We present a framework for analyzing “model persuasion,” where persuaders influence people’s beliefs by proposing models (likelihood functions) that suggest how to organize past data (e.g., on investment performance) to make predictions (e.g., about future returns). Receivers are assumed to find models more compelling when they better explain the past. A key tradeoff persuaders face is that models that better fit the data induce less movement in receivers’ beliefs. Model persuasion sometimes makes the receiver worse off than if he interprets data through the lens of a default model. The receiver is most misled by persuasion when there is a lot of publicly available data that is open to interpretation and exhibits randomness, as this gives the persuader “wiggle room” to highlight false patterns. Even when the receiver is exposed to the true model, the wrong model often wins because it better fits the past. Competition more broadly pushes persuaders towards overfitting available data and as a result tends to neutralize the data by leading receivers to view it as unsurprising. With multiple receivers, a persuader is more effective when receivers share similar priors and default interpretations. We illustrate with examples from finance, politics, and law.
Keywords: model persuasion; Data and Data Sets; Forecasting and Prediction; Mathematical Methods; Framework;
Citation:
Schwartzstein, Joshua, and Adi Sunderam. "Using Models to Persuade." NBER Working Paper Series, No. 26109, July 2019. View Details