What is Potential Outcomes?
Potential Outcomes are the outcomes a unit would exhibit under each well-defined treatment or intervention level, whether or not that treatment is actually assigned, and they provide a formal basis for defining causal effects.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Define treatment versions and time before writing Y(a)
A label such as “uses the assistant” can hide different doses, models, permissions, start times, and adherence patterns. Potential Outcome Y_i(a) is meaningful only when intervention a, eligibility, time zero, outcome window, and handling of switching or censoring are specified. Otherwise different versions are collapsed into one symbol and Consistency cannot be defended.
SUTVA combines treatment-version and interference assumptions
The Stable Unit Treatment Value Assumption commonly requires a unique outcome for each stated treatment and no unmodeled dependence on other units' assignments. Network effects, shared capacity, contagion, ranking competition, and marketplace equilibrium can violate no-interference assumptions. The remedy is to redefine the exposure and assignment unit or model interference, not merely add more covariates.
Population effects are identifiable more often than individual effects
Naimi and Whitcomb show how Potential Outcomes define Average Treatment Effects and how Consistency, Exchangeability, and Positivity connect them to observed quantities. Even when an average is identified, the joint pair Y_i(1), Y_i(0) for one unit is not observed; personalized effects require additional structural assumptions and calibrated validation.
Key Characteristics
- Assigns a hypothetical outcome to each unit-treatment combination
- Defines causal effects by contrasts across treatment states
- Leaves at least one counterfactual outcome unobserved for each unit
- Requires well-defined treatment versions and outcome timing
- Makes interference assumptions explicit through SUTVA or extensions
- Supports population estimands without promising identifiable individual effects
Common Use Cases
- Defining the effect of a product feature versus a fixed control
- Separating observed outcomes from missing counterfactual outcomes
- Specifying ATE, ATT, ATC, or subgroup treatment effects
- Diagnosing interference in networks or shared-capacity systems
- Checking whether an observational study has a coherent causal question
Example
Loading code...Frequently Asked Questions
What is the difference between a Potential Outcome and an observed outcome?
A Potential Outcome is defined for a unit under a specified intervention whether or not it occurs. The observed outcome equals the Potential Outcome corresponding to the treatment actually received under Consistency; other treatment-state outcomes remain counterfactual.
Why can both Potential Outcomes not be observed for one unit?
At one decision time, a unit normally receives only one mutually exclusive treatment. Observing its outcome under treatment prevents simultaneous observation under control with everything else unchanged. This is the Fundamental Problem of Causal Inference.
What does SUTVA require?
In its common form, SUTVA requires no hidden versions of each treatment and no interference between units. If outcomes depend on peers' assignments or treatments vary materially in dose or implementation, the exposure and Potential Outcomes must be redefined.
Are Potential Outcomes the same as model predictions?
No. Potential Outcomes are causal quantities defined by interventions. A model prediction is an estimate based on data and assumptions. Predicting both treatment states does not make them identified, accurate, or individually verifiable.
Can Potential Outcomes define effects for continuous or multiple treatments?
Yes. The notation extends to `Y(a)` for multiple, continuous, dynamic, or time-varying strategies, but the intervention, feasible range, timing, adherence, and interference assumptions become more demanding and must be stated explicitly.