Randomized experiments should be designed to predict individual treatment effects, not just average effects. Different sampling methods and models can affect the accuracy of these predictions. Problems with generalizability can lead to bias in predictive models. Sometimes, predicting individual effects is more accurate than estimating average effects.