Algorithms and Bureaucrats: Evidence from Tax Audit Selection in Senegal

Can algorithms enhance the work of bureaucrats in developing countries? Developing economies are often data-poor environments, where individual bureaucrats have substantial discretion to take key decisions, such as selecting taxpayers for audits. Exploiting a trove of newly digitized microdata, we conduct a field experiment across tax offices in Senegal, whereby half of the annual audit program is selected by inspectors and the other half is selected by a risk-scoring algorithm. We document three sets of results. First, inspector-selected audits are 18 ppt more likely to be conducted and detect higher amounts of evasion. Second, algorithm-selected audits are less cost-effective and do not generate less corruption. Third, even an ex-post optimized algorithm, trained on outcome data, would have increased detected evasion by only 5% compared to the inspector selection. This is consistent with the inspectors’ high skill level and the imperfection of the available data.

(with Pierre Bachas, Alipio Ferreira and Bassirou Sarr)