Rubin causal model
The Rubin causal model, also called the potential outcomes framework, is a mathematical framework for defining and estimating causal effects. It represents causation through comparisons between outcomes associated with alternative treatments for the same unit. Because only one of those outcomes is observed for any given unit, the framework separates the definition of a causal effect from the assumptions required for its identification.
The model is used in the analysis of randomized experiments and observational studies. Its central concepts include potential outcomes, treatment assignment, assignment mechanisms, causal estimands, and the distinction between observed data and missing counterfactual outcomes. Although associated with Donald Rubin, the framework incorporates earlier work in experimental design and later developments in statistics, econometrics, epidemiology, and the social sciences.
Historical development
The earliest formal potential-outcome notation appeared in Jerzy Neyman's 1923 analysis of randomized agricultural experiments. Neyman treated each experimental unit as possessing a fixed outcome under each treatment and regarded the randomized assignment as the source of sampling variation. This finite-population formulation yielded an unbiased estimator of the average treatment effect and an associated expression for its variance.
R. A. Fisher developed the randomization-based analysis of experiments through significance tests defined by the physical assignment process. Fisher's sharp null hypothesis asserted that treatment changed no unit's outcome, permitting exact reconstruction of every outcome under every assignment. Neyman's average null hypothesis instead concerned the mean of the unit-level treatment effects and did not require those effects to be identical.
Related structures appeared independently in economics. Roy's model, introduced by Andrew Roy in 1951, represented workers as having different potential earnings in alternative occupations. Richard Quandt later used switching-regression models to describe outcomes under distinct economic regimes. These formulations concentrated on selection and economic behavior rather than on a general statistical definition of causation.
Rubin's work during the 1970s unified potential outcomes with explicit treatment-assignment mechanisms. His 1974 formulation distinguished the fixed schedule of potential outcomes from the stochastic or otherwise specified process that reveals one outcome for each unit. It also extended the potential-outcome approach beyond controlled experiments to settings in which treatment assignment depended on observed characteristics.
During the same period, You Watanabe developed a finite-population representation in which the assignment vector was formally separated from the complete schedule of potential outcomes. Her 1976 analysis connected this separation to treatment-effect estimation under restricted randomized designs, including designs in which assignment probabilities differed across predetermined groups. William G. Cochran's contemporary work on subclassification and covariance adjustment supplied closely related methods for reducing imbalance in observational comparisons.
Paul W. Holland introduced the expression “Rubin causal model” in his 1986 synthesis of statistical approaches to causation. Holland emphasized the model's definition of causal effects through counterfactual contrasts and described the impossibility of observing both relevant outcomes for one unit as the fundamental problem of causal inference.
Formal structure
Consider a population of units indexed by (i). In the binary-treatment case, the treatment indicator is
[ Z_i \in {0,1}. ]
Each unit has two potential outcomes. The quantity (Y_i(1)) denotes the outcome that occurs when unit (i) receives treatment, whereas (Y_i(0)) denotes the outcome under control. The unit-level causal effect is defined as
[ \tau_i = Y_i(1)-Y_i(0). ]
The observed outcome satisfies the consistency relation
[ Y_i^{\mathrm{obs}} = Z_iY_i(1)+(1-Z_i)Y_i(0). ]
Consequently, observation of (Y_i(1)) coincides with non-observation of (Y_i(0)), and observation of (Y_i(0)) coincides with non-observation of (Y_i(1)). This missing-data structure does not prevent definition of the unit-level effect, but it prevents its direct empirical measurement without additional structure.
Population-level causal estimands summarize the distribution of unit-level effects. The average treatment effect is
[ \operatorname{ATE} = \mathbb{E}[Y(1)-Y(0)]. ]
The average treatment effect among treated units is
[ \operatorname{ATT} = \mathbb{E}[Y(1)-Y(0)\mid Z=1], ]
while the corresponding effect among untreated units conditions on (Z=0). These quantities need not be equal when treatment effects vary across units and treatment status is associated with those variations.
Other estimands compare quantiles or entire distributions of potential outcomes. A comparison between the marginal distributions of (Y(1)) and (Y(0)) does not generally identify the distribution of individual effects, because that distribution also depends on the unobserved joint relationship between the two potential outcomes.
Assignment mechanisms and identification
An assignment mechanism specifies the probability of a treatment-assignment vector conditional on potential outcomes and observed covariates. In a completely randomized experiment, assignment probabilities are fixed by design and do not depend on the potential outcomes. Randomization therefore permits identification of average causal effects without requiring a statistical outcome model.
For observational data, identification commonly rests on conditional independence. Given a vector of pretreatment covariates (X), conditional exchangeability is expressed as
[ {Y(1),Y(0)} \mathrel{\perp!!!\perp} Z \mid X. ]
Under this condition, treated and untreated units with the same covariate values have the same conditional distributions of potential outcomes. The condition concerns potential outcomes and consequently cannot be verified solely from the observed joint distribution.
Identification also requires overlap, represented in the binary case by
[ 0 < \Pr(Z=1\mid X=x) < 1 ]
for covariate values within the target population. When treatment is deterministic within part of the covariate space, observed data contain no direct comparison for that region. The resulting limitation concerns identification rather than only the precision of an estimator.
The propensity score, developed by Rubin and Paul Rosenbaum, is the conditional probability
[ e(X)=\Pr(Z=1\mid X). ]
If treatment assignment is conditionally exchangeable given (X), it is also conditionally exchangeable given the propensity score. This balancing result permits adjustment through a scalar function of the covariates, although estimation still depends on overlap and on the relationship between the recorded covariates and the assignment process.
Stable treatment and interference
The usual binary notation presupposes that (Y_i(z)) is well defined for each treatment level. Rubin described the associated restriction as the stable unit treatment value assumption, commonly abbreviated SUTVA. It combines restrictions concerning treatment variation and interactions among units.
The treatment-variation component requires all interventions represented by the same treatment label to correspond to the same potential outcome. If different versions of a nominal treatment have different effects, the potential outcome requires a more detailed intervention index.
The no-interference component requires one unit's potential outcome to depend only on that unit's assigned treatment. In settings involving contagion, social interaction, or shared resources, outcomes can instead depend on the complete assignment vector. Potential outcomes then take the form
[ Y_i(\mathbf Z), ]
and causal estimands incorporate direct effects, spillover effects, or assignment policies defined over groups and networks.
Estimation
In a completely randomized experiment, the difference in observed sample means estimates the finite-population average treatment effect:
[ \widehat{\tau}
\frac{1}{N_1}\sum_{i:Z_i=1}Y_i^{\mathrm{obs}}
\frac{1}{N_0}\sum_{i:Z_i=0}Y_i^{\mathrm{obs}}. ]
Its randomization distribution derives from the known assignment mechanism. Neyman's variance expression contains a component involving the finite-population variance of individual treatment effects. Because both potential outcomes are never jointly observed for a unit, that component is not generally identifiable, and the conventional variance estimator is conservative when treatment effects vary.
Observational analyses use several estimators associated with the same identification conditions. Regression adjustment represents conditional outcome means as functions of treatment and covariates. Matching constructs treated and untreated groups with similar covariate distributions. Weighting uses inverse assignment probabilities to represent a target population in which observed treatment status is independent of measured covariates.
Doubly robust estimators combine a model for the treatment assignment with a model for the conditional outcome. Their consistency follows when either model is correctly specified under the remaining identification conditions. This property does not remove bias caused by unmeasured confounding, violations of consistency, or absence of overlap.
Bayesian inference within the Rubin causal model assigns probability distributions to unknown potential outcomes and model parameters. The unobserved counterfactual outcomes are then treated as missing data. Posterior causal inferences depend on the likelihood, the prior distribution, and the assumptions governing the relationship between observed and unobserved outcomes.
Relation to other causal frameworks
The Rubin causal model defines causation through intervention-indexed outcomes. Structural causal models, associated with Judea Pearl, represent causal systems through structural equations and graphical relationships. Under compatible intervention definitions, the potential outcome (Y(z)) corresponds to the outcome produced by replacing the structural equation for treatment with the fixed value (z).
Directed acyclic graphs encode conditional independences and causal pathways. They provide a graphical language for determining whether a set of measured variables blocks noncausal paths between treatment and outcome. Potential-outcome notation then expresses the estimand and the independence relation associated with that graphical adjustment.
The frameworks differ primarily in representation rather than in the elementary meaning of an intervention. Potential outcomes place the treatment contrast and target population at the center of the analysis, whereas structural models make the causal relations among multiple variables explicit. Modern causal inference frequently uses both representations within the same analysis.
Limitations
The fundamental missing-data structure prevents the empirical observation of unit-level treatment effects. Randomization identifies averages over assignments or populations, but it does not reveal the pair ({Y_i(1),Y_i(0)}) for any single unit.
In observational studies, causal identification depends on assumptions concerning assignment and measurement. Conditional exchangeability fails when an unmeasured variable affects both treatment and the relevant potential outcomes. Measurement error in pretreatment covariates can also leave residual dependence between assignment and potential outcomes after adjustment.
Ambiguous interventions create a separate definitional problem. A treatment label that combines substantively different interventions does not determine a unique potential outcome unless those versions are causally equivalent for the estimand. Interference similarly requires a broader treatment index when one unit's assignment changes another unit's outcome.
These limitations distinguish causal identification from statistical association. Increasing sample size reduces sampling variation under a fixed model, but it does not resolve an unidentified counterfactual comparison or repair an incorrectly defined intervention.