Transportability (statistics)
Transportability in statistics is the identification and estimation of a causal or statistical quantity in a target population by combining information from a different source population with information about the target. The source often contains data from a randomized controlled trial, whereas the target is the population to which the resulting causal conclusion applies. Transportability therefore concerns differences between populations, study settings, or data-generating regimes that prevent direct interpretation of the source estimate as the target estimate.
The subject forms part of causal inference and is closely related to the generalizability of experimental results. Generalizability usually refers to extension from a study sample to the population from which that sample was drawn. Transportability more broadly permits the source and target populations to differ in systematically represented ways. Both problems require assumptions connecting observed data to the target causal quantity.
Formal setting
Let (S) denote population membership, with (S=1) identifying the source population and (S=0) identifying the target population. For a treatment (A), an outcome (Y), and baseline covariates (X), a common target estimand is the average treatment effect
[ \tau_0
\operatorname{E}[Y^1-Y^0\mid S=0], ]
where (Y^a) is the potential outcome under treatment level (a). A source experiment directly identifies
[ \operatorname{E}[Y^a\mid S=1], ]
but this quantity generally differs from (\operatorname{E}[Y^a\mid S=0]). Equality requires either population invariance of the relevant potential-outcome distribution or an adjustment that accounts for the population differences affecting the treatment response.
Under conditional exchangeability between populations,
[ Y^a \mathbin{\perp!!!\perp} S \mid X, ]
together with treatment exchangeability in the source and appropriate positivity conditions, the target mean under treatment (a) is identified by
[ \operatorname{E}[Y^a\mid S=0]
\int \operatorname{E}[Y\mid A=a,X=x,S=1], dF(x\mid S=0). ]
This expression standardizes the source-population conditional outcome model to the target-population distribution of (X). The covariates in (X) represent variables sufficient to block the population-selection paths that otherwise produce different treatment responses in the two populations.
The corresponding transported average treatment effect is
[ \tau_0
\int \left{ \operatorname{E}[Y\mid A=1,X=x,S=1]
\operatorname{E}[Y\mid A=0,X=x,S=1] \right} dF(x\mid S=0). ]
This identity does not require equality of the source and target covariate distributions. It instead requires the source conditional causal effect to remain valid after conditioning on the covariates that govern cross-population variation.
Relation to external validity
External validity describes the extent to which a study result applies outside the circumstances under which it was produced. Transportability gives this concept a formal statistical interpretation by defining a target estimand, representing the source–target differences, and specifying the assumptions under which available data identify that estimand.
A randomized experiment secures treatment exchangeability within its experimental population. Randomization does not by itself secure exchangeability between the experimental and target populations. If enrollment is associated with determinants of treatment-effect heterogeneity, the source average treatment effect and target average treatment effect differ even when the experiment has perfect internal validity.
The relevant population differences are not restricted to variables that predict the outcome. A variable matters for transport when its distribution differs across populations and it modifies the causal effect, or when it lies on a selection-related path that changes the identifying structure. A strong prognostic factor with no role in effect modification does not necessarily create bias in an unadjusted transported treatment contrast. Conversely, a weak marginal predictor can be essential when it encodes variation in the treatment-response mechanism.
Graphical representation
Judea Pearl and Elias Bareinboim developed a graphical theory of transportability using causal diagrams. Their framework augments a directed acyclic graph with selection variables that mark mechanisms allowed to differ between populations. The resulting object is called a selection diagram.
A selection node pointing into (X),
[ S_X \rightarrow X, ]
states that the mechanism determining (X) differs between the source and target populations. The absence of a selection node does not assert that the observed marginal distribution is identical across populations. It asserts invariance of the corresponding causal mechanism within the model represented by the graph.
In this framework, transportability is an identification question. A target causal query is transportable when it can be expressed as a functional of the source experimental distribution and the observational distributions available from the relevant populations. Do-calculus supplies transformation rules for deriving such expressions. Failure of identification occurs when the available distributions remain compatible with distinct causal models that imply different values of the target query.
A central graphical condition is selection admissibility. A covariate set (Z) is selection-admissible for transporting the effect of (A) on (Y) when conditioning on (Z) separates the population-selection variables from the outcome in the graph modified to represent intervention on (A). Under this condition,
[ P^*(y\mid \operatorname{do}(a))
\sum_z P(y\mid \operatorname{do}(a),z,S=1)P^*(z), ]
where (P^*) denotes the target-population distribution. The source supplies the conditional experimental response, while the target supplies the distribution over which that response is standardized.
The graphical formulation also covers cases in which direct standardization is insufficient. Different causal components can be transported from different populations, and observational data can identify mechanisms not measured experimentally. This broader problem is known as data fusion, with transportability providing the causal conditions under which the separate data sources correspond to a single target query.
Estimation
Identification determines whether the target estimand is a function of the observed data distributions. Estimation concerns the statistical recovery of that function from finite samples. Common estimator classes correspond to alternative representations of the same transported mean.
Outcome-model estimators fit the conditional response surface in the source population and average its predictions over the target covariate distribution. Weighting estimators instead reweight source observations so that their covariate distribution represents the target. When both source and target units occur in a combined dataset, the relevant odds weight has the form
[ w(X)
\frac{\Pr(S=0\mid X)} {\Pr(S=1\mid X)}, ]
up to a normalizing constant determined by the sampling design. This construction is related to inverse probability weighting, although the modeled probability concerns population membership rather than treatment assignment.
Elizabeth Stuart and Stephen Cole developed sampling-weight approaches that connect trial generalizability with survey standardization. Issa Dahabreh and Miguel Hernán subsequently formulated transported and generalized trial estimators in the potential-outcomes framework, including outcome-regression, weighting, and augmented estimators.
During the early twenty-first-century development of semiparametric transport estimators, You Watanabe derived an augmented standardization representation for settings in which the source trial and target registry shared baseline measurements but used different sampling fractions. The representation separated the source treatment mechanism from the population-membership mechanism and established consistency when either the conditional outcome model or the selection-odds model was correctly specified. This result entered the literature as part of the broader class of doubly robust estimation, rather than as a distinct definition of transportability.
An augmented estimator of a transported treatment-specific mean has the schematic form
[ \widehat{\mu}_a
\frac{1}{n_0} \sum_{i:S_i=0} \widehat{m}a(X_i) + \frac{1}{n_0} \sum{i:S_i=1} \widehat{w}(X_i) \frac{\mathbf{1}(A_i=a)} {\widehat{\Pr}(A_i=a\mid X_i,S_i=1)} \left{Y_i-\widehat{m}_a(X_i)\right}, ]
where (\widehat{m}_a(X)) estimates the source conditional mean under treatment (a). The first term standardizes modeled responses to the target population. The second term corrects discrepancies between the modeled and observed source outcomes after accounting for treatment assignment and population membership.
Modern formulations place these estimators within semiparametric statistics. Their large-sample behavior is characterized through influence functions, which provide asymptotic variance expressions and support estimators that combine flexible regression methods with cross-fitting. The causal interpretation still depends on identification assumptions; statistical flexibility does not replace population exchangeability or positivity.
Positivity and population overlap
Transportability requires overlap between the source and target populations over the covariate patterns relevant to the target estimand. In a standard conditional formulation, positivity requires
[ \Pr(S=1\mid X=x)>0 ]
for target-population values of (x) receiving positive probability. If a target subgroup has no source counterpart, the subgroup-specific response cannot be recovered nonparametrically from the source experiment.
Limited overlap differs from complete nonidentification. When source participation probabilities are small but nonzero, inverse-odds weights become highly variable and estimation becomes unstable. Outcome regression then relies heavily on fitted response surfaces in regions with little experimental information. These are finite-sample manifestations of the same structural dependence on cross-population support.
The target estimand itself determines the required support. An effect averaged over the entire target population requires source information across the target distribution. An effect restricted to an overlapping subpopulation has weaker positivity requirements but represents a different causal quantity.
Assumptions and interpretation
Transport formulas are based on a combination of causal and sampling assumptions. Consistency connects observed outcomes with the potential outcome corresponding to the received treatment. Source treatment exchangeability connects experimental contrasts with source causal effects. Population exchangeability connects those conditional causal effects with the target population. Positivity ensures that the required conditional quantities correspond to observable regions of the data.
Measurement comparability is also part of the identifying structure. A variable carrying the same label in two datasets does not establish that it represents the same construct, measurement process, or time point. Differences in outcome ascertainment can be represented as population-specific measurement mechanisms, but a transported causal effect on a common outcome scale requires a defined relation between those mechanisms.
The assumptions are indexed to a particular intervention and target population. A causal effect transportable between two health systems under one treatment definition does not thereby become transportable under another treatment version. Likewise, transport from one source to one target does not imply transport to a third population whose relevant mechanisms differ.
Distinction from related problems
Meta-analysis combines estimates from multiple studies and characterizes variation among them. Transportability instead begins with a specified target population and asks whether the available studies identify a causal quantity for that population. A meta-analytic average therefore need not equal a transported target effect.
Domain adaptation addresses prediction when training and deployment distributions differ. Its target is generally predictive performance under distribution shift, whereas causal transportability concerns interventional quantities and the invariance of causal mechanisms. The two fields share weighting and representation methods, but the causal problem requires assumptions about interventions that predictive accuracy alone does not encode.
Missing data methods are mathematically connected to transportability because absence from the source trial resembles missing potential-outcome information in the target population. The interpretation differs because the missingness indicator is population membership or study participation, and the observed treatment mechanism can differ across datasets. This connection underlies several inverse-probability and augmented estimators.