Methods & sources · 方法与来源

How these datasets were built

数据构建说明

Each dataset's construction, source files, validation steps, and known limitations are documented below. Click any section to expand it.

Supporters · 国家支持

Supporters methodology支持数据方法

1. What the dataset records, and how it was compiled

No countries have officially "joined" the Global Security Initiative, but many have publicly referenced the initiative in joint statements or at bilateral meetings. These statements and meetings range in the language used to discuss the initiative. Three researchers searched Chinese, English, French, Spanish, and German-language media for official mentions of the GSI by a country's leader or foreign minister. If a country's "support" was only mentioned by China's state media or a Chinese official but was not supported by an official readout of the statement, we did not include it.

2. The 0–5 ranking scheme

The researchers ranked each country's bilateral level of support for the GSI from 0–5. The ranking was cross-checked by three researchers.

  • 0 — No mention. This may be because there has been no opportunity for the initiative to be brought up by China.
  • 1 — Non-committal "acknowledgement".
  • 2 — Positive language such as "welcome" or "appreciate" but no explicit support.
  • 3 — Simple "support" for the initiative.
  • 4 — High support for the initiative, such as "strongly" or "resolutely" support.
  • 5 — Some kind of agreement to implement or pursue further cooperation under the initiative.

In the paper for simplicity I combined 1 and 2 as noncommittal, and 3–5 as support, but also broke down whether this was initiated in a multilateral forum or a bilateral/unilateral statement.

3. Validation via China's partnership hierarchy

To test the accuracy of this ranking, I examined how China's strategic partners respond to the initiative. According to Ketian Zhang (2025), countries have seven possible partnerships with China, in a hierarchy from a simple "comprehensive cooperative partnership" (with Liberia, for example) to the highest "comprehensive strategic co-ordinated partnership" (Russia only). We would expect countries at the top of the hierarchy to be more likely to explicitly join the initiative, while those with no partnership would be less likely to be invited to do so.

Coding these from 0 (no partnership) to 7 (Russia), suggests that the ranking is broadly accurate: a closer partnership with China means a country takes a stronger (bilateral) stance toward the GSI.

4. Overlays (regime, ally definitions, etc.)
  • V-Dem v15 regime classification (year-specific where available; 2022 fallback)
  • EIU Democracy Index score (year-specific; 2022 fallback)
  • US ally (RAND 2017) — RAND RRA1066-1 Appendix B
  • OECD / China's neighbour / BRICS binary flags
  • US and China arms imports 2022 — SIPRI TIV
5. Limitations

"No support" is the absence of a positive signal, not a documented refusal. Several Sahel and Pacific countries have indicated multilateral support that has not yet appeared in retrievable public statements. Bilateral support is credited only when the language explicitly references the GSI.

Media sentiment · 媒体话语

Media-sentiment methodology媒体数据方法

1. Corpus & sourcing

Every article from China's two English-language state-media services that mentions the Global Security Initiative (全球安全倡议), 2022-02-12 through 2026-06-17. n = 1,695 articles; 8,721 sentences after filtering.

  • Xinhua General News Service — the English news wire (1,133 articles).
  • Ministry of Foreign Affairs, English website — news releases plus Spokesperson's Remarks, i.e. the daily regular press conferences (562 articles).

The 2022 – June 2025 base corpus is a LexisNexis export of both services. The June 2025 – June 2026 extension re-runs the same Nexis searches for the Xinhua wire and MFA News; because the Nexis "Spokesperson's Remarks" feed thins out over that window, the GSI press-conference Q&As for the extension are drawn instead from a scrape of the MFA's own daily-briefing transcript archive. Genuine duplicate re-exports (the same article filed twice) are removed.

2. Source-group tagging

Each sentence is tagged with one regional class based on the countries / organisations it references:

  • USA/NATO — references to the United States, NATO, or named NATO member states
  • UN — references to UN bodies, the UNSC, or the Secretary-General
  • ASEAN — references to ASEAN members or ASEAN-led mechanisms
  • BRICS, SCO, EU, AU — analogous
3. Sentiment scoring

Sentence-level lexicon score: (positive − negative opinion words) ÷ number of content words, using the Bing/Liu opinion lexicon. Range roughly −1 to +1; 0 is neutral.

The whole corpus (2022–2026) is scored with this one method, so the extension is directly comparable to the historical series. (The original 2022–2025 release used a different, undocumented lexicon; re-scoring everything uniformly preserves the by-source ordering — USA/NATO most negative, ASEAN most positive — while compressing magnitudes slightly toward zero.)

% negative vs mean. Dividing by sentence length, plus the uniformly positive tone of Chinese state media, makes the mean small and tightly clustered (most country means fall between +0.03 and +0.07). The share of negative sentences (% scoring below 0) is much more variable and interpretable — USA/NATO ≈ 40% negative vs ≈ 6% for Global South groups; conflict states (Iraq, Syria, Afghanistan, Ukraine) highest — so it is the default metric on the map and time series, with the mean offered alongside.

4. Country-level map threshold

The "Mentions and sentiment by country" map only shades countries with ≥25 mentions across the corpus, to prevent noisy single-mention countries from dominating the colour scale.

5. Limitations

English-language coverage only. Pure Chinese-language coverage in People's Daily and other domestic channels is not included here. Some sentences are double-tagged where they reference more than one source group; these appear as separate rows.

Expenditure · 安全开支

Expenditure methodology安全资金数据方法

1. Source data

All figures are drawn from AidData's CLG-Global Dataset v1.0 (Custer et al. 2025), which records 33,593 Chinese government-financed projects committed to recipient countries between 2000 and 2023. Amounts are expressed in constant 2023 U.S. dollars using AidData's Adjusted_Amount_Constant_USD_2023 field. Pledges that never became formal commitments and projects marked Cancelled or Suspended are excluded.

2. Security assistance dataset (604 deals; USD 1.85B)

The security dataset was constructed in three stages.

First, we hand-coded all CLG records for which at least one of the following applied:

  • the recipient agency was a military, police, gendarmerie, customs, border, or interior-ministry entity;
  • the funding agency was a Chinese defence state-owned enterprise (NORINCO, AVIC, Poly Technologies, CETC) or the People's Liberation Army;
  • the project narrative explicitly referenced military equipment, police equipment, demining, peacekeeping deployment, or counter-narcotics / counter-terrorism support.

Second, we used Claude Opus to cross-check the Security flag.

Third, we removed false positives through manual review.

Excluded categories include: oil/gas pipelines and refineries, crude-oil donations, power-generation infrastructure, mining and metallurgical projects, civilian transport infrastructure, civilian telecommunications, civic buildings, agricultural commodity transfers, debt rescheduling, and surveillance technology (held separately).

The final dataset covers 170 recipient countries. Note that many of the included cases had no actual monetary value (as they referenced training or equipment donations). They are included in the database but not in the figures in the paper.

3. Surveillance financing dataset (107 deals; USD 5.78B, 2003–2021)

The surveillance financing dataset was assembled separately.

First, we hand-coded deals whose title or narrative explicitly described:

  • Safe-City or Smart-City systems
  • video-surveillance platforms, CCTV networks
  • facial-recognition systems, biometric-identification systems
  • electronic border (e-border) installations
  • national-security-communication networks, dedicated security information systems
  • maritime surveillance or intelligence-interception equipment

Equipment supplied by named surveillance vendors (Hikvision, Dahua Technology, Megvii, SenseTime, CloudWalk, Yitu) was included.

Second, we used Claude Opus to cross-check the Surveillance flag.

Third, we removed false positives through manual review.

Excluded categories include: air-traffic surveillance, public-health (disease) surveillance, water-utility plant monitoring, and mining-site monitoring. Huawei and ZTE telecommunications-backbone projects (national broadband, 4G/LTE rollouts, fibre-optic backbones, data centres) are treated as commercial telecom infrastructure and kept as a separate dual-use series rather than counted as surveillance.

4. Limitations

This is committed finance, not delivered. Some deals are still in implementation. Arms transfers (SIPRI TIV) are tracked separately from this dataset.