Stable Unit-Treatment-Value Assumption

(Wikipedia)

The assumption in Rubin’s Causal Model that the response on one unit should be unaffected by the assignment of treatments to the other units, no matter what treatments other units receive. In other words 1) there is no inference across units and 2) treatment means the same thing for every unit. This assumption requires us to be clear about units, treatments, and potential outcomes.

SUTVA is violated when, for example

• there exist unrepresented versions of treatment (e.g. because no agreement on which version of treatment should be used)
• a treatment unit crosses over to control - drop the treatment
• a treatment interacts/interferes with a control unit

SUTVA simplifies the problem of quantifying the observed outcome for different pairs of units. If we don’t have SUTVA, the problem will be complicated.