Discretion vs. Algorithms

Abstract

The implementation of government programs requires a list or register of individuals who are eligible for the program. Building these registers accurately is a challenge in contexts with low administrative capacity, and is often the responsibility of bureaucrats. We study the impacts of giving such bureaucrats more (or less) discretion in building these registers in the context of the creation of the first digitized property tax register in Senegal. We randomly assign neighborhoods to valuation methods with different degrees of bureaucrat discretion and compare the registered property values against a benchmark of market values provided by licensed real estate assessors. Bureaucrats in full discretion areas undervalue properties, and more so for higher-value properties, resulting in a regressive tax profile. The median tax rate is 3.8% in the lowest quintile and 1.7% in the top quintile, instead of the expected 4.4% and 8.6% rates based on the tax code. In contrast, a rule-based system where bureaucrats record property characteristics (not values) that an algorithm then uses to compute values, significantly reduces this tax gap. A pure rule with no bureaucrat inputs yields the highest accuracy and equity. We show this is due to bureaucrats’ lack of knowledge about high-end properties and their fairness concerns, and not due to collusion between bureaucrats and property owners.