Skip to content
Causal Inference & Experimentation

When Randomization Fights Back

Interference, Switchbacks & Networks

10 min+100 XPStep 1/10

SUTVA dies in marketplaces

SUTVA's no-interference clause — my treatment can't touch your outcome — holds fine for isolated content features. It collapses anywhere users share a resource:

  • Two-sided marketplaces — treated riders book more, drivers shift supply, control riders' wait times change. The control 'condition' no longer exists.
  • Social products — a feature that makes treated users post more changes the feeds of their control friends.
  • Shared finite resources — support capacity, inventory, ad auctions, ranking slots.

The signature symptom: the control group's behavior moves when treatment turns on — not because they saw the feature, but because the world around them shifted. Once controls are contaminated, the A/B difference measures 'difference between more-treated and less-treated worlds', which isn't your estimand.

warning

Interference doesn't always make experiments conservative — contamination can inflate *or* deflate estimates depending on which direction the shared resource flows. Assume the bias sign is unknown until modeled.