You work on a collaborative design product and are testing a change to sharing or commenting behavior. Because users work in teams on the same files and projects, one user’s treatment experience could affect another user’s outcomes even if that second user is assigned to control. You are concerned the experiment may violate the usual independence assumptions.
When would you worry about network interference in this experiment, and how could it distort the measured treatment effect? How would you redesign or analyze the test to account for that risk before deciding whether to ship?
Users collaborate in shared workspaces or projectsTreatment can change behavior of untreated collaboratorsNaive user-level randomization may violate SUTVAThe right answer should discuss redesign, not just post-hoc analysis