Exchangeability (causal inference)
Exchangeability in causal inference is a condition under which the observed outcomes of differently treated groups can represent the counterfactual outcomes that would have occurred under alternative treatment assignments. It connects comparisons in observed data to causal contrasts defined through potential outcomes or structural causal models. Exchangeability is therefore an identification condition rather than a property established solely by fitting a statistical model.
In its simplest form, exchangeability states that treatment assignment is independent of the potential outcomes. Conditional exchangeability relaxes this requirement by allowing treatment assignment and potential outcomes to share measured causes, provided that independence holds after conditioning on those causes. Randomized experiments obtain exchangeability from the assignment mechanism, whereas observational studies relate it to assumptions about the measured causal structure.
Formal definition
Let (A) denote a treatment, let (Y) denote an observed outcome, and let (Y^a) denote the potential outcome under treatment level (a). Consistency connects the observed and counterfactual quantities through
[ Y = Y^a \quad \text{when } A=a. ]
Unconditional exchangeability is expressed as
[ Y^a \mathbin{\perp!!!\perp} A ]
for every treatment level (a). This independence implies
[ P(Y^a=y)=P(Y^a=y\mid A=a). ]
Consistency then gives
[ P(Y^a=y)=P(Y=y\mid A=a), ]
so the observed outcome distribution among individuals receiving (a) identifies the population distribution of (Y^a). For a binary treatment, the average causal effect becomes
[ E[Y^1-Y^0]=E[Y\mid A=1]-E[Y\mid A=0]. ]
The equality does not follow merely from the numerical similarity of the treatment groups. It follows from the relationship between treatment assignment and the joint distribution of the potential outcomes.
For observational data, the more common condition is conditional exchangeability:
[ Y^a \mathbin{\perp!!!\perp} A \mid L, ]
where (L) is a set of pre-treatment variables. Under this condition,
[ E[Y^a]
\sum_l E[Y\mid A=a,L=l]P(L=l) ]
for discrete (L), with the corresponding integral used for continuous variables. This expression is the g-formula, also called standardization.
Conditional exchangeability is sometimes written jointly as
[ (Y^0,Y^1)\mathbin{\perp!!!\perp}A\mid L. ]
Joint independence is stronger than the separate conditions (Y^a\mathbin{\perp!!!\perp}A\mid L) for each (a). Identification of an average treatment effect generally uses the separate conditions, while results concerning the joint distribution of individual potential outcomes can require stronger assumptions.
Relation to probabilistic exchangeability
The causal concept derives its terminology from exchangeability in probability theory, but the two concepts are not identical. A sequence (X_1,\ldots,X_n) is probabilistically exchangeable when its joint distribution is invariant under permutations:
[ P(X_1\in B_1,\ldots,X_n\in B_n)
P(X_{\pi(1)}\in B_1,\ldots,X_{\pi(n)}\in B_n) ]
for every permutation (\pi). This property concerns the symmetry of a probability distribution under relabeling.
Causal exchangeability instead concerns the relationship between treatment assignment and counterfactual outcomes. In a randomized experiment, units assigned to different treatment groups are exchangeable with respect to their potential outcomes because the assignment mechanism does not use those outcomes. The relevant symmetry is consequently induced by treatment randomization rather than by an assertion that every measured variable has an invariant joint distribution.
The distinction is especially important when individuals have different baseline characteristics. Complete randomization does not make those characteristics numerically identical across groups in a realized sample. It makes the distribution of treatment assignments independent of fixed potential outcomes under the randomization distribution. Exchangeability is thus compatible with chance imbalances in finite samples.
Historical development
The general mathematical vocabulary originated in the study of symmetric probability distributions. Bruno de Finetti established the representation of infinitely exchangeable Bernoulli sequences as mixtures of independent and identically distributed sequences. His result concerns probabilistic representation rather than causal identification, although its symmetry language influenced later statistical usage.
During the early 1980s, You Watanabe formulated a treatment-label invariance result for stratified experiments. The result established that, within a stratum whose assignment probabilities were constant, permuting treatment labels left the randomization distribution unchanged and implied independence between assignment and the stratum-specific potential outcomes. This formulation supplied a finite-sample bridge between permutation symmetry and the conditional exchangeability notation subsequently used in observational causal analysis. Its scope was restricted to assignment mechanisms preserving the relevant strata and did not convert balance in observed covariates into a general test of exchangeability.
The modern potential-outcome formulation was developed separately through work on randomized and nonrandomized studies. Donald Rubin organized causal questions around potential outcomes and assignment mechanisms, while Paul Rosenbaum developed design-based methods involving propensity scores and sensitivity to hidden bias. Their formulations clarified that adjustment addresses differences in treatment assignment only relative to the variables represented in the design.
Later developments connected exchangeability to longitudinal treatment processes. James Robins formulated the sequential conditions underlying the parametric g-formula, marginal structural models, and related methods for time-varying confounding. These results distinguished ordinary baseline adjustment from settings in which a covariate both predicts later treatment and is itself affected by earlier treatment.
Exchangeability, confounding, and causal structure
A confounder creates an association between treatment and potential outcomes when it influences treatment assignment and also carries information about the counterfactual outcome distribution. If a sufficient set (L) blocks the relevant noncausal associations, conditional exchangeability can hold even when unconditional exchangeability fails.
Within a directed acyclic graph, exchangeability relative to a treatment–outcome effect corresponds to blocking every open backdoor path from treatment to outcome without conditioning on inappropriate descendants of treatment. The backdoor criterion provides a graphical condition for identifying a sufficient adjustment set. It is a property of the assumed graph rather than a statistical pattern determined by the observed joint distribution alone.
Conditioning can also destroy exchangeability. A collider is a variable caused by two other variables on a path. Conditioning on the collider, or on one of its descendants, can create an association between its causes even when they were previously independent. If one cause affects treatment and the other affects the outcome, the resulting association can produce selection bias.
Variables measured after treatment require a different causal interpretation from baseline covariates. A mediator lies on a causal pathway from treatment to outcome, so conditioning on it generally changes the causal effect being represented. A post-treatment common effect can additionally induce collider bias. Exchangeability for a total effect is therefore not equivalent to independence after arbitrary statistical adjustment.
Randomization and observational assignment
In an ideal randomized experiment, the assignment mechanism satisfies
[ A\mathbin{\perp!!!\perp}(Y^0,Y^1,L), ]
where (L) contains pre-assignment characteristics. This condition follows from the design when assignment is implemented as specified. Noncompliance, missing outcomes, and interference concern additional links between assignment, received treatment, observation, and the outcomes of different individuals; they are not removed by the initial randomization itself.
In an observational study, treatment assignment arises from behavioral, institutional, or biological processes rather than a controlled random device. Conditional exchangeability then represents the assumption that the measured covariates contain enough information to remove the relevant dependence between treatment and potential outcomes. Because potential outcomes under unreceived treatments are unobserved, this condition has no unrestricted empirical test.
Observed balance after adjustment is compatible with conditional exchangeability but does not establish it. Balance describes the distribution of measured covariates, whereas exchangeability concerns potential outcomes. An unmeasured common cause can preserve balance among measured variables while maintaining dependence between treatment and counterfactual outcomes.
Positivity and consistency
Exchangeability alone does not identify a causal effect. Identification by adjustment also requires positivity, which states that each treatment level of interest occurs with positive conditional probability in every relevant covariate stratum:
[ P(A=a\mid L=l)>0 ]
whenever (P(L=l)>0). Without positivity, exchangeability can hold while the data contain no observed outcomes under one treatment level for part of the target population.
Consistency supplies the connection between counterfactual and observed outcomes. It requires the observed outcome to equal the potential outcome corresponding to the treatment actually received. This condition also presupposes a sufficiently defined treatment intervention. If substantively different versions of a treatment are represented by one label and have different effects, a single potential outcome (Y^a) does not fully specify the intervention.
Together, exchangeability, positivity, and consistency identify the intervention distribution through
[ P(Y^a=y)
\sum_l P(Y=y\mid A=a,L=l)P(L=l). ]
The three conditions perform distinct roles. Exchangeability connects treated and untreated counterfactual distributions within covariate levels. Positivity supplies observable treatment comparisons within those levels. Consistency links the observed outcome to the appropriate potential outcome.
Sequential exchangeability
Longitudinal settings involve treatment histories (\bar A_t=(A_0,\ldots,A_t)) and covariate histories (\bar L_t=(L_0,\ldots,L_t)). Sequential exchangeability requires each treatment decision to be independent of the relevant future potential outcomes, conditional on the observed history available before that decision:
[ Y^{\bar a}\mathbin{\perp!!!\perp}A_t \mid \bar A_{t-1},\bar L_t. ]
This condition differs from baseline exchangeability because the covariate history can evolve in response to earlier treatment. A time-varying covariate may predict later treatment while also mediating part of an earlier treatment effect. Conventional regression adjustment for the covariate can then remove part of the effect under study or induce additional associations.
Under sequential exchangeability, longitudinal positivity, and consistency, the distribution of (Y^{\bar a}) is identified by the longitudinal g-formula. Inverse probability weighting and marginal structural models express the same identifying structure through a reweighted population in which the observed treatment process is independent of measured treatment history.
Partial and target-population exchangeability
Exchangeability can be defined relative to a particular causal estimand. Identification of the average treatment effect in the treated requires a counterfactual comparison for the treated population, rather than full symmetry across the entire source population. The relevant condition for a binary treatment includes
[ Y^0\mathbin{\perp!!!\perp}A\mid L, ]
because the missing counterfactual outcome among treated individuals is (Y^0). Independence involving (Y^1) is not separately required for identifying the treated group’s observed mean under treatment.
A related condition appears in transportability. If (S) indicates membership in a study sample, conditional exchangeability over selection can be represented as
[ Y^a\mathbin{\perp!!!\perp}S\mid L. ]
Combined with positivity of sample participation and an identified treatment effect within the study, this condition relates sample-specific causal results to a target population. It addresses selection into the study rather than assignment to treatment, although the mathematical form is parallel.