Section 3: Principles for building in machine learning
Researchers and applied scientists need to take a revised approach to experimental design if they want to leverage the advantages of machine learning. Field experiments in applied sciences are routinely optimized for detecting average effects, and not subgroup (heterogeneous) effects. They are powered and designed to understand whether interventions work on average. This limits our ability to learn more about what interventions work and whether impacts vary among subgroups. Designing experiments to find out how to personalize interventions needs a tailored approach, which deviates from existing best practices. In this section, we outline principles for how to design for machine learning.