Scenario
You are planning an A/B test for a new recommendation algorithm in a networked product where users interact with friends and creators. Because user outcomes can be affected by others' assignments (network spillovers), standard individual randomization may violate the "no-interference" assumption.
Task
Design a clustered random sampling approach to select test and control users that limits interference.
Questions
-
How would you construct clusters and assign treatment/control at the cluster level?
-
Which clustering choices make sense in this context (e.g., markets, friend/interaction graphs), and why do they reduce interference?
-
What trade-offs or risks does clustered sampling introduce (e.g., intracluster correlation and power loss)?
Hints
-
Consider geographic/language markets and social/interaction graphs.
-
Aim to minimize cross-cluster ties that carry spillovers.
-
Intracluster correlation (ICC) inflates variance; anticipate power loss.
-
Think about analysis with cluster-robust methods and stratification.