced panel. Consequently, the number of consecutive observed years varies across firms. Missing values are not replaced with zeros; a zero value indicates that the corresponding indicator was truly not observed or not disclosed for that firm‑year.

Potential data errors may arise from strategic disclosure in annual reports, measurement bias between AI keywords and actual AI adoption levels, name‑matching errors for patent applicants, incomplete coverage of informal cooperative relationships by joint patents, and differences in update timings and statistical definitions across databases. To mitigate these errors, we implement name standardisation, code matching, duplicate verification, outlier detection, 1% and 99% winsorisation, cross‑validation across multiple databases, and alternative measures of AI application.

We also exclude low‑quality disclosure samples, replace the network position indicator, and conduct various robustness checks to validate the results. The dataset reflects observable formal green technology cooperative relationships and cannot fully capture informal linkages such as tacit knowledge exchanges, non‑patent collaborations, technology alliances, or supply‑chain synergies.Data files are recommended to be saved in UTF‑8 encoded CSV format and Microsoft Excel workbook (.xlsx) format; graphical or network visualisation outputs may be saved in PNG format.

All are common, widely accessible file formats. CSV files can be opened with Microsoft Excel, WPS Spreadsheet, Python, R, Stata, and other software; XLSX files can be opened with Microsoft Excel or WPS Spreadsheet, without reliance on specialised or niche software. Data organisation, word‑frequency calculation, social network indicator computation, PCA, and econometric regressions can be performed in a conventional personal computing environment.

To ensure research reproducibility, the submitted data package should include the data dictionary, data‑cleaning rules, variable construction explanations, and analysis code, along with a supplementary note stating the software names, version numbers, and operating environment actually used by the authors. Due to licensing restrictions of commercial databases such as Wind, CSMAR, CNRDS, and IncoPat, as well as the usage policies for listed company annual reports and patent data, the public data package should preferentially provide anonymised and processed derived variables, network edge tables, variable dictionaries, and reproducible code.

For raw data that cannot be made publicly available, the legal access channels and conditions for obtaining them should be clearly stated.