Website CV Paper
Fields
Theory, Political Economy
Job market paper
Communicating with Models and Data
Many real-world decisions rely on models and data supplied by experts, like consultants, researchers, and AI chatbots. I study how model simplification distorts data collection when experts want to make decisionmakers optimistic. A receiver decides whether to adopt an initiative and, if so, must make an implementation choice. However, she is uncertain about the data-generating process (DGP) mapping these choices to outcomes. A sender, who wants to maximize the probability of adoption, samples data and estimates a model of this DGP. The sender is either an accurate or simple type. The accurate sender recovers the true DGP. The simple sender is misspecified and assumes the curvature of the DGP has a predetermined value. I show that if the receiver suspects the sender might be simple, she must also entertain alternative DGPs, which the sender can influence through his choice of underlying data. In equilibrium, a simple sender samples data only about extreme choices the receiver is unlikely to make. This leaves room for alternative DGPs that imply higher outcomes at intermediate choices the receiver is likely to make, making her more likely to adopt the initiative. An accurate sender benefits from being misperceived as simple because the receiver entertains alternative DGPs that make her more likely to adopt the initiative than if she knew the truth. The accurate sender benefits most when simple senders are "polarized" and estimate only extremely low- or high-curvature models. In the limit, these correspond to linear regressions and "worst-case scenarios."