Reducing tax evasion is a key challenge for governments around the world, particularly in developing countries. This paper presents a methodology to generate information to optimize audit strategies. Randomly selected taxpayers receive a deterrence message. Comparing their subsequent tax payments to a control group allows estimating what types of taxpayers are more likely to respond to an increase in perceived audit probability. This information can be used to target audits toward taxpayers that respond particularly strongly, and to construct risk indicators to predict taxpayers’ responses. We show results from an application in Chile and describe lessons learned during the implementation.