The Algorithmic Assignment of Incentive Schemes

Abstract

The assignment of individuals with different observable characteristics to different
treatments is a central question in designing optimal policies. We study this question
in the context of increasing workers’ performance via targeted incentives, using machine
learning algorithms with worker demographics, personality traits, and preferences as input.
Running two large-scale experiments we show that (i) performance can be predicted by
accurately measured worker characteristics, (ii) a machine learning algorithm can detect
heterogeneity in responses to different schemes, (iii) a targeted assignment of schemes to
individuals increases performance significantly above the level of the single best scheme,
and (iv) algorithmic assignment is more effective for workers who have a high likelihood to
repeatedly interact with the employer, or who provide more consistent survey answers.

Keywords: Randomized Controlled Trial, Incentives, Heterogeneity, Treatment Effects, Selection, Algorithm

JEL Classification: C21, C93, M52