Abstract: Systems are provided to select targets for communications, designed to cause or provoke the targets to perform a desired action, using bias-corrected models that predict the effect of such communications on the targets. Data regarding previous actions taken by potential targets for communications—also referred to as “candidates”—can be analyzed to determine whether there is a difference in actions taken by candidates who have received prior communications in comparison with candidates who have not received the prior communications. Biases in the selection of candidates to receive the prior communications can be corrected by weighting the data associated with the selected candidates to more closely match the distribution of candidates not selected to receive the prior communications.