People often make choices with imperfect knowledge of how the variables in their decision problem are related. We study such choices when individuals face menus of conflicting and possibly misspecified models that link these variables. Do they discard inaccurate models, what types of inaccuracies do they detect, and how? Or do they instead follow models that sound appealing at face value, and what determines that appeal? Our experiment yields two main findings. First, many individuals readily intuit the models’ predicted correlations and reject models that contradict the data. Performance is high because the required inference is qualitative rather than quantitative. Second, when unable to identify the correct model, most choose cautiously by focusing on worst-case outcomes. This behavior contradicts the Narrative Competition literature’s assumption of best-case maximization, but a failure of contingent reasoning when interpreting models’ payoff implications can mimic that assumption. Our results are robust to tripled stakes.
Ambuehl et al. (Thu,) studied this question.