SUTVA and No-Interference Failures

Why No Interference Matters

A causal effect is usually described by comparing what would happen to the same unit under different treatment assignments. That comparison becomes ambiguous if one person’s outcome also depends on whether other people are treated. A vaccine can affect an unvaccinated person through reduced transmission; a peer program can affect students who do not participate; and a service intervention can change how a limited resource is distributed. In each case, the treatment assignment of one unit may alter another unit’s outcome.

The no-interference condition rules out that dependence. It is part of the Stable Unit Treatment Value Assumption, usually abbreviated SUTVA. The condition is not a claim that people never interact or that outcomes are statistically independent. It is a specific claim about potential outcomes: holding a unit’s own treatment fixed, changing other units’ assignments does not change that unit’s outcome.

Definition: Let $A_i$ denote unit $i$’s treatment assignment, and let $A_{-i}$ denote the assignments of all other units. Write $Y_i(a_i,a_{-i})$ for the potential outcome of unit $i$ under the full assignment vector. No interference holds if, for every unit $i$, every own assignment $a_i$, and every pair of assignments $a_{-i}$ and $a'_{-i}$ to the other units, $$ Y_i(a_i,a_{-i})=Y_i(a_i,a'_{-i}). $$ When this condition holds, the potential outcome can be written $Y_i(a_i)$: other units’ assignments do not need to appear in its arguments.

The full-vector notation is useful for stating what the assumption requires, even when the study’s usual notation uses only $Y_i(0)$ and $Y_i(1)$. Those two potential outcomes are adequate only if changing other units’ assignments cannot change either outcome. If interference is possible, there may instead be several potential outcomes for unit $i$ under the same own assignment, depending on the assignments around that unit.

Connection to SUTVA and the Usual Notation

SUTVA is commonly described as including two related requirements: no interference between units and no hidden versions of treatment. The first concerns whether other units’ assignments affect unit $i$’s outcome. The second concerns whether the treatment label itself stands for materially different versions. For example, “received the program” may describe different program intensities or delivery methods. A well-defined treatment is needed to interpret its potential outcomes, but it does not by itself guarantee no interference.

Under no interference, the potential outcome for unit $i$ under own assignment $a_i$ is unaffected by the rest of the allocation. The observed outcome is then the potential outcome corresponding to that unit’s realized assignment. In full-vector notation, if the realized assignment is $(A_i,A_{-i})$, the observed outcome is $Y_i(A_i,A_{-i})$. No interference lets us simplify this to $Y_i(A_i)$. This simplification is a statement about the outcome-generating structure, not a consequence of random assignment.

Y_i(A_i,A_{-i})=Y_i(A_i) \quad\text{under no interference.}

A short implication follows directly from the definition. Fix unit $i$ and own assignment $a_i$. If no interference holds, the equality in the definition says that $Y_i(a_i,a_{-i})$ has the same value for every possible $a_{-i}$. Therefore it has a single value for that $a_i$, which can be denoted $Y_i(a_i)$. If even two other-unit assignments produce different values, that simplification is not valid for that unit and assignment.

Diagnosing Interference in Practice

A practical diagnostic starts by stating the unit, the treatment assignment, and the outcome precisely. Then ask: Could changing another unit’s assignment, while holding this unit’s assignment fixed, change this unit’s outcome? The answer should be considered across plausible ways treatment could spread or resources could be shared, not only through direct face-to-face contact. Potential pathways include transmission, information sharing, competition, congestion, displacement, and changes to a common environment.

The key is the counterfactual comparison. If the answer is yes for a relevant assignment configuration, no interference fails. This is not established merely because units are connected or because their outcomes are correlated. Conversely, a study may have interference even if the investigator did not measure a connection between units: a shared resource or common exposure can transmit effects without a recorded social tie.

Worked Example: Vaccination and Infection Risk

Consider a hypothetical household with an index person $i$, who is not vaccinated, and one other household member $j$. Let $A_i=0$ mean the index person is unvaccinated, and let $A_j=0$ or $1$ indicate whether the other member is unvaccinated or vaccinated. Suppose the hypothetical potential infection outcomes for the index person are:

Index assignment $A_i$	Other member assignment $A_j$	Potential infection outcome $Y_i(A_i,A_j)$
0	0	1
0	1	0

These are deliberately simplified, hypothetical binary outcomes: 1 means infection and 0 means no infection. Holding the index person’s assignment fixed at $A_i=0$, compare the two potential outcomes:

$$ Y_i(0,1)-Y_i(0,0)=0-1=-1. $$

The value changes when the other household member’s assignment changes. Thus, for this hypothetical assignment pattern, the index person’s outcome depends on another unit’s treatment, and no interference fails. The example represents a possible indirect pathway: vaccinating the other person may reduce the index person’s exposure. It does not establish that this particular pattern occurs in real households, nor does one hypothetical pair quantify a population effect.

Worked Example: A Peer Program in a Study Group

Suppose two students are in a hypothetical study group. The outcome is a quiz score in points. Student $i$ does not participate in a tutoring program, so $A_i=0$. Let $A_j=1$ mean the other student participates, and suppose the hypothetical potential scores for student $i$ are $Y_i(0,0)=62$ when neither participates and $Y_i(0,1)=68$ when only the peer participates.

The own assignment remains fixed at zero. The change in the potential score across the peer assignments is:

Y_i(0,1)-Y_i(0,0)=68-62=6\text{ points}.

Because this difference is not zero, the no-interference condition fails for this hypothetical student. A plausible pathway is that the participating peer shares study strategies or notes. A study analyzing only each student’s own participation would not capture that dependence in the usual $Y_i(A_i)$ notation. The example diagnoses a possible failure; the six-point contrast is not evidence about actual students or a general estimate for a peer program.

Worked Example: A Shared Appointment Resource

Imagine a hypothetical clinic with two patients and a limited number of early appointment slots. Patient $i$ does not receive a scheduling intervention, so $A_i=0$. Patient $j$ either does not receive the intervention ($A_j=0$) or receives priority scheduling ($A_j=1$). The outcome for patient $i$ is waiting time in days. Suppose the hypothetical potential outcomes are $Y_i(0,0)=2$ days and $Y_i(0,1)=5$ days because the priority appointment uses a slot that would otherwise have been available to patient $i$.

With patient $i$’s assignment held fixed, the difference is:

Y_i(0,1)-Y_i(0,0)=5-2=3\text{ days}.

The outcome differs by three days under the two assignments to patient $j$, so the no-interference condition fails in this hypothetical setting. The pathway is resource displacement rather than social contact. This example shows why a review should ask whether units compete for a shared resource, not only whether treatment can spread directly from one person to another.

What a Violation Does—and Does Not—Mean

When no interference fails, a conventional pair $Y_i(0),Y_i(1)$ may omit information needed to define the potential outcomes. For example, $Y_i(0)$ could differ depending on whether none, some, or many other units are treated. There is then no single value for $Y_i(0)$ unless the relevant context is specified or the analysis adopts a different representation. Treating the usual notation as if it were complete can therefore obscure which counterfactual comparison is intended.

A violation does not automatically make all causal analysis impossible. It means the assumptions and estimand must reflect how assignments can affect outcomes across units. The diagnosis should be followed by a clear description of the plausible pathways and of what is held fixed in the intended comparison. Merely fitting a more elaborate model or adjusting for measured covariates does not, by itself, restore no interference.

It is also important to separate interference from several nearby ideas:

Confounding: Confounding concerns common causes of treatment and outcome. Interference concerns whether another unit’s assignment changes a unit’s potential outcome. Randomization may address confounding from treatment assignment while leaving interference possible.
Outcome dependence: Outcomes can be correlated because units share background causes, even when changing one unit’s assignment would not change another unit’s outcome. Correlation alone is not a violation of no interference.
Contact or network connection: Contact can make interference plausible, but contact alone does not prove that another unit’s treatment changes the outcome. Interference can also occur through shared resources without direct contact.
Treatment versions: Different forms of a treatment can create ambiguity about what “treated” means. That issue is related to SUTVA but is distinct from dependence on other units’ assignments.

Common Mistake

A common mistake is to assume that randomizing individuals guarantees no interference: “Treatment was randomized, so each person’s outcome depends only on their own assignment.” Randomization determines how assignments are generated; it does not prevent one assignment from affecting another person’s outcome. In the vaccination example, assignments could be randomized and the unvaccinated person’s infection outcome could still depend on whether a household member was vaccinated.

The correction is to assess the potential-outcome condition directly. Hold a unit’s own assignment fixed and ask whether changing the assignments of other units could change its outcome. If yes, randomization alone does not justify writing the outcome as $Y_i(A_i)$. A second mistake is to treat any observed similarity between peers as proof of interference. Similar outcomes may instead arise from shared causes; the defining issue is a causal change under another unit’s assignment, not resemblance in observed data.

Key takeaway: No interference requires each unit’s potential outcome to remain unchanged when other units’ treatment assignments change, with its own assignment held fixed. Before using ordinary unit-level potential-outcome notation, consider transmission, peer influence, competition, and shared resources that could violate this condition.

Check Your Understanding

For each question, focus on whether changing another unit’s assignment could change the outcome while the focal unit’s own assignment stays fixed.

In a hypothetical vaccination study, what potential-outcome comparison would reveal interference for an unvaccinated focal person?
Two classmates have correlated test scores, but there is no pathway by which one student’s assignment can affect the other’s score. Does correlation alone show that no interference fails? Explain.
A program gives some patients priority access to a limited number of appointments. Identify a plausible way another patient’s assignment could affect the focal patient’s outcome.
Why does random assignment of treatment not, by itself, establish no interference?
How is dependence on other units’ assignments distinct from ambiguity caused by multiple versions of a treatment?