Targeting for Long-Term Outcomes

Decision makers often want to target interventions so as maximize an outcome that is observed only in the long term. This typically requires delaying decisions until or relying on simple short-term proxies for long-term outcome. Here, we build statistical surrogacy and policy learning literatures impute missing outcomes then approximate optimal targeting imputed via a doubly robust approach. We first show conditions validity of average treatment effect estimation with are also sufficient valid evaluation optimization; furthermore, these can be somewhat relaxed optimization. apply our approach two large-scale proactive churn management experiments at The Boston Globe by discounts its digital subscribers aim maximizing revenue. Using experiment, evaluate this empirically comparing learned using ground-truth, outcomes. performance policies statistically indistinguishable, rule out large losses from surrogates. Our outperforms In second field implement additional randomized exploration, which allows us update future subscribers. Over three years, had net-positive revenue impact range $4–$5 million compared status quo. paper was accepted Eric Anderson, marketing. Funding: work supported Media. Supplemental Material: online appendix data available https://doi.org/10.1287/mnsc.2023.4881 .