Why SUTVA Matters
A causal effect for a unit compares what would happen to that same unit under different treatment states. For the comparison to be meaningful, however, the treatment states need to have clear meanings, and the outcome assigned to one unit cannot silently vary with other units’ treatment. These requirements are collected under the Stable Unit Treatment Value Assumption, or SUTVA.
For a binary treatment, write \(A_i\in\{0,1\}\) for unit \(i\)'s treatment and \(Y_i(1)\) and \(Y_i(0)\) for that unit’s potential outcomes under treatment and control. The unit-level causal contrast is \(Y_i(1)-Y_i(0)\). The earlier material on consistency explains how the observed outcome connects to the potential outcome for the treatment actually received. SUTVA addresses a different question: whether those potential outcomes are well defined in the first place.
The Two Requirements Are Different
The first requirement concerns what treatment a unit receives. Suppose “treated” could mean either a brief phone call or several in-person sessions. If those versions can affect outcomes differently, the label \(A_i=1\) does not select one unique potential outcome. The issue remains even if units cannot influence one another: the treatment itself is not sufficiently specified.
The second requirement concerns whose treatment can affect a unit’s outcome. Even when treatment is precisely defined, unit \(i\)'s outcome may depend on whether a roommate, classmate, or nearby community member is treated. In that case, the outcome cannot generally be written as a function of \(i\)'s own treatment alone.
Formally, let \(\mathbf a=(a_1,\ldots,a_n)\) denote a treatment-assignment vector for \(n\) units, and let \(Y_i(\mathbf a)\) be unit \(i\)'s potential outcome under that entire assignment. No interference requires that changing the assignments of other units, while keeping \(a_i\) fixed, does not change unit \(i\)'s outcome:
This notation assumes the treatment levels themselves are well defined. If the treatment has versions with different effects, adding the assignment vector does not resolve the ambiguity; the intervention or its version must also be specified.
Worked Example: A Treatment Label With Multiple Versions
Imagine a hypothetical study in which “received nutrition counseling” is coded as \(A_i=1\). Some treated participants receive a single 20-minute phone call; others receive four in-person sessions. Suppose a participant’s potential outcome would be 6 units of a health score after the phone call and 9 units after the sessions. The control outcome is 4 units.
For this participant, the treatment contrast under the phone-call version is \(6-4=2\) score units. Under the session version it is \(9-4=5\) score units. If \(A_i=1\) does not specify which version is meant, the single value \(Y_i(1)\) is ambiguous: it could be 6 or 9. Consequently, the treatment contrast \(Y_i(1)-Y_i(0)\) is also ambiguous: it could be 2 or 5 score units.
This example concerns the well-defined-treatment requirement. It does not require any participant to affect another participant’s outcome. A study could avoid this particular ambiguity by specifying one version as the treatment, or by defining a treatment that includes a clear rule for which version is delivered.
Worked Example: An Outcome Depends on Another Unit’s Assignment
Consider two hypothetical neighbors, 1 and 2, in a study of a precisely specified mosquito-control service. Let \(Y_1(a_1,a_2)\) be neighbor 1’s potential number of mosquito bites in a week, where \(a_1\) and \(a_2\) indicate whether each neighbor receives the service. Suppose the potential outcomes for neighbor 1 are:
| Neighbor 1 assignment \(a_1\) | Neighbor 2 assignment \(a_2\) | \(Y_1(a_1,a_2)\), bites per week |
|---|---|---|
| 0 | 0 | 12 |
| 0 | 1 | 8 |
| 1 | 0 | 6 |
| 1 | 1 | 4 |
Hold neighbor 1’s assignment at \(a_1=0\). If neighbor 2 is untreated, neighbor 1’s outcome is 12 bites; if neighbor 2 is treated, it is 8 bites. Thus \(Y_1(0,0)\ne Y_1(0,1)\), so the no-interference condition fails. The service may reduce mosquitoes that reach both properties, making neighbor 2’s assignment relevant to neighbor 1’s outcome.
The treatment is precisely specified, so the problem is not hidden treatment versions. The contrast for neighbor 1 also depends on the other neighbor’s assignment: it is \(Y_1(1,0)-Y_1(0,0)=6-12=-6\) bites per week when neighbor 2 is untreated, and \(Y_1(1,1)-Y_1(0,1)=4-8=-4\) bites per week when neighbor 2 is treated. A single contrast indexed only by neighbor 1’s treatment does not capture these possibilities.
Why the Requirements Matter for Causal Contrasts
The usual unit-level contrast \(Y_i(1)-Y_i(0)\) compares two potential outcomes for the same unit. It is interpretable only if each potential outcome refers to a clear treatment state and is not additionally indexed by other units’ assignments. The logic is direct: if treatment “1” contains several outcome-relevant versions, \(Y_i(1)\) has no single value; if another unit’s assignment matters, the relevant potential outcome is \(Y_i(\mathbf a)\), not merely \(Y_i(a_i)\).
This distinction also clarifies why SUTVA is not a claim that all units respond identically. Units can have different \(Y_i(1)\), different \(Y_i(0)\), and different causal effects. SUTVA says that the potential outcomes for a given unit are defined with respect to the stated intervention and, under no interference, that the other units’ assignments do not change those outcomes. It does not rule out variation in treatment effects across units.
Worked Example: Diagnosing Which Requirement Is at Risk
In a hypothetical school study, the treatment is described as “access to a tutoring program.” Schools can choose either weekly small-group sessions or a self-guided online package. Students also share materials and study together, so one student’s access may affect classmates’ learning. Consider the two questions separately.
Are treatment values well defined? Not from this description alone. The treated label includes at least two forms of tutoring, and they may have different effects. Unless the study specifies a version or defines a single intervention that includes the choice rule, the treated potential outcome is not uniquely described.
Is there interference? It is plausible. A student’s outcome may depend on whether classmates receive access, because students share materials or help one another. To see the implication, suppose a student’s hypothetical score is 70 with no one treated, 74 when only that student is treated, 73 when only a classmate is treated, and 78 when both are treated. The student’s outcome under own treatment \(a_i=0\) changes from 70 to 73 as the classmate’s assignment changes, so no interference fails in this illustration.
The scenario raises both concerns, but they are logically distinct. Specifying one tutoring version would address the treatment-definition concern; it would not remove peer effects. Conversely, eliminating peer contact would not make the two tutoring versions equivalent or specify which version “treated” means.
Assessment, Alternatives, and Limits
SUTVA is an assumption about how the causal system is represented, not a statistical test that can ordinarily be confirmed from observed outcomes alone. For each unit, only one treatment state is observed. The unobserved potential outcomes cannot generally reveal whether a different treatment version would have produced another outcome or whether changing a neighbor’s assignment would have changed the unit’s outcome.
Study design and subject-matter knowledge can reveal warning signs. Before analysis, ask what exactly counts as treatment, whether different providers or delivery modes might matter, whether units interact, and whether treatment can affect shared environments. Maps, contact patterns, delivery records, and protocol deviations can inform these questions, but a lack of obvious interactions does not prove that interference is absent. Likewise, a single recorded treatment label does not prove that all treated units received substantively identical interventions.
If the standard assumptions are implausible, the response is not simply to ignore the issue. One option is to refine the intervention so that treatment versions are specified. Another is to represent relevant exposure to others’ treatment—for example, by defining outcomes according to a unit’s own treatment and a clearly specified measure of peer exposure. Designs that assign treatment to groups can also be considered when spillovers are expected. These approaches require defining different causal quantities; they do not make the original no-interference assumption true by wording alone.
| Approach | What it represents | Key caution |
|---|---|---|
| Standard SUTVA formulation | One well-defined treatment per level; each outcome depends on own assignment only | May be implausible when treatment versions or spillovers matter |
| Specify the intervention | A clearer treatment level with a defined delivery or version | Does not by itself eliminate interference |
| Represent other-unit exposure | Outcomes may depend on own treatment and specified exposure to others’ treatment | Requires a defensible exposure definition and suitable data |
A common mistake is to treat “no interference” and “no hidden treatment versions” as interchangeable. They are not: the first concerns dependence across units, while the second concerns ambiguity within a treatment label. Another mistake is to infer SUTVA from random assignment. Random assignment concerns how treatment is allocated; it does not itself ensure that treatments have one well-defined version or that one unit cannot affect another. SUTVA also does not guarantee that treatment assignment is unconfounded in an observational study, nor does it ensure adequate data or eliminate measurement error.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A study calls a treatment “exercise support,” but treated participants receive either a wearable activity monitor or weekly coaching. Which SUTVA requirement is in question, and why?
- Suppose \(Y_i(0,0)=15\) and \(Y_i(0,1)=11\), where the first argument is unit \(i\)'s assignment and the second is a neighbor’s assignment. What does this imply?
- Can SUTVA fail even if treatment is randomized? Explain.
- In one sentence each, distinguish the two SUTVA requirements.
- If a study has no evidence of spillovers in its observed sample, has it established no interference? Explain.
Solutions. (1) The well-defined-treatment requirement is in question: “exercise support” combines two potentially outcome-relevant versions, so the treated potential outcome may be ambiguous. (2) No interference fails, because the neighbor’s assignment changes unit \(i\)'s outcome while \(i\)'s own assignment remains 0; the outcome changes by \(11-15=-4\) units. (3) Yes. Randomization governs assignment, but it does not guarantee a single treatment version or prevent one unit’s treatment from affecting another unit’s outcome. (4) Well-defined treatment requires each treatment level to identify a sufficiently specified intervention; no interference requires a unit’s outcome not to depend on other units’ assignments. (5) No. Observed data alone generally cannot establish the unobserved counterfactual outcomes needed to rule out spillovers; study design and substantive knowledge can assess plausibility but do not turn absence of observed evidence into proof.