When Incentives Matter Too Much: Explaining Significant Responses to Irrelevant Information

When agents observe a continuous variable and a discrete signal based on that variable, theory suggests that the signal should not impact behavior conditional on the variable. Numerous empirical studies, many based on regression discontinuity design, contradict this basic prediction. We propose two rationalizations with testable implications. One is based on information acquisition costs and the other on learning and imperfect information. Using education data from North Carolina and exploiting a pay-for-performance system, we find support for the model of learning. This implies that rational responses to policy interventions may emerge gradually, and evaluations with short-term data may understate treatment effects.