Does Algorithm Aversion Exist in the Field? An Empirical Analysis of Algorithm Use Determinants in Diabetes Self-Management
41 Pages Posted: 25 Aug 2021 Last revised: 28 Sep 2021
Date Written: July 23, 2021
Advancements in algorithms hold promise to better operations by improving users’ decision-making. However, humans may exhibit so-called “algorithm aversion,” which would be a barrier to achieving such improvements. Using the bolus calculator (algorithm) use behavior in over 306,000 bolus insulin decisions from a field experiment on Type 1 Diabetes self-management (Aleppo et al. 2017), we contribute field analysis to identify drivers of algorithm use. We first focus on an influential experimental finding from Dietvorst et al. (2015) that we refer to as dynamic algorithm aversion – an asymmetric usage response to performance feedback that favors humans over the algorithm. Using panel data models, we reject this hypothesis, instead finding an asymmetric usage response in favor of the algorithm over the human. Moreover, we find that previous algorithm use strongly predicts future algorithm use, and that algorithm use declines from morning to evening. We explore three additional factors that affect algorithm use: one’s need to be precise, deviations from algorithm recommendations, and exposure to multiple, potentially conflicting algorithm input sources. Using linear probability models, we find that algorithm use increases as one’s need for precision increases, and that previous deviations from algorithm recommendations lead to lower future algorithm use. Finally, we leverage an experimental design feature from the original field data with a differences-in-differences analysis to show that increasing the number of measurements provided to the user for a single algorithm input decreases algorithm use. Our field results complement experimental findings and generate new insight into levers that affect algorithm use.
Funding: None to declare
Declaration of Interest: None to declare
Ethical Approval: We obtained approval from the University of Southern California Institutional Review Board that our research is considered non-human subjects research as it meets the criteria for coded private information (study ID: UP-20-00856). That is, our data uses de-identified (publicly available) data from a published field study (REPLACE-BG).
Keywords: healthcare operations, behavioral operations, algorithm aversion, Type 1 Diabetes, precision medicine
Suggested Citation: Suggested Citation