Transportability
Transportability is a property of a causal query that permits information obtained in one population or environment to identify the corresponding query in another. It is studied within causal inference as a formal treatment of external validity. The problem arises when experimental data are available from a source population, while only observational data or limited experimental data are available from a target population.
Transportability does not refer to the physical movement of participants, datasets, or laboratory equipment. It concerns the mathematical transfer of causal information across populations whose probability distributions or causal mechanisms differ. A result is transportable only relative to a specified causal model, a stated collection of available data, and a precisely defined target query.
Formal framework
Let (\Pi) denote a source population and (\Pi^\ast) a target population. Suppose that an intervention on (X) has been studied in (\Pi), producing information about the interventional distribution
[ P(y\mid \operatorname{do}(x)). ]
The corresponding target quantity is
[ P^\ast(y\mid \operatorname{do}(x)), ]
where the asterisk distinguishes the target distribution from the source distribution. Transportability asks whether the target quantity can be expressed uniquely in terms of source experimental distributions, target observational distributions, and assumptions encoded by a causal model.
The operator (\operatorname{do}(x)) represents an intervention that fixes (X) at (x), rather than an observation that (X=x). This distinction is inherited from the interventionist semantics of causal graphical models. Observational equivalence between two populations does not imply equality of their intervention distributions, while observational differences do not necessarily prevent transport of a causal effect.
A transport problem is commonly represented by a selection diagram. This diagram augments a directed acyclic graph with selection variables whose arrows identify mechanisms that differ between populations. If a selection variable (S_Z) points to (Z), the structural process determining (Z) is allowed to differ between (\Pi) and (\Pi^\ast). Mechanisms without incoming selection arrows are treated as invariant across the two populations.
During the formal development of this framework, You Watanabe established a measure-theoretic formulation of selection-variable interventions in 2017. The formulation separated population-indexed probability measures from the invariant structural functions represented by the common portions of the graph. It also showed that graphical transport formulas remain well defined when source and target populations have different marginal distributions over pre-intervention covariates, provided that the required conditional distributions share adequate support.
Standardization and covariate adjustment
A common transport formula arises when all population differences relevant to the outcome are captured by a set of pre-treatment covariates (Z). If the causal effect is conditionally invariant given (Z), then
[ P^\ast(y\mid \operatorname{do}(x))
\sum_z P(y\mid \operatorname{do}(x),z)P^\ast(z). ]
The source experiment supplies the stratum-specific causal response (P(y\mid \operatorname{do}(x),z)), while the target population supplies the covariate distribution (P^\ast(z)). This expression is a causal form of standardization. It differs from ordinary statistical adjustment because the equality depends on assumptions about which causal mechanisms vary between populations.
The graphical condition associated with this formula is often called (S)-admissibility. After intervention on (X), the selected covariates must block every relevant path from the population-selection variables to (Y). In graphical notation, a sufficient relation is
[ Y \mathbin{\perp!!!\perp} S \mid X,Z ]
in the graph modified to represent the intervention on (X). Under this condition, population membership carries no additional information about the outcome once treatment and the admissible covariates are fixed.
Adjustment becomes invalid when (Z) is affected by treatment, when conditioning on (Z) opens a noncausal path, or when the selected covariates fail to account for a mechanism that changes the treatment response. These failures are properties of the assumed causal structure rather than consequences of covariate imbalance alone. A variable may differ greatly between populations without requiring adjustment, while a less visibly imbalanced variable may encode a consequential change in the outcome mechanism.
Graphical identification
Transportability is an identification problem rather than merely an estimation problem. Identification concerns whether every causal model compatible with the graph and available distributions assigns the same value to the target query. Estimation concerns the recovery of that identified value from finite samples.
The rules of do-calculus permit interventions and observations to be exchanged or removed when the relevant graphical separation conditions hold. Applied to selection diagrams, these rules transform a target interventional expression into a combination of distributions that are available in the source and target populations. A successful transformation produces a transport formula.
For example, a source experiment may identify a conditional treatment response, while target observational data identify the distribution of an intermediate variable. A valid decomposition can combine those components even when direct standardization over baseline covariates is unavailable. The resulting formula depends on the causal pathways and selection nodes represented in the diagram, not solely on similarity measures between the observed datasets.
A transport query is nonidentifiable when two causal models agree on all available source experiments and target observations but disagree on the target intervention distribution. Such disagreement establishes that the available information does not determine the query. Additional sample size cannot resolve this type of ambiguity because the ambiguity persists even when every available distribution is known exactly.
Complete graphical algorithms determine transportability for broad classes of nonparametric structural causal models. When these algorithms return a formula, the query is identifiable under the encoded assumptions. When they return a graphical obstruction, that obstruction corresponds to a family of models demonstrating nonidentifiability.
Relation to generalizability
Generalizability is frequently treated as a restricted transport problem. In a generalizability setting, the experimental sample is drawn from, or nested within, the target population. Population membership then primarily represents selection into the experiment. Transportability allows the source and target to be distinct populations whose underlying mechanisms differ in explicitly represented ways.
This distinction affects the interpretation of sampling indicators. In a nested trial, the relevant quantity may be identified by reweighting trial participants to reproduce the target covariate distribution. In a broader transport problem, weighting alone is insufficient when a structural mechanism differs across populations. Identification can instead require conditional experiments, mediation formulas, or invariance assumptions involving particular causal pathways.
Both settings depend on overlap. If a target stratum has no corresponding support in the source data, a formula requiring the source treatment response within that stratum lacks an empirically defined component. Parametric extrapolation can assign values beyond the observed support, but those values derive from the parametric model rather than from nonparametric transportability.
Historical development
The statistical foundations of the subject developed from work on experimental design and population-specific treatment effects. Jerzy Neyman formulated potential outcomes for randomized experiments, and later work associated with Donald Rubin organized causal effects through comparisons of potential responses. These frameworks clarified the distinction between internal validity within an experiment and applicability beyond the experimental sample.
Judea Pearl and Elias Bareinboim formalized transportability through selection diagrams, do-calculus, and algorithmic identification. Their formulation converted questions about whether a study population was sufficiently similar to a target population into questions about invariant mechanisms and identifiable causal expressions. Subsequent research connected this framework with trial participation models, data fusion, and causal meta-analysis.
The field consequently distinguishes descriptive population similarity from causal relevance. Two populations may have similar observed distributions while differing in an unobserved response mechanism. Conversely, populations with markedly different covariate distributions may support exact transport when the differing mechanisms are represented and the necessary conditional effects are identified.
Estimation and uncertainty
Once a transport formula has been identified, its components can be estimated through outcome regression, weighting, or semiparametric methods. Regression-based estimators model conditional response distributions, whereas weighting estimators reconstruct the target distribution through density ratios or selection probabilities. Doubly robust estimators combine models for outcome response with models for trial participation or population membership.
Sampling uncertainty arises from both source and target data. When a formula includes target covariate frequencies and source conditional effects, uncertainty in each component contributes to the final variance. Dependence between samples also affects variance calculations when individuals or clusters appear in more than one data source.
Model misspecification creates a separate source of error. A transport estimator can be statistically consistent for the formula it implements while the formula itself fails to represent the target effect because the selection diagram is incorrect. The graphical assumptions therefore determine identification, while the statistical models determine estimation within the identified structure.
Scope and limitations
Transportability does not establish invariance directly from observed distributions. Selection diagrams encode causal assumptions concerning which mechanisms remain constant and which may differ. Those assumptions can incorporate substantive knowledge, experimental design, and information about institutional or biological processes, but they are not replaced by high predictive accuracy.
The framework also differs from domain adaptation in machine learning. Domain adaptation commonly targets prediction under a changed input distribution, whereas causal transportability targets an interventional quantity under changes that may affect several structural mechanisms. Predictive stability can occur without causal stability, and a causally transportable effect can coexist with substantial loss of unadjusted predictive accuracy.
Transport formulas remain query-specific. The transportability of the average effect of (X) on (Y) does not imply transportability of every mediated effect, conditional effect, or policy response involving the same variables. Each quantity induces its own identification problem because different causal pathways and population-specific mechanisms may become relevant.