# Who sees, who decides? In 2015, China moved decision rights for some branches from the centre to prefecture supervisors. Follow the same information through report, centre and intervention. _CEPR Discussion Paper 21846 · 16 August 2026 · not peer reviewed. documented bank case · model mechanism · transferable question._ ## The change was not only geographic The 2015 reform did not create autonomous regional regulators. It moved decision rights over local-bank branches to prefecture-level supervisors while keeping them accountable to the centre. That separates the question about information from a slogan about decentralization. The reader sees what the study reports first. Only then comes a reconstruction of the proposed mechanism and a transfer sheet for other systems. ## Move the stamp, keep the hierarchy Choose who decides in the same scene. The observation stays local. The point where information becomes actionable changes. ### Old architecture The local office gathers information about the branch, but the decision remains at the centre. Local detail travels through a report and must be translated for the decision point. ### Delegated decision Decision rights for local-bank branches move to the prefecture supervisor, who remains accountable to the centre. Local detail reaches the point that can order the intervention directly. The scene explains the mechanism. The study reports aggregate results, not public before-and-after matches for individual cases. ## More interventions. A more differentiated pattern. If delegation only meant a tougher local hand or suddenly riskier branches, we should see a different profile. The authors look for those alternatives. ### 01. More interventions **observed result.** Branches affected by the reform were 57–80% more likely to receive a supervisory intervention. The detailed estimate is +6.3 to +8.8 percentage points, against an unconditional probability of 11.1%. ### 02. Less uniform fines **observed result.** Average fines fall while fine dispersion rises. The pattern is consistent with detecting less severe cases and making more differentiated responses. This is a result pattern, not proof that every additional intervention was a real violation newly discovered. ### 03. Information distance matters **observed result.** Effects are larger where the centre’s informational disadvantage is more plausible, for example farther from Beijing. These heterogeneous effects fit the authors’ proposed mechanism, but do not directly observe what was lost in an individual report. ### 04. Lending changes **downstream behaviour.** Later, local branches charge loan spreads about 31 basis points higher and make smaller loans. This behaviour is consistent with more conservative risk-taking. It does not show that the result is socially optimal. ### 05. What we cannot see **unknown.** The data do not contain the universe of real violations that went undetected or unsanctioned. The improved-detection interpretation and any efficiency gain remain limited by the model and by what the design can measure. **The difference between a recorded intervention and a real violation matters. The dataset sees enforcement, not the universe of events that never reached the register.** ## Transfer the question, not the result Choose a domain. The sheet does not say who should decide and does not assign China’s effect to Romania. It shows what to verify before proposing a move in authority. ### bank supervision The starting case: the bank, branch, prefecture and centre hold different information and rights. **evidence to look for.** intervention register, decision authority, spreads and loan size 01. **Where does the highest-resolution observation appear?** Find where the case can be seen before it becomes a category. 02. **Who can turn an observation into a case?** Separate the right to report from the right to open a procedure. 03. **Who can decide the intervention?** Name the formal authority, not only the institution on the chart. 04. **What must travel upward before action can begin?** Note what becomes a summary, category or threshold. 05. **What preserves control and accountability?** Check feedback, audit, capture and effects beyond the territory. _There is no score. Local authority is not automatically better, and centralization is not automatically worse._ ## Three levels of evidence The page separates the empirical result, the author’s public summary and the open draft used for detailed checks. Transfer to other domains is a research proposal, not a finding of the study. ### S1 · CEPR Discussion Paper 21846 Published 16 August 2026. A discussion paper, not peer reviewed. Supports the reform, sample, aggregate results and main limits. [open source](https://cepr.org/publications/dp21846) ### S2 · Wouter Dessein · publications The author’s public summary describes the model of bias and information loss and the distinction between delegation within a hierarchy and autonomous decentralization. [open source](https://wd2179.github.io/publications/) ### S3 · Authors’ November 2025 draft The open version used to check the detailed estimate, empirical architecture and tables referenced by the CEPR summary. [open source](https://sbfc.sydney.edu.au/2025/papers/SBFC2025_7B2_P234.pdf) ### What this page cannot say - It does not show the universe of real violations that went undetected. - It does not show that local authorities were less biased. - It does not show that more interventions mean a safer society or socially optimal supervision. - It does not estimate what would happen in Romania or another domain. - It does not turn a model reconstruction into a record of an actual report. _Facts about the reform and aggregate results are marked as documented. The hypothesis about information being translated or lost is marked as a model mechanism. Questions for other systems are marked as transfer._ --- A public object by Marius Comper. [about the author](https://mariuscomper.uk/).