算法与官僚：来自塞内加尔税务审计选择的证据（英）

11205 Algorithms and Bureaucrats Evidence from Tax Audit Selection in Senegal Pierre Jean BachasAnne BrockmeyerAlipio FerreiraBassirou Sarr Development EconomicsDevelopment Research GroupSeptember 2025 Policy Research Working Paper11205 Abstract Can algorithms enhance bureaucrats’ work in developingcountries? In data-poor environments, bureaucrats oftenexercise discretion over key decisions, such as audit selec-tion. Exploiting newly digitized micro-data, this studyconducted an at-scale field experiment whereby half of Sen-egal’s annual audit program was selected by tax inspectorsand the other half by a transparent risk-scoring algorithm. The algorithm-selected audits were 18 percentage pointsless likely to be conducted, detected 89% less evasion, wereless cost-effective, and did not reduce corruption. Moreover,even a machine-learning algorithm would only have mod-erately raised detected evasion. These results are consistentwith bureaucrats’ expertise, the task complexity, and inher-ent data limitations. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developmentissues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry thenames of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely thoseof the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank andits affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. AlgorithmsandBureaucrats: EvidencefromTaxAuditSelectioninSenegal PierreBachas,AnneBrockmeyer,AlipioFerreira,BassirouSarr* *Pierre Bachas: World Bank Research, pbachas@worldbank.org; Anne Brockmeyer: World Bank, IFS, UCL and CEPR,abrockmeyer@worldbank.org; Alipio Ferreira: Southern Methodist University, alipioferreira@smu.edu; Bassirou Sarr:Senegal Ministry of Finance. We thank Denis Cogneau, Laurent Corthay, L´eo Czajka, Lucie Gadenne, Stephen Hansen,Janet Jiang, Nicola Limodio, Jan Loeprick, Markus Kitzmuller, Justine Knebelmann, Imran Rasul, Dan Rogger, EduardoSouza-Rodrigues, Gabriel Zucman for helpful comments and discussions, and seminar audiences at Berkeley Haas,IFC, IFS, IIPF, CESifo Public Economics Week, INSPER, TARC Exeter, WB Tax Conference, CMI TaxCapDev Con-ference, Paris School of Economics, Oxford Centre for Business Taxation, Oxford Economics Department, PUC Rio,University of Muenster, University of New Mexico, Norwegian School of Economics, NTA Conference, North TexasEconomics Conference, RIDGE Public Economics, and Universidad del Pac´ıfico. We thank Senegal’s Tax Administra-tion (DGID), in particular, Bassirou S. Niasse, Amadou A. Badiane, Oumar D. Diagne, Hady Dieye, Mor Fall, SerigneM. Fall, and Mathiam Thioub. We thank Samba Mbaye, Assane Sylla, and Medoune Sall from the CRDES for theircollaboration, Oumy Thiandoum and Rafael Vilarouca for excellent research assistance, the Paris School of Economicsand CEPREMAP for administrative support.The authors are grateful for financial support from the Knowledge forChange (KCP) Program, administered by the World Bank, and currently funded by The Swedish International Devel-opment Cooperation Agency (SIDA), Agence Franc¸aise de D´eveloppement (AFD) - French Development Agency, theGovernment of Japan, and the European Union; the research support budget of the World Bank’s research group; UKAIDvia its Economic and Development Institutions call for proposal; the Centre for Tax Analysis in Developing Countries(TaxDev) at the Institute for Fiscal Studies; and the UKRI through Brockmeyer’s Future Leaders Fellowship (grant ref-erence MR/V025058/1). The findings, interpretations, and conclusions do not represent the views of the World Bank, itsaffiliated organizations, its Executive Directors, or the governments they represent, nor of the Government of Senegal. 1Introduction Lower-income countries are often data-poor environments, where policy decisions are taken in a dis-cretionary rather than data-driven manner. For instance, individual bureaucrats may decide which tax-payers to audit, which water treatment facilities to inspect, and which manufacturing plants to monitorfor pollution (Khwaja et al., 2011; OECD, 2023). In high-income countries, such decisions are oftenbased on rules and data-driven, with limited input from individual agents. Data-driven decision mod-els leverage available information in a systematic manner but require high-quality data. Discretionarydecision-making, on the other hand, leverages bureaucrats’ private information and experience but isvulnerable to bias. Although high-income countries commonly use data-driven methods for policydecisions, there is limited causal evidence on the conditions under which developing countries canbenefit from re