How Do Geo and Switchback Designs Handle Interference?

Geo cluster and switchback experiments address interference and spillover in marketplace, social, or channel experiments where user level randomization creates contamination. In a two sided marketplace like Uber or Airbnb, treating some buyers but not others in the same city affects supply availability for everyone. Sellers see mixed signals from treated and control buyers, which washes out the treatment effect and inflates variance. The solution is to randomize entire geographic clusters (cities, regions) or time windows so all units in a cluster receive the same treatment. Geo cluster experiments assign entire cities or regions to treatment or control. This eliminates within-market contamination but introduces new challenges. First, sample size is limited by the number of available clusters. Randomizing 40 cities instead of 1 million users drastically reduces power. Second, clusters are heterogeneous; New York and a small rural town have very different baselines. Selection of treatment clusters uses correlation based matching to controls or simulation to minimize minimum detectable effect (MDE). In a retail experiment, selecting 20 treatment stores over 6 weeks with a 52 week preperiod achieved 1.22 percent MDE for order count using synthetic difference in differences and jackknife variance. Switchback designs alternate treatment by time, assigning all units to treatment in odd hours and control in even hours, or treatment on weekdays and control on weekends. This works when carryover effects decay quickly. Switchbacks preserve large sample sizes (all users are measured) while reducing interference within each time window. The challenge is choosing the right window length. Too short (minutes) and carryover contaminates adjacent periods. Too long (weeks) and you lose sample efficiency. DoorDash uses hourly switchbacks for delivery time prediction experiments where carryover is minimal. Analysis of geo and switchback experiments requires specialized methods. Use difference in differences or synthetic control to account for preperiod trends and cluster heterogeneity. Compute cluster robust or jackknife standard errors to reflect the effective sample size of clusters, not individual users. Variance is higher and MDE is larger than user level experiments, often 2x to 5x, but the estimates are unbiased when interference is present. The trade off is between unbiased but noisy geo cluster estimates versus biased and precise user level estimates when contamination exists.

Uber runs geo cluster experiments for driver incentive changes, assigning entire cities to treatment to prevent supply side spillover across nearby riders

Airbnb uses geo randomization for pricing experiments in marketplace settings where host supply affects both treated and control guests in the same city

Instacart uses switchback designs for shopper batch assignment, alternating treatment hourly to reduce interference while maintaining statistical power