Causal Designs for Field Data

This chapter covers how to recover causal effects from field data no one randomized — a feature that ships region by region, a store format that opens in one city, a policy that lands in a single state. It works through difference-in-differences and its parallel-trends assumption, synthetic control's optimizer-built weighted twin (illustrated on Colorado housing and the Zillow Home Value Index after 2014 cannabis legalization), and heterogeneous treatment effects, where a single average lift hides which segments actually pay. Worked cases — a Bean & Basket checkout rollout, a Denver store format, and an income-stratified milk-pricing experiment — show why the headline number is rarely the decision-relevant one, and why an effect is only as trustworthy as the design that produced it.