Interference, Spillovers, SUTVA

SUTVA (Stable Unit Treatment Value Assumption) is a core principle in causal inference that states each person's outcome should only depend on their own treatment, not on what treatment others receive. When this assumption is violated—because of social influence, shared resources, or communication—it's called interference, and the effects that spread to untreated individuals are called spillovers. Think of it like testing a new vaccine by giving it to half the people in a shared household—the untreated family members might still get some protection if the vaccine reduces transmission, making it impossible to isolate the vaccine's true effect. Notice this: When people influence each other, your control group isn't truly controlled.

How Interference Occurs

Interference happens whenever treatments can affect people beyond those directly treated. In social networks, if your friends receive an intervention (like a voter mobilization message), their behavior change might influence you even if you didn't receive the message. In vaccine trials, vaccinating some people in a community can reduce disease transmission, indirectly protecting unvaccinated individuals—a positive spillover. In online platforms, showing some users a new feature can change their behavior in ways that affect friends' experiences, creating network-wide effects.

Why This Matters for Causal Inference

When interference exists, the "average treatment effect" we typically estimate becomes ambiguous. Are we measuring the effect of treating one person while keeping everyone else untreated? The effect of treating everyone compared to treating no one? These are different quantities, and standard experimental designs that randomly assign individuals can produce biased estimates for either one. The comparison between treated and control groups becomes contaminated because control group members are exposed to spillovers from treated individuals.

What You Can Do

Design strategies include cluster randomization (assigning entire groups rather than individuals, like schools or villages), graph cluster randomization (using network structure to assign connected groups together), or spatial separation between treatment conditions. Analysis strategies include modeling spillover effects explicitly, using exposure mappings to define different "effective treatments," or focusing on estimands that acknowledge interference rather than ignoring it. The key is recognizing when SUTVA is implausible and planning accordingly rather than pretending interference doesn't exist.

SUTVA (Stable Unit Treatment Value Assumption) was formalized by Donald Rubin in the potential outcomes framework and serves as a foundational assumption for identifying causal effects from randomized experiments [1]. When SUTVA is violated due to interference between units, the causal quantity of interest becomes ambiguous because potential outcomes depend on the entire treatment assignment vector, not just a unit's own treatment [2].

The main limitation is that most methods for handling interference require strong assumptions about the structure of spillovers (such as that interference only occurs within neighborhoods of a certain distance). These assumptions are often made for mathematical convenience rather than being justified by substantive theory. In many social processes, interference can propagate through networks arbitrarily far through chains of influence, making even neighborhood-based assumptions questionable. Graph cluster randomization reduces bias but can substantially increase variance, especially when clusters are large. There is also a fundamental trade-off: to identify spillover effects, we need variation in treatment assignment patterns across networks, but standard experimental designs often don't provide this. Many proposed methods also require knowing the complete network structure, which is often unavailable or expensive to measure.