Network Interference & Switchback Design

Network Interference, Spillover Effects & Switchback / Cluster Experimentation resolves SUTVA (Stable Unit Treatment Value Assumption) violations in two-sided marketplaces (Uber, DoorDash, Meituan) and social networks; SUTVA Failure: when treatment units consume shared finite supply (e.g. discounting rides to treatment riders depletes drivers, artificially degrading control rider pickup times), standard user-level A/B testing severely overestimates true general equilibrium lift; Experimental Solutions: 1) Switchback Testing (alternating the entire geographic market between control and treatment in discrete time windows, e.g. 30-min blocks separated by 10-min washout buffers to drain residual momentum); 2) Geographic Cluster Randomization (partitioning non-interacting metropolitan clusters); 3) Graph Ego-Network Cluster Partitioning.

Two-sided dispatch algorithms, dynamic surge pricing experimentation, and social viral loop evaluations.

Eliminates catastrophic cannibalization bias in shared-resource supply environments where standard A/B splits violate causal independence.

Explain the dual requirements of SUTVA (No Interference and No Hidden Variations) and why two-sided driver supply pools violate it?

Why must Switchback experiments enforce a 5-10 minute washout buffer between treatment flips to drain carryover state?

How to correct for temporal autocorrelation in Switchback time-block clusters using Cluster-Robust Standard Errors (CRSE)?

How to deploy Synthetic Controls when geographic randomization units are constrained to single-digit major metropolitan clusters?

Explain Two-Sided Randomization designs estimating both Direct and Spillover network effects simultaneously?