A randomized controlled trial (RCT) assigns units to treatment or control by chance. Because assignment is independent of patient characteristics (in expectation), treated and control groups are comparable. The difference in average outcomes estimates the causal effect.

A/B tests in product teams are small-scale RCTs. Clinical trials are RCTs at scale. Randomization is powerful because it reduces the burden of modeling confounders — though it does not eliminate problems like non-compliance or attrition.

Why observational studies dominate

Ethics, cost, and logistics often forbid randomization. We cannot randomly deny patients a standard therapy or randomly expose cities to pollution. Historical data from operations, EHRs, and CRM systems is almost always observational.

Observational causal inference does not pretend data was randomized. It documents assumptions, adjusts or designs around confounding, and stress-tests results.

Quasi-experiments

Sometimes nature or policy creates as-good-as-random variation: eligibility cutoffs (regression discontinuity), timing of law changes (difference-in-differences), or instruments that shift treatment but not outcome directly. These designs are widely used in economics and increasingly in industry.