The Stable Unit Treatment Value Assumption (SUTVA) underpins the potential outcomes framework. It promises that each unit’s observed outcome reflects a single, well-defined potential outcome that depends only on the unit’s own treatment status. When SUTVA breaks, causal estimates target a moving object; versions of the treatment change the estimand, and spillovers smear counterfactual comparisons.
What SUTVA Actually Says
- No multiple versions: If unit $i$ receives treatment $d$, the realized potential outcome is $Y_i(d)$ regardless of how that treatment was delivered. Vaccine $A$ administered in the left arm must be equivalent to the same vaccine in the right arm.
- No interference across units: Potential outcomes depend only on the unit’s own treatment assignment: $Y_i(d) = Y_i(d, D_{-i})$ for all configurations $D_{-i}$ of other units’ treatments.
Combined, these clauses ensure the mapping $D_i \to Y_i(D_i)$ is well-defined. Formally, if you attempt to write $Y_i(d, g(D_{-i}))$ where $g$ summarizes others’ assignments, SUTVA insists $g$ must be irrelevant.
Why It Matters in Practice
- Randomized experiments: Randomization guarantees independence between treatment and potential outcomes only if each potential outcome is uniquely tied to a treatment level. Noncompliance or protocol deviations often signal multiple versions lurking in the background.
- Observational identification: Selection-on-observables or IV strategies inherit the same fragility. If interference is present, conditioning or instrumenting on $D_i$ alone still leaves contamination from $D_{-i}$.
- External validity: When the estimand aggregates over heterogeneous versions of a treatment (e.g., different dosage schedules), reporting an “average effect” masks which version drives the result and complicates replication.
Canonical Violation Patterns
- Network spillovers: Social exposure, peer effects, or infectious diseases let other units’ treatments shift $Y_i$. Cluster-randomized trials often arise explicitly to handle this.
- Implementation heterogeneity: Education programs delivered by multiple teachers or clinics may vary subtly. Without tracking those variants, $Y_i(1)$ becomes a mixture of implementation-specific effects.
- Supply constraints and interference through resources: When treated units consume scarce resources (teachers’ time, ICU beds), untreated units’ outcomes change indirectly.
- Measurement mismatch: If survey instruments or timing differ across arms, you effectively measure different versions of the outcome, violating the spirit of SUTVA.
Diagnosing SUTVA Risks
- Map the causal graph to identify paths where other units’ treatments can reach unit $i$. Group membership, shared markets, or shared staff are danger zones.
- Descriptive spillover checks: Plot outcomes for untreated units against the share of treated peers (neighbors, classmates, co-workers). Monotone relationships signal potential interference.
- Implementation logs: Track compliance, delivery protocols, and exposure intensity. Variation beyond a tolerable threshold means you are pooling multiple versions.
- Qualitative field work: Interviews and process tracing often uncover informal spillovers (information sharing, material lending) invisible in administrative data.
Design Remedies When SUTVA Fails
- Redefine the unit or treatment: Aggregate to classrooms, villages, or clusters where interference is contained, or split treatments into explicit variants (e.g., “SMS reminder + phone call”).
- Explicit exposure mapping: Model $Y_i(d, g(D_{-i}))$ with exposure metrics such as the number of treated neighbors or total treated dosage in a network (see Aronow & Samii, 2017).
- Two-stage randomization: Randomize clusters to saturation levels, then randomize individuals within clusters; estimators separate direct, spillover, and overall effects.
- Instrument for implementation details: Use encouragements or randomized rollout timing to uncover effects of specific treatment versions.
- Report partial interference models: Assume units interfere only within predefined groups. Variance estimators and hypothesis tests must reflect the cluster structure.
Reporting Checklist for SUTVA Transparency
- Describe how the treatment is delivered, who delivers it, and what fidelity monitoring occurred.
- Document potential channels for spillovers and whether the study design blocked or measured them.
- Provide balance tables or diagnostics for exposure variables (e.g., average treated neighbors) if interference is plausible.
- State clearly whether estimated effects should be interpreted as direct, indirect, or total effects in the presence of interference.
- Discuss how alternative treatment versions or rollout intensities might alter the estimand.
For Further Study
- Imbens & Rubin (2015), Causal Inference for Statistics, Social, and Biomedical Sciences, ch. 1-2 for the formal SUTVA statement.
- Hudgens & Halloran (2008) on causal inference in infectious disease settings with interference.
- Aronow & Samii (2017) for exposure mappings and estimands under arbitrary interference.
- Basse & Airoldi (2018) on randomization-based inference with network interference.
Treat SUTVA as a modeling choice instead of a background assumption. By articulating where it might fail–and designing around those failures–you keep causal estimates anchored to interpretable, policy-relevant quantities.