You are evaluating a new collaborative feature in a work management product, such as a change to task assignment notifications or inbox collaboration behavior. In a 14-day experiment, 12,000 users across 2,400 workspaces were eligible. Average workspace size was 5 users. Historical data shows the user-level weekly task completion rate has mean 0.40 and variance 0.24, and the intra-workspace correlation for this metric is 0.18. The feature can affect both the treated user and their collaborators.
How would you determine the right unit of randomization, quantify the statistical cost of user-level randomization versus workspace-level randomization, and decide which design is more valid for estimating the causal effect?