External Transportability Criterion

• External transportability criterion is a framework that specifies conditions for transferring causal effects from one population to another by ensuring covariate alignment and exchangeability.
• It is operationalized using methods such as regression, weighting, matching, and doubly robust estimators to adjust for differences between study and target populations.
• Diagnostic tools like overlap metrics and standardized mean differences help assess if measured effect modifiers support transportability, with failures arising from unmeasured confounding or positivity issues.

Searching arXiv for the core review and closely related transportability papers. The external transportability criterion denotes the set of conditions under which a causal effect identified in one population can be validly transported to another population. In the potential-outcomes formulation used in the review literature, it is the combination of internal validity assumptions with conditional exchangeability of potential outcomes across populations given measured covariates, positivity of selection, and SUTVA for study selection; under these conditions, the target Population Average Treatment Effect (PATE) is identified from study data, possibly combined with target-population covariate information (Degtiar et al., 2021). In the graphical literature, the same idea is expressed through selection diagrams: transport is licensed when the selection mechanism can be blocked by observed covariates and the resulting transport formula can be derived by do-calculus (Pearl et al., 2015).

1. Conceptual location within external validity

Within this literature, internal validity and external validity are distinct. Internal validity concerns whether a study estimate is unbiased for the causal effect in the population from which the study sample is a simple random sample; external validity concerns whether that internally valid estimate is unbiased for a causal effect in a different population. The review on generalizability and transportability treats external validity bias as a form of sample selection bias and divides it into generalizability, where the study population is a subset of the target population, and transportability, where the study population is at least partly external to the target population and may be disjoint from it (Degtiar et al., 2021).

The distinction between study population and target population is central. The study population is the hypothetical population for which the study sample is a simple random sample, given enrollment, inclusion and exclusion criteria, and post-enrollment missingness. The target population is the population about which causal inferences are ultimately desired. In this setup, the study directly identifies the Sample Average Treatment Effect,

$\tau_s = \mathbb{E}(Y^1 - Y^0 \mid S=1),$

whereas the target estimand is the Population Average Treatment Effect,

$\tau = \mathbb{E}(Y^1 - Y^0).$

When the distribution of effect modifiers differs between study and target populations, $\tau_s \neq \tau$, even if the study is internally valid. The external transportability criterion is the formal bridge between these two quantities (Degtiar et al., 2021).

This criterion matters only when treatment effects are heterogeneous in covariates whose distributions differ across populations. If the conditional treatment effect is constant across covariate strata, then external transportability is trivial because the study effect and target effect coincide. This suggests that transportability is not a separate problem from heterogeneity; it is a problem created by the interaction of heterogeneity and cross-population covariate shift (Degtiar et al., 2021).

2. Formal identification criterion

A standard data structure uses treatment $A \in \{0,1\}$, outcome $Y$, baseline covariates $X$, and a study-selection indicator $S \in \{0,1\}$, where $S=1$ denotes study participation and $S=0$ denotes target-population membership outside the study. The target PATE can then be written as

$$\tau = \mathbb{E}_X\!\left( \mathbb{E}(Y \mid S=1,A=1,X) - \mathbb{E}(Y \mid S=1,A=0,X) \right) = \mathbb{E}(Y^1-Y^0),$$

where the outer expectation is taken with respect to the target covariate distribution (Degtiar et al., 2021).

The identification criterion has two layers. The first is internal validity within the study population. The review states three standard conditions:

1. Conditional treatment exchangeability in the study:

$\tau = \mathbb{E}(Y^1 - Y^0).$0

2. Positivity of treatment assignment in the study:

$\tau = \mathbb{E}(Y^1 - Y^0).$1

3. SUTVA for treatment:

$\tau = \mathbb{E}(Y^1 - Y^0).$2

For PATE-oriented analysis, conditional treatment exchangeability may be weakened to mean exchangeability of the treatment effect in the study (Degtiar et al., 2021).

The second layer is the actual external transportability criterion. The review states a parallel trio of assumptions for selection:

4. Conditional exchangeability for study selection:

$\tau = \mathbb{E}(Y^1 - Y^0).$3

This requires that, within strata of $\tau = \mathbb{E}(Y^1 - Y^0).$4, being in the study rather than in the target population does not affect the distribution of potential outcomes. All effect modifiers that differ across populations must therefore be measured and included in $\tau = \mathbb{E}(Y^1 - Y^0).$5.

5. Positivity of selection. For generalizability,

$\tau = \mathbb{E}(Y^1 - Y^0).$6

For transportability between disjoint populations, the positivity requirement shifts: the propensity to be in the target population versus the study must be bounded away from 0 and 1, so that the selection mechanism does not perfectly separate the populations in covariate space.

6. SUTVA for study selection. The review interprets this as requiring no interference between being in the study and being in the target population, and no relevant differences in treatment implementation or outcome measurement across populations that are not encoded in $\tau = \mathbb{E}(Y^1 - Y^0).$7 (Degtiar et al., 2021).

Taken together, these assumptions identify the target causal effect from joint study and target data. The same review summarizes the criterion directly: internal validity assumptions plus conditional exchangeability of potential outcomes across populations given measured covariates, positivity of selection, and SUTVA for selection identify the PATE in the target population (Degtiar et al., 2021).

Under these conditions, identification admits both regression and weighting representations. The standardization form is

$\tau = \mathbb{E}(Y^1 - Y^0).$8

For weighting, the same review gives inverse-probability-of-participation and odds-of-participation formulas. For transportability to $\tau = \mathbb{E}(Y^1 - Y^0).$9,

$\tau_s \neq \tau$0

where $\tau_s \neq \tau$1 and $\tau_s \neq \tau$2. The weighting formula makes the criterion operational: study individuals are reweighted so that their covariate distribution matches that of the target population (Degtiar et al., 2021).

3. Diagnostics, heterogeneity, and failure modes

The literature emphasizes that transportability becomes substantively important when treatment effects are heterogeneous and the distribution of effect modifiers differs between study and target populations. The review therefore treats tests of treatment effect heterogeneity and tests of population difference as diagnostic criteria, not identification assumptions, for assessing whether external transportability assumptions matter in a given application (Degtiar et al., 2021).

Common diagnostics include baseline covariate comparisons, overlap diagnostics, and outcome-based checks. The review lists standardized mean differences,

$\tau_s \neq \tau$3

the distribution of the selection propensity $\tau_s \neq \tau$4, and Tipton’s generalizability index

$\tau_s \neq \tau$5

where $\tau_s \neq \tau$6 and $\tau_s \neq \tau$7 are the target and study proportions in propensity-score bins. In that framework, $\tau_s \neq \tau$8 indicates poor generalizability and $\tau_s \neq \tau$9 high generalizability. Outcome-based diagnostics compare observed outcomes in the target population with reweighted predictions from the study, or compare conditional outcomes between $A \in \{0,1\}$0 and $A \in \{0,1\}$1 given $A \in \{0,1\}$2 and treatment. Differences in these quantities suggest failure of conditional exchangeability for selection (Degtiar et al., 2021).

The principal failure modes are conceptually parallel to failures of internal validity. External transportability fails when there exist unmeasured effect modifiers $A \in \{0,1\}$3 such that $A \in \{0,1\}$4 affects $A \in \{0,1\}$5, differs in distribution across populations, and remains associated with $A \in \{0,1\}$6 after conditioning on observed $A \in \{0,1\}$7. It also fails under positivity violations, where some target strata of $A \in \{0,1\}$8 have no representation in the study, and under contextual differences not captured in $A \in \{0,1\}$9, such as differences in treatment implementation, setting, or outcome measurement. The review gives two canonical examples: cancer randomized trials with almost no African Americans, and attempts to transport findings from teaching hospitals to community clinics using a covariate like hospital type that can deterministically separate populations (Degtiar et al., 2021).

A subtle point concerns the covariate set itself. The review notes that, in transportability, variables that define the population boundary may need to be excluded from the exchangeability set $Y$0 if including them would destroy overlap. This is a positivity nuance rather than a rejection of covariate adjustment. It means that the bridge covariates must block selection-induced outcome differences without perfectly indexing the study-target partition (Degtiar et al., 2021).

4. Estimation strategies that operationalize the criterion

Once the criterion is accepted, estimation reduces to methods that align the study covariate distribution with the target covariate distribution or that model the outcome surface in the study and integrate it over the target population. The review surveys four broad classes: post-stratification, weighting or matching, outcome regression, and doubly robust estimators (Degtiar et al., 2021).

Post-stratification or standardization is the classical low-dimensional implementation. Covariates are partitioned into strata, treatment-specific means are estimated within each study stratum, and those means are then reweighted by target stratum frequencies. In that form, transportability is implemented directly by replacing the study distribution of $Y$1 with the target distribution (Degtiar et al., 2021).

Inverse Probability of Participation Weighting formalizes the same idea through a selection model. For transportability to $Y$2, the review gives normalized odds of participation weighting,

$Y$3

Under conditional exchangeability over selection and correct specification of $Y$4 and $Y$5, this produces a pseudo-population of study units whose covariate distribution matches that of the target population (Degtiar et al., 2021).

Matching methods, including full matching on the selection propensity and fine-balance methods, construct matched or reweighted study samples whose empirical covariate distribution aligns with the target. The inferential logic is the same as IPPW: matched outcomes from the study are interpreted as outcomes in the target under selection exchangeability (Degtiar et al., 2021).

Outcome regression directly models

$Y$6

and then plugs target covariate values into the fitted model. In this class, linear outcome models with treatment–covariate interactions and machine-learning regressions such as BART are used to approximate the conditional outcome surface. The transportability criterion enters through the requirement that the conditional outcome model is invariant across populations given $Y$7 (Degtiar et al., 2021).

Doubly robust estimators combine selection weighting and outcome regression. The review highlights augmented IPPW estimators, which remain consistent if either the outcome regression or the selection model is correctly specified, and TMLE constructions based on the efficient influence function. In the review’s formulation, these estimators explicitly combine outcome invariance across populations with covariate-distribution alignment, making them a direct semiparametric implementation of the criterion (Degtiar et al., 2021).

5. Graphical formulations and major extensions

In the graphical tradition, the criterion is expressed through selection diagrams, which augment a causal DAG with selection nodes $Y$8 pointing into mechanisms that may differ across populations. A target causal effect is transportable when do-calculus can reduce the target intervention expression to a combination of invariant experimental quantities from the source population and observational quantities from the target population. In the simplest sufficient case, if there exists a set $Y$9 such that

$X$0

then the $X$1-specific causal effect is directly transportable, and the target effect is obtained by averaging over the target distribution of $X$2 (Pearl et al., 2015).

Several later developments refine the criterion rather than replace it. One line extends transportability to cases where the source and target do not share the same covariate set. In “Efficient Generalization and Transportation,” the source uses a full covariate vector $X$3, while transportability across populations depends only on a common subset $X$4, under

$X$5

That work derives efficient influence functions and doubly robust estimators for generalization and transportation under this relaxed covariate-sharing structure (Zeng et al., 2023).

Another line studies relative-effect rather than difference-scale transportability. “Learning about treatment effects in a new target population under transportability assumptions for relative effect measures” and “Causal inference under transportability assumptions for conditional relative effect measures” assume equality of conditional relative effect measures across populations,

$X$6

and show how marginal target-population mean differences and ratios can be identified under that weaker scale-specific criterion (Dahabreh et al., 2022, Wang et al., 2024).

A different direction addresses settings where standard positivity fails. “Transportability without positivity: a synthesis of statistical and simulation modeling” keeps the target estimand fixed and replaces full empirical overlap with an overview of observed-data models in overlapping strata and externally specified simulation models for non-overlapping strata. This suggests a model-based extension of the criterion in which lack of support is compensated by explicit structural assumptions about extrapolation rather than by redefining the target population (Zivich et al., 2023).

The criterion has also been specialized to domain-specific settings. “Transportability of Principal Causal Effects” extends the framework to principal stratification, where transported estimands are indexed jointly by target-population membership and latent compliance type, and summarizes the corresponding criterion as

$X$7

plus overlap and internal principal-stratification assumptions (Clark et al., 2024). “Target Aggregate Data Adjustment” develops a two-stage weighting framework for survival outcomes when only target aggregate covariate moments are available, combining method-of-moments participation weights with inverse probability of censoring weights under the assumptions

$X$8

(Yan et al., 2024). “Constructing external comparator groups via transportability in mean or in effect measure” distinguishes transportability in mean

$X$9

from transportability in effect measure

$S \in \{0,1\}$0

using a common comparator treatment to identify external-comparator contrasts in the index population (Ung et al., 21 Apr 2026).

These extensions preserve the core logic of the external transportability criterion: identify a covariate set that blocks population differences relevant to the estimand, ensure overlap on that set, and then standardize, weight, or augment source-population information to match the target population.

6. Practical implementation and scope

The practical workflow begins by explicitly defining the study population, target population, and estimand. Different target populations imply different PATEs, so transportability is always relative to a specified target. The review recommends making the target operational, often via an external dataset or target sample, rather than describing it only verbally (Degtiar et al., 2021).

Design and measurement choices determine whether the criterion will later be plausible. Recommended steps include selecting eligibility criteria that cover the range of effect modifiers expected in the target, measuring all plausible effect modifiers that may differ across populations, and collecting at least some information on nonparticipants or external target samples. Diagnostics such as SMDs, propensity-score overlap, and outcome-based checks are then used to evaluate the plausibility of exchangeability and positivity. If overlap is poor, the analysis should explicitly acknowledge that extrapolation rather than strict transport is occurring (Degtiar et al., 2021).

Estimator choice depends on data structure and on which part of the criterion is most vulnerable. Weighting and matching are natural when overlap is adequate and the target sample contains covariates but not outcomes. Outcome regression is natural when high-dimensional modeling of treatment heterogeneity is feasible. Doubly robust and semiparametric methods are preferred when both model misspecification and efficiency are concerns. In several later developments, only target covariate summaries rather than individual-level target data are required, but the same causal criterion remains in force: selection exchangeability, positivity, and adequate nuisance-function estimation still govern validity (Yu et al., 2024).

When the criterion is not credible, the literature recommends changing the target population to one closer to the study, reporting the coverage of the original target population represented by the overlap region, or retaining the original target and explicitly treating conclusions as extrapolations likely to be biased. This suggests that the external transportability criterion is not merely an identification theorem; it is also a discipline for stating where causal claims are licensed and where they are not (Degtiar et al., 2021).