Across many experimental settings, from precision medicine to science funding, different interventions may be much more effective for some individuals than others. In such cases, both researchers and decision makers would like to know more than the “average treatment effect.” Doing the most good with limited resources can also require targeting interventions toward those who will benefit from them most—if only we could identify those subgroups in advance, or, in other words, if only we could estimate “personalized treatment effects” (PTEs). Consider a school district deeply concerned about devasting learning losses during the pandemic. Preliminary studies indicate that high-dosage tutoring (HDT) is a costly but highly effective form of small group instruction that can help teachers double or triple mathematics learning each year. Another effective yet lower-cost option is to replace some in-person instruction with time on a high-quality computer-assisted-learning platform. This treatment is called “sustainable high-dosage tutoring” (SHDT). On average, students seem to benefit from SHDT as much as HDT. The problem is that certain students hardly benefit from the technology-assisted component at all, and thus should be treated with the more expensive high-dosage tutoring. Personalized treatment effect estimation allows the school district to identify these students. Jens Ludwig at the University of Chicago seeks to advance the estimation of PTEs when data is obtained from large-scale field experiments. His approach builds on fundamental techniques developed by Sloan grantee Susan Athey, whose “random forest” algorithm carries out machine learning by exploring many different “decision trees.” Ludwig will extend these methods, leveraging recent experimental data involving eight U.S. school districts where a total of 20,000 students were randomly assigned to a control group, to high-dosage tutoring (HDT), or to sustainable high-dosage tutoring (SHDT). Ludwig’s team will develop PTE estimation methods for identifying which of the variables observable in advance best predict how students vary in their treatment response to HDT and SHDT. They will devise and test practical rules for helping schools assign different types of students to HDT or SHDT based on these individual characteristics. This personalized approach has the potential to maximize learning gains and optimize resource allocation, offering a cost-effective solution to reversing pandemic learning loss in STEM education and allowing the same tutoring program budget to benefit a greater number of students The project will also produce a practical guide for researchers on PTE estimation methods, helping to ensure robust application of this methodology across other contexts outside of STEM education.