Working Paper: NBER ID: w25976
Authors: M. Kate Bundorf; Maria Polyakova; Ming Tai-Seale
Abstract: Algorithms are increasingly available to help consumers make purchasing decisions. How does algorithmic advice affect human decisions and what types of consumers are likely to use such advice? We use data from a randomized controlled trial of algorithmic advice in the context of prescription drug insurance to examine these questions. We propose that algorithmic recommendations can affect decision-making by influencing consumer beliefs about either product features (learning) or how to value those features (interpretation). We use data from the trial to estimate the importance of each mechanism. We find evidence that algorithms influence choices through both channels. Further, we document substantial selection into the use of algorithmic expert advice. Consumers who we predict would have responded more to algorithmic advice were less likely to demand it. Our results raise concerns regarding the ability of algorithmic advice to alter consumer preferences as well as the distributional implications of greater access to algorithmic advice.
Keywords: algorithmic advice; consumer decision-making; health insurance; Medicare Part D; randomized controlled trial
JEL Codes: D1; D12; D8; D81; D82; D83; D9; D90; D91; G22; H51; I13
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Noisy beliefs about product features (D83) | Suboptimal choices (D01) |
| Noisy beliefs about own utility function parameters (D80) | Suboptimal choices (D01) |
| Exposure to algorithmic recommendations (C91) | Switching likelihood (C34) |
| Exposure to algorithmic recommendations (C91) | Satisfaction with choice process (D80) |
| Exposure to algorithmic recommendations (C91) | Time spent on plan choice (C41) |