SUTVA: The Assumption That Your Treatment Isn't Leaking
SUTVA assumes your treatment on one person doesn't spill over to affect others, and that the treatment is consistent for all. It's a key assumption for A/B tests, but is violated when one person's vaccine protects their unvaccinated neighbor.
WHY IT EXISTS To get a clean, trustworthy answer to "did my change have an effect?", we need to isolate the change. If the treatment for one person affects another, or if the "treatment" itself is inconsistent, we can't be sure what we're actually measuring. SUTVA provides the ground rules for a fair comparison between treatment and control.
THE MENTAL MODEL Think of SUTVA as building walls around your experiment. The first wall, "no interference," prevents the treatment group from leaking its effect onto the control group. The second wall, "consistency of treatment," ensures that everyone in the treatment group is actually getting the exact same treatment. Without these walls, your experiment is contaminated.
HOW IT WORKS SUTVA is an assumption you must verify about your system before running an analysis. It's composed of two sub-assumptions. First, no interference: The outcome for any unit is unaffected by the treatment assignments of other units. My outcome depends only on whether I got the treatment, not whether you did. Second, consistency: For any unit, there is only one version of the treatment. If a unit is assigned the treatment, the outcome is the same regardless of how it was administered.
WHEN TO USE IT You don't "use" SUTVA; you check for it. It is a prerequisite for most standard A/B tests and causal inference. Before comparing the means of a treatment and control group, you must ask if SUTVA plausibly holds. For example, testing a new button color on a website for individual, independent users likely satisfies SUTVA because one user's experience doesn't affect another's.
WHEN NOT TO USE IT SUTVA is often violated, which requires more advanced methods. A common violation is interference in social networks. Giving a new feature to some users might cause their friends (in the control group) to become more engaged, too. Another is in marketplaces: discounting one seller's product can hurt sales of a competing seller. In these cases, you may need cluster-based randomization (e.g., by city instead of by user) to contain the spillover effects.
ONE CANONICAL EXAMPLE A vaccine trial. The treatment is the vaccine. If person A gets the vaccine, they are less likely to get sick. But they are also less likely to transmit the disease to person B, even if person B is in the control group. Person A's treatment has affected person B's outcome. This is a classic violation of the "no interference" part of SUTVA due to herd immunity.
Read the original → en.wikipedia.org
Get five bites like this every day.
Tezvyn delivers a daily feed of 60-second tech bites with quizzes to lock in what you learn.