Biased Updating in Dynamic Experimentation

Abstract

There is ample evidence that people overestimate the control they have over stochastic outcomes. In this model, I demonstrate the consequence of this bias in a situation where an agent learns about the profitability of an action by taking it repeatedly, i.e. experimenting. The agent may or may not have control over the profitability of the action, which is to be learnt. I show that the bias leads to a lower incentive to experiment. Moreover, the bias may cause the «near-miss effect,» where a failure perceived as close to a success results in a higher incentive to experiment.