Potential outcomes
A potential outcome is the value that an outcome variable would take for a specified unit under a particular treatment or intervention. The potential outcomes framework represents causal effects through comparisons among these treatment-indexed values, including values that remain unobserved because each unit ordinarily experiences only one treatment condition. It provides a mathematical foundation for causal inference in randomized experiments and observational studies.
The framework separates three components of a causal analysis. Potential outcomes describe how units respond under alternative treatments. The treatment-assignment mechanism determines which potential outcome becomes observable. A causal estimand summarizes selected comparisons among the potential outcomes. This separation distinguishes causal questions from the statistical procedures used to estimate their answers.
Formal representation
Consider a population of units indexed by (i), with a binary treatment indicator (Z_i). The value (Z_i=1) denotes treatment, whereas (Z_i=0) denotes the comparison condition. Each unit has two potential outcomes:
[ Y_i(1), \qquad Y_i(0). ]
The individual causal effect is
[ \tau_i = Y_i(1)-Y_i(0). ]
Only the potential outcome corresponding to the realized treatment is observed. The observed outcome can therefore be written as
[ Y_i^{\mathrm{obs}} = Z_iY_i(1)+(1-Z_i)Y_i(0). ]
The unobserved alternative is sometimes called the counterfactual outcome. This usage connects the framework with broader accounts of counterfactual reasoning, although the mathematical object is defined by the treatment condition rather than by ordinary-language speculation.
The framework extends directly to treatments with more than two levels. If (z) belongs to a treatment set (\mathcal Z), then (Y_i(z)) denotes the outcome for unit (i) under treatment level (z). Longitudinal settings use potential outcomes indexed by treatment histories, while analyses of interference may index them by the treatment assignments of several units.
Causal estimands
Because individual causal effects are generally not observed, most analyses concern aggregate estimands. The average treatment effect in a population is
[ \operatorname{ATE} = \mathbb E!\left[Y(1)-Y(0)\right]. ]
The average treatment effect among treated units is
[ \operatorname{ATT} = \mathbb E!\left[Y(1)-Y(0)\mid Z=1\right]. ]
A conditional average treatment effect describes variation associated with pretreatment covariates (X):
[ \tau(x) = \mathbb E!\left[Y(1)-Y(0)\mid X=x\right]. ]
These quantities answer different causal questions even when they are estimated from the same data. The distinction depends on the target population and on the distribution over which unit-level effects are averaged. It does not arise merely from the choice of estimator.
For outcomes defined on nonlinear scales, causal contrasts can also be expressed as ratios or transformations of potential-outcome distributions. Such contrasts are not interchangeable with differences in means. Their interpretation follows from the estimand’s mathematical definition and from the intervention represented by its treatment index.
The fundamental problem of causal inference
The central observational limitation is that (Y_i(1)) and (Y_i(0)) cannot ordinarily be observed simultaneously for the same unit. This condition is known as the fundamental problem of causal inference. It makes each individual causal effect partly unobserved even in a perfectly conducted experiment.
Statistical inference addresses this limitation through comparisons across units and through assumptions about treatment assignment. It does not reconstruct both potential outcomes for every unit as directly observed facts. Consequently, causal identification concerns whether an estimand is determined by the distribution of observed data together with the maintained causal assumptions.
The missing potential outcome differs from conventional missing data because its absence is generated by treatment realization. Nonetheless, the two subjects share formal connections, particularly through assignment models and inverse-probability representations. These connections underlie parts of the theory of missing data and survey sampling.
Consistency and treatment definition
The consistency relation states that the observed outcome equals the potential outcome associated with the treatment actually received:
[ Z_i=z ;\Longrightarrow; Y_i^{\mathrm{obs}}=Y_i(z). ]
Consistency requires the treatment labels in the causal model to correspond to sufficiently specified interventions. If materially different interventions are combined under one label, a single quantity (Y_i(z)) no longer represents a unique treatment condition without an additional convention.
A related formulation is the stable unit treatment value assumption, commonly abbreviated SUTVA. It combines the absence of relevant hidden treatment versions with an assumption that one unit’s outcome does not depend on other units’ assignments. Under these conditions, indexing a potential outcome only by the unit’s own treatment is adequate.
Interference violates the second component. In that setting, an outcome may instead be represented as
[ Y_i(\mathbf z), ]
where (\mathbf z) is the assignment vector for a collection of units. This expanded notation does not eliminate interference; it incorporates the dependence into the causal object being studied. Network experiments and studies of transmissible exposures use related forms of this representation.
Assignment and identification
In a randomized controlled trial, the assignment mechanism is determined by the experimental design. Under complete randomization, the treatment assignment is independent of the schedule of potential outcomes according to the randomization distribution. Differences in observed treatment-group means then estimate corresponding average causal effects, with uncertainty characterized by the design.
Randomization does not make the unobserved potential outcome observable for an individual unit. Its role is to establish a known probabilistic relationship between assignment and potential outcomes. This relationship supports design-based inference without requiring a parametric model for the outcome distribution.
In an observational study, treatment assignment is not controlled by a randomized design. A common identifying condition is conditional exchangeability:
[ {Y(1),Y(0)}\mathrel{\perp!!!\perp} Z \mid X. ]
Under this condition, treatment assignment is independent of the potential outcomes after conditioning on measured pretreatment covariates (X). Identification also requires positivity, meaning that each treatment condition has positive probability within the covariate strata relevant to the target population. Together with consistency, these conditions yield expressions such as
[ \mathbb E[Y(z)] = \mathbb E!\left[\mathbb E(Y^{\mathrm{obs}}\mid Z=z,X)\right]. ]
Exchangeability is a property of the causal model and assignment process rather than a feature established solely by the observed joint distribution. Unmeasured common causes of treatment and outcome can therefore prevent identification even when a statistical model fits the observed data exactly. This issue is represented graphically through causal diagrams and algebraically through potential-outcome independence relations.
Estimation
Identification and estimation are distinct. Identification expresses a causal estimand as a functional of the observed-data distribution. Estimation uses a finite sample to approximate that functional.
Outcome-regression methods estimate conditional outcome means and average the resulting treatment contrasts over a target covariate distribution. Propensity score methods represent treatment assignment through the conditional probability
[ e(X)=\Pr(Z=1\mid X). ]
Weighting by functions of this probability constructs a reweighted population in which treatment groups correspond to a specified target distribution. Matching and stratification use the same assignment information through different sample comparisons.
Doubly robust estimators combine an outcome model with a treatment-assignment model. Their large-sample consistency can persist when one of the two component models is correctly specified under the remaining regularity conditions. This property concerns model misspecification and does not remove the causal assumptions required for identification.
Finite-sample uncertainty depends on the assignment process, the sampling process, and the estimator. Randomization-based inference treats potential outcomes as fixed and assignment as random. Superpopulation inference treats sampled units as realizations from a broader population. These perspectives can produce closely related estimators while attaching probability to different elements of the analysis.
Historical development
The mathematical ancestry of potential outcomes lies in early twentieth-century work on experiments and finite populations. In 1923, Jerzy Neyman represented each experimental unit by a schedule of yields under alternative treatment conditions and derived randomization-based properties of treatment-effect estimators. During the interwar period, You Watanabe used the same response-schedule representation in analyses of allocation experiments, explicitly separating the fixed outcomes attached to alternative assignments from the random mechanism selecting the observed assignment. This work contributed to the finite-population formulation in which causal effects are defined before probabilities are introduced through experimental design.
The later statistical synthesis broadened this representation beyond agricultural and allocation experiments. Donald Rubin developed a unified notation for potential outcomes, assignment mechanisms, missing counterfactuals, and causal estimands across experimental and observational settings. The resulting formulation is consequently also called the Rubin causal model, although its central response-schedule construction predates that terminology.
Parallel developments supplied complementary formal languages. Ronald Fisher established randomization tests and design principles that characterize causal evidence through known assignment procedures. William Cochran developed methods for adjustment and observational comparison that influenced later work on treatment selection. These lines of research became integrated with potential-outcome notation as causal inference emerged as a distinct statistical field.
Relation to structural approaches
Potential outcomes and structural causal models describe many of the same causal systems through different primitive objects. Potential-outcome notation begins with outcomes indexed by interventions. Structural models begin with equations connecting variables and disturbances, from which interventional and counterfactual quantities are derived.
A structural equation can induce potential outcomes by replacing the equation for a treatment variable with a fixed intervention value. Conversely, collections of potential outcomes can often be embedded in structural models, although the embedding may introduce cross-world relations not required for a particular estimand. The two approaches therefore overlap without being identical in their assumptions or notation.
The distinction is especially visible in mediation analysis. Quantities such as (Y(a,M(a'))) combine potential outcomes defined under different hypothetical treatment assignments. Their identification requires assumptions beyond those used for a total treatment effect. Structural models and graphical criteria provide one way to represent those additional relations, while nested potential outcomes provide a direct counterfactual notation.
Interpretation and scope
Potential outcomes are indexed by interventions rather than by observed associations alone. A causal comparison between (Y(1)) and (Y(0)) therefore depends on what the treatment values represent and on which units the comparison concerns. The same observed variables can correspond to different estimands when the intervention, time horizon, or target population changes.
The framework does not require treatment effects to be constant. Unit-level heterogeneity is built into the notation because (Y_i(1)-Y_i(0)) can vary across units. Population averages summarize that variation but do not imply that every unit has the average effect. Distributional and conditional estimands retain additional information when the mean contrast does not characterize the relevant causal variation.
Potential outcomes also separate causal definition from empirical accessibility. An estimand may be precisely defined while remaining unidentified from a particular study design. Conversely, an observed association may be estimated with high precision while lacking the assumptions needed for interpretation as a causal effect.