Watch on YouTube

7:48

After randomization: attrition, non-compliance, and spillover

Randomization makes the groups comparable at baseline, and what happens next can undo that. If the people who drop out differ between arms, the comparison is no longer fair, and sensitivity analyses test how much it matters. If people don't take what they were assigned, intention-to-treat answers the policy question of offering the intervention, per-protocol analysis risks selection bias, and complier average effects use randomization as an instrument. If treatment spreads to the untreated, as with school deworming in Kenya, cluster randomization creates a buffer, at the cost of a much larger sample.

Also covered: attrition and sensitivity analysis · intention-to-treat and per-protocol · spillover and SUTVA · intracluster correlation and design effect